content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{sect:intro}
Graphene is a recently isolated material composed of carbon atoms arranged in a truly two-dimensional (2D) honeycomb lattice~\cite{geim_science_2009,castroneto_rmp_2009,geim_pt_2007,geim_natmat_2007,katsnelson_ssc_2007}. States near the Fermi energy of a graphene sheet are described by a massless Dirac equation which has chiral states in which the honeycomb-sublattice pseudospin is aligned
either parallel to or opposite to momentum. The Dirac-like wave equation and the existence of this spin-$1/2$-like quantum degree-of-freedom have a number of very intriguing implications on the properties of this material, most of which have been reviewed in the literature mentioned above.
Graphene has been shown to possess a wealth of tantalizing electronic, mechanical, and optical properties and might well become the material that will replace silicon in the next generation devices~\cite{avouris_natnano_2007}. Current exfoliated samples however suffer from a limited mobility, with typical values around $10.000-20.000~{\rm cm}^2/({\rm V} {\rm s})$: the main source of disorder which is behind these numbers is not yet completely understood and represents a substantial obstacle against the quest for fundamental physical effects and the development of functional devices. The mechanism which is limiting the mobility of the current (exfoliated) samples is actually one of the controversial topics
in this field of research. Two ``schools of thought" can be roughly identified: (i) one which ascribes the main limiting mechanism to charged impurities located in the (${\rm SiO}_2$) substrate~\cite{ando_jpsj_2006,nomura_prl_2006,nomura_prl_2007,hwang_prl_2007,adam_pnas_2007,adam_review_2009}, and (ii) one which instead relies on other scattering mechanisms, such as quenched ripples~\cite{katsnelson_ptrsA_2008}, which are also long-range in nature. Ripples have been seen in suspended membranes~\cite{meyer_nature_2007,bao_naturenanotech_2009} and also in flakes deposited on substrates~\cite{morozov_prl_2006, stolyarova_PNAS_2007,ishigami_nanolett_2007,geringer_prl_2009} and have been studied theoretically by Monte Carlo~\cite{fasolino_naturemat_2007,los_prb_2009} and molecular dynamics~\cite{abedpour_prb_2007,thompson_epl_2009} simulations.
The controversy is enriched by several experiments which have targeted the role of disorder in exfoliated samples~\cite{yacoby_natphys_2008,chen_natphys_2008,jang_prl_2008,bolotin_ssc_2008,du_naturenanotech_2008,ponomarenko_prl_2009,zhang_nature_2009,xia_natnano_2009,hong_prb_2009}. In particular, Bolotin {\it et al.}~\cite{bolotin_ssc_2008} and Du {\it et al.}~\cite{du_naturenanotech_2008} have observed a drastic increase in mobility in suspended samples, in agreement with a scenario in which charged impurities in the substrate are the main source of scattering. On the other hand, Ponomarenko {\it et al.}~\cite{ponomarenko_prl_2009} have studied exfoliated samples deposited on various substrates and found a rather weak dependence of the mobility on the type of substrate. The authors of Ref.~\onlinecite{ponomarenko_prl_2009} have also studied transport in flakes embedded in media with very high dielectric constants, such as glycerol, ethanol, and water, and measured only a small increase in mobility. This experimental study seems thus to suggest that charged impurities are not necessarily the primary source of scattering in current samples. Whatever the leading sources of disorder are, it is of utmost importance to understand how well or poorly these are screened by electrons in graphene.
The induced carrier density in graphene sheets subjected to the long-range potential of one or many charged impurities, in the absence or in the presence of electron-electron interactions, has been extensively studied theoretically~\cite{divincenzo_prb_1986,katnelson_prb_2006, cheianov_prl_2006,shytov_prl_2007,pereira_prl_2007, fogler_prb_2007,terekhov_prl_2008,polini_prb_2008,rossi_prl_2008,brey_prb_2009, fogler_prl_2009}: to the best of our knowledge, similar microscopic studies in the presence of corrugations have not yet appeared. The aim of this article is to cover this gap: we present extensive self-consistent fully-quantum-mechanical calculations of the electronic density profiles of massless Dirac fermions in the external scalar and vector potentials created by the corrugations. Our main findings can be summarized as follows: (i) the spatial density fluctuations induced by the ripples are almost entirely controlled by the scalar potential, especially in graphene sheets that are close to average neutrality; (ii) the contributions to the scalar and vector potentials due to in-plane atomic displacements are as large as those due to out-of-plane ones; and (iii) exchange and correlation contributions to the effective scalar (Kohn-Sham) potential seem to play a minor role in determining the shape of the ripple-induced electron-hole puddles, at least in the range of parameters we have analyzed.
This manuscript is organized as follows. In Sect.~\ref{sect:one} we discuss in detail how we have calculated scalar and vector potentials starting from a corrugated graphene sheet. In Sect.~\ref{sect:two} we introduce the theory and the numerical procedure we have used to calculate the induced carrier density in the presence of the ripple-induced potentials and present our main numerical results. Finally in Sect.~\ref{sect:three} we draw our main conclusions. Appendix~\ref{appendix} reports some technical remarks on the calculation of the density induced by a purely vector potential within linear-response theory.
\section{From ripples to scalar and vector potentials}
\label{sect:one}
The aim of this Section is to describe how we have computed the scalar and vector potentials associated with ripples. For definiteness we focus our attention on ripples generated by thermal fluctuations~\cite{fasolino_naturemat_2007,los_prb_2009,abedpour_prb_2007}. The procedure we have followed is however completely general and applies to any type of ripples, independently of the microscopic, intrinsic or extrinsic, mechanisms that lie at their origin.
\subsection{Microscopic calculation of the average displacements}
\label{sect:onea}
\begin{figure}
\begin{center}
\includegraphics[width=0.95\linewidth]{fig01}
\caption{(Color online) Three-dimensional plot of the corrugated graphene sample used to calculate the average displacements shown in Fig.~\ref{fig:displmesh} and the scalar and vector potentials shown in Fig.~\ref{fig:andopot}. \label{fig:sample}}
\end{center}
\end{figure}
In what follows we consider a specific realization of a corrugated graphene sheet at a temperature $T = 300$~K,
computed with a Monte Carlo simulation as in Ref.~\onlinecite{los_prb_2009}. In Fig.~\ref{fig:sample} we show the three-dimensional bond structure of the sample, which contains 19504 atoms and fulfills periodic boundary conditions in the simulation box.
The computation of the corrugation-induced scalar and vector potentials that we will carry out in Sect.~\ref{sect:oneb} below
requires the knowledge of the displacements $\{ {\bm u}_{i} \}$ of the atomic positions $\{{\bm r}'_{i}\}$ in the sample ($i$ is the atomic label) with respect to a flat reference distribution $\{{\bm r}_{i}\}$.
The latter is defined by applying a dilation/contraction to the honeycomb lattice at $T=0$.
More precisely, we first make sure that the positions, ${\bm r}_{\rm CM}$ and ${\bm r}'_{\rm CM}$, of the center-of-mass of the two distributions coincide, and use in the following the displaced vectors ${\bm r} \to {\bm r} - {\bm r}_{\rm CM}$.
We then dilate/contract the honeycomb lattice at $T=0$ to compensate for the variation of the carbon-carbon bond length
produced by the finite temperature.
The coefficient $\lambda$ in the transformation ${\bm r} \to \lambda {\bm r} $ is obtained by averaging the ratio
$\lambda_{i} = |{\bm r}_{i}'| / |{\bm r}_{i}|$ over all the atoms $i$ such that $|{\bm r}_{i}| > 50.0$~\AA.
The latter restriction reduces the impact of the fluctuations of the atomic positions, produced by the ripples, but does not affect the computation of the overall stretch/compression produced by the temperature.
We find $\lambda \simeq 0.998$ ($<1$: the effect of temperature in this range is indeed to {\it reduce} the carbon-carbon bond length~\cite{zakharchenko_prl_2009}). The variance of $\{ \lambda_{i} \}$ is of order $10^{-3}$, hence the stretch induced by the temperature is the dominant contribution of the atomic displacements from the positions of the bare honeycomb lattice.
In other words, to prepare a sensible reference distribution it is essential to perform the aforementioned stretch, even if the factor $\lambda$ is close to unity.
Finally, we make sure that the sample and the reference distribution are not globally rotated with respect to each other.
We compute the average angular displacement vector
\begin{equation}
{\bm \phi} = \frac{1}{N_{\phi}}\sum_{i} \arccos{\left ( \frac{{\bm r}_{i}' \cdot {\bm r}_{i}}{|{\bm r}'_{i}| |{\bm r}_{i}|} \right ) }
\frac{{\bm r}_{i}' \times {\bm r}_{i}}{| {\bm r}_{i}' \times {\bm r}_{i} |}~,
\end{equation}
where the summation is restricted to the $N_{\phi}$ atoms such that the cosine of the angle between ${\bm r}_{i}'$ and ${\bm r}_{i}$ is larger than $0.9$. In the analyzed sample the modulus of ${\bm \phi}$ is of order $10^{-3}$, hence we conclude that the sample and the reference distribution are properly aligned. We are now in position to compute the displacement vectors ${\bm u}_{i} = {\bm r}_{i}' - {\bm r}_{i}$: thanks to the above mentioned preparation procedures, these will be free of artificial systematic trends and will provide us with an accurate local description of the ripples.
As we solve for the electronic density on a square mesh in the simulation box (see the description of the method in Sec.~\ref{sect:twob}), the knowledge of the displacement of each atom is superabundant. For this reason we average the atomic displacements over square patches
defined on a square mesh. To show that this averaging yields indeed a correct modeling of the physical system, we observe that the problem possesses three length scales: (i) graphene's lattice constant $a = a_0\sqrt{3} \approx 0.25~{\rm nm}$ (here $a_0 = 1.42$~\AA~is the carbon-carbon distance), (ii) the length scale $\lambda_{\rm s}$ of the spatial structures in the specific sample shown in Fig.~\ref{fig:sample}, which is of the order of several nanometers ($\lambda_{\rm s} \approx 8~{\rm nm}$); and (iii) the spatial resolution $\lambda_{\rm res}$ which we have in our continuum-model electronic structure calculations [see Eq.~(\ref{eq:spa_resol})]. For a sample of roughly $22~{\rm nm} \times 22~{\rm nm}$, as the one shown in Fig.~\ref{fig:sample}, $\lambda_{\rm res} \approx 1.5~{\rm nm}$ (see discussion below in Sect.~\ref{sect:twob}).
Since $\lambda_{\rm s} \gg \lambda_{\rm res} \gg a$, the structures in Fig.~\ref{fig:sample} are properly resolved by the mean
displacement vectors ${\bar {\bm u}}({\bm r})$, obtained by averaging the microscopic displacements over square patches of area $\approx \lambda^2_{\rm res}$. The result of this averaging procedure for the sample in Fig.~\ref{fig:sample} is shown in Fig.~\ref{fig:displmesh} where we have plotted ${\bar {\bm u}}({\bm r})$ as calculated on a square mesh with $32 \times 32$ points.
We remark that the in-plane displacements undergo strong variations between neighboring patches as a consequence of the fact that even the in-plane displacements of neighboring atoms in the sample do not present signatures of local correlations.
\begin{figure}
\begin{center}
\includegraphics{fig02}
\caption{\label{fig:displmesh}
(Color online) Average displacements ${\bar {\bm u}}({\bm r})$ calculated as discussed in Sect.~\ref{sect:onea}.
The color scale represents the ${\hat {\bm z}}$ component of the average displacements, varying from $-3.0$~\AA~(blue) to $+3.0$~\AA~(red). The arrows, whose length has been multiplied by a factor ten for better visibility, represent the in-plane components of the average displacements.}
\end{center}
\end{figure}
We now proceed to discuss how we have calculated the deformation tensor and the corrugation-induced scalar and vector potentials.
\subsection{The deformation tensor and the corrugation-induced scalar and vector potentials}
\label{sect:oneb}
We have calculated scalar $V_1$ and vector $V_2 = A_x - i A_y$ potentials according to the standard formulas of the theory of elasticity~\cite{ando_prb_2002,manes_prb_2007}:
\begin{equation}\label{eq:V1}
V_1 = g_1 (u_{xx} + u_{yy})
\end{equation}
and
\begin{equation}\label{eq:V2}
V_2 = g_2 (u_{xx} - u_{yy} + 2 i u_{xy})~,
\end{equation}
where $u_{ij}$ (with $i,j \in \{x,y\}$) is the usual deformation tensor,
\begin{equation}\label{eq:deformation}
u_{ij} = \frac{1}{2} \left(\frac{\partial {\bar u}_i}{\partial x_j} + \frac{\partial {\bar u}_j}{\partial x_i} + \sum_{k \in \{x,y,z\}}
\frac{\partial {\bar u}_k}{\partial x_i} \frac{\partial {\bar u}_k}{\partial x_j}\right)~.
\end{equation}
Here ${\bar u}_i = {\bar u}_i({\bm r})$ with $i \in \{x,y,z\}$ are the Cartesian components of the average displacements. For the coupling constant $g_1$ we have used two values, $g_1 = 3~{\rm eV}$ and $g_1 = 16~{\rm eV}$ (the latter value~\cite{sugihara_prb_1983,ando_prb_2002}, which is based on old transport data on graphite sample,
seems largely overestimated~\cite{guinea_prb_2010}), while
\begin{equation}\label{eq:g2}
g_2 = \frac{3 \kappa \beta}{4}\gamma_0~,
\end{equation}
where $\beta = - \partial \log{(\gamma_0)}/\partial{\log (a_0)} \approx 2$, $\gamma_0 \approx 2.7~{\rm eV}$ is the
nearest-neighbour hopping parameter, and
\begin{equation}\label{eq:kappa}
\kappa \equiv \frac{1}{\sqrt{2}} \frac{\mu_{\rm s}}{B}~.
\end{equation}
For the shear $\mu_{\rm s}$ and bulk $B$ moduli we have used the recently calculated values~\cite{zakharchenko_prl_2009},
$\mu_{\rm s} = 9.95~{\rm eV}$~\AA$^{-2}$ and $B= 12.52~{\rm eV}$~\AA$^{-2}$, at a temperature $T = 300~{\rm K}$.
We thus find that $\kappa \approx 0.56$ at this temperature.
In Fig.~\ref{fig:andopot} we illustrate scalar and vector potentials calculated using Eqs.~(\ref{eq:V1})-(\ref{eq:kappa}) above. While performing the calculation of $V_1$ and $V_2$ we have noticed that the derivatives of the average in-plane displacements ${\bar {\bm u}}_\perp$ are of ${\cal O}(10^{-2})$, while the derivatives of the out-of-plane displacements ${\bar u}_z$ are much bigger, ${\cal O}(10^{-1})$. However, in the deformation tensor (\ref{eq:deformation}) the latter enter only {\it quadratically}. We thus conclude that the contributions from in-plane and out-of-plane displacements are both of the same order, ${\cal O}(10^{-1})$. As a result, no evident correlations link the out-of-plane {\it topographic} corrugations [{\it i.e.} the distribution of the out-of-plane average displacements ${\bar u}_{z}({\bm r})$ shown in the color map in Fig.~\ref{fig:displmesh}] with the scalar and vector potentials illustrated in Fig.~\ref{fig:andopot}.
\begin{figure*}
\begin{center}
\begin{tabular}{c c c}
\includegraphics[width=0.33\linewidth]{fig03a} &
\includegraphics [width=0.33\linewidth]{fig03b} &
\includegraphics [width=0.33\linewidth]{fig03c}
\end{tabular}
\caption{(Color online) Left panel: color plot of the scalar potential $V_1({\bm r})$ (in units of meV) calculated using Eq.~(\ref{eq:V1}) with $g_1 = 3~{\rm eV}$. Central panel: the real part of the potential $V_2({\bm r})$ (in units of meV) calculated using Eq.~(\ref{eq:V2}). Right panel: the imaginary part of the potential $V_2({\bm r})$ (in units of meV).\label{fig:andopot}}
\end{center}
\end{figure*}
\section{Kohn-Sham-Dirac density-functional calculations}
\label{sect:two}
In this Section we present an approximate self-consistent microscopic theory for the carrier density distribution in the corrugation-induced scalar and vector potentials shown in Fig.~\ref{fig:andopot}.
\subsection{Approximate Kohn-Sham-Dirac theory for corrugated graphene sheets}
\label{sect:twoa}
We have generalized the Kohn-Sham-Dirac (KSD) theory described in Ref.~\onlinecite{polini_prb_2008} to deal with situations
in which the massless Dirac fermion liquid is subjected to a space-dependent vector potential ${\bm A}({\bm r})$
(the vector potential introduced below has the physical dimensions of energy) which changes smoothly over many lattice constants.
In this limit the induced density $\delta n({\bm r})$ can be calculated by solving the following single-spin single-valley KSD equation:
\begin{equation}\label{eq:ksd}
\left\{{\bm \sigma}\cdot [v{\bm p} + {\bm A}({\bm r})] + \openone_{\sigma} V_{\rm K S}({\bm r}) \right\}\Phi_\lambda({\bm r})=\varepsilon_{\lambda}\Phi_\lambda({\bm r})~.
\end{equation}
Here ${\bm \sigma}$ is a 2D vector constructed with the $2 \times 2$ Pauli matrices $\sigma_1$ and $\sigma_2$ acting in pseudospin space, $v = 3 \gamma_0 a_0/(2 \hbar) \approx 10^{6}~{\rm m}/{\rm s}$ is the bare Fermi velocity, ${\bm p}=-i\hbar \nabla_{\bm r}$, $\openone_{\sigma}$ is the $2 \times 2$ identity matrix in pseudospin space, and the Kohn-Sham potential,
\begin{equation}\label{eq:kspot}
V_{\rm KS}({\bm r}) = V_{\rm ext}({\bm r}) + \Delta V_{\rm H}({\bm r}) + V_{\rm xc}({\bm r})~,
\end{equation}
is the sum of the external scalar potential $V_{\rm ext}({\bm r})$, the Hartree potential, and the scalar exchange-correlation potential. For
${\bm A} = {\bm 0}$ Eq.~(\ref{eq:ksd}) reduces to the KSD equation introduced in Ref.~\onlinecite{polini_prb_2008}.
Note that Eq.~(\ref{eq:ksd}) neglects exchange-correlation corrections to the vector potential~\cite{vignale_prl_1987} ${\bm A}$, which are beyond the scope of the present paper and which will be addressed in a subsequent publication.
The ground-state density $n({\bm r})$ is obtained as a sum over the KSD spinors $\Phi_\lambda({\bm r})$:
\begin{eqnarray}\label{eq:density}
n({\bm r}) = g \sum_{\lambda}[|\varphi^{(A)}_{\lambda}({\bm r})|^2+|\varphi^{(B)}_{\lambda}({\bm r})|^2]f(\varepsilon_\lambda)~,
\end{eqnarray}
where the factor $g = g_{\rm s}g_{\rm v} = 4$ is due to valley and spin degeneracies, $\{\varphi^{(\sigma)}_{\lambda}({\bm r}), \sigma=A,B\}$ are the pseudospin (sublattice) components of the spinor $\Phi_\lambda({\bm r})$, and $f(x) = \{\exp{[(x-\mu)/(k_{\rm B} T)]}+1\}^{-1}$ is the usual Fermi-Dirac thermal factor at a chemical potential $\mu = \mu(T)$. Equation (\ref{eq:density}) is a self-consistent closure relationship for the KSD equation (\ref{eq:ksd}), since the Kohn-Sham potential $V_{\rm KS}({\bm r})$ is a functional of the ground-state density $n({\bm r})$.
In the absence of any source of external scalar and magnetic fields, the scalar $V_{\rm ext}({\bm r})$ and vector ${\bm A}({\bm r})$ potentials are solely determined by the corrugations:
\begin{equation}
\left\{
\begin{array}{l}
V_{\rm ext}({\bm r}) = V_1({\bm r})\vspace{0.1 cm} \\
{\bm A}({\bm r}) = (\Re e~V_2({\bm r}), -\Im m~V_2({\bm r}))
\end{array}
\right.~.
\end{equation}
The Hartree potential is given by
\begin{equation}\label{eq:hartree}
\Delta V_{\rm H}({\bm r})=\int d^2{\bm r}'\frac{e^2}{\epsilon|{\bm r}-{\bm r}'|} \; \delta n({\bm r}')~,
\end{equation}
where $\epsilon$ is an average dielectric constant
\begin{equation}
\epsilon = \frac{\epsilon_1 + \epsilon_2}{2}~.
\end{equation}
Here $\epsilon_1$ and $\epsilon_2$ are the dielectric constants of the media above and below the graphene flake.
For example $\epsilon \approx 2.5$ for graphene placed on ${\rm SiO}_2$ with the other side being exposed to air, while
$\epsilon \approx 1$ for suspended graphene. The quantity $\delta n({\bm r})=n({\bm r})-n_0$ is the local density measured relative to a ``background" value, $n_0$, which is defined by
\begin{equation}\label{eq:average}
n_0=\frac{2}{{\cal A}_0}+ {\bar n}_{\rm c}~.
\end{equation}
Here $2/{\cal A}_0$ is the density of a neutral graphene sheet,
${\cal A}_0=3\sqrt{3} a^2_0/2 \sim 0.052~{\rm nm}^2$ being the area of the unit cell in the honeycomb lattice,
and ${\bar n}_{\rm c}$ is the spatially averaged carrier density, which can be positive or negative and controlled by gate voltages.
The third term in $V_{\rm KS}({\bm r})$, $V_{\rm xc}({\bm r})$, is the scalar exchange-correlation potential.
This is a functional of the ground-state density, which is known only approximately.
Following Ref.~\onlinecite{polini_prb_2008} we employ the local-density approximation (LDA),
\begin{eqnarray}\label{eq:lda}
V_{\rm xc}({\bm r})&=&\left.v^{\rm hom}_{\rm xc}(n)\right|_{n \to n_{\rm c}({\bm r})}~,
\end{eqnarray}
where $v^{\rm hom}_{\rm xc}(n)$ is the $T=0$ exchange-correlation potential of a uniform 2D liquid of
massless Dirac fermions~\cite{polini_prb_2008,barlas_prl_2007} with carrier density $n$.
$v^{\rm hom}_{\rm xc}(n)$ is related to the ground-state energy per excess carrier
$\delta \varepsilon_{\rm xc}(n)$ by
\begin{equation}
v^{\rm hom}_{\rm xc}(n)=\frac{\partial [n \delta\varepsilon_{\rm xc}(n)]}{\partial n}~.
\end{equation}
The carrier density $n_{\rm c}({\bm r})$ is the density relative to that of a uniform {\it neutral} graphene sheet:
\begin{equation}
n_{\rm c}({\bm r}) \equiv n({\bm r}) - \frac{2}{{\cal A}_0} = \bar n_{\rm c} + \delta n({\bm r})~.
\end{equation}
The expression used for
$\delta \varepsilon_{\rm xc}(n)$ depends on the zero-of-energy, which is normally~\cite{barlas_prl_2007} chosen so that
$v^{\rm hom}_{\rm xc}(n=0)=0$.
\subsection{Technical remarks on the method of solution}
\label{sect:twob}
In order to solve Eq.~(\ref{eq:ksd}) we have followed the same technique adopted in Ref.~\onlinecite{polini_prb_2008}, {\it i.e.} we use a square simulation box of size $L\times L$ with periodic boundary conditions and conveniently expand the spinors $\Phi_{\lambda}(\bm r)$ in a plane-wave basis. We discretize real space restricting ${\bm r}$ to a square mesh ${\bm r}_{i j} = (i \delta, j \delta )$, with $i,j = 1,\dots,N$. Here $\delta = L/N$ is the spacing of the mesh. Fourier transforms ${\widetilde f}({\bm k})$ of real-space functions $f({\bm r})$ are calculated by means of a standard fast-Fourier-transform algorithm~\cite{website} that allows us to compute ${\widetilde f}$ on the set of discrete wavevectors ${\bm k}_{ij}$,
\begin{equation}
{\bm k}_{ij}= (k_{x,i},k_{y,j})
= \frac{2\pi}{L}~(n_{x,i}, n_{y,j})~,
\end{equation}
with $-N/2 \le n_{x,i}, n_{y,j} < N/2 $ (or, equivalently, $0 \le n_{x,i}, n_{y,j} < N$).
In momentum space Eq.~(\ref{eq:ksd}) reads
\begin{equation}\label{eq:ksd_momentum}
\sum_{{\bm k}'}\langle {\bm k} | \{ {\bm \sigma}\cdot [ v{\bm p} + {\bm A}({\bm r})] + {\mathbb I}_\sigma V_{\rm KS}({\bm r})\} |{\bm k}'\rangle {\widetilde \Phi}_{\lambda}({\bm k}')
=\varepsilon_{\lambda} {\widetilde \Phi}_{\lambda}({\bm k})~,
\end{equation}
and the problem is thus mapped into the diagonalization of the KSD matrix ${\cal H}^{\rm KSD}_{{\bm k},{\bm k}'}\equiv \langle {\bm k} |\{{\bm \sigma}\cdot [ v{\bm p} + {\bm A}({\bm r})] + {\mathbb I}_\sigma
V_{\rm KS}({\bm r})\}|{\bm k}'\rangle$.
The matrix elements in Eq.~(\ref{eq:ksd_momentum}) can be computed either analytically or numerically. More specifically, the matrix elements of the kinetic Hamiltonian are given by
\begin{eqnarray}
\langle {\bm k}|~v {\bm \sigma} \cdot {\bm p}~|{\bm k}'\rangle &=&
\hbar v{\bm \sigma} \cdot {\bm k}' \delta_{{\bm k}, {\bm k}'}~.
\end{eqnarray}
The matrix elements of the Hartree term are given by
\begin{eqnarray}\label{eq:hartree_momentumspace}
\langle{\bm k}| \Delta V_{\rm H}({\bm r}) |{\bm k}'\rangle=
\frac{2\pi e^{2}}{\epsilon |{\bm{k}-\bm{k}'}|}\; \delta {\widetilde n}({\bm k} - {\bm k}')~,
\end{eqnarray}
where $\delta {\widetilde n}({\bm k})={\widetilde n}({\bm k})-n_0\delta_{{\bm k}, {\bm 0}}$ is the Fourier transform of the charge neutral density $\delta n({\bm r})$, introduced above.
The matrix elements of the external, vector, and exchange-correlation potentials can be calculated numerically from
\begin{equation}\label{eq:lda_momentum_space}
\langle{\bm k}| f({\bm r}) |{\bm k}'\rangle =\frac{1}{L^2}
\int d^2 {\bm r}~f({\bm r})~e^{-i({\bm k}-{\bm k}')\cdot {\bm r}}~,
\end{equation}
where $f({\bm r})$ is either $V_{\rm ext}({\bm r})$, $V_{\rm xc}({\bm r})$, $A_{x}({\bm r})$, or $A_{y}({\bm r})$.
In practice the diagonalization of the KSD matrix ${\cal H}^{\rm KSD}_{{\bm k},{\bm k}'}$ requires the introduction of a momentum space cut-off~\cite{polini_prb_2008}, $k_{x,i},k_{y,j} \in [-k_{\rm c},+k_{\rm c}]$, which does not exceed the Brillouin-zone boundary defined by our real-space discretization: $k_{\rm c} < \pi /\delta$. $k_{\rm c}$ defines the range of momenta used in the expansion of the Hamiltonian ${\cal H}^{\rm KSD}_{{\bm k},{\bm k}'}$ and thus defines its dimension $d_{\rm H}$:
\begin{equation}
d_{\rm H} = 2\times \left(2\times \frac{Lk_{\rm c}}{2\pi}+1\right)^{2}~.
\end{equation}
The factor of $2$ here is due to the sublattice pseudospin degree-of-freedom.
Given a value of $k_{\rm c}$ the Kohn-Sham-Dirac matrix ${\cal H}^{\rm KSD}_{{\bm k},{\bm k}'}$
has $d_{\rm H}$ eigenvalues, labeled by the discrete index $\lambda=1,\dots,d_{\rm H}$.
Let us consider a neutral-on-average graphene sheet (${\bar n}_{\rm c}=0$) with areal extension $L\times L$.
The total number of electrons in such sheet is
\begin{equation}\label{eq:Nreal}
N_{\rm real} = \frac{2}{{\cal A}_0}\times L^{2}~.
\end{equation}
The total number of electronic states available in our calculations is $g d_{\rm H}$. To simulate a neutral-on-average sheet we clearly need half of these states:
\begin{equation}\label{eq:Nsimul}
N_{\rm simul}=\frac{1}{2}\times g d_{\rm H} = g \times \left(2 \times \frac{Lk_{\rm c}}{2\pi}+1\right)^2~.
\end{equation}
In Ref.~\onlinecite{polini_prb_2008} the authors enforced the following condition
\begin{equation}\label{eq:constraint}
N_{\rm simul} = N_{\rm real}~,
\end{equation}
which physically means that all the electrons in the $\pi$-band are simulated.
This leads to the relation $2 L^2/{\cal A}_0=g~[2 Lk_{\rm c}/(2\pi)+1]^2$
which links the momentum-space cut-off $k_{\rm c}$ and the size of the system $L$. This relationship is however too restrictive since one would need very large values of $k_{\rm c}$ (much larger than those prescribed by the computational limit) to simulate flakes with an areal extension of experimental interest~\cite{note_kc}. Therefore, the requirement (\ref{eq:constraint}) severely affects the possibility of performing quantitative predictions for large systems.
There are also more physical reasons for lifting the requirement (\ref{eq:constraint}): the massless Dirac fermion model~\cite{castroneto_rmp_2009} does not describe all electrons in the $\pi$-bands but only a fraction $\eta' \ll 1$ of them.
We thus have decided to relax the constraint (\ref{eq:constraint}) allowing $N_{\rm simul} \neq N_{\rm real}$, {\it i.e.}
\begin{equation}\label{eq:relaxed_constraint}
N_{\rm simul} = \eta'~N_{\rm real}
\end{equation}
with $0 < \eta'\ll 1$. Letting $\eta'$ be different from unity we can choose $L$ and $k_{\rm c}$ independently.
The factor $\eta'$ can be tuned in order to fulfill Eqs.~(\ref{eq:Nreal}), (\ref{eq:Nsimul}), and (\ref{eq:relaxed_constraint}):
\begin{equation}\label{eq:etapr}
\eta' = \frac{g d_{\rm H}}{4}\ \frac{{\cal A}_0}{L^2} = g~[2 Lk_{\rm c}/(2\pi)+1]^2~\frac{{\cal A}_0}{2 L^2}~.
\end{equation}
For example, we can choose $L \approx 22~{\rm nm}$ (as in the case of Fig.~\ref{fig:sample}) and fix
$k_{\rm c}$ according to our numerical capabilities, say $k_{\rm c} = 15 \times (2\pi/L)$.
Substituting these values for $L$ and $k_{\rm c}$ in Eq.~(\ref{eq:etapr}), one obtains that the fraction of simulated electrons in this case is $\eta' \approx 0.2$, {\it i.e.} $20\%$ of the electrons in graphene's $\pi$-band.
We remark that the existence of a momentum space cut-off $k_{\rm c}$ implies a minimum spatial resolution,
\begin{equation}\label{eq:spa_resol}
\lambda_{\rm res} = \frac{2\pi}{k_{\rm c}}~,
\end{equation}
which in this case would be $\lambda_{\rm res} \sim 1.5~{\rm nm}$, and thus sufficient to resolve rather short-wavelength
spatial structures in the induced carrier density.
The arguments above can be readily generalized to the case of a doped graphene sheet (${\bar n}_{\rm c} \neq 0$): in this case Eq.~(\ref{eq:relaxed_constraint}) reads
\begin{equation}
N_{\rm simul} = \frac{g d_{\rm H}}{2} + {\bar n}_{\rm c} L^2 = \eta'~2 \frac{L^2}{{\cal A}_0} + {\bar n}_{\rm c} L^2~.
\end{equation}
We clearly see that even at finite doping we can arbitrarily choose $L$ and $k_{\rm c}$, with a fraction of simulated electrons which is still
given by Eq.~(\ref{eq:etapr}).
Before concluding this Section we recall that the exchange and correlation potential $v^{\rm hom}_{\rm xc}(n)$ introduced in Sect.~\ref{sect:twob} depends on carrier density ${\bar n}_{\rm c}$ through the dimensionless quantity~\cite{barlas_prl_2007}
$\Lambda = k_{\rm max}/k_{\rm F}$, where $k_{\rm max}$ is an ultraviolet cut-off and
$k_{\rm F} = \sqrt{4 \pi |{\bar n}_{\rm c}|/g}$ is the Fermi wave number. We take $k_{\rm max}$ to be such that
\begin{equation}\label{eq:kmax}
\pi k_{\rm max}^2 = \eta \frac{8\pi^2}{g{\cal A}_0}~,
\end{equation}
where $\eta$ is a dimensionless number, $0<\eta\leq 1$, which should be assigned a value according to the wave vector range over which the continuum model describes graphene~\cite{factor2}. Thus, making use of Eqs.~(\ref{eq:etapr}) and (\ref{eq:kmax}), we find
\begin{equation}
\Lambda = \sqrt{\frac{2\eta}{{\cal A}_0 |\bar n_{\rm c}|}} =
\sqrt{\frac{\eta}{\eta'}}~\sqrt{\frac{g d_{\rm H}}{2 |\bar n_{\rm c}| L^2}}~.
\end{equation}
However, it is physically reasonable to identify $\eta$ and $\eta'$ since they both refer, directly or indirectly, to the range of applicability of the massless Dirac fermion model to describe electrons in graphene. Consequently, we see that, taking $\eta=\eta'$, $\Lambda$ is independent of the choice of $\eta$ while it depends on ${\bar n}_{\rm c} L^2$, {\it i.e.} on the average carrier density in units of $1/L^2$, and on the dimension $d_{\rm H}$ of the KSD Hamiltonian (or equivalently on $k_{\rm c}$).
\subsection{Numerical results}
\label{sect:twoc}
In Fig.~\ref{fig:g13}-\ref{fig:densityonripples} we report our main numerical results obtained from the self-consistent solution of the KSD
equation (\ref{eq:ksd}) with a momentum-space cut-off $k_{\rm c} = 15 \times (2\pi/L)$. The induced density profiles depend on the strength of electron-electron interactions which is measured by the dimensionless fine-structure constant
\begin{equation}
\alpha_{\rm ee} = \frac{e^2}{\epsilon \hbar v}~.
\end{equation}
In Fig.~\ref{fig:g13} we illustrate the fully self-consistent electronic density profile $\delta n({\bm r})$ in the ripple-induced
scalar and vector potentials shown in Fig.~\ref{fig:andopot}. By ``fully self-consistent" we mean that $\delta n({\bm r})$ has been obtained with the inclusion of {\it both} Hartree and scalar LDA exchange-correlation potentials. In this figure we have reported results for two values of graphene's fine structure constant, $\alpha_{\rm ee} = 0.9$ (graphene on ${\rm SiO}_2$) and $2.2$ (suspended graphene). We clearly see electron-hole puddles with a typical size of a few nanometers.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics{fig04a}\\
\includegraphics{fig04b}
\end{tabular}
\caption{(Color online) Top panel: fully self-consistent electronic density profile $\delta n({\bm r})$
(in units of $10^{12}~{\rm cm}^{-2}$) in a corrugated graphene sheet.
The data reported in this figure have been obtained by setting
$g_1 = 3~{\rm eV}$, $\alpha_{\rm ee} = 0.9$ (this value of $\alpha_{\rm ee}$ is the commonly used value for a graphene sheet on a ${\rm SiO}_2$ substrate), and an average carrier density $\bar n_{\rm c} \simeq 0.8\times 10^{12}~{\rm cm}^{-2}$. Bottom panel: same as in the top panel but for $\alpha_{\rm ee} = 2.2$ (this value of $\alpha_{\rm ee}$ corresponds to suspended graphene). \label{fig:g13}}
\end{center}
\end{figure}
In Fig.~\ref{fig:zoom1D} we show one-dimensional cuts of $\delta n ({\bm r})$ for the same system parameters as in Fig.~\ref{fig:g13}
to better address the separate role of Hartree and exchange-correlation potentials. From the top panel in Fig.~\ref{fig:zoom1D} we clearly see what is the role of electron-electron interactions and screening: the amplitude of the density fluctuations is indeed completely controlled by interactions. From the bottom panel we see how, for this particular set of parameters, scalar LDA exchange and correlations effects seem to be playing only a minor (quantitative) role.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics{fig05a}\\
\includegraphics{fig05b}
\end{tabular}
\caption{(Color online) \label{fig:zoom1D} Top panel: a one-dimensional plot of $\delta n ({\bm r})$ (as a function of $x$ in nm for $y=11.3$~nm) for the same set of parameters as in the lower panel of Fig.~\ref{fig:g13}. Here we have reported data for noninteracting electrons (circles), data obtained including only the Hartree term in Eq.~(\ref{eq:kspot}) (triangles), and data obtained including both Hartree and exchange-correlation potentials (squares). Note that electron-electron interactions completely control the magnitude of density fluctuations and that, on this scale, the data obtained including exchange-correlation effects (squares) are indistinguishable from the data obtained with the inclusion of the Hartree potential only (triangles). Bottom panel: same as in the top panel but with the exclusion of data for noninteracting electrons. Differences between data labeled by squares and by triangles can be seen on this scale. These differences are however only quantitative and not qualitative.}
\end{center}
\end{figure}
As in Ref.~\onlinecite{polini_prb_2008}, it is interesting to compare the reduction in the amplitude of density fluctuations seen in the top panel of Fig.~\ref{fig:zoom1D} with what would be expected in a linear screening approximation. Assuming that the biggest role is played by the scalar potential $V_1$ (this assumption will be justified below in Sect.~\ref{sect:twoc}), within linear-response theory (LRT) the induced density change (in Fourier transform) is given by
\begin{equation}\label{eq:LRT}
\delta n({\bm q}) = \frac{\chi_0(q)}{\varepsilon(q)}~V_1({\bm q})~,
\end{equation}
where $\chi_0(q)$ is the static $T=0$ Lindhard function of a homogeneous noninteracting massless Dirac fermion fluid (see for example Ref.~\onlinecite{barlas_prl_2007}),
\begin{equation}\label{eq:lindhard}
\chi_0(q) = - \nu(\varepsilon_{\rm F}) -\frac{gq}{16 \hbar v}F\left(\frac{2k_{\rm F}}{q}\right) + \frac{g k_{\rm F}}{4\pi \hbar v}
G\left(\frac{2k_{\rm F}}{q}\right)~,
\end{equation}
and $\varepsilon(q) =1-v_q\chi_0(q)$ is the static random-phase-approximation dielectric function:
\begin{equation}\label{eq:dielectricfunction}
\varepsilon(q) = 1 + \frac{q_{\rm TF}}{q} + g \frac{\pi}{8} \alpha_{\rm ee}F\left(\frac{2k_{\rm F}}{q}\right)
-\frac{q_{\rm TF}}{2 q} G\left(\frac{2k_{\rm F}}{q}\right)~.
\end{equation}
Here $\nu(\varepsilon_{\rm F})=g k_{\rm F}/(2\pi \hbar v)$ is the density-of-states at the Fermi level, $v_q = 2\pi e^2/(\epsilon q)$ in the Fourier transform of the electron-electron interaction,
$q_{\rm TF} = g \alpha_{\rm ee} k_{\rm F}$ is the Thomas-Fermi screening vector, and, finally,
\begin{equation}\label{eq:F&G}
\left\{
\begin{array}{l}
{\displaystyle F(x) = 1-\frac{2}{\pi}\arcsin\left[\frac{1}{2}(1+x)-\frac{1}{2}|1-x|\right]}\vspace{0.1 cm}\\
{\displaystyle G(x) = \sqrt{1-x^2}~\Theta(1-x)}
\end{array}
\right.~.
\end{equation}
Note that $F(x) = G(x)=0$ for $x>1$ ({\it i.e.} $q < 2k_{\rm F}$).
In Fig.~\ref{fig:zoom1Drpa} we show a comparison between the prediction of LRT, based on the Fourier transform of Eq.~(\ref{eq:LRT}),
and the non-linear screening result based on the solution of Eq.~(\ref{eq:ksd}) with the Hartree potential only.
We thus see that, maybe surprisingly, LRT explains the data quantitatively.
\begin{figure}[t]
\begin{center}
\includegraphics{fig06}
\caption{(Color online) \label{fig:zoom1Drpa} A one-dimensional plot of $\delta n ({\bm r})$ (as a function of $x$ in nm for $y=11.3$~nm) for the same set of parameters as in Fig.~\ref{fig:zoom1D}. Here we compare results based on the solution of Eq.~(\ref{eq:ksd}) with electron-electron interactions treated at the Hartree level (triangles) with those based on linear-response theory (hexagons), Eqs.~(\ref{eq:LRT})-(\ref{eq:F&G}). Linear screening seems to describe very well the data.}
\end{center}
\end{figure}
In Fig.~\ref{fig:g116} we show fully self-consistent electronic density profiles obtained for a much larger value of the scalar $g_1$ constant. These results have to be compared with those reported in Fig.~\ref{fig:g13}. As expected, in the case $g_1 = 16~{\rm eV}$ the amplitude of the density fluctuations is much larger. A direct comparison has been reported in the one-dimensional cuts in Fig.~\ref{fig:comparison_g1}.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics{fig07a}\\
\includegraphics{fig07b}
\end{tabular}
\caption{(Color online) Same as in Fig.~\ref{fig:g13} but for $g_1 = 16~{\rm eV}$. \label{fig:g116}}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics{fig08}
\caption{(Color online) A one-dimensional plot of the fully self-consistent $\delta n({\bm r})$ (as a function of $x$ in nm for $y=15.8$~nm) obtained using $g_1 = 3$~eV (circles) or $g_1 = 16$~eV (triangles). The other parameters are $\alpha_{\rm ee} = 2.2$ and ${\bar n}_{\rm c} \simeq 0.82 \times 10^{12}~{\rm cm}^{-2}$.\label{fig:comparison_g1}}
\end{center}
\end{figure}
The dependence of the self-consistent density profiles on the doping level ${\bar n}_{\rm c}$ is shown in Fig.~\ref{fig:doping}: from this plot, and especially from the inset, we see that the amplitude of the density fluctuations seem to saturate slowly with increasing ${\bar n}_{\rm c}$, as already found~\cite{polini_prb_2008,rossi_prl_2008} in the case of self-consistent screening calculations in the presence of randomly-distributed charged impurities.
\begin{figure}[t]
\begin{center}
\includegraphics{fig09}
\caption{(Color online) One-dimensional plots of the self-consistent density profiles (as functions of $x$ in nm for $y=21.1$~nm) for different values of doping: $\bar n_{\rm c}\simeq 0.8\times 10^{12}~{\rm cm}^{-2}$ (circles), $\bar n_{\rm c}\simeq 3.96\times 10^{12}~{\rm cm}^{-2}$ (triangles), and $\bar n_{\rm c}\simeq 3.17\times 10^{13}~{\rm cm}^{-2}$ (squares). The data reported in this figure have been obtained by setting $g_1 = 3~{\rm eV}$ and $\alpha_{\rm ee} = 2.2$. The inset shows $\delta n ({\bm r})$ (in units of $10^{12}~{\rm cm}^{-2}$) at a given point ${\bm r}$ in space as a function of the average carrier density $\bar n_{\rm c}$ (in units of $10^{12}~{\rm cm}^{-2}$).\label{fig:doping}}
\end{center}
\end{figure}
Before concluding this Section we stress again that there is no evident correlation between the out-of-plane topographic corrugations
and the spatial structures (electron-hole puddles) in the density profiles, as already pointed out in Sect.~\ref{sect:oneb}. This is highlighted in Fig.~\ref{fig:densityonripples}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{fig10}
\caption{(Color online) Three-dimensional plot of the fully self-consistent continuum-model Dirac-Kohn-Sham density profile reported directly on the corrugated graphene sample shown in Fig.~\ref{fig:sample}. More precisely, the color-coding of the hexagonal bonds labels
the local value of $\delta n({\bm r})$ shown in the two-dimensional
color plot reported in the bottom panel of Fig.~\ref{fig:g13}. Note that there is no simple correspondence between the out-of-plane topographic corrugations and the density profile.\label{fig:densityonripples}}
\end{center}
\end{figure}
\subsection{Self-consistent electronic density in the presence of a model ripple}
\label{sect:twod}
As emphasized in Sects.~\ref{sect:oneb} and~\ref{sect:twoc}, in-plane and out-of-plane displacements have the same impact on the corrugation-induced scalar and vector potentials: this results into complicated spatial patterns of the carrier density with no immediate link with the topographic corrugations. In this Section we present the self-consistent electronic density profile in the presence of a simple model ripple which exhibits displacements {\it only} in the ${\hat {\bm z}}$ direction.
For concreteness, following Ref.~\onlinecite{dejuan_prb_2007}, we consider the following Gaussian out-of-plane displacement:
\begin{equation}\label{eq:bump}
u_z({\bm r}) = A~\exp{\left(-\frac{x_{\rm rel}^2 + y_{\rm rel}^2}{b^2}\right)}~,
\end{equation}
where $x_{\rm rel} = x - L/2$ and $y_{\rm rel} = y - L/2$. The scalar and vector potentials can be easily computed from
Eqs.~(\ref{eq:V1}) and (\ref{eq:V2}), leading to the following expressions:
\begin{equation}\label{eq:V1bump}
V_1({\bm r}) = 2 g_1 \frac{A^2}{b^4} (x_{\rm rel}^2 + y_{\rm rel}^2)~\exp{\left(-2\frac{x_{\rm rel}^2 + y_{\rm rel}^2}{b^2}\right)}
\end{equation}
and
\begin{equation}\label{eq:V2bump}
V_2({\bm r}) = 2 g_2 \frac{A^2}{b^4}\left(x_{\rm rel} + i y_{\rm rel}\right)^2~\exp{\left(-2\frac{x_{\rm rel}^2 + y_{\rm rel}^2}{b^2}\right)}~.
\end{equation}
The fully self-consistent density profile $\delta n({\bm r})$ calculated with the use of the potentials (\ref{eq:V1bump}) and~(\ref{eq:V2bump}) is reported in Fig.~\ref{fig:andopotentialsbump}. These data show that when in-plane displacements are neglected the correlation between the density profile and the topography of the corrugated graphene sheet [Eq.~(\ref{eq:bump})] is much more transparent. Note that the oscillations in $\delta n({\bm r})$ stem from the fact that the quantity $|\nabla u_z({\bm r})|^2$, which controls the scalar potential $V_1$, is maximal at $|{\bm r}| \approx b$.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c c}
\includegraphics[width=0.5\linewidth]{fig11a} &
\includegraphics[width=0.5\linewidth]{fig11b} \\
\includegraphics[width=0.5\linewidth]{fig11c} &
\includegraphics[width=0.5\linewidth]{fig11d}
\end{tabular}
\caption{(Color online) Top left panel: color plot of the analytical scalar potential $V_1({\bm r})$ (in units of meV) reported in Eq.~(\ref{eq:V1bump}). The parameters used are: $g_1 = 3$~eV, $A = 0.05 L = 1.1~{\rm nm}$, and $b = 0.2 L=4.5~{\rm nm}$. Top right panel: real part of the analytical potential $V_2({\bm r})$ (in units of meV) in Eq.~(\ref{eq:V2bump}). Bottom left panel: imaginary part of the potential $V_2({\bm r})$ (in units of meV) in Eq.~(\ref{eq:V2bump}). Bottom right panel: fully self-consistent electronic density profile (in units of $10^{12}~{\rm cm}^{-2}$) calculated in the presence of the scalar and vector potentials shown in the other panels. This numerical calculation has been performed using $\alpha_{\rm ee} = 2.2$ and $\bar n_{\rm c}\simeq 3.96\times 10^{12}~{\rm cm}^{-2}$.
\label{fig:andopotentialsbump}}
\end{center}
\end{figure}
\subsection{Comments on the density response to a purely vector potential}
\label{sect:twoe}
A natural question might arise at this point: what is the relative role of $V_1$ and $V_2$ in determining the induced density $\delta n({\bm r})$?
In this Section we study the density response of a system of massless Dirac fermions to a {\it purely vector} potential.
Let us begin for simplicity from a noninteracting system: in this case we can prove that $\delta n({\bm r}) = 0$, independently of doping.
This can be easily seen within the framework of LRT: in this case
\begin{equation}
\delta n({\bm q}) = \sum_{i \in {x,y}} \chi_{n\jmath^i}(q) A_i({\bm q})~,
\end{equation}
where $\delta n({\bm q})$ and $A_i({\bm q})$ are the Fourier transforms of $\delta n({\bm r})$ and $A_i({\bm r})$, and
$\chi_{n\jmath^i}(q) = \lim_{\omega \to 0}\chi_{n\jmath^i}(q,\omega)$
is a static linear-response function. It turns out (see Appendix~\ref{appendix} for a formal proof) that
\begin{equation}
\chi_{n\jmath^i}(q,\omega) = \frac{q_i}{q}~\left[\frac{\omega}{q}\chi_{nn}(q,\omega)\right]~,
\end{equation}
where $\chi_{nn}(q,\omega)$ is the density-density response function of a noninteracting system of massless Dirac fermions (see for example Ref.~\onlinecite{barlas_prl_2007} and references therein). Because $\chi_{nn}(q,\omega)$ is well behaved in the static limit we immediately find that $\chi_{n\jmath^i}(q) =0$.
An identical conclusion can be reached by invoking Furry's theorem~\cite{furry_pr_1937,jackiw_prb_2009}, which applies independently of the strength of the external vector potential ${\bm A}$ (and thus also beyond the regime of applicability of LRT) and in the presence of electron-electron interactions. The theorem, however, is valid only for systems with an electron-hole-symmetric spectrum.
We thus expect $\delta n({\bm r}) = 0$ only in the case of a neutral-on-average system, while we expect a finite induced density for a finite value of ${\bar n}_{\rm c}$.
We have checked these expectations numerically. We have performed calculations in the presence of the scalar $V_1$ component only and compared the calculated induced density, $\delta n_{\rm S}({\bm r})$, with that obtained in the presence of both scalar and vector potentials,
$\delta n_{\rm TOT}({\bm r})$. In Fig.~\ref{fig:totvsscalar} we report the results for $\alpha_{\rm ee}=0$: we clearly see, especially from the bottom panel, that even at finite average carrier density the amplitude of the spatial fluctuations induced by the vector potential only is rather small.
Differences between $\delta n_{\rm S}({\bm r})$ and $\delta n_{\rm TOT}({\bm r})$ have been quantified by the value of the following dimensionless parameter,
\begin{equation}
\varepsilon = \frac{\sqrt{||\delta n_{\rm TOT}({\bm r}) - \delta n_{\rm S}({\bm r})||}}{\sqrt{||\delta n_{\rm TOT}({\bm r})||} + \sqrt{||\delta n_{\rm S}({\bm r})||}}~,
\end{equation}
where
\begin{equation}
||{\cal O}({\bm r}) ||^{2} = \int d^2 {\bm r} |{\cal O}({\bm r})|^2
\end{equation}
is the usual $L^2$ norm. In the case ${\bar n}_{\rm c} = 0$ we find $\varepsilon \simeq 3\times 10^{-4}$, which is below our numerical precision ($0.005$): within the accuracy of the calculation thus $\delta n_{\rm TOT}({\bm r}) = \delta n_{\rm S}({\bm r})$. In the calculations with finite carrier density, however, we find much higher values of $\varepsilon$: for $\bar n_{\rm c} \simeq 3.96\times 10^{12}~{\rm cm}^{-2}$ we find $\varepsilon \simeq 0.02$, while for $\bar n_{\rm c} \simeq 3.17\times 10^{13}~{\rm cm}^{-2}$ we find $\varepsilon \simeq 0.03$.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics{fig12a}\\
\includegraphics{fig12b}
\end{tabular}
\caption{(Color online) Top panel: a one-dimensional plot of the noninteracting ($\alpha_{\rm ee}=0$) density profile $\delta n ({\bm r})$ (as a function of $x$ in nm for $y=12.3$~nm) obtained solving the Dirac equation in the presence of both scalar and vector potentials (circles) or of the scalar potential only (triangles). Bottom panel: a one-dimensional plot of the noninteracting density profile $\delta n ({\bm r})$ (as a function of $x$ in nm for $y=12.3$~nm) obtained solving the Dirac equation in the presence of the vector potential only. The data reported here refer to $g_1 = 3~{\rm eV}$ and $\bar n_{\rm c} \simeq 3.96\times 10^{12}~{\rm cm}^{-2}$. From both panels we conclude that density fluctuations are largely controlled by the scalar potential.
\label{fig:totvsscalar}}
\end{center}
\end{figure}
\subsection{Electronic density in the presence of both ripples and charged impurities}
\label{sect:twof}
Before concluding we would like to briefly illustrate how the presence of the ripples modifies qualitatively the density landscape induced by a random distribution of charged impurities~\cite{polini_prb_2008,rossi_prl_2008}. In this Section we report numerical results based on the self-consistent solution of Eq.~(\ref{eq:ksd}) in the presence of a scalar potential $V_{\rm ext}({\bm r})$ given by:
\begin{equation}\label{eq:ext_pot}
V_{\rm ext}({\bm r}) = V_1({\bm r}) + V_{\rm imp}({\bm r})~.
\end{equation}
Here $V_{\rm imp}({\bm r})$ is a scalar potential due to charged impurities~\cite{polini_prb_2008},
\begin{equation}\label{eq:chargedimp}
V_{\rm imp}({\bm r}) = - \sum_{i=1}^{N_{\rm imp}}\frac{Ze^2}{\epsilon \sqrt{|{\bm r}-{\bm R}_i|^2 + d^2}}~,
\end{equation}
where ${\bm R}_i$ are random positions in the supercell and $d$ is the distance between the graphene sheet and the plane where the impurities are located. For simplicity, all charges have been taken to have the same $Z$ in Eq.~(\ref{eq:chargedimp}).
In Fig.~\ref{fig:ripples&imp} we show fully self-consistent density profiles of massless Dirac fermions subjected to the scalar potential of
$N_{\rm imp}=5$ charged impurities: in the top panel we show $\delta n ({\bm r})$ calculated in the absence of ripples ($g_1=g_2=0$), while in the bottom panel we have included them.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics{fig13a}\\
\includegraphics{fig13b}
\end{tabular}
\caption{(Color online) Top panel: fully self-consistent electronic density profile (in units of $10^{12}~{\rm cm}^{-2}$) calculated from the solution of Eq.~(\ref{eq:ksd}) in the presence of $N_{\rm imp} = 5$ charged impurities with charge $Z = + 1$ (donors). The white circles label the position of the charges on a plane located at a distance $d\simeq 2$~nm from the graphene sheet. Bottom panel: same as in the top panel but in the presence of ripples too. The data reported here have been obtained by setting $g_1 = 3~{\rm eV}$, $\alpha_{\rm ee} = 0.9$, and $\bar n_{\rm c} \simeq 3.96\times 10^{12}~{\rm cm}^{-2}$. \label{fig:ripples&imp}}
\end{center}
\end{figure}
We clearly see how the smooth landscape of electron-hole puddles in the presence of charged impurities only (top panel) is dramatically affected by the presence of corrugations (bottom panel), which induce additional spatial variations with a much smaller length scale (probably well below the current spatial experimental resolution of probes like SET~\cite{yacoby_natphys_2008} or STM~\cite{zhang_nature_2009}). Once again, we would like to emphasize that these small-wavelength carrier-density oscillations are due to a complicated interference between the effects of out-of-plane {\it and} in-plane atomic displacements.
\section{Conclusions}
\label{sect:three}
In summary, we have presented quantitative calculations of scalar and vector potentials induced by corrugations in single-layer graphene sheets. We have found that the contributions from in-plane and out-of-plane atomic displacements are both of the same order and that this does not lead to evident correlations between the out-of-plane topographic corrugations and the induced scalar and vector potentials.
We have then used these potentials to calculate self-consistently the induced electronic density distribution in the presence of electron-electron interactions. To this end we have generalized the Kohn-Sham-Dirac theory of Ref.~\onlinecite{polini_prb_2008} to treat situations with spatial-dependent vector potentials. We have discovered that spatial density fluctuations are largely controlled by the scalar potential, especially in nearly-neutral graphene sheets, and that this creates complicated short-wavelength (a few ${\rm nm}$) electron-hole puddles which do not exhibit evident correlations with the topography of the sheet.
In the future we would like to investigate more deeply the role of the exchange-correlation corrections to the vector potential~\cite{vignale_prl_1987}, especially in view of the fact that the exchange-correlation contribution to the scalar Kohn-Sham potential,
which has been studied here, has been found to play a minor role.
\acknowledgements
M.G. and A.T. have equally contributed to this work. M.P. acknowledges useful discussions with A.I. Milstein.
A.F. and M.I.K acknowledge a support from Stichting voor Fundamenteel Onderzoek der Materie (FOM), The Netherlands.
|
\section{Introduction}
We are entering exciting times in particle physics: the Large Hadron Collider
(LHC) is setting a new landmark at the high-energy frontier and probes,
through the collision of multi-TeV protons, the structure of
matter and space-time at an unprecedented level. There is a lot of circumstantial
evidence that the physics at the TeV scale exploited at LHC will bring decisive insights into fundamental
questions such as the origin of particle masses, the nature of dark matter in the universe, and
the unification of all forces, including gravity.
Indeed, most proposals to embed the Standard Model of particle physics into a more general,
unified framework, notably the ones based on string theory or its low energy incarnations,
supergravity and supersymmetry, predict new heavy, $m\gtrsim 100$~GeV,
particles which may be searched for at TeV colliders. Some of these particles,
prominent examples being neutralinos, are natural candidates for the constituents of
cold dark matter in the form of so-called weakly interacting massive particles (WIMPs).
However, there is also evidence that there is fundamental physics at the sub-eV scale.
Indeed, atmospheric, reactor, and solar neutrino data strongly support the hypothesis
that neutrinos have masses in the sub-eV range. Moreover, the vacuum energy density of
the universe, as inferred from cosmological observations, points to the sub-eV range,
$\rho_\Lambda \sim {\rm meV}^4$. As a matter of fact, many of the above mentioned extensions
of the Standard Model not only predict WIMPs, but also WISPs -- very weakly interacting
sub-eV particles. Prominent candidates for such particles go under the
names axions and axion-like particles, often arising as Nambu-Goldstone bosons associated with
the breakdown of global symmetries. Further WISP candidates are massless or light extra, hidden U(1)
gauge bosons as well as light, chiral fermions charged under this hidden U(1). These particles
are frequently encountered in string embeddings of the Standard Model.
The latter potentially contain also light moduli fields and light gravitinos as further
WISP candidates.
Unlike for WIMPs, TeV colliders are not the best means to search for WISPs.
For this purpose, low energy experiments exploiting lasers, microwave cavities, strong electromagnetic
fields, torsion balances etc. seem to be superior.
It is the purpose of this review to present the physics case and a status report
of this emerging low-energy frontier of fundamental physics.
The organization of this review is as follows. In the following Sect.~\ref{physicscase}
we will argue that
many extensions of the Standard Model predict new particles and phenomena at low energies.
On the one hand new light particles are suggested to solve puzzling experimental
results but on the other hand they also appear to be a generic feature of underlying fundamental theories such as string theory.
In Sect.~\ref{astro} we will turn to current constraints from astrophysics and cosmology. Moreover, we will discuss a
few interesting observations that could be explained by invoking WISPs.
Then, in Sect.~\ref{searches}, we will explore how WISPs can be searched for in a variety of controlled laboratory
experiments. We discuss the advantages of these experiments as well as the challenges they face.
Finally in Sect.~\ref{conclusions} we will summarize the current situation and give an outlook towards the future.
\section{Physics Case for WISPs}\label{physicscase}
\subsection{Axions and Axion-Like Particles}
\subsubsection{The Strong CP Problem and Axions\label{Sec:QCD_axion}}
Quantum chromodynamics (QCD), the non-Abelian gauge theory describing strong
interactions, allows for a CP-violating term in the Lagrangian,
\begin{eqnarray}
{\mathcal L}_{\rm CP-viol.} =
\frac{\alpha_s}{4\pi}\, \theta\, {\rm tr}\, G_{\mu\nu} {\tilde G}^{\mu\nu} \equiv
\frac{\alpha_s}{4\pi}\, \theta\,
\frac{1}{2}\,\epsilon^{\mu\nu\alpha\beta}\,{\rm tr}\, G_{\mu\nu} G_{\alpha\beta},
\label{topterm}
\end{eqnarray}
where $G$ is the gluonic field strength. Similar to the strong coupling constant $\alpha_s$,
the fundamental parameter $\theta$ has to be determined experimentally.
One of the most sensitive probes for it is the electric dipole moment of the neutron, arising from the CP-violating term
given in Eq.~(\ref{topterm}). It should be of order
\begin{eqnarray}
\left| d_n\right| \sim \frac{e}{m_n} \left( \frac{m_q}{m_n}\right)
\left|\bar{\theta}\right|
\sim 10^{-16}\ \left|\bar{\theta}\right| \ e\,{\rm cm},
\end{eqnarray}
where $m_n$ ($m_q$) is the neutron (a light-quark mass), $e$ is the unit electric charge,
and
\begin{equation}
\bar{\theta} \equiv \theta + {\rm arg\ det\ }M,
\end{equation}
with $M$ being the quark mass matrix. $\bar{\theta}$ is the actual physical CP-violating parameter in the Standard Model.
The current experimental
upper bound on $\left|d_n\right|<2.9\times 10^{-26}\ e\,$cm~\cite{Amsler:2008zzb} places an extremely stringent limit on
\begin{equation}
\left|\bar\theta\right| \lesssim 10^{-10}.
\end{equation}
The strong CP problem is the lack of an explanation why the dimensionless parameter $\bar\theta$,
a sum of two contribution of very different origins, is so unnaturally small.
The axion occurs in course of a possible solution of this problem. In essence, the proposal
of Peccei and Quinn~\cite{Peccei:1977hh} was to promote $\theta$ to a dynamical field which can relax
spontaneously to zero. The axion field $a$ is introduced as a dynamical $\theta$ parameter,
which has a shift symmetry,
\begin{eqnarray}
a\to a + {\rm const.},
\end{eqnarray}
broken only by the anomalous CP-violating terms, i.e. its low-energy effective Lagrangian is parametrized as
\begin{eqnarray}
{\mathcal L}_a =
\frac{1}{2} \partial_\mu a \partial^\mu a
+ \frac{\alpha_s}{4\pi f_a}\, a\, {\rm tr}\, G^{\mu\nu} {\tilde G}_{\mu\nu}+
\frac{s\alpha}{8\pi f_a}\, a\,F^{\mu\nu} {\tilde F}_{\mu\nu} + {\mathcal L}_a^{\rm int} \left[\frac{\partial_\mu a}{f_a};\psi\right],
\label{axion_leff}
\end{eqnarray}
where $s$ is a model dependent parameter, $F$ is the electromagnetic
field strength, and $\psi$ denotes generic Standard Model
fields. The dimensionful axion decay constant $f_a$
determines the strength of the interaction of the axion with the
Standard Model particles. The $\theta$-term in the QCD Lagrangian can then be
eliminated by absorbing it into the axion field, $a=\bar{a} - \bar{\theta}
f_a$. Finally, the topological charge density $\propto \langle
{\rm tr}\, G^{\mu\nu} {\tilde G}_{\mu\nu} \rangle \neq 0$, induced
by topological fluctuations of the gluon fields such as QCD
instantons, provides a nontrivial potential for the axion field $\bar{a}$
which is minimized at zero expectation value, $\langle \bar{a}\rangle =0$:
thus, the $\bar{\theta}$ dependence is wiped out by the axion field,
providing a natural explanation why the electric dipole moment of
the neutron is so small. The nontrivial potential around $\langle \bar{a}\rangle =0$ promotes the elementary particle excitation of the
axion field, the axion, to a pseudo Nambu-Goldstone
boson~\cite{Weinberg:1977ma} (which we will now again denote by $a$) with a non-vanishing, but
parametrically small mass. This mass can be calculated via current
algebra and expressed in terms of the light ($u,d$) quark masses,
the pion mass $m_\pi$ and the pion decay constant $f_\pi$~\cite{Weinberg:1977ma} (cf. also \cite{Amsler:2008zzb}),
\begin{eqnarray}
m_a =
\frac{m_\pi f_\pi}{f_a}\frac{\sqrt{m_u m_d}}{m_u+m_d}\simeq { 0.6\, {\rm meV}}
\times
\left(
\frac{10^{10}\, {\rm GeV}}{f_a}\right) .
\label{axionmass}
\end{eqnarray}
For large axion decay constant $f_a$, we see that the axion is a prime
example for a WISP~\cite{Kim:1979if}: it is a very weakly interacting (cf. Eq.~(\ref{axion_leff})) sub-eV mass
particle. In particular, its coupling to photons, which for an axion, like for any other pseudo-scalar, should
be of the form,
\begin{eqnarray}
{\cal L}_{a \gamma \gamma} = - \frac{1}{4}\, g\, a\, F_{\mu \nu}\tilde{F}^{\mu \nu} =
g\, a\, \vec{E}\cdot \vec{B} ,
\end{eqnarray}
is very much suppressed~\cite{Bardeen:1977bd},
\begin{eqnarray}
{ g} = \frac{\alpha}{2\pi f_a}
\left( {\frac{2}{3}\,\frac{m_u+4 m_d}{m_u+m_d} - s }\right)
\sim 10^{-13}\ {\rm GeV}^{-1} \left(
\frac{10^{10}\, {\rm GeV}}{f_a}\right).
\label{axionphotoncoupling}
\end{eqnarray}
Although expected to be small, the guaranteed coupling of axions to
photons, Eq.~(\ref{axionphotoncoupling}), may result, if axions
exist, in observable consequences from processes involving large
electromagnetic fields. These often occur in astrophysical and
cosmological environments (cf. Sect.~\ref{astro}) and can be
prepared in laboratory experiments (cf. Sect.~\ref{searches}).
The proposal of an anomalous Peccei-Quinn shift symmetry is motivated to provide for a solution
of the strong CP problem. This concept has been generalized to other similar WISP candidates
-- axion-like particles (ALPs) -- which may arise as (pseudo) Nambu-Goldstone bosons
from the breaking of other global symmetries such as, for example, family symmetries.
However, in contrast to axions, for generic ALPs a non-zero coupling $g$ to photons is not
guaranteed. Moreover, for them a predictive relation between $g$ and
the mass is missing. Correspondingly, ALP searches, exploiting their interactions with
photons, should try to cover the entire parameter space spanned by $g$ and the mass
of the ALP and not only the restricted parameter space, Eqs.~(\ref{axionmass}) and
(\ref{axionphotoncoupling}), predicted for axions.
\subsubsection{Axions and Axion-Like Particles from String Compactifications\label{Sec:axions_string}}
The existence of axions and ALPs can also be strongly motivated from a top-down
point of view. In fact, when compactifying
the six extra spatial dimensions of string theory they arise quite naturally as Kaluza-Klein zero modes
of antisymmetric tensor fields, which are generically present in all string theories.
Moreover, the (Chern-Simons) couplings of these form fields to the gauge fields,
which are crucially determined by anomaly cancelation conditions, result in the anomalous
CP violating couplings in the low-energy effective Lagrangian, Eq.~(\ref{axion_leff}),
necessary for the solution of the strong CP problem.
Thus, string compactifications suggest plenty of candidates for axions and axion-like
WISPs~\cite{Witten:1984dg,Conlon:2006tq,Svrcek:2006yi,Arvanitaki:2009fg}. However, it is fair to say that they do not
really predict them, because there are several mechanisms known by
which they can be removed from the low-energy spectrum. Only the
ones which escape these mechanisms are WISP candidates.
In the compactification of the weakly coupled heterotic string, a
universal, {\em model-independent} axion appears as the Poincare
dual\footnote{\baselineskip 16pt To perform this dualization explicitly one
introduces the axion field as a Lagrange multiplier for the Bianchi
identity for $H$, $dH=1/(16\pi^2)({\rm tr}R\wedge R-{\rm tr}F\wedge
F)$, and subsequently integrates over
$H$~\cite{Witten:1984dg,Svrcek:2006yi}.} of the (Neveu-Schwarz)
antisymmetric tensor field $B_{\mu\nu}$, with $\mu$ and $\nu$
tangent to 3+1 dimensional Min\-kow\-ski
space-time~\cite{Witten:1984dg}. Its decay constant $f_a$ is quite
independent of the details of the compactification. To compute it,
one considers the action of an $N=1$ supergravity coupled to an
${\rm E}_8\times {\rm E}_8$ pure gauge theory in 9+1 dimensions,
\begin{equation}
S_{\rm H}
=\frac{2\pi M_s^8}{g_s^2}\int d^{10}x \sqrt{-g}R
-\frac{M_s^6}{2\pi g_s^2}\int\frac{1}{4}{\rm tr} F\wedge\star F
-{\frac{2\pi M_s^4}{g_s^2}\int\frac{1}{2} H\wedge\star H}+\ldots ,
\end{equation}
which describes the dynamics of the massless bosonic excitations of the heterotic string
in terms of the Ricci scalar $R$, the
gauge field strength $F$, and the field strength $H$ of the two-form field $B$.
Compactifying this theory
on a 6 dimensional manifold with volume $V_6$, the
resulting effective action can be matched to its standard normalization in 3+1 dimensions
\begin{equation}
S_{\rm 3+1}
=\frac{M_P^2}{2}\int d^4x\sqrt{-g}\,R
-\frac{1}{4g_{\rm YM}^2}\int d^4x\sqrt{-g}\,{\rm tr}\, F_{\mu\nu} F^{\mu\nu}
-{\frac{1}{f_a^2}\int\frac{1}{2} H\wedge\star H} + \ldots\,,
\end{equation}
with
\begin{eqnarray}
M_P^2 = (4\pi/g_s^2) M_s^8 V_6; \hspace{2ex}
g_{\rm YM}^2 = 4\pi g_s^2/(M_s^{6} V_6); \hspace{2ex}
{ f_a^2=g_s^2/(2\pi M_s^4 V_6)}\,,
\label{heterotic_couplings}
\end{eqnarray}
expressing the reduced Planck mass $M_P=2.4\times 10^{18}$~GeV,
the gauge coupling $g_{\rm YM}$, and the axion decay
constant $f_a$ in terms of the string coupling $g_s$, the string scale $M_s=1/\ell_s$, and the volume
$V_6$. Eliminating the volume $V_6$ and the string scale by means of the first two relations
in Eq.~(\ref{heterotic_couplings}),
one ends up with an axion decay constant of order
of the GUT scale~\cite{Choi:1985je},
\begin{eqnarray}
f_a =\alpha_{\rm YM} M_P/(2\pi \sqrt{2})
\simeq 1.1\times 10^{16}\ {\rm GeV}\,,
{\rm \ for\ } \alpha_{\rm YM}=g^2_{\rm YM}/(4\pi)\sim 1/25\,.
\label{f_a_heterotic}
\end{eqnarray}
{\em Model-dependent} axions arise in the context of weakly coupled
heterotic strings from massless excitations of the two-form
$B$-field on the 6 dimensional compact
manifold~\cite{Witten:1984dg}. Correspondingly, their properties
depend much more on the details of the compactification. Nevertheless, a recent exhaustive study has
elucidated~\cite{Svrcek:2006yi} that also in this case the axion
decay constant cannot be smaller than $10^{15}$~GeV. Similar
conclusions have been drawn for the
axions in strongly coupled heterotic string theory~\cite{Svrcek:2006yi}.
These findings can be easily understood physically: it is the string
scale $M_s$ which mainly determines the axion decay
constant~\cite{Conlon:2006tq}. And in the heterotic case, this scale
is large, e.g. $M_s = \sqrt{\alpha_{\rm YM}/(4\pi)} M_P$, for the
weakly coupled heterotic string (cf.
Eq.~(\ref{heterotic_couplings})).
\begin{figure}[t!]
\centerline{\includegraphics[width=0.9\textwidth]{Stringbild18.eps}}
\caption{\baselineskip 16pt In compactifications of type II string theories the
Standard Model is locally realized by a stack of space-time filling D-branes wrapping
topologically non-trivial submanifolds in the compact dimensions. In
general, there can also be hidden sectors localized at different
places. They can arise from branes of different dimension (D3 or D7
branes) which can be either of large extent or localized at
singularities. Light visible and hidden matter particles arise from
strings located at intersection loci and stretching between brane
stacks. }\label{Fig:type_ii_comp}
\end{figure}
This may be different in compactifications of type II string theories which give rise to
``intersecting brane worlds". In these theories, the Standard Model lives on a stack
of D$(3+q)$-branes which are extended along the 3+1 non-compact dimensions and wrap
$q$-dimensional topologically non-trivial submanifolds in the compactification manifold,
while gravity propagates in the bulk, leading to a possibly smaller string scale at the
expense of a larger compactification volume, $M_s\sim g_s M_P/\sqrt{V_6 M_s^6}$ (see Fig.~\ref{Fig:type_ii_comp}).
In type II string theory, the axions come from the massless excitations of the
(Ramond-Ramond) $q$-form gauge field $C_q$. The precise predictions depend
on the particular embedding of the Standard Model, but generically one finds
that the axion decay constant can be substantially
lower than in the heterotic case, varying between~\cite{Conlon:2006tq}
\begin{equation}
f_a\sim \frac{M_P}{\sqrt{V_6 M_s^6}}\sim \frac{M_s}{g_s}\sim 10^{4\div 17}\ {\rm GeV},
\end{equation}
corresponding to a variation of the string scale between the TeV and the GUT scale.
\subsubsection{Scalars and Chameleons}\label{chameleonsect}
Apart from possibly light pseudoscalars, string compactifications generically also predict scalar particles
-- the dilaton and large numbers of moduli -- which appear also massless at the compactification scale.
Essentially massless scalar fields are also often invoked by cosmologists in the context
of dark energy. In fact, a plausible explanation for the apparent acceleration of the cosmic
expansion rate of the universe is provided by the presence of a spatially
homogeneous scalar field which is rolling down a very flat
potential~\cite{Wetterich:1987fm}.
Interactions of very light scalar fields with ordinary matter are strongly
constrained by the non-observation of ``fifth force" effects
leading to {\it e.g.}~violations of the equivalence principle
(cf. Sect.~\ref{Sec:fifth_force}). Correspondingly,
if such particles exist, the forces mediated by them should be either much weaker
than gravity or short-ranged in the laboratory. The latter occurs in theories
where the mass of the scalar field depends effectively on the local
density of matter -- in so-called chameleon field theories~\cite{Khoury:2003aq}.
Depending on the non-linear field self-interactions and on the interactions with the ambient
matter, the chameleon may have a large mass in regions of high density (like the earth),
while it has a small mass in regions of low density (like interstellar space).
Since such particles are able to hide so well from observations and experiments,
they have been named ``chameleons".
\subsection{Ultralight Hidden-Sector Particles\label{Sec:hidden_string}}
Similar to axions and axion-like particles, additional hidden sector U(1) gauge bosons are
also a generic feature arising in string compactifications. These are therefore well motivated
WISP candidates.
In fact, in the standard compactification of the ${\rm E}_8\times {\rm E}_8$ supergravity based
on the heterotic string on a smooth (Calabi-Yau) manifold,
the Standard Model gauge group is embedded in the first ${\rm E}_8$ factor,
whereas the second ${\rm E}_8$ factor comprises a ``hidden gauge group", which interacts with
the first ${\rm E}_8$ factor only gravitationally~\cite{Candelas:1985en}. This second ${\rm E}_8$ factor may be broken in the
course of compactification to products of non-Abelian and U(1) gauge groups.
The occurrence of hidden U(1)s can be studied quite exhaustively in toroidal orbifold compactifications
of the heterotic string, which allow for a systematic scanning of possible gauge group factors and particles
after compactification. Requiring a realistic visible sector, it seems that there are still
a number of models which have possibly massless hidden U(1)s.
In this ``mini-landscape" of orbifold compactifications of the heterotic string~\cite{Lebedev:2008un} one encounters
a breaking of the gauge symmetry to the Standard Model, a hidden sector
non-Abelian gauge symmetry and, typically, at least one hidden U(1), for example, cf.~Ref.~\cite{Lebedev:2009ag},
\begin{equation}
{\rm E}_8\times {\rm E}_8\to
\underbrace{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}_{\rm Standard\ Model}
\times
[{\rm SU(6)}\times {\rm U(1)}] .
\end{equation}
Compactifications of type II string theory also suggest the existence of hidden U(1)s,
be it as Kaluza-Klein zero modes of the previously mentioned
(Ramond-Ramond) form fields or as massless excitations of branes.
In fact,
as illustrated in
Fig.~\ref{Fig:type_ii_comp}, type II compactifications generically involve space-time filling
hidden sector branes not intersecting with the Standard Model branes,
often also for global consistency requirements.
Some of these hidden U(1)s may remain unbroken down to very small energy scales.
In this case their dominant interaction with the photon, which is
encoded in the low-energy effective Lagrangian,
\begin{equation}
\mathcal{L} \supset -\frac{1}{4e^2} F_{\mu \nu} F^{\mu \nu}
- \frac{1}{4g_h^2} X_{\mu \nu} X^{\mu \nu}
+ \frac{\chi}{2 e g_h} F_{\mu \nu} X^{\mu \nu}
+ \frac{m_{\gamma^\prime}^2}{2 g^{2}_h} X_{\mu} X^{\mu},
\label{LagKM}
\end{equation}
with $X_\mu$ denoting the hidden U(1) field with field strength $X_{\mu\nu}$ and gauge coupling $g_{h}$,
will be through kinetic mixing~\cite{Holdom:1985ag}, with mixing parameter $\chi$. Therefore, light hidden U(1)s (``hidden photons") are well motivated WISP candidates, since $\chi$ is expected to be small.
In fact, kinetic mixing is generated at one-loop by the exchange of heavy messengers that couple both
to the visible U(1) as well as to the hidden U(1). In the context of compactifications
of the heterotic string, its size has been estimated as~\cite{Dienes:1996zr}
\begin{equation}
\chi \sim \frac{e g_h}{16\pi^2}\,C\,\frac{\Delta m}{M_P}\sim 10^{-5}\div 10^{-17}\,,
{\rm \ for\ } C\gtrsim 10\,,
\end{equation}
where $\Delta m\sim 10^{5\div 17}$~GeV is the possible range of mass splitting in the messenger sector once supersymmetry is broken.
A great variety for possible values of $\chi$ can also be accommodated in type II compactifications
for the mixing between brane-localized hidden U(1)s and the visible U(1). Here, kinetic
mixing can be understood as originating from the exchange of closed strings through the
bulk~\cite{Lust:2003ky,Abel:2008ai}. Generically, one finds~\cite{Goodsell:2009xc}
\begin{equation}
\label{chieq}
\chi \sim \frac{e g_h}{16 \pi^2},
\end{equation}
where the size of the hidden sector gauge coupling $g_h$ depends on the $q$-dimensional
volume ($0\leq q\leq 6$) of the cycle which the hidden brane wraps,
\begin{equation}
g_h^2 \simeq
\frac{2\pi g_s}{V_q M_s^q}
= 2\pi g_s \left( \frac{4\pi}{g_s^2} \frac{M_s^2}{M_P^2}\right)^{q/6}\,,
\label{gsquaredq}
\end{equation}
leading to a quite large range of possible values for the kinetic mixing,
\begin{equation}
10^{-12}\lesssim \chi \lesssim 10^{-3},
\end{equation}
for the string scale varying between a TeV scale and the GUT scale.
Smaller values of kinetic mixing can be obtained in these setups in special cases where the
one-loop contribution is cancelled or vanishes. Moreover,
exponentially suppressed values can be naturally obtained in flux compactifications
with warped throats~\cite{Abel:2008ai}.
Masses for the hidden photons can arise via the standard Higgs mechanism but also via a Stueckelberg mechanism.
In LARGE volume compactifications small, even sub-eV, masses arise quite naturally~\cite{Goodsell:2009xc}.
If the masses arise from a Stueckelberg mechanism, mass and size of the kinetic mixing are typically linked through one scale, the string
scale, and therefore related to each other. Depending on the specific way in which the cycles wrap the singularities one obtains
expressions for the masses like
\begin{equation}
(m^{\rm Stueck}_{\gamma^{\prime}})^2\simeq\frac{g_{s}}{2}\left(\frac{4\pi}{g^2_{s}}\frac{M^{2}_{s}}{M^{2}_{P}}\right)^z,\quad z=\frac{1}{3},1.
\end{equation}
For example in the case $z=1$ we obtain for a string scale of $M_{s}\sim 1\,{\rm TeV}$, $m_{\gamma^{\prime}}\sim {\rm meV}$ and, from Eqs.~\eqref{chieq}, \eqref{gsquaredq}, a mixing of
$\chi\sim10^{-12}$.
The predictions for masses arising from the Higgs mechanism are less precise, however, they can also be
tiny, if the supersymmetry breaking scale in the hidden sector is much smaller than in the visible sector.
Occasionally, there is also light hidden matter charged under the hidden U(1)s.
After diagonalization of the gauge kinetic terms by a shift
$X\to X + \chi A$ and a multiplicative hypercharge renormalization,
one observes that the hidden sector particles acquire a minihypercharge
$\epsilon = \chi g_h$~\cite{Holdom:1985ag}. In particular, also an eventual hidden Higgs particle may
be searched for by exploiting its effective minihypercharge~\cite{Ahlers:2008qc}.
In a similar way minicharged particles can also arise from hidden sector magnetic monopoles
if the gauge fields mix via a non-diagonal $\theta$-term~\cite{Bruemmer:2009ky}.
Therefore, light minicharged particles (MCPs) are
also very well motivated WISP candidates.
\section{Astrophysical and Cosmological Constraints on WISPs}\label{astro}
As we have reviewed in the last section, there is a strong physics case for the possible existence of WISPs.
Moreover, their possible masses and couplings span a very wide range
in parameter space. Correspondingly, searches for signatures of WISPs have to exploit a wide variety of
observational and experimental techniques, ranging from cosmology and astrophysics
to terrestrial laboratory experiments. As reviewed in this section, the strongest bounds on the existence of
WISPs presently often come from stellar evolution and cosmology, where to the best of our knowledge observations seem
to agree with the standard budget of elementary particles.
However, there are also some intriguing astronomical observations
which are hard to explain by known physics and might be interpreted as indirect hints pointing towards
the existence of WISPs.
\subsection{Bounds from Stellar Evolution}\label{stellar}
Production of WISPs in stars would substantially affect stellar evolution \cite{Raffelt:1996wa}.
WISPs are only rarely produced in the dense plasmas of stellar interiors, but they will easily
escape, contributing directly to the total energy loss of the star.
This has to be contrasted with the standard energy loss of stars, which
occurs mainly due to photons from the stellar surface and neutrinos from the core.
Therefore, the WISP luminosity is enhanced by a huge volume/surface factor, leading to very strong
constraints.
\begin{figure}[t!]
\centerline{\includegraphics[width=.75\textwidth]{axionline.eps}}
\vspace{0.5cm}
\centerline{\includegraphics[width=.6\textwidth]{axionspecial.eps}}
\caption{\baselineskip 16pt Summary of cosmological and astrophysical constraints
for axions (top) (for the mass $m_a$ or decay constant
$f_a$)~\cite{Raffelt:2006cw} and axion-like-particles (bottom) (two
photon coupling $g$ vs. mass $m_{a}$ of the
ALP)~\cite{Andriamonje:2007ew,Schlattl:1998fz,Inoue:2008zp}. See the
text for details. Note that the mass region, where the axion can
be the cold dark matter (the orange regions labeled ``CDM" in the
plots), can be extended towards smaller masses (larger $f_a\lesssim
10^{16}$\,GeV) by anthropic reasoning. Moreover, in the first plot
the areas marked ``ADMX'' and ``CAST'' show the near future search
ranges. In the second plot the axion band is shown hatched.
We have also marked other
areas with interesting astrophysical hints in orange. For
comparision, we also show laboratory limits from photon regeneration
experiments (ADMX and LSW) as discussed in Section~\ref{searches}.
(Both compilations extended from Ref.~\cite{Redondo:2008en}.) Note
that the limit from ADMX is valid only under the assumption that the
local density of ALPs at earth is given by the dark matter density.
}\label{Fig:axions_astro}
\end{figure}
Stars evolve fusing increasingly heavier nuclei in their cores. Heavier nuclei require hotter environments, and when a nuclear species is exhausted in the core, the latter slowly contracts and heatens up until it reaches a new burning phase. WISP emission shortens normal burning phases, since the energy loss rate is higher than standard but the total energy is limited by the number of nuclei. On the other hand, WISP emission prolongs intermediate (red giant) phases, since WISP cooling delays reaching the appropriate temperature during the core contraction.
These effects have been used to constrain a variety of WISPs in different stellar environments~\cite{Raffelt:2006cw,Raffelt:1996wa} for which information on evolutionary time scales is available.
For the standard QCD axions, the best constraints come from white dwarf cooling~\cite{Raffelt:1985nj,Isern:2008nt} through the coupling to electrons and from the duration of the SN1987A neutrino burst~\cite{Raffelt:2006cw} through the coupling
to nucleons (cf. Fig.~\ref{Fig:axions_astro} (top)).
The strongest limits for general ALPs with a two photon coupling and MCPs come from observations of Horizontal Branch (HB) stars in
globular clusters (GC)~\cite{Raffelt:1985nk,Raffelt:1987yu}
(cf. Figs.~\ref{Fig:axions_astro} (bottom) and \ref{Fig:mcp_astro}). For very small masses an even tighter limit on a two-photon coupling of ALPs
can be obtained from the absence of a $\gamma$-ray burst in coincidence with a neutrino burst during the explosion SN 1987a~\cite{Brockway:1996yr}.
The principle behind the
latter bound is that ALPs would be produced in the supernova core from the Primakoff effect and reconverted into $\gamma$-rays inside the galactic magnetic field.
\begin{figure}[t!]
\centerline{\includegraphics[angle=0,width=1.0\textwidth]{MCPbounds.eps}}
\caption{\baselineskip 16pt
Summary of cosmological and astrophysical constraints for minicharged particles (fractional charge
$\epsilon =Q_\epsilon/e$ vs. mass $m_{\epsilon}$) (compilation from Ref.~\cite{Goodsell:2009xc}).
See the text for details. In addition we also show the laboratory limits discussed in Sect.~\ref{searches}. Moreover, at relatively large
masses and couplings we also have the bounds from accelerator and fixed target experiments (SLAC).}
\label{Fig:mcp_astro}
\end{figure}
The sun is less sensitive than these other stars to axion or MCP emission, even though its properties are better known. Solar bounds have been obtained from studies of its
lifetime, helioseismology and the neutrino flux~\cite{Schlattl:1998fz,Gondolo:2008dd}, but although the data is more precise the resulting constraints are weaker.
However, as is apparent in Fig.~\ref{Fig:hp_astro}, this is different for hidden photons:
the region in parameter space excluded by the solar lifetime~\cite{Redondo:2008aa} complements
in this case the one excluded by the lifetime of HB stars~\cite{Redondo:2008ec}.
\begin{figure}[t!]
\centerline{\includegraphics[width=1.0\textwidth]{hpbounds.eps}}
\caption{\baselineskip 16pt
Summary of cosmological and astrophysical constraints for hidden photons (kinetic mixing
$\chi$ vs. mass $m_{\gamma^\prime}$) (compilation from Ref.~\cite{Redondo:priv}).
See the text for details. In addition we also show laboratory limits (see Sect.~\ref{searches} for details on the
constraints in the sub-eV regions; at higher mass we have electroweak precision measurements (EW), bounds from upsilon decays ($\Upsilon_{3S}$)
and fixed target experiments (EXXX)). Areas that are especially interesting are marked in light orange.}
\label{Fig:hp_astro}
\end{figure}
The stellar bounds are very strong but also somewhat vulnerable: they can be considerably relaxed
if the couplings to photons effectively depend on environmental conditions such as the temperature and matter
density~\cite{Jaeckel:2006xm}. This definitely occurs in some specific models, such as the above-mentioned chameleons or in those presented
in Refs.~\cite{Masso:2005ym,Masso:2006gc}.
\subsection{Bounds from Big Bang Nucleosynthesis}
Big Bang Nucleosynthesis (BBN) provides us with a unique probe of
the early universe (for a recent review, see
Ref.~\cite{Iocco:2008va}). At temperatures below $\sim$~MeV, the
weak reactions $p+e^-\leftrightarrow n+ \nu_e$ in the primordial
plasma became ineffective, fixing the neutron/proton density ratio
to $n/p\sim 1/7$. In fact, this ``freeze-out" ratio crucially
depends on the rate of cosmic expansion $H$, which in turn grows with increasing total energy
density $\rho$ of all particles in
the primordial plasma: the larger $\rho$, the sooner the $p$-$n$
freezing the closer n/p becomes to the high temperature value of
1/2. After decoupling, during the proper primordial nucleosynthesis, neutrons are mostly confined into $^4$He nuclei whose
primordial abundance can be measured today, leading to a bound on
the non-standard energy density $\rho_x$ during BBN, usually
expressed as the effective number of extra thermal neutrino species,
\begin{equation}
N_{\nu,x}^{\rm eff}\equiv \frac{4}{7}\frac{30}{\pi^2 T^4}\rho_x .
\end{equation}
A recent determination of
this number~\cite{Simha:2008zj} resulted in
\begin{equation}
N_{\nu,x}^{\rm eff} = -0.6_{-0.8}^{+0.9},
\end{equation}
for three standard neutrinos.
Therefore, while an extra neutral spin-zero particle thermalized during BBN is allowed,
this is not the case for other WISPs like a mini-charged particle, for which
\begin{equation}
N_{\nu,\rm MCP}^{\rm eff}\geq 1,
\end{equation}
or a massive hidden photon, with
\begin{equation}
N_{\nu,\gamma^\prime}^{\rm eff}=21/16 .
\end{equation}
The interactions of MCPs and $\gamma^\prime$s with the standard bath should not allow thermalization before BBN.
MCPs $\psi$ are produced with a rate $\Gamma(e^+ e^-\to \psi\ \overline \psi)\sim \alpha^2 \epsilon^2 T/2$, while $\gamma^\prime$s are produced with rate $\Gamma(\gamma e^\pm \to \gamma' e^\pm)\sim \chi^2_{\rm eff}\Gamma_{\rm C}$ with $\Gamma_{\rm C}$ the standard Compton scattering rate. Here $\chi_{\rm eff}$ is the effective $\gamma-\gamma^\prime$ mixing in the plasma, which for sub-eV $\gamma^\prime$ masses is $\chi_{\rm eff}\simeq \chi(m_{\gamma^\prime}/\omega_{\rm P})^2$. The ratio of the $\gamma^\prime$ mass to the plasma frequency, $m_{\gamma^\prime}/\omega_{\rm P}$, is extremely small before BBN so it suppresses $\gamma^\prime$ production with respect to other WISPs.
Correspondingly, one finds, from a comparison
with the expansion rate $H$, that MCPs with $\epsilon < 2\times 10^{-9}$
would be allowed~\cite{Davidson:2000hf} (cf. Fig.~\ref{Fig:mcp_astro}, labelled ``BBN"), but there are no significant bounds for
hidden photons~\cite{Masso:2006gc}.
\subsection{Bounds from the Cosmic Microwave Background\label{Sec:bounds_cmb}}
The cosmic microwave background (CMB) features an almost perfect
blackbody spectrum with ${\cal O}(10^{-5})$ angular anisotropies. It
is released at a temperature $T\sim 0.1$ eV, but the reactions
responsible for the blackbody shape freeze out much earlier, at
$T\sim$ keV. Reactions like $\gamma+...\to$WISP$+...$ would have
depleted photons in a frequency dependent way, which can be
constrained by the precise FIRAS spectrum
measurements~\cite{Fixsen:1996nj}. This can be used to constrain
light MCPs and ALPs~\cite{Melchiorri:2007sq} as well as hidden
photons~\cite{Jaeckel:2008fi}. More generally~\cite{Mirizzi:2009iz},
(resonant) production of hidden photons leads to distortions in the
CMB spectrum measured by FIRAS strongly constraining their existence
in a wide mass range, as can be seen from Fig.~\ref{Fig:hp_astro}
(similar bounds can be obtained for ALPs but they depend on the
unknown strength of the intergalactic magnetic field~
\cite{Mirizzi:2009nq}). Similarly, in presence of MCPs, when the CMB
photons pass through the magnetic field of clusters this leads to a
local distortions of the CMB spectrum in the direction of the
cluster. Such distortions are constrained by
measurements~\cite{Lawrence} of the so-called Sunaev-Zel'dovich (SZ)
effect and lead to strong bounds on
MCPs~\cite{Burrage:2009yz}\footnote{\baselineskip 16pt Analogously light from distant
supernovae passing through the (less well known) intergalactic magnetic field
would be dimmed by MCP production, again constraining the existence
of such particles~\cite{Ahlers:2009kh} (SN dimming in
Fig.~\ref{Fig:mcp_astro}).}. On the other hand, around $T\sim$ eV
the primordial plasma is so sparse that WISPs would free-stream out
of the density fluctuations, diminishing their contrast. Moreover,
thermal WISPs contribute to the \emph{radiation} energy density,
delaying the matter-radiation equality and reducing the contrast
growth before decoupling. In this respect, they behave identically
to standard neutrinos~\cite{Ichikawa:2008pz}. Therefore, the extra
contribution to the energy density, $\rho_x$ (and the couplings that
would produce it), can again be constrained from the value of
$N_\nu^{\rm eff}$ inferred from analysis of CMB anisotropies and
other large scale structure (LSS) data, e.g. from a recent
analysis~\cite{Simha:2008zj}
\begin{equation}
N^{\rm eff}_{\nu,x} = -0.1^{+2.0}_{-1.4}
.
\label{cmb_lss_bound}
\end{equation}
This argument has been used to constrain axions~\cite{cosmoaxion}
(cf. Fig.~\ref{Fig:axions_astro} (top), labelled ``hot DM") and meV $\gamma^\prime$s~\cite{Jaeckel:2008fi}
(cf. Fig.~\ref{Fig:hp_astro}, labelled ``FIRAS+hCMB").
However, it should be noted that in the determination of the value of
$N^{\rm eff}_{\nu,x}$ in Eq.~(\ref{cmb_lss_bound}),
Ly-$\alpha$ forest data has been deliberately omitted.
Ly-$\alpha$ has systematically favored values of $N_{\nu,x}^{\rm eff}$ larger than
zero~\cite{Seljak:2006bg}, which could be revealing the existence of a cosmic WISP relic density
(cf. Sec.~\ref{Sec:hidden_cmb}). Alternatively, it may be
due to an incorrect treatment of the bias parameter~\cite{Hamann:2007pi}.
\subsection{Possible Indirect Hints for WISPs\label{hints}}
However, there are also other cosmological and astrophysical puzzles which -- interpreted in terms of WISPs --
may indicate that the latter are just around the corner and that it is of high interest to search for them under
controlled laboratory conditions.
\subsubsection{Axions as Cold Dark Matter}\label{axiondm}
First of all, this concerns the possibility that axions constitute
the cold dark matter (CDM) in the universe. In fact, for very weak
coupling, i.e. large decay constant $f_a$, the ultra-light axions
are produced non-thermally in the early universe. At early times,
at temperatures well above the QCD phase transition, the axion is
effectively massless and the corresponding field can take any value,
parameterized by the ``misalignment angle" $\theta_i$. Later, as the
temperature of the primordial plasma falls below the hadronic scale,
$T\lesssim $~GeV, the axion develops its mass $m_a$ due to
non-perturbative (topological instanton) effects. When the mass
$m_a$ becomes of order the Hubble expansion rate, the axion field
will start to oscillate around its mean value $\langle a\rangle =0$.
These coherent and spatially uniform oscillations correspond to a
coherent state of non-relativistic axion particles, whose
contribution to today's energy density, in terms of the critical
energy density, can be estimated\footnote{\baselineskip 16pt It should be noted that
this is a relatively crude estimate. In principle the amount of axions also depends
on the order in which cosmological events took place. In particular, whether the breaking of the Peccei-Quinn symmetry
occurred before or after inflation. In the latter case, for example, there are additional contributions
from the formation and decay of cosmic strings and domain walls. Typically, these contributions
are of similar order to the axion density produced in the misalignment mechanism~\cite{Sikivie:2006ni}.} as~\cite{Preskill:1982cy}
\begin{equation}
\Omega_a h^2 = \kappa_a \left( \frac{f_a}{10^{12}\ {\rm GeV}}\right)^{1.175} \theta_i^2,
\end{equation}
where $0.5\lesssim \kappa_a\lesssim $~few. Therefore, for generic
values of the misalignment angle, $\theta_i=\mathcal{O}(1)$, the
axion could be the main constituent of CDM in the universe,
$\Omega_{\rm CDM} h^2 \sim 0.1$, if its decay constant is of order
$f_a\sim 10^{12}$~GeV, corresponding to a mass in the $10\,\mu$eV range.
Larger values of $f_a$ (cf. Fig.~\ref{Fig:axions_astro} (top);
labeled ``CDM") are excluded because they would lead to an
overclosure of the universe. However, if the initial $\theta_i$ is small,
values of $f_a$ near the GUT or Planck scale are still possible.
This could be due to a finetuning but also due to anthropic
selection~\cite{Hertzberg:2008wr}.
Axion dark matter may indeed explain two additional puzzling
observations: firstly there is an interesting alignment in the multipoles of the CMB \linebreak anisotropies~\cite{Tegmark:2003ve}
and secondly the rotational curves of a number of galaxies provides
evidence for an additional structure in the galactic halo, so-called caustic rings~\cite{Kinney:1999rk}.
Both observations may be explained by Bose-Einstein condensation of dark matter axions~\cite{Sikivie:2009fv}.
Finally, in the context of dark matter it should be noted that axions could also contribute a
hot dark matter component. As for neutrinos the fraction of this hot dark matter is proportional to the mass of
the axion. Recent constraints on the size of a possible hot dark matter component provide a
bound of $m_{a}\lesssim 1.2\,{\rm eV}$ (see Fig.~\ref{Fig:axions_astro} (top))~\cite{cosmoaxion}.
\subsubsection{Non-Standard Energy Loss in White Dwarfs}
A possible non-standard energy loss has been recently identified in the white dwarf luminosity function~\cite{Isern:2008nt}.
As pointed out by the authors this is compatible with the existence of axions with an axion-electron coupling,
\begin{equation}
g_{eea}\simeq 10^{-13},
\end{equation}
suggesting an axion decay constant and axion mass\footnote{\baselineskip 16pt For the mass of an ALP actually everything below
$\sim$ keV would be acceptable.} of
\begin{equation}
f_a\sim g_{eea} m_e \sim {\rm few}\times 10^{9}\ {\rm GeV},\hspace{6ex}
m_a\sim {\rm meV},
\label{fa_benchmark_axion_wd}
\end{equation}
respectively.
In typical models, the coupling to photons is then
\begin{equation}
g\sim \alpha/f_a \sim 10^{-12}\ {\rm GeV}^{-1},
\label{g_benchmark_axion_wd}
\end{equation}
i.e. very close to the stellar evolution bounds.
But, apart from the obvious possibility that
a more conventional explanation may be found for this non-standard energy loss, it should
be noted that this could also be explained in terms of hidden photons coupling via
non-renormalizable operators to electrons, cf. Ref.~\cite{Hoffmann:1987et}.
Nevertheless, the values in Eqs.~(\ref{fa_benchmark_axion_wd}) and
(\ref{g_benchmark_axion_wd}) should be considered as benchmark values for
future axion searches. Fortunately, as we will see in the next section, this region of parameter space
is not completely off the possibilities to be checked in a controlled laboratory experiment
based on photon regeneration.
\subsubsection{Hints for Cosmic Photon Regeneration\label{Sec:hints_cosmic}}
It has been argued that recent observations in TeV gamma astronomy may point towards the
existence of ALPs or at least will help to give sensible new constraints on their existence.
Quite distant astrophysical sources have been observed in TeV gamma rays by H.E.S.S. and
MAGIC. This appeared to be quite puzzling, since the gamma ray absorption rate
due to electron/positron pair production off the extragalactic background light (EBL)
was believed to be too strong to allow
for their observation~\cite{Aharonian:2005gh,Mazin:2007pn}. Clearly, a conventional explanation is either
that the EBL is less dense than expected and/or that the source
spectra are harder as previously thought. Alternatively,
such a high transparency of the universe may also be explained by the conversion of
gamma rays into ALPs in the magnetic fields around the gamma ray sources. These ALPs would travel then unimpeded
until they reach our galaxy and reconvert into photons in the galactic magnetic
fields~\cite{Hochmuth:2007hk}. Alternatively, the conversion/reconversion could take place in
the intergalactic magnetic fields~\cite{DeAngelis:2007dy} (see Ref.~\cite{Mirizzi:2009aj}
for a comprehensive bibliography and the current status).
The intergalactic magnetic fields are not well known but the assumption of being organized in randomly oriented patches would produce
a characteristic scatter in luminosity relations.
A powerful signature of cosmic photon regeneration could therefore emerge if the reconstructed
EBL along different lines of sight towards different TeV gamma sources were to display such a
characteristic scatter~\cite{Mirizzi:2009aj}.
ALPs may also leave their imprints in luminosity relations of active galactic nuclei.
In fact, mixing between photons and ALPs in the random magnetic fields in galaxy clusters
may induce a characteristic scatter in the relations of X-ray vs. optical luminosities of compact sources
in these clusters. Evidence for such an effect has recently been found in an analysis of
luminosity relations of about two hundred active galactic nuclei, providing
a strong hint for the possible existence of a very light axion-like particle~\cite{Burrage:2009mj}.
Furthermore, photon - ALP mixing could also explain the puzzling origin of the (debated)
correlation of arrival directions of ultra high energy cosmic rays and BL-Lac objects observed by AGASA and
HiRes~\cite{Fairbairn:2009zi} (see, however, Ref.~\cite{Albuquerque:2010rq}).
Moreover, the observed alignment of the polarization vectors of very distant quasars
may also be explained by selective photon disappearance from photon-ALP oscillations~\cite{Payez:2008pm}.
Finally, it is also worth nothing that very similar ALPs have been invoked to solve some problematic aspects of the X-ray activity of the Sun,
the longstanding corona problem and the triggering of solar flares~\cite{Zioutas:2008ie}.
Intriguingly, in all the above mentioned hints for cosmic photon regeneration, the required ALP should
have a very small mass, say~\cite{Burrage:2009mj}
\begin{equation}
m_{a}\ll 10^{-12}\,{\rm eV}\div 10^{-9}\,{\rm eV},
\label{mass_benchmark_ALP_transp}
\end{equation}
and a coupling in the range
\begin{equation}
g\sim 10^{-12}\div 10^{-11}\ {\rm GeV}^{-1}.
\label{g_benchmark_ALP_transp}
\end{equation}
Again, these values should be considered as important benchmarks for future
laboratory searches for ALPs. As we will see, photon regeneration experiments can do a very good
job here.
\subsubsection{Dark matter from hidden photons\label{Sec:dm_hidden}}
Dark matter could also be directly related to hidden photons.
Indeed there are several different way in which a hidden U(1) could contribute to dark matter.
First of all in the region $(\chi,m_{\gamma^\prime} )\sim (10^{-12}, 0.1\ {\rm MeV}$), labeled ``luke DM" in
Fig.~\ref{Fig:hp_astro}, the hidden photon itself could be a lukewarm dark matter candidate~\cite{Pospelov:2007mp,Redondo:2008ec}.
For somewhat larger mixing angles $(\chi,m_{\gamma^\prime} )\sim (10^{-11}, \lesssim 100\,{\rm GeV}$), in the region labeled ``Hidden
Photino DM" in Fig.~\ref{Fig:hp_astro}, the supersymmetric partner
of the hidden photon, the hidden photino, is a promising dark matter candidate, if its mass is
in the 10 to 150 GeV range~\cite{Ibarra:2008kn}.
Moreover, for $(\chi,m_{\gamma^\prime} )\sim (10^{-23}, 0$), the hidden photino, with mass in the TeV range, could be a candidate of decaying dark matter, giving rise to
the above mentioned excesses observed in galactic cosmic
ray positrons and electrons~\cite{Shirai:2009kh}.
Finally, in the region $(\chi,m_{\gamma^\prime} )\sim (10^{-4}, {\rm GeV}$), labeled ``Unified DM" in
Fig.~\ref{Fig:hp_astro} the hidden photon
plays an important role in models where the dark matter resides in the
hidden sector~\cite{ArkaniHamed:2008qn}. These models aim at a unified description of unexpected observations
in astroparticle physics, notably the positron excess observed by
the satellite experiment PAMELA \cite{Adriani:2008zr} and the annual modulation signal seen by the direct dark matter search experiment DAMA~\cite{Bernabei:2008yi}.
The massive hidden U(1) can then mediate ``Dark Forces''.
These values are also accessible to accelerator searches~\cite{ArkaniHamed:2008qp,Baumgart:2009tn}
and have been motivated in various supersymmetric scenarios~\cite{Chun:2008by,Baumgart:2009tn}
(see also Ref.~\cite{Suematsu:2006wh}).
\subsubsection{A hidden CMB?\label{Sec:hidden_cmb}}
As discussed in Sect.~\ref{Sec:bounds_cmb}, hidden photons contribute to the effective number of neutrinos.
Indeed, such a contribution would lead to a higher number of effective neutrinos at the time of CMB and large scale structure
formation than at the time of BBN. Interestingly, some global cosmological analyses
that take into account precision cosmological
data on the cosmic microwave background and on the large scale structure of the universe
appear to require some extra radiation energy density from invisible particles apart from the three known neutrino
species. The case for this was strengthened by the recently released WMAP 7 year data~\cite{Komatsu:2010fb}.
Hidden photons in the parameter region $(\chi,m_{\gamma^\prime} )\sim (10^{-6}, 0.2\ {\rm meV}$), labeled ``hCMB" in
Fig.~\ref{Fig:hp_astro}, lead to a natural explanation of this finding~\cite{Jaeckel:2008fi}.
As a side remark we note that these parameter values also allow for interesting technological
applications of hidden photons~\cite{Jaeckel:2009wm}.
\section{WISP Searches with Low-Energy Photons\label{searches}}
One of the most striking features of many new light bosons is that
one can have photon -- light boson oscillations in very much the
same way as the different neutrino species oscillate into each
other. Below we will start with the description of so-called ``light
shining through a wall'' (LSW) experiments, which most directly make use of
this oscillation phenomenon. We will also use this opportunity to
introduce the basic equations governing these oscillation phenomena.
\subsection{Photon Regeneration Experiments}
\subsubsection{Light Shining Through a Wall -- Theory}
One of the most striking consequences of the photon -- light boson
oscillations is the possibility of ``light shining through a wall''~\cite{Okun:1982xi,Anselm:1986gz,VanBibber:1987rq}.
This is exploited in experiments of the same name. A schematic setup
is shown in Fig.~\ref{Fig:lsw}. The idea is as follows. If an
incoming photon is somehow converted into a WISP the latter can
transverse an opaque wall without being stopped. On the other side
of the wall the WISP could then reconvert into a photon.
\begin{figure}[t!]
\begin{center}
\scalebox{0.5}[0.5]{
\begin{picture}(386,130) (159,-207)
\SetWidth{1.0}
\SetColor{Black}
\GBox(336,-206)(368,-78){0.882}
\Line[dash,dashsize=8.4,arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](272,-142)(432,-142)
\Photon(160,-142)(272,-142){7.5}{6}
\Photon(432,-142)(544,-142){7.5}{6}
\Vertex(432,-142){8}
\Vertex(432,-142){8}
\Vertex(272,-142){8}
\SetOffset(0,-10)
\Text(224,-110)[c]{\Large{\Black{$\gamma$}}}
\Text(480,-110)[c]{\Large{\Black{$\gamma$}}}
\Text(304,-110)[c]{\Large{\Black{$X$}}}
\Text(400,-110)[c]{\Large{\Black{$X$}}}
\end{picture}
}
\end{center}
\caption{\baselineskip 16pt
Schematic of a ``light-shining-through a wall'' experiment.
An incoming photon $\gamma$ is converted into a new particle $X$
which interacts only very weakly with the opaque wall. It passes
through the wall and is subsequently reconverted into an ordinary
photon which can be detected.
}\label{Fig:lsw}
\end{figure}
This type of experiment is sensitive to a whole variety of WISPs as shown in Fig.~\ref{Fig:zoo}.
In particular, the classic axion or axion-like particles can be searched for by
employing a magnetic field in the conversion regions.
This facilitates the conversions of photons into
axions via the two photon interaction predicted for axion-like
particles (cf. Fig.~\ref{zooalp}).
\begin{figure}[t!]
\begin{center}
\subfigure[]{\scalebox{0.3}[0.3]{
\begin{picture}(322,141) (95,-63)
\SetWidth{1.0}
\SetColor{Black}
\Photon(96,13)(176,13){7.5}{4}
\GBox(240,-51)(272,77){0.882}
\Line[dash,dashsize=2,arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](176,13)(336,13)
\Photon(176,13)(176,-51){7.5}{3}
\COval(176,-51)(11.314,11.314)(45.0){Black}{White}\Line(170.343,-56.657)(181.657,-45.343)\Line(170.343,-45.343)(181.657,-56.657)
\Photon(336,13)(336,-51){7.5}{3}
\Photon(336,13)(416,13){7.5}{4}
\COval(336,-51)(11.314,11.314)(45.0){Black}{White}\Line(330.343,-56.657)(341.657,-45.343)\Line(330.343,-45.343)(341.657,-56.657)
\end{picture}
}
\label{zooalp}}
\hspace*{0.2cm}
\subfigure[]{\scalebox{0.3}[0.3]{
\begin{picture}(322,130) (95,-73)
\SetWidth{1.0}
\SetColor{Black}
\Photon(96,2)(176,2){7.5}{4}
\GBox(240,-62)(272,66){0.882}
\Photon(336,2)(416,2){7.5}{4}
\ZigZag(176,2)(336,2){7.5}{8}
\SetWidth{3.0}
\Line(160.002,17.998)(191.998,-13.998)\Line(191.998,17.998)(160.002,-13.998)
\Line(320.002,17.998)(351.998,-13.998)\Line(351.998,17.998)(320.002,-13.998)
\end{picture}
}
\label{zoohp} } \hspace*{0.2cm}
\subfigure[]{\scalebox{0.3}[0.3]{
\begin{picture}(386,141) (63,-63)
\SetWidth{1.0}
\SetColor{Black}
\GBox(240,-51)(272,77){0.882}
\Photon(128,13)(64,13){7.5}{3}
\Arc[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](160,13)(32,270,630)
\ZigZag(192,13)(320,13){7.5}{6}
\Arc[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](352,13)(32,270,630)
\Photon(384,13)(448,13){7.5}{3}
\Photon(146,-16)(127,-52){7.5}{2}
\Photon(176,-15)(192,-51){7.5}{2}
\Photon(336,-15)(320,-51){7.5}{2}
\Photon(368,-14)(384,-52){7.5}{2}
\COval(128,-51)(11.314,11.314)(45.0){Black}{White}\Line(122.343,-56.657)(133.657,-45.343)\Line(122.343,-45.343)(133.657,-56.657)
\COval(192,-51)(11.314,11.314)(45.0){Black}{White}\Line(186.343,-56.657)(197.657,-45.343)\Line(186.343,-45.343)(197.657,-56.657)
\COval(320,-51)(11.314,11.314)(45.0){Black}{White}\Line(314.343,-56.657)(325.657,-45.343)\Line(314.343,-45.343)(325.657,-56.657)
\COval(384,-51)(11.314,11.314)(45.0){Black}{White}\Line(378.343,-56.657)(389.657,-45.343)\Line(378.343,-45.343)(389.657,-56.657)
\end{picture}
}
\label{zoomcp}}
\end{center}
\caption{\baselineskip 16pt
Explicit processes contributing to LSW for various WISPs.
From left to right we have photon -- ALP, photon -- hidden photon
and photon -- hidden photon oscillations facilitated by
MCPs.}\label{Fig:zoo}
\end{figure}
These oscillation phenomena can be described using a non-diagonal
mass term in the equations of motion decribing the photon $A$
(with energy $\omega$) and
the new particle $X$ (cf. also Ref.~\cite{Raffelt:1987im}),
\begin{equation}
\label{eom}
[\omega^2\mathbf{1}+\partial^{2}_{z}\mathbf{1}-{\mathcal{M}}^{X}]
\left(
\begin{array}{c}
A \\
X \\
\end{array}
\right)=0,
\end{equation}
where we have suppressed the Lorentz structure. Indeed, for the
types of particles discussed in Sect.~\ref{physicscase} the
equations of motion always separate into the two possible linear
polarizations but the mass matrix may differ for the different
polarization directions.
The solutions to the equations of motion are of the form,
\begin{equation}
v_{1}=\exp(-\mathbf{i}(\omega t-k_{1}z))\left(
\begin{array}{c}
1 \\
\delta \\
\end{array}
\right),
\quad
v_{2}=\exp(-\mathbf{i}(\omega t-k_{2}z))\left(
\begin{array}{c}
-\delta \\
1 \\
\end{array}
\right).
\end{equation}
If the off-diagonal entry in the mass matrix is small we can obtain
simple analytical formulas for the mixing angle,
\begin{equation}
\tan(2\,\delta)=2\frac{{\mathcal{M}}^{X}_{12}}{{\mathcal{M}}^{X}_{11}-{\mathcal{M}}^{X}_{22}},
\end{equation}
and the wave numbers for the two mass eigenstates,
\begin{equation}
k^2_{1}=\omega^2-{\mathcal{M}}^{X}_{11},\quad
k^{2}_{2}=\omega^{2}-{\mathcal{M}}^{X}_{22}.
\end{equation}
Using these it is straightforward to find the transition amplitudes,
\begin{equation}
\label{amplitude} A(\gamma\to
X)=\delta\left[\exp(\mathbf{i}k_{1}z)-\exp(\mathbf{i}k_{2}z)\right],
\end{equation}
from which we can obtain
\begin{eqnarray}
\label{transition} P(\gamma\to X,\ell)\!\!&=&\!\!P(X\to\gamma,\ell)=|A(\gamma\to X)|^2
\\\nonumber
\!\!&=&\!\!|\delta|^2[\exp(-2{\rm Im}(k_{1})\ell)+\exp(-2{\rm Im}(k_{2})\ell)
\\\nonumber
&&\quad\quad\quad\quad\quad\quad\quad\quad\quad-2\exp(-{\rm Im}(k_{1}+k_{2})\ell)\cos({\rm Re}(k_{1}+k_{2})\ell)].
\end{eqnarray}
In a light shining through a wall experiment the photon must convert
into a WISP and back. Therefore the total probability for a photon
to arrive at the detector is\footnote{\baselineskip 16pt Here, we neglect the typically
very small absorption probability of the WISP $X$ inside the wall.},
\begin{equation}
\label{lswprob} P(\gamma\to X,\ell_{1})P(X\to\gamma,\ell_{2}).
\end{equation}
This probability can be enhanced by using optical cavities to
reflect the light back and forth inside the production and
regeneration regions~\cite{Hoogeveen:1990vq,Sikivie:2007qm}. In the production region one can easily imagine that
the probability is enhanced by the number of passes towards the
wall. On the regeneration side such an enhancement is not as
obvious. Nevertheless, a \emph{resonant} cavity on the regeneration
side allows for an additional enhancement corresponding to the
number of passes inside this cavity as well. The probability
including these improvements is therefore,
\begin{equation}
\label{enhanced} P_{\rm LSW}=N_{1}N_{2}P(\gamma\to
X,\ell_{1})P(X\to\gamma,\ell_{2}),
\end{equation}
where $N_1$ ($N_2$) is the effective number of passes in the respective cavity
divided by a factor two.
Let us now consider the three sets of particle discussed in Sect.~3:
axion-like particles, massive hidden photons, and massless hidden
photons with additional minicharged particles.
Axion-like particles
couple to two photons. In this case the photon -- axion-like
particle oscillations have to be facilitated by the presence of a
magnetic field as shown in Fig.~\ref{Fig:lsw} (a) which provides for
one of the two photons in the interaction. The presence of the
magnetic field marks a preferred direction. Therefore the matrix
${\mathcal{M}}$ can now depend on the polarization direction. Indeed one
finds,
\begin{eqnarray}
{\mathcal{M}}^{a^{-}}_{\parallel}\!\!&=&\!\!\left(
\begin{array}{cc}
0 & -gB\omega \\
-gB\omega & m^{2}_{a} \\
\end{array}
\right)
,\quad\quad{\mathcal{M}}^{a^{-}}_{\perp}=0,
\\\nonumber
{\mathcal{M}}^{a^{+}}_{\parallel}\!\!&=&\!\!0,\quad\quad\quad\quad\quad\quad\quad\quad\quad\,\,
{\mathcal{M}}^{a^{+}}_{\perp}=\left(
\begin{array}{cc}
0 & -gB\omega \\
-gB\omega & m^{2}_{a} \\
\end{array}
\right),
\end{eqnarray}
where $a^{-}$ indicates a pseudo-scalar axion-like particle coupling
to $F\tilde{F}$ and $a^{+}$ is a scalar coupling to $F^2$. Moreover,
the subscripts $\parallel,\perp$ indicate the polarization direction
with respect to the magnetic field.
Since the mass matrix is real
the expression for the probability in an LSW experiment, Eqs.~\eqref{transition},
\eqref{lswprob}, simplifies. If the polarization of the laser is at an
angle $\theta$ with respect to the magnetic field, the probability
for a pseudoscalar reads,
\begin{equation}
\label{axionprob}
P_{\rm LSW}=16\frac{(gB\omega\cos(\theta))^4}{m^{8}_{a}}
\sin^2\left(\frac{\ell_{1}m^2_{a}}{4\omega}\right)
\sin^2\left(\frac{\ell_{2}m^{2}_{a}}{4\omega}\right).
\end{equation}
For a scalar the $\cos(\theta)$ has to be replaced by a
$\sin(\theta)$. Moreover, for simplicity we have made the additional
assumption $\omega\gg m_{a}$.
The transition into hidden photons occurs also in absence of a
magnetic field. Accordingly there is no preferred direction and both
polarizations have the same mass-mixing,
\begin{equation}
\label{hiddenmass}
{\mathcal{M}}^{\gamma^{\prime}}=m_{\gamma^\prime}^2\left(
\begin{array}{cc}
\chi^2 & -\chi \\
-\chi & 1 \\
\end{array}
\right).
\end{equation}
The corresponding transitions are depicted in Fig.~\ref{Fig:lsw} (b) and
the transition probability is
\begin{equation}
\label{hpprob}
P_{\rm LSW}=16\chi^4\sin^2\left(\frac{\ell_{1}m_{\gamma^\prime}^2}{4\omega}\right)
\sin^2\left(\frac{\ell_{2}m_{\gamma^\prime}^2}{4\omega}\right).
\end{equation}
Finally, for the combination of massless hidden photons and
minicharged particles we again need a magnetic field to allow for a
transition,
\begin{equation}
{\mathcal{M}}^{\gamma^{\prime}+{\rm MCP}}= -2\omega^2{e^{2}_{{h}}}\Delta
N_{i}\left(
\begin{array}{cc}
+\chi^2 & -\chi \\
-\chi & +1 \\
\end{array}
\right).
\end{equation}
Here, $i=\parallel, \perp$ again
indicate the polarization with
respect to the magnetic field and $\Delta N_{i}$ are the magnetic field dependent, complex refractive indices describing
the refraction and absorption due to the virtual and real production of MCPs (see \cite{Ahlers:2007rd} for details).
\begin{figure}[t!]
\subfigure[]{\includegraphics[width=.5\textwidth]{alpminus.eps}}
\hfill
\subfigure[]{\includegraphics[width=.5\textwidth]{alpplus.eps}}
\subfigure[]{\includegraphics[width=.5\textwidth]{hp.eps}}
\hfill
\subfigure[]{\includegraphics[width=.5\textwidth]{mcp.eps}}
\caption{\baselineskip 16pt
Sensitivities of LSW experiments. Top panels:
pseudoscalar (left) and scalar (right) axion-like particles.
The results from ALPS are preliminary.
Bottom panels: massive hidden photons (left) and massless hidden
photons with an additional minicharged particle (right).
Compilation from Ref.~\cite{ALPS}.}\label{lswresult}
\end{figure}
\subsubsection{Light Shining Through a Wall - Experiments}\label{lswexperiments}
As the name suggests most of these experiments employ laser light in
the optical regime. The typical lengths of the production and
regeneration zones is in the range of a few meters. As we can see
from Eqs.~\eqref{axionprob}, \eqref{hpprob} an oscillation length of
a few meters corresponds to particle masses in the range of meV. For
ALPs these experiments are therefore sensitive for masses up to a
few meV and for hidden photons the sensitivity is usually optimal in
this mass range. A variety of these light shining through walls
experiments has been performed or is currently
running~\cite{Ruoso:1992nx,Cameron:1993mr}.
The results can be seen in
Fig.~\ref{lswresult}. The bounds are in the $g\leq {\rm few} \times
10^{-8}\, {\rm GeV}^{-1}$ range for ALPs and the kinetic mixing and
minicharges probed are typically of the order of ${\rm few}\times
10^{-7}$. Although, this is not yet competitive with the
astrophysical and cosmological arguments -- except for the case of hidden
photons in the meV mass range (cf. Fig.~\ref{lswresult} (c)) --
it should be noted that
these bounds are less model-dependent (cf. \cite{Masso:2005ym,Jaeckel:2006xm,Masso:2006gc}).
Moreover, it is worth noting that so far only two of these
experiments (BFRT and ALPS) have employed mirrors to enhance the
transition probability on the production side\footnote{\baselineskip 16pt Only ALPS
employed a Fabry-Perot cavity whereas BFRT used an optical delay
line. If used only on the production side this does not make a
difference, but for resonant regeneration an optical cavity is
needed.}. Using cavities on both sides~\cite{Hoogeveen:1990vq,Sikivie:2007qm} could lead to significant
improvements in the near future. However, it should be noted that
this presents a technological challenge~\cite{cantatore}
because the cavities have to be tuned to be resonant with each other (and,
of course, a better quality of the cavity reduces the bandwidth
making this even more difficult). Nevertheless, it seems that this
methods paves the way to beat for the first time the astrophysical
bound on the ALP-photon coupling. Moreover, it can even reach
the ALP benchmark point (\ref{g_benchmark_ALP_transp}), $g\sim 10^{-11}$~GeV$^{-1}$,
for small enough ALP mass, therefore possibly testing the ALP
interpretation of the reported hints of cosmic photon regeneration
(cf. Sect.~\ref{Sec:hints_cosmic}).
In order to probe the axion benchmark point (\ref{fa_benchmark_axion_wd}), however, one needs
further ingredients in order to access also higher masses: one possibility
is to exploit alternating magnetic field directions~\cite{VanBibber:1987rq}, another
is to use phase shift plates~\cite{Jaeckel:2007gk}.
Higher masses could also be accessed by exploiting keV photons from X-ray
free-electron lasers or synchrotron radiation sources~\cite{Rabadan:2005dm}.
However, presently the photon
fluxes from these sources appear to be too small to be competitive with
optical photons.
Further improvement in the sensitivity at low masses could also come from
experiments that employ electromagnetic waves in the radiofrequency
range instead of laser light in the visible regime~\cite{Hoogeveen:1992uk}.
The advantage of such ``microwave shining through walls'' is threefold.
First, microwave cavities can be constructed such
that the light is effectively reflected back and forth inside the
cavities up to $10^{11}$ times. This allows for a significant
improvement compared to the currently best optical cavities where
the light can be reflected only up to $10^5$ times. Second, although
still a technological challenge the tuning of the cavities to
resonance is simplified by the fact that extremely frequency-stable
cavities exist in the microwave regime. Finally, using the phase
information of the generator to distinguish between signal and
background one can detect microwaves with intensities less than
$10^{-23}\,{\rm W}$ with commercially available technology. With
currently available technology, sensitivities of the order
$$
g\sim 10^{-10}\,{\rm GeV}^{-1},\quad \chi\sim 10^{-12}
$$
seem achievable. The price to pay is that due to the lower
frequency/energy of the microwaves, the sensitive range is typically
restricted to particle masses less than $0.1\,{\rm meV}$.
Several of these experiments are currently under construction or in
planning stages~\cite{Yale,ADMXcham}.
Finally another variant of the LSW idea is to use static magnetic fields~\cite{Jaeckel:2008sz}.
The setup consists of a highly sensitive magnetometer inside a superconducting shielding.
This is then placed inside a strong (but sub-critical) magnetic field. In ordinary electrodynamics the magnetic field cannot
permeate the superconductor and no field should register on the magnetometer. However, photon -- hidden-sector photon -- photon oscillations would allow to penetrate the superconductor and the magnetic field would ``leak" into the shielded volume and register on the magnetometer.
Compared to the classic LSW setup there are two crucial differences. First, the fields are (nearly) static and the photons involved are virtual.
Second, the magnetometer directly measures the field-strength and not a probability. Correspondingly, the signal is suppressed
only quadratically, $\propto \chi^2$, instead of quartically.
For hidden photon masses in the range 0.002-200 meV the projected sensitivity for the mixing parameter lies in the
$5\times 10^{-9}$ to $10^{-6}$ range.
\subsubsection{Afterglow Experiments}
For chameleon particles (s. Sect.~\ref{chameleonsect}) and in general particles whose mass increases with the local matter/energy density~\cite{Jaeckel:2006xm}, LSW experiments do not
always work. The reason is that the high (compared to the vacuum) density inside the wall increases the mass of the produced WISP
and therefore creates a high potential barrier on which the particle is reflected.
However, this reflection can actually be used in a slightly different type of experiment~\cite{Ahlers:2007st,Gies:2007su}. If the production zone is enclosed by suitably
dense walls on all sides (transparent windows let the photons into the production zone but they are typically too dense to let chameleons escape)
the chameleons produced are trapped inside the production zone and accumulate over time.
After some time the laser is switched off. Now the accumulated chameleons can reconvert into photons which escape through the windows. Hence,
there is an ``afterglow'' after the laser has been switched off. This is shown in Fig.~\ref{fig:Setup}.
\begin{figure}[!t]
\includegraphics[width=4.2cm]{chamcav1.eps}
\hspace*{0.2cm}
\includegraphics[width=4.2cm]{chamcav2.eps}
\hspace*{0.2cm}
\includegraphics[width=4.2cm]{chamcav3.eps}
\caption[]{\baselineskip 16pt Illustration of an afterglow experiment to search for chameleon particles
(from Ref.~\cite{Ahlers:2007st}).
Filling the vacuum tube by means of a laser beam with chameleons
via photon-chameleon conversion in a magnetic field (left).
An isotropic chameleon gas forms (middle).
Afterglow from chameleon-photon conversion in a magnetic field (right).}
\label{fig:Setup}
\end{figure}
Such an experiment has been performed by the GammeV collaboration~\cite{Chou:2008gr} and allows to exclude sub-meV (vacuum mass) chameleons
with couplings in the range $g\sim (10^{-7}-10^{-6})\ {\rm GeV}^{-1}$. Improvements planned by the GammeV collaboration~\cite{gammev}
will further increase the sensitivity. Moreover, by driving the ADMX cavity with an external generator and looking for a ``microwave afterglow'' the ADMX collaboration
has performed a small test experiment for chameleons with masses around $2\,\mu$eV~\cite{ADMXcham}.
\subsubsection{Helioscopes}
As already mentioned in Sect.~\ref{stellar} WISPs can also be
produced inside the sun. This can either occur via the same interactions that lead to the
oscillation phenomena described above or via additional derivative interactions with matter (e.g. for axions).
In any case due to the high number of photons inside the sun and the high total number of interactions,
this typically would lead to the production of a huge number of WISPs.
Helioscopes try to detect these WISPs on earth~\cite{Sikivie:1983ip,vanBibber:1988ge}.
Basically helioscopes employ the same idea as an LSW experiment. However, the production side is replaced by the
WISP production inside the sun. The regeneration side is the same as for a completely laboratory based LSW experiment with the detector pointing towards the
sun. The wall is simply everything in between the solar core and the regeneration side (the rest of the sun, the atmosphere, the walls of the experimental hall etc.).
The enormous total number of interactions inside the sun leads to a large WISP flux even if the coupling is tiny. This
make helioscopes an extremely powerful tool to search for WISPs, however, the same caveats discussed in Sect.~\ref{stellar} apply: if somehow
the production of WISPs inside the sun is suppressed, helioscopes loose their sensitivity.
The typical energy of photons inside the sun is in the keV range, accordingly this is also the typical energy of WISPs produced inside the sun.
Currently two axion helioscopes are running, CAST and SUMICO~\cite{Andriamonje:2007ew,Inoue:2008zp}. The two experiments employ large magnets. Therefore, they are sensitive
to ALPs as well as hidden photons (with and without additional MCPs).
Both experiments have performed searches for
WISPs with keV energies leading to X-ray photons inside the detector. In addition, CAST has also searched for WISPs in the eV regime~\cite{karuza} which
gives weaker but less model dependent constraints on WISPs (cf.~\cite{Jaeckel:2006xm}).
CAST, has recently surpassed the HB constraints for ALPs with a two
photon coupling~\cite{Andriamonje:2007ew} (cf. Fig.~\ref{Fig:axions_astro} (bottom)), and its results have been
used to limit a possible solar $\gamma^\prime$
flux~\cite{Redondo:2008aa,Gninenko:2008pz}.
This is shown in Fig.~\ref{Fig:hp_astro} as part of the purple area. Additional improvements of the CAST apparatus are already in planning~\cite{ferrer}.
Further dedicated helioscopes, an add on to SUMICO~\cite{DedHelio} and a stand alone hidden photon helioscope SHIPS~\cite{SHIPS},
are likely to increase the sensitivity for hidden photons towards smaller masses.
To search for solar axions one could also use the magnetic field of earth for the regeneration and an X-ray satellite
to look through the earth at the solar core~\cite{Davoudiasl:2005nh}.
Alternatively, one could also replace the sun as the source of axions with other cosmic sources~\cite{Fairbairn:2007vj}.
\subsubsection{Direct Axion Dark Matter Searches}
Finally, let us turn to another type of photon regeneration experiment that employs an external source of WISPs: axion dark matter searches.
As the name suggests these experiments are mainly focussed on axions.
As described in Sect.~\ref{axiondm} axions are produced in the early universe and can form all or part of the dark matter.
One can now place the regeneration side of an LSW experiment (with magnet) on earth and wait for dark matter axions to enter the experiment
and be converted into photons. This is the basic idea of an axion dark matter experiment also called axion haloscope~\cite{Sikivie:1983ip}.
In principle this type of experiment is also sensitive to other types of WISPs such as hidden photons. However, due to their different
production mechanisms these typically can only form a small part of the dark matter (which itself puts a constraint on their existence)
which, in addition is typically rather hot, somewhat limiting the sensitivity of haloscope searches for non-axion WISPs. (But a final analysis is still
outstanding.)
The ADMX~\cite{Duffy:2006aa} and CARRACK~\cite{Tada:1999tu} collaborations are currently operating axion haloscopes.
Since the typical masses/energies expected for axion dark matter
lie in the $(1-100)\,\mu$eV range this experiment uses microwave cavities for resonant regeneration (cf. Sect.~\ref{lswexperiments}).
Since dark matter axions are very cold, i.e. their kinetic energy is very small compared to their mass, the
energy of the regenerated photons is basically given by the axion mass. To achieve resonant regeneration the
axion mass therefore has to lie within the bandwidth of the microwave cavities. Since the axion mass is unknown such an experiment scans through a range
of masses by changing the resonance frequency of the cavity.
All in all, as can be seen from Fig.~\ref{Fig:axions_astro}, ADMX achieves enormous sensitivity but currently only for a relatively limited
range of axion masses. Yet, future improvements will significantly increase the scanning speed. With their Phase II upgrade, which
is currently being implemented, ADMX plans to scan the whole range from $10^{-6}$~eV to $10^{-5}$~eV.
It should be noted, however, that regeneration experiments for dark matter axions work only in presence of a background of
axions. Usually, the latter is assumed to be of the size of the local dark matter density and the ADMX bound given in Fig.~\ref{Fig:axions_astro} is obtained under this
assumption.
As an alternative to a microwave cavity experiment one could also
search for dark matter axions using optical cavities filled with laser light.
Photons absorbing dark matter axions would lead to sidebands in the spectrum of the laser light~\cite{Melissinos:2008vn}.
At very small masses (in the finetuning or anthropic region where the initial value of the axion field needs to be small
(cf. Sect.~\ref{axiondm})) the cavities become too large to be practical.
A proposal is to use instead resonant LC-circuits~\cite{Thomas}. This would allow to explore axion dark matter with decay constants up to the GUT scale,
$f_{a}\sim 10^{16}\,{\rm GeV}$.
\subsection{Laser Polarisation Experiments\label{laserpolarisation}}
The experiments described in the previous section search for regenerated photons as a signature of WISPs.
An alternative is to search for the \emph{disappearance} or phase modifications of photons
that would signify the real and virtual production of WISPs~\cite{Maiani:1986md,Gies:2006ca,Ahlers:2007rd}.
This idea is realized in laser polarization experiments.
In these experiments (linearly) polarized laser light is shone through a magnetic field and changes in the polarization are searched for.
The detectable changes are a rotation of the linearly polarized light and also a phase shift between different polarization
directions that modifies linear polarized light into elliptically polarized light.
However, changes in the overall magnitude of the laser field are typically not detectable.
As an example the relevant processes for ALPs and minicharged particles are depicted in Fig.~\ref{polarization}.
\begin{figure}[t!]
\hspace{-0.7cm}
\subfigure[]{
\scalebox{0.8}[0.8]{
\begin{picture}(240,150)(0,0)
\includegraphics[width=9cm]{stream.eps}
\scalebox{0.4}[0.4]{
\SetOffset(-460,260)
\Text(100,-10)[c]{\scalebox{1.6}[1.6]{$\mathbf{B}_{\mathbf{ext}}$}}
\Photon(0,90)(70,90){5}{7.5}
\Photon(70,90)(70,00){5}{10.5}
\DashLine(70,90)(140,90){6}
\SetOffset(-460,262)
\CArc(70,-10)(7.5,0,360)
\Line(75,-5)(65,-15)
\Line(65,-5)(75,-15)
\SetOffset(-460,40)
\Text(100,-10)[c]{\scalebox{1.6}[1.6]{$\mathbf{B}_{\mathbf{ext}}$}}
\Photon(0,90)(70,90){5}{7.5}
\Photon(70,90)(70,0){5}{10.5}
\DashLine(70,90)(130,90){6}
\SetOffset(-460,42)
\CArc(70,-10)(7.5,0,360)
\Line(75,-5)(65,-15)
\Line(65,-5)(75,-15)
\SetOffset(-460,40)
\Photon(130,90)(130,0){-5}{10.5}
\SetOffset(-400,42)
\CArc(70,-10)(7.5,0,360)
\Line(75,-5)(65,-15)
\Line(65,-5)(75,-15)
\SetOffset(-460,40)
\Photon(130,90)(200,90){5}{7.5}}
\end{picture}}
\label{alpconversion}}
\hspace{0.8cm}
\subfigure[]{
\scalebox{0.8}[0.8]{\begin{picture}(240,150)(0,0)
\includegraphics[width=9cm]{stream.eps}
\scalebox{0.4}[0.4]{
\SetOffset(-460,260)
\Text(100,-10)[c]{\scalebox{1.6}[1.6]{$\mathbf{B}_{\mathbf{ext}}$}}
\Photon(0,90)(70,90){5}{7.5}
\Photon(70,40)(70,0){5}{4.5}
\ArrowLine(70,90)(140,90)
\ArrowLine(70,40)(70,90)
\ArrowLine(140,40)(70,40)
\SetOffset(-460,262)
\CArc(70,-10)(7.5,0,360)
\Line(75,-5)(65,-15)
\Line(65,-5)(75,-15)
\SetOffset(-460,40)
\Text(100,-10)[c]{\scalebox{1.6}[1.6]{$\mathbf{B}_{\mathbf{ext}}$}}
\Photon(0,90)(70,90){5}{7.5}
\Photon(70,40)(70,0){5}{4.5}
\ArrowLine(70,90)(130,90)
\ArrowLine(70,40)(70,90)
\ArrowLine(130,40)(70,40)
\ArrowLine(130,90)(130,40)
\SetOffset(-460,42)
\CArc(70,-10)(7.5,0,360)
\Line(75,-5)(65,-15)
\Line(65,-5)(75,-15)
\SetOffset(-460,40)
\Photon(130,40)(130,0){-5}{4.5}
\SetOffset(-400,42)
\CArc(70,-10)(7.5,0,360)
\Line(75,-5)(65,-15)
\Line(65,-5)(75,-15)
\SetOffset(-460,40)
\Photon(130,90)(200,90){5}{7.5}}
\end{picture}}
\label{milliconversion}}
\vspace{-0.6cm}
\caption{\baselineskip 16pt
Rotation (upper half) and ellipticity (lower half) caused by the existence of a light neutral spin-0
boson (left) or a light particle with a small electric charge (right) (figure adapted from \cite{Brandi:2000ty}).}
\label{polarization}
\end{figure}
To calculate the relevant changes in phase and amplitude one can use the same equations of motion as in the previous subsection.
Instead of the photon to WISP transition amplitude the relevant quantity is now the photon to photon amplitude and
the corresponding probability $P(\gamma\to\gamma)$.
Note that in order to obtain an observable effect the changes in the amplitude and phase
of the photons must be different for the different polarizations since changes in the overall magnitude of the laser field are not detectable.
Therefore, in presence of Lorentz symmetry these experiments need
a magnetic field that distinguishes a certain axis in order to generate an observable effect.
For the same reason massive hidden photons without any additional matter, which have a magnetic field and polarization independent mass matrix, Eq.~\eqref{hiddenmass},
are not detectable in laser polarization experiments.
In contrast to LSW experiments which have no Standard Model background\footnote{\baselineskip 16pt Actually this is not quite true. There is a (tiny) background
from the process photon -- graviton -- photon and an additional, wall thickness dependent one, from neutrino tunneling~\cite{Gies:2009wx}.}
there are two processes in QED which provide a background in laser polarization experiments. The first and most prominent is
the QED contribution to the phase shift
. The relevant diagram is the same as in the lower half of Fig.~\ref{milliconversion} just
with ordinary electrons in the loop. Current experiments are only about two orders of magnitude away from the sensitivity required
to measure this effect and the next generation indeed aims to measure it. This limits searches for ALPs via this process
to about $g\gtrsim 10^{-7}\ {\rm GeV}^{-1}$ and for minicharged particles to $\epsilon\gtrsim 10^{-7}$.
The situation is much better for the rotation. Here, the Standard Model background results
from photon splitting, graviton and neutrino production
.
These backgrounds are much much smaller, basically allowing background free discovery potential down to
couplings $g\lesssim 1/M_{P}$ and minicharges $\epsilon \lesssim 10^{-20}$~\cite{Ahlers:2008jt}.
One should note however, that spectacular experimental progress would be needed in order to achieve these sensitivities.
Currently a variety of these experiments have already finished data taking, are running, in construction
or in preparation~\cite{Cameron:1993mr,Zavattini:2005tm}.
Although their basic principle is the same
they differ in that they use pulsed or constant lasers and or magnetic fields.
The polarization experiments described in this section typically only measure the difference in the speed of light between
the two polarization directions. However, using an interferometer where one arm is placed in a magnetic field allows to measure
the change in the speed of light for both polarizations individually. One interesting possibility for this is to equip one
arm of a gravitational wave detector with pulsed magnetic fields~\cite{Dobrich:2009kd}.
In an inhomogeneous magnetic field axions would cause one polarization direction of the laser beam to be split into two parts (similar to a Stern-Gerlach experiment).
Potentially this could also be exploited experimentally to search for axions~\cite{Guendelman:2008jm}.
\subsection{Experiments Using Strong Electromagnetic Fields}
So far we have described the propagation of light using linear equations of motion. All (possibly non-linear) effects of background fields
were included in effective mass matrices that lead to small modifications of the propagation in vacuum.
However, in strong electromagnetic fields non-linear quantum effects can lead to entirely new processes such as, e.g. Schwinger pair production~\cite{Schwinger:1951nm}.
In presence of light particles interacting with the electromagnetic field these effects may set in at much smaller fields than
in ordinary QED where the critical field strength for such effects is set by the electron mass $\sim m^{2}_{e}/e$.
This can be used to search for WISPs. As an example let us consider how minicharged particles can be searched for using Schwinger pair production.
\begin{figure}[t!]
\begin{center}
\includegraphics[angle=0,width=1.0\textwidth]{fig1.eps}
\end{center}
\vspace{-0.6cm} \caption{
\baselineskip 16pt
Schematic illustration of an
\emph{accelerator cavity dark current} (AC/DC) experiment
for searching minicharged particles (from Ref.~\cite{Gies:2006hv}).}
\label{acdc}
\end{figure}
The basic setup \cite{Gies:2006hv} is depicted in Fig.~\ref{acdc}.
In a strong electric field a vacuum pair of charged particles gains energy if the particles are separated by a distance along the lines of the electric field.
If the electric field is strong enough (or the distance large enough) the energy gain can overcome the rest mass, i.e. the virtual particles turn into real particles.
This is the famous Schwinger pair production mechanism~\cite{Schwinger:1951nm}.
After their production the electric field accelerates the particles and antiparticles according to their charge in opposite directions.
This leads to an electric current (dashed line in Fig.~\ref{acdc}).
If the current is made up of minicharged particles the individual particles have very small charges and interact
only very weakly with ordinary matter. Therefore, they can pass even through thick walls nearly unhindered. An electron current, however, would be stopped.
After passage through the wall we can then place an ampere meter to detect the minicharged particle current.
Typical accelerator cavities achieve field strengths of $\gtrsim 25\,{\rm MeV/m}$ and their size is typically of the order of tens of
centimeters.
Precision ampere meters can certainly measure currents as small as $\mu {\rm A}$ and even smaller currents of the order of ${\rm pA}$ seem feasible.
Using the Schwinger pair production
rate we can then estimate the expected sensitivity for such an experiment to be
\begin{equation}
\epsilon_{\rm sensitivity} \sim 10^{-8}\div 10^{-6},\quad{\rm for}\,\,m_{\epsilon}\lesssim {\rm meV}.
\end{equation}
Therefore such an experiment has the potential for significant improvement over the currently best
laboratory
bounds~\cite{Gies:2006ca,Badertscher:2006fm
,Jaeckel:2009dh}, $\epsilon\lesssim{\rm few}\,\, 10^{-7}$ (cf. Fig.~\ref{Fig:mcp_astro}).
\subsection{Fifth Force Experiments\label{Sec:fifth_force}}
Some of the WISP candidates discussed here, such as scalar ALPs or
hidden photons, mediate long range forces between macroscopic
bodies. Therefore, on account of the cumulative effect of a
macroscopic amount of particles, experiments testing the inverse
square law of the Newton and Coulomb forces offer very sensititve
probes of WISPs in a certain mass range~\cite{Adelberger:2009zz}.
Note, that in our discussions of fifth forces we concentrate on
WISPs that interact with photons.
However, searches for
non-Newtonian forces are also sensitive to a whole variety of WISPs
that interact gravitationally or with very weak (effective) Yukawa
interactions. Examples include the dilaton and other moduli,
Kaluza-Klein gravitons but also the chameleons mentioned in
Sect.~\ref{chameleonsect}. For a review
see~\cite{Adelberger:2009zz}.
\begin{figure}[t!]
\begin{center}
\scalebox{0.5}[0.5]{
\begin{picture}(258,162) (111,-95)
\SetWidth{1.0}
\SetColor{Black}
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](112,66)(160,18)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](160,18)(160,-46)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](160,-46)(112,-94)
\Photon(160,-46)(208,-14){7.5}{2}
\Photon(160,18)(208,-14){-7.5}{2}
\Line[dash,dashsize=2](208,-14)(272,-14)
\Photon(272,-14)(320,-46){-7.5}{2}
\Photon(272,-14)(320,18){7.5}{2}
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](368,-94)(320,-46)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](320,-46)(320,18)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](320,18)(368,66)
\end{picture}
}
\end{center}
\caption{\baselineskip 16pt
Exchange of a scalar ALP, coupled two photons via Eq.~(\ref{Lone}),
between two protons, giving rise to a Yukawa-type non-Newtonian force between two neutral test bodies
(from Ref.~\cite{Adelberger:2006dh}).
\label{fig:non_newtonian}}
\end{figure}
\subsubsection{Bounds on ALPs from Searches for Non-Newtonian Forces}
Due to its pseudoscalar (pseudo-)Goldstone boson nature, the exchange of a very light axion
dominantly leads to a spin-dependent force which can not be probed by the most sensitive
experiments exploiting unpolarized bodies. A spin-independent force is generated by the
exchange of two axions\footnote{\baselineskip 16pt Single pseudoscalar exchange leads to long range interactions between
unpolarized test bodies only in presence of CP-violation. For axions
the resulting interaction potential is $\sim \bar{\theta}^2$~\cite{Moody:1984ba,Adelberger:2009zz}.
Although $\bar{\theta}$ is expected to be non-vanishing due to the CP-violation in the weak interactions
its value is very small (the electric dipole moment of the neutron after all tells us $\bar{\theta}<3\times10^{-10}$).},
which leads to a power-law correction to the spin-independent
long-range force between neutral test bodies \cite{Ferrer:1998ue}. The bounds inferred from
corresponding torsion-balance type experiments on $f_a$ are, however, not competitive with
astrophysical bounds~\cite{Adelberger:2006dh}.
This is different for scalar ALPs, whose coupling to two photons occurs via
the effective Lagrangian
\begin{equation}
{\cal L}_1= {g \over 4}\, \phi F^{\mu\nu} F_{\mu\nu} .
\label{Lone}
\end{equation}
This coupling leads, via the radiative corrections shown in Fig.~\ref{fig:non_newtonian},
to a spin-independent non-Newtonian force between test bodies of the Yukawa-type,
\begin{equation}
\propto (gm_p)^2\exp(-m_\phi r).
\end{equation}
From the non-observation of such a force in sensitive
torsion-balance searches for Yukawa violations of the gravitational
inverse-square law one may put a very stringent
limit~\cite{Dupays:2006dp,Adelberger:2006dh},
\begin{equation}
g\lesssim 4\times 10^{-17}\ {\rm GeV}^{-1},
\end{equation}
in the meV mass range and even stronger constraints for smaller masses
(see also Fig.~\ref{fig:non_newtonian_M}).
\begin{figure}[t!]
\centerline{\includegraphics[width=.8\textwidth]{5th-constraints-M.eps}}
\caption{\baselineskip 16pt
Constraints on the two photon coupling $M=1/g$ of a scalar ALP vs.
its mass $m_\phi$ from searches for non-Newtonian forces exploiting torsion balance
techniques (Be/Cu~\cite{Su:1994gu}, Irvine~\cite{Spero:1980zz},
E\"ot-wash~\cite{Kapner:2006si,Adelberger:2006dh},
Stanford~\cite{Chiaverini:2002cb}) or Casimir force searches
(Lamoreaux~\cite{Lamoreaux:1996wh}). Also shown are constraints from
astrophysical observations (Earth/moon~\cite{Dickey:1994}, Astr. 1998~\cite{Onofrio:2006mq}).
Compilation from Ref.~\cite{Dupays:2006dp}.
\label{fig:non_newtonian_M}}
\end{figure}
In order to test for the pseudoscalar axions one can also search for monopole-(spin)dipole ($\sim \bar{\theta}^{1}$) and dipole-dipole
interactions ($\sim \bar{\theta}^{0}$)~\cite{Moody:1984ba,Adelberger:2009zz} (see \cite{Vasilakis:2008yn} for some very recent new data),
where $\bar{\theta}$ is the remaining CP violating angle due to the CP violation in the electroweak sector.
However, for monopole-dipole searches again the smallness of $\bar{\theta}$ limits the sensitivity and dipole-dipole searches are experimentally rather challenging.
\subsubsection{Bounds on WISPs from Searches for Non-Coulomb Forces}\label{fifthforces}
Hidden photons~\cite{Okun:1982xi,Popov:1999} and minicharged particles~\cite{Jaeckel:2009dh} also leave detectable imprints as modifications of Coulomb's law.
For massive hidden photons the non-diagonal mass term in the equations of motion leads to a modification
of the potential between two charges,
\begin{equation}
V(r) = \frac{\alpha}{r}\left( 1 + \chi^2 e^{-m_{\gamma^\prime}r}\right).
\end{equation}
Minicharged particles modify the potential at the one-loop level by the Uehling contribution. At large
distances the deviation is exponential as well,
\begin{equation}
\label{Uehlingapprox}
V(r)\approx \frac{\alpha}{r} \left[1+ \frac{\alpha\epsilon^2}{4\sqrt{\pi}}\frac{\exp(-2mr)}{(mr)^{\frac{3}{2}}}\right]
,\quad\quad\quad\quad\quad\quad{\rm for}\quad mr\gg 1.
\end{equation}
The inverse square-law has been tested in the laboratory by
Cavendish-type experiments, checking for the absence of an electric field inside a
charged conducting sphere~\cite{Williams:1971ms}. These experiments,
originally performed to set limits on the photon mass~\cite{Goldhaber:1971mr} --
which would also give an exponential
deviation from the Coulomb law -- and to constrain also the range and strength of
an electrical fifth force~\cite{Bartlett:1988yy}, give also a strong constraint on
hidden photons, notably in the $\mu$eV range (cf. Fig.~\ref{Fig:hp_astro},
labeled ``Coulomb"). Similarly, they give the best laboratory constraints on minicharged particles in the sub-$\mu$eV range.
This demonstrates that searches for deviations from Coulomb's law are a powerful tool to search for WISPs, giving strong motivation to
improve upon the by now nearly 40 year old experiments.
Static magnetic fields are also modified in presence of hidden photons. In particular, the large scale magnetic fields
of planets can be used to test for hidden photons with very small masses as can be seen from the constraints
``Earth" or ``Jupiter" in Fig.~\ref{Fig:hp_astro} (cf.~\cite{Goldhaber:1968mt,Goldhaber:1971mr}).
\section{Summary and Outlook}\label{conclusions}
For particle physicists the term ``low energy physics'' is usually used for physics below the current high energy frontier, today roughly
the electroweak scale $\sim 100\,{\rm GeV}$. Currently, three areas of ``low energy physics'' are proving to
be fruitful grounds to explore fundamental physics.
First there are experiments just below the electroweak scale, in the regime of roughly $1\,{\rm GeV}$-$100\,{\rm GeV}$. This energy range is particularly interesting for high precision flavor physics
but it could also provide insight into so-called
``Dark Forces'' that have recently attracted a lot of attention in order to resolve puzzling astrophysical observations
(cf. Sect.~\ref{Sec:dm_hidden}).
The second area of interest is to use low energy experiments and observations to test fundamental symmetries, such as e.g. Lorentz symmetry, to an incredibly
high precision (for a review see~\cite{Kostelecky:2007qf}).
Finally, there is the possibility that there exist new particles with very small masses (possibly even sub-eV) but also very weak interactions
with the known Standard Model particles. This is the case discussed in this review.
Light weakly coupled particles with masses below an eV, so-called weakly interacting sub-eV particles (WISPs), are strongly motivated both
from top down as well as from bottom-up considerations. A classic example for a WISP is the axion
(cf. Sect.~\ref{Sec:QCD_axion}).
From a bottom-up point of view the axion is predicted as a consequence of the Peccei-Quinn
solution of the strong CP problem. A new
global symmetry is introduced which is spontaneously broken at a very high energy scale $\sim f_{a}$. As a consequence, the axion as a pseudo-Nambu-Goldstone boson of this symmetry
has both small mass, $m_{a}\sim {\rm meV}(10^{10}\, {\rm GeV}/f_{a})$, and coupling to two photons, $g\sim 10^{-13}\ {\rm GeV}^{-1}(10^{10}\, {\rm GeV}/f_{a})$.
Taking, on the other hand, a top-down approach, in string theory particles with couplings similar to axions, axion-like particles,
seem to be a natural consequence of the compactification of extra dimensions (cf. Sect.~\ref{Sec:axions_string}).
Again, their small couplings arise from the large energy scales involved, the string scale (possibly also the sizes of the extra dimensions).
Searching for such particles is thereby a way to probe very high energy scales far beyond the energy scale in colliders such as LHC.
Axions are by far not the only possible WISP candidates. String models often contain also additional U(1) gauge groups and matter charged under
them (cf. Sect.~\ref{Sec:hidden_string}).
The corresponding extra ``hidden'' photons can mix kinetically with the ordinary electromagnetic photon. Small mixing angles and corresponding small interactions with ordinary matter
arise from high energy scales involved and from the fact that the additional gauge factors are ``far away" in the extra dimensions. Roughly speaking
probing small couplings allows us look beyond our immediate neighborhood and probe the global structure of the compactification.
For extra photons (and matter charged under them) small masses are possible but not necessary.
Astrophysics and cosmology already provide a powerful tool to constrain WISPs and in many regions of parameter space this
sets the standard against which laboratory tests have to measure up (cf. Sect.~\ref{astro}).
The rapidly increasing amount of available data will hopefully further
improve the situation (see also Ref.~\cite{Chelouche:2008ta}).
Beyond this some astrophysical and cosmological observations even provide suggestive hints for the existence of WISPs
(cf. Sect.~\ref{hints}).
Importantly, both axions as well as hidden photons (or particles related to them) may contribute all or part of the dark matter.
In addition to the above there is a number of astrophysical puzzles that can be solved by the presence of WISPs such as, e.g., the
possible observations of high and very high energy cosmic photon regeneration,
the observed alignment of the polarization vectors of very distant quasars,
the energy loss of white dwarfs, and a slight excess in the effective number of neutrino species in the CMB measurements.
These interpretations can and should be tested in the laboratory.
One of the most important features of the WISPs described above is that they have (very weak) interactions with photons.
In combination with their small mass this allows to search for them in photon regeneration experiments,
in laser polarisation experiments, in experiments exploiting strong electromagnetic fields, and also in
fifth force experiments.
Currently, considerable activity takes place in the field of laser light shining through a wall experiments
(cf. Sect.~\ref{lswexperiments}),
which is now entering a new generation.
Important advances in laser technology appear to pave the way to beat the sensitivity of current WISP helioscopes
and to probe the above mentioned explanations of astrophysical puzzles in terms of photon $\leftrightarrow$ WISP oscillations.
Pioneering experiments exploiting instead high-quality microwave cavities for the generation and regeneration of
WISPs are in the commissioning phase. The microwave cavity search for dark matter axions probes a complementary
region in parameter space compared to the other photon regeneration experiments and should provide a definitive
answer whether axions are the dominant part of cold dark matter within the current decade.
Laser polarisation experiments (cf. Sect.~\ref{laserpolarisation}) will continue their quest to detect the QED birefringence. At the same time they
will also considerably improve the bounds they place on WISPs. They will also further help to develop the optical
techniques, such as cavities that are used in light shining through a wall experiments. Moreover, experiments analyzing
the spectrum of laser light in an interferometer also seem to be a promising tool to search for dark matter axions.
Fifth force experiments (cf. Sect.~\ref{Sec:fifth_force}), in particular searches for deviations from the gravitational inverse square law are already
an established tool to test fundamental physics which will continue to improve
significantly, thereby providing new bounds on a variety of WISPs,
in particular light scalar fields. Other WISPs such as hidden photons and minicharged partcles can be probed with high sensitivities
by testing for deviations from Coulomb's law. Although tests of Coulomb's law are already at an impressive precision the currently
best available bounds in the length range of meters are nearly 40 years old. It stands to hope that with current technology considerable improvements
are possible.
All in all using low energy experiments with photons to search for WISPs may give important information about fundamental
particle physics complementary to the one obtainable at high energy colliders.
Already today these experiments provide very strong bounds on light weakly interacting particles.
But even more excitingly the next few years are likely to bring considerable advances and huge discovery potential for new physics.
\section{Acknowledgements}
The authors would like to thank Markus Ahlers, Holger Gies, Mark Goodsell, Axel Lindner, Alessandro Mirrizzi, Javier Redondo and Christoph Weniger
for interesting discussions, helpful suggestions and joyful collaboration on the subjects discussed in this review.
|
\section{Introduction}
Evolutionary game theory \cite{Hof98,Now06}
has led in the past to novel and interesting insights into the complex behavior that can emerge in
a multispecies ecological system (see for example \cite{Bus79,Sza07,Fre09} and references therein).
Cyclic dominance of competing "species" is not restricted to biological systems, but can also be found
in economic and social systems. This competition can lead to a large variety of phenomena, ranging from the
emergence of regular spatio-temporal pattern to chaotic dynamics. Intrinsic fluctuations and omnipresent
nonlinearities yield a rich variety of scenarios, the two extreme cases being ever-lasting biodiversity or rapid
dominance of a single species and extinction of the others.
The three species rock-paper-scissors game
\cite{Fra96a,Fra96b,Pro99,Sza01,Tse01,Ker02,Sza04,Kir04,Rei06,Rei07,Rei08,Cla08,Pel08,Rei08a,Ber09}
is one of the simplest cyclic dominance model where the influence
of fluctuations can be studied systematically. On the level of the mean-field rate equations this system
is characterized by a reactive fixed point, corresponding to species coexistence
where every species yields one third of the total population. When adding fluctuations, one has
to distinguish between spatial and nonspatial systems. In a nonspatial system, the presence of
stochastic fluctuations yields as the final state one of the three absorbing states where only one species survives
whereas the other two have become extinct \cite{Rei06,Ber09}. Interestingly, the 'weakest' species, i.e. the less efficient
predator, survives in the case of asymmetric interactions \cite{Ber09,Fre01}. In a spatial system without diffusion,
species extinction takes place in one dimension \cite{Fra96a,Fra96b,Pro99}, whereas in two dimensions spatio-temporal
pattern can emerge in the form of rotating spirals \cite{Pro99,Tse01,Rei07,Rei08,Pel08,Rei08a}. Recent studies of
two-dimensional lattice models have revealed that the mobility of individuals,
realized in the form of an exchange of individuals, destabilizes the spirals
for larger swapping rates, thus leading to species extinction in the long-time limit \cite{Rei08,Pel08,Rei08a}.
Real-world examples discussed in terms of rock-paper-scissors models range from coral reef invertebrates \cite{Bus79}
to lizard populations in California \cite{Sin96} and from competing bacterial strains \cite{Ker02,Kir04}
to self-organizing Min proteins \cite{Loo08}, to name but a few examples. It is worth noting that some of
the recent theoretical studies of low-dimensional rock-paper-scissors games have been motivated by the intriguing bacteria experiments.
It follows from our brief discussion of the recent work on systems with cyclic dominance
that stochastic fluctuations have important, and sometimes surprising, effects on this type of systems.
Motivated by this observation, we wish to fully elucidate the role of fluctuations in low-dimensional
systems composed of cyclically dominating species.
We therefore propose in the following
to complement this line of research by studying numerically one-dimensional systems of mobile individuals, with species
dependent interaction and swapping rates. As we shall show, biodiversity is made possible in one dimension
through high mobility, which is different to the behavior in two dimensions where an increased mobility
leads to species extinction.
In addition, and in contrast to the nonspatial case, we discover for asymmetric interaction
and swapping rates that the law of the weakest is not strict in the spatial case and that it depends on
the system parameters whether the weakest survives or dies out.
Another motivation for our study of one-dimensional systems comes from the recent studies of bacterial populations
in nanofabricated landscapes (see for example \cite{Key06} where linear arrays of coupled microscale patches of habitat
are discussed). These investigations open the intriguing possibility of a future study of competing bacterial populations
in quasi one-dimensional systems.
Our paper is organized in the following way. After having introduced our model in Section II, we discuss in Section III the
case of symmetric, i.e. species independent, rates and show that a high mobility can lead
in one dimension to the emergence of species coexistence. In Section IV we allow for asymmetric, i.e.
species dependent, interaction and swapping rates and show that this asymmetry leads to an interesting
dynamical phase diagram not observed in the nonspatial case. Finally, we conclude in Section V.
\section{One-dimensional rock-paper-scissors games}
In our spatial rock-paper-scissors games we consider three species living on a one-dimensional support that
are competing in a cyclic way. Calling $A$, $B$, and $C$ the three species, these interactions
can be cast in the following reaction scheme:
\begin{eqnarray}
A+B,~ B+A & \overset{k_{ab}}{\longrightarrow} & A+A \\
B+C,~ C+B & \overset{k_{bc}}{\longrightarrow} & B+B\\
C+A,~ A+C & \overset{k_{ca}}{\longrightarrow} & C+C
\end{eqnarray}
where we allow for species dependent interaction rates $k_{ab}$, $k_{bc}$, $k_{ca}$. In the following we
restrict ourselves to the case where every site is occupied by exactly one individual, such that interactions
only take place between nearest neighbors. In this setting mobility is realized by swapping the positions of
individuals on neighboring sites \cite{Rei08,Pel08}
\begin{eqnarray}
A+B & \overset{s_{ab}}{\rightleftarrows} & B+A \\
B+C & \overset{s_{bc}}{\rightleftarrows} & C+B\\
C+A & \overset{s_{ca}}{\rightleftarrows} & A+C
\end{eqnarray}
where the swapping rates $s_{ab}$, $s_{bc}$, $s_{ca}$ can again be species dependent.
Note that mobility can also be realized through simple diffusion steps if empty sites are allowed and/or
more than one individual can occupy a given lattice site \cite{Fra96b,Pel08,He09}.
In our numerical simulations, every lattice site is initially occupied by any of the three species with the
same probability 1/3. The total number of individuals,
$N_A + N_B +N_C = N$, is a conserved quantity. Here $N_X$ is the number of individuals of species $X$ in the system and $N$ is the
number of lattice sites. We consider periodic boundary conditions and use sequential dynamics. Having selected a neighboring pair
of individuals belonging to species $A$ and $B$, we allow for the following events: $A$ dominates $B$
with probability $k_{ab}$, $A$ and $B$ exchange positions with probability $s_{ab}$, or no
interaction takes place with probability $1 - k_{ab} - s_{ab}$. After having selected $N$ such pairs the time
is increased by one unit.
First studies of one-dimensional systems with $M$ competing species have been
published some time ago \cite{Fra96a,Fra96b,Pro99,Fra98}. If the individuals are immobile,
exact results can be obtained \cite{Fra96a,Fra96b}. For example,
it is found that the individuals organize in single-species domains whose average size $\langle \lambda
\rangle \sim t^\alpha$ grows
algebraically with time. The exponent $\alpha$ governing this domain growth depends on the number of competing
species as well as on the chosen dynamics (parallel or sequential dynamics). For the case most relevant to
our study ($M=3$ and sequential updates) the exact result is $\alpha = 3/4$. Interestingly,
for parallel dynamics an asymptotic equivalence was established between the diffusion-reaction
approach and the lattice model with immobile individuals \cite{Fra96b}. Recent simulations of lattice models
with sequential dynamics, where diffusion
is made possible through multiple occupancy of a site and/or the presence of empty sites \cite{He09},
also indicate that inclusion of simple diffusion does not lead to a qualitative different behavior.
However, as we show in the following, this is no longer the case when the mobility of individuals is realized
through exchanges.
\section{Symmetric interaction and swapping rates}
Let us start our discussion of mobility effects in one-dimensional rock-paper-scissors games
by first looking at the symmetric case where all rates are species independent. We therefore set in the
following $k= k_{ab} = k_{bc} = k_{ca}$ and $s= s_{ab} = s_{bc} = s_{ca}$, with $k + s = 1$.
\begin{figure}[t]
\centerline{\epsfxsize=3.80in\ \epsfbox{figure1.eps}}
\caption{Space-time plots of one-dimensional rock-paper-scissors games with symmetric interaction rates $k$
and swapping rates $s$. White, gray, and black correspond to the three different species.
In the left panel, the individuals have a low mobility, with $s = 0.1$ and $k =0.9$,
yielding segregation of species, similar to what is observed for immobile individuals. In the right panel,
high mobility, with $s =0.9$ and $k = 0.1$, continuously mixes the different species, such giving raise to
coexistence even in one dimension. The system size is 2500 and the first 10000 time steps are shown.}
\label{fig1}
\end{figure}
As already mentioned, earlier studies \cite{Fra96a,Fra96b,Pro99,Fra98} of immobile individuals
have highlighted the existence of species segregation through the formation of single-species domains
that grow algebraically in time. If one allows for mobile individuals, the conclusions drawn in \cite{Fra96a,Fra96b}
remain unchanged for not too high swapping rates $s$. As an example we show in the left panel of Fig. \ref{fig1}
the space-time diagram for a system composed of 2500 individuals where the rates are $k=0.9$ and $s = 0.1$. Segregation sets in
immediately, yielding the formation of numerous small single-species domains. These domains then coarsen, and
the system rapidly decomposes into a few large domains. At this stage, swapping has become
irrelevant and the behavior of the system is determined by the movement of the interfaces between different clusters.
The time evolution of the system is then similar to the time evolution of the system composed
of immobile individuals, yielding the same exponent $\alpha = 3/4$ for the average domain size, see Fig. \ref{fig2}. Eventually,
this process leads to the extinction of two of the three species as one species dominates the others in the long-time limit.
\begin{figure}[t]
\centerline{\epsfxsize=3.20in\ \epsfbox{figure2.eps}}
\caption{Average domain size as a function of time for different swapping rates $s$, with $k+s=1$.
The results shown here have
been obtained after averaging over 40000 independent runs for systems composed of 160000 individuals.
The dashed line indicates the exactly known result that for $s =0$ the time-dependent average domain size increases
algebraically with an exponent 3/4 for large times.}
\label{fig2}
\end{figure}
If one keeps increasing the mobility of the individuals, one eventually reaches a critical swapping rate $s_c = 0.84(1)$ above which
the exchange mechanism very effectively mixes the different species. This yields completely different space-time plots, see the
right panel of Fig. \ref{fig1}, with a mosaic of ordered domains that emerge at some time and vanish again later. In addition to
these regular pattern one also observes more chaotic space-time regions where no large domains are formed. This intriguing
spatial-temporal behavior also shows up in the average domain size, see Fig. \ref{fig2}, as $\langle \lambda \rangle$ does no longer
increase algebraically with time for $s > s_c$, but saturates at a finite, $s$ dependent, value.
In this regime the particle
densities of all three species display an irregular oscillatory behavior, see Fig. \ref{fig3}. Obviously, the mixing
induced by the high exchange rates
promotes the coexistence of species, and therefore biodiversity, in our one-dimensional system.
This observation of coexistence in an ecological system of highly mobile individuals agrees with the
intuitive picture that a mixing of species should lead to a mean-field like behavior. This is different to
the more complex situation in two space dimensions where for medium values of the mobility an intermediate
regime exists in which rotating spirals are destabilized, yielding
rapidly a uniform state
where only one species survives. A coexistence regime similar to that observed by us in one space dimension eventually emerges
in two dimensions for very high
mobilities \cite{Rei08,Pel08,Rei08a}.
Remarkably, a comparison of our results with the few existing results with simple diffusion (made possible through the
presence of empty sites and/or sites with multiple occupancy) \cite{Fra96b,He09} indicates that the mechanism through
which mobility is realized matters. At this stage we can only speculate on the origin of these unexpected differences.
We note that in a field-theoretical description our exchange mechanism yields a nonlinear term, in contrast to simple
diffusion which enters linearly \cite{Tau09}. It is, however, beyond the scope of this article to explore this further
using field-theoretical techniques.
\begin{figure}
\centerline{\epsfxsize=2.60in\ \epsfbox{figure3.eps}}
\caption{Time-dependent particle densities for all three species for a typical run with $k=0.1$ and $s=0.9$ and $N=10000$
lattice sites. In this regime, where $s > s_c$, an irregular oscillatory behavior is observed.}
\label{fig3}
\end{figure}
\section{Asymmetric interaction and swapping rates}
There are usually no compelling reasons that require all interaction rates to be the same for a system with competing species.
On the contrary, one should expect to have species dependent rates in any real system, as has also been
observed in a recent study of the self-organization
of Min proteins \cite{Loo08}. Theoretical studies usually neglect this aspect and exclusively focus on species independent
interaction rates. One exception of which we are aware of is the study of small asymmetries in a four-state
rock-paper-scissors game (where the forth state means that a site is unoccupied) in two dimensions \cite{Pel08}.
For that model, it was shown through
first-order perturbation theory that rotating spirals are robust against small asymmetries.
Similar conclusions where drawn recently from a four-state rock-paper-scissors model that was constructed to
describe the symbiosis between an ant-plant and two protecting ant symbionts \cite{Szi09}.
In nonspatial rock-paper-scissors
games an asymmetry in the reaction rates was shown to lead to a 'law of the weakest', as the less efficient predator
always survives \cite{Ber09}.
In principle, we have to explore for our three species model a six-dimensional parameter space. We will in the following
not try to present a comprehensive study of all possible cases. Instead we fix the interaction and swapping rates between (1) $B$ and $C$
and (2) $C$ and $A$ particles and systematically change the rates for the processes involving $A$ and $B$ particles.
Concretely, we set $k_{bc} = k_{ca} = 0.4$ and $s_{bc} = s_{ca} = 0.4$.
In fact, this two-dimensional slice through our six-dimensional parameter space already captures our main result,
namely that the 'law of the weakest' is not a strict one in spatial games.
We first remark that an asymmetry in the rates yields the dominance of a single species both in the spatial
and in the nonspatial game.
Whereas in nonspatial games the weakest predator, i.e. the species $X$ for which the interaction rate
$k_{xy}$ on another species $Y$ is smaller than the other interaction rates, always survives,
this is different in one space dimension. We discuss in Fig. \ref{fig4} the survival probabilities for the fixed
interaction rate $k_{ab} = 0.45 > k_{bc}$, $k_{ca}$. Writing the survival probability of species $A$ in a total population of $N$
individuals as $P_A(N)$, one readily verifies that the nonspatial game yields for this situation
$P_B(N) \longrightarrow 1$ and $P_C(N), P_A(N) \longrightarrow 0$ in the macroscopic limit $N \longrightarrow \infty$,
i.e. the species $B$, which is dominated by $A$ and dominates $C$, survives.
In the one-dimensional case, however, the situation is much more complex, as the surviving species changes as
a function of the value of the exchange rate $s_{ab}$. As shown in Fig. \ref{fig4}a for a system with 5000 individuals,
we can identify three different regimes. Whereas for small exchange rates $s_{ab} < s_1$ with $s_1 = 0.225(5)$
the species $A$ has the highest
survival probability, this changes for larger values as then either species $B$ (for $s_1 < s_{ab} < s_2$ with
$s_2 = 0.32(1)$) or species $C$
(for $s_{ab} > s_2$) prevails.
The transition at $s_{ab} = s_1$ is a smooth one, as none of the survival probabilities change dramatically when
crossing the transition point. This is different for the transition at $s_{ab} = s_2$ which is characterized by a very abrupt
change of the survival probabilities. The latter transition has therefore typical characteristics of a discontinuous
transition, whereas the behavior at the former transition is reminiscent of that at a continuous transition.
We carefully checked the robustness of our results against a change of the number
of individuals, see
Fig. \ref{fig4}b,c.
\begin{figure}
\centerline{\epsfxsize=3.40in\ \epsfbox{figure4.eps}}
\caption{(Color online) (a) Survival probability of the different species as a function of $s_{ab}$ for $k_{ab} = 0.45$.
The total population is 5000 individuals. The data result from averaging over 16000 different runs with
different realizations of the noise. The dashed lines indicate the locations of the two transitions, see main
text. The errors are comparable to the size of the symbols.
(b) Survival probability as a function of time for $k_{ab} = 0.45$ and $s_{ab} = 0.10$. The different curves
correspond to different system sizes $N$, with $N=2500$ (full black lines), $N=5000$ (grey [green] lines),
and $N=10000$ (dashed lines). (c) The same as (b), but now for $k_{ab} = 0.45$ and $s_{ab} = 0.50$.}
\label{fig4}
\end{figure}
Repeating this for other values of $k_{ab}$, we obtain the dynamical phase diagram shown in Fig. \ref{fig5}.
Note that with our definition of the rates only the triangle defined by $k_{ab} + s_{ab} \leq 1$ is accessible. We
can distinguish three different phases which differ by the dominating species: in phase I species A dominates, in phase
II species B dominates, and in phase III species C dominates. These phases are separated by transition lines. The lower
transition line (dashed line) is characterized by the fact that $P_A = P_B > P_C$, whereas along the line separating phases II and III
we have $P_A = P_B = P_C$. As shown in the inset, the latter line does not show any size dependence. This is different for
the line separating the regimes I and II as here finite-size effects are observed for the smaller system sizes.
In fact, these finite-size dependences are compatible with a continuous transition along the line separating
regimes I and II and a discontinuous transition along the line separating regimes II and III.
Fig. \ref{fig5} seems to indicate the existence at $k_{ab} = 0.355(5)$ and $s_{ab}=0.47(1)$
of a triple point where the three phases meet. Based on the continuous and discontinuous character of the
different transition lines, this triple point should in fact be a tricritical point.
However, the existing data do not allow us to fully characterize the nature of this triple point,
due to subtle finite-size effects.
Systems much larger than those studied here might be needed in order to reliable determine the location
and the character of the triple point.
\begin{figure}
\centerline{\epsfxsize=3.20in\ \epsfbox{figure5.eps}}
\caption{(Color online) Dynamical phase diagram for a two-dimensional slice through the six-dimensional parameter space, with
$k_{bc}= k_{ca} = 0.4$ and $s_{bc} = s_{ca} =0.4$. Three different phases characterized by the dominating species
(species $A$ in I, $B$ in II, and $C$ in III) are identified. Along the dashed resp. full line, one has the following
relation between the survival probabilities: $P_A = P_B > P_C$ resp. $P_A = P_B = P_C$. The vertical dotted line indicates
the line along which the data shown in Fig. \ref{fig4}a have been obtained. The inset shows the transition lines
for different system sizes.}
\label{fig5}
\end{figure}
This complex dynamical phase diagram indicates that the law of the weakest observed in a nonspatial game is not strictly
valid any more for spatial systems. Thus, for $k_{ab} < 0.2$ we indeed observe that the most inefficient predator eventually
survives, as expected for the law of the weakest. However, the situation is much more complex for $k_{ab} >0.2$, as in addition
to the regime where the weakest survives one also encounters regimes where the weakest dies out.
\section{Conclusion}
Cyclic dominance in ecological or chemical systems is known to yield a very rich behavior and to lead to such intriguing
features as rotating spirals. Very recently, a range of new insights have been gained through the study of three species
rock-paper-scissors games, and this in a spatial as well as in a nonspatial setting \cite{Rei06,Rei07,Rei08,Cla08,Pel08,Rei08a,Ber09}.
In this manuscript we complement this recent research activity by a numerical study of one-dimensional rock-paper-scissors games
with mobile individuals. We thereby realize the mobility through the exchange of neighboring individuals. In addition,
we allow for species dependent interaction and exchange rates, thereby bringing our model closer to real systems.
Interestingly, our study indicates that both the mobility and the rate asymmetry yield intriguing new features.
In presence of symmetric rates, a high mobility leads to a well-mixed system characterized by the prevailance of
coexistence in one dimension. The effect of the mobility is therefore different from the effect in two dimensions
where moderate mobility dissolves existing space-time pattern and ultimately leads to species extinction, whereas high
mobility leads to the emergence of coexistence \cite{Rei08,Pel08,Rei08a}, similar to what we observe. Another surprising
result is the observation that mobility realized through the exchange of individuals has a much stronger impact than
mobility realized through simple diffusion \cite{Fra96b,He09}.
For asymmetric, i.e. species dependent, rates we observe that the spatial support induces new phenomena not
observed in nonspatial games. Thus the law of the weakest, which states that the weakest predator survives and which is a
strict one in nonspatial games \cite{Ber09}, is replaced by a dynamic phase diagram with different regimes that are characterized
by the surviving species. A behavior in accordance with the law of the weakest can still be observed for certain
ranges of the system parameters, but it is no longer a strict one.
We are not aware of any (quasi) one-dimensional experimental realization of the rock-paper-scissors games. A possible way
for a future realization of such a system could involve nanofabricated landscapes as those used in \cite{Key06} for a study
of bacterial populations. Our study points out the intriguing phenomena showing up in low dimensions and we hope that
these aspects will be addressed in future experiments.
In conclusion, our numerical study highlights the importance of species specific interaction and mobility rates in ecological systems.
Our results raise some new questions (the most intriguing being the observation that the effect of the mobility depends
on how it is realized and the possible existence of tricritical points in the dynamical phase diagram)
and we plan to focus on these issues in the future.
\begin{acknowledgments}
We thank Uwe C. T\"{a}uber and Royce K. P. Zia for interesting and insightful discussions.
We are also grateful to Qian He, Mauro Mobilia, and Uwe C. T\"{a}uber for showing us their results
for systems with cyclic dominance and diffusion prior to publication.
This work was supported in part by the US National
Science Foundation through DMR-0904999.
\end{acknowledgments}
|
\section{Introduction}
Nuclear star clusters (NSCs) are found at the centers of a majority of
galaxies with Milky Way mass and below
\citep{boker02,carollo02,cote06}. These NSCs have similar sizes to
globular clusters ($\sim$5~pc), but are much more luminous and massive
\citep{walcher05}. Their morphology is often complicated and their star
formation history is typically extended \citep[e.g.][]{seth06}. The
relationship of NSCs to supermassive black holes (BHs) typically found
at the centers of more luminous galaxies is not clear. NSCs and
supermassive BHs co-exist in some galaxies \citep{seth08}, and both
have masses that scale with the mass and dispersion of the galaxy they
live in \citep[e.g.][]{ferrarese06}. Unlike BHs, the histories of
NSCs are recorded in their morphology, kinematics, and stellar
populations, making them excellent tools to study how
mass accumulates at the centers of galaxies.
\section{Nearby Nuclear Star Cluster Survey}
We are conducting an ongoing survey of the nearest NSCs, observing a
sample of 13 galaxies within 5~Mpc. By resolving the kinematics and
stellar populations in these NSCs, we can understand how they
formed and constrain the masses of any BHs within them. Their
proximity makes them among the most promising objects for detecting
intermediate mass BHs, constraining the low mass end of the BH-galaxy
scaling relations and the occupation fraction of low mass galaxies.
Because the BHs in these galaxies are likely to not have undergone
much subsequent evolution, measuring their properties is one of the
only ways of constraining the initial formation mechanism of
supermassive BHs
\citep[e.g.][]{volonteri08}.
We are obtaining adaptive optics assisted integral field spectroscopy
on the entire sample using Gemini/NIFS (PI: Seth) and VLT/SINFONI (PI:
Neumayer). In addition, we are obtaining low resolution optical
spectroscopy to examine the stellar populations in the nucleus and HST
imaging to constrain their morphology. Our first published results
\citep{seth08b}, show a flattened multi-component NSC with strong
rotation in the edge-on spiral galaxy NGC~4244. This NSC must have
formed from repeated accretion of material from the disk of the host
galaxy, a scenario tested in simulations by Markus Hartmann (this
volume).
\begin{figure}
\includegraphics[height=3in]{001aseth_fig1.ps}
\caption{MMT 6.5m optical spectrum of the NGC~404 nuclear cluster
(black line) with the best fit spectral synthesis model (overplotted
gray line). The spectral templates used to fit the spectra (shown in
light gray) are drawn from 11 possible age bins ranging from 1~Myr to 13~Gyr
in age. More than 50\% of the light is in stars with ages 1-3~Gyr,
while $\sim$10\% is at ages $\leq$10~Myr. Residuals are about 1\%.
Vertical gray lines indicate areas excluded from the fit due to
emission lines. \label{aseth_specfig}}
\end{figure}
\section{The Nucleus of NGC~404}
We now focus on the NSC formation and possible intermediate mass BH in
NGC~404, the nearest S0 galaxy (D=3.1~Mpc). The galaxy is nearly
face-on and has a total mass of about $10^9$~M$_\odot$ and $M_V =
-17.35$. NGC~404 also hosts the nearest LINER nucleus, although
whether this observed activity is actually the result of an accreting
BH is a matter of some contention. The results shown here are
presented in much greater detail in a paper submitted to the {\em
ApJ}.
\vspace{0.03in}
\noindent {\bf Nuclear Star Cluster}\\
The surface brightness profile of the inner part of NGC~404 (see left
panel of Fig.~\ref{aseth_nscfig}) can be decomposed into three
distinct components each with distinct kinematics and stellar
populations. The bulge of the galaxy has a mass of $9 \times
10^8$~M$_\odot$ and dominates the light profile outside the central
arcsecond (15~pc). Optical spectra of the stellar populations of this
component show that it is dominated by old stars with ages $>$5~Gyr.
The nuclear star cluster dominates the light in the central arcsecond
and has a half-light radius of 10~pc and a dynamical mass of
$10^7$~M$_\odot$, larger than expected from NSC scaling relations. Its
optical spectrum is dominated by the light of stars with age of
1-3~Gyr, with stellar population synthesis fits suggesting that over
half the mass of the cluster was formed during this period. The
cluster shows mild rotation in the same direction as HI gas at larger
radii in the galaxy \citep{delrio04}, which has been suggested to have
a merger origin. Finally, within the central 0.2'', the central
excess of emission seen in the surface-brigthtness profile, is also
found to be counter-rotating relative to the NSC. The emission of
this component probably results from a combination of young
($<$100~Myr) stars, hot dust emission and perhaps AGN continuum.
The nuclear star cluster in NGC~404 appears to have a quite different
formation from the episodic disk accretion seen in NGC~4244. The
burst of star formation $\sim$1~Gyr ago is suggestive of a merger
origin, a scenario that is also consistent with the difference in
rotation between the central excess and nuclear star cluster.
\begin{figure}
\includegraphics[height=2.5in]{001aseth_fig2.ps}
\includegraphics[height=2.5in]{001aseth_fig3.ps}
\caption{{\em Left --} The surface brightness profile of NGC~404
in $F814W$/$I$ band made from HST imaging. The lines show the
best bulge$+$NSC fit, where both components are assumed fit as
S\'ersic profiles. The central 0.2'' show a clear excess
above these profiles. {\em Right} -- Results from the gas dynamical
model of a disk of molecular H$_2$ gas emission. Shown are
$\chi^2$ contours as a function of BH mass and disk inclination.
The best fitting BH mass is $4.5^{+3.5}_{-2.0} \times
10^5$~M$_\odot$, however the stellar dynamical models suggest a BH
mass upper limit of $1 \times 10^5$~M$_\odot$.
\label{aseth_nscfig}}
\end{figure}
\vspace{0.03in}
\noindent {\bf Possible Intermediate Mass BH}\\
Signs of BH accretion in NGC~404 are conflicting, with the
strongest evidence coming from the presence of variable UV emission
\citep{maoz05}. We also find compact and possibly variable hot dust
emission indicative of BH accretion at the center of the NSC.
From the Gemini/NIFS data we have obtained kinematic measurements of
both the stars (as measured by the CO-bandhead absorption at
2.3$\mu$m) and for a gas disk seen in emission of excited H$_2$ at
2.12$\mu$m. The NIFS adaptive optics observations have a PSF core
with a FWHM=0.09'' containing 50\% of the light. We have created
dynamical models to fit to the kinematics of both tracers. The
starting point of both models is a mass model for the galaxy generated
from HST imaging data using the multi-gaussian expansion method
\citep{emsellem94}. The mass model assumes a constant mass-to-light
($M/L$) ratio.
The stellar dynamical model uses axisymmetric Jeans modeling with the
JAM package \citep[][see also contribution in this
volume]{cappellari08} to fit the CO bandhead velocity and dispersion
in the central 1.5'' (22~pc) of the galaxy. The model fits the NSC
$M/L$, the anisotropy ($\beta_z$), and the BH mass ($M_{\rm BH}$).
Assuming a constant $M/L$, we find a $M/L=0.70$ in the $I$ band and a
3-$\sigma$ upper limit of $M_{\rm BH} < 1 \times 10^5$~M$_\odot$. The
$M/L$ determined from the dynamical model matches the best-fit stellar
population synthesis $M/L$ to within 10\%.
The gas dynamical model is similar to that presented in
\citet{neumayer07}, and assumes a thin disk geometry. It uses the $M/L$ from
the stellar dynamical model and finds the disk inclination and BH mass
that best match the H$_2$ kinematics. The results of this fit are
shown in the right panel of Fig.~\ref{aseth_nscfig}. The best-fit
$M_{\rm BH} = 4.5^{+3.5}_{-2.0}
\times 10^5$~M$_\odot$ (3$\sigma$ errors), somewhat inconsistent with
the BH mass upper
limit from the stellar kinematics. However both BH mass estimates are
comparable to the uncertainty in the central stellar mass of $< 3
\times 10^5$~M$_\odot$ due to possible $M/L$ variations within the
cluster. Despite this uncertainty, the dynamical models constrain the
mass of the possible BH to be smaller than any previously dynamically
measured BH and below several bulge mass scaling relations. We hope
to obtain HST spectroscopy and imaging of the stellar populations
across the nucleus to better refine the mass model and robustly
measure the BH mass in NGC~404.
\bibliographystyle{aipproc}
|
\section{Introduction}
Colors are used to represent and differentiate various chemical, physical, or geometric aspects of a symmetrical structure. For instance, different colors may correspond to different types of atoms or to different orientations of a magnetic moment in a crystal. These applications paved the way to the development of the theory of color symmetry. The theory was firmly established with Shubnikov's work on antisymmetry in the 1950's \cite{Sc, S}. In addition, scientists then were intrigued by the colored tilings of the Dutch graphic artist Escher and began to analyze them mathematically \cite{Mc}. A comprehensive discussion of the history and applications of the theory of color symmetry by Schwarzenberger can be seen in \cite{Sc}. Color symmetry remains to be of interest until today because of its relation with crystallography.
In color symmetry, we not only look at the symmetrical pattern, but also the various ways of symmetrically coloring the pattern. Given a colored symmetrical pattern, three groups are associated to it: the group $G$ of symmetries sending the uncolored pattern to itself, the subgroup $H$ of $G$ consisting of symmetries that induce a permutation of colors in the pattern (called the \emph{color group} associated to the coloring), and the subgroup $K$ of $H$ whose elements fix the colors (called the \emph{symmetry group of the colored pattern}). Since $H$ acts on the set $C$ of colors of the pattern, there exists a homomorphism $f:H\rightarrow P(C)$ where $P(C)$ is the group of permutations of $C$. The kernel of $f$ is $K$. Hence, the group of color permutations $f(H)$ is isomorphic to $H/K$ and this is usually referred to as the \emph{color permutation group} corresponding to the coloring.
Senechal outlined the development of color symmetry as applied to symmetrical patterns and posed open problems in \cite{S}. In the same paper, she pointed out that classifying color groups and classifying colored patterns are not the same since different colored patterns may correspond to the same color group. Roth also distinguished between the notions of equivalence for colored patterns and equivalence for color groups, and gave several illustrations in \cite{Ro1} and \cite{Ro2}.
A colored symmetrical pattern is said to be perfectly colored if $H=G$. This is the most restrictive case since all symmetries of the uncolored pattern effect color permutations. These colorings were first studied by Burckhardt and Van der Waerden in \cite{BVW}. Perfect colorings of different types of patterns have been studied in detail, such as transitive tilings and patterns in the plane \cite{GS} and
hyperbolic tessalations \cite{DLPFL}.
On the other hand, nonperfect colorings have also appeared in some early works in color symmetry, such as colored patterns in \cite{WG}. In \cite{Ro1}, Roth discussed nonperfect transitive colorings (referred to as ``partially symmetric colorings'') of certain patterns and their equivalence. It was suggested by Senechal in \cite{S} that a systematic study of nonperfect colorings might become useful and interesting.
Rigby encountered nonperfect colorings while studying precise colorings of the regular triangular tiling $\{3,n\}$ \cite{R}. He coined the term ``semiperfect coloring'' to describe colorings wherein half of the direct (orientation preserving) and half of the opposite (orientation reversing) symmetries of the uncolored pattern permute the colors. Rigby also used the term ``chirally perfect'' for colorings where all the direct symmetries of the uncolored pattern effect color permutations but none of the opposite symmetries do. In both cases, the corresponding color group $H$ is of index 2 inside the group $G$ of symmetries of the uncolored regular triangular tiling $\{3,n\}$. In this paper, we will look at how to obtain colorings where $[G:H]=2$.
\section{Preliminaries}
Let $G$ be a subgroup of an $n$-dimensional crystallographic group and $X$ be the set of objects in a given pattern to be colored. Suppose
$G$ acts transitively on $X$ such that for all $x\in X$, the stabilizer of $x$ in $G$ is $\{e\}$. If this is the case, then the $G$-orbit of an $x\in X$ is $Gx=\{gx:g\in G\}=X$ and we obtain a one-to-one correspondence between $G$ and $X$ given by $g\leftrightarrow gx$. Hence, we can associate a partition $P=\{P_1,\ldots,P_r\}$ of $G$ with the partition $\{P_1x,\ldots,P_rx\}$ of $X$. Given a set $C=\{c_1,\ldots,c_r\}$ of $r$ colors, we call the assignment of each color $c_i$ to $P_ix$ a \emph{coloring} of $X$ corresponding to the partition $P$. Thus, a coloring of the pattern is treated as a partition $P$ of $G$ where each element of $P$ corresponds to a unique color.
The group $G$ acts on the set of all partitions of $G$ by left multiplication. Denote by $H$ the stabilizer of a partition $P$ of $G$. Hence,
$h\in H\Leftrightarrow hP=P$ and we say that $P$ is an \emph{$H$-invariant partition of $G$}. Geometrically, an element $h\in H$ is said to \emph{permute} the colors in a coloring of $X$ ($h$ is also called a \emph{color symmetry}). This means that all objects in $X$ of a given color is mapped by $h$ onto objects in $X$ of a single color. That is, we can associate $h$ to a permutation of the set of colors. When $[G:H]=1$ or $H=G$, a coloring associated with $P$ is called \emph{perfect}. If $[G:H]=2$, we call a coloring associated with $P$ \emph{semiperfect}.
To obtain semiperfect colorings of patterns, we use the framework by De Las Pe\~nas, Felix, and Quilinguin in \cite{DLPFQ1} and \cite{DLPFQ2}. Let $H\leq G$ and $Y$ a complete set of right coset representatives of $H$ in $G$. A $(Y_i,J_i)-H$ \emph{partition of} $G$ is the partition $\{hJ_iY_i:i\in I, h\in H\}$ of $G$ where $Y=\cup_{i\in I}{Y_i}$ and $J_i\leq H$ $\forall i\in I$. If $P$ is a $(Y_i,J_i)-H$ partition of $G$, then $hP=P$ $\forall h\in H$. That is, all the elements of $H$ will permute the colors in a coloring associated with a $(Y_i,J_i)-H$ partition of $G$.
Using this framework, we may obtain semiperfect colorings of symmetrical patterns by observing the following procedure:
\begin{enumerate}
\item Choose a subgroup $H$ of index 2 in $G$.
\item Choose a complete set $Y$ of right coset representatives of $H$ in $G$.
\item Partition $Y$ either as $Y=Y_1$ or as $Y=Y_1\cup Y_2$.
\item \begin{itemize}
\item If $Y=Y_1$, choose $J_1\leq H$ and form the partition $\{hJ_1Y_1:h\in H\}$. Such partitions will be referred to as
\emph{Type I} partitions and they give rise to colorings having only one orbit of colors.
\item If $Y=Y_1\cup Y_2$, choose $J_1$, $J_2\leq H$ and form the partition $\{hJ_1Y_1:h\in H\}\cup\{hJ_2Y_2:h\in H\}$. Such
partitions will be referred to as \emph{Type II} partitions and they give rise to colorings having at most two orbits of colors.
\end{itemize}
\end{enumerate}
To illustrate, consider the uncolored hexagonal pattern in Figure \ref{fig1}(a). Its symmetry group is $G=\langle a,b\rangle\cong D_6$ where $a$ is the $60^{\circ}$-counterclockwise rotation about the center of the hexagon and $b$ is the reflection along the horizontal line through the center of the hexagon. Note that the pattern may be obtained as the $G$-orbit of the tile labeled ``$e$'' and we obtain an assignment of each element of $G$ to a unique tile in the pattern as shown in Figure \ref{fig1}(a).
Let $H=\langle a^2,b\rangle$, a subgroup of index 2 in $G$, and $Y=\{e,a^3\}$, a complete set of right coset representatives of $H$ in $G$. Write $Y=Y_1\cup Y_2$ where $Y_1=\{e\}$ and $Y_2=\{a^3\}$ and choose the subgroups $J_1=\langle a^2b\rangle$ and $J_2=H$ of $H$. Form the Type II partition $\{hJ_1Y_1:h\in H\}\cup\{hJ_2Y_2:h\in H\}=\{\{e,a^2b\},\{b,a^4\},\{a^4b,a^2\},\{ab, a, a^3b, a^3, a^5b, a^5\}\}$. Upon assigning the color yellow to $\{e,a^2b\}$, the color green to $\{b,a^4\}$, the color blue to $\{a^4b,a^2\}$, and the color red to $\{ab, a, a^3b, a^3, a^5b,a^5\}$, the coloring in Figure \ref{fig1}(b) is obtained.
In Figure \ref{fig1}(b), we see that the reflection $b$ fixes the colors red and blue (that is, all red tiles are mapped by $b$ to red tiles, and the same is true for blue tiles) and interchanges the colors green and yellow (that is, $b$ sends green tiles to yellow tiles and vice-versa). Thus, $b$ permutes the colors in Figure \ref{fig1}(b). On the other hand, the rotation $a$ does not permute the colors. Indeed, some red tiles are mapped by $a$ to yellow tiles while others are mapped to blue (and green) tiles. Looking at the effect of the other symmetries in $G$ on the colors, we conclude that the coloring in Figure \ref{fig1}(b) is semiperfect since only the elements of $H$ permute the colors. Also, the coloring has two orbits of colors, namely, $\{\text{blue},\text{green},\text{yellow}\}$ and $\{\text{red}\}$.
\noindent\begin{figurehere}
\centering{\includegraphics[width=40mm]{Figure1a.eps}}
\centering{(a)}
\begin{minipage}[c]{40mm}
\centering{\includegraphics[width=35mm]{Figure1b.eps}}
\centering{(b)}
\end{minipage}
\begin{minipage}[c]{40mm}
\centering{\includegraphics[width=35mm]{Figure1c.eps}}
\centering{(c)}
\end{minipage}
\caption{Semiperfect (b) and perfect (c) colorings of the hexagonal pattern in (a) }
\label{fig1}
\end{figurehere}
\medskip If $H^*$ is the stabilizer of a $(Y_i,J_i)-H$ partition of $G$ then $H\leq H^*$. Since we chose $H$ such that $[G:H]=2$, the coloring corresponding to the $(Y_i,J_i)-H$ partition of $G$ will be either semiperfect or perfect. For instance, if we partition $Y$ in the previous example as $Y=\{e,a^3\}$ and choose $J=\langle b\rangle$, we generate the perfect coloring in Figure \ref{fig1}(c). In this paper, we give methods on how to determine whether a $(Y_i,J_i)-H$ partition of $G$ where $[G:H]=2$ corresponds to a perfect or semiperfect coloring without doing the actual coloring.
\section{Equivalence of colorings}
In enumerating semiperfect colorings, we will only consider those that are inequivalent to each other. Two colorings of the same symmetrical pattern
are said to be \emph{equivalent} \cite{Ro1} if one of the colored patterns may be obtained from the other colored pattern by
\begin{enumerate}
\item a bijection from $C_1$ to $C_2$ where $C_i$ is the set of colors in the $i$th colored pattern for $i=1,2$, or
\item a symmetry of the uncolored pattern, or
\item a combination of $(1)$ and $(2)$.
\end{enumerate}
When we apply a symmetry of an uncolored pattern to a coloring of the pattern, we obtain a reassignment of the colors to different objects in the pattern. Hence, only a relabelling of the colors is necessary to show that two perfect colorings are equivalent \cite{Ro1}. However, given two
nonperfect colorings of the same pattern, it is usually not easy to determine whether they are equivalent. The following are some suggestions on
how to find out if the two colorings are equivalent or not.
\begin{enumerate}
\item Check that the number of colors used in both colorings are the same. If this is not the case, then they must be inequivalent.
\item Count the number of color orbits formed in both colorings. Note that the patterns formed by the colors belonging to one orbit of color
are necessarily congruent. If the number of color orbits are not equal, then the colorings are inequivalent.
\item If the number of colors and the number of orbits of colors are the same, then using the same set of colors for both colorings
facilitates in distinguishing them.
\item Choose a color in one of the colorings and identify the pattern formed by that color. If the same pattern does not appear for some
color in the other coloring then the two colorings are inequivalent.
\end{enumerate}
For instance, the colorings in Figures \ref{fig1}(b) and \ref{fig1}(c) are clearly inequivalent because Figure \ref{fig1}(b) consists of four colors and two color orbits while Figure \ref{fig1}(c) consists of only three colors and one color orbit.
Consider the uncolored pattern in Figure \ref{fig2} which is assumed to repeat over the entire plane. Its symmetry group is the crystallographic group $G=\langle a,b,x,y\rangle$ of type $p4m$ where $a$ is the $90^{\circ}$-counterclockwise rotation about the indicated point $P$, $b$ is the reflection along the horizontal line passing through $P$, and $x$ and $y$ are translations as indicated.
\noindent\begin{figurehere}
\centering{\includegraphics[width=70mm]{Figure2.eps}}
\caption{Uncolored pattern with symmetry group of type $p4m$}
\label{fig2}
\end{figurehere}
\noindent\begin{center}\begin{figurehere}
\begin{minipage}[c]{50mm}
\includegraphics[width=50mm]{Figure3a.eps}
\centering{(a)}
\end{minipage}\quad
\begin{minipage}[c]{50mm}
\includegraphics[width=50mm]{Figure3b.eps}
\centering{(b)}
\end{minipage}
\caption{Two equivalent colorings of the uncolored pattern in Figure \ref{fig2}}
\label{fig3}
\end{figurehere}\end{center}
\medskip The colorings of the pattern in Figure \ref{fig3} are equivalent. Indeed, suppose we replace the color black by red and the color white by gray in Figure \ref{fig3}(a). We now see that the coloring in Figure \ref{fig3}(b) is just the image of Figure \ref{fig3}(a) (now colored also with red and gray) under the reflection $b$.
We now define equivalence of colorings via partitions of the symmetry group $G$ of the uncolored pattern. If $P$ and $Q$ are partitions of $G$ and there is a $g\in G$ such that $Q=gP$, then we say that $P$ and $Q$, as well as their associated colorings, are \emph{equivalent}.
To illustrate, the partitions of $G=\langle a,b,x,y\rangle$ of type $p4m$ corresponding to the colorings in Figure \ref{fig3}(a) and (b) are $P=\{h\langle ab,x,y\rangle\{e,a\}:h\in\langle ab,a^3b,x,y\rangle\}$ and $Q=\{h\langle a^3b,x,y\rangle\{e,b\}:h\in\langle ab,a^3b,x,y\rangle\}$, respectively. We have $Q=bP$ and hence, the partitions $P$ and $Q$, and the two colorings, are equivalent.
We will look at results on how one should select the subgroup $H$ of index 2 in $G$, the complete set of right coset representatives $Y$ of $H$ in $G$, the partition of $Y$, and the subgroups $J$, or $J_1$ and $J_2$ of $H$, in order to avoid obtaining colorings equivalent to those already generated. We start with the following theorem which is a special case of results in \cite{Q} and \cite{Ro1}.
\begin{theorem}\label{thm1}
Let $G$ be a group and $P$ an $H$-invariant partition of $G$. If $[G:H]=2$ then there are only two partitions of $G$ that are equivalent to
$P$, namely, $P$ and $yP$, for some $y\in G\setminus H$. Moreover, the stabilizer of $yP$ in $G$ is also $H$.
\end{theorem}
\begin{proof}
Let $E=\{gP:g\in G\}$ be the set of partitions of $G$ that are equivalent to $P$. Then $E$ is the orbit of $P$ under the action of $G$ on
the set of partitions of $G$ by left multiplication. By the Orbit-Stabilizer Theorem, $|E|=[G:H]=2$ and $E=\{P,yP\}$, for some $y\in
G\setminus H$. Now, $g(yP)=yP\Leftrightarrow (y^{-1}gy)P=P\Leftrightarrow g\in yHy^{-1}=H$ since $H\unlhd G$.
\end{proof}
\section{Type II partitions}
We first consider Type II partitions of $G$ for a fixed subgroup $H$ of index 2 in $G$. We want to enumerate all inequivalent partitions of $G$ of the form $P=\{hJ_1\{x\}:h\in H\}\cup\{hJ_2\{y\}:h\in H\}$ where $J_1$, $J_2\leq H$, $x\in H$, and $y\in G\setminus H$. However, we can write
$P$ as $P=\{hJ_1':h\in H\}\cup y\{hJ_2':h\in H\}$ where $J_1'={J_1}^{x^{-1}}=x^{-1}J_1x\leq H$, $J_2'={J_2}^{\,y^{-1}}\leq H$. Thus, we can look at instead partitions of $G$ of the form $\{hJ_1:h\in H\}\cup y\{hJ_2:h\in H\}$, where $J_1$, $J_2\leq H$, and $y\in G\setminus H$.
\begin{theorem}\label{thm2}
Let $H\leq G$ with $[G:H]=2$, $J_1$, $J_2\leq H$, $y\in G\setminus H$, and $P=P_1\cup yP_2$ where $P_i=\{hJ_i:h\in H\}$ for $i=1,2$.
\begin{itemize}
\item For all $y'\in G\setminus H$, $P_1\cup y'P_2=P_1\cup yP_2=P$.
\item The two distinct partitions of $G$ equivalent to $P$ are $P$ and $P_2\cup yP_1=\{hJ_2:h\in H\}\cup y\{hJ_1:h\in H\}$.
\item A coloring induced by $P$ is perfect if and only if $J_1=J_2$.
\end{itemize}
\end{theorem}
\begin{proof}
\begin{itemize}
\item[]
\item Since $y'\in G\setminus H=yH$, $y'P_2=yP_2$.
\item By Theorem \ref{thm1}, $P$ is only equivalent to $P$ and $yP=y(P_1\cup yP_2)=P_2\cup yP_1$.
\item The coloring corresponding to $P$ is perfect $\Leftrightarrow$ $gP=P$ $\forall g\in G$. Since $hP=P$ $\forall h\in H$, and
$P_1$ and $P_2$ are partitions of $H$ into left cosets, we conclude that the coloring is perfect $\Leftrightarrow yP=P
\Leftrightarrow P_2\cup yP_1=P_1\cup yP_2\Leftrightarrow P_1=P_2\Leftrightarrow J_1=J_2$. \qedhere
\end{itemize}
\end{proof}
Theorem \ref{thm2} suggests that in order to generate all inequivalent Type II $H$-invariant partitions of $G$ that correspond to semiperfect colorings;
\begin{enumerate}
\item Choose any $y\in G\setminus H$.
\item Take all possible 2-combinations $\{J_1,J_2\}$ of subgroups of $H$.
\item Form the different partitions $\{hJ_1:h\in H\}\cup y\{hJ_2:h\in H\}$ of $G$.
\end{enumerate}
Therefore, if $H$ has $n$ subgroups then there are $\binom{n}{2}$ inequivalent $H$-invariant partitions of $G$ of Type II.
Consider the subgroup $H=\langle a^2,b\rangle$ of index 2 in the symmetry group $G=\langle a,b\rangle$ of the uncolored hexagonal pattern in Figure
\ref{fig1}(a). Since $H$ has 6 subgroups, there are $\binom{6}{2}=15$ inequivalent semiperfect colorings of the hexagonal pattern with $H=\langle a^2,b\rangle$ as its associated color group and two orbits of colors.
Remember that the uncolored repeating pattern in Figure \ref{fig2} has symmetry group $G$ of type $p4m$. Choose $H$ to be a subgroup of index 2 in $G$ that is also of type $p4m$. The group $H$ has 7 subgroups of index 2 and no subgroups of index 3. Hence, there are $\binom{8}{2}+0=28$ inequivalent Type II $H$-invariant partitions of $G$ that correspond to colorings having at most four colors.
We remark here that Theorem \ref{thm2}(c) is equivalent to the result in \cite{DLPF} (see also \cite{DLPP}) that the Type II partition of $G$ of the form $\{hJ_1:h\in H\}\cup\{hJ_2\{y\}:h\in H\}$ corresponds to a perfect coloring if and only if $J_2={J_1}^{y}$. For example, recall that the coloring in Figure \ref{fig1}(b) was obtained from a Type II partition of $G=\langle a,b\rangle$ for which $J_1=\langle a^2b\rangle$, $J_2=H=\langle a^2,b\rangle$, $Y_1=\{e\}$, and $Y_2=\{a^3\}$. Clearly, $J_2\neq {J_1}^{a^3}$ and this is why the coloring in Figure \ref{fig1}(b) is semiperfect.
\section{Type I partitions}
We now turn our attention to partitions of $G$ of Type I.
\begin{lemma}\label{lemma3}
Let $H\leq G$ with $[G:H]=2$. Let $J\leq H$, $g\in G$, $Y$ a complete set of right coset representatives of $H$ in $G$, and
$g^{-1}Y=\{x,y\}$ where $x\in H$, $y\in G\setminus H$. The following statements hold:
\begin{itemize}
\item $\{hJ^gY:h\in H\}=g\{hJ(g^{-1}Y):h\in H\}$
\item $\{hJ\{x,y\}:h\in H\}=x\{hJ^{x^{-1}}\{e,x^{-1}y\}:h\in H\}$
\item If $J'=J^{x^{-1}}$ and $y'\in J'x^{-1}y$, then
\item[] $\{hJ'\{e,x^{-1}y\}:h\in H\}=\{hJ'\{e,y'\}:h\in H\}$.
\item $\{hJ'\{e,y'\}:h\in H\}=\{h(J'\cup J'y'):h\in H\}$
\end{itemize}
\end{lemma}
For a given subgroup $H$ of index 2 in $G$, we want to determine all the inequivalent Type I partitions of $G$ that could give rise
to semiperfect colorings. It is enough to consider a representative $J$ from each conjugacy class of subgroups of $H$ in $G$ by Lemma \ref{lemma3}(a).
From Lemma \ref{lemma3}(b), the complete set of right coset representatives of $H$ in $G$ may be taken as $\{e,y\}$ as long as we consider the conjugates of each $J$ by elements of $H$. This, together with Lemma \ref{lemma3}(d), implies that we only need to look at Type $I$ partitions of $G$ of the form $P=\{h(J\cup Jy):h\in H\}$. Note that $P$ is determined by the set $J\cup Jy$ where $J\leq H$ and $y\in G\setminus H$. In addition, Lemma \ref{lemma3}(c) indicates that when choosing $y$, we only need to take representatives from each right coset of $J$ in $G$ that is contained in $G\setminus H$.
Assume that a representative subgroup $J$ from each conjugacy class of subgroups of $H$ in $G$ has been chosen. From Lemma \ref{lemma3}, we only need to look at all partitions of the form $P^l(r):=\{h(J^l\cup J^lr):h\in H\}$ where $l$ runs over a complete set of left coset representatives of $N_H(J)$ in $H$ and $r$ runs over a complete set of right coset representatives of $J^l$ in $G$ that are not in $H$. Note that there are $[H:N_H(J)]\cdot[H:J]$ such partitions. However, some of these partitions may still be equivalent to each other. We address this problem in the next theorem.
\begin{theorem}\label{thm4}
Suppose $J\leq H\leq G$ with $[G:H]=2$. Let $L$ be a complete set of left coset representatives of $N_H(J)$ in $H$ and for each $l\in L$,
let $R(l)$ be a complete set of right coset representatives of $J^l$ in $G$ that are not in $H$. If $N_G(J)=N_H(J)$ then the semiperfect
colorings associated with the partitions $P^l(r)=\{h(J^l\cup J^lr):h\in H\}$ of $G$ for each $l\in L$, $r\in R(l)$ are inequivalent to each
other. Otherwise, a pairing of equivalent semiperfect colorings is obtained.
\end{theorem}
\begin{proof}
By Theorem \ref{thm1}, if $P^l(r)$ corresponds to a semiperfect coloring then a partition $P^x(y)$ is equivalent to $P^l(r)$ if and only if
$P^x(y)=P^l(r)$ or $P^x(y)=rP^l(r)$. We have $P^x(y)=P^l(r) \Leftrightarrow x=l$ and $y=r$. Now,
\begin{alignat*}{2}
rP^l(r)&=r^{-1}\{h(J^l\cup J^lr):h\in H\}\\
&=\{h(r^{-1}J^l\cup r^{-1}J^lr):h\in H\}\\
&=\{h(J^{r^{-1}l}\cup J^{r^{-1}l}r^{-1}):h\in H\}.
\end{alignat*}
Thus, $P^x(y)=rP^l(r)\Leftrightarrow J^x\cup J^xy=J^{r^{-1}l}\cup J^{r^{-1}l}r^{-1}\Leftrightarrow J^x=J^{r^{-1}l}$ and $y=r^{-1}$. Note
that $J^x=J^{r^{-1}l} \Leftrightarrow J=J^{x^{-1}r^{-1}l}$ and this cannot happen if $N_G(J)=N_H(J)$ because $x^{-1}r^{-1}l\in G\setminus H$.
Hence, none of the partitions $P^l(r)$ are equivalent to each other when $N_G(J)=N_H(J)$. On the other hand, suppose $N_G(J)\neq N_H(J)$,
that is, there exists $g\in G\setminus H$ with $J^g=J$. We have $r^{-1}lg\in H$ and $J^{r^{-1}lg}={(J^g)}^{r^{-1}l}=J^{r^{-1}l}$.
Therefore, $P^l(r)$ is equivalent to $P^x(y)$ where $x=r^{-1}lg$ and $y=r^{-1}$ and when we consider all the colorings corresponding to the
partitions of the form $P^l(r)$, we get a pairing of equivalent semiperfect colorings.
\end{proof}
Note that if $R=\{r_i\}$ is a complete set of right coset representatives of $J$ in $G$ that are not in $H$ then for all $l\in L$, $R^l=\{{r_i}^l\}$ is a complete set of right coset representatives of $J^l$ in $G$ that are not in $H$. Hence, for the rest of this section, we will simply take $R(l)$ to be $R^l$ for each $l\in L$.
Recall that the colorings corresponding to partitions of $G$ of the form $P^l(r^l)$ are either perfect or semiperfect. The next theorem says when each of these situations occur.
\begin{theorem}\label{thm5}
Let $G$ be a group, $H$ a subgroup of $G$ of index 2, $J\leq H$, $l\in H$, and $r\in G\setminus H$.
\begin{itemize}
\item The coloring associated with the partition $P=\{h(J\cup Jr):h\in H\}$ of $G$ is perfect if and only if $rJ=Jr$ (or $r\in
N_G(J)$) and $r^2\in J$.
\item The coloring corresponding to the partition $P=\{h(J\cup Jr):h\in H\}$ of $G$ is perfect if and only if the coloring
corresponding to the partition $P^l(r^l)=\{h(J^l\cup J^lr^l):h\in H\}$ of $G$ is perfect.
\end{itemize}
\end{theorem}
\begin{proof}
\begin{itemize}
\item[]
\item Since $hP=P$ $\forall h\in H$, the coloring corresponding to $P$ is perfect if and only if $rP=P$ in which case $rJ\cup rJr=
J\cup Jr$ or $rJ=Jr$ (both being subsets of $G\setminus H$) and $rJr=J$ (both being subsets of $H$). Now, $rJ=Jr$ and
$rJr=J\Leftrightarrow rJ=Jr$ and $r^2J=J$ or $r^2\in J$.
\item This follows from (a) since $rJ=Jr\Leftrightarrow r^lJ^l=J^lr^l$ and $r^2\in J\Leftrightarrow {(r^l)}^2\in J^l$.\qedhere
\end{itemize}
\end{proof}
Theorem \ref{thm5}(a) implies that the number of partitions $P=\{h(J\cup Jr):h\in H\}$ of $G$, for a fixed subgroup $J$ of $H$ and where $r$ runs over a complete set of right coset representatives of $J$ in $G$ that are not in $H$, that correspond to perfect colorings is \[p(J):=(\text{the number of involutions in }N_G(J)/J)-(\text{the number of involutions in }N_H(J)/J).\] In addition, we have $p(J)=p(J^l)$ $\forall l\in H$ by Theorem \ref{thm5}(b). Therefore, for each representative subgroup $J$ from each conjugacy class of subgroups of $H$ in $G$, the number of inequivalent partitions $P^l(r^l)$ of $G$ that give rise to semiperfect colorings is $[H:N_H(J)]\cdot [H:J]$ if $N_G(J)=N_H(J)$ and $\frac{1}{2}[H:N_H(J)]([H:J]-p(J))$ otherwise.
To illustrate, recall that the uncolored hexagonal pattern in Figure \ref{fig1}(a) has symmetry group $G=\langle a,b\rangle$. Choose the subgroup $H=\langle a^2,b\rangle$ of index 2 in $G$. A complete set of representative subgroups $J$ from each conjugacy class of subgroups of $H$ in $G$ is $\{H,\langle a^2\rangle, \langle b\rangle, \{e\}\}$. We need to consider the different partitions $P^l(r^l)$ for each representative $J$ and determine if the associated coloring for each partition will be perfect or semiperfect. The computations and corresponding results are summarized in Table \ref{Table1}.
In particular, consider the case when we choose $J=\langle b\rangle$, $l=e$, and $r=a^3$. The coloring in Figure \ref{fig1}(c) corresponds to the partition $\{h(J\cup Jr):h\in H\}$ of $G$ and is perfect because $a^3J=Ja^3$ and ${(a^3)}^2=e\in J$. Observe also that we obtain a pairing of equivalent semiperfect colorings when $J=\langle b\rangle$ since $\langle a^3,b\rangle=N_G(J)\neq N_H(J)=\langle b\rangle$.
\begin{tablehere}
\begin{center}
\begin{tabular}{|c|c|c||c|}
\hline
$J$ & $l$ & $r^l$ & Resulting Coloring\\\hline\hline
$H$ & $e$ & $a$ & perfect\\\hline
$\langle a^2\rangle$ & $e$ & $a$ & perfect\\\cline{3-4}
& & $ab$ & perfect\\\hline
& & $a$ & (1) semiperfect\\\cline{3-4}
& $e$ & $a^3$ & perfect\\\cline{3-4}
& & $a^5$ & (2) semiperfect\\\cline{2-4}
& & $a$ & (3) semiperfect\\\cline{3-4}
$\langle b\rangle$ & $a^2$ & $a^3$ & perfect\\\cline{3-4}
& & $a^5$ & equivalent to (1)\\\cline{2-4}
& & $a$ & equivalent to (2)\\\cline{3-4}
& $a^4$ & $a^3$ & perfect\\\cline{3-4}
& & $a^5$ & equivalent to (3)\\\hline
& & $ab$ & perfect\\\cline{3-4}
& & $a$ & (4) semiperfect\\\cline{3-4}
$\{e\}$ & $e$ & $a^3b$ & perfect\\\cline{3-4}
& & $a^3$ & perfect\\\cline{3-4}
& & $a^5b$ & perfect\\\cline{3-4}
& & $a^5$ & equivalent to (4)\\\hline
\end{tabular}
\end{center}
\caption{Subgroups $J$, left coset representatives $l$ of $N_H(J)$ in $H=\langle a^2,b\rangle$, and right coset representatives $r^l$ of $J^l$
in $G=\langle a,b\rangle$ that should be considered to enumerate all inequivalent Type I $H$-invariant partitions of $G$}
\label{Table1}
\end{tablehere}
\medskip
It is convenient to use Theorem \ref{thm5}(a) to determine whether a coloring corresponding to a Type I partition of $G$ is perfect or semiperfect especially when $G$ is infinite. Consider again the uncolored infinite repeating pattern in Figure \ref{fig2} whose symmetry group is $G=\langle a,b,x,y\rangle$. Suppose we choose the subgroup $H=\langle b,a^2b,x,y\rangle$ of type $pmm$ and of index 2 in $G$, the subgroup $J=\langle xa^2b,xy,xy^{-1}\rangle$ of type $cm$ in $H$, and $r=a\in G\setminus H$. Since $J$ does not contain any rotations, $a^2\notin J$. Hence, a coloring obtained from the partition $\{h(J\cup Jr):h\in H\}$ of $G$ must be semiperfect by Theorem \ref{thm5}(a).
Theorem \ref{thm5}(a) has many immediate consequences. For instance, if $J\unlhd G$ then a coloring induced by the partition $P=\{h(J\cup Jr):h\in H\}$ of $G$ is perfect if and only if $r^2\in J$. In particular, if $J=\{e\}$ then $P$ corresponds to a perfect coloring if and only if $r$ is an involution. We also conclude that when enumerating inequivalent Type I $H$-invariant partitions of $G$, there is no need to consider $J=H$ since the partition obtained will always correspond to a perfect coloring. Also, no Type I partition of the dihedral group $D_n=\langle a,b:a^n=b^2={(ab)}^2=e\rangle$ gives rise to chirally perfect colorings (that is, semiperfect colorings where the associated color group is $\langle a\rangle$).
We conclude this section with a useful geometric consequence of Theorem \ref{thm5}(a). It makes use of the \emph{diagram}, $D(S)$, of a set $S$ of isometries in $\mathbf{R}^n$, which is the set of symmetry elements of the isometries in $S$.
\begin{corollary}\label{cor6}
Let $H$ be a subgroup of index 2 in $G$, $J\leq H$, and $r\in G\setminus H$. If $rD(J)\neq D(J)$ then a coloring induced by the partition
$\{h(J\cup Jr):h\in H\}$ of $G$ is semiperfect.
\end{corollary}
\begin{proof}
We know that $D(rJr^{-1})=rD(J)$ (see \cite{CFF}). Hence, if $rD(J)\neq D(J)$ then $rJ\neq Jr$.
\end{proof}
\begin{figurehere}
\centering{\includegraphics[width=65mm]{Figure5.eps}}
\caption{$D(J)$ and $(a^2b)D(J)$ where $J=\langle a^3b,xy,x^{-1}y\rangle$}
\label{fig4}
\end{figurehere}
\medskip For example, consider the subgroup $H=\langle xa,ab,xy,x^{-1}y\rangle$ of type $p4m$ and of index 2 in the symmetry group $G=\langle a,b,x,y\rangle$ of the pattern in Figure \ref{fig2}. We choose $J=\langle a^3b,xy,x^{-1}y\rangle\leq H$ (of type $pm$) and $r=a^2b\in G\setminus H$. Note that $a^2b$ is the reflection whose axis (dotted blue line) is shown in Figure \ref{fig4} and ${(a^2b)}^2=e\in J$. Now, $D(J)$ consists of the entire plane and the axes of reflections shown in black in Figure \ref{fig4}. Hence, the image of $D(J)$ under $a^2b$ is the entire plane and the mirror axes shown in red in Figure \ref{fig4}. We see that $D(J)\neq (a^2b)D(J)$ and so a coloring corresponding to the partition $\{h(J\cup Jr):h\in H\}$ of $G$ will be semiperfect by Corollary \ref{cor6}.
\section{Colorings associated with $(Y_i,J_i)-H$ partitions of $G$ for different subgroups $H$ of index 2 in $G$}
Let $H$ and $H'$ be two distinct subgroups of index 2 in $G$. If $P$ and $P'$ are $H$- and $H'$-invariant partitions of $G$, respectively, then the colorings corresponding to $P$ and $P'$ must be inequivalent by Theorem \ref{thm1}. Hence, we need to consider as $H$ all subgroups of index 2 in $G$ to enumerate all inequivalent partitions of $G$ that correspond to semiperfect colorings.
\noindent\begin{center}\begin{figurehere}
\begin{minipage}[c]{40mm}
\centering{\includegraphics[width=38mm]{Figure4a.eps}}
\medskip\centering{(a)}
\end{minipage}
\begin{minipage}[c]{40mm}
\centering{\includegraphics[width=38mm]{Figure4b.eps}}
\medskip\centering{(b)}
\end{minipage}
\caption{(a) A semiperfect coloring inequivalent to Figure \ref{fig1}(b) ; (b) The reflection $\alpha\in N(G)$ and corresponding labels of fundamental domains of $G=\langle a,b\rangle$}
\label{fig5}
\end{figurehere}\end{center}
\medskip Recall that the color group associated to the semiperfect coloring in Figure \ref{fig1}(b) is $H=\langle a^2,b\rangle$. The coloring of the same hexagonal pattern in Figure \ref{fig5}(a) is also semiperfect since its associated color group $H'=\langle a^2,ab\rangle$ is of index 2 in $G$. Since the associated color groups of the semiperfect colorings in Figure \ref{fig1}(b) and Figure \ref{fig4}(a) are different, they are not equivalent by Theorem \ref{thm1}. This is also evident geometrically because the color patterns in the two colorings are not congruent.
However, if one looks closely, even though the colorings in Figures \ref{fig1}(b) and \ref{fig5}(a) are inequivalent, they seem to be ``similar'' in some respects. This phenomenon was noted by Macdonald and Street in \cite{MacS} and was considered by Roth in \cite{Ro1}, \cite{Ro2}, and \cite{Ro3}.
We will show that even if the colorings in Figures \ref{fig1}(b) and \ref{fig5}(a) are inequivalent, one may be obtained from the other by considering the images of the fundamental domains of $G$ under some element $\alpha$ of the normalizer of $G$, $N(G)$, in the group of isometries of $\mathbf{R}^n$. To achieve this, we make use of the next result.
\begin{theorem}\label{thm7}
Let $H$, $H'$ be subgroups of index 2 in $G$ with $H'=H^{\alpha}$ for some $\alpha\in N(G)$.
\begin{itemize}
\item If $D$ is a fundamental domain of $G$ and $\alpha (D)=D$ then the action of $\alpha$ on
$\{g(D):g\in G\}$ is equivalent to the action of $\alpha$ on $G$ by conjugation.
\item If $P$ and $P'$ are $(Y_i,J_i)-H$ and $({Y_i}^{\alpha},{J_i}^{\alpha})-H'$ partitions of $G$, respectively,
then $P'=P^{\alpha}$. Moreover, a coloring corresponding to $P$ is perfect if and only if a coloring corresponding
to $P'$ is perfect.
\item If $C$ and $C'$ are the sets of colors in colorings corresponding to $(Y_i,J_i)-H$ and $({Y_i}^{\alpha},{J_i}^{\alpha})-H'$
partitions of $G$, respectively, then the action of $H$ on $C$ is equivalent to the action of $H'$ on $C'$.
\end{itemize}
\end{theorem}
\begin{proof}
\begin{itemize}
\item[]
\item There is a bijection between $G$ and the set $\{g(D):g\in G\}$ given by $g\leftrightarrow g(D)$. Hence,
$g^{\alpha}\in G$ is in a one-to-one correspondence with $g^{\alpha}(D)=\alpha(g(D))$ since $\alpha$, and hence $\alpha^{-1}$,
stabilizes $D$.
\item Note that if $J_i\leq H$ then ${J_i}^{\alpha}\leq H'$ for $i\in I$. Also, if $Y=\{e,y\}$ is a complete set of right coset
representatives of $H$ in $G$, then $Y'=\{e,y^{\alpha}\}$ is a complete set of right coset representatives of $H'$ in $G$. Hence,
we obtain
\begin{alignat*}{2}
P^{\alpha}&=\{(\alpha h\alpha^{-1})(\alpha J_i\alpha^{-1})\{e,\alpha y\alpha^{-1}\}:i\in I, h\in H\}\\
&=\{h'{J_i}^{\alpha}\{e,y^{\alpha}\}:i\in I,h'\in H'\}\\
&=P'.
\end{alignat*}
If $P$ is a Type I partition, then by Theorem \ref{thm5}(a), either both $P$ and $P'=P^{\alpha}$ correspond to perfect colorings or
both correspond to semiperfect colorings since $Jy=yJ \Leftrightarrow J^{\alpha}y^{\alpha}=y^{\alpha}J^{\alpha}$ and
$y^2\in J\Leftrightarrow {(y^{\alpha})}^2\in J^{\alpha}$. The same result holds even if $P$ is of Type II because
$J_2={J_1}^y\Leftrightarrow {J_2}^{\alpha}={({J_1}^{\alpha})}^{y^{\alpha}}$.
\item The elements of $C$ and $C'$ are in a one-to-one correspondence with the elements of the partitions $P=\{hJ_iY_i:i\in I,h\in
H\}$ and $P'=\{h'J_i'Y_i':i\in I, h'\in H'\}$ of $G$, respectively, where $J_i'={J_i}^{\alpha}$ and $Y_i'={Y_i}^{\alpha}$ for $i\in
I$. Since $P'=P^{\alpha}$, there is a bijection $f$ from $P$ to $P'$ given by $f(hJ_iY_i):=\alpha(hJ_iY_i)\alpha^{-1}
=h^{\alpha}J_i'Y_i'$. Let $\phi$ be the isomorphism from $H$ to $H'$ given by $\phi(h)=h^{\alpha}$. The groups $H$ and $H'$ act on
$P$ and $P'$, respectively, by left multiplication, and hence $\forall g\in H$, $\phi(g)\cdot
f(hJ_iY_i)=g^{\alpha}(h^{\alpha}J_i'Y_i')=f(g\cdot hJ_iY_i)$.\qedhere
\end{itemize}
\end{proof}
Suppose $H'=H^{\alpha}$ for some $\alpha\in N(G)$. Theorem \ref{thm7}(b) tells us that $(Y_i',J_i')-H'$ partitions of $G$ can be obtained by conjugating the $(Y_i,J_i)-H$ partitions of $G$ by $\alpha$. If $\alpha$ is chosen such that it stabilizes a fundamental domain $D$ of $G$, then by Theorem \ref{thm7}(a), not only can we obtain the colorings corresponding to $(Y_i',J_i')-H'$ partitions of $G$ from the colorings corresponding to $(Y_i,J_i)-H$ partitions of $G$ analytically by looking at the partitions, but also geometrically from the colorings. That is, if $\alpha(D)=D$, then a coloring associated to the $({Y_i}^{\alpha},{J_i}^{\alpha})-H'$ partition of $G$ may be obtained from a coloring associated to the $(Y_i,J_i)-H$ partition of $G$ in the following manner: If $R$ is the object of the pattern contained in $D$ then for all $g\in G$,
\begin{enumerate}
\item associate the color of $g(R)$, in a coloring corresponding to the $(Y_i,J_i)-H$ partition of $G$, to $g(D)$;
\item determine $\alpha(g(D))$;
\item designate the color of $g(D)$ as the new color of $\alpha(g(D))$;
\item assign to each $g^{\alpha}(R)$ the new color of $\alpha(g(D))$ and we obtain a coloring corresponding to the
$({Y_i}^{\alpha},{J_i}^{\alpha})-H'$
partition of $G$.
\end{enumerate}
Note that by Theorem \ref{thm7}(b), the two colorings that we get are either both perfect or both semiperfect.
Finally, even if the colorings induced by the $(Y_i',J_i')-H'$ partitions of $G$ are not equivalent to any of the colorings induced by the $(Y_i,J_i)-H$ partitions of $G$, Theorem \ref{thm7}(c) tells us that the color permutation groups that we obtain for both sets of colorings are isomorphic. We illustrate all of these in the following examples.
Consider again the uncolored hexagonal pattern in Figure \ref{fig1}(a) whose symmetry group is $G=\langle a,b\rangle$. If we denote by $c$ the $30^{\circ}$-counterclockwise rotation about the center of the hexagon, then $N(G)=\langle c,b\rangle\cong D_{12}$.
Let $\alpha=cb\in N(G)$ be the reflection along the angle bisector of the angle formed by the axes of the reflections $b$ and $ab$ (see Figure \ref{fig5}(b)). Observe that $\alpha$ stabilizes the fundamental domain labeled ``$e$'' in Figure \ref{fig5}(b). Also, the subgroups $H=\langle a^2,b\rangle$ and $H'=\langle a^2, ab\rangle$ of index 2 in $G$ are conjugate subgroups in $N(G)$ with $H'=H^{\alpha}$.
Recall that we obtained the coloring in Figure \ref{fig1}(b) by considering the partition
\begin{alignat*}{2}
P&=\{h\langle a^2b\rangle\{e\}:h\in H\}\cup\{hH\{a^3\}:h\in H\}\\
&=\{\{e,a^2b\},\{b,a^4\},\{a^4b,a^2\},\{ab, a, a^3b,a^3, a^5b, a^5\}
\end{alignat*}
of $G$. Now, consider the partition
\begin{alignat*}{2}
P'&=\{h'{\langle a^2b\rangle}^\alpha\{e\}:h'\in H'\}\cup\{h'H^{\alpha}\{{(a^3)}^{\alpha}\}:h'\in H'\}\\
&=\{h'\langle a^5b\rangle\{e\}:h'\in H'\}\cup\{h'H'\{a^3\}:h'\in H'\}\\
&=\{\{e,a^5b\},\{ab,a^2\},\{a^3b,a^4\},\{a, a^2b, a^3, a^4b, a^5, b\}\}
\end{alignat*}
of $G$. We see that Figure \ref{fig5}(a) is a coloring associated to $P'$. Hence, we are able to obtain the coloring in Figure \ref{fig5}(a) from the coloring in Figure \ref{fig1}(b) by computing their corresponding partitions of $G$, as stated in Theorem \ref{thm7}(b). Moreover, both colorings are semiperfect.
If we consider the images of the fundamental domain labeled ``$e$'' by elements of $G$, then we get the labelling of the fundamental domains as shown in Figure \ref{fig5}(b). We now discuss how we can transform the coloring in Figure \ref{fig1}(b) to the coloring in Figure \ref{fig5}(a). We associate the colors in the coloring in Figure \ref{fig1}(b) to their corresponding fundamental domains. Getting the image of each fundamental domain under $\alpha$, we obtain a new assignment of colors to the fundamental domains (see Table \ref{Table2}). Associating the new colors designated to each fundamental domain to the corresponding tile of the pattern, and changing the color yellow to purple, green to white, blue to light blue, and red to gray, we then obtain the coloring in Figure \ref{fig5}(a).
\begin{tablehere}
\begin{center}
\begin{tabular}{|c||c|c|c|}
\hline
Fundamental & Original & Image & New\\
domain & color & under $\alpha$ & color\\\hline\hline
$e$ & yellow & $e$ & yellow\\\hline
$ab$ & red & $b$ & green\\\hline
$a$ & red & $a^5$ & red\\\hline
$a^2b$ & yellow & $a^5b$ & red\\\hline
$a^2$ & blue & $a^4$ & green\\\hline
$a^3b$ & red & $a^4b$ & blue\\\hline
$a^3$ & red & $a^3$ & red\\\hline
$a^4b$ & blue & $a^3b$ & red\\\hline
$a^4$ & green & $a^2$ & blue\\\hline
$a^5b$ & red & $a^2b$ & yellow\\\hline
$a^5$ & red & $a$ & red\\\hline
$b$ & green & $ab$ &red\\\hline
\end{tabular}
\end{center}
\caption{Original color, image under $\alpha$, and new color of the fundamental domains of $G=\langle a,b\rangle$}
\label{Table2}
\end{tablehere}
\medskip Denote the colors blue, green, red, yellow by 1, 2, 3, 4, respectively, and the colors light blue, white, gray, purple by 1', 2', 3', 4', respectively. Then the set of colors of the coloring in Figure \ref{fig1}(b) is $C=\{1,2,3,4\}$ and the set of colors of the coloring in Figure \ref{fig5}(a) is $C'=\{1',2',3',4'\}$. Table \ref{Table3} gives the color permutations induced by the elements of $H$ and $H'$ on their corresponding colorings. Clearly, the action of $H$ on $C$ is equivalent to the action of $H'$ on $C'$.
\begin{tablehere}
\begin{center}
\begin{tabular}{|c|c||c|c|}
\hline
$h\in H$ & Color & $h^{\alpha}\in H'$ & Color \\
& permutation && permutation\\\hline\hline
$e$ & (1) & $e$ & (1')\\\hline
$a^2$ & (124) & $a^4$ & (1'2'4')\\\hline
$a^4$ & (142) & $a^2$ & (1'4'2')\\\hline
$b$ & (24) & $ab$ & (2'4')\\\hline
$a^2b$ & (12) & $a^5b$ & (1'2')\\\hline
$a^4b$ & (14) & $a^3b$ & (1'4')\\\hline
\end{tabular}
\end{center}
\caption{Color permutations corresponding to the action of the elements of $H$ and $H'$ on $C$ and $C'$, respectively}
\label{Table3}
\end{tablehere}
\medskip We also give an example for the infinite repeating pattern in Figure \ref{fig2} whose symmetry group is $G=\langle a,b,x,y\rangle$. Figure \ref{fig6}(a) shows the fundamental domains of $G$. Let $\alpha\in N(G)$ be the reflection along the broken line in Figure \ref{fig6}(a). Note that $\alpha$ stabilizes the fundamental domain labeled ``$e$''. The groups $H=\langle a,ab,xy,x^{-1}y\rangle$ and $H'=\langle
xa,ab,xy,x^{-1}y\rangle$ are both subgroups of index 2 in $G$ and of type $p4m$ with $H'=H^{\alpha}$. Choose $J=\langle a^2b,b,x^2,y^2\rangle\leq H$ (of type $pmm$) and $Y=\{e,xab\}$. Since ${(xab)}^2=xy\notin J$, the coloring induced by the $(Y,J)-H$ partition of $G$ in Figure \ref{fig6}(b) is semiperfect by Theorem \ref{thm5}(a). A semiperfect coloring corresponding to the $(Y^{\alpha},J^{\alpha})-H'$ coloring of $G$ is shown in Figure \ref{fig6}(c).
\noindent\begin{figurehere}
\centering{\includegraphics[width=70mm]{Figure6a.eps}}
\centering{(a)}
\begin{minipage}[c]{50mm}
\centering{\includegraphics[width=50mm]{Figure6b.eps}}
\centering{(b)}
\end{minipage}\quad
\begin{minipage}[c]{50mm}
\centering{\includegraphics[width=50mm]{Figure6c.eps}}
\centering{(c)}
\end{minipage}
\caption{(a) Fundamental domains of $G=\langle a,b,x,y\rangle$ of type $p4m$ and the reflection $\alpha\in N(G)$ ; (b) Semiperfect coloring
induced by the $(Y,J)-H$ partition of $G$, where $H=\langle a,ab,xy,x^{-1}y\rangle$ of type $p4m$, $J=\langle a^2b,b,x^2,y^2\rangle$ of type $pmm$, and $Y=\{e,xab\}$ ; (c) Semiperfect coloring induced by the $(Y^{\alpha},J^{\alpha})-H^{\alpha}$ partition of $G$}
\label{fig6}
\end{figurehere}
\medskip Therefore, to enumerate the semiperfect colorings of the hexagonal pattern in Figure \ref{fig1}(a), we may consider as $H$ only the subgroups $\langle a^2,b\rangle$ and $\langle a\rangle$ of $G$. The resulting 25 inequivalent semiperfect colorings of the hexagonal pattern can be seen in \cite{L}. In the case of infinite repeating patterns, we expect that there will be infinitely many semiperfect colorings. However, one may impose certain restrictions on the colorings, such as the number of colors. For instance, the 44 inequivalent semiperfect colorings of the infinite repeating pattern in Figure \ref{fig2} with at most four colors, one orbit of colors, and associated color group of type $p4m$, are listed in \cite{L}.
\section{Conclusion and Outlook}
In this paper, we considered semiperfect colorings of symmetrical patterns where the objects to be colored in the pattern are in a one-to-one correspondence with the elements of the symmetry group $G$ of the pattern. In order to enumerate all inequivalent semiperfect colorings of the pattern, we looked at the different $(Y_i,J_i)-H$ partitions of $G$ where $H$ is a subgroup of index 2 in $G$. We provided an organized and efficient method of identifying and counting the inequivalent $(Y_i,J_i)-H$ partitions of $G$ that correspond to semiperfect colorings. Moreover, inequivalent semiperfect colorings whose associated color groups are conjugate subgroups with respect to $N(G)$ were related by considering the corresponding partitions and the images of the fundamental domains of $G$ under some suitable element of $N(G)$.
Unfortunately, not all colorings of symmetrical patterns correspond to a partition of the symmetry group of the pattern. The next step would be to determine how to enumerate inequivalent semiperfect colorings of such patterns, examples of which are the Archimedean and hyperbolic tilings. Adapting results in this paper to the framework discussed in \cite{DLPFL} might be effective in achieving this goal.
Semiperfect colorings is just a small part of the broader theory of nonperfect colorings. One might look at the general case, that is, how to obtain all inequivalent colorings of a given symmetrical pattern whose associated color group is of index $n$ in the symmetry group of the pattern.
Lastly, it might be interesting to explore further colorings that are not really equivalent and yet one can be obtained from the other by a symmetry in the normalizer of the symmetry group of the pattern in the group of isometries.
\medskip\noindent\small\emph{Acknowledgments}. The second author gratefully acknowledges the financial assistance given by the University of the Philippines HRDO. Part of this work was carried out by the second author during his stay at the FSPM, University of Bielefeld, Germany.
|
\section{Introduction}
Very weakly interacting sub-eV particles (WISPs) appearing in a hidden sector of nature, i.e. a sector of particles carrying no standard model charges, can mix with photons.
This is the case of the standard graviton, and also of hypothetical particles such as axions, axion-like-particles (ALPs)~\cite{Raffelt:1987im} or hidden photons ($\gamma'$)~\cite{Okun:1982xi}.
In the last case, the mixing can be provided by a non-diagonal kinetic term, a so-called kinetic mixing that after a field redefinition appears as $\gamma-\gamma'$ mass mixing~\cite{Ahlers:2007rd}.
In all the previous cases, mixing cannot occur at tree level (these WISPs have spin different from 1).
However, the existence of WISP couplings to \emph{two} photons can produce an effective mixing term in an background magnetic field.
The WISP-photon mixing gives a non-diagonal contribution to the mass matrix which no longer allows photons to be propagation eigenstates. This leads to the phenomenon of photon oscillations and photon disappearance, analogously to the neutrino case.
The $\gamma\leftrightarrow$WISP conversion probability as a function of propagation length $L$ in a medium of index of refraction $n$ is given by
\begin{equation}
\label{prob}
P(\gamma\to \phi)=
\frac{4\delta^2}{(m_\phi^2-m_\gamma^2)^2+4\delta^2}
\sin^2 \(\frac{((m_\phi^2-m_\gamma^2)^2+4\delta^2)^{1/2}L}{4\omega}\)=\sin^2 2\theta \sin^2 \(\frac{\pi L}{L_{\rm osc}} \)
\end{equation}
where $m_\phi$ is the WISP mass and $m_\gamma^2 \simeq -2 \omega^2 (n-1)$ is an effective photon mass with $\omega$ the photon frequency.
The mixing term $\delta$ depends on the particular WISP.
For gravitons we have $\delta = \sqrt{32\pi} B \omega/M_{\rm Pl}$ where $B$ is the component of the external magnetic field perpendicular to the photon propagation direction and the Planck mass is $M_{\rm Pl}=1.22\times 10^{19}$ GeV.
Axions and ALPs have $\delta = g B \omega$, with $g$ the two photon coupling (widely discussed in this workshop) and \emph{they only mix with one photon polarization}.
For hidden photons we have $\delta_{\gamma'}=\chi m_{\gamma'}^2$ with $\chi$ the kinetic mixing parameter.
String motivated models give plausible values of $\chi$ in the $10^{-16}\sim 10^{-3}$ range~\cite{Dienes:1996z
}.
Note that in vacuum, $\gamma\leftrightarrow$WISP oscillations can be very suppressed if the WISP mass is much larger than $\delta$.
However, inspection of Eq.~(\ref{prob}) reveals that, even in this case, the amplitude of oscillations can be made maximal (1, indeed) in a medium which gives a photon effective mass that matches the WISP mass ($m_\gamma=m_\phi$) producing a resonant effect.
A typical experiment looking for $\gamma\leftrightarrow$WISP oscillations could be:
1) Take a very intense and well understood light source, 2) make it propagate the longest possible distance through 3) a medium whose index of refraction is fairly homogeneous and tunable and 4) try to detect a distortion in the light after propagation.
5) If no distortion is observed, tune another index of refraction (which makes resonant another WISP mass) and try again.
The cosmic microwave background provides an excellent source for such an experiment.
First, it is a very well measured and understood source: the black-body nature of its spectrum is well accounted in terms of QED and standard thermodynamics in the $\Lambda$CDM cosmological model, and was measured to a precision of $10^{-4}$ by the FIRAS on the COBE satellite~\cite{Fixsen:1996nj}. More precise measurements are also under consideration~\cite{Fixsen:2002,Kogut:1996zb}.
Second, the CMB photons travelled through the very homogeneous primordial plasma.
For most of the time, its index of refraction was smaller than one, which is a paramount requirement since otherwise $m_\gamma^2$ would be negative and a resonance impossible.
Moreover, as the universe slowly expanded, the plasma became increasingly sparse and the the index of refraction decreased accordingly.
As we will see, the photon effective mass swept all possible WISP masses.
We can therefore look for signatures of WISPs regardless of their mass.
Finally, CMB photons travel the longest conceivable distance for an experiment, basically the size of the universe, enhancing enormously the conversion probabilities and thus the WISP signatures.
\section{Photon mass in the early universe}
In our studies we have used a simplified model, yet containing the main features, for the effective photon mass in the primordial plasma. The typically dominant part is a positive contribution from the free electrons through the plasma frequency $\omega_{\rm P}$, while a negative (frequency-dependent) part from electrons bound in H atoms plays a role in special cases
\begin{equation}
m_\gamma^2 = \omega_\mathrm{_P}^2(X_e) \times\[1-0.0073 \(\frac{\omega}{\rm eV}\)^2 \(\frac{1-X_e}{X_e}\)\]\ ,
\end{equation}
where $\omega_\mathrm{_P}(X_e) \simeq 1.6\times 10^{-14}(1+z)^{3/2}X^{1/2}_e$~eV is the average plasma frequency in the current $\Lambda$CDM model and $X_e(z)$ is the hydrogen ionization fraction as a function of redshift (taken from~\cite{Seager:1999bc} for recombination and modelled around redshift $z\sim 7-10$ for reionization).
\section{Transition probability in an expanding universe}
The transition probability in Eq.~(\ref{prob}) is only valid in an homogeneous medium. In the expanding universe we should account for the variation in time or redshift of the different physical quantities and the problem becomes substantially more complex. Moreover, if the resonance happens before recombination photon scattering can be important during the resonance and shall be included.
The latter case was explored in a first paper focusing in the $\gamma'$ case~\cite{Jaeckel:2008fi} while resonances after recombination were presented in~\cite{Mirizzi:2009iz}. The ALP case was developed in~\cite{Mirizzi:2009nq}.
While every case is different, it turns out that the results are equivalent in the regime of small transition probability. In these proceedings we present another way of reaching the same result by using the perturbative solution of the dispersionless equations of motion of the $\gamma$-WISP system as presented in~\cite{Raffelt:1987im}.
Using the equivalence between length, time and redshift infinitesimals $dL\simeq dt = H^{-1}(1+z)^{-1}d z$ ($H$ the expansion parameter) we can write
\begin{equation}
P(\gamma\to \phi) =
\left|
\int
dt\, \frac{\delta (t)}{2\omega} {\rm Exp} \left\{\i \int^t dt' \frac{m_\phi^2-m_\gamma^2(t')}{2\omega} \right\} \right|^2\simeq
\pi \frac{\delta^2}{m_\phi^2 \omega H} \left|\frac{d \log m_\gamma^2}{d \log (1+z)}\right|_{z=z_r}^{-1}
\end{equation}
where for evaluating the integral we have used a saddle point approximation so that all quantities are to be evaluated in the resonance point $m_\phi=m_\gamma(z_r)$.
Note that due to the H refraction term several resonances occur for small masses and large frequencies~\cite{Mirizzi:2009iz,Mirizzi:2009nq}. There is however a dominant one for which the last expression makes full sense, but this depends on the specific WISP. Fixing the WISP mass and neglecting the log derivative which amounts an O(1) factor we find that:
1) in the hidden photon case $P\propto (\omega H)^{-1}$ which always decreases with redshift so the \emph{latest} resonance is the most relevant and
2) in the ALP or graviton case $P\propto B^2\omega/H$ so the \emph{earliest} resonance dominates\footnote{Primordial magnetic fields usually increase with redshift faster than $(1+z)$, indeed in our studies we used the most conventional assumption that $B=B_0(1+z)^2$.}.
\section{Signatures of a hidden CMB}
The early $\gamma\to$WISP conversions can leave different footprints in the CMB depending on when the resonant conversion happens. The CMB is unprotected from spectral distortions below a temperature $T\sim$ keV and the $\gamma\to$WISP conversions are frequency dependent so they generally distort the blackbody shape. A careful\footnote{Whenever the resonance happens before recombination we included the re-thermalization processes of the photon spectrum.} $\chi^2$ analysis of the FIRAS monopole results allowed us to set strong constraints on hidden photons and ALPs with masses smaller than $\sim 0.2$ meV. Beyond this mass, the $\gamma\to$WISP resonance happens when photons can regain a thermal distribution by interacting with the primordial plasma. This of course makes the FIRAS bounds disappear.
Nevertheles, the WISPs produced contribute to the dark matter of the universe and therefore affect structure formation.
Their oscillation origin makes these WISPs to have a similar spectrum than photons, so they are in fact hot dark matter relics. As such, they behave in a completely similar fashion to the standard neutrinos by free-streaming out of the primordial over-densities and suppressing the power spectrum at small scales.
Their effects can be included in the number of effective neutrino species
\begin{equation}
N_\nu^{\rm eff}(x)=\frac{N_\nu}{1-x}+\frac{8}{7}\frac{x}{1-x}
\left(\frac{11}{4}\right)^{4/3}
\end{equation}
where $x=\rho_\phi/\rho_\gamma$ is the fraction of the original photon density converted into WISPs during the resonance and $N_\nu$ is the effective number of neutrinos before the resonance. Comparing the value of $N_\nu^{\rm eff}$ recently inferred from WMAP5, other anisotropy probes, large scale structure surveys and supernova data with the standard value $N_\nu = 3.046$ gives $x<0.2$ with $95\%$ C.L.
This limit translates into severe constraints for the $\gamma\to$WISP mixing.
A summary of the bounds obtained in~\cite{Jaeckel:2008fi,Mirizzi:2009iz,Mirizzi:2009nq} is shown in Fig.~\ref{Fig:bounds}. Note that for ALPs we only can constraint the product of the coupling times a sky averaged magnetic field during the resonance $g\langle B^2\rangle^{1/2}$.
These results are of little use currently but can eventually turn into a fabulous diagnosis tool in the case of the discovery of an ALP, in which case one could constraint the existence of primordial magnetic fields (PMFs). Other posibility is that PMFs are discovered by other means in which case out bound will constraint the $\gamma$-ALP coupling. Bounds on $g$ lie around the $10^{-10}$ GeV$^{-1}$ ballpark while those on PMFs are slightly above the nG. If the discovery of any of those is experimentally around the corner our bounds on the other can be very relevant.
This seems to imply that the detection of both $g$ and PMFs is very unlikely in the short term, specially for very small ALP masses.
The graviton case can be read from the ALP graph when $m_\phi\to 0$.
Since the coupling is known one obtains a bound on the primordial field intensity of $69\ \mu$G.
In the hidden photon case, our bounds complement and typically beat the previous bounds from modifications of the Coulomb's law.
\begin{figure}[t]
\centering
{
\psfragscanon
\psfrag{a1}[][l][0.6]{post-recombination}
\psfrag{a2}[][l][0.6]{weak-coupling}
\psfrag{a3}[][l][0.6]{ \hspace{0.2cm}$\mu$}
\psfrag{a4}[][l][0.6]{ \hspace{0.3cm}$\Delta N_{\rm eff}$}
\psfrag{malp}[][l][0.7]{$m_{\phi}$ [eV] }
\psfrag{galp}[][l][0.7]{$g \langle B^2 \rangle^{^{1/2}}\times 10^{10}$ GeV $\times$ nG\vspace{1cm}}
\includegraphics[width=7cm]{redondo_javier_2.fig.1.eps}
}
{\psfragscanon
\psfrag{a}[][l][0.7]{$m_{\gamma'}$ [eV]}
\psfrag{b}[][l][0.7]{$\chi$}
\includegraphics[width=7cm]{redondo_javier_2.fig.2.eps}
}
\vspace{-0.35cm}
\caption{\footnotesize Constraints in the ALP (left) and hidden photon (right) parameter space from the signatures on the CMB left by a primordial $\gamma\to$WISP resonant conversion. $\Delta N_\nu^{\rm eff}$ is bounded by the WMAP5 power spectrum while for lower masses the bounds come from the FIRAS blackbody measurements.
In the hidden photon case, the yellowish region is excluded by experimental searches of deviations of the Coulomb's law~\cite{Bartlett:1988yy}.
}
\label{Fig:bounds}
\vspace{-0.35cm}
\end{figure}
\section*{Acknowledgements}
I would like to thank the organisers of this very interesting meeting and acknowledge support of the DFG cluster of excellence EXC 153 ``Origin and Structure of the Universe''.
\begin{footnotesize}
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
Quantum mechanics (QM) can be instructed either by adopting the schemes
proposed by its inventors (Born, Heisenberg, Schr\"{o}dinger, Jordan) or,
more rigorously, following Dirac, within the Hilbert space framework, or
even using Feynman's path integral approach. As it is believed to be
pedagogically more appealing, almost all textbooks prefer to begin with the
nonrelativistic approach, discussing the wave-particle dualism,
wavefunctions, Schr\"{o}dinger equation, Hilbert space, non-commutative
operators, etc. However, looking at the emblematic dynamical equations of
Schr\"{o}dinger and Dirac, one notes that the Schr\"{o}dinger equation (SE)
is less fundamental than Dirac's relativistic equation, because this one
contains inherently the internal degree of freedom spin, while it is absent
in the former. In between there is the Pauli-Schr\"{o}dinger equation (PSE),
which was derived by Pauli when he applied the low energy approximation in
Dirac equation. Although being nonrelativistic yet the PSE is more complete
than SE because the spin is inherently present, while in the SE the spin
must be added as an extra degree of freedom. See references in \cite%
{HESTENES} for a detailed discussion.
In the beginning of the 1980's the possibility of quantum computation was
foreseen by people like Benioff, Feynman and Deutsch \cite%
{BENIOFF,FEYNMAN,DEUTSCH}, and their work influenced a mini-revolution that
began in the 1990\'s, which threw a new look in QM, mainly in its
interpretation and potentialities to explain new phenomena; in the last 15
years we have witnessed huge theoretical developments along with ingenious
experiments involving single or few atoms or molecules, electrons and
photons. So, the understanding of quantum physics has widen, shaping the new
arena called \emph{quantum information theory (QIT) }that borrowed many
concepts of classical information theory. In this context it seems to exist
a recognition \cite{FUCHS} that QM is a special kind of information theory
immersed in Hilbert space, and characterized by a reversible logic \cite%
{BENNETT,DIVICENZO,FREDKIN}.
In that connection, by using elementary concepts of communication theory, as
bits and gates, however represented in the framework of Hilbert space, I
will show here that a quantum evolution equation for one classical bit (%
\emph{Cbit}$,$\emph{\ }as defined in \cite{MERMIN}) or quantum bit (\emph{%
qubit}) of information can be derived. Then, by asking what could be the
\emph{carrier} of one qubit (or spin\ 1/2), the natural choice is a particle
of mass \emph{m} characterized by its kinetic energy, and this information
is introduced in the qubit dynamical equation. This procedure is sufficient
to derive the PSE that rules the time evolution of both, the qubit and the
particle, its carrier. So, it ceases to be a particle in the Newtonian sense
to become a hybrid compelled to display wave properties and described by a
wavefunction. The qubit/spin evolution acquires an ascendancy over the
particle motion, being at the root of the observed quantum properties of
matter. Last but not least, the Dirac equation and its solution are briefly
discussed in terms of qubits.
\section{Cbits and actions}
In classical information theory the numbers in $\mathbb{Z}_{2}=\left\{
0,1\right\} $ are associated to bits, as in a relay or in a memory storage
device. One can go one step further and associate a particular
representation to the numbers 0 and 1: a column matrix for one classical bit
of information, the Cbit state, $1\longrightarrow \binom{1}{0}$ and $%
0\longrightarrow \binom{0}{1}$, as like the states \textquotedblleft
up\textquotedblright\ and \textquotedblleft down\textquotedblright\ for the
spin 1/2. These states can be written in the more familiar form of Dirac's
kets $\left\vert x\right\rangle $, $\left\vert \bar{x}\right\rangle $ ($\bar{%
x}=1-x$) for $\left\{ x,\bar{x}\right\} \in \mathbb{Z}_{2}$, and the bras
are the transposed, so $\mathcal{H}_{2}^{\times }$ $\equiv \left(
\left\langle 1\right\vert ,\left\langle 0\right\vert \right) $ is the dual
space of $\mathcal{H}_{2}\equiv \left( \left\vert 0\right\rangle ,\left\vert
1\right\rangle \right) $. The qubit $\binom{a}{b}$ is a generalization of
the Cbit, with $a$ and $b$ being complex numbers. The simplest operators to
be used are the identity $\boldsymbol{I}$, $\boldsymbol{I}\left\vert
x\right\rangle =\left\vert x\right\rangle $, and the NOT $\boldsymbol{X}$
that inverts the Cbit state, $\boldsymbol{X}\left\vert x\right\rangle
=\left\vert \bar{x}\right\rangle $. So the 4-uple $\mathcal{Q}=\left\{
\mathbb{Z}_{2},\mathcal{H}_{2},\mathcal{H}_{2}^{\times },\mathcal{L}%
_{2}\right\} $, $\mathcal{L}_{2}=\left\{ \boldsymbol{I},\boldsymbol{X}%
\right\} $ plus the field of complex numbers $\mathbb{C}$ are sufficient
tools for my purposes.
The \emph{action} $\boldsymbol{U}\left( \alpha ,\beta \right) \equiv \alpha
\boldsymbol{I}+\beta \boldsymbol{X}$ is a linear map of a Cbit or qubit into
a qubit, $\boldsymbol{U}\left( \alpha ,\beta \right) \left\vert
x_{0}\right\rangle \rightarrow \left\vert x_{1}\right\rangle =\alpha
\left\vert x_{0}\right\rangle +\beta \left\vert \bar{x}_{0}\right\rangle $,
for arbitrary parameter $\alpha $, $\beta $ in $\mathbb{C}$. Let's first
restrict the parameters values to two numbers: $\beta =$ $\bar{\alpha}%
=1-\alpha $ and $\left\{ \alpha ,\bar{\alpha}\right\} \in \mathbb{Z}_{2}$,
such that $\boldsymbol{U}_{\alpha }\left\vert x_{0}\right\rangle \equiv
\alpha \left\vert x_{0}\right\rangle +\bar{\alpha}\left\vert \bar{x}%
_{0}\right\rangle $ is still a Cbit because $\alpha \bar{\alpha}=0$. The
actions $\left\{ \boldsymbol{U}_{0},\boldsymbol{U}_{1}\right\} $ form a
group: (a) the unit element is $\boldsymbol{I=U}_{0}$ $,$ while $\boldsymbol{%
X=U}_{1}$; (b) the inverse is $\boldsymbol{U}_{\alpha }^{-1}=$ $\boldsymbol{U%
}_{\alpha }$; (c) the product of two elements is an element in the group $%
\boldsymbol{U}_{\alpha _{2}}\boldsymbol{U}_{\alpha _{1}}=\boldsymbol{U}%
_{\beta }=\beta \boldsymbol{I}+\bar{\beta}\boldsymbol{X}$ with $\beta
=\alpha _{2}\alpha _{1}+\bar{\alpha}_{2}\bar{\alpha}_{1}$, and $\bar{\beta}=%
\overline{\alpha _{2}\alpha _{1}+\bar{\alpha}_{2}\bar{\alpha}_{1}}=\alpha
_{2}\bar{\alpha}_{1}+\bar{\alpha}_{2}\alpha _{1}$; (d) the associative
property $\left( \boldsymbol{U}_{\alpha _{3}}\boldsymbol{U}_{\alpha
_{2}}\right) \boldsymbol{U}_{\alpha _{1}}=\boldsymbol{U}_{\alpha _{3}}\left(
\boldsymbol{U}_{\alpha _{2}}\boldsymbol{U}_{\alpha _{1}}\right) $ holds, and
the elements are unitary $\boldsymbol{U}_{\alpha }^{\dagger }=\boldsymbol{U}%
_{\alpha }^{-1}$.
Sequential $n$ actions
\begin{equation}
\boldsymbol{U}_{n}\left( \vec{\alpha}\right) \equiv \prod_{j=1}^{n}\left(
\alpha _{j}\boldsymbol{I}+\bar{\alpha}_{j}\boldsymbol{X}\right) ,
\label{evol}
\end{equation}%
applied on a Cbit maps it into another Cbit, $\boldsymbol{U}_{n}\left( \vec{%
\alpha}\right) \left\vert x_{0}\right\rangle =\left\vert x_{n}\right\rangle $%
, going through the intermediary states $\left\{ \left\vert
x_{1}\right\rangle ,\left\vert x_{2}\right\rangle ,\left\vert
x_{3}\right\rangle ,...,\left\vert x_{n-1}\right\rangle \right\} $. Each set
of numbers $h_{n}=\left\{ \alpha _{n},...,\alpha _{3},\alpha _{2},\alpha
_{1}\right\} $ defines one \emph{history}, or \emph{trajectory}. One can
also write $\boldsymbol{U}\left( \alpha _{1}\right) \left\vert
x_{0}\right\rangle $ by changing the label of the Cbit state, and by a
trivial formal manipulation it is simple to show that $\boldsymbol{U}%
_{\alpha _{1}}\left\vert x_{0}\right\rangle =\left\vert x_{1}\right\rangle
=\left\vert \alpha _{1}x_{0}+\bar{\alpha}_{1}\bar{x}_{0}\right\rangle $,
identifying the label as the mapped bit $x_{1}\equiv \alpha _{1}x_{0}+\bar{%
\alpha}_{1}\bar{x}_{0}$. By induction $x_{n}=\alpha _{n}x_{n-1}+\bar{\alpha}%
_{n}\bar{x}_{n-1}$ for $n=1,2,3,...$. Also holds the transitivity property
expressed by the composition law
\begin{equation}
\boldsymbol{U}_{n}\left( \vec{\alpha}^{\left( 2\right) }\right) \boldsymbol{U%
}_{m}\left( \vec{\alpha}^{\left( 1\right) }\right) =\boldsymbol{U}%
_{n+m}\left( \vec{\alpha}\right) . \label{trans1}
\end{equation}%
The sequence of actions (\ref{evol}) is reversible since each one is
unitary, then $\left\vert x_{0}\right\rangle =\boldsymbol{U}_{n}^{-1}\left(
\vec{\alpha}\right) \left\vert x_{n}\right\rangle $. The reverse history is
given by the sequence $h_{n}^{-1}=\left\{ \alpha _{1},...,\alpha
_{n-2},\alpha _{n-1},\alpha _{n}\right\} $. In summary, (\ref{evol}) carries
the evolution $\left\vert x_{0}\right\rangle \longrightarrow $ $\left\vert
x_{n}\right\rangle $, and $\boldsymbol{U}_{n}^{-1}\left( \vec{\alpha}\right)
=\prod_{j=n}^{1}\left( \alpha _{j}\boldsymbol{I}+\bar{\alpha}_{j}\boldsymbol{%
X}\right) $ does the inverse path, $\left\vert x_{n}\right\rangle
\longrightarrow \left\vert x_{0}\right\rangle $. Formally, $\boldsymbol{U}%
_{n}\left( \vec{\alpha}\right) $ and $\boldsymbol{U}_{n}^{-1}\left( \vec{%
\alpha}\right) $ are the same since each factor in Eq. (\ref{evol}) commutes
with all others.
\subsection{ Coefficients on a circle of unit radius}
I now assume the parameters $\alpha $ and $\beta $ in $\boldsymbol{U}\left(
\alpha ,\beta \right) $ to be real, with $\alpha \beta \neq 0$ and ask $%
\alpha ^{2}+\beta ^{2}=1$, so $\left( \alpha ,\beta \right) \in \mathbb{%
\tilde{R}}_{2}$ is the set of all real numbers on a circle of radius $1$. As
so, acting on a Cbit one gets a qubit, $\boldsymbol{U}\left( \alpha ,\beta
\right) \left\vert x_{0}\right\rangle =\alpha \left\vert x_{0}\right\rangle
+\beta \left\vert \bar{x}_{0}\right\rangle $. Two consecutive operations
give $\boldsymbol{U}\left( \alpha _{2},\beta _{2}\right) \boldsymbol{U}%
\left( \alpha _{1},\beta _{1}\right) =\boldsymbol{U}\left( \alpha _{3},\beta
_{3}\right) $, and as $\alpha _{1}^{2}+\beta _{1}^{2}=$ $\alpha
_{2}^{2}+\beta _{2}^{2}=1$, it follows that $\alpha _{3}^{2}+\beta
_{3}^{2}=1+4\alpha _{2}\alpha _{1}\beta _{2}\beta _{1}\neq 1$, so $\left(
\alpha _{3},\beta _{3}\right) \notin \mathbb{\tilde{R}}_{2}$ and $%
\boldsymbol{U}\left( \alpha _{3},\beta _{3}\right) $ is \emph{not} an
element of the group, unless one of the four coefficients is zero, therefore
any probabilistic interpretation for $\alpha ^{2}$ and $\beta ^{2}$ fails.
Moreover, the inverse action is $\boldsymbol{U}^{-1}\left( \alpha ,\beta
\right) =\tilde{\alpha}\boldsymbol{I}+\tilde{\beta}\boldsymbol{X}$, where
the new parameters $\tilde{\alpha}=\alpha /\left( \alpha ^{2}-\beta
^{2}\right) $ and $\tilde{\beta}=-\beta \left( \alpha ^{2}-\beta ^{2}\right)
$ are \emph{not} in $\mathbb{\tilde{R}}_{2}$. Due to the reality of $\alpha $
and $\beta $, $\boldsymbol{U}\left( \alpha ,\beta \right) $ is a
self-adjoint operator $\boldsymbol{U}^{\dagger }\left( \alpha ,\beta \right)
=\boldsymbol{U}\left( \alpha ,\beta \right) $ however it is\textit{\ }not
unitary since $\boldsymbol{U}^{\dagger }\left( \alpha ,\beta \right) \neq
\boldsymbol{U}^{-1}\left( \alpha ,\beta \right) $. Although the norm $%
\left\Vert \boldsymbol{U}\left( \alpha ,\beta \right) \left\vert
x_{0}\right\rangle \right\Vert =1$ is parameter independent, this is not
true for the inverse $\left\Vert \boldsymbol{U}^{-1}\left( \alpha ,\beta
\right) \left\vert x_{0}\right\rangle \right\Vert =\left\vert \alpha
^{2}-\beta ^{2}\right\vert ^{-1}$. Thus, if we want to construct an
evolution operator $\boldsymbol{U}_{n}\left( \vec{\alpha},\vec{\beta}\right)
=\prod_{j=1}^{n}\left( \alpha _{j}\boldsymbol{I}+\beta _{j}\boldsymbol{X}%
\right) $, with $\alpha _{j}^{2}+\beta _{j}^{2}=1$, that is also reversible,
we are in trouble. Since the inverse of $\boldsymbol{U}\left( \alpha
_{j},\beta _{j}\right) $ is $\boldsymbol{U}\left( \tilde{\alpha}_{j},\tilde{%
\beta}_{j}\right) $, for a sequence of $n$ inverse actions we have $%
\boldsymbol{U}_{n}^{-1}\left( \vec{\alpha},\vec{\beta}\right)
=\prod_{j=n}^{1}\boldsymbol{U}\left( \tilde{\alpha}_{j},\tilde{\beta}%
_{j}\right) $, however as $\tilde{\alpha}_{j}^{2}+\tilde{\beta}%
_{j}^{2}=\left( \alpha _{j}^{2}-\beta _{j}^{2}\right) ^{-2}\neq 1$,
therefore normalization is not possible.
\subsection{Invertibility and complex coefficients}
In order to establish the invertibility of $\boldsymbol{U}\left( \alpha
,\beta \right) $, the domain of $\alpha $ and $\beta $ must be be extended
to the field of complex numbers because the conditions $\left\vert \alpha
\right\vert ^{2}+\left\vert \beta \right\vert ^{2}=1$ and $\alpha ^{2}-\beta
^{2}=1\Longrightarrow \left\vert \beta \right\vert ^{2}+\beta ^{2}=0$ must
be satisfied. This happens for $\alpha $ real and $\beta =-i\left\vert \beta
\right\vert $, a pure imaginary. Since one is left with one free parameter,
a natural parametrization is $\alpha =\cos \xi $ and $\beta =-i\sin \xi $ ($%
\xi $ real), thus $\boldsymbol{U}\left( \alpha ,\beta \right) \equiv
\boldsymbol{U}\left( \xi \right) =\cos \xi \ \boldsymbol{I}-i\sin \xi \
\boldsymbol{X}$ \ is a unitary operator mapping a Cbit or a qubit into a
qubit, $\boldsymbol{U}\left( \xi \right) \left\vert x_{0}\right\rangle =\cos
\xi \ \left\vert x_{0}\right\rangle -i\sin \xi \ \left\vert \bar{x}%
_{0}\right\rangle $. So, the complex nature of $\boldsymbol{U}\left( \xi
\right) $ is due to its invertibility property. A sequence of actions
\begin{equation}
\boldsymbol{U}_{n}\left( \vec{\xi}\right) =\prod_{j=1}^{n}\left( \cos \xi
_{j}\ \boldsymbol{I}-i\sin \xi _{j}\ \boldsymbol{X}\right) , \label{U-1n}
\end{equation}%
on a Cbit $\left\vert x_{0}\right\rangle $ takes it to the qubit $%
\boldsymbol{U}_{n}\left( \vec{\xi}\right) \left\vert x_{0}\right\rangle
=A_{n}\left( \vec{\xi}\right) \left\vert x_{0}\right\rangle +B_{n}\left(
\vec{\xi}\right) \left\vert \bar{x}_{0}\right\rangle =\left\vert \psi
_{n}\right\rangle $, with coefficients $A_{n}\left( \vec{\xi}\right) =\cos
\left( \sum_{j=1}^{n}\xi _{j}\right) $ and $B_{n}\left( \vec{\xi}\right)
=-i\sin \left( \sum_{j=1}^{n}\xi _{j}\right) $. The parameters $\xi _{j}$
are undetermined and their sum is $\phi _{n}=\sum_{j=1}^{n}\xi _{j}$, so one
can write (\ref{U-1n}) in the compact form%
\begin{equation}
\boldsymbol{U}_{n}\left( \vec{\xi}\right) \Longrightarrow \boldsymbol{U}%
\left( \phi _{n}\right) =\exp \left[ -i\phi _{n}~\boldsymbol{X}\right] ,
\label{Uexp}
\end{equation}%
where $\phi _{n}$ is interpreted as a register parameter, it sets the
ordering of the actions. Due to the indetermination of the parameters $\xi
_{j}$ nothing can be said about the intervals between consecutive actions,
see Figure (\ref{sequen}-a),\FRAME{ftbpFU}{2.8784in}{1.6571in}{0pt}{\Qcb{%
{\protect\small (a) Undetermined intervals between sequences of actions. (b)
Uniformization of the intervals.}}}{\Qlb{sequen}}{sequencias 2.jpg}{\special%
{language "Scientific Word";type "GRAPHIC";display "USEDEF";valid_file
"F";width 2.8784in;height 1.6571in;depth 0pt;original-width
9.5998in;original-height 7.1996in;cropleft "0";croptop "1";cropright
"1";cropbottom "0";filename 'sequencias 2.jpg';file-properties "XNPEU";}}
the vertical bars stand for each action, they can be distributed at will,
although obeying an ordered sequence. Imposing the composition law (\ref%
{trans1})\ one has $\boldsymbol{U}\left( \phi _{n}\right) \boldsymbol{U}%
\left( \phi _{m}\right) =\boldsymbol{U}\left( \phi _{n+m}\right) $, and the
form (\ref{Uexp}) implies $\phi _{n}+\phi _{m}=\phi _{n+m}$; as so,
necessarily and uniquely $\phi _{n}$ must be linear in $n$, $\phi _{n}=n\bar{%
\xi}$ with $\bar{\xi}$ some parameter. Thus $\boldsymbol{U}\left( \phi
_{n}\right) $ becomes $\boldsymbol{U}\left( n\bar{\xi}\right) =\exp \left[
-in\bar{\xi}~\boldsymbol{X}\right] $, which stands for a sequence of
actions, or an evolution. The previously undetermined intervals between
actions become equally spaced, see Figure (\ref{sequen}-b), characterizing
the \emph{uniformization} of their distribution. In order to turn the
distribution dense I shall look for a differential equation for $\boldsymbol{%
U}\left( n\bar{\xi}\right) $ by taking first the difference between two
consecutive values of $n$ and then dividing by $\bar{\xi}$,
\begin{equation*}
\frac{\boldsymbol{U}\left( \left( n+1\right) \bar{\xi}\right) -\boldsymbol{U}%
\left( n\bar{\xi}\right) }{\bar{\xi}}=\left( \frac{e^{-i\bar{\xi}\boldsymbol{%
X}}-1}{\bar{\xi}}\right) \exp \left[ -in\bar{\xi}~\boldsymbol{X}\right] .
\end{equation*}%
The limit to a continuous parameter is obtained for $n\gg 1$ and $\bar{\xi}%
\ll 1$, keeping however the product $n\bar{\xi}=\tau $ finite. A linear
differential equation results, $id\boldsymbol{U}\left( \tau \right) /d\tau =%
\boldsymbol{XU}\left( \tau \right) $, and $\boldsymbol{U}\left( \tau \right)
=e^{-i\tau \boldsymbol{X}}$, where $\tau $ is the continuous ordering
parameter of the actions, or a local\emph{\ time}\textit{\ }in arbitrary
units, that should be set according to the clock to be used. Writing $%
\left\vert x_{\tau }\right\rangle =\boldsymbol{U}\left( \tau \right)
\left\vert x_{0}\right\rangle $ the evolution equation $id\left\vert x_{\tau
}\right\rangle /d\tau =\boldsymbol{X}\left\vert x_{\tau }\right\rangle $
says how a qubit evolves due to the action of $\boldsymbol{X}$, which is the
generator of the changes.
Defining a more general generator, $\boldsymbol{G}=\mu \boldsymbol{I}+\nu
\boldsymbol{X}$, $\mu $ and $\nu $ being two real parameters, the evolution
equation writes
\begin{equation}
i\frac{d\left\vert \psi _{\tau }\right\rangle }{d\tau }=\boldsymbol{G}%
\left\vert \psi _{\tau }\right\rangle , \label{eqsch2}
\end{equation}%
with $\boldsymbol{U}\left( \tau \right) =e^{-i\tau \mu \boldsymbol{I}%
}e^{-i\tau \nu \boldsymbol{X}}$ for the evolution operator. Differently from
the factor $e^{-i\tau \nu \boldsymbol{X}}$ that do really affect the
evolution of a qubit, the phase factor $e^{-i\tau \mu \boldsymbol{I}}$ is
apparently no significant because, besides a global phase factor, it does
not entail any change when acting on Cbit or qubit. The eigenvalues and
eigenstates of $\boldsymbol{G}$ are respectively $G_{\pm 1}=\mu \pm \nu $, $%
\left\vert x_{\pm 1}\right\rangle =\left( \left\vert 0\right\rangle \pm
\left\vert 1\right\rangle \right) /\sqrt{2}$. A general solution to Eq. (\ref%
{eqsch2}) is $\left\vert \psi _{\tau }\right\rangle =\sum_{\sigma =\pm
1}e^{-iG_{\sigma }\tau }c_{\sigma }\left\vert x_{\sigma }\right\rangle $
where $G_{\sigma }=\mu +\sigma \nu $. Now conjecturing about the qubit
carrier, I assume it a massive particle \cite{LANDAUER91} and the parameter $%
\mu $ is chosen to represent its energy; thus the change $\boldsymbol{%
X\rightarrow }$ $\boldsymbol{G}$ is important because it allows the
introduction of that particle property. $\boldsymbol{G}$ can be identified
as a hamiltonian, and for an arbitrary initial condition the mean value is $%
\left\langle \psi _{\tau }\right\vert \boldsymbol{G}\left\vert \psi _{\tau
}\right\rangle =\mu +\nu \left( \left\vert c_{+1}\right\vert ^{2}-\left\vert
c_{-1}\right\vert ^{2}\right) $; while $\mu $ is the particle kinetic
energy, the second term is the qubit energy that exists only when it is
coupled to some field ($\nu \neq 0$). {\LARGE \ }
\section{The qubit carrier and the Pauli-Schr\"{o}dinger equation}
The spatial localization of the carrier must be introduced into Eq. (\ref%
{eqsch2}), thus for a qubit state $\left\vert \psi _{0}\right\rangle
=a_{0}\left\vert x_{0}\right\rangle +b_{0}\left\vert \bar{x}%
_{0}\right\rangle $ the parameters $a_{0},b_{0}$ should depend on the
position $q$, namely, $\left\vert \psi _{0}\left( q\right) \right\rangle
=a_{0}\left( q\right) \left\vert x_{0}\right\rangle +b_{0}\left( q\right)
\left\vert \bar{x}_{0}\right\rangle $ becomes the state of the whole system,
with normalization $\int dq\ \left\vert a_{0}\left( q\right) \right\vert
^{2}+\int dq\ \left\vert b_{0}\left( q\right) \right\vert ^{2}=1$. The qubit
state is correlated to the particle position that influences its probability
outcomes $\left\vert a_{0}\left( q\right) \right\vert ^{2}$ and $\
\left\vert b_{0}\left( q\right) \right\vert ^{2}$. Coordinate dependence
should also be present in the generator, so $\boldsymbol{G}\left( q\right)
=\mu \left( q\right) \boldsymbol{I}+\nu \boldsymbol{X}$ and the parameter $%
\nu $ is assumed $q$-independent because interaction between both degrees of
freedom is not considered. The evolved state is $\boldsymbol{U}\left( \tau
\right) \left\vert \psi _{0}\left( q\right) \right\rangle =\left\vert \psi
\left( q,\tau \right) \right\rangle =a_{\tau }\left( q\right) \left\vert
x_{0}\right\rangle +b_{\tau }\left( q\right) \left\vert \bar{x}%
_{0}\right\rangle $, with amplitudes
\begin{eqnarray}
a_{\tau }\left( q\right) &=&e^{-i\tau \mu \left( q\right) }\left(
a_{0}\left( q\right) \cos \nu \tau -ib_{0}\left( q\right) \sin \nu \tau
\right) \ \label{atau} \\
b_{\tau }\left( q\right) &=&e^{-i\tau \mu \left( q\right) }\left(
-ia_{0}\left( q\right) \sin \nu \tau +b_{0}\left( q\right) \cos \nu \tau
\right) . \label{btau}
\end{eqnarray}%
with $a_{0}\left( q\right) $ and $b_{0}\left( q\right) $ as initial values.
So, the qubit was merged with the spatial motion of its carrier within a
single equation, meaning that the joint evolution -- the qubit sequence of
actions as well as the change in the spatial configuration of the carrier --
is measured by a single clock. To determine the parameters $a_{0}\left(
q\right) $ and $b_{0}\left( q\right) $ they should obey some differential
equation for the variable $q$, then $\mu \left( q\right) $ must depend also
on $\partial /\partial q$ and/or its powers. However, instead of trying to
guess the functional form, it is better to take advantage of the available
information from hamiltonian mechanics, so I define $\mu $ as the kinetic
energy of a non-relativistic particle, $\mu \Longrightarrow T\left( p\right)
$ $=p^{2}/2m$, where $p$ is the linear momentum in some reference frame. Eq.
(\ref{eqsch2}) becomes $i\kappa _{0}d\left\vert \tilde{\psi}\left( p,\tau
\right) \right\rangle /d\tau =\left[ T\left( p\right) \boldsymbol{I}%
+\varepsilon _{0}^{\prime }\nu \boldsymbol{X}\right] \left\vert \tilde{\psi}%
\left( p,\tau \right) \right\rangle $. Since $T\left( p\right) $ has units
of energy, the second term in brackets should also have the same units. So
the constants $\kappa _{0}$ and $\varepsilon _{0}^{\prime }$, have both
units of energy. One can also choose some unit to measure the dimensionless
time $\tau $, $\tau =t/t_{0}$, so the dynamical equation becomes%
\begin{equation}
ih_{0}\frac{d\left\vert \tilde{\psi}\left( p,t\right) \right\rangle }{dt}=%
\left[ T\left( p\right) \boldsymbol{I}+\varepsilon _{0}\boldsymbol{X}\right]
\left\vert \tilde{\psi}\left( p,t\right) \right\rangle , \label{schreq2}
\end{equation}%
where $h_{0}=\kappa _{0}t_{0}$, $\varepsilon _{0}=\varepsilon _{0}^{\prime
}t_{0}\nu $. Note that the constant $h_{0}$ has units of energy $\times $
time and $\varepsilon _{0}$ has units of energy. An arbitrary initial
condition assumes that the particle momentum and the qubit state are
correlated and the probability amplitude
\begin{equation}
\left\vert \tilde{\psi}_{0}\left( p,0\right) \right\rangle =\left\vert
\tilde{\psi}_{0}\left( p\right) \right\rangle =\tilde{a}_{0}\left( p\right)
\left\vert x_{0}\right\rangle +\tilde{b}_{0}\left( p\right) \left\vert \bar{x%
}_{0}\right\rangle , \label{supmom}
\end{equation}%
depends on the particle momentum and it contains all the available
information. In momentum space the evolution operator is $\boldsymbol{U}%
\left( t\right) =\exp \left[ -it\left( T\left( p\right) \boldsymbol{I}%
+\varepsilon _{0}\nu \boldsymbol{X}\right) /h_{0}\right] $ and the solution
to Eq. (\ref{schreq2}) is%
\begin{equation*}
\left\vert \tilde{\psi}\left( p,t\right) \right\rangle =e^{-itT\left(
p\right) /h_{0}}\left[ \tilde{a}_{t}\left( p\right) \left\vert
x_{0}\right\rangle +\tilde{b}_{t}\left( p\right) \left\vert \bar{x}%
_{0}\right\rangle \right]
\end{equation*}%
with $\tilde{a}_{t}\left( p\right) =\cos \left( \varepsilon _{0}\nu
t/h_{0}\right) \tilde{a}_{0}\left( p\right) -i\sin \left( \varepsilon
_{0}\nu t/h_{0}\right) \tilde{b}_{0}\left( p\right) $ and $\tilde{b}%
_{t}\left( p\right) =\cos \left( \varepsilon _{0}\nu t/h_{0}\right) \tilde{b}%
_{0}\left( p\right) -i\sin \left( \varepsilon _{0}\nu t/h_{0}\right) \tilde{a%
}_{0}\left( p\right) $, and the particle mean energy is $\left\langle \tilde{%
\psi}\left( p,t\right) \right\vert \boldsymbol{H}\left( p\right) \left\vert
\tilde{\psi}\left( p,t\right) \right\rangle =T\left( p\right) +2\varepsilon
_{0}{Re}\left( \tilde{a}_{0}^{\ast }\left( p\right) \tilde{b}_{0}\left(
p\right) \right) $. Since coordinate and momentum are conjugated variables
the statevector in coordinate representation is
\begin{equation*}
\left\vert \psi \left( q,t\right) \right\rangle =\psi _{x_{0}}\left(
q,t\right) \left\vert x_{0}\right\rangle +\psi _{\bar{x}_{0}}\left(
q,t\right) \left\vert \bar{x}_{0}\right\rangle ,
\end{equation*}%
and $\psi _{{x_{0}}}\left( q,t\right) $, $\psi _{\bar{x}_{0}}\left(
q,t\right) $ are the amplitudes associated to the Cbits $\left\vert
x_{0}\right\rangle $, $\left\vert \bar{x}_{0}\right\rangle $;$\ $they can be
written as Fourier transforms
\begin{equation}
\psi _{\binom{x_{0}}{\bar{x}_{0}}}\left( q,t\right) =\int \frac{dp}{2\pi }%
e^{ipq/h_{1}}e^{-itT\left( p\right) /h_{0}}\binom{\tilde{a}_{t}\left(
p\right) }{\tilde{b}_{t}\left( p\right) }. \label{schreq22}
\end{equation}%
The constant $h_{1}$ is introduced to set the correct dimensionality, it has
the same units as $h_{0}$, nonetheless nothing can be said about being the
same constant, unless confirmed by experiment. In Eq. (\ref{schreq22})
\begin{equation}
\binom{\tilde{a}_{t}\left( p\right) }{\tilde{b}_{t}\left( p\right) }=\int
dq^{\prime }e^{-ipq^{\prime }/h_{1}}\binom{a_{t}\left( q^{\prime }\right) }{%
b_{t}\left( q^{\prime }\right) }, \label{psi02}
\end{equation}%
with $a_{t}\left( q^{\prime }\right) =\cos \left( \varepsilon _{0}\nu
t/h_{0}\right) a_{0}\left( q^{\prime }\right) -i\sin \left( \varepsilon
_{0}\nu t/h_{0}\right) b_{0}\left( q^{\prime }\right) $ and $b_{t}\left(
q^{\prime }\right) =\cos \left( \varepsilon _{0}\nu t/h_{0}\right)
b_{0}\left( q^{\prime }\right) -i\sin \left( \varepsilon _{0}\nu
t/h_{0}\right) a_{0}\left( q^{\prime }\right) $. So even not existing a
direct interaction between the qubit and its carrier, the probability for
measuring the qubit in Cbit $\left\vert x_{0}\right\rangle $, or $\left\vert
\bar{x}_{0}\right\rangle $, becomes affected by its position.
Using Eqs. (\ref{schreq22}) and (\ref{psi02}) and manipulating Eq. (\ref%
{schreq2}) it is not hard to verify that one can substitute the c-number $p$
by the derivative $-ih_{1}\partial /\partial q$, and we can rewrite that
equation as
\begin{equation}
ih_{0}\frac{\partial \left\vert \psi \left( q,t\right) \right\rangle }{%
\partial t}=\left[ \frac{1}{2m}\left( -ih_{1}\frac{\partial }{\partial q}%
\right) ^{2}\boldsymbol{I}+\varepsilon _{0}\nu \boldsymbol{X}\right]
\left\vert \psi \left( q,t\right) \right\rangle . \label{schreq3}
\end{equation}%
The terms in brackets stand for the particle and qubit hamiltonian in
coordinate and matrix representation, so the parameter $\mu $ becomes
determined. In the presence of an energy conserving potential $V\left(
q\right) $ the PSE takes its familiar form, with hamiltonian $\boldsymbol{H}%
=H_{0}\boldsymbol{I}+\varepsilon _{0}\nu \boldsymbol{X}$ and $H_{0}=\left[
(-ih_{1}/(2m))({\partial }^{2}/{\partial ^{2}q})+V\left( q\right) \right] $.
The particle described by Eq. (\ref{schreq3}) has now blurred classical
properties (it looses the sharp trajectory it has in phase space), its best
representation is a wavefunction and the appearance of quantum properties
are due to the qubit it is carrying. Any further generalization is trivial
and immediate: (1) from 1-D to 3-D in spacial coordinates, $\partial
/\partial q\rightarrow \nabla $ and (2) since any $2\times 2$ matrix can be
expanded in the basis formed by the unit matrix $\boldsymbol{I}$ and Pauli
matrices ($\sigma _{x}$, $\sigma _{y}$, and $\sigma _{z}$), then $\nu
\boldsymbol{X}\longrightarrow \vec{\nu}\cdot \boldsymbol{\vec{\sigma}}$.
\section{Dirac equation: two qubits of information}
Few words about Dirac equation $i\hbar \partial \left\vert \Psi _{D}\left(
t\right) \right\rangle /\partial t=\boldsymbol{H}_{D}\left\vert \Psi
_{D}\left( t\right) \right\rangle $, its hamiltonian is $\boldsymbol{H}_{D}=c%
\boldsymbol{\vec{\alpha}}\cdot \vec{p}+mc^{2}\boldsymbol{\beta }$ and the
four dimension-4 matrices $\boldsymbol{\vec{\alpha}}$, $\boldsymbol{\beta }$
\ satisfy the relations $\boldsymbol{\alpha }_{k}\boldsymbol{\alpha }_{l}+%
\boldsymbol{\alpha }_{l}\boldsymbol{\alpha }_{k}=2\boldsymbol{I}\delta _{kl}$%
, $\boldsymbol{\vec{\alpha}\beta }+\boldsymbol{\beta \vec{\alpha}=0}$ and $%
\boldsymbol{\beta }^{2}=0$. These matrices can be expressed as tensor
products of dimension-2 matrices, each one acting on its own qubit, $%
\boldsymbol{\vec{\alpha}\ =\ }\boldsymbol{\boldsymbol{X}_{1}\otimes \left(
\boldsymbol{X}_{2},i\boldsymbol{Y}_{2},\boldsymbol{Z}_{2}\right) =X}%
_{1}\otimes \boldsymbol{\vec{\sigma}}_{2}$, so $c\boldsymbol{\vec{\alpha}}%
\cdot \vec{p}=\boldsymbol{X}_{1}\otimes \left( c\vec{p}\cdot \boldsymbol{%
\vec{\sigma}}_{2}\right) $ and $\boldsymbol{\beta =Z}_{1}\otimes \boldsymbol{%
I}_{2}.$ Thus, Dirac's hamiltonian can be written as tensor products acting
on independent D-2 Hilbert subspaces $\boldsymbol{H}_{D}=\boldsymbol{Z}%
_{1}\otimes \left( mc^{2}\boldsymbol{I}_{2}\right) +\boldsymbol{X}%
_{1}\otimes \left( c\vec{p}\cdot \boldsymbol{\vec{\sigma}}_{2}\right) $.
Squaring $\boldsymbol{H}_{D}$ one gets the relativistic energy $\left(
\boldsymbol{H}_{D}\right) ^{2}=E_{p}^{2}\ \boldsymbol{I}_{1}\otimes
\boldsymbol{I}_{2}$, where $E_{p}^{2}=m^{2}c^{4}+c^{2}\vec{p}^{\ 2}$. The
time-dependent equation reduces into direct products of $2\times 2$ matrices
\begin{eqnarray*}
&&\left[ \boldsymbol{I}_{1}\otimes \left( i\hbar \frac{\partial }{\partial t}%
\boldsymbol{I}_{2}\right) -\boldsymbol{Z}_{1}\otimes \left( mc^{2}%
\boldsymbol{I}_{2}\right) -\boldsymbol{X}_{1}\otimes \left( c\vec{p}\cdot
\boldsymbol{\vec{\sigma}}_{2}\right) \right] \\
&\times &\left\vert \Psi _{D}\left( t\right) \right\rangle =0,
\end{eqnarray*}%
which is invariant under Lorentz transformation. The solutions are
\begin{eqnarray}
\left\vert \Psi ^{\lambda }\left( \vec{p},t\right) \right\rangle
&=&N_{\lambda }e^{-i\lambda tE_{p}}\left[ \left\vert 1\right\rangle
_{1}\left\vert \varphi \left( \vec{p}\right) \right\rangle _{2}\right.
\notag \\
&+&\left. \left\vert 0\right\rangle _{1}\frac{c\vec{p}\cdot \boldsymbol{\vec{%
\sigma}}_{2}}{mc^{2}+\lambda E_{p}}\left\vert \varphi \left( \vec{p}\right)
\right\rangle _{2}\right] , \label{fundirac}
\end{eqnarray}%
where
\begin{equation*}
\left\vert \varphi \left( \vec{p}\right) \right\rangle _{2}=\left(
\begin{array}{c}
\varphi _{+}\left( \vec{p}\right) \\
\varphi _{-}\left( \vec{p}\right)%
\end{array}%
\right) _{2}
\end{equation*}%
with $\lambda =\pm 1$ standing for positive and negative energy solutions
and $N_{\lambda }$ is a normalization constant. The qubit 2 in Eq. (\ref%
{fundirac}) represents the particle state whereas the Cbit 1 is apparently
ancillary, it works as a selector: the projector $\left( \left\vert
1\right\rangle \left\langle 1\right\vert \right) _{1}$ selects the
nonrelativistic component $\left\vert \varphi \left( \vec{p}\right)
\right\rangle _{2}$ while $\left( \left\vert 0\right\rangle \left\langle
0\right\vert \right) _{1}$ projects the relativistic complement. Also
interesting is that all the $\gamma _{\mu }$ matrices have the structure of
the direct product of two-qubit operators $\gamma ^{0}=\boldsymbol{Z}%
_{1}\otimes \boldsymbol{I}_{2}$, $\gamma ^{1}=i\boldsymbol{Y}_{1}\otimes
\boldsymbol{X}_{2}$, $\gamma ^{2}=-\boldsymbol{Y}_{1}\otimes \boldsymbol{Y}%
_{2}$, $\gamma ^{3}=i\boldsymbol{Y}_{1}\otimes \boldsymbol{Z}_{2}$. $\quad $
\section{Concluding remarks}
As long as the qubit is not probed (for a spin, there is no external
magnetic field), $\nu =0$, Eq. (\ref{schreq3}) reduces to two uncoupled Schr%
\"{o}dinger equations
\begin{equation}
ih_{0}\frac{\partial }{\partial t}\left(
\begin{array}{c}
\psi _{x_{0}}\left( q,t\right) \\
\psi _{\bar{x}_{0}}\left( q,t\right)%
\end{array}%
\right) =H_{0}\left(
\begin{array}{c}
\psi _{x_{0}}\left( q,t\right) \\
\psi _{\bar{x}_{0}}\left( q,t\right)%
\end{array}%
\right) , \label{schreq5}
\end{equation}%
for two wavefunctions $\psi _{x_{0}}\left( q,t\right) $, $\psi _{\bar{x}%
_{0}}\left( q,t\right) $ in one qubit/spin eigenstate. The two equations are
redundant thus the relevant information resides in one of them only, going
to the usual spinless SE. Although classical physics was crucial to arrive
at Eq. (\ref{schreq5}), by setting $\nu =0$ the particle motion is not ruled
(back) by classical physics (Hamilton equations) but by the usual SE. So why
do quantum properties of the particle still persist even when the
correlation between a qubit/spin and its carrier is broken? The answer is
that even not being activated the qubit/spin is still carried by the
particle, the internal degree of freedom and the particle make one single
object, although not entangled. One is left with an equation (SE) that does
not keep any clue about the presence of a qubit/spin, nonetheless it is
still there although not manifestly evident. That%
\'{}%
s why the SE can be used without any mention to spin if not needed;
otherwise, this internal degree of freedom must be appended in order to
explain observed phenomena. In conclusion, because it is carrying one qubit
of information, to the observer the particle shifts its behavior from the
classical picture, it acquires wave properties with a probabilistic
character where the uncertainty relations represent one facet.
\begin{acknowledgments}
Acknowledgements
\end{acknowledgments}
I wish to express thanks for the support from FAPESP and CNPq, Brazilian
agencies.
|
\section[I]{Introduction}
Weakly interacting massive particles (WIMP) as dark matter in our
galaxy can be trapped inside massive heavenly bodies like sun due to
later's gravity \cite{use2}. This gravitational trapping may happen when such
dark matter in course of its passage through sun undergoes elastic
scattering off the nuclei present in the solar core which causes its
final velocity to fall below its velocity of escape from solar
gravitational pull. These trapped dark matter may undergo the process
of pair-annihilation producing primarily $b$, $c$ and $t$ quarks,
$\tau$ leptons, gauge bosons etc. Mass and composition of
dark matter determine the annihilation products which in turn produce
neutrinos and antineutrinos either through decay or pair annihilation.
Such neutrinos from the dark matter annihilation in solar environment
have also been studied by previous authors (e.g. \cite{khlopov}).
Detection of such solar neutrinos in terrestrial detectors
not only provides indirect evidence of dark matter but
along with the results from the direct detection experiments of dark
matter such as CDMS \cite{cdms}, DAMA \cite{dama},
XENON \cite{xenon} etc. provides more insight into the nature
of dark matter and its interactions. Analysis of recent observations
of two dark matter signal events with the Cryogenic Dark Matter
Search experiment (CDMS II) at the Soudan Underground Laboratory
combined with all previous CDMS II data has been
reported by CDMS collaboration \cite{cdms1}. It sets new upper limits
on the WIMP-nucleon elastic scattering cross section
($\sigma_\chi$) as a function of WIMP mass ($m_\chi$)
\cite{cdms1}. The indirect searches for WIMPs through their annihilation
in sun, with 1679.6 live days of data from SK detector using
neutrino-induced upward through-going muons provide WIMP-induced
upward muon flux limits at SK as a function of WIMP mass \cite{skwimp}.
In this work we use the 90\% Confidence Level (C.L.) limits on
$\sigma_\chi (m_\chi)$ from recent CDMS analysis \cite{cdms1}
to calculate corresponding limits on detection rates of up-going
muons at SK as a function of WIMP mass ($m_\chi$) and compare
them with the results in \cite{skwimp}.
The differential flux of neutrinos of type $i$
($i = \nu_\mu, \bar{\nu}_\mu$) at earth from
WIMP annihilation products in the sun is given by \cite{jungman}
\begin{eqnarray}
{\left ( \frac {d\phi} {dE} \right ) }_i
&=&
\frac {\Gamma_A} { 4\pi R^2} \sum_F B_F
{ \left( \frac {dN} {dE} \right )}_{F,i}
\end{eqnarray}
where $R$ is sun-earth distance and
$B_F$ is the annihilation branch for channel $F$.
${(dN/dE)}_{F,i}$ is the differential spectrum of neutrinos
of type $i$ in the sun for the annihilation channel $F$.
The total rate for WIMP annihilation in the sun,
$\Gamma_A$ is given in \cite{jasonkumar}
\begin{eqnarray}
\Gamma_A
&=&
\frac{1}{2}C\tanh^2[{(aC)}^{1/2}\tau]
\end{eqnarray}
where $\tau \simeq 4.5$ Gyr is the age of sun.
$a = \langle \sigma v \rangle / 4\sqrt{2} V$ is a function of
the average WIMP annihilation cross section
and the effective volume $V$ of WIMPs in the sun
\cite{jasonkumar,use1,use2,use3,use4,use5}. Under
the astrophysical assumptions on density and velocity distribution
as mentioned in \cite{jasonkumar,use1} (local dark matter density,
$\rho_\chi = 0.3 $ GeV cm$^{-3}$, mean velocity of dark matter,
$\bar{v} = 300$ km sec$^{-1}$,
a Maxwellian distribution of velocities etc.)
the dark matter capture rate $C$ in the sun is approximated as a
function of the ratio of the WIMP-nucleus elastic scattering cross section
to the dark matter mass as \cite{jasonkumar,use2}
\begin{eqnarray}
C &\simeq& 10^{29} \frac{\sigma_\chi}{m_\chi}\ {\rm GeV}\ {\rm pb}^{-1}
{\rm sec}^{-1}
\label{eqc}
\end{eqnarray}
For such indirect detections of WIMPs at SK,
neutrinos are detected through up-going muons produced by
charged current interactions of neutrinos
with the rock below the detector. With numerical values of
this cross section, the muon range in the rock and expressing
the energy distribution of neutrino flux in terms of its second
moments, the total detection rate at SK detector of up-going muons
induced by neutrinos from WIMP annihilation in the sun is given by
\cite{jungman}
\begin{eqnarray}
\Gamma_{\rm detect}
&=&
(1.27\times10^{-29} {\rm yr}^{-1} )
\frac{C}{{\rm sec}^{-1}}
\left(\frac{m_\chi}{ {\rm GeV}}\right)^2
\displaystyle\sum_{i=\nu,\bar{\nu}} a_i b_i \sum_F B_F \langle {N_Z}^2 \rangle_{F,i} \times A_{eff}
\label{eqgammadetect}
\end{eqnarray}
where, $A_{eff} (\approx 1200 {\rm m}^2)$
is the muon effective area of the SK detector \cite{effarea},
$a_i$'s are the neutrino scattering coefficients.
The range of neutrino induced muons in the rock are given
as the coefficients $b_i$. These coefficients are given by
$a_\nu = 6.8$, $a_{\bar{\nu}} = 3.1$, $b_\nu = 0.51$,
$b_{\bar{\nu}} = 0.67$ \cite{jungman}. ${\langle {N_Z}^2
\rangle}_{F,i}$'s are the second moments of the spectrum
of neutrino type $i$ for the WIMP annihilation channel $F$ in the sun.
The ${\langle {N_Z}^2 \rangle}_{F,i}$ for different channels relevant
for the present calculations are listed below \cite{jungman,ritz}.
\begin{enumerate}
\item[(a)] $\tau\bar{\tau}$ channel:
\begin{eqnarray}
{\langle {N_Z}^2 \rangle}_i (E_{\rm inj})
&=&
\Gamma_{\tau\to \mu\nu\bar{\nu}} h_{\tau,i}(E_{\rm inj}\tau_i)
\quad (i = \nu_\mu,\bar{\nu}_\mu)
\end{eqnarray}
where $E_{\rm inj}$ is the injection energy of the decaying
WIMP annihilation product
inside the sun, the branching ratio
$\Gamma_{\tau\to \mu\nu\bar{\nu}} \simeq 0.18$,
$\tau_\nu (\tau_{\bar{\nu}}) = 1.01\times 10^{-3}
(3.8\times 10^{-4})$ GeV$^{-1}$
are the stopping coefficients for $\nu (\bar{\nu})$ and
\begin{eqnarray}
h_{\tau,\nu_\mu}(y)
&=&
\frac{4+y}{30(1+y)^4} \nonumber\\
h_{\tau,\bar{\nu}_\mu}(y)
&=&
\frac{168 + 354y + 348y^2 + 190y^3 + 56y^4 + 7y^5}{1260(1+y)}
\end{eqnarray}
\item[(b)] $b\bar{b}$ channel:
\begin{eqnarray}
{\langle {N_Z}^2 \rangle}_i (E_{inj})
&\simeq &
\Gamma_{b\to \mu\nu X} \frac{\langle E_d \rangle^2}{E_i^2}
h_{b,i}\left(\sqrt{\langle E_d^2 \rangle} \tau_i \right)
\end{eqnarray}
where, the branching ratio $\Gamma_{b\to \mu\nu X} = 0.103$.
The hadronization and the decay processes of the quarks from
WIMP annihilation in the sun are characterized by the mean
energy $\langle E_d \rangle$ of the hadron, $\langle E_d \rangle =
E_c \exp\left(\frac{E_c}{E_0}\right)E_1\left(\frac{E_c}{E_0}\right)$,
$E_c = 470$ GeV \cite{ritz} . $E_0 = Z_f E_{\rm inj}$ is the initial hadron
energy for quarks injected with energy $E_{\rm inj}$ and
$Z_f(=0.73)$ is the quenching fraction for $b$-quarks to account for
the loss of energy during hadronization,
$E_1(x) = \int_x^\infty \frac{e^{-y}}{y}dy$. Also
$\langle E_d^2\rangle = E_c (E_0 - \langle E_d \rangle)$ and
$h_{b,i}$ is same as $h_{\tau,i}$.
\item[(c)] $W^+W^-$ and $Z\bar{Z}$ channel:
\begin{eqnarray}
\left .{\langle {N_Z}^2 \rangle}_i (E_{\rm inj})\right|_{W}
&\simeq &
\left. \frac{\Gamma_{W\to \mu\nu}}{\beta}
\frac{2+2E\tau_i(1+\alpha_i)+E^2\tau_i^2\alpha_i(1+\alpha_i)}
{E_{\rm inj}^3 \tau_i^3 \alpha_i (\alpha_i^2 - 1)(1+E\tau_i)^{\alpha_i+1}}
\right|^{E=E_{\rm inj}(1-\beta)/2}_{E=E_{\rm inj}(1+\beta)/2} \\
&& \nonumber\\
\left .{\langle {N_Z}^2 \rangle}_i (E_{\rm inj})\right|_{Z}
&\simeq &
\left. \frac{2\Gamma_{Z\to \nu_\mu\bar{\nu}_\mu}}{\beta}
\frac{2+2E\tau_i(1+\alpha_i)+E^2\tau_i^2\alpha_i(1+\alpha_i)}
{E_{\rm inj}^3 \tau_i^3 \alpha_i (\alpha_i^2 - 1)(1+E\tau_i)^{\alpha_i+1}}
\right|^{E=E_{\rm inj}(1-\beta)/2}_{E=E_{\rm inj}(1+\beta)/2}
\label{eqnzz}
\end{eqnarray}
where,
$\alpha_\nu(\alpha_{\bar{\nu}}) = 5.1(9.0)$ and the branching
ratios $\Gamma_{W\to \mu\nu} = 0.105$ ,
$\Gamma_{Z\to \nu_\mu\bar{\nu}_\mu} = 0.067$.
$\beta$ being the velocity of the gauge bosons.
\end{enumerate}
The injection energy $E_{\rm inj}$ is the
energy with which the WIMP annihilation products
$b\bar {b}$, $\tau\bar {\tau}$, $W$ and $Z$ etc. are produced.
In this work we present our results for two benchmark scenarios namely
$E_{\rm inj} = m_\chi$ and $\frac {m_\chi} {3}$.
Recently the CDMS collaboration announced two dark
matter signal events with 90\% C.L.\cite{cdms1}
which set new limits on dark matter-nucleon
scattering cross-section $\sigma_\chi$ for
different $m_\chi$'s \cite{cdms1}. In
this work we choose our set of $\sigma_\chi$'s and
corresponding $m_\chi$'s from this limit
(exclusion plot for ``Soudan(all)" in Fig. 4 of Ref. \cite{cdms1}).
With these CDMS limits on $\sigma_\chi$ we compute the detection
rates for neutrino induced muons (from the product of the Dark
Matter annihilation in the sun) for different annihilation channels.
Thus we calculate the CDMS induced limits on event rates in SK.
We compute the detection rate for neutrino induced
muons (from the product of WIMP annihilation
in the sun) using Eq. 4.
\begin{figure}[t]
\includegraphics[width=8.0cm, height=7cm, angle=0]{events1.eps}
\vglue -7.0cm \hglue 9.0cm
\includegraphics[width=8.0cm, height=7cm, angle=0]{events3.eps}
\caption{\label{fig:rate}
CDMS upper bounds on WIMP induced muon event rates in different annihilation
channels with $E_{\rm inj} = m_\chi$ (left panel) and
$E_{\rm inj} = m_\chi/3$
(right panel)}
\end{figure}
In this work we present our results for two benchmark scenarios namely
$E_{\rm inj} = m_\chi$ and $\frac {m_\chi} {3}$.
A conservative estimate for neutrino production
from WIMP annihilation in the sun is obtained by
considering the dominating effect of
$b{\bar{b}}$ production in the region $m_b < m_\chi < m_W$, that of
$\tau\bar{\tau}$ and $W, Z$ production in the regions
$m_W < m_\chi < m_t$ and
$m_\chi > m_t$ respectively \cite{jungman},
where $m_b$, $m_t$, $m_W$ are respective masses of
$b$-quark, $t$-quark and $W$-boson.
The estimated limits of WIMP induced muon detection rates at SK using
Eqs.\ (\ref{eqgammadetect}-\ref{eqnzz}) in the annihilation
channels $b{\bar{b}}$ ,$\tau\bar{\tau}$ and $W$, $Z$
as a function of WIMP mass has been shown in
Fig.\ (\ref{fig:rate}). These rates are calculated
for $E_{\rm inj} = m_\chi$
(left panel of Fig.\ \ref{fig:rate})
and $E_{\rm inj} = m_\chi/3$ (right panel of Fig.\ \ref{fig:rate}).
Plots for each channel in Fig.\ (\ref{fig:rate})
are obtained by taking the value of branching fraction
($B_F$) for that channel to be $1$.
This can be justified by the fact that each channel is effective
in a particular $m_\chi$ region as discussed above. However, we have
shown in our results broader $m_\chi$ ranges for each of the channels.
Also shown in this figure is the corresponding
SK limit \cite{skwimp} and our results indicate existence for room
for an $m_\chi$ dependent enhancement of order
$10^{3}-10^{5}$ of the detection rate
in order to interpret SK bounds in terms of the CDMS
$\sigma_\chi(m_\chi)$ limit.
\begin{figure}[t]
\includegraphics[width=8.0cm, height=7cm, angle=0]{bvsm1.eps}
\vglue -7.0cm \hglue 9.0cm
\includegraphics[width=8.0cm, height=7cm, angle=0]{bvsm3.eps}
\caption{\label{fig:b} Plot of $B$ vs $m_\chi$ in each
channel with $E_{\rm inj} = m_\chi$ (left panel) and $E_{\rm inj} = m_\chi/3$
(right panel)}
\end{figure}
In a recent work, Cao et al \cite{caopamela} have shown that to
study the anti-proton fraction of the results of
satellite borne PAMELA experiment \cite{PAMELAnature, PAMELAprl}
using the recent CDMS results
one needs to invoke a boost factor ($\sim 10^3$)
for the annihilation of WIMPs into
quarks. This, in general is consistent with the enhancement needed to
explain the excess of positron fraction as observed by PAMELA
experiment \cite{PAMELAnature}. Our results also
indicate similar trends that if one uses the CDMS limit as is done in
the work, one is allowed to enhance the detection rate for the present
of SK bounds. We redefine $B_F$ in our framework (Eq. 4) as
$B \times B_F'$ where $B_F'$ is the branching fraction for a
particular process and $B$ is some boost factor. For a single channel
to represent the entire neutrino production process, ${B_F}^\prime = 1$.
Then we
estimate the upper limits of $B$'s for different dark matter annihilation
channels in the sun for the recent $\sigma_\chi - m_\chi$ CDMS limit. We do
this by comparing the detection rates of neutrinos --
obtained from the decays of the dark matter annihilation
products in the sun -- at Super-Kamiokande (SK) detector.
The upper limits on WIMP induced up-going muons at SK
as a function of WIMP mass is given by the SK Collaboration
\cite{skwimp}.
In Fig.\ \ref{fig:b} (left panel)
we show the variation of $B$ with $m_\chi$ for
$E_{\rm inj} = m_\chi$. Right panel of
Fig.\ \ref{fig:b} shows similar plots for
$E_{\rm inj} = m_\chi$/3.
From Fig.\ \ref{fig:b} (left panel) we see $B$ varies
(with $m_\chi$) between $\sim 10^5 - \sim 10^4$ for $b\bar{b}$ channel,
$\sim 10^4 - \sim 10^3$ for $\tau\tau$ channel, $\sim 10^3$ for each of W and Z
channels etc. For $E_{\rm inj} = m_\chi/3$ case (right panel of
Fig. \ref{fig:b})
however, the variation of $B$'s for each of W and Z channels
is between $\sim 10^3 - \sim 10^2$
and for $\tau\tau$ channel it is from $\sim 10^4 - 10^2$.
These limits are consistent with the order $10^3$ as discussed earlier.\\
\begin{figure}[t]
\includegraphics[width=8.0cm, height=7cm, angle=0]{bfwz1.eps}
\vglue -7.0cm \hglue 9.0cm
\includegraphics[width=8.0cm, height=7cm, angle=0]{bfwz3.eps}
\caption{\label{fig:wz} Iso $m_\chi$ plot in
$B\times {B_F}^\prime$ ($W$ channel)
-- $B\times {B_F}^\prime$ ($Z$ channel) plane interpreting SK limits
with $E_{\rm inj} = m_\chi$ (left panel) and $E_{\rm inj} = m_\chi/3$
(right panel).}
\end{figure}
Although in Figs.\ (\ref{fig:rate}-\ref{fig:b}) we have shown the
results for $W$ and $Z$ channels separately but as discussed earlier
both are dominant in the region $m_\chi > m_t$. Therefore
we consider the effect of both of them together in calculating the rate
using Eq. (4).
In Fig.\ \ref{fig:wz} (left panel)
we present in $B\times {B_F}^\prime$ ($W$ channel)
-- $B\times {B_F}^\prime$ ($Z$ channel) plane the iso $m_\chi$ plots for
$E_{\rm inj} = m_\chi$ that represent the required rate to interpret the
SK limits \cite{skwimp}. Right panel of
Fig. \ref{fig:wz} gives the similar plots but with
$E_{\rm inj} = m_\chi/3$.
In the present work we compute the WIMP induced upward going muon
rates at SK. They are induced by neutrinos from WIMP annihilation
products in the sun. We use recent CDMS bounds on WIMP-nucleon
scattering cross sections for different WIMP masses in our
rate calculation. It is observed that representation of SK
upper bounds on WIMP induced up-going muon rates allows an
enhancement in the calculated rates in all individual channels.
We made an estimation of this enhancement as a function of WIMP mass
assuming branching fractions for each different channels to be 1 (maximum).
Interpretation of PAMELA data in terms of CDMS limits
also demand such enhancements of WIMP annihilation.
\vskip 5mm
The authors like to acknowledge the hospitality and facilities at
XI Workshop on High Energy Physics Phenomenology (WHEPP XI)
at Physical Research Laboratory, Ahmedabad, India during which this work
was initiated and progressed.
The authors also thank S. Kundu and A. Ghosal for some useful discussions.
|
\section{Introduction}
The undisputed galactic origin of cosmic rays at energies below the so-called knee implies an existence of a non-thermal population of galactic sources which effectively accelerate protons and nuclei to TeV-PeV energies. The distinct signatures of these cosmic accelerators are high energy neutrinos and gamma rays produced through hadronic interactions with ambient gas or photoproduction on intense photon fields near the source. While gamma rays can be produced also by directly accelerated electrons, high-energy neutrinos provide unambiguous and unique information on the sites of the cosmic accelerators and hadronic nature of the accelerated particles.
ANTARES (http://antares.in2p3.fr/) is a deep-sea neutrino telescope, designed for the detection of all flavours of high-energy neutrinos emitted by both Galactic (supernova remnants, micro-quasars etc.) and extragalactic (gamma ray bursters, active galactic nuclei, etc.) astrophysical sources. The telescope is also sensitive to neutrinos produced via dark matter annihilation within massive bodies such as the Sun and the Earth. Other physics topics include measurement of neutrino oscillation parameters, the search for magnetic monopoles, nuclearites etc.
The recently completed ANTARES detector is currently the most sensitive neutrino observatory studying the southern hemisphere and includes the particularly interesting region of the Galactic Centre in its field of view. ANTARES is also a unique deep-sea marine observatory providing continuous, high-bandwidth monitoring from a variety of sensors dedicated to acoustic, oceanographic and Earth science studies.
\section{The ANTARES Detector}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_01}
\caption{The layout of the completed ANTARES detector. The top insert shows an image of an acoustic storey with
its six acoustic sensors and the lower insert an image of an optical storey with its three photomultipliers.}
\label{fig1}
\end{figure}
The ANTARES detector is located at a depth of 2475~m in the Mediterranean Sea, 42~km from La Seyne-sur-Mer in the South of France ($42^\circ48’ N, 6^\circ10’ E$). A schematic of the detector layout is shown in Figure \ref{fig1}. It is equipped with 885 optical sensors arranged on 12 flexible lines. Each line comprises up to 25 detection storeys each equipped with three downward-looking 10-inch photo-multipliers (PMTs), orientated at $45^\circ$ to the line axis. The lines are maintained vertical by a buoy at the top of the 450~m long line. The spacing between storeys in 14.5~m and the lines are spaced by 60-70~m. An acoustic positioning system provides real-time location of the detector elements to a precision of a few centimeters. A system of optical beacons allows in-situ time calibration. The first detection line was installed in 2006. Five lines have been operating since March 2007. Ten lines were operational in December 2007. With the installation of eleventh and twelfth lines in May 2008, the detector construction was completed. An additional line (IL07) contains an ensemble of oceanographic sensors dedicated measurement of the environmental parameters. The twelfth line and the IL07 also includes hydrophone-only storeys dedicated to the study of the ambient acoustic backgrounds; R\&D for possible acoustic detection of ultra-high energy neutrinos.
The ANTARES Collaboration currently comprises 29 particle physics, astrophysics and sea science institutes from seven countries (France, Germany, Italy, Netherlands, Romania, Russia and Spain).
The three-dimensional grid of photomultiplier tubes is used to measure the arrival time and position of Cherenkov photons induced by the passage of relativistic charged particles through the sea water. The reconstruction algorithm relies on the characteristic emission angle of the light (about 43 degrees) to determine the direction of the muon and hence infer that of the incident neutrino. The accuracy of the direction information allows to distinguish upward-going muons, produced by neutrinos, from the overwhelming background from downward-going muons, produced by cosmic ray interaction in the atmosphere above the detector. Installing the detector at great depths serves to attenuate this background and also allows to operate the PMTs in a completely dark environment.
At high energies the large range of the muon allows the sensitive volume of the detector to be significantly greater than the instrumented volume. Although optimised for muon neutrino detection, the detector is also sensitive to the electron and tau neutrinos albeit it with reduced effective area.
The total ANTARES sky coverage is 3.5$\pi$sr, with an instantaneous overlap of 0.5$\pi$sr with that of the Icecube experiment at the South Pole. Together ANTARES and Icecube provide complete coverage of the high-energy neutrino sky.
Compared to detectors based in ice, a water based telescope benefits from a better angular resolution, due to the absence of light scattering on dust and/or bubbles trapped in the ice. On the other hand, it suffers from additional background light produced by beta decay of $^{40}$K salt present in the sea water as well as bioluminescent light produced by biological organisms. Furthermore, the continual movement of the detector lines, in reaction to the changing direction and intensity of the deep-sea currents, must be measured and taken into account in the track reconstruction.
The ANTARES data acquistion \cite{bouwhuis1} is based on the 'all-data-to-shore' concept, in which all hits above a threshold of 0.3 single photon-electrons are digitised and transmitted to shore. Onshore a farm of commodity PCs apply a trigger based on requiring the presence of a 4-5 causally connected local coincidences between pairs of PMTs within a storey. The typical trigger rate is 5-10 Hz, dominated by downgoing muons. In addition, an external trigger generated by the gamma-ray bursts coordinates network (GCN) will cause all the buffered raw data (two minutes) to be stored on disk. This offers the potential to apply looser triggers offline on this subset of the data \cite{bouwhuis2}.
\section{Detector Calibration}
The precision on the neutrino direction is limited at low energies by the kinematics of the neutrino interaction. For neutrino energies above 10~TeV the angular resolution is determined by the intrinisic detector resolution i.e. the timing resolution and accuracy of the location of the PMTs. The energy measurement relies on an accurate calibration of the charge detected by each PMT \cite{baret}.
\subsection {Acoustic Positioning}
The positions of the PMTs are measured every two minutes with a high-frequency long-baseline acoustic positioning system comprising fixed acoustic emitters-receivers at the bottom of each line and acoustic receivers distributed along a line \cite{brown}. After triangulation of the positions of the moving hydrophones, the shape of each line is reconstructed by a global fit based on a model of the physical properties of the line and additional information from the tiltmeters and compass sensors located on each storey. The relative positions of the PMTs are deduced from this reconstructed line shape and the known geometry of a storey. The system provides a statistical precision of a few mm. The final precision on the PMT locations is a few cm, smaller than the physical extension of the PMTs, and is limited by the systematic uncertainties on the knowledge of the speed of sound in the sea water.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_02}
\caption{The horizontal movements relative to the bottom of the line,
of all hydrophones on Line 11 for a 6 month period; black points is the
hydrophone on Storey 1, red is on Storey 8, blue is on Storey 14, green
is on Storey 22 and magenta is on Storey 25.}
\label{fig2}
\end{figure}
Figure \ref{fig2} shows the movements of various storeys on a line, relative to its centre axis. The extent of the displacement
depends on the intensity of the sea current. For typical currents of a few centimetres per second, the displacement is a few metres for the topmost storeys.
\subsection {Time Calibration}
The relative time calibration of ANTARES is performed via a number of independent and redundant systems \cite{Gomez}.
The master clock system features a method to measure the transit time of the clock signals to the electronics located in each storey of the detector. The determination of the remaining residual time offsets within a storey, due to the delays in the front-end electronics and transit time of the PMTs, are based on the detection of signals from external optical beacons distributed throughout the detector. The presence of $^{40}$K in the sea water also provides a convenient source of calibrated light which is used to verify the time offsets between the triplet of PMTs within a storey as well as study the long term stablity of the PMTs efficiencies.
Every fifth storey of a line contains an optical beacon emitting in the blue. Each beacon illuminates the neighbouring storeys on its line. Comparison of the arrival hit times within a storey provides the relative inter-storey time offsets. Intra-storey time offsets can also be established after corrections are applied for time walk and 'first photon' effects.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_03}
\caption{Conicidence peak due to $^{40}$K decays for a single pair of photomultipliers in one storey.}
\label{fig3}
\end{figure}
Potassium-40 is a radioactive isotope naturally present in the sea water. The decay $^{40}K \rightarrow e^- \nu_e ^{40} Ca$ yields an electron with an energy up to $1.3$~MeV. This energy exceeds the Cherenkov threshold for electrons in water
(0.26~MeV), and is sufficient to produce up to 150 Cherenkov photons. If the decay occurs in the vincinity of a detector storey, a coincident signal may be recorded by pairs of PMTs on the storey. In Figure \ref{fig3} the distribution of the measured time difference between hits in neighbouring PMTs of the same storey is shown. The peak around 0~ns is mainly due to single $^{40}$K decays producing coincident signals. The fit of the data is the sum of a Gaussian distribution and a flat background. The full width at half maximum of the Gaussian function is about 9~ns. This width is mainly due to the spatial distribution of the $^{40}$K decays around the storey. The positions of the peaks of the time distributions for different
pairs of PMTs in the same storey are used to cross-check the time offsets computed by the optical beacon system. This is illustrated in Figure \ref{fig3} which shows a comparision of the time offsets calculated by the optical beacons and that extracted from the $^{40}$K analysis; an rms of 0.6~ns is obtained.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_04}
\caption{Time offsets for all photomultipliers as extracted using the LED beacons and independently
checked by the K40 conicidence method.}
\label{fig4}
\end{figure}
The rate of genuine $^{40}$K coincidences is given by the integral under the peak of Figure \ref{fig3} and is used to monitor the relative efficiencies of all PMTs and their temporal stability.
\section{Muon Reconstruction}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_05}
\caption{Example display of a neutrino candidate. The 2D plots, one for each line, show on the y axis the vertical position of the PMT with a hit and on the x axis the arrival time of the hit. The fit to the arrival time distribution corresponds to the chi-square algorithm.}
\label{fig5}
\end{figure}
Two alternative algorithms for reconstruction of the muon trajectories have been developed \cite{heijboer}.
In the first approach, a simple chi-square fit is applied to a high purity sample of pre-selected hits. In addition, this algorithm merges the hits observed by the PMTs of the triplet and assumes that they are located on the line axis, i.e. the azimuthal orientation of the storey, measured by the compasses, is ignored. In Figure \ref{fig5} an upgoing neutrino candidate fitted using this algorithm is shown. This algorithm was initially adopted as a fast reconstruction for online monitoring of the detector. Although it provides an non-optimum angular resolution (typically 1-2 degrees above 10~TeV) it has been used in a number of analyses for which the ultimate angular resolution is not crucial.
In the second approach a full maximum likelihood fit is applied, which uses a detailed parameterisation, derived from simulation, of the probability density function for the track residuals. The fit includes most hits in the event and the PDF takes into account the probability that photons arrive late due to Cherenkov emission by secondary particles or light scattering. A number of increasingly sophisticated prefits are used to aid in the location of the correct maxima of the likelihood. This algorithm makes use of the maximum amount of information, including the line shape and storey orientation, and provides
an angular resolution better than 0.3 degrees above 10~TeV.
\section{Atmospheric Muons}
The dominant signal observed by ANTARES is due to the passage of downgoing atmospheric muons, whose
flux exceeds that of neutrino-induced muons by several orders of magnitude. They are produced by high
energy cosmic rays interacting with atomic nuclei of the upper atmosphere, producing charged pions and
kaons, which subsequently decay into muons. Although an important background for neutrino detection, they are
useful to verify the detector response. In particular, with three years of data taking, a deficit in the muon flux in the direction of the moon should allow an important verification of the pointing accuracy of the detector.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_06}
\caption{Attenuation of the flux of muons as a function of depth, as extracted using the adjacent storey coincidence method. The shaded band represents the systematic uncertainties due to detector effects. Predictions from MUPAGE and Corsika Monte Carlo simulations are also shown.}
\label{fig6}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=3.1in]{icrc1319_07}
\caption{Vertical depth intensity relation of atmospheric muons with $E_\mu > 20$~GeV (black points). The
error band represents systematic uncertainties. A compliation of data from other experiments are also shown.}
\label{fig7}
\end{figure}
Two different studies of the vertical depth intensity relation of the muon flux have been performed. In the first, the attenuation of the muon flux as a function of depth is observed as a reduction in the rate of coincidences between adjacent storeys along the length of the detection lines \cite{zaborov}. This method has the advantage that it does not rely on any track reconstruction. In Figure \ref{fig6} the depth dependence of the extracted flux for the 24 inter-storey measurements averaged over all lines is shown.
In the second study, a full track reconstruction is performed and the reconstructed zenith angle is converted to an equivalent slant depth through the sea water \cite{bazotti}. Taking into account the known angular distribution of the incident muons, a depth intensity relation can be extracted (Figure \ref{fig7}). The results ae in reasonable agreement with previous measurements.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_08}
\caption{Identification of energetic electromagnetic showers: (left) Data and Monte Carlo comparison of the number of reconstructed electromagnetic showers for the 5-line data, assuming the nominal primary cosmic-ray composition. (right) dependence on primary cosmic-ray composition (proton versus iron).}
\label{fig8}
\end{figure}
The composition of the primary cosmic rays in the knee region is of particular interest. As the number of identified electromagnetic showers in an event depends on the muon energy and the number of muons present in a muon bundle, it is sensitive to the primary cosmic ray composition. An algorithm has been developed to estimate the number of energetic electromagnetic (EM) showers generated along the muon trajectory \cite{mangano1}. This algorithm relies on the fact that the emission point of Cherenkov photons from the muon are uniformly distributed along the muon trajectory whereas Cherenkov photons orginating from an electromagnetic shower will tend to cluster from a single point. The efficiency and purity of the algorithm to identify a shower depends on the shower energy, for example the efficiency to identify a 1~TeV shower is 20\% with a purity of 85\%. In Figure \ref{fig8} (left) the distribution of the number of reconstructed energetic showers per event in the 5-line data is shown. Good agreement with the Corsika Monte Carlo is obtained when the 22Horandel primary composition model is assumed. In Figure \ref{fig8} (right) the data distribution is compared with that obtained assuming a pure proton or a pure iron primary cosmic ray composition.
A search for a large-scale anistropy in the arrival directions of the atmospheric muons has been performed but with the current statistics is not yet sensitive to the $0.1\%$ level variations reported by other experiments \cite{mangano2}. The possibilty for detection of gamma ray induced air showers with ANTARES is also under evaluation \cite{guillard}.
\section{Search for Cosmic Neutrino Point Sources}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_09}
\caption{Zenith distribution of reconstructed muons in the 2008 data. The Mone Carlo expectation for the atmospheric muon and atmospheric neutrino backgrounds are indicated.}
\label{fig9}
\end{figure}
The muons produced by the interaction of neutrinos can be isolated from the muons generated by the cosmic ray interactions by requiring that the muon trajectory is reconstructed as up-going. In Figure \ref{fig9} the zenith angular distribution of muons in the 2007+2008 data sample by the $\chi^2$ reconstruction algorithm is shown. A total of 750 mulitline up-going neutrinos candidates are found, in good agreement with expectations from the atmospheric neutrino background.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_10}
\caption{Sky map, in geocentric coordinates, of the upgoing neutrino candidates for the 2007 data.}
\label{fig10}
\end{figure}
\begin{table}
\begin{center}
{\footnotesize
\begin{tabular}{|l|r@{.}l|r@{.}l|c|c|c|} \hline
Source name & \multicolumn{2}{c|}{$\delta$ ($^{\circ}$)} & \multicolumn{2}{c|}{RA ($^{\circ}$)} & n$_{{\rm bin.}}$ & p-value & $\phi_{90}$ \\ \hline \hline
PSR B1259-63 & -63&83 & 195&70 & 0 & - & 3.1 \\
RCW 86 & -62&48 & 220&68 & 0 & - & 3.3 \\
HESS J1023-575 & -57&76 & 155&83 & 1 & 0.004 & 7.6 \\
CIR X-1 & -57&17 & 230&17 & 0 & - & 3.3 \\
HESS J1614-518 & -51&82 & 243&58 & 1 & 0.088 & 5.6 \\
GX 339 & -48&79 & 255&70 & 0 & - & 3.8 \\
RX J0852.0-4622 & -46&37 & 133&00 & 0 & - & 4.0 \\
RX J1713.7-3946 & -39&75 & 258&25 & 0 & - & 4.3 \\
Galactic Centre & -29&01 & 266&42 & 1 & 0.055 & 6.8 \\
W28 & -23&34 & 270&43 & 0 & - & 4.8 \\
LS 5039 & -14&83 & 276&56 & 0 & - & 5.0 \\
HESS J1837-069 & -6&95 & 279&41 & 0 & - & 5.9 \\
SS 433 & 4&98 & 287&96 & 0 & - & 7.3 \\
HESS J0632+057 & 5&81 & 98&24 & 0 & - & 7.4 \\ \hline
ESO 139-G12 & -59&94 & 264&41 & 0 & - & 3.4 \\
PKS 2005-489 & -48&82 & 302&37 & 0 & - & 3.7 \\
Centaurus A & -43&02 & 201&36 & 0 & - & 3.9 \\
PKS 0548-322 & -32&27 & 87&67 & 0 & - & 4.3 \\
H 2356-309 & -30&63 & 359&78 & 0 & - & 4.2 \\
PKS 2155-304 & -30&22 & 329&72 & 0 & - & 4.2 \\
1ES 1101-232 & -23&49 & 165&91 & 0 & - & 4.6 \\
1ES 0347-121 & -11&99 & 57&35 & 0 & - & 5.0 \\
3C 279 & -5&79 & 194&05 & 1 & 0.030& 9.2 \\
RGB J0152+017 & 1&79 & 28&17 & 0 & - & 7.0 \\ \hline
IC22 hotspot & 11&4 & 153&4 & 0 & - & 9.1 \\ \hline
\end{tabular}
}
\caption{\small Results of the search for cosmic neutrinos correlated
with potential neutrino sources. The sources are divided into three
groups: galactic (top), extra-galactic (middle) and the hotspot from
IceCube with 22 lines (bottom). The source name and location in
equatorial coordinates are shown together with the number of events
within the optimum cone for the binned search, the p-value of the
unbinned method (when different from 1) and the corresponding upper
limit at 90\% C.L. $\phi_{90}$ is the value of the normalization
constant of the differential muon-neutrino flux assuming an $E^{-2}$
spectrum (i.e. $E^{2} d\phi_{\nu_{\mu}} / dE \le \phi_{90} \times
10^{-10} $ TeV cm$^{-2}$ s$^{-1}$). The integration energy range is 10
GeV - 1 PeV.}
\label{tab:sources}
\end{center}
\end{table}
For a subset of this data (the 5-line data of 2007) the angular resolution has been improved by applying the pdf based track reconstruction, which makes full use of the final detector alignment. After reoptimisation of the selection cuts for the best upper limits, 94 upgoing neutrino candidates are selected \cite{toscano}. The corresponding sky map for these events is shown in Figure \ref{fig10}. An all sky search, independent of assumption on the source location, has been performed on these data. The most significant cluster was found at ($\delta=-63.7^\circ, RA=243.9^\circ$) with a corresponding p-value of 0.3 ($1\sigma$ excess) and is therefore not significant.
A search amongst a pre-defined list of 24 of the most promising galactic and extra-galactic neutrino sources (supernova remnants, BL Lac objects, Icecube hot spot, etc.) is reported in Table \ref{tab:sources}. The lowest p-value obtained (a $2.8 \sigma$ excess, pre-trial) corresponds to the location ($\delta=-57.76^\circ, RA=155.8^\circ$). Such a probablity or higher would be expected in $10\%$ of background-only experiments and is therefore not significant. The corresponding flux upper limits, assuming an $E^{-2}$ flux, are plotted in Figure \ref{fig11} and compared to published upper limits from other experiments. Also shown in Figure \ref{fig11} is the predicted upper limit for ANTARES after one full year of twelve line operation.
\begin{figure}[!t]
\centering
\includegraphics[width=3.1in]{icrc1319_11}
\caption{Neutrino flux upper limits at $90\%$ C.L. obtained by this analysis (solid squares), compared with published results from other experiments (IceCube [24], AMANDA [25], SuperKamiokande [26] and MACRO [27]). The expected sensitivity of ANTARES for one year with twelve lines is also shown (solid line). The source spectrum assumed in these results is $E^{−2}$, except for MACRO, for which an $E^{−2.1}$ spectrum was used.}
\label{fig11}
\end{figure}
\section{Search for Dark Matter}
In many theoretical models a Weakly Interacting Massive Particle (WIMP), a relic from the Big Bang, is proposed to explain the formation of structure in the universe and the discrepancy observed between the measured rotation curves of stars and the associated visible matter distribution in galaxies. A generic property of such WIMPs is that they gravitationally accumulate at the centre of massive bodies such as the Sun or the Earth, where they can self annihilate into normal matter. Only neutrinos, resulting from the decay of this matter, can escape from the body and be detected by a neutrino telescope.
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_12}
\caption{Upper limit on the muon flux from the Sun as a function of neutralino mass. The expected fluxes for a scan of mSUGRA parameters is shown as well as a variety of limits from other experiments.}
\label{fig12}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_13}
\caption{The spin dependent WIMP-proton cross-section versus neutralino mass in the mSUGRA model. The points
are the results of the scan over the range of model parameters. The points in blue would be excluded at 90\% CL after three years of ANTARES operation. Existing upper limits from a variety of direct and indirect direction experiments are shown. }
\label{fig13}
\end{figure}
Within Supersymmetric models with R-parity conservation, the lightest supersymmetric particle (LSP) is stable and is the WIMP candidate. In order to predict the expected solar neutrino fluxes the constrained phenomenological framework of the minimal Supergravity model (mSUGRA, computations using ISASUGRA[5]) has been adopted. Figure \ref{fig10} shows the predicted integrated neutrino fluxes above 10~GeV in ANTARES as a function of neutralino mass for the scan of the model parameters: scalar mass $m_0$ in [0,8000] GeV, gaugino mass $m_{1/2}$ in [0,2000] GeV, tri-linear scalar coupling $A_0$ in [$-3m_0,3m_0$], sign of the Higgsino mixing parameter: $\mu > 0$, ratio of Higgs fields vacuum expectation values $tan \beta$ in [0,60], $m_{top}=172.5~GeV$. The local Dark Matter halo density (NFW-model) was set to $0.3~GeV/cm^3$.
The most favourable models for neutrino telescopes are in the so-called ‘focus point’ region, for which the decays are mainly via $W^+W^-$ leading to a harder neutrino spectra. Thanks to its low-energy threshold ANTARES is ideally suited to address low-mass neutralino scenarios.
A search for neutrinos from the direction of the Sun in the 5-line data \cite{lim}, showed no excess with respect to background expectations. The corresponding derived limit on the neutrino flux is shown in Figure \ref{fig12}. Also shown is the expected limit with 5 years of data taking with the full 12-line detector; a large fraction of the focus point region could be excluded.
Due to the capture of the WIMPs inside the Sun, which is mainly hydrogen, neutrino telescopes are particularly sensitive to the spin-dependent coupling of the WIMPs to standard matter. In Figure \ref{fig13} the corresponding limit on the spin-dependent WIMP-proton cross-section after three years of ANTARES operation is shown. For this case, the neutrino telescopes are significantly more sensitive than the direct direction experiments.
\section{Multi-Messenger Astronomy}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_14}
\caption{Common sky coverage for VIRGO/LIGO/ANTARES in geocentric coordinates. This map shows the combined antenna pattern for the gravitational wave detector network (above half-minimum), assuming that ANTARES has 100\% visibility in its antipodal hemisphere and 0\% elsewhere.}
\label{fig14}
\end{figure}
In order to augment the discovery potential of ANTARES, a program of collaboration with other types of observatory have been established. In this ''multi-messenger'' approach the detection threshold can be lowered in each separate experiment while preserving an acceptable rate of accidental coincidences. One example of such a program is being pursued with the gravitational wave detectors VIRGO and LIGO \cite{elewyck}. Both of these detectors had a data-taking phase during 2007 (VIRGO science run 1 and LIGO S5) which partially coincided with the ANTARES 5-line configuration. A new common science run has also recently started in coincidence with the ANTARES 12-line operation. The common sky coverage for ANTARES-VIRGO+LIGO is signifcant and is shown in Figure \ref{fig14}.
In a similar vein, a collaboration with the TAROT optical telescopes has been established \cite{dornic}. The directions of interesting neutrino triggers (two neutrinos within 3 degrees within a time window of 15 minutes or a single neutrino of very high energy) are sent to the Chile telescope in order that a series of optical follow up images can be taken. This procedure is well suited to maximise the sensitivity for transient events such as gamma ray bursters or flaring sources.
\section {Acoustic Detection}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_15}
\caption{Power spectral density (PSD) of the ambient noise recorded with one sensor. Shown in grey shades of grey is the occurence rate in arbirary units, where dark colours indicate higher occurence. Shown as a white dotted line is the mean value of the in-situ PSD and as a black solid line the noise level recorded in the laboratory before deployment.}
\label{fig15}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.7in]{icrc1319_16}
\caption{Map of the angular directions of the detected transient acoustic signals at the ANTARES site. }
\label{fig16}
\end{figure}
Due to the large attenuation length, $\approx$ 5~km for 10~kHz signals, the detection of bipolar acoustic pressure pulses in huge underwater acoustic arrays is a possible approach for the identification of cosmic neutrinos with energies exceeding 100~PeV. To this end, the ANTARES infrastructure incorporates the AMADEUS system, an R\&D project intended to evaluate the acoustical backgrounds in the deep sea \cite{simeone}. It comprises a set of piezo-electric sensors for broad-band recording of acoustic signals with frequencies ranging up to 125~kHz with typical sensitivities around 1V/$\mu$Pa. The sensors are distributed in six ''acoustic clusters'', each comprising six acoustic sensors that are arranged at distances roughly 1~m from each other. The clusters are installed along the line 1é and the IL07 line at a horizontal distance of 240~m. The vertical spacing within a line range from 15~m to 125~m (see Figure \ref{fig1}).
In Figure \ref{fig15} is shown the measured power spectral density of the ambient noise recorded with an acoustic sensor. Due to wind generated surface noise, the observed noise level is larger than that measured in the laboratory. Strong correlation of the measured acoustic noise with the measured surface wind speed are observed.
In Figure \ref{fig16} an acoustic "sea map" of all transient (not just bipolar) signals detected by the apparatus during a one month period is shown. The acoustic pinger of the ANTARES positioning system are clearly identified in the lower hemisphere. The acoustic activity in the upper hemisphere is presumably due to surface boats and possibly marine mammals.
\section{Conclusion}
After more than a decade of R\&D and prototyping the construction of ANTARES, the first operating deep-sea neutrino telescope has been completed. Since the deployment of the first line in 2006, data taking has proceeded essentially continuously. During this time the methods and procedures to calibrate such a novel detector have been developed, including in-situ time calibration with optical beacons and acoustic positioning of the detector elements with acoustic hydrophones. The presence of the $^{40}K$ in the sea water has proven particularly useful for monitoring the stability of the time calibration as well as the detector efficiency.
Based on data from the intermediate 5-line configuration, a number of preliminary analyses have been presented; measurements of the atmospheric muon vertical depth intensity relation, a search for cosmic neutrino point sources in the southern sky, and a search for dark matter annihilation in the Sun. For the latter two analyses no significant signal was observed and competitive upper limits have been obtained.
The succesful operation of ANTARES, and analysis of its data, is an important step towards KM3NET \cite{km3net}, a future km$^3$-scale high-energy neutrino observatory and marine sciences infrastructure planned for construction in the Mediterranean Sea.
\newpage
|
\section{Introduction}
Most grand design spiral patterns are believed to be due to
interactions with other galaxies \citep{Toomre1972}, or are driven by
bars \citep[][Athanassoula et al., this volume]{Kormendy1979,
Bottema2003}. Here we investigate the first of these scenarios, by
performing numerical simulations of interacting galaxies. In this
instance, there is not necessarily a singular pattern speed for the
spiral arms, as indicated by recent observations of M51
\citep{Meidt2008a}.
\section{Method}
We use the Smoothed Particle Hydrodynamics Code (SPH) to model a
galaxy subject to an interaction. The first galaxy is modeled by
assuming a spherical potential for the halo, with 3 different
scenarios for the galactic disc: a) a live stellar disc, no gas; b) a
logarithmic stellar potential and a live gaseous disc and c) a disc
containing live gaseous and stellar components. The stars and/or gas
are allocated velocities such that the Toomre instability parameter,
Q, is globally 2, but in addition, the galaxy was allowed to evolve in
isolation until any flocculent structure disappears. In all
simulations 1 million particles are used. For the case of stars and
gas there are 500,000 gas particles and 500,000 stellar particles, and
in all cases the total mass of the disc is $5\times 10^9$ M$_{\odot}$.
We adopt a similar approach to \citet{Oh2008} to model the
interaction. The interacting galaxy is represented by a sink particle
\citep{Bate1995} and is of relatively low mass, equal to the mass of
the disc or 2.5\% of the total mass of the first galaxy. The
interacting galaxy takes a parabolic orbit, reaching a closest
approach of 25 kpc after a time of 370 Myr. Initially this galaxy is
at a distance of 50 kpc.
\section{Results}
\subsection{Stellar disc}
\begin{figure}[]
\centering
\resizebox{\hsize}{!}{\includegraphics[clip=true]{dobbs_f01.ps}}
\caption{\footnotesize
The column density is shown for a stellar disc of a galaxy perturbed
by a (point mass) galaxy on a parabolic orbit. The second galaxy is
located 70 kpc away, at the coordinates $(-20,68)$ kpc. The stellar
disc contains broad and diffuse spiral arms.}
\end{figure}
\begin{figure}[]
\centering
\resizebox{\hsize}{!}{\includegraphics[clip=true]{dobbs_f02.ps}}
\caption{\footnotesize
The pattern speeds for each spiral arm are plotted versus radius for
the stellar disc. $\Omega_{\rm p}$ is different for each arm and decreases
with radius. The lower line shows a $1/r$ dependence.}
\end{figure}
\begin{figure}[]
\centering
\resizebox{\hsize}{!}{\includegraphics[clip=true]{dobbs_f03.ps}}
\caption{\footnotesize
The column density is shown for the gas disc perturbed by a galaxy,
at the same time as Fig.~1. The gas shocks produce very narrow,
dense spiral arms compared to Fig.~1.}
\end{figure}
\begin{figure}[]
\centering
\resizebox{\hsize}{!}{\includegraphics[clip=true]{dobbs_f04.ps}}
\caption{\footnotesize
The pattern speeds are plotted for the spiral arms of the gas
disc. $\Omega_p$ is much lower than typical from observations.}
\end{figure}
Fig.~1 shows the column density of a stellar disc, at a time of 820
Myr. The interacting galaxy is 70 kpc from the centre of the plot. The
spiral arms are relatively weak and broad.
We calculate the pattern speed of the spiral arms according to
\begin{equation}
\Omega_{\rm p}=\frac{\phi(\rho_{\rm max})|_{t_2}-\phi(\rho_{\rm max})|_{t_1}}{t_2-t_1}
\end{equation}
where $\rho_{\rm max}$ is the peak density of a particular spiral arm
at a given radius, and $t_1$ and $t_2$ are times during the
simulation. First we select points covering a particular spiral arm at
time $t_1=800$ Myr, when a strong spiral pattern has emerged. Then the
azimuthal angle of the spiral arm is located for different radii. This
process is repeated at time $t_2=925$ Myr, to obtain the change in
azimuthal angle of the spiral arm at each radius, and thus
$\Omega_{\rm p}$. Given there are 2 spiral arms, this method leads to
a pattern speed for each spiral arm. These pattern speeds are shown
versus radius in Fig.~2, the errors reflecting the uncertainty in
locating the peak density of the spiral arm.
The magnitude of the pattern speeds are not dissimilar from some of
those measured for spirals (e.g., \citealt{Clemens2001, Grosbol2004}),
although the spiral arms clearly exhibit different pattern
speeds. This difference is a consequence of the asymmetry of the
system, i.e. that the interaction induces one arm on one side of the
disc first. The pattern speeds for each arm also decrease with radius,
roughly as expected for spiral patterns induced by interactions.
\subsection{Gaseous disc}
We also performed calculations with just gas. The gas constitutes 1\%
of the mass of the galactic disc. This value is unrealistically small,
but the low gas mass is chosen to avoid gravitational instabilities,
which would halt the calculation. Essentially, we are only
investigating the reaction of the gas to the interaction, not the self
gravity of the gas. These calculations are also not a particularly
realistic case as they ignore the perturbation experienced by the
stellar disc from the interaction (instead the stellar disc is
represented by a symmetric potential). However the case with just gas
is explored for completeness.
In Fig.~3, we show the disc when only the gas is included, at the same
time (800 Myr) as Fig.~1. The spiral arms are clearly much narrower,
more prominent, and more dense than for the stellar disc. The
evolution of the gas and stellar discs is also different. The gaseous
spiral arms rotate much slower than the stellar arms. Consequently the
pattern speed is very low, $\sim$4 km s$^{-1}$ kpc$^{-1}$ and does not
show much variation with radius (Fig.~4). The difference compared to
the stellar disc is that one spiral arm is still linked to, and
rotates at the same angular velocity as, the orbiting galaxy.
\begin{figure}[!h]
\centering
\includegraphics[width=.45\textwidth,clip=true]{dobbs_f05a.ps} \\
\includegraphics[width=.45\textwidth,clip=true]{dobbs_f05b.ps} \\
\includegraphics[width=.45\textwidth,clip=true]{dobbs_f05c.ps}
\caption{\footnotesize
The gas distribution is shown from a calculation with gas and stars
at times of 500 (top), 740 (middle) and 1060 Myr (bottom). The
spiral arms are very asymmetric at the earliest time (when the
position of the interacting galaxy is located at $(22.5, 15)$ kpc),
but becomes more symmetric at later times, after the interaction.}
\end{figure}
\begin{figure}[]
\centering
\resizebox{\hsize}{!}{\includegraphics[clip=true]{dobbs_f06.ps}}
\caption{\footnotesize
The pattern speed is shown for the gaseous arms, from the simulation
with stars and gas. The gas is largely coupled to the stars, hence
$\Omega_{\rm p}$ is higher than when only gas particles are
present. Again there is a clear difference in the pattern speeds of
each spiral arm.}
\label{li_vhel}
\end{figure}
\subsection{Evolving both the stars and gas}
Finally we show calculations with gas and stars, again where the gas
represents 1\% of the mass of the stellar disc. Thus the mass of the
stellar disc is $5\times 10^9$ M$_{\odot}$ and the gaseous disc $5
\times 10^8$ M$_{\odot}$, although the actual number of gas and
stellar particles in the calculation are equal. Fig.~5 shows the
column density of gas from the calculation with live stellar and
gaseous components. The morphology evidently changes over time. At an
earlier stage in the interaction, the spiral pattern is very
asymmetric (unlike the calculations with gas alone, where the spiral
pattern is symmetric throughout the simulation). The pattern becomes
more symmetric at later times, tending towards the distribution
without a live stellar disc (Fig.~3). Although not shown on Fig.~5,
the stellar distribution reflects the gas distribution, but the spiral
arms are much weaker and broader.
Finally in Fig.~6 we plot the pattern speed of the gas from the
calculation with both live stellar and gaseous components. The times
selected to calculate the pattern speed are 450 and 500 Myr. The
spiral arms clearly have very different pattern speeds, which is not
surprising given the asymmetry of the disc. The spiral arm with the
lower pattern speed is still associated with the interacting
galaxy. The pattern speeds are more similar in magnitude to the case
when only stars are used (Fig.~2), since the gas is better coupled to
the stars.
\section{Conclusions}
We have performed calculations with a stellar, gaseous and both
stellar and gaseous disc subject to an interaction with an orbiting
galaxy. The pattern speed across the disc is generally not constant,
and pattern speeds in each arm differ. For a stellar disc,
$\Omega_{\rm p}=5-20$ km s$^{-1}$ kpc$^{-1}$, decreasing with radius
approximately as $1/r$. With only gas, the pattern speeds are much
lower ($3-6$ km s$^{-1}$ kpc$^{-1}$). When stars and gas are included,
the gas tends to follow the stellar distribution, thus the pattern
speeds of the gaseous spiral arms are higher ($5-17$ km s$^{-1}$
kpc$^{-1}$).
These calculations may be improved by using a more consistent initial
galaxy set up (e.g., \citealt{Dubinski1995}). A natural extension of
this work would also be to compare with observations by applying the
Tremaine-Weinberg \citep{Tremaine1984} method to these calculations
(see also \citealt{Meidt2008b}).
\begin{acknowledgements}
The calculations presented in these proceedings were performed using
the University of Exeter's SGI Altix ICE 8200 supercomputer, Zen.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
To face the present and forthcoming global energy crisis, human
should search for clean and effective energy source. Recently the
investigations on the basic energy science for this purpose has
received great attention and
experienced impressive progress based on the fundamental physics \cit
{Fleming08,arti}. In photosynthetic process, the structural elegance
and chemical high efficiency of the natural system based on pigment
molecules in transferring the energy of sunlight have stimulated a
purpose driven investigation \cite
{Venturi08,HuXiChe971,Johnson08,Plenio091,Nori09,Guzik09,Nori08,Fleming09,Plenio092,Guzik08,Mukamel09},
finding artificial analogs of porphyrin-based chromophores. These
artificial systems replicate the natural process of photosynthesis
\cite{arti} so that the much higher efficiencies could be gained
than that obtained in the conventional solid systems \cite{arti}. It
is because one of the most attractive features of photosynthesis is
that the light energy can be captured and transported to the
reaction center (RC) within about 100ps and with more than 95\%
efficiency \cite{HuXiChe971,Fleming94}.
Actually, in most of the plants and bacterium, the primary processes
of photosynthesis are almost in common
\cite{Fleming94,HuXiChe972,Venturi08}: Light is harvested by antenna
proteins containing many chromophores; then the electronic
excitations are transferred to the RC sequentially, where
photochemical reactions take place to convert the excitation energy
into chemical energy. Most recent experiments have been able to
exactly determine the time scales of various transfer processes by
the ultra-fast laser technology \cite{exp1,exp2,exp3}. These great
progresses obviously offer us a chance to quantitatively make clear
the underlying physical mechanism of the photosynthesis, so that
people can construct the artificial photosynthesis devices in the
future to reach the photon-energy and photon-electricity conversions
with higher efficiency. For example, quantum interference effects in
energy transfer dynamics \cite{Guzik08} has been studied for the
Fenna-Matthews-Olson (FMO) protein complex, and it was found
\cite{Guzik09} that, for such molecular arrays, the spatial
correlations in the phonon bath and its induced decoherence could
affect on the efficiency of the primary photosynthetic event. The
present paper will similarly study the influences of spatial
structure on the primary processes of photosynthesis for the
light-harvesting complexes II (LH2).
In the past, by making use of the x-ray crystallographic techniques, the
structure of light-harvesting system has been elucidated \cit
{HuXiChe96,Venturi08}. In the purple photosynthetic bacteria, there exist
roughly two types of light-harvesting complexes, referred to as
light-harvesting complex I (LH1) and light-harvesting complex II (LH2). In
LH1, the RC is surrounded by a B875 bacteriochlorophyll (BChl) ring with
maximum absorption peak at 875 nm. The LH2 complex, however, does not
contain the RC, but can transfer energy excitation to the RC indirectly
through LH1. In the purple bacteria, LH2 is a ring-shaped aggregate built up
by $8$ (or $9$) minimal units, where each unit consists of an $\alpha \beta
-heterodimer, three BChls, and one carotenoid. The $\alpha \beta
-heterodimers, i.e., $\alpha $-apoproteins and $\beta $-apoproteins
constitute the skeleton of LH2, while the BChls are embedded in the scaffold
to form a double-layered ring structure. The top ring including $16$ (or $18
) BChl molecules is named as B850 since it has the lowest-energy
absorption maximum at 850 nm. The bottom ring with 8 BChls is called
B800 because it mainly absorbs light at 800 nm. In every minimal
unit, the carotenoid connects B800 BChl with one of the two B850
BChls. Excitation is transferred
from one pigment to the neighbor one through the F\"{o}ster mechanism \cit
{HuXiChe971}, while the electron is spatially transferred via the
Marcus mechanism \cite{Leegwater96}. Generally, it is independent of
the global geometry configuration of the system.
In the present paper, we will study the energy transfer procedure in LH2 by
considering the structure dimerization of the B850 ring. It has been
conjectured that the dimerized inter-pigment couplings can cause the energy
gap to protect the collective excitations \cite{HuXiChe972}. Indeed, like
the the Su-Schrieffer-Heeger model for the flexible polyacetylene chain \cit
{SSH}, the dimerization of the spatial configuration with the Peierls
distorted ground state will minimize the total energy for the phonon plus
electron. As it is well known, this model exhibits a rich variety of
nonlinear phenomena and topological excitations including the topological
protection of the quantum state transfer \cite{Song}. Similarly, we will
show that, when the B850 ring in LH2 is dimerized the excitation transfer
efficiency may be enhanced to some extent.
Based on the open quantum system theory, we simply model the excited
and un-excited states of a BChl pigment as a quasispin. The
excitation transfer is represented by the total leakage from a LH2.
Using the master equation, we calculate the efficiency of excitation
transfer and the average transfer time in low temperature for
various initial states with different superposition properties. The
results explicitly indicate that the dimerization of couplings
indeed enhances the quantum transport efficiency and shortens the
average transfer time.
\begin{figure}[ptb]
\includegraphics[bb=49 366 550 770, width=8 cm, clip]{fig1.eps}
\caption{(color online) The model setup of the light-harvesting
complex II constructed by 8 unit cells. The couplings between the
neighboring
quasi-spins in the B850 ring is dimerized as $J_{2}(1+\protect\delta)$ and
J_{2}(1-\protect\delta)$. $g_{1}$ denote the nearest couplings
between B850 BChls and B800 BChls, while $g_{2}$ denote the next
nearest couplings between B850 BChls and B800 BChls. (a)
Illustration of the whole system with $g_{2}=0 $. (b) Detailed
drawing of three unit cells and their non-local couplings. (c)
Legends.} \label{model1}
\end{figure}
This paper is organized as follows. In Sec. II, a double-ring XY model with
N$ unit cells is presented to simulate the LH2 system. In Sec. III, the
energy transfer process is described by the quantum master equation. The
transfer efficiency $\eta \left( t\right) $ and the average transfer time
\tau $ are introduced to characterize the dynamics of the system. In Sec.
IV, we represent the master equation in the momentum space and show that
only the $\left( k,k\right) $-blocks of the density matrix are relevant to
energy transfer. In Sec. V, it is found that the transfer efficiency $\eta
\left( t\right) $ and the average transfer time $\tau $ of an arbitrary
initial state can be obtained through the channel decomposition. Some
numerical analysis of $\eta ^{\left[ A,k\right] }\left( t_{0}\right) $ and
\tau ^{\left[ A,k\right] }$ for all the $k$-channels are presented in Sec.
VI. They show that a suitable dimerization of the B850 BChl ring can enhance
the transfer efficiency and shorten the average transfer time. Conclusions
are summarized at the end of the paper. In Appendix A, we provide an
alternative way to deal with the energy leakage problem. In Appendix B,
detail derivations of transforming the master equation from the real space
to the $k$-space are given. The approximate solution of $\tau ^{\left[ A,
\right] }$ for $k=0$ and $k=\pm \pi $ channel is shown in Appendix C.
\section{Model setup}
The simplified model of LH2 is\ shown in Fig. \ref{model1}. All the
bacteriochlorophylls (big and small green squares) are modeled by the
two-level systems with excited state $|e_{j}^{[c]}\rangle $, ground state
|g_{j}^{[c]}\rangle $,\ and energy level spacing $\Omega _{c}$. The raising
and lowering quasi-spin operators of the $j$th two-level system on the $[c]
\ ring is expressed as
\begin{equation}
\sigma _{j}^{+[c]}=|e_{j}^{[c]}\rangle \langle g_{j}^{[c]}|\text{, }\sigma
_{j}^{-\left[ c\right] }=|g_{j}^{[c]}\rangle \langle e_{j}^{[c]}|,
\end{equation
where $[c]=$\ $\left[ a\right] $ ($\left[ b\right] $) denotes the B800
(B850) BChl ring. Approximately, all the couplings are supposed to be of XY
type \cite{Johnson08}. This simplification enjoys the main feature of
excitation transfer. \ The Hamiltonians
\begin{equation}
H_{a}=\frac{\Omega _{a}}{2} \sum_{j=1}^{N}\sigma _{j}^{z\left[
a\right] }+J_{1}\sum_{j=1}^{N}\left( \sigma _{j}^{+\left[ a\right]
}\sigma _{j+1}^{-[a]}+\mathrm{H.c.}\right) \label{h1}
\end{equation
and
\begin{align}
H_{b}& =\frac{\Omega _{b}}{2} \sum_{j=1}^{2N}\sigma _{j}^{z\left[
b\right] }+J_{2}\sum_{j=1}^{N}[\left( 1+\delta \right) \sigma _{2j-1}^{
\left[ b\right] }\sigma _{2j}^{-\left[ b\right] } \notag \\
& +\left( 1-\delta \right) \sigma _{2j}^{+\left[ b\right] }\sigma _{2j+1}^{
\left[ b\right] }+\mathrm{H.c.}] \label{h2}
\end{align
with $N=8$,\ describe the excitations of the B800 and\ B850 BChl rings,\
respectively.\ In the B850 BChl ring, the parameter $\delta \neq 0$
characterizes the dimerization due to the spatial deformation of the
flexible B850 BChl ring in LH2. The coupling constants of $H_{b}$ are
dimerized as $J_{2}\left( 1+\delta \right) $ and $J_{2}\left( 1-\delta
\right) $ since the intra-unit and inter-unit Mg-Mg distance between
neighboring B850 BChls may be different. The non-local XY type interaction
\begin{align}
H_{ab}& =g_{1}\sum_{j=1}^{N}\left[ \sigma _{j}^{+\left[ a\right] }\left(
\sigma _{2j-1}^{-\left[ b\right] }+\sigma _{2j}^{-\left[ b\right] }\right)
\mathrm{H.c.}\right] \notag \\
& +g_{2}\sum_{j=1}^{N}\left[ \sigma _{j}^{+\left[ a\right] }\left( \sigma
_{2j-3}^{-\left[ b\right] }+\sigma _{2j-2}^{-\left[ b\right] }+\sigma
_{2j+1}^{-\left[ b\right] }+\sigma _{2j+2}^{-\left[ b\right] }\right)
\mathrm{H.c.}\right] \label{Hspin}
\end{align
is used to describe the interaction between the B800 and B850 BChl rings.
In the single excitation case, the quasi-spin can be represented with a
spinless fermion with the mappin
\begin{equation}
\sigma _{j}^{+\left[ a\right] }\leftrightarrow A_{j}^{\dag },\sigma
_{2j-1}^{+\left[ b\right] }\leftrightarrow B_{j}^{\dag },\sigma _{2j}^{
\left[ b\right] }\leftrightarrow C_{j}^{\dag }
\end{equation
from the spin space $V_{s}=C_{2}^{\otimes 3N}$ to the subspace $V_{F}$ of\
the Fermion Fock space spanned by
\begin{equation}
\left\{ \left\vert O,j\right\rangle =O_{j}^{\dag }\left\vert 0\right\rangle
\text{ }|\text{ }O=A,B,C;j=1,2,\cdots ,N\right\} .
\end{equation
Hereafter, let us represent the site index as $\left( O,j\right) $, where $j$
refers to a unit cell shown in Fig. \ref{model1}, and $O=A,B,C$ to a
position type inside the unit cell. In the subscripts, the site index
\left( O,j\right) $ is written as $Oj$ for simplicity. The vacuum state of
the Fermion system $\left\vert 0\right\rangle $ corresponds to the state
that all the quasi-spins are in their ground states,
\begin{equation}
\left\vert 0\right\rangle \leftrightarrow \prod_{j=1}^{N}\left\vert g_{j}^{
\left[ a\right] }\right\rangle \otimes \prod_{j=1}^{2N}\left\vert g_{j}^
\left[ b\right] }\right\rangle .
\end{equation
Then the total Hamiltonian $H_{S}=H_{a}+H_{b}+H_{ab}$ of LH2 is mapped into
\begin{align}
H_{S}& =\sum_{j=1}^{N}\left[ \Omega _{a}A_{j}^{\dag }A_{j}+\Omega _{b}\left(
B_{j}^{\dag }B_{j}+C_{j}^{\dag }C_{j}\right) \right] \notag \\
& +\sum_{j=1}^{N}\left\{ J_{1}A_{j}^{\dag }A_{j+1}+g_{1}A_{j}^{\dag }\left(
B_{j}+C_{j}\right) \right. \notag \\
& +J_{2}[\left( 1+\delta \right) B_{j}^{\dag }C_{j}+\left( 1-\delta \right)
C_{j}^{\dag }B_{j+1}] \notag \\
& \left. +g_{2}A_{j}^{\dag }\left( B_{j+1}+B_{j-1}+C_{j+1}+C_{j-1}\right)
\mathrm{H.c.}\right\} . \label{Hparticle}
\end{align
In the present work, no multi-fermion interactions are considered for
simplicity.
On the other hand, we use the Holstein-Primakoff transformation \cit
{HPtransform} to map the quasi-spin into bosons. The excitations of
the BChls can be described by quasi-spins with the total angular
momentum $S$. Then $D=A,B,C$ can be regarded as the annihilation
operators
of bosons for the Fock space spanned b
\begin{eqnarray}
\{\left( D_{j}^{\dag }\right) ^{n_{D,j}}\left\vert 0\right\rangle \text{ }
\text{ }D &=&A,B,C;j=1,\cdots ,N; \notag \\
n_{D,j} &=&0,1,\cdots ,2S\}.
\end{eqnarray
For $S>1/2$, one local bacteriochlorophyll has more than one excited states.
In this case, higher order coherence could be included for further
generalization.
In the following, we focus on the single excitation case. Then the
temperature should be suitable to ensure there is no higher order
excited state.
\section{Transfer efficiency and average transfer time via the master
equation}
Next we consider the energy transfer from an initial state
\begin{equation}
\widehat{\rho }\left( 0\right) =\sum_{j,l}\rho _{Aj,Al}\left( 0\right)
\left\vert A,j\right\rangle \left\langle A,l\right\vert ,
\end{equation
which is a coherent superposition or a mixture of those local states
\left\vert A,j\right\rangle $ on the B800 ring. As time goes by, the initial
state will evolves a state distributing around both the B800 and the B850
rings. Since there exists a difference of chemical potential. $\Delta \Omega
=\Omega _{a}-\Omega _{b}$, energy is transferred between the two rings
during the time evolution. For an isolated LH2 system, such energy transfer
is coherent, namely, the system oscillates between the B800 and the B850
rings . However, when a LH2 is coupled to a heat reservoir with infinite
degrees of freedom, irreversible energy transfer occurs. As illustrated in
Fig. \ref{model2}, in the real photosynthetic system, energy is transferred
from one LH2 to another LH2 or LH1 through the B850 ring \cite{HuXiChe972}.
Therefore, we regard the first excited LH2 as an open system, and the sum of
others as the a heat reservoir. The energy transfer now can be manipulated
as the energy leakage from the B850 ring to the environment.
\begin{figure}[tbp]
\includegraphics[bb=27 374 563 777, width=8 cm, clip]{fig2.eps}
\caption{(color online) The first excited LH2 is treated as an open system
while the other LHs are regarded as heat reservoirs. The energy transfer
process is equivalent to the excitation leakage from the B850 BChl ring of
the LH2 system to the environment.}
\label{model2}
\end{figure}
In order to describe such a procedure that the excitations are finally
transferred from the B850 ring to the heat reservoir, the Markovian master
equation
\begin{equation}
\frac{d\widehat{\rho }}{dt}=-i\left[ H_{S},\widehat{\rho }\right] +{\mathcal
L}}\left( \widehat{\rho }\right) \label{mastereq}
\end{equation
in the Lindblad form is employed for determining the time-evolution of the
density matrix. Here two kinds of loss processes, dissipation and dephasing,
are considered as Lindblad terms
\begin{equation}
{\mathcal{L}}\left( \widehat{\rho }\right) =\sum_{j=1}^{N}\left[ {\mathcal{L
}_{\mathrm{diss},j}\left( \widehat{\rho }\right) +{\mathcal{L}}_{\mathrm{dep
},j}\left( \widehat{\rho }\right) \right] .
\end{equation
We suppose that each quasi-spin on the B850 ring is coupled to an
independent heat reservoir \cite{Guzik09}, which reflects the local modes of
phonons and other local fluctuations. Then the dissipation from the $j$th
unit cell is described as
\begin{equation}
{\mathcal{L}}_{\mathrm{diss},j}\left( \widehat{\rho }\right) =\Gamma
_{j}\sum_{O=B,C}(O_{j}\widehat{\rho }O_{j}^{\dag }-\frac{1}{2}\left\{
O_{j}^{\dag }O_{j},\widehat{\rho }\right\} ), \label{dissipation}
\end{equation
where $\left\{ \cdot ,\cdot \right\} $ denotes the anti-commutator.
Here, the sink rate $\Gamma _{j}$ at the $j$th point may be site
dependent. For the dynamics constrained on the subsystem described
by $O$ operators, the last term of
${\mathcal{L}}_{\mathrm{diss},j}\left( \widehat{\rho }\right) $
gives contribution $-\Gamma _{j}\rho _{Oj,Oj}$ to $d\rho
_{Oj,Oj}/dt$, thus dissipation results\ in the reduction of the
total population. Therefore, the dissipation term Eq.
(\ref{dissipation}) represents the incoherent transfer of energy
into the environment.
On the other
hand, the dephasing term read
\begin{equation}
{\mathcal{L}}_{\mathrm{deph},j}\left( \widehat{\rho }\right) =\Gamma
_{j}^{\prime }\sum_{O=B,C}(O_{j}^{\dag }O_{j}\widehat{\rho }O_{j}^{\dag
}O_{j}-\frac{1}{2}\left\{ O_{j}^{\dag }O_{j},\widehat{\rho }\right\} ).
\label{dephasing}
\end{equation
Compared with the dissipation term, the dephasing one ${\mathcal{L}}_
\mathrm{deph},j}\left( \widehat{\rho }\right) $ does not contribute\
to any time local change of the probability distribution, i.e., the
derivative of the diagonal elements of the density matrix is
irrelevant to this term. Thus the total population
$\sum_{O=A,B,C}\sum_{j}\rho _{Oj,Oj}$ would be conserved if only the
dephasing term were present. However, the dephasing process is also
incoherent since it make the nondiagonal elements of the density
matrix tend to zero.
The above two contributions force the LH2 system to ultimately reach a
steady state $\widehat{\rho }_{\mathrm{steady}}=\left\vert 0\right\rangle
\left\langle 0\right\vert =\widehat{\rho }_{v,v}$, namely, in the long-time
limit, all excitations are sinked away. The same steady state is obtained
from Eq. (\ref{mastereq}) in the super-operator for
\begin{equation}
\frac{d}{dt}[\rho ]=M[\rho ],
\end{equation
where $[\rho ]$ denotes the column vector defined by all matrix elements in
some order, and the super-operator $M$ is determined by
\begin{equation}
M[\rho ]=[-i\left[ H_{S},\widehat{\rho }\right] +{\mathcal{L}}\left(
\widehat{\rho }\right) ].
\end{equation
In this sense the\ steady state is just the non-trivial eigenstate of $M$
with vanishing eigen-energy. Usually, from $\det M=0,$ the steady state can
be found.
However, we are interested in the system dynamics on a short timescale,
i.e., how soon can the excitations be transferred from one LH2 to the other
light-harvesting complexes? To this end, the transfer efficiency $\eta
\left( t\right) $ is defined as the population $\rho _{v,v}\left( t\right) $
of the vacuum state $\left\vert 0\right\rangle $ at time $t$,
\begin{equation}
\eta \left( t\right) =\rho _{v,v}\left( t\right) .
\end{equation
The corresponding master equation (\ref{mastereq}
\begin{align}
\frac{d\rho _{v,v}}{dt}& =\sum_{j=1}^{N}\Gamma _{j}\left\langle 0\right\vert
\left( B_{j}\widehat{\rho }B_{j}^{\dag }+C_{j}\widehat{\rho }C_{j}^{\dag
}\right) \left\vert 0\right\rangle \notag \\
& =\sum_{j=1}^{N}\Gamma _{j}\sum_{O=B,C}\rho _{Oj,Oj}
\end{align
means that only the first term of ${\mathcal{L}}_{\mathrm{diss},j}\left(
\rho \right) $ contributes to the time derivatives of $\rho _{v,v}\left(
t\right) $. The transfer efficiency is given by the integral of the above
formula \cite{Johnson08,Guzik09,Plenio092}
\begin{equation}
\eta \left( t\right) =\int_{0}^{t}\sum_{j=1}^{N}\Gamma _{j}\sum_{O=B,C}\rho
_{Oj,Oj}\left( t^{\prime }\right) dt^{\prime }. \label{Eeta}
\end{equation
The average transfer time $\tau $ is further defined as \cit
{Guzik09,Johnson08}
\begin{eqnarray}
\tau &=&\lim_{t\rightarrow \infty }\frac{1}{\eta \left( t\right)
\int_{0}^{t}t^{\prime }\sum_{j=1}^{N}\Gamma _{j}\sum_{O=B,C}\rho
_{Oj,Oj}\left( t^{\prime }\right) dt^{\prime } \notag \\
&=&\frac{1}{\overline{\eta }}\int_{0}^{\infty }t^{\prime
}\sum_{j=1}^{N}\Gamma _{j}\sum_{O=B,C}\rho _{Oj,Oj}\left( t^{\prime }\right)
dt^{\prime }, \label{Etau}
\end{eqnarray
where usually
\begin{equation}
\overline{\eta }=\lim_{t\rightarrow \infty }\eta \left( t\right) =1.
\end{equation
Therefore, an efficient energy transfer requires not only a perfect
transmission efficiency $\eta $ but also a short average time $\tau $.
In Appendix A, we present an equivalent non-Hermitian Hamiltonian method,
which can also be utilized to study the dynamics of the open system.
\section{$k$-space representation of the master equation}
In this section we present the $k$-space representation of the above master
equation, so that we can reduce the dynamics of time evolution in some
invariant subspace. If all the dissipation and dephasing rates are
homogeneous on the B850 BChl ring, i.e., $\Gamma _{j}=\Gamma $ and $\Gamma
_{j}^{\prime }=\Gamma ^{\prime }$, the whole system has translational
symmetry. For each unit cell containing three BChls shown in Fig. \re
{model1}, we introduce the Fourier transformation,
\begin{equation}
O_{k}^{\dag }=\frac{1}{\sqrt{N}}\sum_{j=1}^{N}e^{ikj}O_{j}^{\dag }
\label{fourier}
\end{equation
for $O=A,B,C$. Then in the $k$-space the Hamiltonian (\ref{Hparticle}) is
represented as $H_{S}=\sum_{k}H_{k}$ wit
\begin{align}
H_{k}& =2J_{1}\cos kA_{k}^{\dag }A_{k}+\{\left( g_{1}+2g_{2}\cos k\right)
\left( A_{k}^{\dag }B_{k}+A_{k}^{\dag }C_{k}\right) \notag \\
& +J_{2}\left[ \left( 1+\delta \right) +\left( 1-\delta \right) e^{-ik
\right] B_{k}^{\dag }C_{k}+\mathrm{H.c.}\}.
\end{align
Here $k$ are chosen as discrete values
\begin{equation}
k=\frac{2\pi l}{N},\text{ for }l=1,2,\cdots ,N.
\end{equation}
\begin{figure}[ptb]
\includegraphics[bb=28 408 463 780, width=8 cm, clip]{fig3.eps}
\caption{(color online) Configuration of the density matrix of the $N=8$
system in the subspace expanded by $\{\left\vert 0\right\rangle , \left\vert
O,k\right\rangle \}$ with $O=A,B,C$ and $k=(2\protect\pi/8)
\times1,2,\cdots,8$. An initial state localized in the $(k_{1},k_{2})$-block
can be evolved to other $(k,k_{2}+k-k_{1})$-blocks (black hollow dot-dash
squares). Only the diagonal $(k,k)$-blocks (green solid squares) are related
to the average transfer time.}
\label{romatrix}
\end{figure}
In the subspace of the single excitation plus the vacuum with the basi
\begin{equation}
\{\left\vert 0\right\rangle ,\left\vert O,k\right\rangle \equiv O_{k}^{\dag
}\left\vert 0\right\rangle |k=\frac{2\pi l}{N};l=1,2,\cdots N;O=A,B,C\},
\end{equation
the general density matrix is decomposed into
\begin{equation}
\widehat{\rho }=\widehat{\rho }_{v,v}+\sum_{k_{1},k_{2}}\widehat{\rho
_{k_{1},k_{2}}+\sum_{k}\left( \widehat{\rho }_{v,k}+\widehat{\rho
_{k,v}\right) . \label{rho}
\end{equation
where
\begin{equation}
\widehat{\rho }_{v,v}=\rho _{v,v}\left\vert 0\right\rangle \left\langle
0\right\vert
\end{equation
is\ the vacuum block while
\begin{equation}
\widehat{\rho }_{k_{1},k_{2}}=\sum_{O,O^{\prime }=A,B,C}\rho
_{Ok_{1},O^{\prime }k_{2}}\left\vert O,k_{1}\right\rangle \left\langle
O^{\prime },k_{2}\right\vert .
\end{equation
is called the $\left( k_{1},k_{2}\right) $-block. For fixed $k_{1}$ and
k_{2}$, $\rho _{Ok_{1},O^{\prime }k_{2}}$ form a matrix
\begin{equation}
\left(
\begin{array}{ccc}
\rho _{Ak_{1},Ak_{2}} & \rho _{Ak_{1},Bk_{2}} & \rho _{Ak_{1},Ck_{2}} \\
\rho _{Bk_{1},Ak_{2}} & \rho _{Bk_{1},Bk_{2}} & \rho _{Bk_{1},Ck_{2}} \\
\rho _{Ck_{1},Ak_{2}} & \rho _{Ck_{1},Bk_{2}} & \rho _{Ck_{1},Ck_{2}
\end{array
\right) .
\end{equation
The $k$-space representation of the density matrix\ is illustrated in Fig.
\ref{romatrix} for the $N=8$ system.
In the $k$-space, the master equation (\ref{mastereq}) is reduced to
\begin{align}
& \frac{d\widehat{\rho}_{k_{1},k_{2}}}{dt}=-i\left( H_{k_{1}}\widehat{\rho
_{k_{1},k_{2}}-\widehat{\rho}_{k_{1},k_{2}}H_{k_{2}}\right) \notag \\
& +\sum_{O=B,C}\left\{ \frac{\Gamma^{\prime}}{N}\sum_{k}O_{k}^{\dag}O_{k_{1}
\widehat{\rho}_{k_{1},k_{2}}O_{k_{2}}^{\dag}O_{k_{2}+k-k_{1}}\right. \notag
\\
& \left. -\frac{1}{2}\left( \Gamma+\Gamma^{\prime}\right) \left(
O_{k_{1}}^{\dag}O_{k_{1}}\widehat{\rho}_{k_{1},k_{2}}+\widehat{\rho
_{k_{1},k_{2}}O_{k_{2}}^{\dag}O_{k_{2}}\right) \right\} \label{drok1k2}
\end{align}
for all the $k_{1},k_{2}$,
\begin{align}
\frac{d\widehat{\rho}_{k,v}}{dt} & =-iH_{k}\widehat{\rho}_{k,v}-\frac{1}{2
\left( \Gamma+\Gamma^{\prime}\right) \sum_{O=B,C}O_{k}^{\dag}O_{k}\widehat
\rho}_{k,v}, \notag \\
\frac{d\widehat{\rho}_{v,k}}{dt} & =i\widehat{\rho}_{v,k}H_{k}-\frac{1}{2
\left( \Gamma+\Gamma^{\prime}\right) \sum_{O=B,C}\widehat{\rho
_{v,k}O_{k}^{\dag}O_{k}
\end{align}
for all the $k$, and
\begin{equation}
\frac{d\widehat{\rho}_{v,v}}{dt}=\sum_{k}\sum_{O=B,C}\Gamma O_{k}\widehat
\rho}_{k,k}O_{k}^{\dag}. \label{dro00}
\end{equation}
The details of the calculation are shown in Appendix B.
We notice that the equations about $\widehat{\rho }_{k,v}$ and $\widehat
\rho }_{v,k}$ are completely decoupled from $\widehat{\rho }_{k_{1},k_{2}}$
and $\widehat{\rho }_{v,v}$. It follows from Eq. (\ref{drok1k2}) that when
no dephasing exists, i.e., $\Gamma ^{\prime }=0$, the $\left(
k_{1},k_{2}\right) $-block $\widehat{\rho }_{k_{1},k_{2}}$ is decoupled with
other $\widehat{\rho }_{k_{1}^{\prime },k_{2}^{\prime }}$ for $\left(
k_{1}^{\prime },k_{2}^{\prime }\right) \neq \left( k_{1},k_{2}\right) $.
Thus $\widehat{\rho }_{k_{1},k_{2}}$ only evolves in the $\left(
k_{1},k_{2}\right) $-block. However, when the dephasing is present ($\Gamma
^{\prime }\neq 0$), the term
\begin{equation}
\sum_{O=B,C}\frac{\Gamma ^{\prime }}{N}\sum_{k}O_{k}^{\dag }O_{k_{1}
\widehat{\rho }_{k_{1},k_{2}}O_{k_{2}}^{\dag }O_{k_{2}+k-k_{1}}
\end{equation
actually induces the coupling between the $\left( k_{1},k_{2}\right) $-block
and the $\left( k,k_{2}+k-k_{1}\right) $-block. The initial $\widehat{\rho
_{k_{1},k_{2}}$ may evolves to $\widehat{\rho }_{k,k_{2}+k-k_{1}}$ as time
goes by. A typical example of $\left( k,k_{2}+k-k_{1}\right) $-blocks are
shown by the black hollow dot-dash squares in Fig. \ref{romatrix}. The
momentum difference $k_{1}-k_{2}$ is conserved during the evolution sinc
\begin{equation}
k_{2}-k_{1}=(k_{2}+k-k_{1})-k.
\end{equation
In addition, Eq. (\ref{dro00}) means that only the $\left( k,k\right)
-blocks of the density matrix result in energy transfer, which are marked by
the $8$ green solid squares in Fig. \ref{romatrix}. All the other $k_{1}\neq
k_{2}$ blocks do not affect the transfer efficiency $\eta \left( t\right) $
and average transfer time $\tau $ at all. Especially, the initial component
\widehat{\rho }_{k_{1},k_{2}}$ with $k_{1}\neq k_{2}$ will not influence
\eta \left( t\right) $ or $\tau $ at any time $t$ afterwards since
it cannot evolve to the blocks with $k_{1}=k_{2}$. Therefore, only
considering the dynamics of the $\left( k,k\right) $-blocks are
enough for the present purpose.
\section{Transfer efficiency and average transfer time with channel
decomposition}
In this section we use the $k$-space representation of master equation to
calculate the average transfer time and transfer efficiency by the standard
open quantum system method. As a highly organized array of chlorophyll
molecules, the LH2 acts cooperatively to shuttle the energy of photons to
elsewhere when sunlight shines on it. In this sense, we use the density
matri
\begin{equation}
\widehat{\rho }\left( 0\right) =\sum_{k_{1},k_{2}}\sum_{O,O^{\prime
}=A,B,C}\rho _{Ok_{1},O^{\prime }k_{2}}\left( 0\right) \left\vert
O,k_{1}\right\rangle \left\langle O^{\prime },k_{2}\right\vert
\label{initstate1}
\end{equation
to describe the excitations in the initial state. From the discussions in
the last section, only the $k_{1}=k_{2}=k$ blocks relevant to energy
transfer.\ Therefore, there exists an equivalence class of initial state
\begin{equation}
\left[ \widehat{\rho }^{\prime }\left( 0\right) \right] =\left\{ \widehat
\rho }\text{ }|\text{ }\left\langle O,k\right\vert \widehat{\rho }\left\vert
O^{\prime },k\right\rangle =\rho _{Ok,O^{\prime }k}\left( 0\right) \right\}
\end{equation
that results in the same transfer efficiency and average transfer time as
that for $\widehat{\rho }\left( 0\right) $. For further use, a special
density matrix is chosen from the equivalence class $\widehat{\varrho
\left( 0\right) \in \left[ \widehat{\rho }^{\prime }\left( 0\right) \right]
\begin{eqnarray}
\widehat{\varrho }\left( 0\right) &=&\sum_{k}\sum_{O,O^{\prime }=A,B,C}\rho
_{Ok,O^{\prime }k}\left( 0\right) \left\vert O,k\right\rangle \left\langle
O^{\prime },k\right\vert \notag \\
&=&\sum_{k}\widehat{\varrho }^{\left[ k\right] }\left( 0\right) ,
\end{eqnarray
which satisfies $\left\langle O,k_{1}\right\vert \widehat{\varrho }\left(
0\right) \left\vert O^{\prime },k_{2}\right\rangle =\rho _{Ok,O^{\prime
}k}\left( 0\right) $ for $k_{1}=k_{2}=k$, and $\left\langle
O,k_{1}\right\vert \widehat{\varrho }\left( 0\right) \left\vert O^{\prime
},k_{2}\right\rangle =0$ for $k_{1}\neq k_{2}$. $\widehat{\varrho }\left(
0\right) $ plays an equivalent role for determining the transfer efficiency
and average transfer time. Here,
\begin{equation}
\widehat{\varrho }^{\left[ k\right] }\left( 0\right) =\sum_{O,O^{\prime
}=A,B,C}\rho _{Ok,O^{\prime }k}\left( 0\right) \left\vert O,k\right\rangle
\left\langle O^{\prime },k\right\vert
\end{equation
is called as the $k$-channel component of the density matrix. According to
the above observation, we first choose every $\widehat{\varrho }^{\left[
\right] }\left( 0\right) $ as the initial state to obtain the final state
\widehat{\varrho }^{\left[ k\right] }\left( t\right) $, which gives the $k
-channel transfer efficiency at time $t$
\begin{equation}
\eta ^{\left[ k\right] }\left( t\right) =\Gamma \int_{0}^{t}\sum_{k^{\prime
}}\sum_{O=B,C}\varrho _{Ok^{\prime },Ok^{\prime }}^{\left[ k\right] }\left(
t^{\prime }\right) dt^{\prime }, \label{etak}
\end{equation
and the $k$-channel average transfer tim
\begin{equation}
\tau ^{\left[ k\right] }=\frac{\Gamma }{\overline{\eta }}\int_{0}^{\infty
}t^{\prime }\sum_{k^{\prime }}\sum_{O=B,C}\varrho _{Ok^{\prime },Ok^{\prime
}}^{\left[ k\right] }\left( t^{\prime }\right) dt^{\prime }. \label{tauk}
\end{equation
Then we prove a general proposition:
\textit{For an arbitrary initial state }$\widehat{\rho }\left( 0\right)
\textit{\ (Eq. \ref{initstate1}) of the LH2\ complex, the transfer
efficiency at time }$t$\textit{\ and the average transfer time are the sum
of }$\eta ^{\left[ k\right] }\left( t\right) $\textit{\ and }$\tau ^{\left[
\right] }$\textit{\ over all }$k$\textit{-channels, respectively.
\begin{eqnarray}
\eta \left( t\right) &=&\sum_{k}\eta ^{\left[ k\right] }\left( t\right)
\notag \\
\tau &=&\sum_{k}\tau ^{\left[ k\right] }. \label{etatau1}
\end{eqnarray}
In order to prove the above proposition we notice that the effective initial
state $\widehat{\varrho }\left( 0\right) $ evloves to
\begin{equation}
\widehat{\varrho }\left( t\right) =\sum_{k}\widehat{\varrho }^{\left[
\right] }\left( t\right) . \label{etatau2}
\end{equation
Since the corresponding transfer efficiency and average transfer time of
\widehat{\varrho }\left( 0\right) $ ar
\begin{eqnarray}
\eta \left( t\right) &=&\Gamma \int_{0}^{t}\sum_{k^{\prime
}}\sum_{O=B,C}\varrho _{Ok^{\prime },Ok^{\prime }}\left( t^{\prime }\right)
dt^{\prime } \notag \\
\tau &=&\frac{\Gamma }{\overline{\eta }}\int_{0}^{\infty }t^{\prime
}\sum_{k^{\prime }}\sum_{O=B,C}\varrho _{Ok^{\prime },Ok^{\prime }}\left(
t^{\prime }\right) dt^{\prime }, \label{etatau3}
\end{eqnarray
Eq. (\ref{etatau1}) is obtained from Eqs. (\ref{etak}), (\ref{tauk}), (\re
{etatau2}), and (\ref{etatau3}). Namely, $\eta \left( t\right) $ and $\tau $
are the sum of $\eta ^{\left[ k\right] }\left( t\right) $ and $\tau ^{\left[
k\right] }$ for different momentum $k$ channels.
The present experimental observations \cite{HuXiChe972} have provided some
potential pathways for light-harvesting. One of them originates from the
excitations on the B800 BChl ring. It shows that the excitations are
transferred to the RC through B800 (LH2) $\rightarrow $ B850 (LH2)
\rightarrow $ B850 (another LH2) $\rightarrow \cdots \rightarrow $ B875
(LH1) $\rightarrow $ RC. As to our model, the initial state is specialized a
\begin{equation}
\widehat{\rho }\left( 0\right) =\sum_{k_{1},k_{2}}\rho
_{Ak_{1},Ak_{2}}\left( 0\right) \left\vert A,k_{1}\right\rangle \left\langle
A,k_{2}\right\vert . \label{initstate2}
\end{equation
Accordingly, the $k$-channel component of the effective initial state
\widehat{\varrho }\left( 0\right) $ become
\begin{equation}
\widehat{\varrho }^{\left[ k\right] }\left( 0\right) =\rho _{Ak,Ak}\left(
0\right) \left\vert A,k\right\rangle \left\langle A,k\right\vert =\rho
_{Ak,Ak}\left( 0\right) \widehat{\varrho }^{\left[ A,k\right] }\left(
0\right) .
\end{equation
Takin
\begin{equation}
\widehat{\varrho }^{\left[ A,k\right] }\left( 0\right) =\left\vert
A,k\right\rangle \left\langle A,k\right\vert \label{initAk}
\end{equation
as the initial state, we obtain the transfer efficiency and the average
transfer time
\begin{eqnarray}
\eta ^{\left[ A,k\right] }\left( t\right) &=&\Gamma
\int_{0}^{t}\sum_{k^{\prime }}\sum_{O=B,C}\varrho _{Ok^{\prime },Ok^{\prime
}}^{\left[ A,k\right] }\left( t^{\prime }\right) dt^{\prime } \notag \\
\tau ^{\left[ A,k\right] } &=&\frac{\Gamma }{\overline{\eta }
\int_{0}^{\infty }t^{\prime }\sum_{k^{\prime }}\sum_{O=B,C}\varrho
_{Ok^{\prime },Ok^{\prime }}^{\left[ A,k\right] }\left( t^{\prime }\right)
dt^{\prime }.
\end{eqnarray
Hereafter, the superscript $\left[ A,k\right] $ denotes that the initial
state is Eq. (\ref{initAk}). Similar to the above analysis about the
proposition, we present a corollary:
\textit{The transfer efficiency }$\eta \left( t\right) $\textit{\ and
average transfer time }$\tau $\textit{\ of the initial state in Eq. (\re
{initstate2}) are the weighted average of }$\eta ^{\left[ A,.k\right]
}\left( t\right) $\textit{\ and }$\tau ^{\left[ A,k\right] }$\textit{,
respectively. \
\begin{eqnarray}
\eta \left( t\right) &=&\sum_{k}\rho _{Ak,Ak}\left( 0\right) \eta ^{\left[
A,.k\right] }\left( t\right) \notag \\
\tau &=&\sum_{k}\rho _{Ak,Ak}\left( 0\right) \tau ^{\left[ A,k\right] }.
\label{corollary}
\end{eqnarray
In the following, we will show the analytical and numerical results of $\eta
^{\left[ A,.k\right] }\left( t\right) $ and $\tau ^{\left[ A,k\right] }$.
First we consider the case without dephasing, i.e., $\Gamma ^{\prime }=0$.
The time evolution from initial state $\widehat{\varrho }^{\left[ A,k\right]
}\left( 0\right) $ only takes place in the $\left( k,k\right) $-block.
According to Eq. (\ref{drok1k2}), the master equation of $\widehat{\rho
_{k,k}$
\begin{align}
\frac{d\widehat{\rho }_{k,k}}{dt}& =-i\left( H_{k}\widehat{\rho }_{k,k}
\widehat{\rho }_{k,k}H_{k}\right) \notag \\
& -\frac{\Gamma }{2}\sum_{O=B,C}\left( O_{k}^{\dag }O_{k}\widehat{\rho
_{k,k}+\widehat{\rho }_{k,k}O_{k}^{\dag }O_{k}\right) . \label{drokk}
\end{align
gives the average transfer tim
\begin{equation}
\tau ^{\left[ A,k\right] }=\frac{\Gamma }{\overline{\eta }}\int_{0}^{\infty
}t^{\prime }\sum_{O=B,C}\varrho _{Ok,Ok}^{\left[ A,k\right] }\left(
t^{\prime }\right) dt^{\prime }.
\end{equation}
When $k=0$, Eq. (\ref{drokk}) about $\rho _{Ok,O^{\prime }k}$ is rearranged
as a system of differential equations about $v_{j}\left( t\right) $ (
j=1,\cdots ,4$) and $v_{5}\left( t\right) =\left[ v_{4}\left( t\right)
\right] ^{\ast }$:
\begin{align}
v_{1}& =\rho _{Ak,Ak}, \notag \\
v_{2}& =\rho _{Bk,Bk}+\rho _{Ck,Ck},v_{3}=\rho _{Bk,Ck}+\rho _{Ck,Bk},
\notag \\
v_{4}& =\rho _{Ak,Bk}+\rho _{Ak,Ck},v_{5}=\rho _{Bk,Ak}+\rho _{Ck,Ak}.
\end{align
It is
\begin{align}
\frac{d}{dt}v_{1}\left( t\right) & =ig_{+}\left[ v_{4}\left( t\right)
-v_{5}\left( t\right) \right] , \notag \\
\frac{d}{dt}v_{2}\left( t\right) & =-ig_{+}\left[ v_{4}\left( t\right)
-v_{5}\left( t\right) \right] -\Gamma v_{2}\left( t\right) , \notag \\
\frac{d}{dt}v_{3}\left( t\right) & =-ig_{+}\left[ v_{4}\left( t\right)
-v_{5}\left( t\right) \right] -\Gamma v_{3}\left( t\right) , \notag \\
\frac{d}{dt}v_{4}\left( t\right) & =2ig_{+}v_{1}\left( t\right) -ig_{+}\left[
v_{2}\left( t\right) +v_{3}\left( t\right) \right] \notag \\
& -\left[ 2i\left( J_{1}-J_{2}\right) +i\Delta \Omega +\frac{\Gamma }{2
\right] v_{4}\left( t\right) , \label{vvv}
\end{align
with initial condition
\begin{equation}
v_{1}\left( 0\right) =1,v_{2}\left( 0\right) =v_{3}\left( 0\right)
=v_{4}\left( 0\right) =v_{5}\left( 0\right) =0.
\end{equation
Here $g_{+}=\left( g_{1}+2g_{2}\right) $. Solving the above differential
equations, we obtai
\begin{align}
\tau ^{\left[ A,k=0\right] }& =\frac{\Gamma }{\overline{\eta }
\int_{0}^{\infty }t^{\prime }v_{2}\left( t^{\prime }\right) dt^{\prime }
\notag \\
& =\frac{g_{+}^{2}+\left( J_{1}-J_{2}+\Delta \Omega /2\right) ^{2}+\Gamma
_{0}^{2}/4}{g_{+}^{2}\Gamma _{0}}, \label{tauk0}
\end{align
with $\Gamma _{0}=\Gamma /2$, which is independent of the dimerization
parameter $\delta $. Similarly, when $k=\pm \pi $, the average transfer time
of $\widehat{\varrho }^{\left[ A,k\right] }\left( t_{0}\right) $ i
\begin{equation}
\tau ^{\left[ A,k=\pm \pi \right] }=\frac{g_{-}^{2}+\left( J_{1}+J_{2}\delta
-\Delta \Omega /2\right) ^{2}+\Gamma _{0}^{2}/4}{g_{-}^{2}\Gamma _{0}},
\label{taukpi}
\end{equation
where $g_{-}=\left( g_{1}-2g_{2}\right) $. It is a quadratic function with
respect to $\delta $. The optimal parameter $\delta $ with the shortest
transfer time satisfies
\begin{equation*}
\delta _{\mathrm{opt}}^{\left[ A,k=\pm \pi \right] }=\frac{\Delta \Omega
/2-J_{1}}{J_{2}}.
\end{equation*
When $g_{1}=2g_{2}$, Eq. (\ref{taukpi}) shows that $\tau ^{\left[ A,k=\pm
\pi \right] }=\infty $, corresponds to $\overline{\eta }=0$, the energy
transfer is prevented at this time.
\begin{figure}[tbp]
\includegraphics[bb=20 200 520 770, width=8 cm, clip]{fig4.eps}
\caption{(color online) The average transfer time $\protect\tau ^{[A,k]}$ of
$\widehat{\protect\varrho}^{[A,k]}(0) $ (blue scatter lines) and $\protec
\tau _{\mathrm{mix}}$ of the initial mixed state $\widehat{\protect\rho} _
\mathrm{mix}}^{A}(0)$ (red solid lines) with respect to the
dimerization degree $\protect\delta $ of the B850 BChl ring. Here
$N=8$, $J_{1}/\Gamma=0.3$, $J_{2}/\Gamma=1$, $g_{1}/\Gamma=0.5$, $
\Delta\Omega / \Gamma=0.1$, $\Gamma^{\prime }/\Gamma=1$,
$g_{2}/\Gamma=0$ (upper panel) and $g_{2}/\Gamma=0.125$ (lower
panel). $\protect\tau $ is in the
unit of $(1/\Gamma )$ and $\overline{\protect\eta }=1$. It shows that each
\protect\tau ^{[A,k]}$ ($k\neq 0$) curve has a minimum at $\protect\delta _
\mathrm{opt}}^{\left[ A,k\right] }\neq 0$. $\protect\tau _{\mathrm{mix}}$ is
the equal-weighted average of a complete set of $\left\{ \protect\tau
^{[A,k]}\right\} $.}
\label{figtau}
\end{figure}
If the dephasing is present, i.e., $\Gamma ^{\prime }\neq 0$, we can provide
approximate solutions for $\tau ^{\left[ A,k=0\right] }$ and $\tau ^{\left[
A,k=\pm \pi \right] }$
\begin{eqnarray}
\tau ^{\left[ A,k=0\right] } \!\!&=& \!\! \frac{g_{+}^{2}\!\left( 4\Gamma
_{s} \! -\Gamma ^{\prime }\right) \! \!/\Gamma \! +\!\left( 2J_{1} \!
-2J_{2} \! +\Delta \Omega \right) ^{2}\! \! +\Gamma _{s}^{2}/4}
2g_{+}^{2}\Gamma _{s}} \notag \\
\tau ^{\left[ A,k=\pm \pi \right] } \!\! &=& \!\! \frac{g_{-}^{2}\!\left(
4\Gamma _{s} \! -\Gamma ^{\prime }\right) \!\! /\Gamma \! +\!\left( 2J_{1}
\! +2J_{2}\delta \! -\Delta \Omega \right) ^{2} \!\! +\Gamma _{s}^{2}/4}
2g_{-}^{2}\Gamma _{s}}, \label{tauappro}
\end{eqnarray}
where $\Gamma _{s}=\Gamma +\Gamma ^{\prime }$. They almost exactly agree
with the numerical calculation below, and can also be confirmed by Eq. (\re
{tauk0}) and (\ref{taukpi}) when $\Gamma ^{\prime }=0$. The details are
shown in Appendix C.
\section{Energy transfer efficiency and average transfer time in numerical calculation}
For a general $k$, the analytical solution of $\eta ^{\left[
A,k\right] }\left( t\right) $ and $\tau ^{\left[ A,k\right] }$ is
not easy to get. Nevertheless, the numerical results of $\tau
^{\left[ A,k\right] }$ as a function of $\delta $ are plotted as
blue scatter lines in Fig. \ref{figtau}. Here we have chosen
\begin{eqnarray}
N &=&8,\frac{J_{1}}{\Gamma }=0.3,\frac{J_{2}}{\Gamma }=1, \notag \\
\frac{g_{1}}{\Gamma } &=&0.5,\frac{\Delta \Omega }{\Gamma }=0.1,\frac{\Gamma
^{\prime }}{\Gamma }=1,
\end{eqnarray
$g_{2}/\Gamma =0$ for the upper panel, $g_{2}/\Gamma =0.125$ for the
lower panel, and $t$ is in the unit of $\left( 1/\Gamma \right) $
and is long enough to ensure $\overline{\eta }=1$. It shows that
when $k\neq 0$
and $\delta $ varies from $-1$ to $1$, there always exist optimum cases
\delta _{\mathrm{opt}}^{\left[ A,k\right] }\neq 0$ with and shorter
average transfer time. This fact reflects the enhanced effect of
dimerization.
We then take the mixed initial density matrix $\widehat{\rho }\left(
0\right) =\widehat{\rho }_{\mathrm{mix}}^{A}\left( 0\right) $ as an example
\begin{eqnarray}
\widehat{\rho }_{\mathrm{mix}}^{A}\left( 0\right) &=&\sum_{j=1}^{N}\rho
_{Aj,Aj}\left( 0\right) \left\vert A,j\right\rangle \left\langle
A,j\right\vert \notag \\
&=&\sum_{k_{1},k_{2}}\rho _{Ak_{1},Ak_{2}}\left( 0\right) \left\vert
A,k_{1}\right\rangle \left\langle A,k_{2}\right\vert .
\end{eqnarray
The weight $\rho _{Ak,Ak}$ always satisfie
\begin{equation}
\rho _{Ak,Ak}\left( 0\right) =\frac{1}{N}\sum_{j=1}^{N}\rho _{Aj,Aj}\left(
0\right) =\frac{1}{N}.
\end{equation
From Eq. (\ref{corollary}), the transfer efficiency and the average transfer
time of $\widehat{\rho }_{\mathrm{mix}}^{A}$ i
\begin{eqnarray}
\eta _{\mathrm{mix}}\left( t\right) &=&\frac{1}{N}\sum_{k}\eta
^{\left[
A,k\right] }\left( t\right) \\
\tau _{\mathrm{mix}} &=&\frac{1}{N}\sum_{k}\tau ^{\left[ A,k\right] },
\end{eqnarray
$\tau_{\mathrm{mix}}$ is also verified numerically and shown in Fig.
\ref {figtau} as the red solid lines.
In order to see the dynamics of the transfer process clearly, we
plot $\eta _{\mathrm{mix}}$ with respect to the dimerization degree
$\delta$ and time $t$ in Fig. \ref{figeta}(a), i.e., $\eta
_{\mathrm{mix}}=\eta _{\mathrm{mix}}(\delta,t)$. At a certain
instant $t_{0}=12$, $\eta _{\mathrm{mix}}(\delta,t_{0})$ as a
function of $\delta$ is plotted in Fig. \ref{figeta}(b), while for a
certain dimerization degree $\delta_{0}=-0.5$, $\eta
_{\mathrm{mix}}(\delta_{0},t)$ as a function of $t$ is plotted in
Fig. \ref{figeta}(c). Here the parameters are chosen as same as the
ones in Fig. \ref{figtau} except that $g_{2}/\Gamma=0.125$. The
contour map Fig. \ref{figeta}(a) and the profiles of $\eta
_{\mathrm{mix}}(\delta,t)$ in Fig. \ref{figeta}(b) and (c) show that
(1) $\eta _{\mathrm{mix}}(\delta,t)$ increases monotonously as time
goes by. In the large $t$ limit, $\eta _{\mathrm{mix}}(\delta,t)$
equals to $1$. (2) At any certain short instant, an optimum $\delta$
can enhance the transfer efficiency.
Similar to $\eta _{\mathrm{mix}}$ and $\tau _{\mathrm{mix}}$, in
general, there exists an
optimal $\delta _{\mathrm{opt}}\neq 0$ for an arbitrary initial $\widehat
\rho }\left( 0\right) $, which means that a suitable distortion of the B850
ring is helpful for the excitation transfer. This result agrees with the
x-ray observation that the Mg-Mg distance between neighboring B850 BChls is
9.2\AA\ within the $\alpha \beta $-heterodimer and 8.9\AA\ between the
heterodimers reported in Ref. \cite{HuXiChe96}. The B850 ring is indeed
dimerized in nature.
\begin{figure}[ptb]
\includegraphics[bb=16 280 530 743, width=8 cm, clip]{fig5.eps}
\caption{(color online) (a) The contour map of the transfer
efficiency of the initial mixed state $\eta
_{\mathrm{mix}}(\delta,t)$ as a function of dimerization degree
$\delta$ and time $t$ for the same setup as that in the lower panel
of Fig. \ref{figtau}. (b) The profile of $\eta _{\mathrm{mix}}$
along $t_{0}=12$ (black dashed line in (a)). (c) The profile of
$\eta _{\mathrm{mix}}$ along $\delta_{0}=-0.5$ (red dot line in
(a)). It shows that $\eta _{\mathrm{mix}}(\delta,t)$ increases over
time, and an optimum $\delta$ can enhance the transfer efficiency.}
\label{figeta}
\end{figure}
As shown in Fig. \ref{figtau}, $\tau ^{\left[ A,k=0\right] }$ and $\tau ^
\left[ A,k=\pm \pi \right] }$ are particular since nearly all the other
\tau ^{\left[ A,k\right] }$ are within the range of $\left[ \tau ^{\left[
A,k=0\right] },\tau ^{\left[ A,k=\pm \pi \right] }\right] $, so is the
average transfer time $\tau $ of an arbitrary $\widehat{\rho }\left(
0\right) $. Besides, the absolute value of $\delta _{\mathrm{opt}}^{\left[
A,k=\pm \pi \right] }$ for the $k=\pm \pi $ case is larger than the one of
other $\rho \left( 0\right) $, i.e., $\left\vert \delta _{\mathrm{opt
}\right\vert \leq \left\vert \delta _{\mathrm{opt}}^{\left[ A,k=\pm \pi
\right] }\right\vert $.\ Hence, once we have known the properties of $\tau ^
\left[ A,k=0\right] }$ and $\tau ^{\left[ A,k=\pm \pi \right] }$, the
behavior of a general $\tau $ can be conjectured to some extend. Compared
the lower panel of Fig. \ref{figtau} with the upper panel, a larger
g_{2}/\Gamma $ can increase $\tau ^{\left[ A,k=\pm \pi \right] }$ but
decrease $\tau ^{\left[ A,k=0\right] }$. In the $g_{2}/\Gamma =0.125$ case,
the homogeneous pure state $\widehat{\varrho }^{\left[ A,k=0\right] }\left(
0\right) $ is better than the mixed state $\widehat{\rho }_{\mathrm{mix
}^{A}\left( 0\right) $ for energy transport. However, the upper panel with
g_{2}/\Gamma =0$ gives the contrary result.
The minimal $\tau ^{\left[ A,k=\pm \pi \right] }$ is reachable at $\delta _
\mathrm{opt}}^{\left[ A,k=\pm \pi \right] }=\left( \Delta \Omega
/2-J_{1}\right) /J_{2}$,
\begin{equation}
\tau _{\min }^{\left[ A,k=\pm \pi \right] }=\frac{\left( g_{1}-2g_{2}\right)
^{2}\left( 4\Gamma +3\Gamma ^{\prime }\right) +\Gamma \left( \Gamma +\Gamma
^{\prime }\right) ^{2}/4}{2\left( g_{1}-2g_{2}\right) ^{2}\Gamma \left(
\Gamma +\Gamma ^{\prime }\right) }.
\end{equation
In the toy model illustrated in Fig. \ref{figtau}, $J_{2}>J_{1}$ and
g_{2}<g_{1}$. When
\begin{equation}
g_{2}/g_{1}=\gamma _{g}=\frac{1}{2}+\xi ^{2}-\xi \sqrt{1+\xi ^{2}},
\end{equation
we have $\tau _{\min }^{\left[ A,k=\pm \pi \right] }=\tau ^{\left[ A,k=
\right] }$, where
\begin{equation}
\xi =\frac{\left( \Gamma +\Gamma ^{\prime }\right) }{2\left(
2J_{2}-2J_{1}-\Delta \Omega \right) }.
\end{equation
On the side of $0<g_{2}/g_{1}<\gamma _{g}$, $\tau _{\min }^{\left[ A,k=\pm
\pi \right] }<\tau ^{\left[ A,k=0\right] }$, while on the\ other side
\gamma _{g}<g_{2}/g_{1}<1$, $\tau _{\min }^{\left[ A,k=\pm \pi \right]
}>\tau ^{\left[ A,k=0\right] }$.
In general, the shortest average transfer time of an arbitrary initial
\widehat{\rho }\left( 0\right) $ is within the range of $\left[ \tau _{\min
}^{\left[ A,k=\pm \pi \right] },\tau ^{\left[ A,k=0\right] }\right] $. The
mean value of $\tau _{\min }^{\left[ A,k=\pm \pi \right] }$ and $\tau ^
\left[ A,k=0\right] }$ can roughly reflect the influence of parameters on
the transfer process
\begin{equation}
\overline{\tau }=\frac{1}{2}\left( \tau _{\min }^{\left[ A,k=\pm \pi \right]
}+\tau ^{\left[ A,k=0\right] }\right) .
\end{equation
In Fig. \ref{tao2}, we plot $\overline{\tau }$ with respect to $g_{1}/\Gamma
$ for different $J_{2}/\Gamma =0,1,2,3$. Here, $J_{1}/\Gamma =0.3$, $\delta
=\left( \Delta \Omega /2-J_{1}\right) /J_{2}$, $g_{2}/g_{1}=0.25$, $\Delta
\Omega /\Gamma =0.1$, $\Gamma ^{\prime }/\Gamma =0.5$, and $\overline{\tau }$
is in the unit of $\left( 1/\Gamma \right) $. It shows that $\overline{\tau }
$ decreases monotonously as $g_{1}/\Gamma $ increases. In the short
g_{1}/\Gamma $ limit, $\overline{\tau }$ tends to infinity, which is
reasonable since the two BChl rings are decoupled in this case. Moreover,
\overline{\tau }$ is larger when $J_{2}/\Gamma $ is larger.
\begin{figure}[ptb]
\includegraphics[bb=30 357 550 772, width=8 cm, clip]{fig6.eps}
\caption{(color online) Plots of $\overline{\protect\tau}$ as a function of
the dissipation ratio $g_{1}/\Gamma$ with $J_{2}/\Gamma=0,1,2,3$, where
J_{1}/\Gamma=0.3$, $\protect\delta=(\Delta\Omega/2-J_{1})/J_{2}$,
g_{2}/g_{1}=0.25$, $\Delta\Omega / \Gamma=0.1$, $\Gamma^{\prime}/\Gamma=0.5
, and $\overline{\protect\tau}$ is in the unit of $(1/\Gamma) $. It shows
that $\overline{\protect\tau}$ decreases as $g_{1}/\Gamma$ increases, but
increases with the increasing of $J_{2}/\Gamma$. }
\label{tao2}
\end{figure}
\section{Conclusion}
In summary, we have studied the craggy transfer in light-harvesting
complex with dimerization. We employed the open quantum system
approach to show that the dimerization of the B850 BChl ring can
enhance the transfer efficiency and shorten the average transfer
time for different initial states with various quantum superposition
properties. Actually our present investigation only focuses on a
crucial stage in photosynthesis -- the energy transfer, which is
carried by the coherent excitations in the typical light-harvesting
complex II (LH2). Here the LH2 is modeled as two coupled
bacteriochlorophyll (BChl) rings. With this modeling, the ordinary
photosynthesis is roughly described as three basic steps: 1)
stimulate an excitation in LH2; 2) transfer it to another LH2 or
LH1; 3) the energy causes the chemical reaction that converts carbon
dioxide into organic compounds. Namely, the excitations are
transferred to the RC through B800 (LH2) $\rightarrow $ B850 (LH2)
$\rightarrow $ B850 (another LH2) $\rightarrow \cdots \rightarrow $
B875 (LH1) $\rightarrow $ RC. Obviously, the first two\ are of
physics, thus our present approach can be generalized to investigate
these physical processes. Although photosynthesis happens in
different fashions for different species, some features are always
in common from the point of view of physics. For example, the
photosynthetic process always starts from the light absorbing and
energy transfer.
Another important issues of the photosynthesis physics concerns
about the quantum natures of light \cite{Glauber1,Glauber2}. Since
the experiments have illustrated the role of the quantum coherence
of collective excitations in LH complexes, it is quite natural to
believe that the excitation coherence may be induced by the higher
coherence of photon. Therefore, in a forthcoming paper we will
report our systematical investigation on how the statistical
properties of quantum light affects the photosynthesis.
\acknowledgments
This work is supported by NSFC No. 10474104, 60433050, 10874091 and No.
10704023, NFRPC No. 2006CB921205 and 2005CB724508.
|
\section{Introduction}
In many applications it is useful to have explicit error bounds
in the prime number theorem. \textsc{Rosser}~\cite{Rosser39,Rosser41}
developed an analytic method which combines a numerical verification
of the \textsc{Riemann} hypothesis with a zero-free region and derived explicit estimates for some number theoretical functions.
The aim of this paper is to find sharper bounds for the
\textsc{Chebyshev}'s functions $\psi(x)$, the logarithm of the least
common multiple of all integers not exceeding $x$, and
$\theta(x)$, the product of all primes not exceeding $x$:
$$\theta(x)=\sum_{p\leqslant x}\ln p,\hspace{1cm}
\psi(x)=\sum_{\substack{{p,\alpha}\\{p^\alpha\leqslant x}}}\ln p$$
where sum runs over primes $p$ and respectively
over powers of primes $p^\alpha$.
The Prime Number Theorem could be written as follows:
$$\psi(x)=x+o(x),\;\;x\rightarrow +\infty.$$
An equivalent formulation of the above theorem should be:
for all $\varepsilon>0$, there exists $x_0=x_0(\varepsilon)$
such that
$$|\psi(x)-x|<\varepsilon\,x\hspace{0.5cm}\mbox{ for }x\geqslant x_0$$
or
$$|\theta(x)-x|<\varepsilon\,x\hspace{0.5cm}\mbox{ for }x\geqslant x_0.$$
Under Riemann Hypothesis (RH), \textsc{Schoenfeld} \cite{Schoenfeld:MathOfComp:Sharper} gives interesting results.
Without the assumption of the RH, the results are not so accurate and depend on the knowledge about Riemann Zeta function.
This article hangs up on some known results:
the most important works on effective results have been shown by
\textsc{Rosser} \& \textsc{Schoenfeld}
\cite{Rosser&Schoenfeld:Ill:Approx,Rosser&Schoenfeld:MathOfComp:Sharper,Schoenfeld:MathOfComp:Sharper},
\textsc{Robin}~\cite{Robin:Acta:estim} \& \textsc{Massias}~\cite{Massias:NP:effectiv}
\and \textsc{Costa Pereira}~\cite{Pereira:MathOfComp:Chebyshev}.
The proofs for estimates of $\psi(x)$
in \cite{Rosser&Schoenfeld:MathOfComp:Sharper}
are based on the verification of {\sc Riemann} hypothesis
to a given height
and an explicit zero-free region for $\zeta(s)$
whose form is essentially that the classical one of
\textsc{De la vall\'ee Poussin}. \textsc{Rosser \& Schoenfeld}
have shown that the first 3~502~500 zeros of $\zeta(s)$
are on the critical strip.
\textsc{Van de Lune }{\it et al}~\cite{Lune86:ZerosZeta} have shown that
the first 1~500~000~000 zeros are on the critical strip.
Recently, \textsc{Wedeniwski}~\cite{zetagrid} then \textsc{Gourdon} \cite{Gourdon} manage to compute zeros in a parallel way and prove that the Riemann Hypothesis is true at least for first $10^{13}$ nontrivial zeros.
This will improve bounds~\cite{dusart:eps} for $\psi(x)$ and $\theta(x)$ for large values of $x$.
We will prove the following results:
\begin{eqnarray*}
\theta(x)-x&<&\frac{1}{36\,260}x
\hspace{1cm}\mbox{ for }x>0,\\
|\theta(x)-x|&\leqslant&0.2\frac{x}{\ln^2 x}
\hspace{1cm}\mbox{ for }x\geqs3\,594\,641.
\end{eqnarray*}
We apply these results on $p_k$, the $k^{th}$ prime, and $\theta(p_k)$.
Let's denote by $\ln_2 x$ for $\ln\ln x.$
The asymptotic expansion of $p_k$ is well known;
\textsc{Cesaro}~\cite{Cesaro} then \textsc{Cipolla}~\cite{Cipolla:determinazione}
expressed it in 1902:
$$p_k=k\left\{\ln k+\ln_2 k-1+\frac{\ln_2 k-2}{\ln k}-\frac{\ln_2^2k-6\ln_2 k+11}{2\ln^2k}
+O\left(\left(\frac{\ln_2k}{\ln k}\right)^3\right)\right\}.$$
A more precise work about this can be find in \cite{Robin:MathOfComp:perm,Salvy:??:Fast}.
The results on $p_k$ are:
\begin{eqnarray*}
p_k&\leqslant &k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2}{\ln k}\right)
\hspace{1cm}\mbox{ for }k\geqslant 688\,383,\\
p_k&\geqslant &k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2.1}{\ln k}\right)
\hspace{1cm}\mbox{ for }k\geqslant 3.
\end{eqnarray*}
We use the above results to prove that, for $x\geqslant 396\,738$, the interval
$$\left[x,x+x/(25\ln^2 x)\right]$$ contains at least one prime.
Let's denote by $\pi(x)$ the number of primes not greater than $x$.
We show that
$$\frac{x}{\ln x}\left(1+\frac{1}{\ln x}\right)\;
{\stackrel{\leqslant}{_{_{\scriptscriptstyle x\geqs599}}}}\;\pi(x)\;
{\stackrel{\leqslant}{_{_{\scriptscriptstyle x>1}}}}\;
\frac{x}{\ln x}\left(1+\frac{1.2762}{\ln x}\right).$$
More precise results on $\pi(x)$ are also shown:
\begin{eqnarray*}
\pi(x)&\geqslant&\frac{x}{\ln x-1} \mbox{ for } x\geqs5\,393,\\
\pi(x)&\leqslant&\frac{x}{\ln x-1.1} \mbox{ for } x\geqs60\,184,\\
\pi(x)&\geqslant&\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2}{\ln^2 x}\right)
\mbox{ for } x\geqslant 88\,783,\\
\pi(x)&\leqslant&\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2.334}{\ln^2 x}\right)
\mbox{ for } x\geqslant 2\,953\,652\,287.
\end{eqnarray*}
\section{Exact computation of $\theta$}
From the well-known identity
\begin{equation}
\label{DefPsi2}
\psi(x)=\sum_{k=1}^\infty \theta(x^{1/k}),
\end{equation}
we have
$$\theta(x)=\psi(x)-\sum_{k=2}^\infty \theta(x^{1/k}).$$
From some exact values of $\psi(x)$ computed by \cite{deleglise:psi}, we obtain Tables \ref{table_theta1} \& \ref{table_theta2} (Exact values of $\theta(x)$)
\section{On the difference between $\psi$ and $\theta$}
As $\theta(2^-)=0$, the summation (\ref{DefPsi2}) ends:
$$\psi(x)=\sum_{k=1}^{\lfloor\frac{\ln x}{\ln 2}\rfloor}\theta(x^{1/k})=\theta(x)+\theta(\sqrt{x})+\sum_{k=3}^{\lfloor\frac{\ln x}{\ln 2}\rfloor}\theta(x^{1/k}).$$
\subsection{Lower Bound}
\begin{prop}
For $x\geqs121$, we have
\begin{equation}
\label{MinDiffPsi}
0.9999\sqrt{x}<\psi(x)-\theta(x)
\end{equation}
\end{prop}
\begin{proof}
By Theorem 24 of \cite{Rosser&Schoenfeld:Ill:Approx} p.73, (\ref{MinDiffPsi}) is verified for $121\leqslant x \leqslant 10^{16}$.
Now by \cite{Pereira:MathOfComp:Chebyshev} p.~211,
$$\psi(x)-\theta(x)=\psi(\sqrt{x})+\sum_{k\geqslant 1}\theta(x^\frac{1}{2k+1}),$$
hence
$$\psi(x)-\theta(x)\geqslant\psi(\sqrt{x})+\theta(x^{1/3}).$$
By Theorem~19 of \cite{Rosser&Schoenfeld:Ill:Approx} p.72, we have
$$\theta(x^{1/3})>\sqrt[3]{x}-2x^{1/6}\mbox{ for }(1423)^3\leqslant x\leqslant (10^8)^3,$$
and we have for $x\geqslant \exp(2b)$,
$$\psi(\sqrt{x})>\sqrt{x}-\varepsilon_b \sqrt{x}=0.9999\sqrt{x}+(0.0001-\varepsilon_b) \sqrt{x}.$$
where $\varepsilon_b$ can be find in Table~\ref{table_eps} (or Table p.358 of \cite{Schoenfeld:MathOfComp:Sharper}).
We verify that
$$(0.0001-\varepsilon_b)\sqrt{x}+\sqrt[3]{x}-2x^{1/6}>0$$
for $10^{16}\leqslant x\leqslant e^{50}$ by intervals (we use $b=18.42,\;20,\;22$).
For $y\geqslant e^{25}$, Table~\ref{table_eps} gives $|\psi(y)-y|<0.00007789y.$ Hence we have by Th.13 of \cite{Rosser&Schoenfeld:Ill:Approx},
\begin{eqnarray*}
|\theta(y)-y|&\leqslant&|\psi(y)-y|+|\theta(y)-\psi(y)|<0.00007789y+1.43\sqrt{y}\\
&<&0.00009y
\end{eqnarray*}
For $x>e^{50}$, we apply the previous result with $y=\sqrt{x}$ to obtain
$$\psi(x)-\theta(x)>\theta(\sqrt{x})\geqslant 0.9999\sqrt{x}.$$
\end{proof}
\subsection{Upper Bound}
\begin{prop}
\label{maj_diff_psi_theta}
For $x>0$,
$$\psi(x)-\theta(x)<1.00007\sqrt{x}+1.78\sqrt[3]{x}.$$
\end{prop}
\begin{proof}
We use (\ref{DiffPsiSqrt}) and Proposition~\ref{prop_eta0}.
\end{proof}
\begin{lemma}
For $x>0$, we have
\begin{equation}
\label{DiffPsiSqrt}
\psi(x)-\theta(x)-\theta(\sqrt{x})<1.777745 x^{1/3}
\end{equation}
\end{lemma}
\begin{proof}
For $x>0$, we have $\theta(x)<1.000081x$ by \cite{Schoenfeld:MathOfComp:Sharper} p.360. Hence
\begin{eqnarray*}
\sum_{k=3}^{\lfloor\frac{\ln x}{\ln 2}\rfloor}\theta(x^{1/k})&<&1.000081\sum_{k=3}^{\lfloor\frac{\ln x}{\ln 2}\rfloor}x^{1/k}\\
&<&1.000081\left(x^{1/3}+\left(\left\lfloor\frac{\ln x}{\ln 2}\right\rfloor-4\right)x^{1/4}\right)\\
&<&1.2\;x^{1/3}\mbox{ for }x>(10^{11})^3.
\end{eqnarray*}
For small values, we have (\ref{DiffPsiSqrt}) by direct computation (Maximal value reaches for $x$=2401).
\end{proof}
\section{Useful Bounds}
\begin{eqnarray}
\label{bound1} p_k\leqslant k\ln p_k &\mbox{ for }&k\geqslant 4,\\
\label{bound2} \ln p_k\leqslant \ln k+\ln_2 k +1&\mbox{ for }&k\geqslant 2.
\end{eqnarray}
\begin{proof}
We deduce (\ref{bound1}) from $\pi(x)>\frac{x}{\ln x}$ (Corollary~1 of \cite{Rosser&Schoenfeld:Ill:Approx}).
By Theorem~3 of \cite{Rosser&Schoenfeld:Ill:Approx}, we have
$p_k<k(\ln k+\ln_2 k-1/2)$ hence $p_k<e k\ln k$ for $k\geqslant 2$.
\end{proof}
\section{On the differences between $\theta$ and identity function}
\begin{prop}
\label{prop_eta0}
$\theta(x)-x<\frac{1}{36\,260} x$ for $x>0$.
\end{prop}
\begin{proof}
By table~\ref{table_012}, we have
$\theta(x)<x$ up to $8\cdot10^{11}$.
With (\ref{MinDiffPsi}) and $8\cdot10^{11}\leqslant x\leqslant e^{28}$,
$$\theta(x)<\psi(x)-0.9999\sqrt{x}<(1.00002841-0.9999/\sqrt{e^{28}})x<1.00002758x.$$
We conclude by computing $\varepsilon_{28}\leqslant 0.00002224$.
\end{proof}
\begin{theorem}
\label{th:eta_k}
We have
$$|\theta(x)-x|<\eta_k \frac{x}{\ln^k x}\quad\mbox{ for }x\geqslant x_k$$
with
$$\begin{array}{|r|c|c|c|c|c|c|c|}
\hline
k & 0 & 1 & 1 & 2 & 2 & 2 & 2 \\
\hline
\eta_k & 1 & 1.2323 & 0.001 & 3.965 & 0.2 &0.05 &0.01 \\
\hline
x_k & 1 & 2 & 908\,994\,923 & 2 &3\,594\,641 & 122\,568\,683&7\,713\,133\,853\\
\hline
\end{array}$$
and
$$\begin{array}{|r|c|c|c|c|c|}
\hline
k& 3 &3 &3 &3 &4\\
\hline
\eta_k&20.83 &10 &1 &0.78 &1300\\
\hline
x_k&2 &32\,321 &89\,967\,803 & 158\,822\,621 &2\\
\hline
\end{array}$$
\end{theorem}
\begin{proof}
We use the estimates of $|\psi(x)-x|$ with Proposition~\ref{maj_diff_psi_theta}.
In particular, we can choose $\eta_2=0.05$ because
$$\scriptstyle(0.00006788+1.00007/\sqrt{10^{11}}+1.78/(10^{11})^{2/3})*26^2
< 0.04809.$$ We obtain Tables~\ref{table_eta1} \& \ref{table_eta2} step by step up to $b=5000$.
For each line, the value is valid between $b_i$ and $b_{i+1}$. Hence, by example, $\eta_2=4.42E-3$ should be chosed for $x\geqslant e^{32}$.
Using Theorem~1.1 of \cite{dusart:eps},
we have $\eta_k\geqslant\sqrt{8/\pi}(\sqrt{\ln(x_0)/R})^{1/2}\cdot e^{-\sqrt{\ln(x_0)/R}}\cdot\ln^k(x_0)$ to
obtain for $x\geqslant x_0= \exp(5000)$,
\begin{eqnarray*}
\eta_0&=&1.196749447941324988148958471 E-12,\\
\eta_1&=&0.000000005983747239706624940744792353,\\
\eta_2&=& 0.00002991873619853312470372396176,\\
\eta_3&=& 0.1495936809926656235186198088,\\
\eta_4&=&747.9684049633281175930990441.
\end{eqnarray*}
Specials constants:\\
for $x\geqslant 1$, $\eta_0<(1-\theta(1^-))/1=(2-\theta(2^-))/2=1$.\\
for $x\geqslant 2$, $\eta_1<(11-\theta(11^-))/11\cdot\ln(11)\approx 1.23227674$.\\
for $x\geqslant 2$, $\eta_2<(59-\theta(59^-))/59\cdot\ln^2(59)\approx 3.964809$\\
for $x\geqslant 2$, $\eta_3<(1423-\theta(1423^-))/1423\cdot\ln^3(1423)\approx 20.8281933$
\end{proof}
\section{Some applications on number theory functions}
\subsection{Estimates of primes
\subsubsection{Estimates of $\theta(p_k)$}
We have an asymptotic development of $\theta(p_k)$:
$$\theta(p_k)=\mbox{Li}^{-1}(k)+O(k^{1/2}\ln^{3/2}k)$$ whose the first terms by \cite{Cipolla:determinazione} are
$$\theta(p_k)=k\left(\ln k+\ln_2 k-1 +\frac{\ln_2 k-2}{\ln k}-\frac{\ln_2^2k-6\ln_2k+11}{2\ln^2 k}+0\left(\frac{\ln_2^3 k}{\ln^3 k}\right)\right)$$
\begin{remark}
We have
\begin{equation}
\label{maj_theta_pk}
\theta(p_k)\leqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2}{\ln k}\right)\quad\mbox{ for } k\geqslant 198.
\end{equation}
by Th.~B(v) of \cite{Massias:NP:effectiv}.
\end{remark}
\begin{prop}
\label{min_theta_pk}
$$\theta(p_k)\geqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2.050735}{\ln k}\right)\quad\mbox{ for }p_k\geqslant 10^{11}$$
$$\theta(p_k)\geqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2.04}{\ln k}\right)\quad\mbox{ for }p_k\geqslant 10^{15}$$
\end{prop}
\begin{proof}
Let $f_\beta$ defined by
$$n\mapsto n\left(\ln n+\ln_2 n-1+\frac{\ln_2 n-\beta}{\ln n}\right).$$
We want to prove that $\theta(p_n)\geqslant f_\beta(n)$.
Define $h_a$ by $h_a(n):=n\left(\ln n+\ln_2 n -a\right).$
Suppose there exist $a$ such that $p_k\geqslant h_a(k)$ for $k\geqslant k_0$.
Hence
$$\theta(p_k)-\theta(p_{k_0})=\sum_{n=k_0+1}^k\ln p_n\geqslant \sum_{n=k_0+1}^k\ln h_a(n).$$
We have $f'_\beta\leqslant \ln h_a$ if
\begin{equation}
\label{eq:beta2}
\frac{\ln_2 n-\beta+1}{\ln n}-\frac{\ln_2 n-\beta-1}{\ln^2 n}\leqslant \ln\left(1+\frac{\ln_2 n-a}{\ln n}\right).
\end{equation}
We can rewrite (\ref{eq:beta2}) as
\begin{equation}
\label{eq:beta3}
\beta(1-1/\ln k)\geqslant 1+\ln_2 k-\ln\left(1+\frac{\ln_2 k -a}{\ln k}\right)\ln k -\frac{\ln_2 k -1}{\ln k}.
\end{equation}
For $a\in[0.95,1]$ and $t\geqslant 22$, the function
$t\mapsto (\ln t-t\ln\left(1+\frac{\ln t -a}{t}\right) -\frac{\ln t -1}{t})/(1-1/t)$ is decreasing.
By \cite{dusart:pk}, we can choose $a=a_0=1$.
For $k\geqslant e^{100}$, the value $\beta=2.048$ satisfies (\ref{eq:beta3}).
For $\pi(10^{11})\leqslant k \leqslant e^{100}$, the value $\beta_0=2.094$ satisfies (\ref{eq:beta3}).
Hence
$$\theta(p_k)\geqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-\beta_0}{\ln k}\right).$$
Then
$p_k\geqslant \theta(p_k)-\eta_2\frac{k}{\ln k}$ by (\ref{th:eta_k}) \& (\ref{bound1}), hence $p_k\geqslant h_{a_1}(k)$ with $a_1=1-\frac{\ln_2 k-(\beta_0+\eta_2)}{\ln k}$.
Splitting the interval of $k$, we use different values of $a$
with adapted values of $\eta_2$. By itering the process, we obtain $\beta=2.050735$ for $k\geqslant k_0=\pi(10^{11})$. This value of $\beta$ verifies $\theta(p_{k_0})\geqslant f_{\beta}(k_0)$.
By same way, we obtain $\beta=2.038$ for $k\geqslant 10^{15}$.
\end{proof}
\begin{prop}
\label{prop:upper:thetaPk}
For $k\geqslant 781$,
$$\theta(p_k)\leqslant k\left(\ln k+\ln_2 k-1 +\frac{\ln_2 k-2}{\ln k}-\frac{0.782}{\ln^2 k}\right)$$
\end{prop}
\begin{proof}
Use Lemma~\ref{lem:upper:pk} and Lemma~\ref{lem:upper:thetaPk}.
\end{proof}
\begin{lemma}
\label{lem:upper:thetaPk}
Let two integers $k_0,k$ and $\gamma>0$ real.
Suppose that for $k_0\leqslant n \leqslant k$,
$$p_n\leqslant n\left(\ln n+\ln_2 n-1+\frac{\ln_2 n-1.95}{\ln n}\right).$$
Let $s(k)=k\left(\ln k+\ln_2 k-1 +\frac{\ln_2 k-2}{\ln k}-\frac{\gamma}{\ln^2 k}\right)$.
Let $f(k)=s(k)-(\ln k +\ln_2 k+1).$ If $\theta(p_{k_0-1})\leqslant f(k_0)$ then $\theta(p_k)\leqslant s(k)$ for all $k\geqslant k_0$.
\end{lemma}
\begin{proof}
Let $S_a(n)$ be an upper bound for $p_n$ for $k_0\leqslant n\leqslant k$ where
$$S_a(n)=n\left(\ln n+\ln_2 n-1+\frac{\ln_2 n-a}{\ln n}\right).$$
Now, for $2\leqslant k_0\leqslant k$, we write
$$\theta(p_{k-1})-\theta(p_{k_0}-1)=\sum_{n=k_0}^{k-1}\ln p_n\leqslant\sum_{n=k_0}^{k-1}\ln S_a(n)\leqslant\int_{k_0}^k\ln S_a(n)dn.$$
We need to prove that $\ln S_a(n)\leqslant f'(n)$.
We have $$\ln S_a(n)= \ln n+\ln_2 n+\ln(1+u(n))$$ with $u(n)=\frac{\ln_2 n-1}{\ln n}+\frac{\ln_2 n-a}{\ln^2 n}$ and
$$f'(n)=\ln n+\ln_2 n+\frac{\ln_2 n-1}{\ln n}-\frac{\ln_2 n+\gamma-3}{\ln^2 n}+\frac{2\gamma}{\ln^3 n}-\frac{1}{n}(1+1/\ln n).$$
Let $\beta<1/2$ such that $\ln(1+u(n))\leqslant u(n)-\beta u^2(n)$ for $n\geqslant k_0$.
Then $\ln S_a(n)\leqslant f'(n)$ if
$$\beta\left(\frac{\ln_2 n-1}{\ln n}+\frac{\ln_2 n-a}{\ln^2 n}\right)^2 -\frac{2\ln_2 n+\gamma-3-a}{\ln^2 n}+2\gamma/\ln^3 n-1/n-1/(n\ln n)\geqslant 0,$$ that we can simplify in
$$\frac{A}{\ln^2 n}+2\frac{B}{\ln^3 n}+\beta\frac{\ln_2^2 n -2a\ln_2 n+a^2}{\ln^4 n}-1/n-1/(n\ln n)\geqslant 0$$
where
$$A=\beta\ln_2^2 n-2(\beta+1)\ln_2 n+3+a+\beta-\gamma$$
$$B=\beta\ln_2^2 n-\beta(a+1)\ln_2 n+a\beta+\gamma$$
We have $1/n+1/(n\ln n)\leqslant 0.02/\ln^3 n$ for $n\geqslant 10^5$.
We study each parts, denoting $\ln_2 n$ by $X$:
\begin{itemize}
\item[$\bullet$] $\beta X^2-2(\beta+1) X +3+a+\beta-\gamma\geqslant 0$ for all $X$ if $\gamma-a-1+1/\beta\leqslant 0$,
\item[$\bullet$] $X^2-(a+1)X+ (a+\gamma/\beta+0.02)\geqslant 0$ for all $X$ if $a^2-2a+1-4(\gamma/\beta+0.02)\leqslant 0$,
\item[$\bullet$] $X^2-2aX+a^2=(X-a)^2\geqslant 0$.
\end{itemize}
We choose $\gamma$ such that $\gamma-a-1+1/\beta=0$.
We choose $\beta=\frac{u(k_0)-\ln(1+u(k_0))}{u^2(k_0)}$.
With $a=1.95$ and $k_0=178974$, we have $\beta=0.461291475\cdots$ and $\gamma=0.78217325\cdots$.
Hence
$\theta(p_{k-1})-f(k)\leqslant \theta(p_{k_0}-1)-f(k_0)$.
As $\theta(p_{k_0}-1)\leqslant f(k_0)$, we have $\theta(p_{k-1})-f(k)\leqslant 0$.
We obtain the upper bound
$\theta(p_k)=\theta(p_{k-1})+\ln p_k \leqslant f(k)+\ln p_k<s(k)$ by (\ref{bound2}).
\end{proof}
\subsubsection{Estimates of $p_k$}
\begin{lemma}
\label{lem:upper:pk}
For $k\geqslant 178\,974$,
$$p_k\leqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-1.95}{\ln k}\right).$$
\end{lemma}
\begin{proof}
Substituing $x$ by $p_k$ in $|\theta(x)-x|\leqslant \eta_2\frac{x}{\ln^2 x}$,
we obtain
$$|p_k- \theta(p_k)|\leqslant\eta_2\frac{p_k}{\ln^2 p_k}.$$
By (\ref{bound1}), we have
$\frac{p_k}{\ln^2 p_k}\leqslant \frac{k}{\ln k}$ and
\begin{equation}
\label{diff_theta_pk_pk}
|p_k - \theta(p_k)|\leqslant \eta_2\frac{k}{\ln k}.
\end{equation}
Using the upper bound (\ref{maj_theta_pk}) of $\theta(p_k)$,
we have
$$p_k\leqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2+\eta_2}{\ln k}\right).$$
We use $\eta_2=0.05$ for $x\geqslant 10^{11}$.
\end{proof}
\begin{prop}
\label{prop:upper:pk}
For $k\geqslant 688\,383$,
$$p_k\leqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2}{\ln k}\right).$$
\end{prop}
\begin{proof}
Use Proposition~\ref{prop:upper:thetaPk} with $\eta_3=0.78$ of Theorem~\ref{th:eta_k} for $\ln p_k>27$. A computer verification concludes the proof.
\end{proof}
\begin{prop}
For $k\geqslant 3$,
$$p_k\geqslant k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-2.1}{\ln k}\right).$$
\end{prop}
\begin{proof}
Using (\ref{diff_theta_pk_pk}),
we have
$$p_k\geqslant \theta(p_k)-\eta_2\frac{k}{\ln k}.$$
By Proposition~\ref{min_theta_pk} and $\eta_2=0.04913$, we conclude the proof.
\end{proof}
\subsubsection{Smallest Interval containing primes}
We already know the result of
{\sc Schoenfeld} \cite{Schoenfeld:MathOfComp:Sharper} showing that,
for $x\geqs2010759.9$,
the interval $]x,x+x/16597[$ contains at least one prime.
We improve this result with the following proposition.
You can see also \cite{Ramare:JNT:short}.
\begin{prop}
For all $x\geqslant 396\,738$, there exists a prime $p$ such that
$$x<p\leqslant x\left(1+\frac{1}{25\ln^2 x}\right).$$
\end{prop}
This result is better that {\sc Rosser} \& {\sc Schoenfeld}'s one
for $x\geqslant e^{25.77}$. The method used in \cite{Ramare:JNT:short} gives better results (if we compare with the same order of $k$, i.e. $k=0$).
\begin{proof}
Let $0<f(x)<1$ for $x\geqslant x_0$.
\begin{eqnarray*}
\theta\left(\frac{1}{1-f(x)}x\right)-\theta(x)&\geqslant&\frac{1}{1-f(x)}x
-\eta_k\frac{\frac{x}{1-f(x)}}{\ln^k\left(\frac{x}{1-f(x)}\right)}
-\left(x+\eta_k\frac{x}{\ln^k x}\right)\\
&>&\left(\frac{1}{1-f(x)}-1\right)x-2\eta_k\left(\frac{1}{1-f(x)}\right)\frac{x}{\ln^k x}
\end{eqnarray*}
Choose $f(x)=\frac{2\eta_k}{\ln^k x}$ hence
$$\theta\left(\frac{1}{1-\frac{2\eta_k}{\ln^k x}}x\right)-\theta(x)>0.$$
For $k=2$, we have $n_2=0.0195$ and
$\frac{1}{1-2\cdot0.0195/\ln^2 x}\leqslant 1+1/(25\ln^2 x)$ for $\ln x\geqslant 28$.
According to \cite{Schoenfeld:MathOfComp:Sharper} p.~355,
$$p_{n+1}-p_n\leqslant 652\mbox{ for }p_n\leqs2.686\cdot10^{12},$$
hence the result is also valid from $x\geqslant 3.8\cdot 10^6$.
\end{proof}
\subsection{Estimates of function $\pi$
Remember that
$$\pi(x)=\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2}{\ln^2 x}
+O\left(\frac{1}{\ln^3 x}\right)\right).$$
\begin{theorem}
\label{Ross:th:pi}
\begin{eqnarray}
\label{pi_ordre1}
\frac{x}{\ln x}\left(1+\frac{1}{\ln x}\right)\leqslant&\pi(x)&\leqslant\frac{x}{\ln x}\left(1+\frac{1.2762}{\ln x}\right)\\
\nonumber \scriptstyle{ x\geqslant 599}&&{\scriptstyle x>1}\\
\nonumber\lefteqn{\mbox{(the value 1.2762 is chosen for }x=p_{258}=1627). }\\ %
\nonumber&&\\
\label{pi_ordre1_5}
\frac{x}{\ln x-1}\leqslant&\pi(x)&\leqslant\frac{x}{\ln x-1.1}\\
\nonumber \scriptstyle x\geqs5\,393&&\scriptstyle x\geqs60\,184\\
\nonumber&&\\
\label{pi_ordre2}\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2}{\ln^2 x}\right)\leqslant&
\pi(x)&\leqslant\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2.334}{\ln^2 x}\right)\\
\nonumber \scriptstyle x\geqslant 88\,783&&\scriptstyle x\geqs2\,953\,652\,287
\end{eqnarray}
\end{theorem}
\begin{proof}
We consider the last inequality.
Let
$$x_{0}=10^{11},\quad K=\pi(x_{0})-\frac{\theta(x_{0})}{\ln x_{0}}.$$
Write
$$J(x;\eta_k)=K+\frac{x}{\ln x}+\eta_k\frac{x}{\ln ^{k+1} x}
+\int_{x_{0}}^x \left(\frac{1}{\ln^2 y}+\frac{\eta_k}{\ln^{k+2} y}\right)dy$$
Since
$$\pi(x)=\pi(x_{0})-\frac{\theta(x_{0})}{\ln x_{0}}
+\frac{\theta(x)}{\ln x}+\int_{x_{0}}^{x}\frac{\theta(y)dy}{y\ln^2 y}$$
and $|\theta(x)-x|\leqslant \eta_k\frac{x}{\ln^k x}$ for $x\geqslant x_{0}$,
we have, for $x\geqslant x_{0}$,
$$J(x;-\eta_k)\leqslant \pi(x)\leqslant J(x;\eta_k).$$
Write $M(x;c)=\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{c}{\ln^2 x}
\right)$ for upper bound's function for $\pi(x)$.
Let's write the derivatives of $J(x;a)$ and of $M(x;c)$ with
respect to $x$:
$$J'(x;a)=\frac{1}{\ln x}+\frac{\eta_k}{\ln^{k+1} x}-k\frac{\eta_k}{\ln^{k+2} x},$$
$$M'(x;c)=\frac{1}{\ln x}+\frac{c-2}{\ln^3 x}-\frac{3c}{\ln^4 x}.$$
For $k=2$, we must choose $c\geqslant (2+\eta_2-2\eta_2/\ln x_0)/(1-3/\ln x_0)$ to have $J'<M'$ for $x\geqslant x_0$.
With $\eta_2=0.05$, we choose $c=2.321$. We verify by computer that
$J(10^{11};0.05)<M(10^{11};2.334)$.
By direct computation for small values of $x$ to obtain
$$\pi(x)<\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2.334}{\ln^2 x}
\right)\quad\mbox{ for }x\geqslant 2\,953\,652\,287.$$
Now write
$$m(x;d)=\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{d}{\ln^2 x}
\right).$$
We study the derivatives: we choose $k=3$, $d=2$ and $\eta_3(1-3/\ln x)<6$ to have $J'>m'$.
As $m(x_{0};2)<J(x_{0};-6)$ and by direct
computation for small values, we obtain
$$ \pi(x)>\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{2}{\ln^2 x}
\right)\quad\mbox{ for }x\geqslant 88\,783.$$
The others inequalities follows:
$(\ref{pi_ordre2})\Rightarrow(\ref{pi_ordre1_5})\Rightarrow(\ref{pi_ordre1})$ for large $x$.
\end{proof}
\subsection{Estimates of sums over primes
Let $\gamma$ be Euler's constant ($\gamma\approx0.5772157$).
\begin{theorem}
\label{Rosser:th:InvP}
Let $B= \gamma+
\sum_{p}\left(\ln(1-1/p)+1/p\right)\approx0.26149~72128~47643$.
For $x>1$,
$$-\left(\frac{1}{10\ln^2 x}+\frac{4}{15\ln^3x}\right)\leqslant\sum_{p\leqslant x}\frac{1}{p}-\ln_{2}x -B
.$$
For $x\geqs10372$,
$$\sum_{p\leqslant x}\frac{1}{p}-\ln_{2}x -B\leqslant\frac{1}{10\ln^2
x}+\frac{4}{15\ln^3x}.$$
\end{theorem}
\begin{proof}
By (4.20) of \cite{Rosser&Schoenfeld:Ill:Approx},
$$\sum_{p\leqslant x}\frac{1}{p}=\ln_{2}x +B+\frac{\theta(x)-x}{x\ln x}
-\int_{x}^\infty \frac{(\theta(y)-y)(1+\ln y)}{y^2\ln^2 y}dy.$$
Hence
$$|\sum_{p\leqslant x}\frac{1}{p}-\ln_{2}x -B|\leqslant
\frac{|\theta(x)-x|}{x\ln x}
+\int_{x}^\infty \frac{|\theta(y)-y|(1+\ln y)}{y^2\ln^2 y}dy.$$
As $|\theta(x)-x|\leqslant \eta_k x/\ln^k x$ (Theorem~\ref{th:eta_k}) and
$$\int_{x}^\infty\frac{1+\ln y}{y\ln^{k+2} y}dy=\frac{1}{k\ln^k x}
+\frac{1}{(k+1)\ln^{k+1} x},$$
we have the result
\begin{equation}
\label{eq:sumInvP}
\left|\sum_{p\leqslant x}\frac{1}{p}-\ln_{2}x -B\right|
\leqslant \frac{\eta_k/k}{\ln^k x}+\frac{\eta_k(1+\frac{1}{k+1})}{\ln^{k+1} x}.
\end{equation}
For $k=2$ and $\eta_2=0.2$, the result is valid for $x\geqslant 3594641$. We conclude by computer's check.
\end{proof}
\begin{theorem}
\label{Rosser:th:lnPInvP}
Let $E= -\gamma-\sum_{n=2}^{\infty}\sum_{p}(\ln p)/p^n
\approx-1.33258~22757~33221$.
For $x>1$,
$$-\left(\frac{0.2}{\ln x}+\frac{0.2}{\ln^2x}\right)
\leqslant\sum_{p\leqslant x}\frac{\ln p}{p}-\ln x -E.$$
For $x\geqslant 2974$,
$$\sum_{p\leqslant x}\frac{\ln p}{p}-\ln x -E\leqslant\frac{0.2}{\ln x}
+\frac{0.2}{\ln^2x}.$$
\end{theorem}
\begin{proof}
By (4.21) of \cite{Rosser&Schoenfeld:Ill:Approx},
$$\sum_{p\leqslant x}\frac{\ln p}{p}=\ln x +E+\frac{\theta(x)-x}{x}
-\int_{x}^\infty \frac{\theta(y)-y}{y^2}dy.$$
Hence
$$|\sum_{p\leqslant x}\frac{\ln p}{p}-\ln x -E|\leqslant\frac{|\theta(x)-x|}{x}
+\int_{x}^\infty \frac{|\theta(y)-y|}{y^2}dy.$$
As
$$\int_{x}^\infty\frac{dy}{y\ln^k y}=\frac{1}{(k-1)\ln^{k-1} x},$$
Theorem~\ref{th:eta_k} yields the result for $x\geqslant 3594641$ with $k=2$.
We conclude by computer's check.
\end{proof}
\subsection{Estimates of products over primes}
\begin{theorem}
\label{Rosser:th:ProdP}
For $x>1$,
$$\prod_{p\leqslant x}\left(1-\frac{1}{p}\right)<\frac{e^{-\gamma}}{\ln x}
\left(1+\frac{0.2}{\ln^2 x}\right)$$
and for $x\geqslant 2\,973$,
$$\frac{e^{-\gamma}}{\ln x}
\left(1-\frac{0.2}{\ln^2 x}\right)
<\prod_{p\leqslant x}\left(1-\frac{1}{p}\right)
$$
For $x>1$,
$${e^{\gamma}}{\ln x}
\left(1-\frac{0.2}{\ln^2 x}\right)<\prod_{p\leqslant x}\frac{p}{p-1}.
$$
and for $x\geqslant 2\,973$,
$$\prod_{p\leqslant x}\frac{p}{p-1}<{e^{\gamma}}{\ln x}
\left(1+\frac{0.2}{\ln^2 x}\right).$$
\end{theorem}
\begin{proof}
By definition of $B$ and (\ref{eq:sumInvP}), we have
$$\left|-\gamma-\ln_{2}x-\sum_{p>x}\frac{1}{p}-\sum_{p}\ln(1-1/p)\right|
\leqslant\frac{\eta_k/k}{\ln^k x}+\frac{\eta_k(1+\frac{1}{k+1})}{\ln^{k+1} x}.$$
Let
$S=\sum_{p>x}\left(\ln(1-1/p)+1/p\right)=
-\sum_{n=2}^\infty\frac{1}{n}\sum_{p>x}\frac{1}{p^n}.$
We have
$$-\gamma-\ln_{2}x-\sum_{p\leqslant x}\ln(1-1/p)-S
\geqslant-\frac{\eta_k}{k\ln^k x}-\frac{(k+2)\eta_k}{(k+1)\ln^{k+1} x}.$$
Take the exponential of both sides to obtain
$$\prod_{p\leqslant x}\left(1-\frac{1}{p}\right)\leqslant
\frac{e^{-\gamma}}{\ln x}\exp\left( -S+\frac{\eta_k}{k\ln^k x}
+\frac{(k+2)\eta_k}{(k+1)\ln^{k+1} x}\right).$$
We use lower bound for S given in \cite{Rosser&Schoenfeld:Ill:Approx}
p.~87:
$$-S<\frac{1.02}{(x-1)\ln x}.$$
Hence, for $k=2$, $\eta_2=0.2$ and $x\geqslant 3\,594\,641$,
$$\prod_{p\leqslant x}\left(1-\frac{1}{p}\right)\leqslant
\frac{e^{-\gamma}}{\ln x}\exp(0.11/\ln^2 x).$$
We have also
$$\prod_{p\leqslant x}\frac{p-1}{p}\geqslant
{e^{\gamma}}{\ln x}\exp(-0.11/\ln^2 x).$$
In the same way, as
$$-\gamma-\ln_{2}x-\sum_{p\leqslant x}\ln(1-1/p)-S
\leqslant\frac{\eta_k}{k\ln^k x}+\frac{(k+2)\eta_k}{(k+1)\ln^{k+1} x},$$
we obtain the others inequalities since $S\leqs0$.
\end{proof}
\begin{table}[!hp]
\caption{Values of $\theta(x)$ for $10^{6} \leq x \leq 10^{10}$}
\protect{\label{table_theta1}}
\begin{center}
\normalsize
\begin{tabular}{||c|r|r||} \hline
x & $\theta(x)$ & $\psi(x)-\theta(x)$ \\
\hline
$ 1E+06 $ & $998484.175026 $ & $ 1102.422470 $\\
$ 2E+06 $ & $1998587.722137 $ & $ 1527.324070 $\\
$ 3E+06 $ & $2998107.530452 $ & $ 1892.449541 $\\
$ 4E+06 $ & $3997323.492084 $ & $ 2167.364713 $\\
$ 5E+06 $ & $4998571.086801 $ & $ 2400.053428 $\\
$ 6E+06 $ & $5996983.791998 $ & $ 2665.785692 $\\
$ 7E+06 $ & $6997751.998535 $ & $ 2823.187880 $\\
$ 8E+06 $ & $7997057.246292 $ & $ 3064.486910 $\\
$ 9E+06 $ & $8997625.570065 $ & $ 3224.678815 $\\
$ 1E+07 $ & $9995179.317856 $ & $ 3360.085490 $\\
$ 2E+07 $ & $19995840.882153 $ & $ 4759.143006 $\\
$ 3E+07 $ & $29994907.240152 $ & $ 5797.041942 $\\
$ 4E+07 $ & $39994781.014188 $ & $ 6699.200805 $\\
$ 5E+07 $ & $49993717.861720 $ & $ 7489.482783 $\\
$ 6E+07 $ & $59991136.134174 $ & $ 8172.843038 $\\
$ 7E+07 $ & $69991996.348980 $ & $ 8786.853393 $\\
$ 8E+07 $ & $79988578.197461 $ & $ 9388.261229 $\\
$ 9E+07 $ & $89985867.940581 $ & $ 9992.336337 $\\
$ 1E+08 $ & $99987730.018022 $ & $ 10512.778605 $\\
$ 2E+08 $ & $199982302.435783 $ & $ 14725.068769 $\\
$ 3E+08 $ & $299981378.219200 $ & $ 18000.443659 $\\
$ 4E+08 $ & $399982033.338736 $ & $ 20744.718991 $\\
$ 5E+08 $ & $499983789.813730 $ & $ 23200.125087 $\\
$ 6E+08 $ & $599976282.577668 $ & $ 25426.013243 $\\
$ 7E+08 $ & $699976911.639135 $ & $ 27402.910397 $\\
$ 8E+08 $ & $799969331.209833 $ & $ 29215.380561 $\\
$ 9E+08 $ & $899953849.181850 $ & $ 30963.754721 $\\
$ 1E+09 $ & $999968978.577566 $ & $ 32617.412861 $\\
$ 2E+09 $ & $1999941083.684486 $ & $ 46075.813369 $\\
$ 3E+09 $ & $2999937036.966284 $ & $ 56255.144708 $\\
$ 4E+09 $ & $3999946136.165586 $ & $ 64858.831531 $\\
$ 5E+09 $ & $4999906575.362844 $ & $ 72411.275590 $\\
$ 6E+09 $ & $5999930311.133705 $ & $ 79301.775139 $\\
$ 7E+09 $ & $6999917442.519773 $ & $ 85715.065356 $\\
$ 8E+09 $ & $7999890792.693956 $ & $ 91420.172461 $\\
$ 9E+09 $ & $8999894889.497541 $ & $ 97066.566501 $\\
$ 1E+10 $ & $9999939830.657757 $ & $ 102289.175716 $\\
\hline
\end{tabular}
\end{center}
\end{table}
\pagebreak
\begin{table}[!hp]
\caption{Values of $\theta(x)$ for $10^{10} \leq x \leq 10^{15}$}
\protect{\label{table_theta2}}
\begin{center}
\normalsize
\begin{tabular}{||c|r|r||} \hline
x & $\theta(x)$ & $\psi(x)-\theta(x)$ \\
\hline
$ 1E+10 $ & $9999939830.657757 $ & $ 102289.175716 $\\
$ 2E+10 $ & $19999821762.768212 $ & $ 144339.622582 $\\
$ 3E+10 $ & $29999772119.815419 $ & $ 176300.955450 $\\
$ 4E+10 $ & $39999808348.775748 $ & $ 203538.541084 $\\
$ 5E+10 $ & $49999728380.731899 $ & $ 227474.729168 $\\
$ 6E+10 $ & $59999772577.550769 $ & $ 249003.320704 $\\
$ 7E+10 $ & $69999769944.203933 $ & $ 268660.720820 $\\
$ 8E+10 $ & $79999718357.195652 $ & $ 287365.266118 $\\
$ 9E+10 $ & $89999644656.090911 $ & $ 304250.688854 $\\
$ 1E+11 $ & $99999737653.107445 $ & $ 320803.322857 $\\
$ 2E+11 $ & $199999695484.246439 $ & $ 453289.609568 $\\
$ 3E+11 $ & $299999423179.995211 $ & $ 554528.646163 $\\
$ 4E+11 $ & $399999101196.308601 $ & $ 640000.361434 $\\
$ 5E+11 $ & $499999105742.583455 $ & $ 715211.001138 $\\
$ 6E+11 $ & $599999250571.436655 $ & $ 783167.715577 $\\
$ 7E+11 $ & $699998999499.845475 $ & $ 845911.916175 $\\
$ 8E+11 $ & $799999133776.084743 $ & $ 904203.190001 $\\
$ 9E+11 $ & $899998818628.952024 $ & $ 958602.924046 $\\
$ 1E+12 $ & $999999030333.096225 $ & $ 1009803.669232 $\\
$ 2E+12 $ & $1999998755521.470649 $ & $ 1427105.865316 $\\
$ 3E+12 $ & $2999997819758.987859 $ & $ 1746299.820370 $\\
$ 4E+12 $ & $3999998370195.717561 $ & $ 2016279.693623 $\\
$ 5E+12 $ & $4999998073643.711478 $ & $ 2253672.042145 $\\
$ 6E+12 $ & $5999997276726.877147 $ & $ 2467566.593710 $\\
$ 7E+12 $ & $6999996936360.165729 $ & $ 2665065.541181 $\\
$ 8E+12 $ & $7999997864671.383505 $ & $ 2848858.049155 $\\
$ 9E+12 $ & $8999996425300.244577 $ & $ 3021079.319393 $\\
$ 1E+13 $ & $9999996988293.034200 $ & $ 3183704.089025 $\\
$ 2E+13 $ & $19999995126082.228688 $ & $ 4499685.436490 $\\
$ 3E+13 $ & $29999995531389.845427 $ & $ 5509328.368277 $\\
$ 4E+13 $ & $39999993533724.316829 $ & $ 6359550.652121 $\\
$ 5E+13 $ & $49999992543194.263655 $ & $ 7109130.001413 $\\
$ 6E+13 $ & $59999990297033.626198 $ & $ 7785491.725387 $\\
$ 7E+13 $ & $69999994316409.871731 $ & $ 8407960.376833 $\\
$ 8E+13 $ & $79999990160858.304239 $ & $ 8988688.375101 $\\
$ 9E+13 $ & $89999989501395.073897 $ & $ 9531798.550749 $\\
$ 1E+14 $ & $99999990573246.978538 $ & $ 10045400.569463 $\\
$ 2E+14 $ & $199999983475767.543204 $ & $ 14201359.711421 $\\
$ 3E+14 $ & $299999986702246.281944 $ & $ 17388356.540338 $\\
$ 4E+14 $ & $399999982296901.085038 $ & $ 20074942.600622 $\\
$ 5E+14 $ & $499999974019856.236519 $ & $ 22439658.012185 $\\
$ 6E+14 $ & $599999983610646.997632 $ & $ 24580138.242324 $\\
$ 7E+14 $ & $699999971887332.157455 $ & $ 26545816.027402 $\\
$ 8E+14 $ & $799999964680836.091645 $ & $ 28378339.693784 $\\
$ 9E+14 $ & $899999961386694.231242 $ & $ 30098146.961102 $\\
$ 1E+15 $ & $999999965752660.939840$ & $31724269.567843$\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!hp]
\caption{Values of $\epsilon(x)$ for $\psi$ and $\theta$}
\label{table_eps}
\normalsize
$$\begin{array}{||l|l|l||l|l|l||}
\hline
b&\epsilon_\psi&\epsilon_\theta & b &\epsilon_\psi&\epsilon_\theta\\
\hline
20 & 6.123 E-4 & 6.606 E-4 & 100 & 2.903 E-11 & 2.903 E-11 \\
21 & 4.072 E-4 & 4.363 E-4 & 200 & 2.838 E-11 & 2.838 E-11 \\
22 & 2.706 E-4 & 2.881 E-4 & 300 & 2.772 E-11 & 2.772 E-11 \\
23 & 1.792 E-4 & 1.897 E-4 & 400 & 2.706 E-11 & 2.706 E-11 \\
24 & 1.183 E-4 & 1.247 E-4 & 500 & 2.641 E-11 & 2.641 E-11 \\
25 & 7.789 E-5 & 8.172 E-5 & 600 & 2.575 E-11 & 2.575 E-11 \\
\ln(10^{11})& 6.788 E-5 & 7.112 E-5 & 1000 & 2.315 E-11 & 2.315 E-11 \\
26 & 5.121 E-5 & 5.352 E-5 & 1250 & 2.153 E-11 & 2.153 E-11 \\
27 & 3.368 E-5 & 3.507 E-5 & 1500 & 1.991 E-11 & 1.991 E-11 \\
28 & 2.224 E-5 & 2.308 E-5 & 2000 & 1.671 E-11 & 1.671 E-11 \\
29 & 1.451 E-5 & 1.502 E-5 & 2200 & 1.544 E-11 & 1.544 E-11 \\
30 & 9.414 E-6 & 9.724 E-6 & 2500 & 1.355 E-11 & 1.355 E-11 \\
31 & 6.099 E-6 & 6.287 E-6 & 2800 & 1.169 E-11 & 1.169 E-11 \\
32 & 3.944 E-6 & 4.057 E-6 & 3000 & 1.047 E-11 & 1.047 E-11 \\
33 & 2.545 E-6 & 2.614 E-6 & 3200 & 9.267 E-12 & 9.267 E-12 \\
34 & 1.640 E-6 & 1.682 E-6 & 3300 & 8.658 E-12 & 8.658 E-12 \\
\ln(10^{15}) & 1.293 E-6 & 1.325 E-6 & 3400 & 8.083 E-12 & 8.083 E-12 \\
35 & 1.055 E-6 & 1.080 E-6 & 3455 & 7.750 E-12 & 7.750 E-12 \\
36 & 6.775 E-7 & 6.928 E-7 & 3500 & 7.488 E-12 & 7.488 E-12 \\
37 & 4.348 E-7 & 4.441 E-7 & 3600 & 6.930 E-12 & 6.930 E-12 \\
38 & 2.793 E-7 & 2.849 E-7 & 3700 & 6.351 E-12 & 6.351 E-12 \\
39 & 1.805 E-7 & 1.839 E-7 & 3750 & 6.080 E-12 & 6.080 E-12 \\
40 & 1.163 E-7 & 1.184 E-7 & 3800 & 5.821 E-12 & 5.821 E-12 \\
41 & 7.414 E-8 & 7.539 E-8 & 3850 & 5.533 E-12 & 5.533 E-12 \\
42 & 4.723 E-8 & 4.799 E-8 & 3900 & 5.259 E-12 & 5.259 E-12 \\
43 & 3.011 E-8 & 3.057 E-8 & 3950 & 4.999 E-12 & 4.999 E-12 \\
44 & 1.932 E-8 & 1.960 E-8 & 4000 & 4.751 E-12 & 4.751 E-12 \\
45 & 1.234 E-8 & 1.251 E-8 & 4050 & 4.496 E-12 & 4.496 E-12 \\
46 & 7.839 E-9 & 7.941 E-9 & 4100 & 4.231 E-12 & 4.231 E-12 \\
47 & 5.026 E-9 & 5.088 E-9 & 4150 & 3.981 E-12 & 3.981 E-12 \\
48 & 3.190 E-9 & 3.228 E-9 & 4200 & 3.746 E-12 & 3.746 E-12 \\
49 & 2.038 E-9 & 2.061 E-9 & 4300 & 3.308 E-12 & 3.308 E-12 \\
50 & 1.301 E-9 & 1.315 E-9 & 4400 & 2.844 E-12 & 2.844 E-12 \\
55 & 1.481 E-10 & 1.492 E-10 & 4500 & 2.445 E-12 & 2.445 E-12 \\
60 & 3.917 E-11 & 3.926 E-11 & 4700 & 1.774 E-12 & 1.774 E-12 \\
70 & 2.929 E-11 & 2.929 E-11 & 5000 & 9.562 E-13 & 9.562 E-13 \\
75 & 2.920 E-11 & 2.920 E-11 & 10000 & 6.341 E-18 & 6.341 E-18 \\
\hline
\end{array}$$
\end{table}
\begin{table}[!hp]
\caption{ Values of $\eta_k$ valid for $\exp(b_i)\leqslant x\leqslant \exp(b_{i+1})$.}
\protect{\label{table_eta1}}
\normalsize
$$\begin{array}{||r|r|r|r|r||}
\hline
b_i&\eta_1&\eta_2&\eta_3&\eta_4\\
\hline
20 & 1.388 E-2 & 2.914 E-1 & 6.118 E+0 & 1.285 E+2 \\
21 & 9.597 E-3 & 2.112 E-1 & 4.645 E+0 & 1.022 E+2 \\
22 & 6.625 E-3 & 1.524 E-1 & 3.505 E+0 & 8.061 E+1 \\
23 & 4.553 E-3 & 1.093 E-1 & 2.623 E+0 & 6.294 E+1 \\
24 & 3.116 E-3 & 7.790 E-2 & 1.948 E+0 & 4.869 E+1 \\
25 & 2.070 E-3 & 5.243 E-2 & 1.328 E+0 & 3.364 E+1 \\
\ln(10^{11}) & 1.849 E-3 & 4.808 E-2 & 1.250 E+0 & 3.250 E+1 \\
26 & 1.445 E-3 & 3.902 E-2 & 1.054 E+0 & 2.845 E+1 \\
27 & 9.820 E-4 & 2.750 E-2 & 7.699 E-1 & 2.156 E+1 \\
28 & 6.693 E-4 & 1.941 E-2 & 5.629 E-1 & 1.633 E+1 \\
29 & 4.504 E-4 & 1.352 E-2 & 4.054 E-1 & 1.216 E+1 \\
30 & 3.015 E-4 & 9.344 E-3 & 2.897 E-1 & 8.980 E+0 \\
31 & 2.012 E-4 & 6.437 E-3 & 2.060 E-1 & 6.592 E+0 \\
32 & 1.339 E-4 & 4.418 E-3 & 1.458 E-1 & 4.811 E+0 \\
33 & 8.887 E-5 & 3.022 E-3 & 1.028 E-1 & 3.493 E+0 \\
34 & 5.807 E-5 & 2.006 E-3 & 6.928 E-2 & 2.393 E+0 \\
\ln(10^{15}) & 4.637 E-5 & 1.623 E-3 & 5.680 E-2 & 1.988 E+0 \\
35 & 3.888 E-5 & 1.400 E-3 & 5.039 E-2 & 1.814 E+0 \\
36 & 2.564 E-5 & 9.484 E-4 & 3.509 E-2 & 1.299 E+0 \\
37 & 1.688 E-5 & 6.412 E-4 & 2.437 E-2 & 9.259 E-1 \\
38 & 1.112 E-5 & 4.333 E-4 & 1.690 E-2 & 6.591 E-1 \\
39 & 7.354 E-6 & 2.942 E-4 & 1.177 E-2 & 4.707 E-1 \\
40 & 4.853 E-6 & 1.990 E-4 & 8.157 E-3 & 3.345 E-1 \\
41 & 3.167 E-6 & 1.330 E-4 & 5.586 E-3 & 2.346 E-1 \\
42 & 2.064 E-6 & 8.872 E-5 & 3.815 E-3 & 1.641 E-1 \\
43 & 1.345 E-6 & 5.918 E-5 & 2.604 E-3 & 1.146 E-1 \\
44 & 8.818 E-7 & 3.968 E-5 & 1.786 E-3 & 8.036 E-2 \\
45 & 5.752 E-7 & 2.646 E-5 & 1.218 E-3 & 5.599 E-2 \\
46 & 3.733 E-7 & 1.755 E-5 & 8.245 E-4 & 3.875 E-2 \\
47 & 2.442 E-7 & 1.173 E-5 & 5.627 E-4 & 2.701 E-2 \\
48 & 1.582 E-7 & 7.749 E-6 & 3.797 E-4 & 1.861 E-2 \\
49 & 1.031 E-7 & 5.151 E-6 & 2.576 E-4 & 1.288 E-2 \\
50 & 7.229 E-8 & 3.976 E-6 & 2.187 E-4 & 1.203 E-2 \\
55 & 8.952 E-9 & 5.371 E-7 & 3.223 E-5 & 1.934 E-3 \\
60 & 2.748 E-9 & 1.924 E-7 & 1.347 E-5 & 9.425 E-4 \\
70 & 2.197 E-9 & 1.648 E-7 & 1.236 E-5 & 9.268 E-4 \\
75 & 2.920 E-9 & 2.920 E-7 & 2.920 E-5 & 2.920 E-3 \\
\hline
\end{array}$$
\end{table}
\begin{table}[!hp]
\caption{Values of $\eta_k$ (continued)}
\label{table_eta2}
\normalsize
$$\begin{array}{||r|r|r|r|r||}
\hline
b_i&\eta_1&\eta_2&\eta_3&\eta_4\\
\hline
100 & 5.805 E-9 & 1.161 E-6 & 2.322 E-4 & 4.644 E-2 \\
200 & 8.512 E-9 & 2.554 E-6 & 7.661 E-4 & 2.299 E-1 \\
300 & 1.109 E-8 & 4.434 E-6 & 1.774 E-3 & 7.094 E-1 \\
400 & 1.353 E-8 & 6.765 E-6 & 3.383 E-3 & 1.692 E+0 \\
500 & 1.585 E-8 & 9.505 E-6 & 5.703 E-3 & 3.422 E+0 \\
600 & 2.575 E-8 & 2.575 E-5 & 2.575 E-2 & 2.575 E+1 \\
1000 & 2.893 E-8 & 3.616 E-5 & 4.520 E-2 & 5.650 E+1 \\
1250 & 3.229 E-8 & 4.843 E-5 & 7.265 E-2 & 1.090 E+2 \\
1500 & 3.982 E-8 & 7.963 E-5 & 1.593 E-1 & 3.185 E+2 \\
2000 & 3.675 E-8 & 8.084 E-5 & 1.779 E-1 & 3.913 E+2 \\
2200 & 3.859 E-8 & 9.646 E-5 & 2.412 E-1 & 6.029 E+2 \\
2500 & 3.794 E-8 & 1.063 E-4 & 2.975 E-1 & 8.328 E+2 \\
2800 & 3.507 E-8 & 1.053 E-4 & 3.157 E-1 & 9.469 E+2 \\
3000 & 3.351 E-8 & 1.073 E-4 & 3.431 E-1 & 1.098 E+3 \\
3200 & 3.058 E-8 & 1.010 E-4 & 3.331 E-1 & 1.099 E+3 \\
3300 & 2.944 E-8 & 1.001 E-4 & 3.403 E-1 & 1.157 E+3 \\
3400 & 2.793 E-8 & 9.648 E-5 & 3.334 E-1 & 1.152 E+3 \\
3455 & 2.713 E-8 & 9.494 E-5 & 3.323 E-1 & 1.163 E+3 \\
3500 & 2.696 E-8 & 9.704 E-5 & 3.494 E-1 & 1.258 E+3 \\
3600 & 2.565 E-8 & 9.488 E-5 & 3.511 E-1 & 1.299 E+3 \\
3700 & 2.382 E-8 & 8.931 E-5 & 3.350 E-1 & 1.256 E+3 \\
3750 & 2.311 E-8 & 8.780 E-5 & 3.337 E-1 & 1.268 E+3 \\
3800 & 2.241 E-8 & 8.628 E-5 & 3.322 E-1 & 1.279 E+3 \\
3850 & 2.158 E-8 & 8.416 E-5 & 3.282 E-1 & 1.280 E+3 \\
3900 & 2.078 E-8 & 8.205 E-5 & 3.241 E-1 & 1.281 E+3 \\
3950 & 2.000 E-8 & 7.997 E-5 & 3.199 E-1 & 1.280 E+3 \\
4000 & 1.924 E-8 & 7.793 E-5 & 3.156 E-1 & 1.279 E+3 \\
4050 & 1.844 E-8 & 7.557 E-5 & 3.099 E-1 & 1.271 E+3 \\
4100 & 1.756 E-8 & 7.286 E-5 & 3.024 E-1 & 1.255 E+3 \\
4150 & 1.672 E-8 & 7.022 E-5 & 2.950 E-1 & 1.239 E+3 \\
4200 & 1.611 E-8 & 6.927 E-5 & 2.979 E-1 & 1.281 E+3 \\
4300 & 1.456 E-8 & 6.404 E-5 & 2.818 E-1 & 1.240 E+3 \\
4400 & 1.280 E-8 & 5.759 E-5 & 2.592 E-1 & 1.167 E+3 \\
4500 & 1.150 E-8 & 5.401 E-5 & 2.539 E-1 & 1.194 E+3 \\
4700 & 8.868 E-9 & 4.434 E-5 & 2.217 E-1 & 1.109 E+3 \\
\hline
\end{array}$$
\end{table}
\begin{table}[!hp]
\label{table_012}
\caption{Values for $\theta(x)$}
\normalsize
$$
\begin{array}{||l||l|l||l|l||l|l||}
\hline
& a_0 & b_0 & a_1 & b_1 & a_2 & b_2 \\
\hline
1E+08 & 0.99985 & 0.99998 & 0.00275 & -0.00044 & 0.05062 & -0.00851 \\
2E+08 & 0.99989 & 0.99997 & 0.00201 & -0.00065 & 0.03847 & -0.01275 \\
3E+08 & 0.99991 & 0.99998 & 0.00165 & -0.00057 & 0.03256 & -0.01131 \\
4E+08 & 0.99993 & 0.99998 & 0.00124 & -0.00049 & 0.02447 & -0.00988 \\
5E+08 & 0.99992 & 0.99998 & 0.00152 & -0.00052 & 0.03039 & -0.01061 \\
6E+08 & 0.99994 & 0.99999 & 0.00103 & -0.00038 & 0.02095 & -0.00785 \\
7E+08 & 0.99993 & 0.99998 & 0.00126 & -0.00051 & 0.02577 & -0.01054 \\
8E+08 & 0.99994 & 0.99998 & 0.00110 & -0.00044 & 0.02268 & -0.00916 \\
9E+08 & 0.99994 & 0.99998 & 0.00117 & -0.00050 & 0.02398 & -0.01050 \\
1E+09 & 0.99995 & 0.99999 & 0.00096 & -0.00021 & 0.02009 & -0.00455 \\
2E+09 & 0.99996 & 1.00000 & 0.00067 & -0.00018 & 0.01427 & -0.00401 \\
3E+09 & 0.99997 & 1.00000 & 0.00062 & -0.00015 & 0.01340 & -0.00333 \\
4E+09 & 0.99997 & 1.00000 & 0.00050 & -0.00017 & 0.01087 & -0.00390 \\
5E+09 & 0.99997 & 1.00000 & 0.00046 & -0.00018 & 0.01026 & -0.00410 \\
6E+09 & 0.99998 & 1.00000 & 0.00040 & -0.00013 & 0.00900 & -0.00313 \\
7E+09 & 0.99998 & 1.00000 & 0.00045 & -0.00018 & 0.01011 & -0.00415 \\
8E+09 & 0.99998 & 1.00000 & 0.00034 & -0.00016 & 0.00774 & -0.00367 \\
9E+09 & 0.99998 & 1.00000 & 0.00037 & -0.00010 & 0.00840 & -0.00249 \\
1E+10 & 0.99998 & 1.00000 & 0.00038 & -0.00008 & 0.00876 & -0.00203 \\
2E+10 & 0.99998 & 1.00000 & 0.00025 & -0.00006 & 0.00584 & -0.00152 \\
3E+10 & 0.99999 & 1.00000 & 0.00020 & -0.00005 & 0.00473 & -0.00122 \\
4E+10 & 0.99999 & 1.00000 & 0.00018 & -0.00007 & 0.00431 & -0.00183 \\
5E+10 & 0.99999 & 1.00000 & 0.00019 & -0.00004 & 0.00458 & -0.00119 \\
6E+10 & 0.99999 & 1.00000 & 0.00015 & -0.00006 & 0.00356 & -0.00167 \\
7E+10 & 0.99999 & 1.00000 & 0.00014 & -0.00004 & 0.00338 & -0.00117 \\
8E+10 & 0.99999 & 1.00000 & 0.00011 & -0.00006 & 0.00276 & -0.00164 \\
9E+10 & 0.99999 & 1.00000 & 0.00011 & -0.00004 & 0.00262 & -0.00114 \\
1E+11 & 0.99999 & 1.00000 & 0.00014 & -0.00002 & 0.00341 & -0.00058 \\
2E+11 & 0.99999 & 1.00000 & 0.00008 & -0.00002 & 0.00206 & -0.00057 \\
3E+11 & 0.99999 & 1.00000 & 0.00007 & -0.00001 & 0.00170 & -0.00031 \\
4E+11 & 0.99999 & 1.00000 & 0.00008 & -0.00002 & 0.00188 & -0.00061 \\
5E+11 & 0.99999 & 1.00000 & 0.00006 & -0.00001 & 0.00138 & -0.00038 \\
6E+11 & 0.99999 & 1.00000 & 0.00005 & -0.00002 & 0.00131 & -0.00070 \\
7E+11 & 0.99999 & 1.00000 & 0.00006 & -0.00001 & 0.00144 & -0.00039 \\
\hline
\end{array}
$$
where the constants satisfy for $n\cdot10^k\leqslant x\leqslant (n+1)\cdot10^k$
$$
\begin{array}{rcl}
a_0 x&\leqslant \theta(x)\leqslant& b_0 x\\
x-a_1 \frac{x}{\ln x}&\leqslant\theta(x)\leqslant& x+b_1 \frac{x}{\ln x}\\
x-a_2\frac{x}{\ln^2 x}&\leqslant\theta(x) \leqslant& x+b_2\frac{x}{\ln^2 x}
\end{array}
$$
up to $8\cdot 10^{11}$.
\end{table}
\begin{table}[!hp]
\label{table_834}
\caption{Values for $p_k$ and $\theta(p_k)$}
\normalsize
$$
\begin{array}{||l||l|l||l|l||l|l||}
\hline
& a_8 & b_8 & a_3 & b_3 & a_4 & b_4 \\
\hline
1E+08 & 2.07947 & 2.07516 & 0.95665 & 0.95433 & 2.07207 & 2.03783 \\
2E+08 & 2.07517 & 2.07280 & 0.95517 & 0.95405 & 2.06330 & 2.04341 \\
3E+08 & 2.07281 & 2.07122 & 0.95493 & 0.95379 & 2.06236 & 2.04535 \\
4E+08 & 2.07123 & 2.07005 & 0.95459 & 0.95403 & 2.06210 & 2.05123 \\
5E+08 & 2.07006 & 2.06909 & 0.95448 & 0.95357 & 2.06053 & 2.04534 \\
6E+08 & 2.06910 & 2.06833 & 0.95448 & 0.95374 & 2.06271 & 2.05150 \\
7E+08 & 2.06834 & 2.06767 & 0.95421 & 0.95342 & 2.05976 & 2.04701 \\
8E+08 & 2.06768 & 2.06710 & 0.95411 & 0.95342 & 2.05999 & 2.04871 \\
9E+08 & 2.06711 & 2.06660 & 0.95395 & 0.95336 & 2.05808 & 2.04762 \\
1E+09 & 2.06661 & 2.06350 & 0.95409 & 0.95319 & 2.06243 & 2.04925 \\
2E+09 & 2.06351 & 2.06183 & 0.95355 & 0.95311 & 2.05913 & 2.05126 \\
3E+09 & 2.06184 & 2.06070 & 0.95336 & 0.95297 & 2.05821 & 2.05036 \\
4E+09 & 2.06071 & 2.05985 & 0.95322 & 0.95295 & 2.05684 & 2.05159 \\
5E+09 & 2.05986 & 2.05917 & 0.95314 & 0.95288 & 2.05600 & 2.05103 \\
6E+09 & 2.05918 & 2.05862 & 0.95313 & 0.95287 & 2.05643 & 2.05149 \\
7E+09 & 2.05863 & 2.05815 & 0.95305 & 0.95276 & 2.05517 & 2.04994 \\
8E+09 & 2.05816 & 2.05774 & 0.95300 & 0.95283 & 2.05489 & 2.05168 \\
9E+09 & 2.05775 & 2.05739 & 0.95300 & 0.95277 & 2.05540 & 2.05073 \\
1E+10 & 2.05740 & 2.05515 & 0.95301 & 0.95264 & 2.05571 & 2.04968 \\
2E+10 & 2.05516 & 2.05395 & 0.95283 & 0.95263 & 2.05364 & 2.04983 \\
3E+10 & 2.05396 & 2.05313 & 0.95275 & 0.95262 & 2.05225 & 2.04964 \\
4E+10 & 2.05314 & 2.05252 & 0.95272 & 0.95260 & 2.05143 & 2.04899 \\
5E+10 & 2.05253 & 2.05203 & 0.95272 & 0.95258 & 2.05127 & 2.04869 \\
6E+10 & 2.05204 & 2.05163 & 0.95269 & 0.95260 & 2.05060 & 2.04875 \\
7E+10 & 2.05164 & 2.05129 & 0.95270 & 0.95260 & 2.05060 & 2.04865 \\
8E+10 & 2.05130 & 2.05099 & 0.95267 & 0.95262 & 2.04978 & 2.04877 \\
9E+10 & 2.05100 & 2.05073 & 0.95269 & 0.95262 & 2.04982 & 2.04879 \\
1E+11 & 2.05074 & 2.04910 & 0.95271 & 0.95259 & 2.05014 & 2.04734 \\
2E+11 & 2.04911 & 2.04821 & 0.95273 & 0.95266 & 2.04851 & 2.04668 \\
3E+11 & 2.04822 & 2.04761 & 0.95276 & 0.95270 & 2.04775 & 2.04621 \\
4E+11 & 2.04762 & 2.04716 & 0.95278 & 0.95271 & 2.04692 & 2.04601 \\
5E+11 & 2.04717 & 2.04680 & 0.95282 & 0.95276 & 2.04670 & 2.04579 \\
6E+11 & 2.04681 & 2.04651 & 0.95283 & 0.95279 & 2.04617 & 2.04543 \\
7E+11 & 2.04652 & 2.04625 & 0.95286 & 0.95280 & 2.04597 & 2.04524 \\
\hline
\end{array}$$
where the constants satisfy for $n\cdot10^m\leqslant p_k\leqslant (n+1)\cdot10^m$
$$\begin{array}{rcl}
k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-a_8}{\ln k}\right)&\leqslant \theta(p_k)
\leqslant & k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-b_8}{\ln k}\right)\\
k\left(\ln k+\ln_2 k-a_3\right)&\leqslant p_k\leqslant& k\left(\ln k+\ln_2 k-b_3\right)\\
k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-a_4}{\ln k}\right)&\leqslant p_k\leqslant& k\left(\ln k+\ln_2 k-1+\frac{\ln_2 k-b_4}{\ln k}\right)\\
\end{array}$$
up to $8\cdot 10^{11}$.
\end{table}
\begin{table}[!hp]
\label{table_567}
\caption{Values for $\pi(x)$}
\normalsize
$$
\begin{array}{||l||l|l||l|l||l|l||}
\hline
& a_5 & b_5 & a_6 & b_6 & a_7 & b_7 \\
\hline
1E+08 & 1.12379 & 1.13015 & 2.36474 & 2.40986 & 1.06139 & 1.06514\\
2E+08 & 1.12113 & 1.12429 & 2.35944 & 2.38213 & 1.06022 & 1.06184\\
3E+08 & 1.11922 & 1.12158 & 2.35351 & 2.37715 & 1.05917 & 1.06074\\
4E+08 & 1.11802 & 1.11954 & 2.35815 & 2.36977 & 1.05878 & 1.05964\\
5E+08 & 1.11642 & 1.11818 & 2.34796 & 2.36727 & 1.05793 & 1.05906\\
6E+08 & 1.11556 & 1.11706 & 2.35087 & 2.36678 & 1.05756 & 1.05858\\
7E+08 & 1.11455 & 1.11587 & 2.34345 & 2.36058 & 1.05696 & 1.05793\\
8E+08 & 1.11363 & 1.11492 & 2.34295 & 2.35791 & 1.05656 & 1.05745\\
9E+08 & 1.11327 & 1.11405 & 2.34153 & 2.35399 & 1.05647 & 1.05700\\
1E+09 & 1.10903 & 1.11346 & 2.33374 & 2.35650 & 1.05441 & 1.05673\\
2E+09 & 1.10679 & 1.10928 & 2.33043 & 2.34118 & 1.05336 & 1.05467\\
3E+09 & 1.10527 & 1.10683 & 2.32501 & 2.33314 & 1.05265 & 1.05342\\
4E+09 & 1.10395 & 1.10535 & 2.32169 & 2.32929 & 1.05195 & 1.05272\\
5E+09 & 1.10306 & 1.10408 & 2.31900 & 2.32487 & 1.05152 & 1.05208\\
6E+09 & 1.10226 & 1.10314 & 2.31746 & 2.32381 & 1.05114 & 1.05162\\
7E+09 & 1.10150 & 1.10234 & 2.31325 & 2.32076 & 1.05074 & 1.05124\\
8E+09 & 1.10096 & 1.10161 & 2.31378 & 2.31816 & 1.05049 & 1.05084\\
9E+09 & 1.10054 & 1.10103 & 2.31191 & 2.31673 & 1.05034 & 1.05057\\
1E+10 & 1.09706 & 1.10062 & 2.30164 & 2.31683 & 1.04856 & 1.05041\\
2E+10 & 1.09520 & 1.09708 & 2.29665 & 2.30469 & 1.04764 & 1.04858\\
3E+10 & 1.09396 & 1.09524 & 2.29308 & 2.29816 & 1.04703 & 1.04767\\
4E+10 & 1.09295 & 1.09398 & 2.28955 & 2.29441 & 1.04651 & 1.04706\\
5E+10 & 1.09220 & 1.09298 & 2.28817 & 2.29108 & 1.04616 & 1.04655\\
6E+10 & 1.09157 & 1.09223 & 2.28576 & 2.28917 & 1.04585 & 1.04619\\
7E+10 & 1.09100 & 1.09158 & 2.28458 & 2.28741 & 1.04556 & 1.04587\\
8E+10 & 1.09050 & 1.09102 & 2.28282 & 2.28500 & 1.04531 & 1.04558\\
9E+10 & 1.09010 & 1.09052 & 2.28203 & 2.28347 & 1.04511 & 1.04532\\
1E+11 & 1.08738 & 1.09012 & 2.27312 & 2.28311 & 1.04375 & 1.04514\\
2E+11 & 1.08583 & 1.08739 & 2.26810 & 2.27413 & 1.04297 & 1.04377\\
3E+11 & 1.08477 & 1.08585 & 2.26471 & 2.26852 & 1.04244 & 1.04299\\
4E+11 & 1.08398 & 1.08478 & 2.26238 & 2.26507 & 1.04205 & 1.04245\\
5E+11 & 1.08335 & 1.08399 & 2.26052 & 2.26261 & 1.04174 & 1.04206\\
6E+11 & 1.08281 & 1.08337 & 2.25875 & 2.26085 & 1.04147 & 1.04175\\
7E+11 & 1.08236 & 1.08282 & 2.25747 & 2.25901 & 1.04124 & 1.04148\\
\hline
\end{array}$$
where the constants satisfy for $n\cdot10^k\leqslant x\leqslant (n+1)\cdot10^k$
$$\begin{array}{rcl}
\frac{x}{\ln x}\left(1+\frac{a_5}{\ln x}\right)&<
\pi(x)\leqslant& \frac{x}{\ln x}\left(1+\frac{b_5}{\ln x}\right)\\
\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{a_6}{\ln^2 x}\right)&\leqslant
\pi(x)\leqslant&\frac{x}{\ln x}\left(1+\frac{1}{\ln x}+\frac{b_6}{\ln^2 x}\right)\\
\frac{x}{\ln x-a_7}&\leqslant
\pi(x)\leqslant&\frac{x}{\ln x-b_7}
\end{array}$$
up to $8\cdot 10^{11}$.
\end{table}
|
\section{Introduction}
The interplay between disorder and strong electronic correlations is
recognized as a very interesting issue in condensed matter physics.
There are now many experimental evidences showing the very important
role of the disorder in $f$-electron systems in addition to the RKKY
or Kondo interactions \cite{Coqblin}. As a result, it can appear
complex phase diagrams which show spin glass (SG) phases in addition
to the onset of antiferromagnetism (AF) or ferromagnetism (FE),
regions dominated by the Kondo effect, the presence of Quantum
Phase Transitions (QPT) and exotic regions which present non-Fermi
liquid behavior (NFL) \cite{NFL}.
Earlier experimental results can illustrate the mentioned
complexity. For instance, in $CeAu_{1-x}Co_{x}Si_{3}$ alloys
\cite{Sampa}, when $Au$ is replaced by $Co$, it first appears a SG
phase, then there is the onset of an AF phase with the N\'eel
temperature decreasing towards a Quantum Critical Point (QCP). Thus,
the glassy behaviour tends to decrease with the increase of $x$ and
finally, for $x>0.9$, there is a complete screening of magnetic
moments due to the Kondo effect.
More recently, experimental findings in $CePd_{1-x}Rh_{x}$
\cite{Wester1,Wester2} and $CeNi_{1-x}Cu_{x}$
\cite{Marcano1,Marcano2} have enlarged the set of non-trivial
behaviour in disordered $f$-electron systems. In both systems, there are
strong indications that a glassy behaviour is present in a suitable
range of doping and this behaviour has been recently identified to
appear as a cluster glass state. In the well studied
$CeNi_{1-x}Cu_{x}$ case, the Kondo interaction is dominating for $x$
smaller than approximately 0.2 \cite{Garcia}. However, the
intermediate doping regime has been extensively studied both
experimentally and theoretically and finally a complex scenario is
obtained when the temperature is decreased. In the first
experimental studies on $CeNi_{1-x}Cu_{x}$ alloys with $x$ typically
between 0.3 and 0.6, a SG phase has been obtained below the
paramagnetic state and then there is a transition to a ferromagnetic
phase at lower temperatures.
More sophisticated experiments have recently shown that dynamic
magnetic clusters are developping at low temperatures below the
paramagnetic state. More precisely, there is the formation of
clusters due to short range ferromagnetic correlations below a
certain temperature $T^{*}$. The volume fraction of these clusters
increases as temperature is lowered and they become frozen at
$T_{cl}$ well below $T^{*}$ and, therefore, it appears an
inhomogeneous ferromagnetic order at very low temperatures
\cite{Marcano1,Marcano2}. Thus, there is a change, below the
paramagnetic phase, from a cluster spin glass to a disordered
ferromagnetic order without any sharp transition, but with a mixed
and disordered intermediate phase.
A Kondo-Cluster-Glass state has been also recently evidenced in
$CePd_{1-x}Rh_{x}$ alloys at very low temperatures. This system
exhibits a continuous evolution from a ferromagnetic order in
$CePd$, with a Curie temperature $T_c=6.6 K$, to an
intermediate-valence ground state in $CeRh$. The Curie temperature
decreases continuously with increasing $x$ and tends to 25 mK at the
value $x=0.87$. Despite pronounced non-Fermi-liquid behavior in the
proximity of this concentration for specific heat and thermal
expansion, it was concluded from the analysis of the Gruneisen ratio
that there is no QCP \cite{Pikul}. On the
opposite, a ``Kondo-cluster-glass" state was found for $x$ larger than
0.65: there is firstly the formation of clusters with predominantly
ferromagnetic couplings of the $f$-moments below a given temperature
$T^{*}$ and then a random freezing of the cluster moments below a
smaller temperature $T_{cl}$ \cite{Wester1,Wester2}. Thus, there are
clearly similarities between the low temperature behaviors of
$CeNi_{1-x}Cu_{x}$ and $CePd_{1-x}Rh_{x}$ alloys, but both a more
profound analysis of the different data and the role of the Kondo
effect have to be precised in these two systems.
Several theoretical studies have tried, since already some time, to
account for the previous experimental data. A Kondo lattice with an
additional Ising term and a random coupling between localized
spins, called here the Kondo-Ising Lattice (KIL) model
\cite{Alba1,Magal1,MagalAF1}, has been firstly used to study the
competition between the Kondo effect and magnetism when disorder is
present within the Static Approximation (SA) \cite{Moore}. It
appears that, for $CeAu_{1-x}Co_{x}Si_{3}$ alloys, a Gaussian random
distributed bond would be adequate as can be seen in Refs.
\cite{MagalAF1,MagalAF2}. The same model has been also firstly
used to describe the case of $CeNi_{1-x}Cu_{x}$ alloys, where the
disorder has been introduced within the classical
Sherrington-Kirkpatrick (SK) model \cite{SK} by taking Gaussian
random intersite coupling $J_{ij}$ with a mean value $J_{0}$
different from zero to describe the ferromagnetic ordering
\cite{Magal1}. The phase diagram giving the temperature $T$ {\it
versus} the strength $J_{K}$ of the Kondo interaction has been
computed and we have obtained, besides the Kondo state, magnetic
phases like Spin Glass (SG), Ferromagnetic (FE) and a mixed phase
(SG+FE). For this particular solution, the ferromagnetic order
occurs with replica symmetry breaking. This phase diagram
could be, therefore, a good starting point to describe the scenario found in
$CeNi_{1-x}Cu_{x}$ alloys. Unfortunately, for this particular kind
of disorder, the Curie temperature $T_{c}$ is always higher than the
freezing one, which is a scenario opposite to the experimental
situation observed in $CeNi_{1-x}Cu_{x}$ system.
Thus, in order to solve the preceding difficulty, a completely
different perspective has been adopted in reference \cite{Magal2}.
The theoretical description of the disorder has been modified from a
bond disordered coupling to a site disordered one. In that case, the
$J_{ij}$ coupling is a generalization of the Mattis model
\cite{Mattis} used extensively to study complex systems \cite{Amit},
given as
$J_{i_{}j_{}}=\frac{J}{2N}\sum_{\mu=1}^{p}\xi_{i}^{\mu}\xi_{j}^{\mu}$,
where $\xi_{i}^{\mu}$ is a random variable which follows a bimodal
distribution.
One important aspect is that, in the corresponding mean field
approach using such $J_{ij}$ values, it is possible to introduce a
parameter which allows to control the level of frustration in the
problem \cite{Magal2}. The first interesting result is that the
Kondo solution is robust in the large $J_{K}$ limit, no matter what
is the level of frustration. For weak frustration and small $J_{K}$,
below a certain temperature, it appears a SG solution. When the
temperature is further decreased, the SG solution is replaced by
Mattis states which have the same thermodynamics as a FE phase
\cite{Amit}. This result suggests that the situation found in
$CeNi_{1-x}Cu_{x}$ alloys would be an example of weak frustration.
Nevertheless, there is an important difference between this model
and the approach of the Gaussian distributed $J_{ij}$. For the kind
of disorder given by this generalized Mattis model, there is no
mixed phase solution for the order parameters, but on the contrary,
there is a first order phase transition between the SG and Mattis
states; such solutions can obviously coexist, but one of
them is always metastable. In conclusion, our previous Mattis-like model gives the SG phase above the FE phase,
but it cannot yield a real SG+FE mixed Phase \cite{Magal2}.
Thus, in order to improve the preceding description and to have,
therefore, a better agreement with experiment, we introduce here, in
our previously used KIL model, a new kind of site disordered
coupling $J_{ij}$, originally introduced by van Hemmen (vH) to study
the Spin Glass in the classical Ising model \cite{van Hemmen}. The
phase diagram obtained from such a classical model displays not only
SG, FE+SG and FE phases, but also they can appear in that order when
temperature is decreased. In this particular case, the SG+FE phase is characterized by both non
zero magnetization and SG order parameters. Recently, a work \cite{Ricardo1} has
studied a mean field solution of a quantum version of the vH model
with an applied transverse field $\Gamma$ and it shows that some
aspects of its classical counterpart can still be preserved in the
quantum vH model, and in particular the SG+FE phase. However, spin
flipping introduced by the presence of $\Gamma$ in the quantum vH
model can modify the phase diagram, suppressing for instance the
presence of SG+FE phase \cite{Ricardo1}. However, it is well known
that an additional transverse field in the KIL model with a Gaussian
random bond coupling between the localized Ising spins operators can
produce important consequences as, for instance, a QCP \cite{Alba2}.
In the present work, we will, therefore, study the KIL model with
both the vH type of disorder for the intersite exchange interaction
$J_{ij}$ and a transverse field $\Gamma$ which allows also to
investigate the possible consequences for the phase diagram with the
spin flipping. There is also another very important aspect related
to the vH type of disorder introduced in the present work:
in the previous approaches using the Gaussian random bond SK-type
$J_{ij}$ \cite{Alba1,Magal1,MagalAF1,MagalAF2,Alba2} or the site
disorder type given by the product of random variables
$\xi_{i}^{\mu}$, the disorder is treated using the so called replica
symmetry solution for the SG order parameters \cite{ReviewBinder}.
This solution is well known to have a serious flaw, because it is
locally unstable below the freezing temperature \cite{Almeida}.
Certainly, that problem could be overcomed by the use of replica
symmetry breaking schemes \cite{Parisi}. However, this kind of
scheme increases the number of order parameters in such a way that
the search for order parameter solutions in the KIL model becomes
extremely complicated. Nevertheless, that is not the only problem
with the use of replicas to treat the disorder in the KIL model.
There are also indications that the presence of one or other
magnetic solutions could be dependent on the particular kind of
replica symmetry breaking schemes \cite{ScesHouston}. By contrast,
that is not the case for the disordered $J_{ij}$ given in the vH
model (see following equation (\ref{eq3})). The disorder can be
treated without the use of replica technique as demonstrated in the
classical and quantum vH models \cite{van Hemmen,Ricardo1}. Thus,
the present use of the van Hemmen description of the disorder in the
KIL model improves considerably the description of the Kondo-Spin
glass-Ferromagnetism competition in disordered Kondo systems.
It is important to remark that the present work
is typically a mean field theory as in
Refs. \cite{Alba1,Magal1,MagalAF1,MagalAF2,Alba2}.
In particular, the Static and saddle point approximations are used here.
The use of the first approximation can be justified since our goal
is mainly to describe phase boundaries as
discussed in Ref. \cite{Alba2}.
The saddle point method is in fact exact here, as a
consequence of the long range nature of the vH coupling.
This paper is structured as follows. In the next section, we
introduce the model and calculate the corresponding thermodynamics.
The following section is dedicated to discuss the numerical
solutions of the saddle point equations for the order parameters and
to derive the phase diagram. Finally, the last section is reserved
to the conclusions.
\section{General Formulation}
The starting Hamiltonian in the KIL model is given by:
\begin{equation}
\begin{split}
H={\displaystyle\sum_{ij,s}}t_{ij}\hat{n}_{is}^{d}+\epsilon_{0}
{\displaystyle\sum_{i,s}}\hat{n}_{is}^{f}
+J_{K}{\displaystyle\sum_{i}}[\hat{S}_{fi}^{+}\hat{s}_{di}^{-}+\hat{S}_{fi}^{-}\hat{s}_{di}^{+}]
\\
-{\displaystyle\sum_{i,j}}J_{ij}\hat{S}_{fi}^{z}S_{fi}^{z}-
2\Gamma\sum_{i}\hat{S}_{fi}^{x}.
\label{eq2}
\end{split}
\end{equation}
In Eq. (\ref{eq2}), $\hat
S_{fi}^{z}=\frac{1}{2}[\hat{n}_{i\uparrow}^{f}-\hat{n}_{i\downarrow}^{f}]$,
$\hat S_{fi}^{+}=f_{i\uparrow}^{\dagger}f_{i\downarrow}$,
$\hat{S}_{fi}^{-}=(\hat{S}_{fi}^{+})^{\dagger}$,
$\hat S_{fi}^{x}=f_{i\uparrow}^{\dagger}f_{i\downarrow}+f_{i\downarrow}^{\dagger}f_{i\uparrow}$,
$\hat s_{di}^{+}=d_{i\uparrow}^{\dagger}d_{i\downarrow}$,
$\hat{s}_{di}^{-}=(\hat{s}_{si}^{+})^{\dagger}$,
$\hat{n}_{is}^{f}=f_{is}^{\dagger}f_{is}$,
$\hat{n}_{is}^{d}=d_{is}^{\dagger}d_{is}$ where
$f_{is}^{\dagger}~(f_{is})$ and $d_{is}^{\dagger}~(d_{is})$ are
fermionic creation (destruction) operators of $f$ and $d$ electrons,
respectively. The spin projections are indicated by $s=\uparrow$ or
$\downarrow$.
The random coupling $J_{ij}$ in Eq. (\ref{eq2}) is given as in the
vH model by:
\begin{equation}
J_{ij}=\frac{J}{N}(\xi_{i}\eta_{j}+\eta_{i}\xi_{j})+\frac{J_{0}}{N}
\label{eq3}
\end{equation}
where $\xi_{i}$ and $\eta_{i}$ in Eq. (\ref{eq3}) are random
variables which follow the bimodal distribution:
\begin{equation}
P(x)=\frac{1}{2}[\delta(x-1)+\delta(x+1)].
\label{eq4}
\end{equation}
In Eq. (\ref{eq4}), $\delta(x)$ is the Dirac delta function.
As discussed in the previous section,
the coupling $J_{ij}$ given in Eq. (\ref{eq3}) is an infinite long range
coupling which
gives
exact solutions in the thermodynamical limit for the
saddle point approximation used below.
The partition function is expressed within functional formalism
using anticommuting Grassmann variables $\varphi_{is}(\tau)$
and $\psi_{is}(\tau)$ associated to the $f$ and $d$ electrons, respectively
as \cite{Alba1,Alba2}:
\begin{equation}
Z = \int D(\psi^{\ast}\psi) D(\varphi^{\ast}\varphi)
\exp\left[A_{VH} + A_K
+ A_0 \right].
\label{e8}
\end{equation}
In the static approximation (SA) \cite{Moore}
the actions in Eq (\ref{e8}) are given as:
\begin{equation}
\begin{split}
A_0 =\sum_{\omega} \sum_{i,j}
\left[(\underline{\psi})_{i}^{\dagger}(\omega)
(i\omega -\beta \varepsilon_{0}+\beta\Gamma \underline{\sigma}_x)
\delta_{ij} \underline{\psi}_{i}(\omega)\right.
\\
\left. +
\underline{\varphi}_{i}^{\dagger} [(i\omega+\mu_{d})\delta_{ij} -
\beta t_{ij}]\underline{\varphi}_{j}
(\omega) \right]
\label{A0}
\end{split}
\end{equation}
where, in the first term of Eq. (\ref{A0}), the chemical potential
$\mu_{f}$ has been absorbed in $\varepsilon_{0}$.
\begin{equation}
\begin{split}
A^{stat}_K \approx \frac{J_K}{N} \sum_{is} \sum_{\omega}
\left[ \varphi_{i-s}^{\ast}(\omega)
\psi_{i-s}(\omega)\right]
\sum_{js}\sum_{\omega^{'}}\left[ \psi_{js}^{\ast}(\omega^{'})
\varphi_{js}(\omega^{'})\right],
\end{split}
\end{equation}
\begin{equation}
A^{stat}_{VH} ={\displaystyle \sum_{ij}} J_{ij} S_{fi}^{z}S_{fj}^{z}
\label{e9}
\end{equation}
with
\begin{eqnarray}
S_{fi}^{z}=\frac{1}{2}{\displaystyle \sum_{\omega}}\underline{\psi}_{i }^{\dagger}(\omega)\underline{\sigma}^{z}\psi_{i }(\omega).
\label{eq10}
\end{eqnarray}
The action $A_{K}^{stat}$ is given in the mean field approximation (see Ref \cite{Alba1}).
In the remaining components of the action $A_{0}$ and $A_{VH}^{stat}$, spinors are used :
\begin{eqnarray}
\begin{array}{ccc}
\underline{\varphi}_{i}(\omega)=\left( \begin{array}{c} \varphi_{i\uparrow}(\omega) \\ \varphi_{i\downarrow}(\omega)\end{array}\right), & &
\underline{\psi}_{i}(\omega)=\left(\begin{array}{c} \psi_{i\uparrow}(\omega) \\ \psi_{i\downarrow}(\omega)\end{array}\right)
\end{array}
\label{ee369b}
\end{eqnarray}
and the Pauli matrices are given as usual by:
\begin{eqnarray}
\begin{array}{ccccc}
\underline{\sigma}_{x}=\left(\begin{array}{cc} 0 & 1 \\ 1 & 0
\end{array}
\right)& &\underline{\sigma}_{y}=\left(\begin{array}{cc} 0 & -i \\ i & 0
\end{array}
\right) &\underline{\sigma}_{z}=\left(\begin{array}{cc} 1 & 0 \\ 0 & -1
\end{array}
\right).\end{array}
\label{ee369c}
\end{eqnarray}
We follow a procedure close to that introduced in Ref. \cite{Alba1}.
Therefore, the Kondo order parameter
$\lambda_{\sigma}\approx\lambda=\frac{1}{N}{\displaystyle\sum_{j,\omega}}\left\langle
\psi_{j\sigma}^{*}(\omega)\varphi_{i\sigma}(\omega)\right\rangle$
can be introduced in the partition function $Z$ defined by Eqs.
(\ref{e8})-(\ref{eq10}). Then, the $\varphi$ fields are integrated
and we obtain the following result:
\begin{equation}
Z/Z^{0}_{d}=\exp(-2N\beta \lambda \lambda^{*})Z_{eff}
\label{eq5}
\end{equation}
where $Z^{0}_{d}$ is the partition function of free $d$ electrons and
\begin{equation}
\begin{split}
Z_{eff}={\displaystyle \int D(\psi^{*}\psi)} \exp\left( A^{stat}_{VH}
{+\displaystyle\sum_{\omega\sigma}}{\displaystyle\sum_{i,j}}\underline{\psi}_{i\sigma}^{*}
(\omega)\underline{g}_{ij}^{-1}(\omega)
\underline{\psi}_{j\sigma}(\omega)
\right)
\label{eq6}
\end{split}
\end{equation}
with
\begin{equation}\begin{split}
\underline{g}_{ij}^{-1}(\omega)=[(i\omega-\beta\epsilon_{0})\underline{I}+\beta\Gamma\underline{\sigma_{z}}]\delta_{ij}
-
\frac{\beta^{2}J_{k}^{2}\lambda^{2}}{(i\omega+ \mu_{d})\delta_{ij}-\beta t_{ij}}\underline{I}.
\label{eq7}
\end{split}\end{equation}
In Eq.(\ref{eq7}), we use the notation $|\lambda^{2}|\equiv\lambda^{2}$ and $\underline{I}$ means the unitary matrix.
Introducing $J_{ij}$ given by Eq. (\ref{eq3}), the action
$A^{stat}_{vH}$ becomes composed of two terms: one randomic and the
other one ferromagnetic. They can be rearranged to introduce SG and
FE order parameters in $Z_{eff}$. The details of such calculations
are shown in the Appendix.
The free energy is, therefore, given by:
\begin{equation}
\beta F= 2\beta J_{K} \lambda^{2}- \lim_{N\rightarrow\infty} \frac{1}{N}\ln Z_{eff}.
\label{eq100}
\end{equation}
Using the saddle point solution for $Z_{eff}$ (see Eqs.
((\ref{eq11})-(\ref{eq19}))), the free energy in Eq. (\ref{eq100})
becomes:
\begin{equation}
\begin{split}
\beta F=2\beta J_{K} \lambda^{2}+2\beta J
q_{1}q_{2} + \beta J_{0}m^{2}
- {\displaystyle\sum_{\omega}}\ln(\det
\underline{G}_{ij}^{-1}(\omega|h_{j})).
\label{eq25}
\end{split}
\end{equation}
The Green function $\underline{G}_{ij}^{-1}(\omega|h_{j})$
is given in Eqs. (\ref{eq20})-(\ref{eq15}).
In order to proceed to the calculations, we use in the last term of Eq. (\ref{eq25})
the approximation introduced in Ref. \cite{Alba1}
which decouples the
random magnetic field $h_{j}$ from the
Kondo lattice.
Thus, we obtain:
\begin{equation}
\ln \det\left(
\underline{G}_{ij}^{-1}(\omega|h_{j})\right)\approx\frac{1}{N}
{\displaystyle\sum_{j}}\ln\left[\det\underline{\Gamma}_{\mu
\nu}^{-1}(\omega|h_{j})\right]
\label{eq23}
\end{equation}
with
\begin{equation}
\begin{split}
\Gamma_{\mu\nu}^{-1}(\omega|h_{j})=\left[ (i\omega - \beta
\epsilon_{0})\underline{I}-\underline{\sigma_{z}} h_{j}
+\beta\Gamma\underline{\sigma_{x}} \right]\delta_{\mu \nu}
\\
-\beta^{2} J_{k}^{2}\lambda^{2}\frac{1}{N}
{\displaystyle\sum_{\vec{k}}} \frac{ e^{i\vec{k}\vec{R}_{\mu \nu}}
}{(i \omega+\mu_{d})-\beta\epsilon_{k}}\underline{I}.
\label{eq24}
\end{split}
\end{equation}
Now, in the last term of Eq. (\ref{eq25}), we can use
self-averaging property $\frac{1}{N}\sum_{j}
f(\eta_{j};\xi_{j})=\left\langle \left\langle f(\eta; \xi)
\right\rangle \right\rangle_{\xi\eta}$.
Therefore:
\begin{equation}
\begin{split}
\frac{1}{N}{\displaystyle\sum_{j}} \ln\left[
\frac{1}{N}{\displaystyle\sum_{\vec{k}}}(\sum_{\omega}\Gamma_{\sigma}(\vec{k},h_{j}))
\right]
= \left\langle \left\langle\ln\left[
\frac{1}{N}{\displaystyle\sum_{\vec{k}}}(\sum_{\omega}\Gamma_{\sigma}(\vec{k},h))\right]
\right\rangle \right\rangle _{\xi\eta}
\label{eq29}
\end{split}
\end{equation}
where
\begin{equation}
\left\langle \left\langle
f(\xi,\eta)
\right\rangle \right\rangle _{\xi \eta}= {\displaystyle \int d\xi d\eta P(\xi, \eta) f(\xi,\eta)}
\label{eq30}
\end{equation}
Then, by assuming that $\mu_{d}=0$ and $\varepsilon_{0}=0$, the free
energy can be found as
\begin{equation}
\begin{split}
\beta F= 2\beta J_{k}\lambda^{2} + 2\beta Jq_{1}q_{2} + \beta J_{0}m^{2}
\\
-
\left\langle \left\langle
\frac{1}{\beta D}
{\displaystyle \int_{-\beta D}^{+ \beta D}} dx \ln
\left[
\cosh\left(\frac{x+H}{2} \right)
\right.\right.\right.
\\
\left.\left.\left.
+\cosh \sqrt{\frac{1}{4}(x-H)^2+\beta^{2}J_{k}^{2}\lambda^{2}}
\right]\right\rangle \right\rangle _{\xi\eta}
\label{eq33}
\end{split}
\end{equation}
with
\begin{equation}
H = \beta\sqrt{\left[2J(\eta q_{2}+\xi q_{1})+ 2J_{0}m
\right]^{2}+\Gamma^{2} }.
\label{eq211}
\end{equation}
In Eq. (\ref{eq33}), the sums over the Matsubara frequencies and
over $\vec{k}$ have been done in a way similar to Ref. \cite{Alba1}.
We have also used here the usual approximation of a constant density of states
for
the $d$ electrons, $\rho=\frac{1}{2D}$ for $-D<\epsilon<D$.
The use of this density of states allows a direct comparison
of phase diagrams obtained in this work with previous ones given
in Refs. \cite{Alba1,Magal1,Magal2}.
Finally,
assuming that the probability distribution $P(\xi,
\eta)=P(\xi)P(\eta)$, we can compute $\left\langle\left\langle...
\right\rangle \right\rangle_{\xi \eta}$ in Eq. (\ref{eq33}) using
Eqs. (\ref{eq4}) and (\ref{eq30}).
\begin{figure}[t]
\includegraphics[height=\columnwidth,angle=-90]{figure1.ps}
\caption{Phase diagram $T/J$ {\it versus} $J_K/J$ for $J_0/J=1.6$ and $\Gamma/J=0$.}
\label{fig1}
\end{figure}
\section{Numerical results}
The coupled saddle point equations for $q_{1}$, $q_{2}$, $m$ and
$\lambda$ can be obtained directly from Eqs.
(\ref{eq33})-(\ref{eq211}). The numerical solutions for such order
parameters allow us to obtain the following phases: (i)
paramagnetism (PARA) given by $q_{1}=q_{2}= 0$, $m=0$ and
$\lambda=0$; (ii) the SG phase given by $q_{1}=q_{2}\neq 0$, $m=0$
and $\lambda=0$; (iii) the mixed phase (SG+FE) given by
$q_{1}=q_{2}\neq 0$, $m\neq 0$ and $\lambda=0$; (iv) ferromagnetism
(FE) given by $q_{1}=q_{2}= 0$, $m\neq 0$ and $\lambda=0$; (v) Kondo
state where only $\lambda$ is different from zero. For
numerical results, $D/J=12$ is used.
Phase diagrams giving temperature $T$ {\it versus} $J_K$ (in units
of $J$) can be built for several values of $J_{0}/J$ and $\Gamma/J$.
In Figure (\ref{fig1}), such a phase diagram is displayed for
$J_0/J=1.6$ and $\Gamma/J=0$. For this case, in the large $J_K$
region there is only one solution which corresponds to the Kondo
state. When $J_K$ decreases, the Kondo solution disappears.
Actually, it is substituted by the magnetic solutions PARA, SG and
FE which appear in that order when $T$ is lowered. In Figure
(\ref{fig2}), we take $J_0/J=1.3$ and $\Gamma=0$. This decrease of
$J_0/J$ from 1.6 to 1.3 does not affect the Kondo state, but changes
a lot the magnetic solutions. In Figure (\ref{fig2}), the solution
FE is replaced by the mixed phase SG+FE, while the size of the
region where the SG solution exists remains almost the same as in
Figure (\ref{fig1}). In Figure (\ref{fig3}), the transverse field
$\Gamma$ is maintained equal to 0 and we take an intermediate value
$J_0/J=1.4$. As in the two previous cases, the Kondo state is not
really affected in the large $J_K$ region, but the region of the
magnetic solutions in the phase diagram is again modified. Besides
the existence of SG and SG+FE solutions, when the temperature is
decreased, there is also an additional FE solution at much lower
temperatures. In other words, in a small range of $J_0/J$
($1.3\leq J_0/J \leq 1.6$), the phase diagrams present several
scenarios concerning the existence of magnetic solutions. In
contrast, the Kondo state is robust to such changes of $J_0/J$.
Furthermore, a new situation is obtained when the transverse field
is turned on, as can be seen in Figure (\ref{fig4}). For instance,
for $\Gamma=1.0$, the Kondo solution is obtained for a value of
$J_K/J$ a little larger than that found previously for $\Gamma=0.4$
or $\Gamma=0$ and simultaneously, the range of $J_K/J$ where the
magnetic solutions are found is increased. Moreover, for such a
decrease of $\Gamma$ from 1 to 0, the transition temperatures
between the magnetic phases are clearly depressed. But the most
important feature observed with the increase of $\Gamma$ concerns
the magnetic solutions, because the SG+FE and FE phases disappear
completely and it remains only the SG phase for a sufficiently large
$\Gamma$ value.
\begin{figure}[t]
\includegraphics[height=\columnwidth,angle=-90]{figure2.ps}
\caption{Phase diagram $T/J$ {\it versus} $J_K/J$ for $J_0/J=1.3$ and $\Gamma/J=0$.}
\label{fig2}
\end{figure}
\begin{figure}[ht]
\includegraphics[height=\columnwidth,angle=-90]{figure3.ps}
\caption{Phase diagram $T/J$ {\it versus} $J_K/J$ for $J_0/J=1.4$ and $\Gamma/J=0$ .}
\label{fig3}
\end{figure}
\begin{figure}[ht]
\includegraphics[height=\columnwidth,angle=-90]{figure4.ps}
\caption{Phase diagrams $T/J$ {\it versus} $J_K/J$ for $J_0/J=1.4$
and three values of $\Gamma/J:$ 0, 0.4 and 1.0. The dashed, dotted and full lines are
results for $\Gamma/J=1.0$, $\Gamma/J=0.4$ and $\Gamma/J=0.0$, respectively.
The critical lines for
$\Gamma/J=0$ occur at higher temperatures than those ones for
$\Gamma/J=0.4$ and $\Gamma/J=1.0$. In particular, for $\Gamma/J=1.0$ there is no more SG+FE and FE solutions.
} \label{fig4}
\end{figure}
\section{Conclusions}
In the present work, the KIL model has been studied with assuming
that the inter-site spin coupling $J_{ij}$ between localized spins is a
random coupling given by the van Hemmen model as given in Eq.
(\ref{eq3}). It has also been added to the model a transverse field $\Gamma$
which mimics a Heisenberg spin-flipping term.
The results are shown in Figures (\ref{fig1})-(\ref{fig4}). For
$\Gamma=0$, they basically display two regimes when the strength
$J_K$ of the Kondo interaction is varied in units of the
component $J$ of the coupling $J_{ij}$ (see Eq. (\ref{eq3})). In the
first regime obtained for large $J_K$ values, there is only the
Kondo phase. In contrast, the second regime with only the magnetic
solutions SG, SG+FE and FE exists when $J_{K}$ is decreased. One
important point is the order in which the magnetic phases are found
when the temperature is decreased. For instance, the SG phase is
found at higher temperature. Then, it can appear a SG+FE phase. The
pure FE phase is found only at the lowest temperatures. It is also
important to notice that the existence of the different solutions
SG, SG + FE or FE depends on the strength of the ferromagnetic
component $J_0$ (given in units of $J$) of the coupling $J_{ij}$, as
can be seen in the Figures (\ref{fig1})-(\ref{fig3}). When $\Gamma$
is different from zero, the two regimes discussed previously are
affected. While the Kondo solution needs larger values of $J_K$ to
be found, the magnetic solutions found at lower temperatures
disappear rapidly when $\Gamma$ is increased.
It should be emphasized that the present approach using the $J_{ij}$
coupling given by the vH model yields two important improvements
with respect to previous approaches. The first one concerns the use
of the replica method which is not necessary here to generate the
thermodynamics. This is an important improvement with respect to the
previous approaches using the bond disorder given by the SK-like
Gaussian random $J_{ij}$ in the KIL model
\cite{Alba1,Alba2,MagalAF1,MagalAF2,Magal1} or using the previous
Mattis-like approach \cite{Magal2}. For instance, the presence of
magnetic solutions in these approaches is quite dependent on which
particular scheme of replica solution is used, as explained in the
discussion of Ref. \cite{ScesHouston}.
As our present results suggest, the second improvement
concerns the
particular kind of site disorder given by the vH model introduced in
the KIL model with a certain range of $J_0$, which allows to obtain
magnetic solutions SG, SG+FE and FE phases when the temperature is
decreased. In that sense, the weakness of the approach proposed in
Reference \cite{Magal2} is overcome and we are able to introduce
here a mixed phase SG+FE.
Thus, our present calculation using the van Hemmen site
disorder can describe Cerium disordered physical systems such as
$CeNi_{1-x}Cu_{x}$ or $CePd_{1-x}Rh_{x}$ alloys. In particular,
Figures 3 and 4 can describe the phase diagram of
$CeNi_{1-x}Cu_{x}$ with $J_K$ increasing with an increasing
Nickel concentration, by explaining the Kondo behavior observed
for $x$ close to $1$ and by proposing a good interpretation of the complicate magnetic behavior
observed for smaller $x$ values.
These are indications that the use of the van Hemmen site disorder could
be useful to describe physical systems such as $CeNi_{1-x}Cu_{x}$ or
$CePd_{1-x}Rh_{x}$ alloys, although the low temperature phase is in
these alloys a Kondo-cluster-glass followed by a disordered
ferromagnetic one. However, it is important to notice that canonical
spins have been used in the present work. This description is
obviously not sufficient to capture the complexity of the cluster glass
state. However, earlier results for a mean field formulation of the
cluster glass indicate that there are no essential differences
between canonical spins and clusters of spins, as far as the phase
boundaries are concerned \cite{Sokoulis}. One can, therefore, expect
that most of the previous discussion concerning
the sequence of magnetic orders as a function of $J_K$
can
be preserved even if the problem is formulated in terms of clusters
of spins instead of canonical spins as it is done in the present
work.
On the other hand, we are presently working on a theoretical
description of the Kondo-Cluster-Glass, by solving exactly the
problem in a small cluster with $n_s$ atoms interacting between them
by a disorder spin glass-like interaction. We have already solved
the problem with only $n_s=3$ and a disorder intercluster bonding given by the
Sherrington-Kirkpatrick interaction \cite{ScesBuzios}. We think that
the van Hemmen approach is easier to treat and we are presently
working on clusters with a larger number $n_s$, in order to have
finally a more local description of the Kondo-Cluster-Glass observed
in some disordered Kondo Cerium systems.
In conclusion, we have to remark that our van Hemmen-Kondo
description yields considerable improvements with respect to
previous theoretical models in the two following points, the non
consideration of the replica method and the problem of the mixed SG+FE
phase. The validity of the van Hemmen model, which does not use the replica trick method, has been
discussed in detail and it has been shown that this model is
perfectly able to describe the spin glass experiments and that it is
simpler than the other models for a mathematical treatment
\cite{Choy,van Hemmen1}. On the other side, our van Hemmen-Kondo model gives with decreasing temperature a SG phase, a SG+FE one and finally a ferromagnetic phase and the intermediate SG+FE phase is a real mixed phase with together non zero SG and FE order parameters. This model gives a good account for the experimental phase diagrams of disordered Cerium systems, such as $CeNi_{1-x}Cu_{x}$ alloys,
and can be used to have a more local description of the Kondo-Cluster-Glass phase.
\section*{Acknowledgments}
B. Coqblin acknowledges the European Cost P16 Action for financial
support.
S.G. Magalhaes and F.M. Zimmer acknowledge the CNPq for financial support.
|
\section{Introduction\label{Intoduction}}
The purpose of this paper is to develop the semi-quantitative extension of
the BCS theory of superconductivity that describes strongly disordered
conductors which normal state is a weak Anderson insulator~\cite{AndersonLoc}
or a very poor metal. The paper focuses on the case of "uniformly
disordered" materials, which do not contain morphological structures such as
grains coupled by tunnel junctions. Below in this section we briefly review
the existing theoretical models of the superconductor-insulator transition
(SIT), compare them with the results of the experimental studies of the
uniformly disordered films and choose the appropriate model for the
superconductor-insulator transition in these materials. The conclusion of
this introductory part is that this quantum transition in uniformly
disordered films can be described by BCS pairing of electrons which
single-particle states are close to the mobility edge of Anderson
localization~\cite{AndersonLoc}. Because BCS pairing is most relevant for
electrons close to the Fermi surface, the transition occurs when Fermi-level
$E_{F}$ is located in the region of localized single-electron states but
close to the mobility edge. In the vicinity of the transition localization
length is longer than typical distance between carriers while the relevant
single electron states have the statistical properties of "critical
wavefunctions". Section \ref{Model} formulates the model in more detail and
discusses the issue of \textit{wavefunction fractality}, which turns out to
be very important for the theory of superconductor-insulator transition
developed in this work.
The important difference between our approach and many other works on
superconductor--insulator transition is neglect of the effects of Coulomb
interaction but consistent treatment of moderately strong disorder. In this
respect our work is the extension of the approaches developed originally by
Ma and Lee~\cite{MaLee}, Kapitulnik and Kotliar~\cite{KapitulnikKotliar1986
, Bulaevskii and Sadovskii~\cite{BulaSad}, and more recently by Ghosal,
Trivedi and Randeria~\cite{Ghosal2001} that have considered competition of
superconducting pairing and Anderson localization without explicit account
for Coulomb interaction. We give detailed arguments which justify
applicability of our approach to disordered films such as amorphous InO$_{x}$ and TiN on
both phenomenological and microscopic levels below in subsection \re
{Experimental results}. Briefly, one should distinguish two possible effects
of the Coulomb interaction: suppression of paring interaction between
individual fermions which occurs at short scales and enhancement of the
phase fluctuations of the order parameter at large scales. The first would
lead to a gap suppression in a direct contradiction with data while the
second would lead to the phenomenology similar to that of Josephson junction
arrays which display markedly different behavior. Our theory can be also
applied to the cold atoms in optical and magnetic lattices with a controlled
disorder. In such systems interaction is always attractive and effectively
short-range. Moreover, it is tunable by magnetic field due to the Feshbach
resonance~\cite{atoms} which would allow to test directly the detailed
predictions of the developed theory.
The important consequence of the wave function fractality is the formation
of the strongly bound electron pairs which survive deep in the insulating
regime. In this situation the pairing interaction reduces the mobility of
individual electons leading to "superconductivity-induced" insulator. We
discuss this behavior in section \ref{Insulating state}. Theory of Cooper
pairing of electrons populating critical fractal states is developed in
section \ref{Cooper instability}. Here we develop three different
approximations for the computation of the superconducting transition
temperature and other properties. Section \ref{Superconducting state} gives
the physical properties of the superconductor in this regime. We show that
three different approximations developed in section \ref{Cooper instability}
agree with each other and predict parametrically strong enhancement of
T_{c} $ with respect to its value given by the "Anderson theorem". Another
distinguishing feature of the emerging superconducting state is extremely
strong spatial inhomogeneity of superconducting order parameter and the
presence of a well-defined global $T_{c}$.
Section \ref{Superconductivity with a pseudogap} presents the theory of
superconductivity coexisting with a strong pseudogap. Here we show that
superconductivity survives when $E_{F}$ is located much deeper in the
localized band than it was previously expected. In this regime the
superconductivity develops against the background of the pseudogap and,
thus, is characterized by a number of unconventional properties. We present
the specific results for the tunneling density of states and tunneling
conductance, Andreev point-contact conductance, and spectral weight of
high-frequency conductivity in this pseudogap superconductive state.
Finally, section \ref{Summary of results} reviews the main results and
discusses a number of open problems. The Appendices present technical
details of virial expansion that was used as the one of methods for the
determination of $T_{c}$.
\subsection{Theoretical models\label{Theoretical models}}
Quantum phase transitions from superconducting to insulating state in
disordered conductors and artificial structures were studied intensively
since mid-1980s, for review see e.g. \cit
{ReviewGoldman,Finkelstein1994,FazioZant}. A number of theoretical models
describing such transitions were proposed and studied but a fully coherent
theoretical picture of this phenomena has not been established. We believe
that such transitions might be driven by different mechanisms in different
materials and thus belong to different universality classes which should be
described by different models. Below in this section we briefly review the
alternative mechanisms and the theoretical methods employed for their
description (see also short review~\cite{Larkin1999}). To avoid confusion,
we note that the first and the second of the mechanisms described below will
not be studied in the main part of our paper. We describe them in some
detail mostly because we need these details in order to argue below in this
section that they are not relevant for the superconductor-insulator
transition in homogeneous amorphous films.
\subsubsection{Coulomb blockade versus Superconductivity in Josephson
junction arrays. \label{Coulomb blockade}}
Macroscopic conductor that looks homogeneous at a macroscopic scale might be
in fact composed of small grains or islands of \ a good superconducting
metal with transition temperature $T_{c0}$. These grains are coupled to each
other via low-transparency insulating tunnel barriers~\cit
{ReviewGoldman,FazioZant}, characterized by dimensionless tunnel
conductances $G_{ij}=h/(2e)^{2}R_{ij}$. The transition is due to the
competition of the charging and Josephson energies. Exactly the same physics
is realized in the artificial Josephson junction arrays, the only difference
between the inhomogeneous films and artificial array is that the number of
neighbors and of plaquette areas in the former are random. The simplest
model describing this physics is given by effective Hamiltonian written in
terms of phases $\phi _{j}$ and charges $Q_{j}$ assigned to grains:
\begin{equation}
H_{1}=\frac{(2e)^{2}}{2}\sum_{ij}C_{ij}^{-1}N_{i}N_{j}-\sum_{ij}E_{J}^{ij
\cos (\phi _{i}-\phi _{j}) \label{H1}
\end{equation
where $N_{j}=Q_{j}/2e$ is the number of Cooper pairs on the $j$-th grain,
C_{ij}$ is the matrix of mutual capacitances, and $E_{J}^{ij}$ is the
Josephson coupling energy, correspondingly. The model is often further
simplified by assuming that $E_{J}^{ij}=E_{J}$ is a non-zero constant only
for nearest-neighboring grains while\textbf{\ }the capacitance matrix is
diagonal, $C_{ij}=C_{0}\delta _{ij}$. This model neglects the effects of the
quasiparticles which might be (sometimes) justified for small grains at low
temperatures $T\ll T_{c0}$, due to exponentially small density of normal
electrons. The key parameter of the problem is the energy ratio
x=E_{J}/E_{C}$ where $E_{C}=(2e)^{2}/2C_{0}$ is the Coulomb charging energy
due to $2e$ charge transfer. As was shown in the paper~\cite{Efetov1980} the
ground-state of the Hamiltonian (\ref{H1}) is insulating at $x\ll 1$ and
superconducting at $x\gg 1$, thus a phase transition(s) takes place at
x\sim 1$. An essence of this phase transition is the Mott-Hubbard
localization of Cooper pairs, taking place when tunneling matrix element of
a pair ($E_{J}$) is much less than on-site repulsion $E_{C}$.
However, the model (\ref{H1}) is unlikely to describe correctly the physics
of superconductor-insulator transition in Josephson arrays or granular
materials, especially in its simplified version with diagonal capacitance
matrix. It has two important deficiencies.
First, realistic Josephson junction arrays and disordered films are poorly
described by the model of diagonal capacitance matrix $C_{ij}=C_{0}\delta
_{ij}$ because normally the charging effects are controlled by capacitances
of junctions $C\gg C_{0}$, not by the ground capacitances of the islands
(see Ref.~\cite{FazioZant}). It is in fact impossible to have a capacitance
matrix dominated by the ground capacitance in the arrays which dimensionless
normal state conductance is $G=h/[(2e)^{2}R_{T}]\gtrsim 1$ because in these
arrays the capacitance of the junctions cannot be small. The reason for this
is that apart from purely geometrical contribution $C_{\mathrm{geom}}=4\pi
S/d$, junction capacitance $C=C^{\mathrm{geom}}+C^{\mathrm{ind}}$ contains
additional induced term $C_{\mathrm{ind}}=\frac{3}{16}Ge^{2}/\Delta $, (this
expression is valid at $T\ll \Delta $). This induced contribution is due to
virtual electron transition across the gap~\cite{LarkinOvchinnikov1983,AES}.
As a result, the charging energy can not be made arbitrary large
(equivalently capacitance cannot be small): $E_{C}\leq 32\Delta /3G$.
Josephson energy of the symmetric junction at low temperatures is
E_{J}=G\Delta /2$, thus the condition $E_{C}\geq E_{J}$ cannot be realized
at large $G$. Furthermore, at temperatures above the parity effect threshold
$T^{\ast }$, (see~\cite{FeigKor}) an additional contribution to the
screening of Coulomb interaction between Cooper pairs comes from
single-electron tunneling. Thus, in all cases the effect of capacitance
renormalization is that the ratio $x=E_{J}/E_{C}$ in granular arrays is
controlled by the dimensionless conductance $G$ in such a way that Coulomb
effects are always weak at $G\geq 1$. Ground capacitance larger than the
junction capacitance thus implies that the transition into the insulating
state would occur in the arrays characterized by very small $G\ll 1$. Such
behavior was never observed experimentally in Josephson arrays (see section
\ref{Experimental results}).
More realistic model involves the capacitance matrix that is dominated by
the junctions capacitances with a small contribution from the ground
capacitance of each grain. In this case the arguments of preceding paragraph
show that the transition between the insulating and superconducting state
should occur at $G_{c}\sim 1$.\cite{Orr1986} Moreover, deep in the
insulating phase the electrostatic interaction between the charges in 2D
Josephson array becomes logarithmic in distance, similar to the one between
the vortices in the superconducting phase. Assumption of the full duality
between vortices in the superconducting state and charges in the insulating
state allows one to make a number of predictions.\cit
{MFisher1990a,MFisher1990b} For instance, because the current of vortices
generates voltage while the current of pairs implies the electrical current
one expects that superconducting-insulator transition is characterized by
the universal value of $G_{c}=1$. \cite{MFisher1990b}
Unfortunately, despite a significant experimental effort the universal value
of the resistance was never experimentally confirmed for Josephson arrays
and for most disordered films. Moreover, in non-zero magnetic field the
Josephson arrays often show a large regime of the temperature-independent
resistance. The reason for this is likely to be due to the important
physical effects missed by the model (\ref{H1}), namely the presence of
random induced charge on the superconducting islands. As was shown in a
number of Josephson junction studies (see e.g. \cite{Zimmerman2008} ) the
induced charge on each island exhibits very slow random fluctuations and is
therefore inherently random variable. Most likely the time dependence of
these fluctuations can be neglected and the induced charge can be regarded
as a quenched random variable $q_{i}$ that should be added to the
Hamiltonian (\ref{H1}):
\begin{equation}
H_{1}=\frac{(2e)^{2}}{2}\sum_{ij}C_{ij}^{-1}(N_{i}-q_{i})(N_{j}-q_{j})
\sum_{ij}E_{J}^{ij}\cos (\phi _{i}-\phi _{j}) \label{H_JJ}
\end{equation
The properties of the model (\ref{H_JJ}) are not currently well understood;
in particular, it seems likely (see \cite{MullerIoffe}) but was not proven
that insulating state has many glassy features responsible for intermediate
'normal' phase. It is, however, clear that in this model the transition, or
a series of transitions occurs at $E_{J}/E_{C}\sim 1$, which corresponds to
G\sim 1$ in the normal state. Near the critical point the gapless
excitations correspond to collective modes build of electron pairs while the
electron spectrum remains fully gapped. In analogy with spin glasses, one
expects that effective frustration introduced by random charges and magnetic
field leads to a large density of low energy states. This might explain the
observed temperature independent resistance that varies at least by one order of
magnitude around $G\sim 1$ as a function of magnetic field.\cite{FazioZant}
As we show in section \ref{Experimental results} both the data and
theoretical expectations for models (\ref{H1},\ref{H_JJ}) differ markedly
from the behavior of the homogeneously disordered films.
A spectacular property of the superconductor-insulator transition in
granular materials is that a strong magnetic field applied to the system in
the insulating regime results in a dramatic increase of the conductance.
Qualitatively, the reason for this behavior in granular systems is that in
the absence of the field the single electron excitations are absent due to
superconducting gap while pairs are localized due to Coulomb energy and
random induced charges. Large field suppresses superconducting gap, which
allows transport by individual electrons that is characterized by a much
larger tunneling amplitude and lower (by a factor of four) effective
charging energy. This effect was reported by \cite{Gerber1997} where strong
magnetic field was applied to Al grains immersed into Ge insulating matrix
and giant negative magneto-resistance was observed. Similar behavior was
reported for homogeneous films of InO deep in the insulating regime\cit
{Gantmakher1996}. This similarity indicates that the main reason for this
effect, which is that the pairing of the electrons survives deep in the
insulating state, also holds for homogeneous films of InO. The quantitative
theory of negative magnetoresistance in granular superconductors was
developed in~\cite{BeloborodovLarkinEfetov} for the case of relatively large
inter-grain conductances $G_{ij}\approx G\gg 1$, in which case the negative
magnetoresistance effect is small as $1/G$. Recent review of theoretical
results on normal and superconductive granular systems can be found in \cit
{Beloborodov}.
\subsubsection{Coulomb suppression of $T_{c}$ in uniformly disordered thin
films. \label{Coulomb suppression}}
The scenario described above assumes that superconductivity remains intact
inside each grain. An alternative mechanism for the suppression of
superconductivity by Coulomb repulsion was developed by Finkelstein~\cit
{Finkelstein1987,Finkelstein1994}, building upon earlier perturbative
calculations~\cite{CoulPert}. Finkelstein effect becomes important for very
thin strongly but homogeneously disordered films, as well as quasi-1D wires
made out of such films ~\cite{Oreg1999}. Contrary to the Coulomb blockade
scenario, the system is supposed to be "uniformly disordered", with no
superstructures such as grains coupled together by weak junctions. Somewhat
similar idea was proposed~\cite{AndersonMuttalibRamakrishnan1983,Coffey} for
three-dimensional materials near the localization threshold. The essence of
Finkelstein effect is that Coulomb repulsion between electrons gets enhanced
due to slow diffusion of electrons in highly disordered film, which results
in the \textit{negative} contribution to the effective Cooper attraction
amplitude at small energy transfer $\varepsilon $:
\begin{equation}
\lambda (\varepsilon )=\lambda _{0}-\frac{1}{12\pi g}\ln \frac{1}
\varepsilon \tau _{\ast }} \label{lambdaFin}
\end{equation
where $g=h/(2e)^{2}R_{\Box }$ is dimensionless film conductance, $\lambda
_{0}$ is the "bare" Cooper attraction constant defined at the scale of Debye
frequency $\omega _{D}$, and $\tau _{\ast }=\max {\tau ,\tau (b/l)^{2}}$,
where $\tau $ and $l=v_{F}\tau $ are the mean scattering time and mean free
path, and $b$ is the film thickness. The suppression of attraction constant
Eq.(\ref{lambdaFin}) leads immediately (we assume here $\omega _{D}\sim
1/\tau _{\ast }$) to the result obtained early on in the leading order of
the perturbation theory~\cite{CoulPert}
\begin{equation}
\frac{\delta T_{c}}{T_{c}}=\frac{\delta \lambda }{\lambda ^{2}}=-\frac{1}
12\pi g}\ln^{3}\frac{1}{T_{c0}\tau _{\ast }} \label{Tc1}
\end{equation
The leading terms in the perturbation theory for $T_{c}$ can be summed by
means of the renormalization group method developed in~\cite{Finkelstein1987
. In the leading order over $1/g\ll 1$ one gets
\begin{equation}
\frac{T_{c}\tau _{\ast }}{\hbar }=\left[ \frac{\sqrt{8\pi g}-\ln (\hbar
/T_{c0}\tau _{\ast })}{\sqrt{8\pi g}+\ln (\hbar /T_{c0}\tau _{\ast })}\right]
^{\sqrt{2\pi g}}, \label{Tc-Fin}
\end{equation
According to Eq.~(\ref{Tc-Fin}), $T_{c}$ vanishes at the critical
conductance $g_{c\mathrm{F}}=\ln ^{2}(\hbar /T_{c0}\tau _{\ast })/(8\pi )$,
(which needs to be large enough for the theory to be self-consistent). At
lower (but still large compared to unity) conductances, the material never
becomes superconducting; it stays metallic at least down to very low
temperatures $T_{loc}\sim (\hbar /\tau _{\ast })\exp (-4\pi g)$ where weak
localization crosses over into the strong localization~\cite{Gershenson}.
This mechanism of superconductivity suppression, described by Eq.(\re
{Tc-Fin}), might be called "fermionic", as opposed to the "bosonic"
mechanism discussed in previous subsection~\cite{Larkin1999}. Within this
mechanism, superconductivity is destroyed at relatively large conductances
g_{cF}\geq 1$, thus a {\it direct} superconductor-insulator transition does
not seem to be a natural option.
The theory of the superconducting-insulator transition outlined above
neglects the mesoscopic fluctuations of the interaction constant, ${g
\mathbf{r)}}$. This assumption was questioned on phenomenological grounds by
Kowal and Ovadyahu \cite{Kowal1994}. The role of these fluctuations become
larger when superconductivity is strongly suppressed because as follows from
(\ref{lambdaFin}), in this regime even small mesoscopic fluctuations of $g
\mathbf{r})$ lead to a large spatial fluctuations of the effective coupling
constant $\lambda (\varepsilon ,\mathbf{r})$. In its turn, the fluctuations
of the effective coupling lead to the local spatial fluctuations of the
transition temperature $T_{c}(\mathbf{r})$, which becomes very strong,
\delta T_{c}/\bar{T_{c}}\geq 1$, for nearly-critical conductance $g\approx
g_{c\mathrm{F}}$, as shown in Ref.~\cite{Skvortsov2005}. Note that
mesoscopic fluctuations of $T_{c}$ remain small in the universal case of
short range disorder if Coulomb suppression of superconductivity is not
taken into account, even in the vicinity of the upper critical field
H_{c2}(0)$ at very low temperatures. \cit
{SpivakZhou1995,GalitskiiLarkin2001}
These results demonstrate the inherent inhomogeneity of superconducting
state near the critical point at which it is destroyed by disorder. Thus, it
seems likely that the regime close to the superconductor-insulator
transition is described by the effective model of superconducting islands
(appeared due to spatial fluctuations of local attraction constant) coupled
by weak SNS junctions. The theoretically consistent description of this
physics is still lacking, the difficulty can be traced to the absence of
tunnel barriers separating fluctuation-induced superconducting islands from
the surrounding media. In the absence of these barriers charging effects
become non-local which makes the determination of the effective Coulomb
energy a nontrivial problem, furthermore the presence of a large normal part
implies dissipative (non-local in time) dynamics of the superconducting
phase. A toy model of this type was solved in~\cit
{FeigelmanLarkinSkvortsov2001}; this model describes artificial
superconductive islands in a good contact with disordered thin film.
In conclusion, in the fermionic mechanism the Coulomb interaction is
enhanced by disorder leading to the suppression of superconductivity. The
state formed when the superconductivity is suppressed is likely to be a poor
metal characterized by a large resistivity and finite density of states at
the Fermi level.
\subsubsection{Localization versus Superconductivity. \label{Localization
versus superconductivity}}
The third alternative mechanism for the superconductor-insulator transition
is due to the localization of single electrons. In this scenario the effects
of Coulomb interaction are neglected, whereas local (in space) Cooper
attraction is treated within standard BCS approximation. Here we focus on
the case of bulk disordered materials or sufficiently thick films in which
localization remains a three dimensional effect. Abrikosov and Gor'kov~\cit
{AG1959} and Anderson~\cite{Anderson1959} have shown that potential disorder
does not affect thermodynamic properties of usual s-wave superconductors.
More precisely, this statement (based upon the presence of time-reversal
symmetry and called "Anderson theorem") means that the parameter $T_{c}\tau
/\hbar $ does not appear in BCS theory as long as magnetic field and/or
supercurrent are absent. However, localization of single-electron
eigenstates at very strong disorder leads to appearance of an additional
energy scale $\delta _{L}=1/\nu _{0}L^{3}$, where $L$ is the single-electron
localization length and $\nu _{0}$ is the density of states (per single spin
projection). The meaning of $\delta _{L}$ is just average level spacing
inside typical volume where wavefunction is localized. One expects that at
large $\delta _{L}$ the superconducting pairing between electrons is
suppressed. Competition between superconductivity and Anderson localization
was studied originally in mid-80's~\cite{MaLee,BulaSad,KapitulnikKotliar1986
. Their major conclusion was that Anderson theorem is valid and
superconductivity survives provided that the condition
\begin{equation}
T_{c}\gg \delta _{L} \label{cond1}
\end{equation
is satisfied. The reasoning leading to Eq.(\ref{cond1}) is that for Cooper
instability to develop, characteristic energy spacing between hybridized
Cooper pairs (which are supposed to be localized in the same region of size
L$) should be smaller than typical energy scale $T_{c}$ corresponding to the
Cooper instability. On the contrary, no superconducting long-range pairing
seems possible when level spacing $\delta _{L}$ strongly exceeds $T_{c}$, in
spite of the presence of inter-electron attraction (we assume that Cooper
attraction survives when the single electron states are localized as long as
$\delta _{L}$ is much smaller than Debye energy $\omega _{D}$).
We will show below that the analysis presented in~\cit
{MaLee,BulaSad,KapitulnikKotliar1986} is not complete in two important
respects. First, the absence of long-range superconductive order does not
necessarily mean that Cooper pairing is totally negligible; we will show
that in the range $T_{c}\ll \delta _{L}\ll \omega _{D}$ Cooper correlations
leads to formation of the hard-gap insulator instead of usual variable-range
one. This gap is of the same origin as the "parity gap" studied by Matveev
and Larkin~\cite{Matveev1997} in the context of ultra-small grains of good
superconductive metal. Second, the notion of \textit{eigenfunction fractalit
} (that was not known when the theory\cit
{MaLee,BulaSad,KapitulnikKotliar1986} was developed) has to be taken into
account and leads to important physical consequences. In the present paper
we are going to fill both these gaps; it will be shown that fractality of
electron eigenfunctions changes qualitative features of superconductive
state, and even modifies the condition $\delta _{L}\approx T_{c}$ for the
critical region where superconductivity is finally destroyed.
More recently the issue of competition between localization and
superconductivity was reconsidered in the important paper by Ghosal, Trivedi
and Randeria~\cite{Ghosal2001}. They considered two-dimensional lattice
model of superconductivity with moderately strong local attraction
(negative-U Hubbard model) and on-site disorder and studied it numerically
by two methods: \ by solving the self-consistent Bogolyubov-de Gennes
equations, and by solving BCS pairing equations in the basis of exact
single-electron eigenstates. They demonstrated that with increase of local
disorder superconducting state is transformed into the insulating one. The
latter possesses sharp gap in the density of states but does not show
coherence peaks. The energy gap was found to be non-monotonic as function of
the disorder strength. It was also shown that superconductivity is very
inhomogeneous in the crossover region, with disorder-generated "islands" of
large pairing amplitude. We will see below that qualitative features of the
results obtained in Ref.~\cite{Ghosal2001} are very robust and survive in a
continuum weak-coupling BCS model that we consider in the present paper (see
section \ref{Model}). The drawback of the treatment developed in Ref.~\cit
{Ghosal2001} is that it does not allow the quantitative analysis of the
physical properties as function of main parameters of the problem (coupling
strength $\lambda \ll 1$ and proximity of the Fermi-energy to the
localization edge, $|E_{F}-E_{c}|\ll E_{F}$), due to limitations imposed by
purely numerical methods. A major drawback of most conventional numerical methods
is their inability to study the regime characterized by dramatically different
energy scales, in particular $T_c \ll E_F$. Development of the method to study
this regime is the goal of the present paper.
The importance of the analytical treatment of the weak coupling regime
$T_c \ll E_F$ is demonstrated in particular, by the numerical work
\cite{Shepelyansky2002} which has studied the 3D disordered Hubbard model with
strong local attraction (4 times larger than bandwidth). In this regime the electrons
are strongly bound to each other even in translationally invariant systems. One
expects that the mobility of the formed pairs is less than the mobility of the original
electrons; this enhances the effect of the disorder. This expectation
is conformed by the data\cite{Shepelyansky2002}. We show that in the
physically relevant regime of \textit{weak} attraction, the situation is opposite:
the superconductivity survives deep in the regime of localized states.
\subsection{Experimental results on S-I transitions. \label{Experimental
results}}
Phenomenologically one should distinguish at least three types of materials
that display superconductivity suppression with the increase of the
disorder: granular systems \cite{Gerber1997,Goldman2002}, nominally
homogeneous films that exhibit superconductor-metal-insulator transition and
homogeneous films that show direct superconductor-insulator transition.\cit
{BaturinaReview2007} We shall discuss only the latter class in this paper,
the materials that exhibit it are thick (more than 20 nm) InO$_{x}$ films,
thinner (abound 4nm) TiN films and very thin (few atomic layers) Be
films. Recent work reports similar behavior also in disordered epitaxial
films of NbN with varying disorder~\cite{Pratap}. The goal of this section
is twofold: to discuss the data that allow to exclude Coulomb mechanisms
(both fermionic and bosonic) of the superconductor-insulator transition in
these films and to briefly summarize the data on these films that need
theoretical explanation. We begin with the first.
The most direct experimental evidence that allows to exclude the 'fermionic'
mechanism of superconductivity suppression discussed in section \ref{Coulomb
suppression} in homogeneously disordered films that exhibit direct
superconductor-insulator transition comes from the recent tunneling data. In
these experiments one observed that the suppression of the superconductivity
either by disorder or temperature is {\it not} accompanied by the
suppression of the gap, which remains intact or even increases reaching
2\Delta _{1}/T_{c}\in 6-9$. \cite{Sacepe2007} Instead, as the insulator is
approached, one observes the disappearance of the local coherence peaks
which, in addition, vary strongly from one point to another ( see also
Ref.~\cite{TiN-STM}) . This behavior
(and especially the temperature dependence of tunnelling conductance) is in
a striking contrast to what is expected for the fermionic mechanism. Less
direct evidence is provided by the data \cite{Shahar2004,Baturina2007a}
showing that superconductivity exists up to a very strong disorder
corresponding to $g\approx 1$, which is at least a factor of two smaller
than the one expected in the fermionic mechanism (\ref{Tc-Fin}).
As explained in section \ref{Coulomb suppression} the basis of the fermionic
mechanism is the idea that Coulomb repulsion is enhanced by disorder which
results in the effective suppression of the attractive interaction that
leads to superconductivity. The actual equations are derived in the
assumption that bare Coulomb repulsion is very strong but is reduced by
screening to the universal limit in which the effective Coulomb repulsion
constant is equal to unity. The dimensionless parameter characterizing the
strength of the Coulomb interaction and its screening is $2\sigma
/(T_{c}\kappa )$, where $\sigma =(e^{2}k_{F}/6\pi ^{2})\,(k_{F}l)$ is the
residual conductivity and $\kappa $ is the dielectric constant due to
electrons far from the Fermi energy ($|E-E_{F}|>\omega _{D}$). Coulomb
interaction is effectively strong provided that $2\sigma /(T_{c}\kappa )\sim
(\xi _{0}/a_{\mathrm{scr}})^{2}/\kappa \gg 1$, where $a_{\mathrm{scr}}$ is
the Thomas-Fermi screening length, $\xi _{0}$ is the coherence length in a
dirty superconductor. If instead $2\sigma /(T_{c}\kappa )\ll 1$ the
coefficient in front of the logarithm in (\ref{lambdaFin},\ref{Tc-Fin})
becomes small so that enhancement of Coulomb repulsion become important only
at exponentially low energy scales. In a very dirty metal (such InO$_{x}
film) with a short mean free path $k_{F}l\sim 0.3$ and low carrier density
e^{2}k_{F}\sim 5000K$ (see Ref.\cite{Shahar1992}) the ratio $\sigma
/T_{c}\sim 10$. Thus, the effects of Coulomb interaction in these films
become unimportant if dielectric constant $\kappa \gg 10$. The direct
measurements of the dielectric constant deep in the insulator regime give
\kappa \geq 30$ ~\cite{Zvi-Insulator}, one expects that it can be only
larger in the vicinity of superconductor-insulator transition, so $2\sigma
/(T_{c}\kappa )\ll 1$ in these films which makes the fermionic mechanism
irrelevant.
The microscopic origin of this large dielectric constant is likely to be due
to the low density of the carriers, $n_{e}\sim 10^{21}cm^{-3}$ in these
conductors, and the peculiar structure of their density of states in which
the Fermi level is located in a large dip.\cite{Anisimov2009}. In this
situation the density of the electrons distant from the Fermi level,
|E-E_{F}|>\omega _{D},$ is high, which results in a large screening of the
Coulomb interaction.
Now we turn to the possibility of the Coulomb driven transition similar to
the one of Josephson arrays (section \ref{Coulomb blockade}). This physics
is due to the long range nature of the Coulomb interaction; the estimates
of the associated energy scales given below show that Coulomb interaction
is sufficiently well screened even in poor conductors so that the corresponding
energy is too small compared with all other energy scales; this
rules out this possibility. These arguments are quite general and apply to
other effects that originate from the long range part of the Coulomb interaction.
We begin with energy scales that are known to be relevant experimentally for
this problem. Namely, the superconducting gap on the ordered side and activation
energy on the insulating side. These energies are large: the superconducting gap
is around $\Delta \sim 5K$ while the activation gap is even larger $T_{0}\sim
10-15K$ (see below). We now estimate the Coulomb interaction at the scale of the
superconducting coherence length, $\xi_0$. Because $T_c/E_F \lesssim 10^{-3}$ in
these materials, this length, even for a very poor conductor, cannot be too
short: $\xi _{0}\gtrsim 10nm$. Large value of the dielectric constant in the
parent insulating compounds, $\kappa \geq 30$ , implies that in the absence of
the mobile electrons with $|E-E_{F}|<\omega _{D}$, the effective charging energy
at scales $\xi _{0}$ would be $E_{c0}=e^{2}/\kappa \xi _{0}\lesssim 50\; K$.
Screening by conduction electrons with energies $T_c<|E-E_{F}|<\omega _{D}$
decreases it further. To estimate this effect we note that at scales less than
$\xi_{0}$ the properties of the electrons are similar to those at the mobility
threshold. At the threshold, the dielectric constant grows with scale, $L$,
according to the scaling law $(L/l)^x$ with $x\gtrsim 1$. This results in the
additional factor $(l/\xi _{0})^{x}$ in the effective Coulomb energy at scales
$\xi_0$ which reduces it to $E_{c}\lesssim 1K$. Thus, the effective Coulomb energy is
much smaller that all relevant energy scales and cannot be the driving force of
the transition. Note that this estimate becomes incorrect in the presence of
thin insulating barriers between the grains, thus allowing for Coulomb driven
transition in inhomogeneous materials. The absence of structural inhomogeneities
that might lead to such barriers in films of InO$_{x}$ was shown in
\cite{Kowal1994}; later studies\cite{Baturina2008} also reported the absence
of inhomogeneities in TiN films.
Another argument against Coulomb driven transition is provided by a
completely different phenomenology of the transition in the films and in
Josephson arrays: in the former one observes direct transition to the
insulating state characterized by a large gap and activation behavior of
resistivity, the transition can be driven either by the increase of disorder
or by magnetic field. In contrast, in Josephson arrays the transition driven
by the field is characterized by a large intermediate regime of temperature
independent resistivity.\cite{FazioZant,Serret2002}. Furthermore, there is
no reason to expect the disappearance of coherence peaks in some places and
not in others as one approaches the transition (reported in \cite{Sacepe2007
) in the array of superconducting grains.
Finally, Coulomb mechanism cannot explain the large value of the activation
energy in the insulating state in the vicinity of the transitions. One
expects that when charging energy becomes large enough to suppress transport
by Cooper pairs, single electrons should dominate transport which implies
activation energy equal to superconducting gap. Instead one often observes
large activation energies: $T_{0}\approx 15K$ in InO$_{x}$ films (see Fig.1
of \cite{Gantmakher1996}) and even larger in Be ones\cite{Wu05}.
We now summarize (see also the review \cite{BaturinaReview2007}) the
phenomenology of the direct superconductor--insulator transition in
homogeneous films as exhibited by three different systems: thick amorphous
InO$_{x}$ films \cit
{Shahar1992,Kowal1994,Gantmakher1996,Shahar2004,Shahar2005,Steiner2005},
thin TiN films ~\cite{Baturina2004,Baturina2007a,Baturina2007} and
extra-thin (below 1 nm) Berillium films~\cite{Wu01,Wu05,Adams2001};
somewhat similar phenomena were observed recently in the patterned Bi film
with honeycomb array of holes~\cite{Valles2009}.
\begin{enumerate}
\item {\ On insulating side of SIT, low-temperature resistivity curves show
simple activated behavior, $R(T)\propto \exp (T_{0}/T)$, which crosses over
into Mott~\cite{Kowal1994}, $R(T)\propto \exp (T_{M}/T)^{1/4}$, or
Efros-Shklovsky~\cite{Baturina2007a,Wu05}, $R(T)\propto \exp
(T_{ES}/T)^{1/2} $ variable-range hopping {\it at higher} temperatures.
This behavior is highly unusual: in hopping insulators where the activation
is frequently observed at high temperatures it crosses over to some
fractional (variable-range) behavior upon the temperature decrease. }
\item {\ At high magnetic fields the films on both sides of the SIT\ show
large negative magnetoresistance \cit
{Gantmakher1996,Shahar2004,Shahar2005,Steiner2005,Adams2001,Baturina2007,Baturina2007a
. It is important that such behavior was observed even in films
characterized by a very high activation energy. }
\item {\ At low fields all insulating films close to SIT\ show positive
magneto-resistance at low fields \cit
{Gantmakher1996,Baturina2007,Baturina2007a,Adams2001,Wu012}. }
\item {\ Ultra-low-temperature measurements on nearly-critical samples of
a-InO$_{x}$~\cite{Shahar2005} and TiN~\cite{Baturina2007a} revealed a very
sharp jump (by several orders of magnitude in current) in nonlinear $I(V)$
curves. }
\item {\ The resistance of Be \cite{Adams2001} and TiN films~\cit
{Baturina2007} approach the quantum resistance $h/e^{2}$ at very strong
magnetic fields and low temperatures. Very recently the importance of Zeeman
pair-breaking for the properties in this regime was demonstrated in Ref.
\cite{Adams2009}. }
\item The properties of the quantum critical point that separates
superconductor and insulator are not fully established. {The samples
corresponding to the critical disorder which separates superconducting and
insulating behavior~ display insulating behavior of the $R(T)$\cit
{Gantmakher1998,Baturina2007a,Wu05}; this suggests that the disorder driven
superconductor-insulator transition is not described by a self-dual theory
proposed in~\cite{MFisher1990a} or that the critical regime where this
behavior sets in is very narrow. Weakly superconducting samples can be
driven into insulating state by the application of magnetic field. Scaling
(or a lack of thereof) near this quantum critical point is subject of
controversy in both the value of the critical resistance and the scaling
exponents. Some works\cite{Baturina2007,Gantmakher1998,Sacepe2007} report
critical value of the resistance larger than }$R_{Q}=6.5k\Omega $ and
exponents that do not agree with the theoretical predictions\cit
{MFisher1990b}. In contrast, recent paper\cite{Steiner2008} reports both the
critical value of the resistance and exponent in a perfect agreement with
the theoretical predictions based on dirty boson scenario\cite{MFisher1990b}.
\end{enumerate}
These results can be summarized by the low temperature phase diagram in
(H,G)$ plane sketched in Figure~\ref{PhaseDiagram}.
\begin{figure}[tph]
\includegraphics[width=12cm]{PhaseDiagramSketch}
\caption{(Color online) Sketch of the experimental phase diagram of
homogeneously disordered films (upper panel) at $T\rightarrow 0$ and its
interpretation in the theory of superconductivity developed in the electron
system close to mobility edge (lower panel).}
\label{PhaseDiagram}
\end{figure}
It is tempting to explain these observations by the scenario in which the
disorder destroys the global superconducting coherence while preserving the
local superconducting gap, $\Delta $, even in the insulating phase. Indeed,
in the absence of pair coherence the conductivity is due to the thermally
excited fermionic quasiparticles. The density of these excitations would be
additionally suppressed by local superconductive gap $\Delta $ that is
formed at low temperatures: $n_{1}(T)\sim \exp (-\Delta /T)$ . This would
explain the crossover to activation behavior at low $T$. High magnetic field
suppresses local $\Delta (H)$ thus leading to very large negative
magnetoresistance. Low-field positive magneto-resistance could be then
associated with frustration induced by magnetic field which eliminates the
last vestiges of superconducting coherence thereby shifting the array
further into insulating side (whereas not yet affecting local
superconducting gap).
This scenario would be naturally realized in the granular material where
superconductivity remains intact in each grain whereas global coherence
appears only due to Josephson coupling which competes with charging energy
(see section \ref{Coulomb blockade})~\cite{Dubi06}. However, this assumption
of hidden granularity of the SIT\ films is not plausible for many reasons
explained above.
The correct theory should be also able to explain the large value of the
activation energy on the insulating side of the transition. Assumption that
it is due to the local superconducting gap in weakly coupled grains is not
sufficient because this energy is higher than the gap in a less disordered
film of the same series that shows superconductivity. Indeed, maximal
observed values of $T_{0}$ were about $15K$ in work\cite{Gantmakher1996} and
$11K$ in~work\cite{Kowal1994} which are significantly larger than maximal
superconducting gap $\Delta \approx 5K$ shown in Fig.1 of Ref.~\cit
{Shahar1992} (note that samples studied in~\cite{Gantmakher1996} were from
the same source as in~\cite{Shahar1992,Kowal1994}). Furthermore, this
assumption would lead to the conclusion that the transition is due to
competition between Coulomb and Josephson energies which are both much
smaller than $T_{0}$ ; because $E_{J}=G\Delta /2$ this would imply that
G\ll 1$ in a direct contradiction with the data.
Large value of the activation energy, together with inefficiency of
electron-phonon coupling at low $T$ provides a natural explanation for a
large jump in $I(V)$ observed in the insulating state~\cite{Jumps}. The
detailed predictions of the theory \cite{Jumps} were recently verified
experimentally in~\cite{SacepeShahar09}. The important ingredient of this
theory is a hard gap for single electron excitations, similar to the one
directly observed in \cite{Sacepe2007} and inferred from the resistivity
data. Thus, the phenomena of $I(V)$ jumps does not impose additional
constraints on microscopic theory. Moreover, very similar jumps were observed
previously in the system which seems to have nothing to do with superconductivity
~\cite{Sanquer1996}.
Real challenge to the theory is presented by the magnetoresistance data from
Ref.\cite{Gantmakher1996}: straightforward interpretation of the negative
magnetoresistance data as being due to superconducting gap suppression in
individual grains leads to unphysically large values of critical field
needed to destroy the superconductivity in such grains. For instance, the
field of $8$ T was observed to produced only moderate ( $R(H=8\mathrm{T
)/R(0)\approx 0.5$ )\ negative magneto-resistance in a sample characterized
by $T_{0}\approx 15\,\mathrm{K}$ (as determined in the temperature range
1.3-5$ K). Interpreting this effect as being due to the suppression of the
pairing gap, we find $T_{0}-T_{0}(H=8\mathrm{T})\approx 0.7\,\mathrm{K}$,
which is about 5\% of $T_{0}$. Interpolating this dependence we find that
the field necessary to destroy completely the superconductivity in each
grain is huge: $H_{cg}^{\exp }\sim 50-80\,\mathrm{T}$. \footnote
Data of Ref.\cite{Steiner2005} show that 32 Tesla field is not sufficient to
fully suppress negative magnetoresistance.}
Such large values of the critical fields are impossible for realistic grain
sizes. Indeed, critical orbital magnetic field for a small (radius $R<\xi $,
where $\xi =\sqrt{\hbar D/\Delta }$ is the coherence length) superconducting
grain is~\cite{Larkin1965,BeloborodovLarkinEfetov}
\begin{equation*}
H_{cg}^{est}\approx \frac{1000\,\mathrm{T}}{R\xi }
\end{equation*
where $R$ and $\xi $ are measured in nanometers. Using a typical diffusion
constant for a poor metal $D\approx 1\mathrm{cm}^{2}/\mathrm{s}$ and
allowing for a very high gap value $\Delta =10\,\mathrm{K}$, we find $\xi
=8.5\,\mathrm{nm}$. Together with the lowest bound for the grain radius
R=6\,\mathrm{nm}$ in which the distance between the levels does not exceed
the superconducting gap
\begin{equation*}
\delta =(4\nu _{0}R^{3})^{-1}<\Delta
\end{equation*
it leads to $H_{cg}^{est}\lesssim 20\,\mathrm{Tesla}$\thinspace\ which is
still much smaller than $H_{cg}^{\exp }$ above. These estimates did not take
into account the spin effect of magnetic field that would further decrease
the value of $H_{cg}^{est}$.
We conclude that the observed activation gap cannot be explained as BCS gap
in small grains composing the material: it is too wide and too stable with
respect to magnetic field.
A number of works argued that mesoscopic fluctuations might lead to the
appearance of inhomogeneous superconductivity (self-induced granularity)\ in
the vicinity of the transition even in the absence of structural granularity
\cite{Kowal1994,BaturinaReview2007,Baturina2007a}. The direct computation
shows that these speculations are correct for the fermionic mechanism of the
superconductivity suppression in two dimensions ~\cite{Skvortsov2005}. We
expect that the self-induced granularity that appears due to this mechanism
does not lead to thin insulating barriers. It is therefore characterized by
a small value of the Coulomb interaction between the 'grains'. Thus, it can
be ruled out as a mechanism of direct superconductor-insulator transition in
homogeneous films by the energy scale arguments given above.
An important unresolved issue is the nature of carriers responsible for the
transport in the insulating state: are they Cooper pairs or single
electrons?\ One expects that the presence of the superconducting gap in the
insulating state implies that the transport is dominated by Cooper pairs in
the vicinity of the transition and this was indeed observed in ultrathin Bi
films\cite{Valles2e}. However, one expects that the transport is dominated
by single electrons further in the insulating state where activation
behavior was observed. Unfortunately there are no data to confirm this.
To summarize: experimental data on SIT in amorphous materials call for a new
mechanism of a gap formation, which is somehow related to the
superconductivity, but is different from the usual BCS gap formation. In the
vicinity of SIT this mechanism should lead to a "pseudo-gap" features in
R(T)$ behavior and tunneling data.
\bigskip
\subsection{Main features of the fractal pseudospin scenario. \label{Main
features}}
The fractal pseudospin mechanism (briefly presented in~\cite{FIKY2007})
should be viewed as alternative to both 'boson' and 'fermion' scenarios.
Here we argue that it is fully compatible with the data. The key elements of
this approach are in fact quite old: (i) Anderson's reformulation~\cit
{Anderson1958} of the BCS theory in terms of "pseudo-spins", (ii) Matveev -
Larkin theory~\cite{Matveev1997} of parity gap in ultra-small
superconducting grains and (iii) fractal properties of single-electron
eigenfunctions with energies near the Anderson mobility edge~\cit
{Chalker1990,KrMut1997,MirlinReview2000}.
Qualitatively, in this scenario the electrons near the mobility edge form
strongly coupled but localized Cooper pairs (notion first introduced in~\cit
{Gantmakher1996,Gantmakher1998}, see also~\cite{Ghosal2001}) due to the
attraction of two electrons occupying the same localized orbital state.
These pairs are characterized by a large binding energy which is responsible
for the single electron gap $T_{0}$ observed in transport measurement in the
insulating state. At temperatures below $T_{0}$ the system can be described
as a collection of Anderson's $S=\frac{1}{2}$ pseudo-spins, whose $S_{j}^{z}$
components measure the Cooper pair occupation number and $S_{j}^{\pm }$
components correspond to pair creation/annihilation operators.
Superconductivity in this system is due to the tunneling of Cooper pairs
from one state to another. It is essential that it competes not with the
Coulomb repulsion but with the random energy of the pair on each orbital
state. In spin language it is described as a formation of non-zero averages
\langle S^{\pm }\rangle $ due to "off-diagonal"
S_{i}^{-}S_{j}^{+}+S_{i}^{+}S_{j}^{-}$ coupling in the effective Hamiltonian
which competes with random field in $z-$direction term $h_{i}S_{i}^{z}$.
Large values of the binding energy and off-diagonal interactions are due to
the properties of localized nearly-critical wavefunctions; the main features
of these wavefunctions are their strong correlations both in real space and
in energy space, and their sparsity in real space (see section \re
{Fractality and correlations}). The resulting phase diagram is shown in
Figure \ref{PhaseDiagram}.
The fractal pseudospin scenario has many common features with the bosonic
mechanism, but it is distinct from it in a few important respects:
\begin{itemize}
\item {pseudo-gap energy scale $T_{0}$ is independent from the collective
energy gap $\Delta $}
\item {fractal nature of individual eigenstates implies a large
"coordination number" $Z\gg 1$ of interacting pseudospins away from the
superconductor-insulator transition. Close to the transition $Z$ drops,
resulting in very inhomogeneous superconductive state and an abrupt decrease
of $T_c$.}
\item {distribution of superconducting order parameter in real space is
extremely inhomogeneous, thus usual notion of space-averaged order parameter
$\overline{\Delta }$ is useless even qualitatively, and the "Anderson
theorem" is not applicable.}
\end{itemize}
In the main part of the paper we present theoretical arguments in support of
this new scenario. We restrict our discussion to the three-dimensional
problem which is appropriate for the electron wave function behavior in most
films. It is possible that the physics in the near vicinity of the
transition is dominated by large scales where the two dimensional nature of
the films become important, the details of the crossover to this critical
regime is beyond the developed theory. We will assume below that
localization effects are not very strong, allowing for the presence of
phonon-induced attraction between electrons. Clearly, the necessary
condition for that is $\delta _{L}\ll \omega _{D}$.
The theory that we develop starts with the single electron states of the
non-interacting problem, so it is not applicable to describe the physics in
high magnetic fields where these states change significantly. Thus,
interesting physics of the metallic state with resistance approaching
h/e^{2}$ is beyond the applicability limits of our theory.
\section{Model \label{Model}}
\subsection{BCS Hamiltonian for electrons in localized eigenstates. \labe
{BCS Hamiltonian}}
We consider simplest model of space-local BCS-type electron-electron
attraction, $V_{int}=g\delta (\mathbf{r})$. It is assumed, as usual, that
this attraction is present for electrons with energies $E$ in the relatively
narrow stripe $E\in E_{F}\pm \omega _{D}$ around Fermi energy. However, we
will see below that in contrast with the usual BCS theory, the parameter
\omega _{D}$ will not enter our final results. The Hamiltonian represented
in the basis of exact single-electron eigenstates $\psi _{j}(\mathbf{r})$
become
\begin{equation}
H=\sum_{j\sigma }\xi _{j}c_{j\sigma }^{\dagger }c_{j\sigma }-\frac{\lambda }
\nu _{0}}\sum_{i,j,k,l}M_{ijkl}c_{i\uparrow }^{\dagger }c_{j\downarrow
}^{\dagger }c_{k\downarrow }c_{l\uparrow }\,, \label{Ham1}
\end{equation
where
\begin{equation}
M_{ijkl}=\int d\mathbf{r}\psi _{i}^{\ast }(\mathbf{r})\psi _{j}^{\ast }
\mathbf{r})\psi _{k}(\mathbf{r})\psi _{l}(\mathbf{r})\,, \label{Melements1}
\end{equation
$\xi _{j}=E_{j}-E_{F}$ is the single-particle energy of the eigenstate $j$
counted from Fermi level, $c_{j\sigma }$ is the corresponding electron
annihilation operator for the spin projection $\sigma $, $\nu _{0}$ is the
density of states (per single spin projection) and $\lambda =g\nu _{0}\ll 1$
is dimensionless Cooper coupling constant.
Note that writing Hamiltonian in the form (\ref{Ham1}) we omitted the
Hatree-type terms which do not contribute directly to the Cooper
instability. Such terms are known to be negligible when single-electron
states are extended (see discussion in Ref.~\cite{MaLee}); the issue of
their importance for critical and, especially, localized states is more
delicate. Below in section \ref{Transition temperature} we present results
for the superconducting transition temperature obtained with and without
account of the Hatree-type terms. The comparison, shown in Fig.~\re
{TcVersusE}, demonstrates that these terms, while changing the quantitative
results somewhat, do not affect our main qualitative conclusions. Therefore,
in order to keep the arguments as simple as possible, we neglect Hatree
terms in the main part of the following text. \footnote
However, those terms are important for a quantitative description of
superconductivity in the region of localized single-particle states, as
shown in Ref.~\cite{Ghosal2001} where 2D problem in the limit of strong
disorder and strong attraction was studied numerically. In particular, they
might lead to additional inhomogeneous broadening of the coherence peaks
observed in tunneling experiments, see section \ref{Point contact tunneling}}
Unless specified, we will not consider magnetic field effects, thus
eigenfunctions $\psi _{j}(\mathbf{r})$ can be chosen real. In the following
we will use frequently a simplified Hamiltonian (\ref{Ham1}) where only
pair-wise terms $i=j$ and $k=l$ are taken into account:
\begin{equation}
H_{2}=\sum_{j\sigma }\xi _{j}c_{j\sigma }^{\dagger }c_{j\sigma }-\frac
\lambda }{\nu _{0}}\sum_{jk}M_{jk}c_{j\uparrow }^{\dagger }c_{j\downarrow
}^{\dagger }c_{k\downarrow }c_{k\uparrow }\,, \label{Ham2}
\end{equation
where
\begin{equation}
M_{jk}=\int d\mathbf{r}\psi _{j}^{2}(\mathbf{r)}\psi _{k}^{2}(\mathbf{r})\,.
\label{Melements2}
\end{equation
Eq.(\ref{Ham2}) is the minimal Hamiltonian that includes hopping of pairs
necessary to establish a global superconducting order. It plays the same
role for our theory as the BCS Hamiltonian with $i=\mathbf{p}$, $j={-\mathbf
p}}$, $k=\mathbf{p^{\prime }}$, $l=-\mathbf{p^{\prime }}$ for usual theory
of superconductivity. We will discuss the accuracy of this approximation
below in Sec.~\ref{Ginzburg-Landau functional}
\subsubsection{Ultra-small metallic grain. \label{Ultra-small metallic grain
}
Here we rederive the known results for the model (\ref{Ham2}) applied to
ultra-small metal grains; this derivation will provide the starting point
for our solution of the model Hamiltonian (\ref{Ham1}) or (\ref{Ham2}) for
the electrons with Fermi level near mobility edge.
Pairing correlations in metallic grains of very small volume $V$, with level
spacing $\delta =(\nu _{0}V)^{-1}$ comparable to the bulk superconductive
gap $\Delta $ were considered in many papers, see review~\cite{UltraSmall}.
The issue which is most important for this work is the parity gap introduced
in~\cite{Matveev1997} to characterize pairing effects in ultra-small grains
with $\delta \ll \Delta $. The work \cite{Matveev1997} assumed the
simplified Hamiltonian (\ref{Ham2}) with identical matrix elements
M_{jk}=1/V$.
More complete treatment of a weak electron-electron interaction in small
metallic grains is given by~\cite{Kurland00} where it was argued that to the
leading order in the small parameter $\tau _{0}\delta $, where $\tau _{0}$
is the flight (or diffusion) time for electron motion inside grain, all
off-diagonal terms $M_{ijkl}$ can be neglected. In the relevant terms the
indices of $M_{ijkl}$\ should be pairwise equal. Because for small grains
the wavefunctions $\psi _{j}(\mathbf{r})$ are essentially random Gaussian
variables subject only to orthogonality and normalization conditions, the
matrix elements (\ref{Melements2}) appearing in the Hamiltonian (\ref{Ham2})
are given by
\begin{equation}
M_{j\neq k}=\frac{1}{V}\,\quad M_{jj}\equiv M_{j}=\frac{3}{V}\,, \label{Mg}
\end{equation
so that the full Hamiltonian can be expressed in terms of the total number
of electrons $\hat{n}$, the total spin $\hat{\mathbf{S}}$ and the operator
\hat{T}=\sum_{k}c_{k\downarrow }c_{k\uparrow }$ related to Cooper pairing
correlations:
\begin{equation}
H_{uni}=\lambda \delta \left[ 2\hat{\mathbf{S}}^{2}-\frac{1}{2}\hat{n}^{2
\right] -\lambda \delta T^{\dagger }T \label{kurland}
\end{equation
It is essential for the validity of (\ref{kurland}) that all matrix elements
$M_{jk}$ with $i\neq j$ are equal to $1/V$, while diagonal terms are three
times larger, only in this case it is possible to represent Eq.(\ref{kurland
) in terms of the total density, spin and pairing operators.
Attractive interaction implies that $S=0$ in the ground state. Because $n$
is conserved for isolated grain the properties of this model in $n=0$, $S=0$
sector are equivalent to the properties of the simplified Matveev-Larkin
model (\ref{Ham2}) which takes into account only the last term in (\re
{kurland}). \ The first term in (\ref{kurland}) is important for the correct
evaluation of the coefficient of the interaction term with $j=k$ in Eq.(\re
{Ham2}) because only $1/3$ of it should be assigned to the interaction in
the Cooper channel, since other $2/3$ contribute to the "n" and "S" terms of
the Hamiltonian.
In the limit $\delta \gg \Delta $ one can use the perturbation theory with
respect to pairing Hamiltonian (\ref{kurland}). In the lowest order in
\lambda $, neglecting all terms except diagonal ones, one finds that the
energy of two identical grains with even number of electrons, $n=2k,$ and
zero spin is by $\Delta E=3\lambda \delta $ less than the energy of the same
two grains with $2k+1$ and $2k-1$ electrons and spin $1/2$. \footnote
For different grains containing different number of particles one needs to
take into account different chemical potentials in these grains but the
final conclusion remains unchanged.} It means~\cite{Matveev1997} that the
average ground-state energy of a grain with even number of electrons is
lower by "parity gap" $\Delta _{P}=\frac{3}{2}\lambda \delta $ than the
energy of the same grain with odd number of electrons. Note that Cooper
pairing contributes $1/3$ of this energy difference. This result is valid
only in the limit of a very small coupling constant $\lambda $, when all the
terms with $j\neq k$ in Eq.(\ref{Ham2}) can be neglected. In a more general
case these terms must be taken into account which leads~\cite{Matveev1997}
to the renormalization of the coefficient of the $T^{\dagger }T$ term in the
Hamiltonian (\ref{kurland}) which becomes
\begin{equation}
\lambda _{R}=\lambda /(1-\lambda \ln (\omega _{D}/\delta )).
\label{lambdaR}
\end{equation
After this renormalization the coefficient of the Cooper pairing becomes
dominant. Introducing bulk energy gap $\Delta =\omega _{D}e^{-1/\lambda }$,
one finds~\cite{Matveev1997} parity gap which is valid for all $\Delta \ll
\delta $:
\begin{equation}
\Delta _{P}=\frac{\delta }{2\ln \frac{\delta }{\Delta }}+\lambda \label{ML}
\end{equation
where the second term is due to the first term in the Hamiltonian (\re
{kurland}) and is small compared to the main term. The result (\ref{ML})
shows that parity gap grows with the decrease in the grain size. Note that
parity gap (\ref{ML}) would not appear if one does note take into account
the double-diagonal terms in (\ref{Ham2}), which are totally irrelevant in
the usual BCS theory of bulk superconductivity. Below we will find somewhat
similar behavior in the case of bulk Anderson insulators.
\subsubsection{Vicinity of the mobility edge. \label{Vicinity of the
mobility edge}}
In the bulk Anderson insulator with the Fermi energy near the mobility edge,
the typical energy scale replacing $\delta $ is
\begin{equation}
\delta _{L}=1/(\nu _{0}L_{loc}^{3}), \label{delta_L_def}
\end{equation
where $L_{loc}$ is the localization length. It was argued in~\cite{MaLee}
that localization is irrelevant for superconductivity if $T_{c}\gg \delta
_{L}$. In the opposite limit $T_{c}\ll \delta _{L}$ pairing correlations
between electrons localized on different orbitals are irrelevant.
Localization length depends on the Fermi-energy (in the scaling region
L_{loc}\gg \ell $) as
\begin{equation}
L_{loc}\approx \ell \left( \frac{E_{0}}{{E_{c}-E_{F}}}\right) ^{\nu }\,,
\label{Lloc}
\end{equation
where $E_{c}$ is the position of the mobility edge, $\nu $ is the
localization length exponent. Numerical data \cite{nu} show that in a very
narrow vicinity of the mobility edge ($({E_{c}-E_{F})/E}_{c}\lll 1$) of 3D
Anderson model the localization length obeys the scaling dependence (\re
{Lloc}) with $\nu \approx 1.57$. As we shall see below, the range of
energies relevant for the superconductor-insulator transition is relatively
wide, $({E_{F}-E_{c})/E}_{c}\lesssim 0.5$; in this broader range the
localization length follows the same scaling behavior (\ref{Lloc}) but with
as somewhat different exponent $\nu \approx 1.2$. We illustrate this by Fig.
\ref{nuFig} that shows the localization length obtained for 3D Anderson
model with Gaussian disorder (see section~\ref{Fractality and correlations}
and Eq.(\ref{G-d}) below).
\begin{figure}[tbp]
\includegraphics[width=8cm]{Inv_xi_t.eps}
\caption{Inverse localization length as function of proximity to the
mobility edge, obtained numerically for 3D Anderson model with Gaussian
disorder of width $W=4$. The values shown here were extracted from the
numerical computation of the inverse participation ratio in this model and
its conversion into the localization length by Eqs.(\protect\ref{delta_L_def}
\protect\ref{d-L-IPR}). Full line is a fit to
$1/L_{loc}=1.87\cdot (E/E_{c}-1)^{1.2}$.}
\label{nuFig}
\end{figure}
The parameter $\ell $ in Eq.(\ref{Lloc}) is the short-scale cutoff of the
order of the elastic scattering length. The associated energy scale
\begin{equation}
E_{0}=1/(\nu _{0}\ell ^{3})
\end{equation
depends on the microscopic details of the model of disorder and can be small
compared to Fermi-energy $E_{F}$ (see next subsection for the discussion of
this issue).
We will assume that Fermi energy is not too close to the mobility edge so
that
\begin{equation}
t\equiv \frac{E_{c}-E_{F}}{E_{0}}\gg \frac{T_{c}}{E_{0}} \label{tau_bound1}
\end{equation
The condition (\ref{tau_bound1}) means that localization properties of
eigenstates $\psi _{j}(\mathbf{r})$ do not vary appreciably within the
energy stripe $E_{F}\pm T_{c}$ mainly responsible for development of
superconducting correlations. Violation of the condition (\ref{tau_bound1})
implies the absence of the particle-hole symmetry in the superconducting
state; we expect anomalous Hall effect in the superconducting state and near
transition in this regime. Using Eq.(\ref{Lloc}) we find
\begin{equation}
\delta _{L}=E_{0}t^{3\nu } \label{delta_tau}
\end{equation
Because $3\nu \approx 4\gg 1$, the condition (\ref{tau_bound1}) is
compatible with $t\ll 1$ in a wide range of parameters which include both
small and large ratios $\delta _{L}/T_{c}$. In the limiting case $\delta
_{L}\ll T_{c}$ the scale set by localization is much larger than the one set
by superconductivity, so all relevant statistical properties of matrix
elements (\ref{Melements2}) can be computed at $t=0$. We will refer to this
case as the critical regime. We will see in section \ref{Cooper instability}
that deep in the limit $\delta _{L}/T_{c}\rightarrow 0$, the transition
temperature approaches its limiting value, which we denote as $T_{c}^{0}$.
In contrast with the conclusions of Ref.~\cite{MaLee}, we will find that
T_{c}^{0}$ may differ substantially from the usual BCS value $T_{c}^{\mathrm
BCS}}=\omega _{D}e^{-1/\lambda }$ for the metal with the same value of the
Cooper attraction constant. Moreover, we find that the values of $T_{c}^{0}$
are typically \textit{larger} than $T_{c}^{\mathrm{BCS}}$ for weak couplings
$\lambda \ll 1$. This unexpected result is related to the \textit{fractality}
of electron wavefunctions with energies close to the mobility edge.
\subsection{Fractality and correlations of the wave functions near the
mobility edge. \label{Fractality and correlations}}
The exact single-particle eigenfunctions $\psi _{j}(\mathbf{r})$ and
eigenvalues $E_{j}$ that enter the model Hamiltonian Eq.(\ref{Ham1})should
be found from the single-particle Hamiltonian with disorder. The
conventional models of disorder are the continuous model of free electrons
in a Gaussian random potential $U(\mathbf{r})$:
\begin{equation}
H_{1a}=\frac{\mathbf{p}^{2}}{2m}+U(\mathbf{r}), \label{cont-Ham}
\end{equation
or the tight-binding model with on-site energies $\varepsilon _{\mathbf{n}}$
being random variables with the probability distribution $\mathcal{P
(\{\varepsilon _{\mathbf{n}}\})=\prod_{\mathbf{n}}p(\varepsilon _{\mathbf{n
})$. The latter is known as the Anderson model, it is described by the
Hamiltonian
\begin{equation}
H_{1b}=\sum_{\mathbf{n}}\varepsilon _{\mathbf{n}}\,a_{\mathbf{n}}^{\dagger
}a_{\mathbf{n}}-\sum_{\mathbf{n},\mathbf{m}=\mathbf{n}+\mathbf{a}}a_{\mathbf
n}}^{\dagger }a_{\mathbf{n}+\mathbf{a}}. \label{AM}
\end{equation
The most common choice of the distribution function $p(x)$ are the box
distribution
\begin{equation}
p(\varepsilon )=\left\{
\begin{matrix}
W^{-1},\;\;\mathrm{if}\;\;x<|W/2|\cr0,\;\;\mathrm{if}\;\;x>|W/2|
\end{matrix
\right. \label{box}
\end{equation
or the Gaussian distribution
\begin{equation}
p(\varepsilon )=\frac{1}{\sqrt{2\pi }\,W}\,\mathrm{exp}\left[ -\frac
\varepsilon ^{2}}{2W^{2}}\right] . \label{G-d}
\end{equation}
Increasing the disorder parameter $W$ in the $3d$ Anderson model (\ref{AM}),
\ref{box}) at a fixed Fermi energy $E_{F}$ results in the Anderson
localization transition at the critical disorder $W=W_{c}$ (at $E_{F}=0$ the
critical value of disorder is $W_{c}=16.5$ for the box distribution, Eq.~\re
{box}). Alternatively, the localization transition occurs when $E_{F}$ is
increased at a fixed disorder $W<W_{c}$ beyond the mobility edge $E_{c}$; in
the following we will mainly use the Gaussian model, Eq.~(\ref{G-d}). The
same type of transition takes place in the continuous model Eq.(\re
{cont-Ham}). The changes in the statistics of wavefunctions resulting from
this transition do not merely reduce to their localization. Well before all
wavefunctions become localized they acquire a certain structure where a
wavefunction occupies not all available space but a certain \textit{fractal}
inside the \textit{correlation radius} $L_{\mathrm{corr}}$. The global
picture of an extended wavefunction resembles a "mosaic" made of such pieces
of fractal with the characteristics size $L_{\mathrm{corr}}$. This peculiar
phase (called the "multifractal metal" in Ref.\cite{CueKra}) appears in the
vicinity of the Anderson transition. It persists down to relatively weak
disorder as long as the decreasing correlation length $L_{\mathrm{corr}}$
exceeds a microscopic length $\ell $ which has a meaning of the minimal
length (a pixel) of the fractal structure. In Anderson model with the box
probability distribution the length $\ell \approx a\,W_{c}^{1/3}\geq 2.5a$,
where $a$ is the lattice constant; fractal effects disappear in this model
at $W<3\ll W_{c}$ only, see Ref.~\cite{CueKra}. For the continuous model
defined by Eq.(\ref{cont-Ham}) it is of the order of the elastic scattering
mean free path.
As one approaches the mobility edge or the critical value of disorder, the
correlation radius $L_{\mathrm{corr}}$ diverges so that the critical
wavefunctions are pure fractal (or, strictly speaking \textit{multifractal}
\cite{CueKra}). On the localized side of the transition the wavefunctions
inside the \textit{localization radius} $L_{\mathrm{loc}}$ resemble the one
inside an element of the mosaic structure of the multifractal metal. This
"multifractal insulator" \cite{CueKra} exists in the vicinity of the
Anderson transition and becomes an ordinary insulator at strong disorder
when $L_{\mathrm{loc}}<\ell $.
\subsubsection{Wavefunction correlations at the mobility edge: algebra of
multi-fractal states. \label{Wavefunction correlations}}
We start by describing the multi-fractal correlations of wavefunctions
exactly at the mobility edge. To avoid confusion we note that for a finite
3d$ sample of the size $L\times L\times L$ the mobility edge is smeared out.
The critical multi-fractal states live in a spectral window around $E_{c}$
of the width $\delta E\propto L^{-1/\nu }$, where $\nu $ is the exponent of
the localization (correlation) length $L_{\mathrm{loc}}(L_{\mathrm{corr
})\propto |E-E_{c}|^{-\nu }$, such that the value of $L_{\mathrm{loc}}(L_
\mathrm{corr}})$ inside this window is larger than $L$. The number of
single-particle states in this window is proportional to $L^{3(1-\frac{1}
3\nu })}$. Because $\nu $ is definitely larger than $\frac{1}{3}$ (in fact,
Harris criterion tells that $\nu \geq 2/d=2/3)$ it tends to infinity as
L\rightarrow \infty $, .
There is a vast numerical and analytical evidence \cite{MirlinNewRep} that
the critical wavefunctions at the mobility edge obeys the multifractal
statistics. This can be seen, for instance, in the behavior of the moments
of the inverse participation ratio:
\begin{equation}
P_{q}=\nu _{0}^{-1}\,\sum_{j}\int d^{d}\mathbf{r}\,|\psi _{j}(\mathbf{r
)|^{2q}\,\delta (E-E_{j}). \label{mom-IPR}
\end{equation
The moments (\ref{mom-IPR}) describe the effective volume occupied by the
the wave function. At the mobility edge they scale with the size of the
sample
\begin{equation}
\langle P_{q}\rangle \sim \ell ^{-(d-d_{q})(q-1)}L^{-d_{q}(q-1)}\propto
L^{-d_{q}(q-1)}, \label{def-multi}
\end{equation
where $d_{q}\leq 3$ is the corresponding fractal dimension. For the 3d
Anderson model of the orthogonal symmetry class (real Hamiltonian) we obtain
by numerical diagonalization the following results for the first two fractal
dimensions: \footnote
The fractal dimensions for bigger sample sizes have been studied recently by
Rodriguez, Vasquez and Roemer\cite{Roemer}. They have found $d_{2}=1.24\pm
0.07$, $d_{4}=0.63\pm 0.07$, $d_{2}^{\mathrm{typ}}=1.35\pm 0.07$, $d_{4}^
\mathrm{typ}}=1.02\pm 0.2$. They also point out on a large systematic error
for $d_{4}^{\mathrm{typ}}$ related with the finite-size effect. In view of
the fact that the critical $q_{c}\approx 2.1...2.2$ is close to 2, the
typical $d_{4}^{\mathrm{typ}}$ should be found from the condition Eq.(\re
{typ}). This gives an estimate $d_{4}^{\mathrm{typ}}=0.84\pm 0.04$.}
\begin{equation}
d_{2}\approx 1.29\pm 0.1,\,\,\,\,d_{4}\approx 0.72\pm 0.1. \label{d-2-4}
\end{equation
The fact that the fractal dimensions $d_{q}$ depend on the order of the
moment $q$ implies the \textit{multiractality} of the wave functions. The
scaling arguments show that such behavior of $\langle P_{q}\rangle $ implies
the power-law correlations of wavefunction amplitudes at different space
points:
\begin{equation}
C_{q}(0,\mathbf{r})=\langle |\psi _{j}(\mathbf{r})|^{2q}\,|\psi
_{j}(0)|^{2q}\rangle \sim L^{-2qd}\,(L/\ell )^{\beta _{q}}\,\left( \frac{L}{
}\right) ^{d-\alpha _{q}}, \label{corr-dif-r}
\end{equation
where $\ell <r<L$, and the exponents are equal to
\begin{eqnarray}
\alpha _{q} &=&d_{2q}(2q-1)-2d_{q}(q-1), \label{alpha} \\
\beta _{q} &=&2(q-1)(d-d_{q})
\label{alpha1}
\end{eqnarray
Note that the sign of $\alpha _{q}$ is positive provided that the moments
P_{q}$ are only \textit{moderately} fluctuating, so that the scaling
behavior of $\langle P_{q}^{2}\rangle $ and $\langle P_{q}\rangle ^{2}$ is
the same. This follows from the inequality
\begin{equation*}
P_{q}^{2}=\left( \sum_{\mathbf{r}}|\psi (\mathbf{r})|^{2q}\right)
^{2}>P_{2q}=\sum_{\mathbf{r}}|\psi (\mathbf{r})|^{4q}
\end{equation*
and the definition of the fractal dimensions Eq.(\ref{def-multi}). However,
from Eq.(\ref{d-2-4}) is follows that
\begin{equation}
\alpha _{2}=3d_{4}-2d_{2}\approx -0.43\pm 0.5. \label{ineqality}
\end{equation
Although the error bars are rather large, it is likely that $\alpha _{2}$ is
\textit{negative}.
This means that the second moment $P_{2}$ is strongly, not moderately
fluctuating, and the $L$-scaling of $\langle P_{2}^{2}\rangle $ is different
from that of $\langle P_{2}\rangle ^{2}$ in agreement with the early
conjecture\cite{Ioffe1985}. Indeed, our numerical simulations on the 3D
Anderson model of the orthogonal symmetry class show that
\begin{equation*}
\langle P_{2}^{2}\rangle \propto L^{-2.16\pm 0.1},\;\;\;\;\;\;\langle
P_{2}\rangle ^{2}\propto L^{-2.58\pm 0.1}.
\end{equation*
This is consistent with the observation \cite{Mirlin2002} that the
distribution function of the second moment $\mathcal{P}_{2}(P_{2})\propto
P_{2}^{-p_{2}}$ has a power-law tail with $p_{2}\approx 2.6$ at relatively
large values of $P_{2}\gg \langle P_{2}\rangle \sim L^{-d_{2}}$ (which is
cut at $P_{2}=1$). As a result the average $\langle P_{2}^{2}\rangle $ is
dominated by the far tail where the one-parameter scaling $\mathcal{P
_{2}(P_{2})=f_{2}(P_{2}/\langle P_{2}\rangle )$ is no longer true. The
distribution function $\mathcal{P}_{4}(P_{4})$ of the fourth moment $P_{4}$
possesses even stronger tail with $\mathcal{P}_{4}\propto P_{4}^{-2.0}$ at
P_{4}\gg \langle P_{4}\rangle $. In this case the average $\langle
P_{4}\rangle $ is considerably contributed by the rare events in the far
tail of the distribution.
In the situation where the assumption of moderate fluctuations of $P_{q}$ no
longer holds and the rare events are important for averaging \cit
{MirlinNewRep} the correct physical quantity is typical average $\langle
P_{q}\rangle _{\mathrm{typ}}=\mathrm{exp}[\langle \ln P_{q}\rangle ]$
instead of the usual one and the corresponding fractal dimensions should be
defined by
\begin{equation} \label{d-typ}
d_{q}^{\mathrm{typ}}\,(q-1)=-d\ln \langle P_{q}\rangle _{\mathrm{typ}}/d\ln
L.
\end{equation
Unlike the usual average, the typical average is determined by the body of
the corresponding distribution function $\mathcal{P}_{q}(P_{q})$ provided
that the tail exponent $p_{q}$ is larger than $1$. Computing the typical
averages $\langle P_{q}\rangle_{\mathrm{typ}}$ and using Eq.(\ref{d-typ}) we
obtained (see also a footnote $^4$ and Ref.~\cite{Roemer} for comparison)
for the orthogonal 3D Anderson model:
\begin{eqnarray}
d_{2}^{\mathrm{typ}} &\approx &1.40\pm 0.1,\;\;\;\;d_{4}^{\mathrm{typ
}\approx 1.08\pm 0.1, \label{alpha-typ} \\
\alpha _{2}^{\mathrm{typ}} &=&3d_{4}^{\mathrm{typ}}-2d_{2}^{\mathrm{typ
}\approx +0.44\pm 0.5. \notag
\end{eqnarray
It is interesting to note \cite{MirlinNewRep} that if the tail exponent for
a given $q$-th moment lies in the region $1<p_{q}<2$ (i.e. the typical
fractal dimension is determined by the body of the distribution for the
corresponding moment $P_{q}$ but the averaged moment is dominated by the
rare events) the dependence of the typical fractal dimension on $q$ is
linear. \footnote
This follows from the fact that the spectrum of fractal dimensions given by
the Legendre transformation $f(\alpha )=-\tau_{q}+q\alpha,\;\;\;\alpha=\frac
d\tau_{q}}{dq},\;\;\;\tau_{q}=(q-1)d_{q}$ determines the scaling with the
system size $L$ of the number of sites $M\propto L^{f(\alpha)}$ where $|\psi
\mathbf{r}_{i})|^{2}\propto L^{-\alpha}$. As a typical event cannot occur at
a number of sites $M<1$ (which is only possible for a rare event), $f^
\mathrm{typ}}(\alpha)$ found from $d^{\mathrm{typ}}_{q}$ must be either
positive (if $\langle P_{q}\rangle=\langle P_{q}\rangle_{\mathrm{typ}}$) or
zero (if $\langle P_{q} \rangle$ is dominated by the rare events and is
different from $\langle P_{q}\rangle_{\mathrm{typ}}$). In the latter case
the Legendre transformation implies $[1-qd/dq]\{(q-1)d^{\mathrm{typ
}_{q}\}=0 $, leading to the fact that $(q-1)d^{\mathrm{typ}}_{q}$ is
proportional to $q $.}
\begin{equation} \label{typ}
d_{q}^{typ}(q-1)=q\,\alpha_{+},\;\;\;\;q>q_{c},
\end{equation}
where $\alpha_{+}$ is a constant.
Substituting Eq.(\ref{typ}) into Eq.(\ref{alpha}) we immediately obtain:
\begin{equation}
\alpha _{q}^{\mathrm{typ}}=(2q-1)d_{2q}^{\mathrm{typ}}-2(q-1)d_{q}^{\mathrm
typ}}=0,\;\;\;\;q>q_{c}. \label{alpha-typ1}
\end{equation
The equality (\ref{alpha-typ1}) is valid if the condition $q>q_{c}$ is
satisfied for both terms entering (\ref{alpha}); in all other cases $\alpha
_{q}^{\mathrm{typ}}>0$. Given also that for $q>1$ $d_{2q}<d_{q}<d$, we
obtain:
\begin{equation}
d>\alpha _{q}^{\mathrm{typ}}>0. \label{pos-typ}
\end{equation
For 3D Anderson model of orthogonal symmetry class $q_{c}\approx 2.2\pm 0.1$
is very close to (but larger than) 2 (see Ref.\cite{Roemer}). This means
that in a typical sample the $r$-dependence of the correlation function Eq.
\ref{corr-dif-r}) is such that the integral over $r$ is dominated by large
distances. This will be important for the estimations of the matrix elements
done below.
Long expressions Eqs.(\ref{alpha},\ref{alpha1}) for the exponents $\alpha _{q}$ and
\beta _{q}$ reflect few simple rules: \newline
\textbf{Rule(i)}: for $r\sim L$ the two wavefunctions are statistically
independent and the result can be found using Eq.(\ref{def-multi}):
\begin{equation*}
\langle |\psi _{j}(0)|^{2q}\rangle ^{2}\sim =[L^{-d}\langle P_{q}\rangle
]^{2}\sim L^{-2d_{q}(q-1)-2d}
\end{equation*}
\textbf{Rule (ii)}: for $r\sim \ell $ the result should be the same as for
the coincident points $r=0$, i.e.
\begin{equation*}
\langle |\psi _{j}(\mathbf{0})|^{4q}\rangle =L^{-d}\langle P_{2q}\rangle
\sim L^{-d_{2q}(2q-1)-d}.
\end{equation*}
These rules can be generalized to the case of correlation of \textit
different} eigenfunctions
\begin{equation*}
C_{q}(\omega ,\mathbf{r})=\langle |\psi _{i}(\mathbf{r})|^{2q}\,|\psi
_{j}(0)|^{2q}\rangle
\end{equation*
with the energy difference $E_{i}-E_{j}=\omega $. In this case the
eigenfunctions become statistically independent for $|\mathbf{r}|>L_{\omega
} $, where
\begin{equation}
L_{\omega }^{d}=\frac{1}{\nu _{0}\omega }. \label{Lom}
\end{equation
With this correction, applying the rules (i) and (ii) one can show that at
\omega >\delta $ the only change in Eq.(\ref{corr-dif-r}) is that the $
\mathbf{r}|$-dependent factor becomes $(L_{\omega }/|\mathbf{r}|)^{d-\alpha
_{q}}$ instead of $(L/|\mathbf{r}|)^{d-\alpha _{q}}$.
This allows us to estimate the \textit{local} averages of \textit{different
wave functions}. In these averages the power law correlations will be seen
in the energy space. Indeed, substituting $|\mathbf{r}|\sim \ell $ in
(L_{\omega }/|\mathbf{r}|)^{d-\alpha _{q}}$ one obtains:
\begin{eqnarray}
C_{q}(\omega ,0) &=&\nu _{0}^{-2}\sum_{i,j}\langle |\psi _{i}(\mathbf{0
)|^{2q}\,|\psi _{j}(\mathbf{0})|^{2q}\,\delta (E-E_{i})\delta (E^{\prime
}-E_{j})\rangle \notag \label{corr-dif-E} \\
&\sim &L^{-2qd}\,(L/\ell )^{\beta _{q}}\,\left( \frac{E_{0}}{\omega }\right)
^{\gamma _{q}},\;\;\;\gamma _{q}=1-\frac{\alpha _{q}}{d}.
\end{eqnarray
where $(\nu _{0}L^{d})^{-1}=\delta <\omega <E_{0}=(\nu _{0}\ell ^{d})^{-1}$.
For $\omega <\delta $ the correlation function $C_{q}(\omega )$ saturates.
Here, essentially the same rules work: \newline
\textbf{(i)} at large energy separations $\omega \sim E_{0}$ the
wavefunctions are statistically independent and\newline
\textbf{(ii)} at small energy separations $\omega \sim \delta =(\nu
_{0}L^{d})^{-1}$ the result should be of the same order as for the one
single wavefunction ($\omega =0$).
This immediately gives rise to the relationship between $\alpha_{q}$ and
\gamma_{q}$ given in Eq.(\ref{corr-dif-E}). In particular for $q=1$ one
finds:
\begin{equation} \label{Chalk}
\gamma_{1}\equiv\gamma=1-\frac{d_{2}}{d}.
\end{equation}
The scaling relationship Eq.(\ref{Chalk}) has been first suggested by
Chalker \cite{Chalker1990} and checked numerically in a number of works \cit
{CueKra}. Fig.\ref{FigC2} gives an evidence that the scaling relationship
Eq.(\ref{corr-dif-E}) between $\gamma_{q}$ and $\alpha_{q}$ holds true for
q=2$ as well. An important point is that finite-size corrections are quite
large and should be taken into account for an accurate determination of the
exponents $\gamma_q$ (an example is presented in Figs.\ref{Cw2},\ref{E01}
below).
\begin{figure}[tbp]
\includegraphics[width=8cm]{C2-3DAMxx}
\caption{(Color online) Scaling relationship between $\protect\gamma_{2}$
and $\protect\alpha_{2}$ for the 3D Anderson model of the orthogonal
symmetry class at two different system sizes L=14 and L=24. The theoretical
slope computed from Eqs.(\protect\ref{corr-dif-E}),(\protect\ref{d-2-4})
gives $\protect\gamma_{2}\approx 1.14\pm 0.03$.}
\label{FigC2}
\end{figure}
So far we have discussed the correlation functions containing $|\psi_{i}
\mathbf{r})|^{2}$ and thus independent of the random eigenfunction \textit
phase}. The simplest \textit{phase-dependent} correlation function is
involved in the density-density correlation function:
\begin{eqnarray} \label{d-d}
K(\omega,\mathbf{r-r^{\prime }})&=&\nu_0^{-2}\sum_{i,j}\langl
\delta(E-E_{i})\delta(E^{\prime }-E_{j}) \notag \\
&\times& \psi_{i}(\mathbf{r})\psi_{i}^{*}(\mathbf{r}^{\prime })\psi_{j}^{*}
\mathbf{r})\psi_{j}(\mathbf{r}^{\prime })\rangle
\end{eqnarray}
According to Ref.\cite{Chalker1990} the Fourier-transform of this
correlation function is equal to:
\begin{equation}
K(\omega ,\mathbf{q})=\frac{1}{2\pi \nu _{0}}\,\frac{D(\omega ,q)q^{2}}
(D(\omega ,q)q^{2})^{2}+\omega ^{2}}, \label{d-d-ChAn}
\end{equation
which is a generalization of the correlation function in the diffusion
approximation for the case where the diffusion coefficient may depend on $q$
and $\omega $. The main assumption \cite{Chalker1990} here is that at the
critical point and for $q\ll \ell ^{-1}$:
\begin{equation}
D(\omega ,q)=q^{d-2}\,F(qL_{\omega })=\left\{
\begin{matrix}
\begin{array}{cc}
L_{\omega }^{2-d}, & qL_{\omega }\ll 1 \\
q^{d_{2}}\,L_{\omega }^{2-d+d_{2}}, & qL_{\omega }\gg
\end{array
\end{matrix
.\right. \label{D-ChAn}
\end{equation
One can easily see that this assumption is equivalent to the statement that
at distances $\ell <|\mathbf{r-r^{\prime }}|<L_{\omega }$ the correlation
function $K(\omega ,\mathbf{r-r^{\prime }})$ has the same $\mathbf
r-r^{\prime }}$-dependence Eq.(\ref{corr-dif-r})(but with $L\rightarrow
L_{\omega }$ as explained above) as the correlation function of the local
DoS $C_{1}(\omega ,\mathbf{r-r^{\prime }})$ that we discussed above. This
requires non-trivial \textit{phase correlations} in the eigenfunctions which
allow to replace the \textit{phase-dependent} average by its \textit
phase-independent} counterpart:
\begin{equation}
\langle \psi _{i}(\mathbf{r})\psi _{i}^{\ast }(\mathbf{r}^{\prime })\psi
_{j}^{\ast }(\mathbf{r})\psi _{j}(\mathbf{r}^{\prime })\rangle \Rightarrow
\langle |\psi _{i}(\mathbf{r})|^{2}|\psi _{j}(\mathbf{r}^{\prime})|^2\rangle
\label{replacement}
\end{equation
The replacement Eq.(\ref{replacement}) illustrates the general rule \newline
\textbf{Rule (iii)}: To estimate the averages that cannot be expressed in
terms of $|\psi |^{2}$ (\textit{phase-dependent averages}) one should find
by permutation of space and/or energy variables the corresponding \textit
phase-independent average} and apply rules (i) and (ii) in order to estimate
the latter.
One should note, however, that this rule does not apply for $|\mathbf{r
|>L_{\omega}$. At such distances the phase correlations are no longer
present, and the phase-dependent average vanishes exponentially while the
\mathbf{r}$-dependence of its phase-independent counterpart saturates.
\subsubsection{Scaling estimates for matrix elements: mobility edge. \labe
{Scaling estimates A}}
Now let us show how to use phenomenology of multifractal wavefunction
statistics to estimate the matrix elements of local interaction $M_{ijkl}$
given by Eq.(\ref{Melements1}). The simplest one $M_{i}\equiv M_{iiii}$, so
called \textit{super-diagonal} matrix element, is proportional to the
inverse participation ratio $P_{2}$:
\begin{equation} \label{i}
\langle M_{i}\rangle \approx 3 \ell^{-(d-d_{2})}\,L^{-d_{2}}.
\end{equation}
Here and below we define microscopic length-scale $\ell$ via its relation to
the upper energy cutoff $E_0 = 1/\nu_0\ell^3$. The factor 3 in the above
equation is of the same origin as in the statistics of eigenvectors of real
Gaussian random matrices. Applicability of such a relation to the case of
fractal wavefunctions was demonstrated in Ref.~\cite{MirlinFyodorov97} for
the case of Anderson transition on a Cayley tree. Qualitative picture,
emerging from their analysis, is that typical wavefunctions can be
represented as products $\psi_j(\mathbf{r}) = \chi_j(\mathbf{r})\Phi_\xi
\mathbf{r})$ where $\chi_j(\mathbf{r})$ describes fast-oscillating functions
specific for each eigenstate $j$, but insensitive to the vicinity of
Anderson transition, whereas $\Phi_\xi(\mathbf{r})$ has a meaning of smooth
envelope function, which is however sensitive to the proximity of energy
\xi $ to the mobility edge $E_c$. Combinatorial factor 3 in Eq.(\ref{i}) is
due to the fast-fluctuating ``Gaussian'' component $\chi_j(\mathbf{r})$.
The \textit{diagonal} matrix element $M_{ij}\equiv M_{ijij}$ with $i\neq j$
can be easily found from Eq.(\ref{corr-dif-E});
\begin{equation} \label{ij}
\langle M_{ij} \rangle \equiv \mathcal{V}^{-1}\, M(\omega) \approx \mathcal{
}^{-1}\, \left(\frac{E_{0}}{\omega}\right)^{\gamma},
\end{equation}
where $\mathcal{V}=L^{d}$ and $\omega = \xi_i-\xi_j$. The detailed numerical
analysis of this matrix element has been done in Ref.\cite{CueKra}. Here we
present in Fig.\ref{Cw2} data for $C_1(\omega) \equiv M(\omega)$ correlator
for 3D Anderson model with Gaussian disorder ($W=4$) at the mobility edge,
computed for three different system sizes. The plot of the density of states
for the same model is presented in Fig.~\ref{FigDoS1}. The values of $\gamma
$ extracted for all three sizes are shown in the inset, together with
extrapolation to $L\to \infty$, which leads to
\begin{equation}
\gamma \approx 0.57\pm 0.02 \label{gamma}
\end{equation}
Below we will use this value of $\gamma$ in our analysis. as well as the
value of the pre-factor
\begin{equation}
E_0 = 2.08 \pm 0.25 \label{E00}
\end{equation}
which is also extracted via large-$L$ extrapolation, as shown in Fig.\re
{E01}. In addition, we present in Fig.\ref{Cw2} data for the same
correlation function at $E=8.0$ in the localized part of the spectrum (to be
discussed below in Sec.~\ref{Matrix elements} and used in Sec.~\re
{Superconductivity with a pseudogap}). Here we just note that the
logarithmic slope of the $M(\omega)$ function (rather deep inside localized
band) is close to its value $\gamma$ for the critical eigenstates.
\begin{figure}[tbp]
\includegraphics[width=8cm]{DoS-3D-Gauss-W-40xx}
\caption{Density of states for the 3D Anderson model with Gaussian disorder
with the width $W=4$ for the system size $L=10$ (black squares) and $L=20$
(blue stars).}
\label{FigDoS1}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{CwGauss2bxx}
\caption{(Color online) Correlation function $M(\protect\omega)$ for 3DAM
with Guassian disorder and lattice sizes $L=10,14,20$ at the mobility edge
E=5.5$ (red squares, blue triangles and black diamonds) and at the energy
E=8$ inside localized band (green dots). Inset shows $\protect\gamma$ values
for $L=10.12.14.16.20$.}
\label{Cw2}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{E0vsLGauss1}
\caption{Pre-factor $E_0$ for 3DAM with Guassian disorder width $W=4$ at
the mobility edge $E_c = 5.5$, for
lattice sizes $L=10,12,14,16,20$}
\label{E01}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{3d-AM-crit-off-diagonalxx}
\caption{(Color online) Scaling with the system size of the off-diagonal
matrix element. The theoretical exponent is $-3/2$ according to Eq.(\protect
\ref{off-square22}).}
\label{FigScalingOff1}
\end{figure}
A straightforward generalization of Eq.(\ref{ij}) is the matrix element
\begin{equation} \label{multi-ij}
\int d^{d}\mathbf{r}\,\langle |\psi_{i_{1}}(\mathbf{r})|^{2}...|\psi_{i_{n}}
\mathbf{r})|^{2}\rangle \sim \mathcal{V}^{-(n-1)}\,\left(\frac{E_{0}}{\omega}
\right)^{\gamma_{n}(n-1)},
\end{equation}
where $\gamma_{n}$ is defined in Eq.(\ref{corr-dif-E}), and all energy
differences $|\xi_{i_{m}}-\xi_{i_{m^{\prime }}}|$ are assumed to be of the
same order $\omega$.
The \textit{off-diagonal} matrix element $M_{ijkl}$ with all indices
different and all energy differences of the same order $\omega $ strongly
fluctuates and therefore has zero average value. However, its average square
modulus can be computed as following. We start by writing
\begin{eqnarray}
&&\langle |M_{ijkl}|^{2}\rangle =\int d^{d}\mathbf{r}d^{d}\mathbf{r^{\prime
} \label{off-square1} \\
&&\langle \psi _{i}^{\ast }(\mathbf{r})\psi _{i}(\mathbf{r^{\prime }})\psi
_{j}^{\ast }(\mathbf{r})\psi _{j}(\mathbf{r^{\prime }})\psi _{k}^{\ast }
\mathbf{r}^{\prime })\psi _{k}(\mathbf{r})\psi _{l}^{\ast }(\mathbf
r^{\prime }})\psi _{l}(\mathbf{r})\rangle \notag
\end{eqnarray
Now we apply the rule $(iii)$ of the previous subsection and define the
phase-independent correlation function:
\begin{equation}
\langle |\psi _{i}(\mathbf{r})|^{2}|\psi _{j}(\mathbf{r}^{\prime
})|^{2}|\psi _{k}(\mathbf{r})|^{2}|\psi _{l}(\mathbf{r^{\prime }
)|^{2}\rangle \label{off-square-phase-ind}
\end{equation
Using the rules $(i),(ii)$, we estimate the phase-independent correlator as:
\begin{equation}
\frac{1}{\mathcal{V}^{4}}\left( \frac{L_{\omega }}{|\mathbf{r-r^{\prime }}|
\right) ^{d-\alpha _{2}}\;\left( \frac{L_{\omega }}{\ell }\right)
^{2(d-d_{2})} \label{off1}
\end{equation
In order to obtain an estimation for $\langle |M_{ijkl}|^{2}\rangle $ one
has to integrate Eq.(\ref{off1}) over $\mathbf{r}$ and $\mathbf{r-r^{\prime
}$. The first integration is trivial and results in the factor $\mathcal{V}
. To estimate the result of the second integration we note that according to
Eq.(\ref{pos-typ})in a typical sample the exponent $d-\alpha _{2}^{\mathrm
typ}}$ in the power law in Eq.(\ref{off1}) is smaller than $d$. This means
that the integral over $\mathbf{r-r^{\prime }}$ is dominated by large
distances $|\mathbf{r-r^{\prime }}|\sim L_{\omega }$. Finally we obtain:
\begin{equation}
\langle |M_{ijkl}|^{2}\rangle _{\mathrm{typ}}\sim \frac{L_{\omega }^{d}}
\mathcal{V}^{3}}\,\left( \frac{E_{0}}{\omega }\right) ^{2\gamma }.
\label{off-square22}
\end{equation}
\begin{figure}[tbp]
\includegraphics[width=8cm]{3d-AM-semi-diag1xx}
\caption{(Color online) Scaling with energy difference $|E_{i}-E_{j}|\sim
|E_{i}-E_{k}|\sim |E_{i}-E_{l}|\sim...\sim \protect\omega$ of the
off-diagonal $\langle |M_{ijkl}|^{2} \rangle$ (blue dots) and the product of
semi-diagonal $\langle M_{iikl}M_{jjkl}\rangle$ (black squares) matrix
elements. The theoretical exponent found from Eqs.(\protect\ref{off-square22
),(\protect\ref{off-dia}) and (\protect\ref{d-2-4})is equal to $-2.13\pm 0.1
.}
\label{FigScalingOff2}
\end{figure}
The same rules applied to the correlation function $\langle
M_{ikil}M_{jljk}\rangle$ of \textit{semi-diagonal} matrix elements give
exactly the same answer:
\begin{equation} \label{off-dia}
\langle M_{ikil}M_{jljk}\rangle_{\mathrm{typ}}\sim\frac{L_{\omega}^{d}}
\mathcal{V}^{3}}\,\left( \frac{E_{0}}{\omega}\right)^{2\gamma}.
\end{equation}
\begin{figure}[tbp]
\includegraphics[width=8cm]{3d-AM-crit-semi-off-diagonalxx}
\caption{(Color online) System size scaling of semi-diagonal matrix element
for the 3D AM. The theoretical exponent found from Eq.(\protect\ref{semi-dia
),(\protect\ref{d-2-4}) is equal to $-1.21\pm 0.02$ for the simple averaging
and $-1.24\pm 0.02$ for the typical averaging. }
\label{FigScalingOff3}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{Eq-32-33-AMcxx}
\caption{(Color online) Energy difference scaling of semi-diagonal matrix
element for the 3D AM with two system sizes $L=14$ and $L=24$. The
theoretical exponent found from Eq.(\protect\ref{semi-dia}),(\protect\re
{d-2-4}) is equal to $-1.57\pm 0.03$. }
\label{FigScalingOff4}
\end{figure}
However, the square modulus of the semi-diagonal matrix elements is much
larger:
\begin{equation}
\langle |M_{ikil}|^{2}\rangle _{\mathrm{typ}}\sim L_{\omega }^{d}\ell
^{-(d-d_{2})}\mathcal{V}^{-2-d_{2}/d}\,(E_{0}/\omega )^{\gamma }.
\label{semi-dia}
\end{equation}
The relevant phase-dependent correlation function and its phase-independent
counterpart according to rule (iii) are:
\begin{eqnarray}
&&\langle |\psi _{i}(\mathbf{r})|^{2}\psi _{k}(\mathbf{r})\psi _{l}^{\ast }
\mathbf{r})|\psi _{i}(\mathbf{r^{\prime }})|^{2}\psi _{k}^{\ast }(\mathbf{r
^{\prime })\psi _{l}(\mathbf{r^{\prime }})\rangle \Rightarrow \notag
\label{semi-off} \\
&&\langle |\psi _{i}(\mathbf{r})|^{2}|\psi _{i}(\mathbf{r^{\prime }
)|^{2}|\psi _{k}(\mathbf{r})|^{2}|\psi _{l}(\mathbf{r^{\prime }})|^{2}\rangle
\end{eqnarray
The latter for $\ell \ll |\mathbf{r-r^{\prime }}|\ll L_{\omega }$ can be
found in the following form:
\begin{eqnarray}
&&\langle |\psi _{i}(\mathbf{r})|^{2}|\psi _{i}(\mathbf{r^{\prime }
)|^{2}|\psi _{k}(\mathbf{r})|^{2}|\psi _{l}(\mathbf{r^{\prime }
)|^{2}\rangle \sim \notag \label{semi-off-conj} \\
&\sim &\frac{1}{\mathcal{V}^{4}}\,\left( \frac{L_{\omega }}{|\mathbf
r-r^{\prime }}|}\right) ^{\nu _{1}}\left( \frac{L_{\omega }}{\ell }\right)
^{\nu _{2}}\,\left( \frac{L}{|\mathbf{r-r^{\prime }}|}\right) ^{\nu _{3}}.
\end{eqnarray
At $|\mathbf{r-r^{\prime }}|>L_{\omega }$ we assume that the $\mathbf
r-r^{\prime }}$ dependence in the first factor in r.h.s. of Eq.(\re
{semi-off-conj}) saturates.
Indeed, according to rule (i), at $|\mathbf{r-r^{\prime }}|\sim L\gg
L_{\omega}$ the following decoupling can be done:
\begin{eqnarray} \label{semi-off-large}
&&\mathcal{V}^{4}\langle |\psi_{i}(\mathbf{r})|^{2}|\psi_{i}(\mathbf
r^{\prime }})|^{2}|\psi_{k}(\mathbf{r})|^{2}|\psi_{l}(\mathbf{r^{\prime }
)|^{2} \rangle\approx \notag \\
&& \mathcal{V}^{4}\langle |\psi_{i}(\mathbf{r})|^{2}|\psi_{k}(\mathbf{r
)|^{2}\rangle \langle|\psi_{i}(\mathbf{r^{\prime }})|^{2}|\psi_{l}(\mathbf
r^{\prime }})|^{2} \rangle\sim \left( \frac{E_{0}}{\omega}\right)^{2\gamma}.
\end{eqnarray}
This suggests that $\nu_{2}=2(d-d_{2})$.
At $|\mathbf{r-r^{\prime }}|\sim \ell$ according to rule (ii) all
wavefunctions are effectively in one space point:
\begin{eqnarray} \label{semi-off-small}
&& \mathcal{V}^{4}\langle |\psi_{i}(\mathbf{r})|^{2}|\psi_{i}(\mathbf
r^{\prime }})|^{2}|\psi_{k}(\mathbf{r})|^{2}|\psi_{l}(\mathbf{r^{\prime }
)|^{2} \rangle \notag \\
&\approx & \mathcal{V}^{4}\langle |\psi_{i}(\mathbf{r})|^{4}|\psi_{k}
\mathbf{r})|^{2}|\psi_{l}(\mathbf{r})|^{2} \rangle \notag \\
&\sim &\left( \frac{L}{\ell}\right)^{d-d_{2}}\,\left( \frac{L_{\omega}}{\ell
\right)^{-3d_{4}+d_{2}+2d}
\end{eqnarray}
Comparing Eqs.(\ref{semi-off-small}),(\ref{semi-off-large}) with Eq.(\re
{semi-off-conj}) we find:
\begin{equation} \label{nu-nu}
\nu_{1}=3(d_{2}-d_{4}),\;\;\;\;\nu_{2}=2(d-d_{2}),\;\;\;\;\;\nu_{3}=d-d_{2}.
\end{equation}
The $|\mathbf{r-r^{\prime }}|$ dependence in Eq.(\ref{semi-off-conj}) is the
power law $|\mathbf{r-r^{\prime }}|^{-(\nu_{1}+\nu_{3})}$ with the exponent
(d-\alpha_{2})$. As in the case of the off-diagonal matrix elements
considered above, for a typical realization of disorder this exponent is
smaller than $d$. Therefore, the integral of the phase-dependent correlation
function over $(\mathbf{r-r^{\prime }})$ which determines the matrix element
$\langle|M_{ikil}|^{2} \rangle_{\mathrm{typ}}$ is dominated by large
distances $|\mathbf{r-r^{\prime }}|\sim L_{\omega}$ (for $|\mathbf
r-r^{\prime }}|>L_{\omega}$ the phase-dependent correlation function decays
exponentially). Thus plugging $|\mathbf{r-r^{\prime }}|\sim L_{\omega}$ in
Eq.(\ref{semi-off-conj}) and multiplying the result by $L_{\omega}^{d}$ and
\mathcal{V}$ (for integration over the remaining coordinate $\mathbf{r}$) we
arrive at the announced result Eq.(\ref{semi-dia}).
We conclude our analysis by noting that if instead of typical averaging the
full ensemble averaging is performed, the exponent $\alpha
_{2}=3d_{4}-2d_{2} $ may be (and in the 3D Anderson model of orthogonal
symmetry class is) negative, due to the contribution of rare untypical
realizations. In this case the principle contributions to all correlations
functions come from small distances $|\mathbf{r-r^{\prime }}|\sim \ell $.
This changes the estimates in Eqs.(\ref{off-square22}),(\ref{off-dia}),(\re
{semi-dia}): all those equations will acquire an extra factor
\begin{equation}
(\omega /E_{0})^{\alpha _{2}/d} \label{extra}
\end{equation
which explicitly depends on the fractal dimensionality $d_{4}$ and not only
on $d_{2}$. Note, however, that in the particular case of the 3D orthogonal
symmetry class, the exponent $|\alpha _{2}/d|\sim 0.1$ is extremely small,
so that an extra factor Eq.(\ref{extra}) is of order one for most of the
practical purposes.
The predictions made on the basis of rules (i)-(iii) of the algebra of
multifractal states are checked by numerical diagonalization of the 3d
Anderson model of the orthogonal symmetry class and summarized in figures
Fig.\ref{FigScalingOff1}-Fig.\ref{FigScalingOff4}. One can see a very
satisfactory agreement for exponents of various power laws which were found
numerically and derived theoretically using only one fractal dimension
d_{2} $.
Closing this subsection we conclude that the rules $(i)-(iii)$ and the
definition of multifractal dimensions Eq.(\ref{def-multi}) constitute the
full set of rules necessary to estimate any correlation function of critical
wavefunctions. This set of algebraic operations will be the main \textit
analytical} tool to deal with the strongly disordered case considered in
this paper.
\subsubsection{Scaling estimates for matrix elements: multifractal
insulator. \label{Scaling estimates B}}
Estimates for the matrix elements on the localized side of the Anderson
transition can be obtained from the corresponding formulae of the preceding
subsection provided that the localization radius is larger than the
characteristic length $\ell $ (multifractal insulator). The only
modification to be done in all the local averages $\langle ...\rangle $ is
to replace $L$ by $L_{\mathrm{loc}}$ and to add a factor $(L_{\mathrm{loc
}/L)^{nd}$ (where $n$ is the number of different eigenfunctions in the
average) that accounts for the probability for a point $\mathbf{r}$ to be
inside the localization radius of each of the wavefunctions. To apply these
simple rules for the matrix elements of different wavefunctions with the
energy separations $\omega $, one has also to make sure that all the
corresponding length scales $L_{\omega }=(\nu _{0}\omega )^{-1/d}$ are
smaller than the localization radius $L_{\mathrm{loc}}$. This sets up an
important condition which determines the frequency domain of the
multifractal correlations (\textit{multifractal frequency domain}):
\begin{equation}
E_{0}>\omega >\delta _{L}=(\nu _{0}L_{\mathrm{loc}}^{d})^{-1}.
\label{delta-xi}
\end{equation
In particular, under this condition we have:
\begin{equation}
\langle M_{i}\rangle =3\ell ^{-(d-d_{2})}\,L_{\mathrm{loc}}^{-d_{2}},
\label{Mi}
\end{equation
which is much larger than the corresponding critical value Eq.(\ref{i}). The
combinatorial factor 3 in Eq.(\ref{Mi}) arises because of the statistics of
phase of wave function as explained above, see note after Eq.(\ref{i}).
However, the estimate for the \textit{average} diagonal matrix element
M_{ij}$ does not change and is still described by Eq.(\ref{ij}). The reason
is the additional (compared to $M_{i}$) factor $(L_{\mathrm{loc}}/L)^{d}$
due to a small probability for \textit{two} wavefunctions being localized
close in space which exactly compensates for the replacement of $\mathcal{V
\rightarrow L_{\mathrm{loc}}^{d}$ in Eq.(\ref{ij}). One can easily check
that the estimates for the average square of the off-diagonal matrix element
Eq.(\ref{off-square22}), and the average of different semi-diagonal matrix
element (\ref{off-dia}) also remain unchanged under the condition Eq.(\re
{delta-xi}).
The average square of the semi-diagonal matrix element gets enhanced with
respect to the critical case Eq.(\ref{semi-dia}):
\begin{equation}
\langle |M_{ikil}|^{2}\rangle _{\mathrm{typ}}\sim \frac{1}{\mathcal{V}^{2}
\ell ^{-(d-d_{2})}\,L_{\mathrm{loc}}^{-d_{2}}\,L_{\omega
}^{d}\;(E_{0}/\omega )^{\gamma }. \label{semi-dia-ins}
\end{equation
This enhancement factor of $(L/L_{\mathrm{loc}})^{d_{2}}$ results in a
regular scaling of $\langle |M_{ikil}|^{2}\rangle \propto \mathcal{V}^{-2}$
with the total volume.
Note, however, that in all cases the typical value of a matrix element
|M^{typ}|$ of \textit{well overlapping states} is much greater than the
typical \textit{average} value $\langle |M| \rangle_{\mathrm{typ}}$, and it
can be obtained from the corresponding expression for $\langle M\rangle_
\mathrm{typ} }$ (or $\langle M^{2}\rangle_{\mathrm{typ} }$) by replacing the
total volume $\mathcal{V}$ by the localized volume $L_{\mathrm{loc}}^{d}$.
In particular, the typical value of the diagonal matrix element for well
overlapping states is:
\begin{equation}
M_{ij}^{typ}\sim L_{\mathrm{loc}}^{-d}\,\left( \frac{E_{0}}{\omega
\right)^{\gamma}.
\end{equation}
This difference between the average value of a matrix element and the
typical value for well overlapping states is due to the fact that the most
of matrix elements in insulator are very small due to poor overlap of the
corresponding states.
One can easily check that at $\omega=\delta_{L}$ the typical absolute value
of matrix elements of well overlapping states $i,j,k,l$ does not depend on
the number of different states and has a order of magnitude of the inverse
participation ratio: $|M^{typ}|\sim \ell^{-d}\,(\ell/L_{\mathrm{loc
})^{d_{2}}$. For $\omega>\delta_{L}$ the matrix elements $M^{typ}$ of well
overlapping states get suppressed, and the suppression is stronger when the
number of different states in the matrix element increases.
\subsubsection{Scaling estimates for matrix elements: multifractal metal.
\label{Scaling estimates C}}
On the metal side of the Anderson transition, the wave function is globally
not a fractal (or multifractal), as the moments $P_{q}$ are proportional to
L^{-d(q-1)}$. However, the correlations of different eigenfunctions (with
the energy difference $\omega $) show the same power-law $\omega $-behavior
as the critical eigenfunctions \cite{CueKra} provided that $\ell \ll
L_{\omega }\ll L_{\mathrm{corr}}$. The physical meaning of the correlation
length $L_{\mathrm{corr}}$ is a typical size of a fractal element of which
the entire eigenfunction support is built. Thus, locally a wavefunction in a
multifractal metal is identical to the one in a multifractal insulator
inside the localization radius. However, the global normalization $\sum_
\mathbf{r}}|\psi (\mathbf{r})|^{2}=1$ requires the reduction of $|\psi |^{2}$
by a factor of $L_{\mathrm{corr}}^{d}/\mathcal{V}$ compared to the case of
insulator. Thus, we can formulate the rule (valid provided that $\ell \ll
L_{\omega }\ll L_{\mathrm{corr}}$) for the estimation of the matrix elements
in the multifractal metal if their counterparts in the multifractal
insulator are known. One has (i) to multiply the result for insulator by a
factor $(L_{\mathrm{corr}}^{d}/\mathcal{V})^{q}$, where $q$ is the total
number of $|\psi |^{2}$ in the matrix element in order to take into account
the normalization and (ii) to eliminate the overlap probability factor $(L_
\mathrm{corr}}^{d}/\mathcal{V})^{n}$, where $n$ is the number of different
wavefunctions, which is no longer needed for the extended metal states. Thus
the overall factor to add is $(L_{\mathrm{corr}}^{d}/\mathcal{V})^{q-n}$.
This immediately leads to
\begin{eqnarray}
&&\langle M_{i}\rangle \sim \frac{1}{\mathcal{V}}\,(L_{\mathrm{corr}}/\ell
)^{d-d_{2}}, \label{mm} \\
&&\langle |M_{ikil}|^{2}\rangle _{\mathrm{typ}}\sim \frac{1}{\mathcal{V}^{3}
\,(L_{\mathrm{corr}}/\ell )^{(d-d_{2})}\,L_{\omega }^{d}\,(E_{0}/\omega
)^{\gamma }.
\end{eqnarray
The averages $\langle M_{ij}\rangle $, $\langle |M_{ijkl}|^{2}\rangle $, and
$\langle M_{ikil}M_{jljk}\rangle $ do not change and have the same order of
magnitude on the both side of the transition provided that the condition Eq.
\ref{delta-xi}) (with $L_{\mathrm{corr}}$ replacing $L_{\mathrm{loc}}$) is
respected.
\subsubsection{Matrix elements of the off-critical states beyond the
multifractal frequency domain. \label{Matrix elements}}
As we have seen in the previous subsections, in the multifractal metal and
insulator characterized by the large correlation/localization length $L_
\mathrm{loc}},L_{\mathrm{corr}}\gg \ell $, the wavefunction correlations are
very similar to those of the critical multifractal states at the Anderson
transition point. However, this correspondence is only valid if the energy
separation $\omega $ between the states lies in the multifractal frequency
domain, $\delta _{L}<\omega <E_{0}$, bounded by effective level spacing,
\delta _{L}$, and high frequency cutoff, $E_{0}$ . If one or several energy
separations are beyond the multifractal frequency domain, the
frequency-dependent factors in the estimates change. If all energy
separations are larger than $E_{0}$ the frequency-dependent factors decrease
very fast, so $E_{0}$ provides the high-frequency cut-off for all wave
function correlations.
More delicate is the case where all energy separations are smaller than
\delta _{L}$. The behavior of the matrix elements in this region is
different in the multifractal metal and in the insulator. In the
multifractal metal, the $\omega $-dependent factors simply saturate \cit
{CueKra}. In contrast, in the multifractal insulator they acquire additional
powers of $\ln (\delta _{L}/\omega )$.
The logarithmic enhancement factor appearing in the insulator reflects the
Mott's physics of the resonance mixing of states; this effect is also
responsible for the logarithmic factors in the expression for the
low-frequency conductivity: $\sigma \sim \omega ^{2}\,\ln ^{d+1}(\delta
_{L}/\omega )$. The key element of this phenomenon is that localized states
with close energies cannot be considered independent even when the distance
between centers of localization is large compared to $L_{\mathrm{loc}}$. To
understand the origin of the correlations, we repeat the Dyson arguments for
the level statistics. Consider two typical localized states with small
energy separation $\delta E<\delta _{L}$ at distance $R>L_{\mathrm{loc}}$
from each other and vary the disorder potential. The states cease to become
orthogonal to each other, their typical overlap decreases as $t\sim \delta
_{L}\exp (-R/L_{\mathrm{loc}})$ with distance. The energy splitting of the
two states becomes $\omega =((\delta E)^{2}+t^{2})^{1/2}$. If $R\lesssim L_
\mathrm{loc}}\,\ln (\delta _{L}/\omega )$ the typical overlap between two
states at distance $R$ is larger than $\omega $ which implies that small
energy splittings are dominated by rare events when $t$ $\sim \delta E\sim
\omega $. This implies that in a typical situation the states at distance $R$
hybridize forming superpositions $\psi _{\pm }(\mathbf{r})=\cos \alpha _{\pm
}\psi _{1}(\mathbf{r})+\sin \alpha _{\pm }\psi _{2}(\mathbf{r})$ of the
parent states $\psi _{1,2}(\mathbf{r})$ with $\alpha _{\pm }\sim 1$. This
resonance hybridization makes even remote in space parent states mix with
each other, provided that the distance between their centers of localization
is less than the optimal one $R\lesssim R_{0}\sim L_{\mathrm{loc}}\,\ln
(\delta _{L}/\omega )\gg L_{\mathrm{loc}}$.
The definition of $R_{0}$ convenient for numerical study is given in terms
of the dipole-moment matrix element:
\begin{equation}
R_{0}^{2}(\omega )=d\,\frac{\sum_{i\neq j}x_{ij}^{2}M_{ij}\delta (\omega
-\xi _{i}+\xi _{j})}{\sum_{i\neq j}M_{ij}\delta (\omega -\xi _{i}+\xi _{j})}
\label{R-X}
\end{equation}
where
\begin{equation}
x_{ij}=2\int d^{d}\mathbf{r}\,\psi _{i}^{\ast }(\mathbf{r})\,x\,\psi _{j}
\mathbf{r}). \label{X}
\end{equation}
The definition (\ref{R-X},\ref{X}) is useful in the range of relatively
well-localized states, at $\omega \ll \delta_L$. One can easily check that
for the extreme strong localization $|\psi _{1,2}(\mathbf{r})|^{2}=\delta (
\mathbf{r}-\mathbf{r}_{1,2})$ the matrix element $x_{ij}=x_{1}-x_{2}$
corresponding to $\psi _{i,j}=\psi _{\pm }$ is equal to the $x$-component of
the distance between centers of localization of the parent states. The
results of numerical computation of (\ref{R-X}), (\ref{X}) for the 3D
Anderson model are presented in Fig.\ref{FigR-0} for $E_{F}=8.0$.
\begin{figure}[tbp]
\label{FigR-0}\includegraphics[width=8cm]{tilde-Rw-3D-hw-10c}
\caption{Logarithmic dependence of $R_{0}$ on $\protect\omega $ for the 3D
Anderson model with $W=4$ and $E_{F}=8.0$ well in the localized region. At
this value of the energy the localization radius $L_{\mathrm{loc}}=(\protec
\nu _{0}\protect\delta_{L})^{-1/3}\approx 3$ (this estimate follows from the
data for $\protect\nu_{0}$ shown in Fig.~\protect\ref{FigDoS1} and for
\protect\delta _{L}$ shown in Fig.~\protect\ref{TcDeltaPdeltaL}). The linear
in $\ln \protect\omega $ fit at small $\protect\omega $ corresponds to $R
\protect\omega )=A\ln \protect\delta _{L}/\protect\omega +B$ with $A=1.9\pm
0.05$, $B=-8.1\pm 0.4$. It is valid in the region where $R(\protect\omega )$
is larger than $2L_{\mathrm{loc}}$. }
\end{figure}
As expected, $R_{0}(\omega )$ is linear in $\ln (\omega )$ at very low
values of $\omega /\delta _{L}$. However, there is a broad transient regime
with essentially non-linear in $\ln \omega$ behavior for moderately small
\omega $. It is important to note that the values of $R(\omega )$ in this
regime are smaller or comparable than $2L_{\mathrm{loc}}$ (estimated from
the inverse participation ratio, with the use of data for the DoS value and
typical level spacing $\delta _{L}$ ) which is the minimum distance between
centers of localization where the Mott's physics of resonance mixing
strictly applies.
It was shown in Ref.\cite{CueKra} that the matrix element $M_{ij}$ is
proportional to $R_{0}^{d-1}(\omega )$. This result can be also justified by
the perturbation theory arguments similar to those used in the derivation
R_{0}\sim L_{\mathrm{loc}}\,\ln (\delta _{L}/\omega )$. Consider the parent
states that are strongly localized at sites $m,n$ and have a distance $R=
\mathbf{R}_{m}-\mathbf{R}_{n}|$ between centers of their localization. We
now vary the realization of disorder and treat the matrix element of the
corresponding change of the disorder potential $H_{mn}$ as a perturbation.
The amplitude of the eigenfunction $|\psi _{m}(\mathbf{R_{n}})|^{2}$ at a
(remote from its center of localization $\mathbf{R}_{m}$) site $\mathbf{R
_{n}$ becomes of the order of
\begin{equation*}
|\psi _{m}(\mathbf{R_{n}})|^{2}\sim \frac{|H_{nm}|^{2}}{(E_{n}-E_{m})^{2}
\sim \left( \frac{\delta _{L}}{\omega }\right) ^{2}\,e^{-R/L_{\mathrm{loc}}}.
\end{equation*
The matrix element is given by
\begin{equation*}
M_{ij}\propto \sum_{\mathrm{parent}\;\;\mathrm{states}}\sum_{r}|\psi _{m}
\mathbf{r})|^{2}|\psi _{n}(\mathbf{r})|^{2}\approx 2|
\psi _{m}(\mathbf{R}_{n})|^2\propto \left( \frac{\delta _{L}}{\omega }\right)
^{2}\int_{R_{0}}^{\infty }dR\,R^{d-1}\,e^{-R/L_{\mathrm{loc}}}.
\end{equation*
Here instead of summing over the parent states we integrate over the
distance between the centers of localization $R$ taking into account the
statistical repulsion of centers of localization at $R<R_{0}\sim L_{\mathrm
loc}}\,\ln (\delta _{L}/\omega )$. The final expression in terms of the
integral over $R$ is similar to the corresponding expression for the Mott's
frequency-dependent conductivity but differs from it by an extra $R^{2}$
because of the square of the dipole moment matrix element and an extra
\omega ^{2}$ in front of the integral. The estimate of the integral in the
limit $R_{0}\gg L_{\mathrm{loc}}$ finally gives $M_{ij}\propto R_{0}^{d-1}$
and $\sigma (\omega )\propto \omega ^{2}\,R_{0}^{d+1}$.
Summarizing the results of this analysis we conclude that the Eq.(\ref{ij})
for $M_{ij}$ in the multifractal insulator becomes
\begin{equation}
\langle M_{ij}\rangle =\mathcal{V}^{-1}\,\left( \frac{E_{0}}{\delta _{L}
\right) ^{\gamma ^{\mathrm{eff}}}\,\left\{
\begin{array}{cc}
\ln ^{2}\left( \frac{\delta _{L}}{\omega }\right) , & \omega \ll \delta _{L}
\\
\left( \frac{\delta _{L}}{\omega }\right) ^{\gamma ^{\mathrm{eff}}}, &
\omega \gg \delta _{L
\end{array
\right. \label{MI-ij}
\end{equation
A simple analytic expression that smoothly interpolates between these two
asymptotes in Eq.(\ref{MI-ij}) can be written for the function $M(\omega )
\mathcal{V}\langle M_{ij}\rangle $ with $\omega =\xi _{i}-\xi _{j}$:
\begin{equation}
M(\omega )=\frac{\left( \frac{E_{0}}{\delta _{L}}\right) ^{\gamma ^{\mathrm
eff}}}\,\ln ^{2}\left( \frac{\delta _{L}}{\omega }+c\right) }{\left( \frac
\omega }{\delta _{L}}\right) ^{\gamma ^{\mathrm{eff}}}\;\ln ^{2}\left( \frac
\delta _{L}}{\omega }+c\right) +1}, \label{interp}
\end{equation
where the constant $c>1$. One should also take into account small variations
of the effective fractal dimension $d_{2}^{\mathrm{eff}}$ as $E_{F}$ moves
away from the mobility edge or the disorder parameter $W$ moves away from
the critical value \cite{CueKra}. This results in the dependence of the
effective exponent:
\begin{equation}
\gamma ^{\mathrm{eff}}=1-d_{2}/d+a\,(\delta _{L}/E_{0}),\;\;\;\;\;\;a>0.
\label{gamma-eff}
\end{equation
The energy scale $\delta _{L}$ can be expressed through the inverse
participation ratio $\langle M_{i}\rangle $ in the form which is convenient
for numerical simulations:
\begin{equation}
\delta _{L}=E_{0}\,\left( \frac{\langle M_{i}\rangle }{3\nu _{0}\,E_{0}
\right) ^{\frac{3}{d_{2}}} \label{d-L-IPR}
\end{equation
Finally, the upper fractality scale $E_{0}$ can be found from the condition
\mathcal{V}\,\langle M_{ij}\rangle \approx 1$ at $\omega =E_{0}$.
The enhancement, Eqs.(\ref{MI-ij},\ref{interp}), of the overlap of two
localized wavefunctions results from the fact that the eigenfunctions which
are anomalously close to each other in the energy space (closer than $\delta
_{L}$ for the localized wavefunctions) are automatically well overlapping,
as they are just the symmetric and anti-symmetric combinations of the one
and the same pair of localized wavefunctions with the optimal distance
between their centers of localization being $R_{\omega } \sim L_{\mathrm{loc
}\,\ln (\delta _{L}/\omega )$. The most important consequence of this
phenomenon is that in 3D case the correlation function $M(\omega )$
continues to grow at low frequencies as $\log ^{2}\frac{\delta _{L}}{\omega }
$, which makes it possible to establish the superconductive order in some
part of the domain $T_{c}\ll \delta _{L}$, as will be discussed below in
section \ref{Superconductivity with a pseudogap}.
\section{Insulating state.\label{Insulating state}}
In this section we discuss the physical properties of the insulating state
in the vicinity of superconductor-insulator transition and show that it is
characterized by the large single particle gap which is responsible for the
activation temperature dependence of conductivity, $\sigma (T)\propto \exp
(-T_{I}/T)$, observed~in many works\cite{Shahar1992,Kowal1994,Gantmakher1996}
at low temperatures on the insulating side of the transition. We start by
assuming that Fermi energy $E_{F}$ is deep inside the region of
localized states, so that the interaction of electrons from different
localized orbitals is weak and leads merely to a small perturbation. Then
the Hamiltonian (\ref{Ham2}) can be further simplified to
\begin{equation}
H_{3}=\sum_{j\sigma }\xi _{j}c_{j\sigma }^{\dagger }c_{j\sigma
}-g\sum_{j}M_{j}c_{j\uparrow }^{\dagger }c_{j\downarrow }^{\dagger
}c_{j\downarrow }c_{j\uparrow }\,, \label{Ham3}
\end{equation
where the scaling estimate for the typical value of matrix elements
M_{j}=\int dr\psi _{j}^{4}(r)$ is given by Eq.(\ref{Mi}) above. We will
refer to the last term in (\ref{Ham3}) as to the \textit{local pairing
coupling}; formally it looks like the \textquotedblleft
negative-U\textquotedblright\ local attraction considered in~\cit
{Ghosal2001}.
The operator product in the last term in Eq.(\ref{Ham3}) is identical to the
occupation number product $n_{j\uparrow } n_{j\downarrow}$ which is equal to
1 if both available electrons states are populated, and to 0 otherwise. Thus
the only role of the interaction term in (\ref{Ham3}) is to shift down
energies of all double-occupied orbitals. Note that one does not encounter
such a term in usual theory of disordered superconductors, since $M_{i}$
vanishes in the thermodynamic limit for delocalized electronic states
(compare Eqs.(\ref{Mi}),(\ref{mm})).
Let us order all eigenstates $\psi _{j}$ according to the increase of
eigenvalues $\xi _{j}$. Then the last filled eigenstate $\psi _{m}$ of the
Fermi-sea (at $T=0$) for \textit{even} total number of electrons is defined
by inequality
\begin{equation}
2\xi _{m}-gM_{m}<0<2\xi _{m+1}-gM_{m+1} \label{Fermisea}
\end{equation
(we count all single-particle energies from the Fermi-energy). In a
macroscopic system, the energy interval in (\ref{Fermisea}) vanishes as the
inverse volume, $1/\mathcal{V}$. Within the "even subspace" of the whole
Hilbert space (i.e. each orbital is either empty of double-occupied), the
local pairing can be fully accounted for by the redefinition $\xi
_{j}\rightarrow \tilde{\xi _{j}}=\xi _{j}-\frac{g}{2}M_{j}$. However,
single-occupied orbitals are not involved in this interaction.
The increase of thermodynamic potential $\Omega $ due to addition of \textit
odd} electron to the ground-state is
\begin{eqnarray}
\delta \Omega _{\mathrm{oe}} &=&\xi _{m+1}=\xi _{m+1}-\tilde{\xi}_{m+1}
\tilde{\xi}_{m+1}= \\
&&\frac{g}{2}M_{m+1}+O(\mathcal{V}^{-1}) \notag \label{deltaOmega1}
\end{eqnarray
Using the Eq.(\ref{Mi}) we estimate typical value of $\delta \Omega _
\mathrm{oe}}$:
\begin{equation}
\delta \Omega _{\mathrm{oe}}^{typ}=\frac{3}{2}g\ell ^{-3}(L_{loc}/\ell
)^{-d_{2}}. \label{deltaOmega2}
\end{equation
where $L_{loc}$ is the localization length for the states with $E=E_{F}$.
Consider now two single-electron excitations on top of a fully paired
Fermi-sea defined by Eq.(\ref{Fermisea}), which are produced by transferring
of one electron from the $m$-th state to the $m+1$-th one. The energy of the
two single-particle excitations (which results from the depairing) is
\begin{equation}
2\Delta _{P}^{(m)}=\xi _{m+1}-\xi _{m}+gM_{m}=\frac{g}{2}(M_{m}+M_{m+1})+O
\mathcal{V}^{-1}) \label{exit1}
\end{equation
thus the typical value of the pairing energy $\Delta _{P}$ is also given by
Eq.(\ref{deltaOmega2}):
\begin{equation}
\Delta _{P}=\frac{3}{2}g\ell ^{-3}(L_{loc}/\ell )^{-d_{2}}=\frac{3\lambda }{
}E_{0}\left( \frac{E_{c}-E_{F}}{E_{0}}\right) ^{\nu d_{2}} \label{DeltaP}
\end{equation
In the right-hand side of (\ref{DeltaP}) we reintroduced the dimensionless
coupling $\lambda =g\nu _{0}$ and the parameter $E_{0}=1/(\nu _{0}\ell ^{3})$
which determines the high-energy cutoff for fractal correlations and we also
employed Eq.(\ref{Lloc}). The average single-particle density of states $\nu
(\varepsilon )$ is determined by the probability distribution $\mathcal{P
(M) $ of inverse participation ratios $M_{j}$. Namely, the probability to
find an excitation with the energy $\varepsilon $ coincides (in the limit
\mathcal{V}\rightarrow \infty $ ) with the probability to find a value of
inverse participation ratio $M\leq \frac{2\varepsilon }{g}$. Therefore
\begin{equation}
\nu (\varepsilon )=\nu _{0}\int_{0}^{2\varepsilon /g}\mathcal{P}(M)dM
\label{DoS}
\end{equation
We have generated the distribution function $\mathcal{P}(M)$ numerically,
using the three dimensional Anderson model with the Gaussian distribution of
local energies, Eq.(\ref{G-d}) with $W=4$. The mobility edge in such a model
is located at $|E|=5.5$.
\begin{figure}[tbp]
\includegraphics[width=8cm]{ggg6-3xx}
\caption{(Color online) Distribution of the inverse participation ratios
P_{2}$ for the 3d Anderson model at the Fermi energy $E_{F}=9.0$ on the
insulating side of SIT (the mobility edge $E_{c}=5.5$) for different system
sizes.}
\label{FigP22}
\end{figure}
Numerical data for the distribution function $\mathcal{P}(M)$ are shown in
Fig.\ref{FigP22} for several sizes $L$ and the Fermi energy $E=9.0$ in the
localized part of the spectrum. These data demonstrate a sharp drop of
\mathcal{P}(M)$ at the values of $M$ much smaller than the typical value $M^
\mathrm{typ}}$, as well as a considerable size-dependence of the slope. Fig
\ref{FigP22} shows that the low $M$ tail can be well approximated by the
exponential dependence
\begin{equation}
\mathcal{P}(M)\propto e^{-c(M^{\mathrm{typ}}/M)}\quad \mathrm{for}\,\,M\ll
M^{\mathrm{typ}} \label{PM0}
\end{equation
where the coefficient $c$ depends on the energy $E_{F}$. We used the
finite-size scaling together with an extrapolation to large $L$ limit in
order to get the values of $c$ appropriate for a macroscopic system. The
results obtained for energies $E_{F}=8.0$ and $9.0$ are shown in Fig.\re
{FigP22a}.
\begin{figure}[tbp]
\includegraphics[width=8cm]{ggg6a-2xx}
\caption{(Color online) The exponent $c$ of the small-IPR tail of the
distribution function $P(M/M^{typ})$ at the energies $E_{F}=8.0$, and $9.0$
for different sizes $L$, together with extrapolated value $c_{\infty }$ for
each energy.}
\label{FigP22a}
\end{figure}
The data shown in Fig.\ref{FigP22a} demonstrate that the extrapolated (at
L\rightarrow \infty $) value of the coefficient $c$ saturates in the
interval
\begin{equation}
c\approx 0.73\pm 0.05 \label{c}
\end{equation
for the Fermi energies deep enough in the localized band. This is in
agreement with the one-parameter character of the distribution $\mathcal{P}
(M)\equiv \mathcal{P}(M/M^{\mathrm{typ}})$ in the $L\rightarrow \infty $
limit.
In Fig. \ref{PDelta} we also present the linear scale data for the
distribution function $P(y)$ of local gaps $\Delta _{P}$,(at the
Fermi-energy $E_{F}=8.0$), normalized to the typical value $\Delta _{P}^{
\mathrm{typ}}$. Note that the data both for large and for small $y$ can be
fitted quite accurately by an analytical expression shown on the plot which
contains exponential factors and $1/y^{2}$ dependence relevant for the
intermediate $1<y<6$. This modification leads to a somewhat lager
coefficient $c$ in the exponential dependence $e^{-c/y}$.
\begin{figure}[tbp]
\includegraphics[width=8cm]{PDeltaPnor20}
\caption{(Color online) Distribution function for the normalized local gaps
y=\Delta _{P}/\Delta _{P}^{\mathrm{typ}}$, at the Fermi energy $E_{F}=8.0$
on the insulating side of SIT (the mobility edge $E_{c}=5.5$).}
\label{PDelta}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{IntP2-3D-AM-Gauss-L-32xx}
\caption{(Color online) Gap in the one-particle density of states due to
local paring of electrons on the localized states at the energies $E_F=8.0$
(black stars) and $E_F=9.0$ (blue circles). Scaling with $x = M_{i}/M_{typ}$
for relatively small $x$ is demonstrated. }
\label{FigDoS}
\end{figure}
We use the numerical data for $\mathcal{P}(M)$ to obtain, according to Eq.
\ref{DoS}), the average density of single-electron states. The results are
shown (for the two Fermi-level positions $E_{F}=8.0$ and $E_{F}=9.0$) in Fig
\ref{FigDoS}. Both plots for DoS $\nu(\varepsilon)$ nearly coincide while
expressed as functions of the reduced variables $\varepsilon/\Delta_P$ in
the most interesting region $x \leq 1.5 $. At the lowest energies DoS decays
exponentially, $\nu_{<<}(\varepsilon) \approx \nu_0
e^{-c\Delta_P/\varepsilon}$. The similar results for the low energy DoS were
obtained in ~\cite{Ghosal2001} (see Fig.16b of Ref.~\cite{Ghosal2001}).
Practically, the shape shown in Fig.\ref{FigDoS} implies the existence of a
nearly hard gap
\begin{equation}
\Delta _{1}=c_{1}\Delta _{P}=\frac{3c_{1}}{2}\lambda E_{0}\left( \frac
E_{c}-E_{F}}{E_{0}}\right) ^{\nu d_{2}}\quad c_{1}\approx 0.2-0.3
\label{hard}
\end{equation}
The average density of states could be measured directly by the tunneling
conductance via a large-area tunnel junction. The problem with such
measurement in an insulator is that electrons should be evacuated somewhere
after tunneling, otherwise the strong Coulomb blockade would make the
measurement impossible. A possible method to avoid the Coulomb blockade
might be to study tunneling conductance through a relatively thin film with
both its surfaces weakly coupled to metal contacts.
The local tunneling conductance measured by STM is expected to show a
threshold behavior with $\Delta _{P}^{(m)}$ corresponding to the state $m$
localized near the observation point in real space and with $\tilde{\xi}_{m}$
close to the Fermi energy. The local gap $\Delta _{P}^{(m)}$ fluctuates from
point to point and is distributed according to Fig.\ref{FigDoS}. We are not
aware of such measurements in the insulating state, the data on the
superconducting side of SIT are given in Ref.~\cite{Sacepe2007}, for
temperatures both above and below $T_{c}$. We present detailed results for
the tunneling conductance as a function of temperature in section~\re
{Tunneling conductance}.
Above the transition temperature the data \cite{Sacepe2007} show large
suppression of the density of states without any coherence peak that appear
at the gap edges below $T_{c}$. The absence of coherence peaks at the gap
threshold shows the qualitative difference between the local pairing gap due
to simple binding of two localized electrons and the BCS gap appearing due
to the \textit{many-body correlations} within the energy range $\sim \Delta
_{\mathrm{BCS}}$ around the Fermi-level that one expects\ in the small grain
of superconductor. In the case of fractal localized states \textit{all}
filled (double-occupied) levels are shifted down in energy, thus the total
number of states near the Fermi-energy (in a stripe of several $\Delta _{P}$
width) is \textit{not} conserved and there is no reason for appearance of a
peak above the gap. In that sense, the local pairing (due to attraction!)
plays the role similar to that of the Coulomb repulsion in suppressing the
tunneling conductance.
The above results were obtained neglecting all matrix elements of
interaction except from \textit{super-diagonal} ones, $M_{j}$. We now
discuss the validity of this approximation. Indeed, in the problem of
ultra-small grains treated by Matveev and Larkin~\cite{Matveev1997}, the
\textit{diagonal} matrix elements $M_{ij}$ were taken into account while
calculating the parity gap (\ref{ML}), via the renormalization (\ref{lambdaR
) of the local pairing energy. This renormalization was necessary (even for
\lambda \ll 1$) due to the ultraviolet (UV) divergency in the Cooper loop
diagrams. Pairing on the fractal eigenstates is of different nature: the
frequency-dependence (\ref{ij}) of matrix elements $M_{ij}$ eliminates the
UV divergence, thus the virtual $i\rightarrow j$ transitions can be
neglected and the high-energy cutoff $\omega _{D}$ is not necessary as long
as $\delta _{L}\gg T_{c}^{0}$. Here $T_{c}^{0}$ is the superconducting
transition temperature for $E_{F}$ at the mobility edge $E_{c}$ given
(within modified mean-field approximation developed below in section \re
{Cooper instability}) by Eq.(\ref{Tc11}). Using of Eqs.(\ref{d-L-IPR},\re
{hard},\ref{Tc11}) one can eliminate the model parameters $E_{0}$ and
\lambda $ and express the local gap $\Delta _{1}$ via the observable
quantities $T_{c}^{0}$ and $\delta _{L}$:
\begin{equation}
\Delta _{1}=\frac{3c_{1}}{2C^{\gamma }}\frac{\delta _{L}}{\left( {\delta _{L
}/{T_{c}^{0}}\right) ^{\gamma }}\approx 0.2\frac{\delta _{L}}{\left( {\delta
_{L}}/{T_{c}^{0}}\right) ^{\gamma }} \label{D01}
\end{equation
where for the 3D Anderson transition $\gamma =0.57$ and $C\equiv
C(0.57)\approx 3.1$. The relation (\ref{D01}) shows that in the insulating
region $\delta _{L}\gg T_{c}^{0}$ due to the nonzero fractal exponent
\gamma $ the local gap value $\Delta _{1}$ grows with $\delta _{L}$ much
weaker than $\delta _{L}$ itself. In the limit $\gamma \rightarrow 0$ Eq.
\ref{D01}) transforms into the analog of the Matveev-Larkin relation (\re
{ML}), after a sub-leading term in the denominator of (\ref{D01}) is taken
into account via a substitution $\gamma \,(\delta _{L}/T_{c}^{0})^{\gamma
}\rightarrow (\delta _{L}/T_{c}^{0})^{\gamma }-1$.
Now we turn to the discussion of the intrinsic low-temperature conductivity
of the insulator with localized pairs. The binding of electrons into local
pairs diminishes the single-particle DoS and thus suppresses the
variable-range hopping conductivity. A classical example of such an effect
is the "Coulomb gap" due to Efros and Shklovsky: the soft gap $\nu (E)\sim
E^{2}$ in the average DoS leads to a transformation of the Mott law $\sigma
(T)\propto e^{-(T_{M}/T)^{1/4}}$ into the Efros-Shklovsky law $\sigma
(T)\propto e^{-(T_{ES}/T)^{1/2}}$. In our case the low-energy states are
exponentially rare, and their account leads to the logarithmic in
temperature corrections to the activation energy determined by the hard gap
\ref{hard}):
\begin{equation}
\sigma (T)\propto \mathrm{exp}\left[ -\frac{T_{0}}{T\,\ln \left(
T_{0}/T\right) }\right] . \label{sigma}
\end{equation
where
\begin{equation}
T_{0}=\frac{c}{3}\Delta _{P}\approx 0.25\Delta _{P} \label{T01}
\end{equation
Note that the nearly activated behavior given by Eqs.(\ref{sigma}) and (\re
{T01}) can hardly be distinguished from the purely activate one with the
activation gap $\Delta _{1}$ given by (\ref{hard}) or (\ref{D01}) in the
limited temperature range available in most experiments.
We thus associate the spectral gap $\Delta _{1}$ with the measured~\cit
{Shahar1992,Kowal1994} activation energy $T_{I}$. The external parameter
(E_{c}-E_{F})$ representing the disorder strength in Eq.~(\ref{hard}) can be
replaced with an experimentally more accessible parameter $(1-\sigma /\sigma
_{c})\propto (E_{c}-E_{F})/E_{0}$. Here $\sigma $ is the high temperature
conductivity and $\sigma _{c}$ is the value of the conductivity where the
parity gap $\Delta _{P}$ first develops. We obtain
\begin{equation}
T_{I}=A(1-\sigma /\sigma _{c})^{\nu d_{2}},\quad A\approx 0.5\lambda E_{0}
\label{ti}
\end{equation
where $A$ is conductivity-independent. This equation predicts a moderate
increase of $T_{I}$ with disorder strength in agreement with the
experimental data \cite{Shahar1992}, see Fig. \ref{DataFit}.
\begin{figure}[th]
\includegraphics[width=2in]{DataFit}
\caption{Experimental values of the gap from Ref.\protect\cite{Shahar1992},
T_{I}$ (boxes) and a fit to the equation (\protect\ref{ti}) with $\protec
\nu _{\mathrm{av}}=1$, $d_{2}=1.3$.}
\label{DataFit}
\end{figure}
The only fitting parameter used in Fig. \ref{DataFit} was the value of the
constant $A=0.5\lambda E_{0}\approx 10K$. Assuming the applicability of the
BCS theory for (less disordered) superconductive InO$_{x}$ samples~\cit
{Shahar1992}, and using~\cite{ZviDan} the estimates $\omega _{D}\approx 500K$
for Debye frequency, we find $\lambda \approx 0.2$ and $E_{0}\approx 100K$.
Applicability of the scaling formulas to the above analysis is not obvious.
Indeed, the value of $1-\sigma /\sigma _{c}\approx 2/3$ for the most
insulating sample on the plot of Fig.\ref{DataFit}. We have demonstrated,
however, that the correlation function $M(\omega)$ (see Fig~\ref{Cw2})
which is closely related to the inverse participation ratio,
is approximately described by the
critical scaling even deeply inside the insulating phase. Then assuming that
Eq.(\ref{Lloc}) also holds approximately in a relatively wide interval of
1-\sigma /\sigma _{c}$ (although, perhaps, with some "average" exponent $\nu
_{\mathrm{av}}$ instead of $\nu $) we obtain an extrapolation of Eq.(\ref{ti
) for an entire interval of $\sigma /\sigma _{c}$ relevant for the
experiment. Note that the classical experiments on Si:P system have shown
that unless the special care was taken to shrink the interval of $1-\sigma
/\sigma _{c}$ to be much smaller than 1, the observed scaling of
localization radius was characterized by the exponent $\nu _{\mathrm{av}}<
1.5$, see Fig.~\ref{nuFig}. That is why it is not surprising that reasonable fit to the
experimental data shown in Fig. \ref{DataFit} corresponds to the choice $\nu
_{\mathrm{av}}=1$.
There are a few reasons why the estimated value of $E_{0}\approx 100K$
turned out to be low in comparison with the Fermi energy $E_{F}\sim 0.3eV$
\thinspace\ for amorphous InO$_{x}$ samples with the electron density in the
range of $10^{21}cm^{-3}$. This value implies that ratio $E_{0}/E_{F}\sim
0.03$ is one order of magnitude lower than our estimate (\ref{E00}) obtained
for the 3D Anderson model with Gaussian disorder (which, however, should not
be expected to be quantitatively correct for a-InO$_{x}$). First, this
estimate of $E_{0}$ uses BCS formula to related the value of the critical
temperature in the less disordered samples $T_{c}\approx 3K$ \cit
{Shahar1992} \ to the value of the coupling constant in all materials of
these series. The BCS formula is however qualitatively wrong for the samples
close to mobility edge, which are in the regime discussed in the section \re
{Cooper instability}. In this regime the critical temperature is given by
\ref{Tc11}) which leads to a much lower value of the interaction constant
\lambda \approx 0.02$. The estimate for the upper energy cutoff $E_{0}$
becomes $1000K$ which is in agreement with (\ref{E00}). This estimate
neglects the effect of thermal fluctuations discussed in section \re
{Ginzburg parameters}, the effect of these fluctuations is to reduce $T_{c}$
so the actual value of the interaction constant corresponding to the sample
close to the mobility edge with $T_{c}\approx 3K$ might be slightly larger
\lambda \approx 0.03-0.05$ corresponding to $E_{0}\approx 400-600K$. These
values of $E_{0}/E_{F}\sim 0.1-0.3$ are roughly what one expects from the
analysis of the three dimensional Anderson model. Finally, we note that
appearance of low energy scales in the insulating samples of InO$_{x}$ was
conjectured in the early paper \cite{Zvi1986} for completely different
reasons. All these arguments demonstrate that the values of $\lambda $ and
E_{0}$ that we obtain from the fit of the experimental data are roughly what
one should expect in these samples.
The reasonable fit to the data was made possible by a small value of the
exponent in Eq.(\ref{ti}) which is substantially less than $\nu d\approx 3$
due to eigenfunction fractality. Thus, the data \cite{Shahar1992} provide
the indirect evidence for the eigenfunctions fractality in the insulating
samples of InO$_{x}$.
\section{Cooper instability near the mobility edge: the formalism. \labe
{Cooper instability}}
In this section we develop approximation schemes to treat the Cooper
instability and superconducting order formation in the regime when the
Anderson theorem~\cite{AG1959,Anderson1959} is not valid due to very strong
disorder. Namely, in section \ref{Modified mean-field} we give two versions
of the modified mean-field approximations, MFA, for determination of the
superconductive transition temperature $T_{c}$; in section \re
{Ginzburg-Landau functional} we derive an analog of the Ginzburg-Landau
functional which is necessary to estimate the role of fluctuations beyond
the modified MFA, and in section \ref{Virial expansion} we give an
alternative method of $T_{c}$ determination: the virial expansion. The
nature of approximations involved in the modified MFA and the virial
expansion methods are very different, thus reasonable agreement between the
results obtained by these methods indicates the validity of both of them. We
show that in a wide region near the mobility edge the dependence of the
superconducting critical temperature on the interaction constant can be
found \textit{analytically} with the accuracy up to a pre-factor of the
order of one despite of the presence of strong thermal and mesoscopic
fluctuations. In a more disordered sample the critical temperature can be
determined using the semi-analytical approach of virial expansion applied to
the \textit{pseudo-spin Hamiltonian}. The results of this section allow to
suggest the pseudo-spin Hamiltonian that provides a unified description of
superconductivity in homogeneously disordered (\textit{non-granular})
systems including the BCS regime and the region close to the
superconductor-insulator transition.
\subsection{Modified mean-field approximation. \label{Modified mean-field}}
The goal of this subsection is to develop two versions of a modified
mean-field approximations that can be used to determine the critical
temperature of a fractal superconductor. We start from the standard
Abrikosov-Gor'kov-Anderson~\cite{AG1959,Anderson1959} mean-field equation
for $T_{c}$ of a disordered superconductor:
\begin{equation}
\Delta (\mathbf{r})=\int d^{d}\mathbf{r^{\prime }}\,K(\mathbf{r},\mathbf
r^{\prime }})\,\Delta (\mathbf{r^{\prime }}). \label{MF}
\end{equation
The kernel $K(\mathbf{r},\mathbf{r^{\prime }})$ is defined by
\begin{equation}
K(\mathbf{r},\mathbf{r^{\prime }})=\frac{g}{2}\,\sum_{ij}\eta _{ij}\,\psi
_{i}(\mathbf{r})\psi _{j}(\mathbf{r})\psi _{j}(\mathbf{r}^{\prime })\psi
_{i}(\mathbf{r}^{\prime }), \label{K}
\end{equation
where $\psi _{j}(\mathbf{r})$ are exact single-electron wavefunctions and
\begin{equation}
\eta _{ij}=\frac{\tanh (\xi _{i}/2T)+\tanh (\xi _{j}/2T)}{\xi _{i}+\xi _{j}
,\;\;\;\;\eta _{i}\equiv \eta _{ii}. \label{eta}
\end{equation}
Within the standard mean-field approximation, one neglects the spatial
variations of $\Delta(\mathbf{r})$. Then integrating over $\mathbf{r}$ in
Eq.(\ref{MF}) and using the orthogonality and normalization of different
single-particle wavefunctions one can eliminate all $\psi_{i}(\mathbf{r})$
out of Eq.(\ref{MF}). Note that it is only possible if the wave functions
are real, $\psi_{i}^{*}(\mathbf{r})=\psi_{i}(\mathbf{r})$, i.e. the
time-reversal symmetry is preserved. This leads immediately to the
well-known equation for $T_c$ in terms of DoS function $\nu(\xi)$:
\begin{equation}
1 = \frac{g}{2}\sum_i \eta_i \equiv \frac{\lambda}{2\nu_0}\int d\xi
\nu(\xi)\eta(\xi) \label{MF0}
\end{equation}
where $\eta_i \equiv \eta_{ii} = \xi_i^{-1}\tanh(\xi_i/2T)$.
However, the approximation of a constant $\Delta $ cannot be used under
strong disorder conditions near the mobility edge. Physically it is due to
strong mesoscopic fluctuations of the local DoS function $\nu (\xi ,\mathbf{
})$, see \cite{BulaSad}. Below we propose a modification of the MFA scheme
which makes it possible to account for the major part of mesoscopic DoS
fluctuations.
To construct the modified MFA, we note that the appearance of the solution
to the Eq.(\ref{MF}) is equivalent to the divergence of the series $\mathrm
Tr}(1-K)^{-1}=\sum_{n=0}^{\infty }\mathrm{Tr}K^{n}$. By d'Alambert criterion
the latter is equivalent to the condition
\begin{equation}
\lim_{n\rightarrow \infty }\frac{\mathrm{Tr}K^{n+1}}{\mathrm{Tr}K^{n}}=1.
\label{Dal}
\end{equation
Explicitly, the trace $\mathrm{Tr}K^{n}$ can be written as
\begin{eqnarray}
\mathrm{Tr}K^{n} &=&(g/2)^{n}\sum_{i_{1}j_{1}...i_{n}j_{n}}\eta
_{i_{1}j_{1}}...\eta _{i_{n}j_{n}} \\
&&\times
M_{j_{n}i_{n}i_{1}j_{1}
\,M_{j_{1}i_{1}i_{2}j_{2}}...M_{j_{n-1}i_{n-1}i_{n}j_{n}}, \notag
\label{tr-tr}
\end{eqnarray
where $M_{ijkl}$ is defined in Eq.(\ref{Melements1}).
Now we make a crucial approximation: we neglect all the \textit{off
-diagonal matrix elements} in Eq.(\ref{Melements1}) with more than two
different indices. The \textit{diagonal} matrix elements
M_{ij}=M_{iijj}=M_{ijij}=M_{ijji}$ with only two different indices will be
retained. By so doing, we neglect thermal fluctuations of the
superconductive order parameter; we also neglect a part of mesoscopic
fluctuations, whereas the most important type of the them (i.e. fluctuations
of the local density of states) will be taken into account. Thus our
approach has a meaning of a modified mean-field approximation. The neglected
off-diagonal terms determine the strength of fluctuations on top of the
modified MFA solution. We derive the corresponding Ginzburg-Landau
functional and estimate the role of such fluctuations in subsection \re
{Ginzburg-Landau functional} below.
There are two ways to impose the \textquotedblleft diagonal
constraints\textquotedblright\ the for matrix elements $M_{ijkl}$: (i) to
set $i_{l}=j_{l}$ thus imposing $n$ constrains in Eq.(\ref{tr-tr}), or (ii)
to set $j_{l}=j_{l^{\prime }}$ or $j_{l}=i_{l^{\prime }}$ with $l\neq
l^{\prime }$. In this case one imposes $2n-2$ constrains, so that only 2
summations will remain. As any summation gives a macroscopically large
number of terms, one can neglect all terms corresponding to (ii) in the
thermodynamic limit $N\rightarrow \infty $. Thus, upon neglect of
off-diagonal matrix elements one finds eventually
\begin{equation}
\mathrm{Tr}K^{n}\approx (g/2)^{n}\,\sum_{i_{1}...i_{n}}\eta _{i_{1}}\eta
_{i_{2}}...\eta _{i_{n}}\;M_{i_{n}i_{1}}M_{i_{1}i_{2}}...M_{i_{n-1}i_{n}},
\label{diag-app}
\end{equation
where $M_{ik}$ are defined in Eq.(\ref{Melements2}). It is easy to show now
that the condition given by Eq.(\ref{Dal}), with $\mathrm{Tr}K^{n}$ defined
by Eq.(\ref{diag-app}), is equivalent to the solvability condition for the
equation
\begin{equation}
\Delta _{i}=\frac{g}{2}\sum_{k}\Delta _{k}\,\eta _{k}\,M_{ki},
\label{tilde-fin}
\end{equation
which is the basis of the modified MFA we will be using below.
The set of new order parameters $\Delta _{i}$, entering Eq.(\ref{tilde-fin
), represent the superconducting "ordering field" acting onto a pair of
electrons occupying the $i$-th orbital $\psi _{i}(\mathbf{r})$. The idea of
the approximation is that, instead of using a constant (in real space) order
parameter $\Delta (\mathbf{r})=\Delta $, we assume a \textit{smooth}
dependence of $\Delta _{i}$ on the single-electron energies $\xi _{i}$.
Under such an assumption, together with the replacement of $M_{ik}$ matrix
elements by their averages (according to Eq.(\ref{ij})), the Eq.(\re
{tilde-fin}) can be transformed into the integral equation
\begin{equation}
\Delta (\xi )=\frac{\lambda }{2}\int d\zeta \eta (\zeta )M(\xi -\zeta
)\Delta (\zeta ) \label{MMFA}
\end{equation
Eq.(\ref{MMFA}) is a natural generalization of the BCS mean field equation
which follows from it at $M(\omega )=1$. However, allowing for the energy
dependence of $M(\omega )$ may lead to drastic consequences. Indeed, a
simple scaling analysis of this equation with "critical" $M(\omega
)=(E_{0}/\omega )^{\gamma }$ given by Eq.(\ref{ij}) leads to the power-law
dependence of the critical temperature on the attraction interaction
constant $\lambda $:
\begin{equation}
T_{c}\propto E_{0}\lambda ^{1/\gamma }\,, \label{Tc0}
\end{equation
where $\gamma $ is defined in Eqs.(\ref{Chalk},\ref{gamma}).
We encounter an unexpectedly strong increase of $T_{c}$ in the small
\lambda $ limit with respect to the usual result, $T_{c}\propto \mathrm{exp
(-1/\lambda )$ for a conventional BCS superconductor (with the same $\lambda
$). The price to pay for this increase is a very strong inhomogeneity in the
real space of the local pairing amplitude:
\begin{equation}
\tilde{\Delta}(\mathbf{r})=\frac{g}{2}\sum_{k}\Delta _{k}\eta _{k}\psi
_{k}^{2}(\mathbf{r}). \label{Deltar}
\end{equation
The corresponding analysis will be presented in subsection \ref{Pairing
amplitude}.
To demonstrate the validity of the solution (\ref{Deltar}), one needs to
plug it into Eq.(\ref{MF}) and to use "diagonal approximation"
\begin{equation*}
\int d^{d}\mathbf{r}\,\psi_{l}^{2}\psi_{j}(\mathbf{r})\psi_{i}(\mathbf{r
)=M_{il}\,\delta_{ij}
\end{equation*}
together with Eq.(\ref{tilde-fin}).
The transformation from Eq.(\ref{tilde-fin}) to Eq.(\ref{MMFA}) is not
exact: we replaced the fluctuating matrix elements $M_{ik}$ by their
averages according to Eq.(\ref{ij}). Thus Eq.(\ref{MMFA}) contains an
additional (with respect to Eq.(\ref{tilde-fin}) mean-field-type
approximation. One can eliminate this additional approximation. For this
purpose, let us define the new matrix $\hat{Q}$ and vectors $\phi _{i}$:
\begin{equation}
Q_{ik}=\frac{g}{2}\sqrt{\eta _{i}\eta _{k}}M_{ik}\qquad \phi _{i}=\Delta _{i
\sqrt{\eta _{i}} \label{Qmat}
\end{equation
Then the solvability condition for Eq.(\ref{tilde-fin}) transforms into the
condition that the largest eigenvalue $k_{\mathrm{max}}$ of the symmetric
matrix $\hat{Q}$ becomes equal to unity. This condition is equivalent to the
instability onset with respect to formation of a superconducting order
parameter $\Delta _{i}$, as will be seen below in subsection \ref{Transition
temperature: coefficient a(T)}.
In the thermodynamic limit the condition $k_{\mathrm{max}}=1$ implies global
superconductivity only if the corresponding eigenvector of the matrix $\hat{
}$ is \textit{extended}. Otherwise, $k_{\mathrm{max}}=1$ means formation of
local "islands" of new phase with uncorrelated local order parameters in
these islands, it does not immediately result in any global order parameter
(for the discussion of relevant examples see ~\cit
{IoffeFeigelman1985,DFI1990}). Eigenfunctions of the $\hat{Q}$ matrix may be
localized, if the effective "coordination number" (the bandwidth) $Z$ of
this matrix stays finite in thermodynamic limit. General estimate for the
coordination number is $Z\sim \nu _{0}T_{c}L_{loc}^{d}$ because the relevant
energy window populated by "active" single-particle states ( taking part in
formation of the many-body superconductive state) is of the order of $T_{c}
, and each eigenstate $|\,i>$ is coupled by the the matrix elements $M_{ij}$
to all neighbors $|\,j>$ in the localization volume $L_{loc}^{d}$. This
argument shows that as long as the single electron eigenstates are
delocalized, $E_{F}<E_{c}$ and \textit{within diagonal approximation}, the
effective coordination number $Z=\infty $ and no localization of the $\hat{Q}
$ eigenstates is possible. However, this conclusion becomes invalid when the
off-diagonal matrix elements $M_{ijkl}$ are taken into account: the lowest
eigenstates of the full kernel $K(r,r^{\prime })$ get localized. The most
important physical consequence of this localization is that, as the
temperature is decreased, the superconductivity first appears in small, well
separated regions, similarly to the situation realized in superconductors
with inhomogeneous $T_{c}$ \cite{IoffeLarkin1981}. In the regime of
delocalized single electron states a further decrease of temperature results
in a more homogeneous superconductivity. As the Fermi energy is increased
past the mobility edge, single electron states get localized and the issue
of the $\hat{Q}$-eigenfunction localization and resulting inhomogeneity
become relevant even within the "diagonal approximation", we discuss this
regime in section \ref{Superconductivity with a pseudogap}.
Here we present, as an example, the numerically obtained spectrum of the
\hat{Q}$ matrix for a finite system at the mobility edge. Fig.~\ref{FigDosK}
shows the averaged (over 2000 realizations) density of states for the $\hat{
}$ matrix generated for the 3D Anderson model with Gaussian on-site disorder
of the strength $W=4$, at the energy $E=5.5$ corresponding to the mobility
edge. Three system sizes, $L=14,19,22$ were analyzed. In a finite system the
peaks of the finite width are seen in the density of states $\rho (k)$ near
k=1$. One sees that with the system size increase, the peaks become more
narrow, so it is natural to assume that they evolve into a $\delta
-function peak in the $L\rightarrow \infty $ limit. The corresponding
temperature (chosen so that to pin the DoS peak position to $k=1$) is thus
associated with $T_{c}$.
In subsection \ref{Comparison of Tc values} we compare results for the
critical temperature $T_{c}$ obtained by the three methods: modified
analytical MFA equation (\ref{MMFA}), numerical generation of the spectrum
of the temperature-dependent matrix $\hat{Q}$, with the temperature $T=T_{c}$
adjusted so to locate the peak in $\rho (k)$ at $k=1$, and virial expansion
method, to be described in subsection \ref{Virial expansion}..
\begin{figure}[tbp]
\includegraphics[width=8cm]{kk100aGaussxx}
\caption{(Color online) Density of states $\protect\rho(k)$ of the matrix
\hat{Q}$, with peak near $k=1$, for $L=14$ (black squares), $L=19$ (blue
dots) and $L=22$ (red diamonds). }
\label{FigDosK}
\end{figure}
Our modified MFA method is similar in spirit to the usual BCS mean-field
approximation: it neglects both thermal and mesoscopic fluctuations of the
pairing field $\Delta (\mathbf{r})$ with respect to its "background
configuration". The essence of the modification we used is that our
background configuration is \textit{not uniform in real space}. Instead it
is given by the function $\tilde{\Delta}(\mathbf{r})$ defined in Eq.(\re
{Deltar}). In order to estimate the strength of fluctuations we neglected,
we will develop, in the next subsection \ref{Ginzburg-Landau functional}, a
Ginzburg-Landau description for the order parameter field configurations
\Delta (\mathbf{r})$ which are \textit{close to the background field}
\tilde{\Delta}(\mathbf{r})$.
\subsection{Ginzburg - Landau functional. \label{Ginzburg-Landau functional}}
Here we derive the free energy functional $F[\Psi (\mathbf{r})]$ defined in
terms of a \textit{smooth} envelope function
\begin{equation}
\Psi (\mathbf{r})=\Delta (\mathbf{r})/\tilde{\Delta}(\mathbf{r}) \label{Psi}
\end{equation
and valid near the transition temperature $T_{c}$ for $\Psi (\mathbf{r})$
field which smoothly varies in space.
We start by writing the action in terms of the Grassmann fields $\varphi
_{n} $
\begin{eqnarray}
S &=&\sum_{n}\varphi _{n}^{\ast }(-i\omega _{n}+H_{1})\varphi
_{n}+\sum_{n}\varphi _{-n}^{\ast }(i\omega _{n}+H_{1})\varphi _{-n}
\label{ac} \\
&&-g\sum_{n_{1},n_{2},n_{3}}\varphi _{n_{1}+n_{3}}^{\ast }\varphi
_{n_{1}-n_{3}}^{\ast }\varphi _{-n_{1}+n_{2}}\varphi _{-n_{1}-n_{2}}. \notag
\end{eqnarray
where $H_{1}$ is the single-particle part of the Hamiltonian, given e.g. by
Eq.(\ref{cont-Ham}) or Eq.(\ref{AM}); $g$ is the interaction constant and $n$
and $n_{1,2,3}$ are the Matsubara frequency summation indices. We decouple
the interaction term via the Hubbard-Stratonovich transformation involving
the auxiliary field $\Delta (\mathbf{r};\Omega _{n})$. Below we focus on the
static $n=0$ component of this field. In doing so we neglect quantum
fluctuations of the order parameter but take into account the thermal and
mesoscopic fluctuations.
We expand the Grassmann fields over the eigenstates $\psi _{i}(\mathbf{r})$
of the single-electron problem
\begin{equation}
\varphi _{n}(\mathbf{r})=\sum_{j}\chi _{j}^{(n)}\psi _{j}(\mathbf{r})
\label{varphi}
\end{equation
and represent the thermodynamic potential $S[\Delta ,\Phi ]$ as
\begin{equation}
S=\frac{1}{g}\sum_{r}|\Delta (\mathbf{r})|^{2}+\sum_{n}\sum_{i,j}\bar{\Phi
_{i}^{(n)}\,\left(
\begin{array}{cc}
(-i\omega _{n}+\xi _{i})\delta _{ij} & \Delta _{ij} \\
-\Delta _{ij}^{\ast } & (-i\omega _{n}-\xi _{i})\delta _{ij
\end{array
\right) \Phi _{j}^{(n)}, \label{HS2}
\end{equation
where
\begin{equation*}
\Phi _{i}^{(n)}=\left(
\begin{array}{c}
\chi _{i}^{(n)} \\
\chi _{i}^{(-n)\ast
\end{array
\right) ,\;\;\;\;\;\bar{\Phi}_{i}^{(n)}=(\chi _{i}^{(n)\ast }\;\;\chi
_{i}^{(-n)}).
\end{equation*
and
\begin{equation}
\Delta _{ij}=\sum_{\mathbf{r}}\Delta (\mathbf{r})\,\psi _{i}(\mathbf{r})\psi
_{j}(\mathbf{r})\equiv \sum_{k}D_{k}R_{kij}. \label{Dab}
\end{equation
Here we defined
\begin{equation}
R_{kij}=\int d\mathbf{r}\psi _{k}(\mathbf{r})\psi _{i}(\mathbf{r})\psi _{j}
\mathbf{r})\quad D_{k}=\int d\mathbf{r}\Delta (\mathbf{r})\psi _{k}(\mathbf{
}) \label{newnot}
\end{equation}
Note that the way we decoupled the interaction term is specific to the
superconductive correlations. A generic Hubbard-Stratonovich field should
contain also a component $V(\mathbf{r};\Omega _{n})$ coupled to the
combination $(\chi _{i}^{(n)})^{\ast }\chi _{i}^{(n)}$ which corresponds to
interaction in the particle-hole channels. We neglect such interactions in
this subsection and will study their effect in subsection \ref{Transition
temperature} in connection with the $S_{i}^{z}S_{j}^{z}$ terms in the
\textit{pseudo-spin} Hamiltonian.
The matrix variables $\Delta _{ij}$, which appeared naturally in the second
term of the action \textit{are not mutually independent}. Indeed, the number
of independent components of the (discretized) field $\Delta (\mathbf{r})$
scales with system volume as $\propto \mathcal{V}$, whereas the number of
matrix elements $\Delta _{ij}$ scales as $\mathcal{V}^{2}$. The matrix
elements $\Delta _{ij}$ are mutually constrained due to the ortho-normality
conditions $\int d\mathbf{r}\psi _{i}(\mathbf{r})\psi _{j}(\mathbf{r
)=\delta _{ij}$.
Now one can complete the Hubbard-Stratonovich transformation by performing
the Gaussian integration over $\Phi _{j}^{(n)}$ to obtain the $\mathrm{Tr
\ln $ of the corresponding matrix in Eq.(\ref{HS2}). In addition, one can
represent the first term in Eq.(\ref{HS2}) in terms of the coefficients
D_{i}$ of expansion of $\Delta (\mathbf{r})$ over the full set of single
particle wave functions $\psi _{i}(\mathbf{r})$. Expanding the $\mathrm{Tr
\ln $ over $D_{i}$ up to fourth order, we obtain free energy functional in
the form
\begin{equation}
F[D_{i}]=\frac{1}{g}\sum_{ij}D_{i}^{\ast }\mathcal{K}_{ij}D_{j}+\frac{1}{4
\,\sum_{ijkl}D_{i}D_{j}^{\ast }\,\mathcal{J}_{ijkl}\,D_{k}D_{l}^{\ast },
\label{F1}
\end{equation
where
\begin{equation}
\mathcal{K}_{ij}=\delta _{ij}-\frac{g}{2}\sum_{\mu \nu }R_{i\mu \nu }\eta
_{\mu \nu }R_{j\mu \nu }. \label{calK}
\end{equation
In Eqs.(\ref{F1}),(\ref{calK}) we denote
\begin{equation}
\mathcal{J}_{ijkl}=\sum_{\nu _{1},\nu _{2},\nu _{3}\nu _{4}}R_{i\nu _{4}\nu
_{1}}R_{j\nu _{1}\nu _{2}}^{\ast }R_{k\nu _{2}\nu _{3}}R_{l\nu _{3}\nu
_{4}}^{\ast }\,\;\zeta _{\nu _{1}\nu _{2},\nu _{3}\nu _{4}}\, \label{J}
\end{equation
where the function $\zeta (...)$ is defined by
\begin{eqnarray}
&&\zeta _{\nu _{1}\nu _{2},\nu _{3}\nu _{4}}=\zeta _{\nu _{2}\nu _{1},\nu
_{4}\nu _{3}}=\zeta _{\nu _{3}\nu _{4},\nu _{1}\nu _{2}}= \label{zeta} \\
&&\frac{1}{(\xi _{\nu _{1}}-\xi _{\nu _{3}})(\xi _{\nu _{4}}-\xi _{\nu _{2}}
}\,\left[ \eta _{\nu _{1}\nu _{2}}+\eta _{\nu _{3}\nu _{4}}-\eta _{\nu
_{3}\nu _{2}}-\eta _{\nu _{1}\nu _{4}}\right] . \notag
\end{eqnarray}
Below we will use Eqs.(\ref{F1})-(\ref{zeta}) to derive the effective
Ginzburg-Landau functional $F[\Psi (\mathbf{r})]$ in the form
\begin{equation}
F_{GL}[\Psi (\mathbf{r})]=\nu _{0}T_{c}^{2}\int d\mathbf{r}\left( a(\mathbf{
})\Psi ^{2}(\mathbf{r})+\frac{b}{2}\Psi ^{4}(\mathbf{r})+C|\nabla \Psi
\mathbf{r})|^{2}\right) \label{FGL}
\end{equation
The factor $\nu _{0}T_{c}^{2}$ in (\ref{FGL}) is introduced so that to keep
\Psi $ and $a(T)$ dimensionless. The physical properties of the
superconductor described by functional (\ref{FGL}) are controlled by four
parameters: $a=\langle a(\mathbf{r})\rangle $, $b$, $C$, and by the strength
$W$ of fluctuations $\delta a(\mathbf{r})=a(\mathbf{r})-a$ defined as
\begin{eqnarray}
W &=&\int d\mathbf{r}\,\langle \delta a(\mathbf{r})\delta a(\mathbf
r^{\prime }})\rangle = \label{W} \\
&=&\mathcal{V}^{-1}\int d\mathbf{r}_{1}d\mathbf{r}_{2}d\mathbf{r}_{3}
\mathbf{r}_{4}\langle \delta K(\mathbf{r}_{1},\mathbf{r_{2}})\delta K
\mathbf{r}_{3},\mathbf{r}_{4})\rangle . \notag
\end{eqnarray
where $\delta K(\mathbf{r,r^{\prime }})=K(\mathbf{r,r^{\prime }})-\langle K
\mathbf{r,r^{\prime }})\rangle $. We are able to describe the spatial
fluctuation by one parameter, $W$, because $\delta a(\mathbf{r})$
correlations are short-ranged compared to the typical scale of $\Psi
\mathbf{r})$ variations.
\subsubsection{Transition temperature: coefficient $a(T)$. \label{Transition
temperature: coefficient a(T)}}
We begin by evaluating the coefficient $a$ that vanishes at the transition
point, $a(T)=\tilde{a}(T-T_{c})$. For this calculation it is sufficient to
use a constant $\Psi (\mathbf{r})=\Psi $. Now let us transform the first
term of Eq.(\ref{F1}), using Eqs.(\ref{calK},\ref{Deltar},\ref{Dab},\re
{newnot}). We assume here that $\Delta _{j}$ obey the matrix equation (\re
{tilde-fin}), i.e. the local order parameter coincides with $\tilde{\Delta}
\mathbf{r})$ defined in Eq.(\ref{Deltar}). The result reads
\begin{equation}
\frac{F_{2}}{\Psi ^{2}}=\frac{g}{4}\sum_{ij}\Delta _{i}\Delta _{j}\eta
_{i}\eta _{j}M_{ij}-\frac{g^{2}}{8}\sum_{ijk}\Delta _{i}\Delta _{j}\eta
_{i}\eta _{j}\eta _{kl}M_{iikl}M_{kljj} \label{F21}
\end{equation
where we used the "fusion rule" following from the completeness of the set
of $\psi _{j}(\mathbf{r})$:
\begin{equation}
\sum_{i}R_{kli}R_{pqi}=M_{klpq}. \label{fus}
\end{equation
and an equivalent form of Eq.(\ref{tilde-fin}) written in terms of $D_{i}$
and $\Delta _{jk}$ variables:
\begin{equation}
D_{i}=\frac{g}{2}\sum_{kl}\eta _{kl}\Delta _{kl}\,R_{ikl}. \label{D-Dij}
\end{equation}
To determine the coefficient $a$ we neglect the off-diagonal terms $M_{iikl}$
with $k\neq l$ in Eq.(\ref{F21}) and thus reduce it to
\begin{equation}
F_{2}^{(0)}=\frac{1}{2}\Psi ^{2}\sum_{ij}\left[ \left( \hat{Q}\right)
_{ij}-\left( \hat{Q^{2}}\right) _{ij}\right] \phi _{i}\phi _{j} \label{F22}
\end{equation
where $\hat{Q}$ and $\phi _{i}$ are defined in Eq.(\ref{Qmat}). In the
continuum limit one writes $\xi _{i}\rightarrow \xi $, $M_{ij}\rightarrow
\frac{1}{\mathcal{V}}M(\xi -\zeta )$, $\sum_{i}\rightarrow \mathcal{V}\nu
_{0}\int d\xi $ and thus the bilinear form (\ref{F22}) transforms into
\begin{equation}
F_{2}^{(0)}=\frac{\nu _{0}\mathcal{V}}{2}\Psi ^{2}\int \int d\xi d\zeta
L(\xi ,\zeta )\phi (\xi )\phi (\zeta )\, \label{F23}
\end{equation
where
\begin{equation}
L(\xi ,\zeta )=Q(\xi ,\zeta )-\int d\xi _{1}Q(\xi ,\xi _{1})Q(\xi _{1},\zeta
)\,, \label{L2}
\end{equation
with the kernel
\begin{equation}
{Q}(\xi ,\zeta )=\frac{\lambda }{2}\sqrt{\eta (\xi )\eta (\zeta )}M(\xi
-\zeta ) \label{kernel}
\end{equation
The function $\phi (\xi )$ is, by construction, an eigenfunction of the
integral equation with the kernel $Q(\xi ,\zeta )$; its eigenvalue $k(T)$
approaches 1 at $T=T_{c}$. At this stage we need to specify the
normalization condition for the function $\phi (\xi )$:
\begin{equation}
\int d\xi \Delta ^{2}(\xi )\eta (\xi )\equiv \int d\xi \phi ^{2}(\xi )=\frac
2T_{c}^{2}}{\gamma } \label{norm1}
\end{equation
The form of the normalization condition Eq.(\ref{norm1}) was chosen to
enable a smooth crossover to the conventional (non-fractal) BCS case $\gamma
\rightarrow 0$, when $\phi _{\gamma =0}=T\sqrt{\eta (\xi )}$. To get the
equations valid in the crossover regime we need to introduce the high-energy
cutoff $\Omega _{D}$ so that all integrals over $d\xi $ go over the range
|\xi |<\Omega _{D}$. Then a straightforward modification of the
normalization condition
\begin{equation}
\int_{-\Omega _{D}}^{\Omega _{D}}d\xi \Delta ^{2}(\xi )\eta (\xi )=\frac
2T_{c}^{2}\ln \frac{\Omega _{D}}{T_{c}}}{\gamma \ln \frac{\Omega _{D}}{T_{c}
+1} \label{norm2}
\end{equation
(that reduces to (\ref{norm1}) when $\gamma \gg 1/\ln \frac{\Omega _{D}}
T_{c}}=\lambda $) allows one to recover the BCS limit at at a fixed $\Omega
_{D}$ and $\gamma \rightarrow 0$. We are mainly interested in the case of
intermediate to strong fractality, $\gamma \sim 0.6$, and weak interaction
\lambda \ll 1$, so we shall use the simplified condition (\ref{norm1})
appropriate in this limit.
Making use of Eq.(\ref{norm1}) we reduce the expression (\ref{F23}) to
\begin{equation}
F_{2}^{(0)}=\nu _{0}\mathcal{V}T^{2}\Psi ^{2}\gamma ^{-1}[k(T)-k^{2}(T)]
\label{F24}
\end{equation
Since $1-k(T)\ll 1$ at $T\approx T_{c}$, it is sufficient to evaluate the
derivative $dk/dT|_{k=1}$. We find it by rewriting the kernel (\ref{kernel})
in a dimensionless form:
\begin{equation*}
\bar{Q}(x,y)=\frac{\lambda _{T}}{2|x-y|^{\gamma }}\sqrt{\frac{\tanh (x)\tanh
(y)}{xy}}\,,
\end{equation*
where $\lambda _{T}=\lambda (E_{0}/2T)^{\gamma }$. Clearly, $d\ln k(T)/d\ln
(T)=-\gamma $, and, finally, the coefficient $a(T)$ is expressed only in
terms of the ratio of $T/T_{c}$:
\begin{equation}
a(T)=\frac{T-T_{c}}{T_{c}} \label{alphaT}
\end{equation
Note that the order parameter $\Psi $ introduced in Eq.(\ref{Psi}) is
dimensionless, whereas $\Delta $ has a dimension of energy.
\subsubsection{Quartic term: coefficient $b$}
Now we turn to the computation of the coefficient $b$ in the expansion (\re
{FGL}). Our starting points are Eqs.(\ref{F1}) and (\ref{J},\ref{zeta}). As
in the previous part we may use here $\Psi (\mathbf{r})=\mathrm{const}$,
thus $\Delta (\mathbf{r})$ is proportional to $\tilde{\Delta}(\mathbf{r})$
defined in Eq.(\ref{Deltar}). Therefore using Eqs.(\ref{Deltar}), (\re
{newnot}) we may substitute $D_{i}=\frac{g}{2}\sum_{a}\Delta _{a}\eta
_{a}R_{iaa}$ in the fourth order in $D_{i}$ term of Eq.(\ref{F1}) reducing
this term to the following form:
\begin{eqnarray}
F_{4} &=&\frac{1}{4}\left( \frac{g}{2}\right)
^{4}\sum_{ijkl}\sum_{abcd}\sum_{\nu _{1}..\nu _{4}}\Delta _{a}\Delta
_{b}\Delta _{c}\Delta _{d}\,\eta _{a}\eta _{b}\eta _{c}\eta _{d}\,
\label{F40} \\
&&\times R_{i\nu _{4}\nu _{1}}R_{j\nu _{1}\nu _{2}}R_{k\nu _{2}\nu
_{3}}R_{l\nu _{3}\nu _{4}}\,R_{iaa}R_{jbb}R_{kcc}R_{ldd}\,\zeta _{\nu
_{1}\nu _{2},\nu _{3}\nu _{4}} \notag
\end{eqnarray
Now we do summations over $i,j,k,l$ using the fusion rule (\ref{fus}), and
then neglect the off-diagonal matrix elements, i.e. we set everywhere
M_{\nu _{1}\nu _{2}aa}\rightarrow \delta _{\nu _{1}\nu _{2}}M_{\nu _{1}a}$.
Summations over $a,b,c,d$ can be done now with the use of modified MFA
equations (\ref{tilde-fin}). Finally, we proceed from the last remaining
summation over $\nu _{1}$ to integration, and find
\begin{equation}
F_{4}=\frac{\nu _{0}\mathcal{V}}{32T^{2}}\Psi ^{4}\int \frac{d\xi }{2T}\Phi
\left( \frac{\xi }{2T}\right) \Delta ^{4}(\xi ) \label{F4}
\end{equation
where
\begin{equation*}
\Phi (x)\equiv 16T^{3}\lim_{\forall \xi _{\alpha }\rightarrow 2Tx}\zeta
_{ij,kl}=\frac{1}{x^{2}}\left( \frac{\tanh (x)}{x}-\frac{1}{\cosh ^{2}(x)
\right)
\end{equation*
with the function $\zeta _{ijkl}$ defined in Eq.(\ref{zeta}). In the limit
\gamma \rightarrow 0$ we have $\Delta (\xi )=\mathrm{const}$ and the
integral in (\ref{F4}) reduces to
\begin{equation*}
\frac{1}{32}\int dx\Phi (x)=\frac{7\zeta (3)}{16\pi ^{2}}=\frac{1}{2}b_
\mathrm{BCS}}\,,
\end{equation*
leading to the standard BCS result for the $b$ coefficient in Eq.(\ref{FGL
). In general case we find
\begin{equation}
b=\frac{1}{4}\int dx\,\Phi (x)\,x^{2}\coth ^{2}(x)\,\bar{\phi}^{4}(x)
\label{beta}
\end{equation
where $\bar{\phi}(x)=T^{-1/2}\phi (2Tx)$ and $\phi (\xi )$ is the $k=1$
eigenfunction of the kernel (\ref{kernel}) subject to the normalization
condition (\ref{norm1}).
\subsubsection{Gradient term: coefficient $C$.\label{Gradient term}}
We now turn to calculation of the gradient term of the functional (\ref{FGL
). Note that previously employed "diagonal approximation" for the matrix
elements $M_{ijkl}$ is not sufficient for that purpose. In order to find the
coefficient $C$ we need to take into account the off-diagonal matrix
elements $M_{ijkl}$ which contain pairs of different levels with nearby
energies: $|\xi _{i}-\xi _{j}|\sim |\xi _{k}-\xi _{l}|\sim \omega (q)$, with
$\omega (q\rightarrow 0)\rightarrow 0$. The reason for that is similar to
the one which applies to the standard BCS Hamiltonian with the reduced
interaction term $g\,\sum_{\mathbf{p,p^{\prime }}}c_{\uparrow ,\mathbf{p
}\,c_{\downarrow ,\mathbf{-p}}\,c_{\downarrow ,\mathbf{p^{\prime }
}^{\dagger }\,c_{\uparrow ,\mathbf{-p^{\prime }}}^{\dagger }$. One has to
allow for the non-zero momentum of a pair and thus go beyond the "diagonal
approximation" in $\mathbf{p}$ and $\mathbf{p^{\prime }}$ in order to be
able to compute the \textit{phase rigidity} $C$ which is related with the
supercurrent.
We first illustrate this statement using the standard theory of disordered
superconductors as an example, see e.g.~\cite{DeGennesBook}. In that case,
quadratic term of the free energy expansion (\ref{F1}) can be represented in
the coordinate space:
\begin{equation}
F[\Delta (\mathbf{r})]=\int d\mathbf{r}d\mathbf{r^{\prime }}\left[ \frac{1}{
}\delta (\mathbf{r}-\mathbf{r^{\prime }})-K(\mathbf{r},\mathbf{r^{\prime }}
\right] \Delta (\mathbf{r})\Delta (\mathbf{r^{\prime }}) \label{FGLs}
\end{equation
where
\begin{equation}
K(\mathbf{r},\mathbf{r^{\prime }})=\nu _{0}T\sum_{\omega _{n}}\sum_{ij}\frac
\psi _{i}(\mathbf{r})\psi _{j}(\mathbf{r})\psi _{i}^{\ast }(\mathbf
r^{\prime }})\psi _{j}^{\ast }(\mathbf{r^{\prime }})}{(\xi _{i}-i\omega
_{n})(\xi _{j}+i\omega _{n})} \label{Kold}
\end{equation
Here $\omega _{n}=\pi T(2n+1)$ and (\ref{Kold}) is the equivalent form of
Eqs.(\ref{K},\ref{eta}). As the standard procedure goes, one averages the
kernel (\ref{Kold}) over disorder in the semiclassical approximation and
obtains the Fourier-transformed kernel
\begin{equation}
\overline{K}(q)=2\pi \nu _{0}T\sum_{\omega _{n}}\frac{1}{Dq^{2}+2|\omega
_{n}|} \label{Kq}
\end{equation
where $D$ is the diffusion coefficient. Finally one obtains for the
coefficient $C$ in Eq.(\ref{FGL}) the well known resul
\begin{equation}
C^{(0)}=-\frac{1}{\nu _{0}}\left. \frac{dK(q)}{dq^{2}}\right\vert
_{q\rightarrow 0}=\frac{\pi }{8}\frac{D}{T_{c}} \label{gamma0}
\end{equation
In order to generalize this derivation to the case of strong disorder in is
convenient to introduce two particle spectral function
\begin{eqnarray}
g(q,\Omega ) &=&\frac{1}{\mathcal{V}\nu _{0}}\sum_{ij}\int d\mathbf{r}
\mathbf{r^{\prime }}\,e^{i\mathbf{q}(\mathbf{r}-\mathbf{r^{\prime }
)}\,\delta (\xi _{i}-\xi _{j}-\Omega ) \notag \label{gq0} \\
&\times &\langle \psi _{i}(\mathbf{r})\psi _{j}(\mathbf{r})\psi _{i}^{\ast }
\mathbf{r^{\prime }})\psi _{j}^{\ast }(\mathbf{r^{\prime }})\rangle
\,\,\delta (E_{F}-\xi _{i}).
\end{eqnarray
The average kernel in Eq.(\ref{Kold}) can be expressed through this two
particle function:
\begin{equation}
\overline{K}(q)=\nu _{0}T\sum_{\omega _{n}}\int \frac{d\xi d\xi ^{\prime
}g(q,\xi -\xi ^{\prime };E)}{(\xi -i\omega _{n})(\xi ^{\prime }+i\omega _{n}
} \label{Kq2}
\end{equation
In the diffusive limit the spectral function is
\begin{equation}
g(q,\Omega )=\frac{1}{\pi }\frac{Dq^{2}}{\Omega ^{2}+(Dq^{2})^{2}}.
\label{gq}
\end{equation
and its substitution into Eq.(\ref{Kq2}) gives back Eq.(\ref{Kq}).
This derivation of the standard result demonstrates that the small $q
-dependence of the averaged kernel $\overline{K}(q)$ comes from matrix
elements of the operator $e^{iq(\mathbf{r}-\mathbf{r^{\prime }})}$ between
eigenstates $\psi _{i}(\mathbf{r}),\psi _{j}(\mathbf{r})$ with nearby
energies $|\xi _{i}-\xi _{j}|\sim Dq^{2}$. Notice that diagonal
approximation used in previous sections corresponds to $i=j$ in Eq.(\re
{Kold}). In this approximation one neglects the off-diagonal terms $M_{ijkl}$
completely and gets $g(q,\Omega )=\delta (\Omega )f(q)$, where
\begin{equation*}
f(q)=\int d\mathbf{r}d\mathbf{r^{\prime }}\,\langle \psi _{i}^{2}(\mathbf{r
)\psi _{i}^{2}(\mathbf{r^{\prime }})\,e^{i\mathbf{q}(\mathbf{r}-\mathbf
r^{\prime }})}\rangle
\end{equation*
For delocalized wavefunctions which occupy all the available volume one can
write
\begin{equation*}
\mathcal{V}^{2}\,\langle \psi _{i}^{2}(\mathbf{r})\psi _{i}^{2}(\mathbf
r^{\prime }})\rangle \approx \mathcal{V}^{2}\,\langle \psi _{i}^{2}(\mathbf{
})\rangle \,\langle \psi _{i}^{2}(\mathbf{r^{\prime }})\rangle =1+O(1/g),
\end{equation*
where $g\gg 1$ is the dimensionless conductance Then $f(q=0)=1+O(1/g)$ and
f(q\neq 0)=O(1/g)$. The jump in $f(q)$ at $q=0$ implies that in the diagonal
approximation the coefficient $C$ is infinite in a metal which is obviously
wrong. However, the value of $K$ at $q=0$ can be determined correctly in the
diagonal approximation because at $q=0$ the spectral function $g(q,\Omega
)\rightarrow \delta (\Omega )$ both in the diagonal and in the diffusive
approximation Eq.(\ref{gq}).
The correct derivation of the coefficient $C~$\ in case of a strong disorder
has to take into account a non-uniformity of the "background configuration"
of the superconducting order parameter given by Eq.(\ref{Deltar}). We follow
the same logics as in the derivation of the $a$ and $b$ coefficients in
previous sections. Note that the $q$-dependence comes from the second term
of the matrix (\ref{calK}) which leads to the second sum in Eq.(\ref{F21}).
We focus on this term and repeat the steps of the derivation which led to
Eq.(\ref{F21}) using now a weakly modulated $\Psi (\mathbf{r})=\Psi _{q}e^{
\mathbf{qr}}$ with small $q$. The second term of (\ref{F21}) becomes (in the
continuous form, after replacing $\sum_{j}\rightarrow \mathcal{V}\nu
_{0}d\xi $):
\begin{eqnarray}
\int d\mathbf{q}\,F_{2}^{(2)}(q) &=&-\int d\mathbf{q}\,|\Psi _{q}|^{2}\
\frac{g^{2}}{8}\int d\xi d\omega d\omega _{1}d\omega _{2}\Delta (\xi
_{+}+\omega _{1})\,\Delta (\xi _{-}+\omega _{2})\; \label{Fq1} \\
&&\times \eta (\xi _{+},\xi _{-})\,\eta (\xi _{+}+\omega _{1})\,\eta (\xi
_{-}+\omega _{2})\;M_{4}(q;\omega ,\omega _{1},\omega _{2}) \notag
\end{eqnarray
where $\xi _{\pm }=\xi \pm \omega /2$; the function $\Delta (\xi )=\phi (\xi
)/\sqrt{\eta (\xi )}$ obeys Eq.(\ref{MMFA}) and is normalized according to
Eq.(\ref{norm1}), and the function $M_{4}(q;\omega ,\omega _{1},\omega _{2})$
is defined by
\begin{equation*}
M_{4}(q;\omega ,\omega _{1},\omega _{2})=\mathcal{V}^{-1}\left\langle
\sum_{ijkl}\delta (\xi _{ij}-\omega )\,\delta (\xi _{ki}-\omega
_{1})\,\delta (\xi _{lj}-\omega _{2})\,\delta (E-\xi
_{i})\,I_{ijkl}(q)\right\rangle ,
\end{equation*
where $\xi _{ij}=\xi _{i}-\xi _{j}$
\begin{equation}
I_{ijkl}(q)=\int d\mathbf{r}d\mathbf{r^{\prime }}\,e^{i\mathbf{q(r-r^{\prime
})}}\,\psi _{i}(\mathbf{r})\psi _{j}(\mathbf{r})\psi _{i}(\mathbf{r^{\prime
})\psi _{j}(\mathbf{r^{\prime }})\psi _{k}^{2}(\mathbf{r})\psi _{l}^{2}
\mathbf{r^{\prime }}) \label{Iijkl}
\end{equation}
The function $M_{4}(q;\omega ,\omega _{1},\omega _{2})$ replaces function
g(q,\omega )$ employed in the standard derivation of the $|\nabla \Psi |^{2}$
term for diffusive superconductors. The additional terms $\psi _{k}^{2}
\mathbf{r})$ and $\psi _{l}^{2}(\mathbf{r^{\prime }})$ (and the summation
over the corresponding states $k$ and $l$) arise due to the \textit
essentially $\mathbf{r}-dependent$} "background configuration" which
\mathbf{r}$-dependence is expressed through $\psi ^{2}(\mathbf{r})$ by Eq.
\ref{Deltar}).
The main source of fluctuations of $M_{4}$ are the spatial fluctuation of
the eigenfunctions whereas the spectral fluctuations are relatively weak.
Neglecting spectral fluctuations, one can switch from summation over indices
$i,j,k,l$ to integration over the energy variables $\xi _{i}$, $\xi _{j}$,
\xi _{k}$, $\xi _{l}$ with the constant DoS and consider the average
I(q,\{\omega \})=\langle I_{ijkl}(q)\rangle $ as a function of all energy
differences $\omega $, $\omega _{1}$, $\omega _{2}$. Then the expression for
the coefficient $C$ in the Ginzburg-Landau functional Eq.(\ref{FGL}) takes
the form
\begin{eqnarray}
C &=&\frac{1}{8T_{c}^{2}}\lambda ^{2}\nu _{0}\int d\xi d\omega d\omega
_{1}d\omega _{2}\,\Delta (\xi +\omega /2+\omega _{1})\,\Delta (\xi -\omega
/2+\omega _{2})\, \label{C} \\
&&\times \eta (\xi +\omega /2,\xi -\omega /2)\,\eta (\xi +\omega /2+\omega
_{1})\,\eta (\xi -\omega /2+\omega _{2})\;J(\{\omega \}), \notag
\end{eqnarray}
where $J(\{ \omega\})= -\mathcal{V}^{3}\left.\frac{dI(q,\{\omega\})}{dq^{2}
\right|_{q=0}$.
The next step is to estimate $I(\{\omega \})$ using the rules formulated in
section \ref{Fractality and correlations}.\ We begin by considering the
critical wavefunction statistics. Applying the rule $\mathbf{(iii)}$ we find
the phase-independent counterpart to the combination of eigenfunctions in
Eq.(\ref{Iijkl}). It happens to coincide with the one for the square of the
off-diagonal matrix element $M_{ijkl}$ and is given by Eqs.(\re
{off-square-phase-ind}),(\ref{off1}). Then expanding Eq.(\ref{Iijkl}) up to
q^{2}$ and integrating over $(\mathbf{r}-\mathbf{r^{\prime }})$ up to $
\mathbf{r}-\mathbf{r^{\prime }}|\sim L_{\omega }$ we obtain an expression
similar to Eq.(\ref{off-square22}) but containing an extra factor $L_{\omega
}^{2}$ due to the expansion of $e^{i\mathbf{q(r-r^{\prime })}}$ up to $q^{2}
:
\begin{equation}
J(\{\omega \})\sim \mathcal{V}^{3}\,L_{\omega }^{2}\,\langle
|M_{ijkl}|^{2}\rangle _{\mathrm{typ}}\sim L_{\omega }^{d+2}\,\left( \frac
E_{0}}{\omega }\right) ^{2\gamma }. \label{deltaI}
\end{equation
In Eq.(\ref{deltaI}) we assumed all energy differences to be of the same
order $\omega _{1}\sim \omega _{2}\sim \omega $. The estimate Eq.(\re
{deltaI}) holds true on the insulator side of the Anderson transition as
long as $\omega \gg \delta _{L}$ (see the discussion in section \ref{Scaling
estimates C} and Eq.(\ref{delta-xi})).
Now we have to estimate the result of integration over the energy variables
in Eq.(\ref{C}). Using the asymptotic behavior of $\Delta (\omega )\propto
\omega ^{-\gamma }$ \ which follows from (\ref{MMFA}), (see section \re
{Modified mean-field}) and $\eta (\omega )\propto \omega ^{-1}$ and Eq.(\re
{deltaI}) power counting shows that for $T_{c}>\delta _{L}$ the integral is
dominated by $\omega \sim \omega _{1}\sim \omega _{2}\sim T_{c}$ and can be
estimated by
\begin{equation}
C\sim \left( \frac{1}{\nu _{0}T_{c}}\right) ^{\frac{2}{d}}\,\lambda
^{2}\left( \frac{E_{0}}{T_{c}}\right) ^{2\gamma }\sim \left( \frac{1}{\nu
_{0}T_{c}}\right) ^{\frac{2}{d}},\;\;\;\;(T_{c}\gg \delta _{L}). \label{CC}
\end{equation
Remarkably, the enhancement factor depending on the fractal exponent $\gamma
$ is canceled by the coupling constant $\lambda $ due to Eq.(\ref{Tc0}), and
the result is not sensitive to fractality and is essentially the same as in
Refs.\cite{BulaSad,KapitulnikKotliar1986}. This cancellation occurs due to the
presence of the additional (with respect to the standard expression Eq.~\re
{gq0}) factors $\psi _{k}^{2}(\mathbf{r})$ and $\psi _{l}^{2}(\mathbf{r})$
in the integral $I_{ijkl}(q)$ defined in Eq.(\ref{Iijkl}); without these
factors, the final result for $C$ would contain extra small factor $\sim
(T_{c}/E_{0})^{\gamma }$.
Note that the coefficient $C$ is dominated by the off-diagonal matrix
elements only in metal or in very weak insulator. As one moves towards a
strong insulator where $T_{c}<\delta _{L}$, the main contribution to $C$
becomes the one that originates from the diagonal approximation and can be
roughly estimated as
\begin{equation}
C\sim L_{\mathrm{loc}}^{2},\;\;\;\;(\delta _{L}>T_{c}) \label{C-ins}
\end{equation
in the diagonal approximation. The estimate (\ref{C-ins}) is based on the
simplest picture that wavefunctions localized at distances larger than
localization length $L_{\mathrm{loc}}$ do not overlap. In fact, as we
discuss in section \ref{Transition temperature}, this approximation is a bit
too crude as it misses an important logarithmic factor which increases $C$
values in the range $T_{c}\ll \delta _{L}$, see Eq.(\ref{Zeff}).
\subsubsection{Mesoscopic fluctuations: coefficient $W$. \label{Mesoscopic
fluctuations}}
Disorder always leads to spatial fluctuations of parameters which enter the
Ginzburg-Landau functional; the major effect is due to fluctuations of $a(T
\mathbf{r})$. Universal mesoscopic fluctuations (which provide a lower bound
for the strength of this effect) were studied in Ref.~\cite{BulaSad} for
usual disordered superconductors and more recently in~\cite{Skvortsov2005}
for 2D films with the strong Finkelstein effect. Here we follow the same
general approach but implement it in the Fock space instead of the
coordinate space.
We start from the Eq.(\ref{F21}) for the quadratic part of the free energy.
Previously we neglected off-diagonal matrix elements $M_{ijkk}$ entering the
second term of (\ref{F21}); now our goal is to estimate the strength of
mesoscopic fluctuations produced by these matrix elements. Thus we represent
the second term as
\begin{equation}
F_{22}=-\frac{\Psi ^{2}}{2}\sum_{ij}\left( (\hat{Q}^{2})_{ij}+\Gamma
_{ij}\right) \phi _{i}\phi _{j}\, \label{F2m}
\end{equation
where $\hat{Q}$ is defined in Eq.(\ref{Qmat}) and
\begin{equation}
\Gamma _{ij}=\frac{g^{2}}{4}\sum_{k\neq l}M_{iikl}M_{kljj}\,\eta _{kl}(\eta
_{i}\eta _{j})^{\frac{1}{2}} \label{Gamma}
\end{equation
The matrix $\Gamma _{ij}$ contains corrections from the (previously
neglected) off-diagonal matrix elements. Its average $\overline{\Gamma _{ij}}
$ contributes to the shift of the critical temperature which can be
estimated using scaling arguments developed in section \ref{Scaling
estimates A}. The result is that the relative shift due to the off-diagonal
matrix elements is $(\delta _{\mathrm{off}}T_{c})/T_{c}\sim 1$.
Here we are interested in the strength of \textit{mesoscopic fluctuations}
W $ defined in Eq.(\ref{W}) and thus calculate the free energy cumulant:
\begin{eqnarray}
(\nu _{0}T^{2})^{2}\,\,W &=&\frac{1}{\Psi ^{4}\mathcal{V}}\langle \langle
F_{22}F_{22}\rangle \rangle = \\
&&\frac{1}{4\mathcal{V}}\sum_{i_{1}i_{2}i_{3}i_{4}}\phi _{i_{1}}\phi
_{i_{2}}\phi _{i_{3}}\phi _{i_{4}}\,\langle \langle \Gamma
_{i_{1},i_{2}}\Gamma _{i_{3},i_{4}}\rangle \rangle \notag \label{F22cu}
\end{eqnarray
Using Eq.(\ref{Gamma}) and switching to a continuum representation, we
obtain the irreducible correlation function $\langle \langle \Gamma \Gamma
\rangle \rangle $
\begin{gather*}
\langle \langle \Gamma _{i_{1},i_{2}}\Gamma _{i_{3},i_{4}}\rangle \rangle =
\\
\left( \frac{g^{2}}{4}\right) ^{2}\left( \nu _{0}\mathcal{V}\right) ^{4
\sqrt{\eta _{i_{1}}\eta _{i_{2}}\eta _{i_{3}}\eta _{i_{4}}}\int ...\int d\xi
_{m}d\xi _{n}d\xi _{k}d\xi _{l}\,\eta _{mn}\eta _{kl}\int ...\int d^{d
\mathbf{r}_{1}d^{d}\mathbf{r}_{2}d^{d}\mathbf{r}_{3}d^{d}\mathbf{r}_{4} \\
\langle \langle \psi _{i_{1}}^{2}(\mathbf{r}_{1})\psi _{i_{2}}^{2}(\mathbf{r
_{2})\psi _{i_{3}}^{2}(\mathbf{r}_{3})\psi _{i_{4}}^{2}(\mathbf{r}_{4})\psi
_{m,n}^{(2)}(\mathbf{r}_{1})\psi _{m,n}^{(2)}(\mathbf{r}_{2})\psi
_{k,l}^{(2)}(\mathbf{r}_{3})\psi _{k,l}^{(2)}(\mathbf{r}_{4})\rangle \rangle
,
\end{gather*
where the double brackets $\langle \langle ...\rangle \rangle $ denote the
cumulant average defined by (we use below indices $i_{1}...$ instead of
\psi _{i_{1}}^{2}...$ etc.): $\langle \langle
i_{1}i_{2}i_{3}i_{4}mnkl\rangle \rangle =\langle
i_{1}i_{2}i_{3}i_{4}mnkl\rangle -\langle i_{1}i_{2}mn\rangle \langle
i_{3}i_{4}kl\rangle $ and $\psi _{m,n}^{(2)}(\mathbf{r})\equiv \psi _{m}
\mathbf{r})\psi _{n}(\mathbf{r})$.
We now estimate the coefficient $W$ for the critical states near the
Anderson transition point. First of all we note that the decoupled average
\langle i_{1}i_{2}mn\rangle \langle i_{3}i_{4}kl\rangle $ depends only on
the two of the three independent differences in coordinates $R_{s,s^{\prime
}}=|\mathbf{r}_{s}-\mathbf{r}_{s^{\prime }}|$ ($s,s^{\prime }=1,...4$),
while the cumulant average depends on all the three of them and vanishes
when $R_{s,s^{\prime }}>L_{\omega }$ (as before we assume that all the
energy differences are of the same order $\omega $). In the region where all
$R_{s,s^{\prime }}<L_{\omega }$ the decoupled average is smaller than the
cumulant one. Assuming all the differences of coordinates are of the same
order $R_{s,s^{\prime }}\sim R$ and applying the rules $\mathbf{(i)-(iii)}$
of section \ref{Wavefunction correlations} we obtain at $R<L_{\omega }$:
\begin{gather*}
\left\langle \left\langle \psi _{i_{1}}^{2}(\mathbf{r}_{1})\psi _{i_{2}}^{2}
\mathbf{r}_{2})\psi _{i_{3}}^{2}(\mathbf{r}_{3})\psi _{i_{4}}^{2}(\mathbf{r
_{4})\psi _{m,n}^{(2)}(\mathbf{r}_{1})\psi _{m,n}^{(2)}(\mathbf{r}_{2})\psi
_{k,l}^{(2)}(\mathbf{r}_{3})\psi _{k,l}^{(2)}(\mathbf{r}_{4})\right\rangle
\right\rangle \sim \\
\left\langle \psi _{i_{1}}^{2}(\mathbf{r}_{1})\psi _{i_{2}}^{2}(\mathbf{r
_{2})\psi _{i_{3}}^{2}(\mathbf{r}_{3})\psi _{i_{4}}^{2}(\mathbf{r}_{4})\psi
_{m,n}^{(2)}(\mathbf{r}_{1})\psi _{m,n}^{(2)}(\mathbf{r}_{2})\psi
_{k,l}^{(2)}(\mathbf{r}_{3})\psi _{k,l}^{(2)}(\mathbf{r}_{4})\right\rangle
\sim \\
\left\langle \psi _{i_{1}}^{2}(\mathbf{r}_{1})\psi _{i_{2}}^{2}(\mathbf{r
_{2})\psi _{i_{3}}^{2}(\mathbf{r}_{3})\psi _{i_{4}}^{2}(\mathbf{r}_{4})\psi
_{m}^{2}(\mathbf{r}_{1})\psi _{n}^{2}(\mathbf{r}_{2})\psi _{k}^{2}(\mathbf{r
_{3})\psi _{l}^{2}(\mathbf{r}_{4})\right\rangle \sim \\
\sim \mathcal{V}^{-8}\left( \frac{L_{\omega }}{R}\right) ^{3d-2\alpha
_{2}-\alpha _{4}}\,\left( \frac{L_{\omega }}{\ell }\right) ^{4(d-d_{2})},
\end{gather*
where $\alpha _{2}=3d_{4}-2d_{2}$, $\alpha _{4}=7d_{8}-6d_{4}$ . Indeed, at
R\sim L_{\omega }$ eigenfunctions in different space points are
statistically independent and the corresponding averages can be decoupled.
The result is $\mathcal{V}^{-8}(L_{\omega }/\ell )^{4(d-d_{2})}$ in
agreement with Eq.(\ref{corr-dif-E}). At $R\sim \ell $ all eigenfunctions
can be considered in one space point and the averaging then gives $\mathcal{
}^{-8}(L_{\omega }/\ell )^{7(d-d_{8})}$. The last estimate coincides with
Eq.(\ref{def-multi}) at small energy separations $L_{\omega }\sim L$ where
one can consider all eigenfunctions to be identical, and it corresponds to
all eigenfunctions independently averaged at large energy separations when
L_{\omega }\sim \ell $. The decoupled average estimated in the same way
using Eqs.(\ref{off-square-phase-ind}),(\ref{off1}) is of the order of
\mathcal{V}^{-8}\,(L_{\omega }/R)^{2d-2\alpha _{2}}\,(L_{\omega }/\ell
)^{4(d-d_{2})}$ and thus is smaller at $R\ll L_{\omega }$ than the cumulant
average by the factor $(L_{\omega }/R)^{d-\alpha _{4}}$.
Now we estimate the result of the four spatial integrations: over difference
of coordinates $R_{12}=|\mathbf{r}_{1}-\mathbf{r}_{2}|$, $R_{13}=|\mathbf{r
_{1}-\mathbf{r}_{3}|$ and $R_{23}=|\mathbf{r}_{3}-\mathbf{r}_{2}|$ and one
free integration which results in the factor $\mathcal{V}$. At this point it
is important that the dependence of the cumulant average on
R_{12},R_{13},R_{23}$ is symmetric and such that the power of $R$ is
typically smaller than $3d$, as $\alpha _{q}^{\mathrm{typ}}>0$ (see Eq.(\re
{pos-typ})). This means that the main contribution to the integrals over
R_{12},R_{13},R_{23}$ comes from the region $R_{12}\sim R_{13}\sim
R_{23}\sim L_{\omega }$. This is a crucial circumstance that eliminates the
dependence on higher fractal dimensions.
Thus we arrive at
\begin{gather*}
\int d\{\mathbf{r}\}\left\langle \left\langle \psi _{i_{1}}^{2}(\mathbf{r
_{1})\psi _{i_{2}}^{2}(\mathbf{r}_{2})\psi _{i_{3}}^{2}(\mathbf{r}_{3})\psi
_{i_{4}}^{2}(\mathbf{r}_{4})\psi _{m,n}^{(2)}(\mathbf{r}_{1})\psi
_{m,n}^{(2)}(\mathbf{r}_{2})\psi _{k,l}^{(2)}(\mathbf{r}_{3})\psi
_{k,l}^{(2)}(\mathbf{r}_{4})\right\rangle \right\rangle \\
\sim \mathcal{V}^{-7}\,L_{\omega }^{3d}\,\left( \frac{E_{0}}{\omega }\right)
^{4\gamma }.
\end{gather*
The remaining energy integration is estimated in the same way as in the
previous subsection. As before, the dominant contribution comes from the
energies within a strip of width $T_{c}$ near the Fermi energy. The final
result reads
\begin{equation}
\langle \langle F_{22}F_{22}\rangle \rangle \sim \lambda ^{4}\,\nu _{0
\mathcal{V}\,T_{c}^{3}\,\left( \frac{E_{0}}{T_{c}}\right) ^{4\gamma }\sim
\nu _{0}\mathcal{V}\,T_{c}^{3}. \label{fluct-fin}
\end{equation
Again, as in Eq.(\ref{CC}), the enhancement factor which depends on the
fractal exponent $\gamma $ cancels out exactly by the coupling constant
\lambda $, and the final result for the coefficient $W$ is
\begin{equation}
W\sim \frac{1}{\nu _{0}T_{c}},\;\;\;\;(\delta _{L}\ll T_{c}). \label{WW}
\end{equation
This result was obtained using the definition (\ref{W}) which makes sense if
the spatial scale $L_{a}$ of $a(\mathbf{r})$ fluctuations is small compared
to the scale of $\Psi (\mathbf{{r})}$ variation. We expect that the same
estimate (\ref{WW}) is valid if both length-scales are of the same order.
In the limit of strong insulator, one can repeat the above analysis to
arrive at
\begin{equation}
W\sim L_{\mathrm{loc}}^{3},\;\;\;\;\;(\delta _{L}>T_{c}). \label{W-ins}
\end{equation
This result is obtained in the diagonal approximation.
\subsubsection{Ginzburg parameters for thermal and mesoscopic fluctuations
\label{Ginzburg parameters}}
Now we use the results given by Eqs.(\ref{alphaT},\ref{beta},\ref{CC},\re
{WW}) to estimate the relative width of the fluctuation region near the
thermal transition into a fractal superconductor state. First we estimate
the Ginzburg parameter $\mathrm{Gi}$ which determines the reduced
temperature range $|1-T/T_{c}|<\mathrm{Gi}$ where thermal fluctuations are
strong in a $3D$ system \cite{LarkinVarlamovBook}:
\begin{equation}
\mathrm{Gi}\sim \frac{b^{2}}{C^{3}(\nu _{0}T_{c})^{2}}\sim 1,\;\;\;\;(\delta
_{L}\ll T_{c}) \label{Gi}
\end{equation
The relative width of \textquotedblleft smearing\textquotedblright\ of the
superconductive transition due to positional disorder is given by the
parameter $\mathrm{Gi_{d}}$ defined as follows (see e.g. Ref.\cit
{IoffeLarkin1981}):
\begin{equation}
\mathrm{Gi}_{d}\sim \frac{W^{2}}{C^{3}}\sim 1,\;\;\;\;(\delta _{L}\ll T_{c})
\label{Gid}
\end{equation}
The estimates (\ref{Gi}) and (\ref{Gid}) demonstrate that the modified
mean-field approximation developed in this section can be used (with
relative accuracy of the order of unity) in order to estimate $T_{c}$ of a
fractal superconductor. Based on this result we conclude that in the region
of extended and weakly localized single-particle states with $\delta _{L}\ll
T_{c}$ the thermal fluctuations of the order parameter phase and the
mesoscopic fluctuations of the local $T_{c}$ which were not taken into
account in the modified MFA can at most reduce the global $T_{c}$ by a
factor of the order of one compared to the modified MFA result, Eq.(\ref{Tc0
) but can hardly lead to a modification of the functional dependence of
T_{c}$ on the interaction constant $\lambda $. This is the most important
conclusion of this subsection.
Another conclusion concerns the \textit{role of the off-diagonal matrix
elements} in the region of extended and weakly localized single-particle
states. These matrix elements are completely neglected in modified MFA which
nevertheless gives accurate results. They are, however, necessary for the
correct account of the thermal fluctuations and the local $T_{c}$ mesoscopic
fluctuations. The off-diagonal matrix elements also determine the
electromagnetic response and are necessary for calculation of the critical
current, as both properties are related with the gradient term in the
Ginzburg-Landau functional Eq.(\ref{FGL}).
In the region $\delta _{L}\gg T_{c}$ the use of the simplified estimate
C\sim L_{\mathrm{loc}}^{2}$ would lead to the conclusion that the parameters
$\mathrm{Gi}$ and $\mathrm{Gi_{d}}$ are large:
\begin{equation*}
\mathrm{Gi}\sim \mathrm{Gi_{d}}\sim \frac{1}{(\nu _{0}T_{c}L_{\mathrm{loc
}^{3})^{2}}\sim \frac{1}{Z_{\mathrm{eff}}^{2}},\;\;\;\;(\delta _{L}>T_{c})
\end{equation*
where $Z_{\mathrm{eff}}\sim T_{c}/\delta _{L}$ is the effective coordination
number to be discussed later on in relation with the \textit{pseudo-spin
Hamiltonian}. In fact, more accurate estimate for $Z_{\mathrm{eff}}$ given
by Eq.(\ref{Zeff}) in section \ref{Transition temperature} below shows that
both $\mathrm{Gi}$ and $\mathrm{Gi_{d}}$ remain of the order of one in the
broad range of large $\delta _{L}/T_{c}$ ratios.
\subsection{Pseudo-spin Hamiltonian. \label{Pseudospin Hamiltonian}}
As we will see below, there is a sufficiently wide range of parameters on
the insulating side of the Anderson transition where the superconducting
transition temperature $T_{c}$ is of the order of its value $T_{c}^{(0)}$
right at the Anderson transition point while the paring gap $\Delta _{P}$
introduced in section \ref{Insulating state} is much larger than $T_{c}$.
This means that practically the entire region where $T_{c}$ gradually
decreases with increasing disorder falls into this \textit{pseudo-gap regime
. The modified MFA does not work in this regime because Ginzburg parameters
\mathrm{Gi}$ and $\mathrm{Gi_{d}}$ are larger than 1. However, the problem
can be significantly simplified by making use of the large value of the gap
\Delta _{P}$ between states with even and odd number of particles occupying
any orbital $\psi _{i}(\mathbf{r})$. Namely, as the creation of the \textit
odd state} with one particle on an orbital takes a large energy $2\Delta
_{P} $ to break the pair, the transitions between the even (having two or no
particles on any orbital) and odd states described by the \textit
off-diagonal} matrix elements can be neglected. Technically, it is
equivalent to neglecting all the off-diagonal matrix elements and
considering the sectors of Hilbert space with even and odd states as
completely decoupled. Restricting ourselves to the low-energy \textit{even}
sector one can rewrite \cite{Anderson1959} the Hamiltonian Eq.(\ref{Ham1})
in the form of a \textit{pseudo-spin Hamiltonian} with the spin operators
S_{i}^{\pm }$, $S_{i}^{z}$ acting in the Fock space of orbitals $|i\rangle
\equiv \psi _{i}(\mathbf{r})$ and the \textit{diagonal matrix elements}
M_{ij}$ playing a role of the coupling matrix. The most general form of this
Hamiltonian is:
\begin{eqnarray}
H_{PS} &=&\sum_{j}\xi _{j}(2S_{j}^{z}+1)-\frac{g_{\perp }}{2
\sum_{ij}M_{ij}(S_{i}^{+}S_{j}^{-}+S_{i}^{-}S_{j}^{+}) \notag \\
&-&g_{\parallel }\sum_{ij}M_{ij}S_{i}^{z}S_{j}^{z}. \label{HamSpin}
\end{eqnarray
where the set of operators
\begin{equation}
S_{j}^{+}=c_{j,\uparrow }^{+}c_{j,\downarrow }^{+}\,\quad
S_{j}^{-}=(S_{j}^{+})^{\dagger }\quad S^{z}=\frac{1}{2}\left( \sum_{\sigma
}c_{j,\sigma }^{+}c_{j,\sigma }-1\right)
\end{equation
is equivalent to the set of spin-$\frac{1}{2}$ operators $\mathbf{S}=\frac{
}{2}\mathbf{\sigma }$. Here $\xi $ is random energy distributed with density
$\nu $ in some interval around $0.$
The Hamiltonian Eq.(\ref{HamSpin}) is the basis of theory of the \textit
pseudogap superconductivity} we will develop in section \re
{Superconductivity with a pseudogap}. However, it is valid in \textit{any}
case where the off-diagonal matrix elements may be neglected for this or
another reason. It was originally suggested by Anderson \cite{Anderson1959}
for a BCS superconductor where the off-diagonal matrix elements are small in
the parameter $T_{c}/E_{F}$. It is parametrically justified in the
pseudo-gap region where the energy denominator associated with the even-odd
transitions is large since $\Delta _{P}\gg T_{c}$. However, the estimate of
the Ginzburg parameters $\mathrm{Gi}\sim \mathrm{Gi_{d}}\sim 1$ (see
discussion in section \ref{Transition temperature} and (\ref{Zeff})) shows
that it is also useful for a \textit{semi-quantitative} (up to a factor of
order one) determination of the transition temperature of the fractal
superconductor near the Anderson localization transition. We will use this
Hamiltonian in order to obtain the phase diagram as a function of $(T,E_{F})$
in the entire region spanning BCS and the pseudo-gap regime.
\subsection{Virial expansion method. \label{Virial expansion}}
In this subsection we develop a new approach based on the virial expansion
method applied to the pseudo-spin Hamiltonian Eq.(\ref{HamSpin}) an use it
to determine the superconductive transition temperature and thus the full
phase diagram of the disordered superconductor. The approximations implied
by this method are completely different from the ones of modified MFA. It
will turn out to be very useful in order to find the precise limit of
applicability of modified MFA and determine $T_{c}$ in the region of
localized single-particle states where the parameters $\mathrm{Gi}$ and
\mathrm{Gi_{d}}$ can be large.
Developing our scheme, we follow the approach of Larkin and Khmelnitsky~\cit
{Virial} who first used the virial expansion method to determine the
temperature of magnetic phase transition in metallic alloys (see also~\cit
{FeigTsvel1979}). The idea of this method is to express the free energy as a
series, where each term contains an \textit{exact} contribution from a fixed
number of local variables (e.g. spins for the problem of magnetic
impurities):
\begin{gather}
F=\sum_{n=1}^{\infty }\mathcal{F}^{(n)}=\sum_{i}F_{i}
\sum_{i>j}(F_{ij}-F_{i}-F_{j}) \label{virial1} \\
+\sum_{i>j>k}(F_{ijk}-F_{ij}-F_{jk}-F_{ik}+F_{i}+F_{j}+F_{k})+\dots \notag
\end{gather
Here $F_{i}=-T\ln \mathrm{Tr}e^{-H_{i}/T}$ is the free energy of a single $i
-th spin in a field, $F_{ij}=-T\ln \mathrm{Tr}e^{-(H_{i}+H_{j}+H_{ij})/T}$
is the exact free energy of two interacting spins, etc. The terms in the
brackets in the second sum and higher order sums cancel each other for large
space separations (e.g. $|\mathbf{r}_{i}-\mathbf{r}_{j}|$ for the second
order term) so that corresponding variables become essentially independent.
When the system is approaching a phase transition, all terms of the virial
expansion become relevant and an actual calculation of critical
singularities becomes impossible. However, the virial expansion method can
be used in order to find an approximate \textit{location} of the transition
point. Indeed, consider the virial expansion for some susceptibility
\begin{equation}
\chi (T)=\sum_{n=1}^{\infty }\chi _{n}(T) \label{vir-ser}
\end{equation
which must diverge at $T=T_{c}$. The value of $T_{c}$ can be determined from
the condition that $T_{c}$ corresponds to the \textit{limit of convergence }
of the series Eq.(\ref{vir-ser}):
\begin{equation}
\lim_{n\rightarrow \infty }\frac{\chi _{n+1}(T_{c})}{\chi _{n}(T_{c})}=1.
\label{virial2}
\end{equation
In practice, an exact \textit{analytical} calculation of $\chi _{n}$ with
large $n$ is very cumbersome, such calculations being usually limited by the
first few terms, $n=1,2,3...$. In the following we will use instead of Eq.
\ref{virial2}) the approximate truncated criterion
\begin{equation}
\chi _{2}(T_{c})=\chi _{3}(T_{c}). \label{virial3}
\end{equation
This step constitutes the key approximation of the virial method of
calculation of $T_{c}$. It consists in an extrapolation (in general,
uncontrolled) into the thermodynamic limit of properties found with exact
treatment of few-spin systems.
We now apply this idea to the specific problem of the calculation of
superconducting transition temperature corresponding to the Hamiltonian Eq.
\ref{HamSpin}). The relevant susceptibility is defined with respect to the
"ordering" field $\Delta $ which enters the Hamiltonian via the source term
\begin{equation}
V_{\Delta }=-\sum_{j}(\Delta S_{j}^{+}+\Delta ^{\ast }S_{j}^{-})
\label{Vdelta}
\end{equation
Transition into a superconducting state is signaled by the divergence of the
Cooper susceptibility
\begin{equation}
\chi (T)=-\frac{\partial ^{2}F}{\partial \Delta \partial \Delta ^{\ast }
=\sum_{n=1}^{\infty }\chi _{n}(T) \label{chi1}
\end{equation
Below we will describe calculation of the lowest-order virial expansion
terms $\chi _{n}(T)$ with $n=1,2,3$ defined in accordance with Eq.(\re
{virial1}) as:
\begin{eqnarray}
\chi _{1} &=&\sum_{i}\chi _{i}^{(1)} \label{chi2} \\
\chi _{2} &=&\sum_{n>m}(\chi _{nm}^{(2)}-\chi _{n}^{(1)}-\chi _{m}^{(1)})
\notag \\
\chi _{3} &=&\sum_{n>m>l}(\chi _{nml}^{(3)}-\chi _{nl}^{(2)}-\chi
_{ml}^{(2)}-\chi _{nm}^{(2)}+\chi _{n}^{(1)}+\chi _{m}^{(1)}+\chi _{l}^{(1)})
\notag
\end{eqnarray
where $\chi _{i}^{(1)},\chi _{ij}^{(2)},\chi _{ijl}^{(3)}...$ are the Cooper
susceptibilities of the system of 1, 2, 3 ... spins. These susceptibilities
can be expressed through the eigenvalues $\lambda _{\alpha }(\Delta )$ of
the $N$-spin Hamiltonian:
\begin{equation}
\chi ^{(N)}=-\frac{1}{Z_{0}}\,\sum_{\alpha }e^{-\lambda _{\alpha
}^{(0)}/T}\,\gamma _{\alpha }, \label{chi-N}
\end{equation
where $Z_{0}=\sum_{\alpha }e^{-\lambda _{\alpha }^{(0)}/T}$, and $\gamma
_{\alpha }$ is the sensitivity of the $\alpha $-th eigenvalue of the N-spin
Hamiltonian to the $\Delta $ perturbation:
\begin{equation}
\lambda _{\alpha }(\Delta )=\lambda _{\alpha }^{(0)}+|\Delta |^{2}\,\gamma
_{\alpha }+o(|\Delta |^{2}). \label{eigen}
\end{equation
In order to compute the $N$-spin susceptibility $\chi _{nml...N}^{(N)}$ one
has to represent the $N$-spin Hamiltonian as a $2^{N}\times 2^{N}$ matrix.
The single-spin susceptibility can be found easily:
\begin{equation}
\chi _{1}(T)=\sum_{i}\frac{1}{2\xi _{i}}\tanh \frac{\xi _{i}}{T}=\nu _{0}\ln
\left( \frac{4e^{\mathrm{\mathbf{C}}}E_{b}}{\pi T}\right) \label{1spin}
\end{equation
where $E_{b}$ the the upper energy cutoff and $\mathrm{\mathbf{C}}=0.577...$
is the Euler constant.
The two-spin Hamiltonian is given by the 4 by 4 matrix
\begin{eqnarray}
H^{(2)} &=&\left(
\begin{array}{cccc}
\xi _{+}-\frac{1}{2}J_{12}^{\parallel } & \Delta & \Delta & 0 \\
\Delta ^{\ast } & -\xi _{-}+\frac{1}{2}J_{12}^{\parallel } & -(J_{12}^{\perp
})^{\ast } & \Delta \\
\Delta ^{\ast } & -J_{12}^{\perp } & \xi _{-}+\frac{1}{2}J_{12}^{\parallel }
& \Delta \\
0 & \Delta ^{\ast } & \Delta ^{\ast } & -\xi _{+}-\frac{1}{2
J_{12}^{\parallel
\end{array
\right) \label{H-2} \\
&\equiv &H_{0}^{(2)}+V, \notag
\end{eqnarray
where $\xi _{\pm }=\xi _{1}\pm \xi _{2}$, $J_{ij}^{\parallel }=\frac
g_{\parallel }}{2}\,M_{ij}$, $J_{ij}^{\perp }=\frac{g_{\perp }}{2}\,M_{ij}$
; and the perturbation term $V\propto \Delta $. Using the Hamiltonian (\re
{H-2}) we calculate $\lambda _{\alpha }^{(0)}$ and $\gamma _{\alpha }$ and
then use Eq.(\ref{chi-N}) to obtain second virial term $\chi _{2}(T)$. We
give here its simplest form corresponding to $J_{12}^{\parallel }=0$, the
full answer being given in the Appendix A:
{\small
\begin{eqnarray}
\chi _{2} &=&\sum_{i>j}\left[ \frac{J_{ij}^{\perp }}{2E_{+}E_{-}}\,\tanh
\left( \frac{E_{+}}{T}\right) \,\tanh \left( \frac{E_{-}}{T}\right) +\frac{
}{2E_{+}}\,\tanh \left( \frac{E_{+}}{T}\right) \right. \label{2-vir} \\
&&\left. +\frac{1}{2E_{-}}\,\tanh \left( \frac{E_{-}}{T}\right) -\frac{1}
2\xi _{i}}\,\tanh \left( \frac{\xi _{i}}{T}\right) -\frac{1}{2\xi _{j}
\,\tanh \left( \frac{\xi _{j}}{T}\right) \right] \notag
\end{eqnarray
} where
\begin{equation}
E_{\pm }=\frac{1}{2}(\xi _{i}+\xi _{j})\pm \frac{1}{2}\sqrt{(\xi _{i}-\xi
_{j})^{2}+|J_{ij}^{\perp }|^{2}}. \label{E+-}
\end{equation
are the exact energies of the two-spin problem.
The three-spin susceptibility $\chi _{3}$ requires the solution of the cubic
equations which is too long to be written here. The corresponding derivation
is given in Appendix \ref{Virial expansion in pseudospin subspace} in a form
suitable for numerical calculations.
For the standard BCS problem with $M_{ij}=\frac{1}{\mathcal{V}}\rightarrow 0
, the renormalized energies $E_{\pm }$ coincide with the bare ones $\xi _{i}$
and $\xi _{j}$ , and the first term in Eq.(\ref{2-vir}) is the leading one.
In this case the summations over the two energy variables become independent
and one gets $\chi _{2}(T)=g^{2}\nu _{0}\ln ^{2}\left( \frac{4e^{\mathrm
\mathbf{C}}}E_{b}}{\pi T}\right) $. Applying the simplest truncated
criterion $\chi ^{(1)}(T_{c})=\chi ^{(2)}(T_{c})$, one finds the correct
result: $T_{c}=(4e^{\mathrm{\mathbf{C}}}/\pi )E_{b}e^{-1/\lambda }$. In the
same way one can show that \textit{any} truncated criterion $\chi
_{n}(T_{c})=\chi _{n+1}(T_{c})$ gives the same correct result for $T_{c}$ of
the BCS problem.
This is no longer true for the case where $M_{ij}$ has essential energy
dependence, as it has in the fractal case where $M_{ij}$ is given by Eq.(\re
{ij}). In this situation the double sum in $\chi _{2}$ and the corresponding
multiple sums in the higher-order susceptibilities $\chi _{n}$ \ do not
factorize. Moreover, all $\chi _{n}$ with $n\geq 2$ are dominated by the
low-energy region $\xi \sim T$ and do not contain any logarithmic
divergence, in contrast to $\chi _{1}$. This is why the simplest truncated
criterion $\chi _{1}(T)=\chi _{2}(T)$ does not make much sense: the
corresponding solution depends on the upper energy cutoff $E_{b}$, while for
any other truncated criterion $\chi _{n}(T)=\chi _{n+1}(T)$ with $n\geq 2$
the solution is independent of $E_{b}$. It is for this reason that we used
Eq.(\ref{virial3}) to find an approximate transition temperature. Note,
however, that the temperature $T_{1-2}$ found numerically for the 3D
Anderson model with critical disorder from the simplest truncated criterion
\chi _{1}(T_{1-2})=\chi _{2}(T_{1-2})$ gives values of $T_{1-2}$ which are
close to $T_{2-3}$ found from Eq.(\ref{virial3}) for a meaningful values of
the upper energy cutoff $E_{b}$.
So far we completely neglected the presence of the odd sector of the Hilbert
space. This is justified in the region of the pseudo-gap superconductivity
where the paring gap $\Delta _{P}\gg T_{c}$. However, even when being
decoupled from the even sector, the odd sector cannot be ignored completely
in a general case of $\Delta _{P}\sim T_{c}$. The reason for that can be
easily seen from Eq.(\ref{chi-N}) which is valid in a general case too.
Decoupling of the two sectors manifests itself in vanishing of all "partial"
susceptibilities $\gamma _{\alpha }$ corresponding to the states $\alpha $
of the odd sector. Thus only the eigenvalues $\lambda _{\alpha }^{(0)}$
corresponding to the even sector appear in the sum in Eq.(\ref{chi-N}). This
is not true for the partition function $Z_{0}$, where \textit{all} states
contribute. For $\Delta _{P}\gg T$ one can neglect $e^{-\lambda _{\mathrm{od
}}/T}$ compared to $e^{-\lambda _{\mathrm{even}}/T}$ and arrive at the
result given in Appendix \ref{Virial expansion in pseudospin subspace}.
However, in the BCS and the fractal region where $\Delta _{P}<T_{c}$ the
susceptibility will be substantially decreased by "statistical dilution" of
the even states by the odd ones.
From the standard theory of the BCS superconductivity it is well known that
in all expressions the temperature dependence appears as the factor $\tanh
\left( \frac{\xi }{2T}\right) $. In contrast to it, the Eqs.(\ref{1spin},
\ref{2-vir}) contain factor $\tanh \left( \frac{\xi }{T}\right) $. This
gives a simple rule to adjust the virial expansion scheme to the case of
vanishing $\Delta _{P}$:
\begin{equation}
\tanh \frac{E_{i}}{T}\rightarrow \tanh \frac{E_{i}}{2T} \label{factor2}
\end{equation
Thus the temperature $T_{2-3}$ found from the formalism described in
Appendix \ref{Virial expansion in pseudospin subspace} should be simply
divided by 2 to find the corresponding approximation for the transition
temperature in the region of extended single-particle states. In Appendix
\ref{Virial expansion including single-occupied states} we present the
derivation of the virial susceptibilities for the general case of an
arbitrary paring gap $\Delta _{P}$ (still, neglecting off-diagonal matrix
elements, which mix the "even" and "odd" subspaces of the Hilbert space). In
particular, we show there that as long as the renormalization of energy
levels (i.e. the difference between $E_{+}$ and $\xi _{i}$ in Eq.(\ref{E+-
)) can be neglected, one can find (see Eq.(\ref{1-full})) the whole family
of distribution functions $F(\varepsilon ,\Delta _{P})$ that interpolate
between the two limiting cases, $\Delta _{P}\gg T_{c}$ and $\Delta _{P}\ll
T_{c}$.
In the region of extended and weakly localized single-particle states the
renormalization of energy levels is indeed negligible which justifies the
approach made in the derivation of (\ref{1-full}) in this regime. This
renormalization is not small in the region of a strong pseudo-gap but in
this regime one can project the pseudo-spin Hamiltonian Eq.(\ref{HamSpin})
into a truncated Hilbert space that consists of the \textit{even subspace}
only. Thus, both these limiting cases should be well described by a simple
replacement rule:
\begin{equation}
\tanh \left( \frac{\varepsilon }{T}\right) \rightarrow F(\varepsilon ,\Delta
_{P})=\frac{\sinh \left( \frac{\varepsilon }{T}\right) }{\cosh \left( \frac
\varepsilon }{T}\right) +e^{-\Delta _{P}/T}}, \label{interPolate}
\end{equation
where $\Delta _{P}=\langle \Delta _{P}\rangle $. Application of this
replacement rule to all the formulae obtained in the pseudo-spin model is
expected to take correctly into account the effect of "dilution" by the odd
sector. With this modification the pseudo-spin Hamiltonian Eq.(\ref{HamSpin
) becomes a unified tool to describe (semi-quantitatively or better) the
entire region spanning weakly disordered BCS superconductor to the Anderson
insulator.
\section{Superconducting state very close to mobility edge. \labe
{Superconducting state}}
In this section we derive the main physical properties of the three
dimensional disordered superconductor with the Cooper attraction
characterized by the single-electron states that are critical or very weakly
localized. As explained above, a very important property of these wave
functions is their fractality which is responsible for the unusual
properties of these superconductors. In the regime considered here the
number of paired electrons in localization volume is still large, which
allows us to neglect the level spacing $\delta _{L}$ compared to other
relevant energy scales:\
\begin{equation*}
\delta _{L}=E_{0}\left( \frac{E_{c}-E_{F}}{E_{0}}\right) ^{3\nu }\ll T_{c}.
\end{equation*}
The most important quantity that characterizes a superconductor is its
transition temperature and the single particle gap. To compute them we
analyze the modified MFA developed in subsection \ref{Modified mean-field}
and determine the critical temperature $T_{c}$ and the gap function $\Delta
(\xi )$. Then in subsection \ref{Comparison of Tc values} we compare $T_{c}$
obtained by the modified MFA with the results of two other methods based on
the "diagonal" approximation": the spectral analysis of the discrete
modified MF equation with the {\it fluctuating} kernel $\hat{Q}$, and the
virial expansion method. We show that the results for the transition
temperature $T_{c}$ obtained by different methods are in a good mutual
agreement but exhibit a strongly enhancement of $T_{c}$ compared to the
expectations based on Anderson theorem. Finally, in subsections \ref{Pairing
in the modified mean-field}-\ref{Superfluid density and critical current} we
use Ginzburg-Landau functional derived in subsection \ref{Ginzburg-Landau
functional} in order to estimate the fluctuation corrections to these
results, to calculate the distribution function of the local order parameter
near $T_{c}$, the local density of states at low temperatures and the
superfluid density.
\subsection{Pairing in the modified mean-field approximation. \label{Pairing
in the modified mean-field}}
The linearized modified mean-field equation for fractal superconductor was
derived in subsection \ref{Modified mean-field}, see Eq.(\ref{MMFA}). It can
be generalized for an arbitrary $T<T_{c}$ where a non-zero gap function
\Delta (\xi )$ develops:
\begin{equation}
\Delta (\xi )=\frac{\lambda }{2}\int d\zeta \frac{M(\xi -\zeta )\Delta
(\zeta )}{\sqrt{\zeta ^{2}+\Delta ^{2}(\zeta )}}\tanh \frac{\sqrt{\zeta
^{2}+\Delta ^{2}(\zeta )}}{2T} \label{MMFA2}
\end{equation
A straightforward calculation shows that near the transition temperature Eq.
\ref{MMFA2}) leads to the same result as the Ginzburg-Landau functional with
coefficients found in section \ref{Ginzburg-Landau functional} above. The
properties of the superconducting phase in the vicinity of the transition
temperature can be found from the solution of the linearized integral
equation, equivalent to (\ref{MMFA}):
\begin{equation}
\Delta (\xi )=\frac{\lambda }{2}\int d\zeta \frac{M(\xi -\zeta )\Delta
(\zeta )}{\zeta }\tanh \frac{\zeta }{2T} \label{MMFA_linear}
\end{equation}
Eqs.(\ref{MMFA2},\ref{MMFA_linear}) with the power-law kernel $M(\omega )$
given by Eq.(\ref{ij}) can be solved numerically. Its solution depends on
the fractal exponent $\gamma $ which controls the power of the kernel. In
particular, the critical temperature is found to be
\begin{equation}
T_{c}^{0}(\lambda ,\gamma )=E_{0}\lambda ^{1/\gamma }C(\gamma ) \label{Tc11}
\end{equation
where the function $C(\gamma )$ is plotted in Fig.\ref{Cgamma}. As was
already mentioned above, the power-law dependence on the interaction
constant $\lambda $, instead of the exponential dependence $e^{-\frac{1}
\lambda }}$ in the standard BCS theory, implies a dramatic enhancement of
T_{c}$ by disorder if interaction constant $\lambda $ is small. This result
could be important for observation of superconductivity in a system of cold
fermions with weak attraction trapped in a disordered lattice.
\begin{figure}[tbp]
\includegraphics[width=8cm]{Cgamma}
\caption{Numerical prefactor $C(\protect\gamma )$ appearing in the equation
for the transition temperature (\protect\ref{Tc11}) obtained by numerical
solution of linearized integral equation (\protect\ref{MMFA_linear}).}
\label{Cgamma}
\end{figure}
The solution for $\Delta (\xi )$ at $T=T_{c}$ is shown in Fig.\ref{DeltaTc}a
for $\gamma =0.57$ corresponding to the 3D Anderson model.
\begin{figure}[tbp]
\includegraphics[width=6cm]{Delta_at_Tc.eps}
\includegraphics[width=6cm]{Delta_Asymptotics_at_Tc}
\caption{(a) Functional dependence of the gap function $\Delta (\protect\xi )
$ for $T=T_{c}$ at $\protect\gamma =0.57$ that corresponds to the largest
eigenvalue of the linearized integral gap equation (\protect\ref{MMFA_linear
). (b) Asymptotic behaviour of $\Delta (\protect\xi )$ at large $\protect\xi
/T$.}
\label{DeltaTc}
\end{figure}
At large $\xi \gg T_{c}$ this function decays as $\Delta (\xi )\propto \xi
^{-\gamma }$, as demonstrated in Fig.\ref{DeltaTc}b.
The maximum value of the function $\Delta (\xi )$ is attained at $\xi=0$
for all temperatures. In the $T\rightarrow 0$ limit we find
\begin{equation}
\Delta _{0}(0)=E_{0}\lambda ^{1/\gamma }D(\gamma ) \label{Delta00}
\end{equation
where the function $D(\gamma )$ is plotted in Fig.\ref{DeltaGamma}a. The
ratio $2D(\gamma )/C(\gamma )$, which characterizes the ratio of
low-temperature spectral gap to the transition temperature, is shown as
function of $\gamma $ in Fig.\ref{DeltaGamma}b. Coincidentally, this ratio for the
3D Anderson model with the fractal exponent $\gamma =0.57$ turns out to be
rather close to the BCS value $3.5$. This implies that unfortunately it is
difficult to distinguish the fractal 3D superconductors from conventional
BCS ones by measuring $2\Delta /T_{c}$ value.
Evident strange feature of $D(\gamma )$ and $2D(\gamma )/C(\gamma )$
behavior is that they do not seem to approach the BCS limit at small $\gamma
$. The reason is the same as was discussed \ in section \ref{Transition
temperature: coefficient a(T)} around Eq. (\ref{norm2}): in order to get the
correct crossover to the non-fractal limit $\gamma \rightarrow 0$ , one
needs to introduce the upper energy cutoff $\Omega _{D}$. Indeed, the BCS
limit is reached at $\gamma \ll \ln ^{-1}\frac{\Omega _{D}}{T_{c}}\ll 1$
which is never satisfied if the upper energy cutoff is infinite as in the
calculations here.
\begin{figure}[tbp]
\includegraphics[width=6cm]{Delta_at_T=0}
\includegraphics[width=6cm]{Delta_over_Tc}
\caption{(a) Prefactor $D(\protect\gamma )$ in the equation (\protect\re
{Delta00}) for the maximum value $\Delta (0)$ of the $T=0$ obtained from the
numerical solution of the non-linear gap equation (\protect\ref{MMFA2}).
(b) Ratio $2\Delta (0)/T_{c}$ that follows from (\ref{Delta00}) with
$D(\protect\gamma )$ shown here and (\ref{Tc11}) with $C(\protect\gamma )$
shown in Fig.\ref{Cgamma} }
\label{DeltaGamma}
\end{figure}
\subsection{Comparison of $T_{c}$ values obtained in three different
approximations. \label{Comparison of Tc values}}
General modified MFA equation (\ref{Tc11}), with the parameter $C(\gamma
)\approx 3.1$ evaluated for $\gamma =0.57$, and the estimate (\ref{E00}) for
$E_{0}$, gives
\begin{equation}
T_{c}^{0}=(6.5\pm 0.8)\lambda ^{1.77} \label{Tc12}
\end{equation
As we explained above, there are two different types of corrections to this
formula: the corrections caused by the off-diagonal matrix elements in the
original Hamiltonian, and the corrections due to approximations made within
the \textquotedblleft diagonal approximation\textquotedblright , in
particular, the continuum approximation that neglects the fluctuation of
spectrum and the matrix elements which are taken to be a power-law function
of the energy difference. An estimate of the contribution from off-diagonal
matrix elements is given in section \ref{Ginzburg parameters}. In order to
clarify the role of the static fluctuations of the spectrum, we determine
the transition temperature for the same 3D Anderson model using two other
methods. We first determine it from the condition $\lambda _{\max }=1$ (see
section \ref{Modified mean-field}) where $\lambda _{\max }$ is the largest
eigenvalue of the matrix kernel $\hat{Q}$ \ defined by Eq.(\ref{Qmat}) that
contains matrix elements $M_{ij}$ corresponding to the\ specific realization
of disorder\ in finite size samples. This method adapts modified MFA to the
case of strong spatial fluctuations. Second, we apply the virial expansion
method as described in section \ref{Virial expansion}. Because here we
consider the regime $\Delta _{P}\ll T_{c}$, we use the replacement rule Eq.
\ref{factor2}) to take account of the dilution of pseudospin states. In both
methods we observe strong finite-size effects as shown in Fig.\ref{KvsL},\re
{K1}.
\begin{figure}[tbp]
\includegraphics[width=8cm]{KvsL10xx}
\caption{Transition temperatures determined via $\protect\rho (k)$ spectrum,
for a number of coupling constants and system sizes between $L=10,12,14$ and
$19$.}
\label{KvsL}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{Virial-K1xx}
\caption{Transition temperatures determined via the virial expansion, for a
number of coupling constants and system sizes between $L=10$ and $L=22$.}
\label{K1}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{Virial-K20xx}
\caption{(Color online) The results for transition temperature as function
of dimensionless coupling: the $\hat{Q}$-kernel analysis (blue dots) and the
virial expansion (black squares), extrapolated to $L=\infty $; The
Fermi-energy is fixed at the mobility edge.}
\label{VirialTc}
\end{figure}
The extrapolation to infinite system sizes gives the values of $T_{c}$ as
function of the dimensionless coupling constant $\lambda $ shown in Fig.\re
{VirialTc}. As one can see, all these methods give the results close to each
other and to the result Eq.(\ref{Tc12}) of the modified MFA at $\gamma =0.57
. The conclusion that we draw from this apparent coincidence is that all
approximations involved in these methods are reasonable. Specifically, it
seems that one can neglect the fluctuations of the single-particle DoS and
the matrix elements $M_{ij}$ at the mobility edge; one can also use the
small-$n$ truncated criterion for $T_{c}$ in the virial expansion method.
The main correction to the results of the \textit{analytical} modified MFA
given by Eqs.(\ref{MMFA2}) comes from the off-diagonal matrix elements.
However, even this correction is able to reduce the transition temperature
at most by a factor of the order of unity leaving all the results of
modified MFA \textit{semi-quantitatively correct}.
\subsection{Pairing amplitude in the real space\label{Pairing amplitude}}
The real-space pairing amplitude corresponding to the solution $\Delta _{i}$
of Eq.(\ref{tilde-fin}) can be determined from Eq.(\ref{Deltar}) for
T\approx T_{c}$. To demonstrate that $\tilde{\Delta}(\mathbf{r})$
corresponds to strongly spatially inhomogeneous solution, we calculate the
averaged square of the pairing amplitude
\begin{equation}
\overline{(\tilde{\Delta}(\mathbf{r}))^{2}}\equiv \frac{1}{\mathcal{V}}\int
d^{d}\mathbf{r}\tilde{\Delta}^{2}(\mathbf{r})=\lambda \int_{0}^{\infty }d\xi
\eta (\xi )\Delta _{c}^{2}(\xi ) \label{D2}
\end{equation
where we used the definition (\ref{Melements2}) and Eqs.(\ref{tilde-fin},\re
{MMFA}) to derive the r.h.s. of the above equation. Then we calculate its
simple average
\begin{equation}
\overline{\tilde{\Delta}(\mathbf{r})}\equiv \frac{1}{\mathcal{V}}\int d^{d
\mathbf{r}\tilde{\Delta}(\mathbf{r})=\lambda \int_{0}^{\infty }d\xi \eta
(\xi )\Delta _{c}(\xi ) \label{D1}
\end{equation
>From Eqs.(\ref{D1},\ref{D2}) we conclude that
\begin{equation}
f=\frac{\left( \overline{\tilde{\Delta}(\mathbf{r})}\right) ^{2}}{\overline{
\tilde{\Delta}(\mathbf{r}))^{2}}}=\lambda \mathcal{Q}(\gamma )=\frac
\mathcal{Q}(\gamma )}{C^{\gamma }(\gamma )}\left( \frac{T_{c}}{E_{0}}\right)
^{\gamma }\ll 1 \label{D12}
\end{equation
where the dimensionless function
\begin{equation}
\mathcal{Q}(\gamma )=\gamma \left( \int_{0}^{\infty }\frac{d\xi }{\xi }\tanh
\frac{\xi }{2T_{c}}\frac{\Delta _{c}(\xi )}{T_{c}}\right) ^{2}\,;\quad
\mathcal{Q}(0.57)=3.1 \label{Q}
\end{equation
The small ratio (\ref{D12}) gives the estimate of the space fraction where
pairing correlations are well established. Indeed, consider a toy model
where $\Delta (\mathbf{r})=\Delta _{0}$ in a fraction $f$ of the whole
space, and zero otherwise. Then we find $\overline{\tilde{\Delta}(\mathbf{r}
}=f\Delta _{0}$ and $\overline{(\tilde{\Delta}(\mathbf{r}))^{2}}=f\Delta
_{0}^{2}$, leading to the ratio (\ref{D12}) equal to $f$. \ In a realistic
case $\Delta (\mathbf{r})$ take a continuum of values, so one can use this
ratio as a proper \textit{definition} of the fraction of space with well
developed superconductive gap.
Note that the regions in space where the superconductive gap is appreciable
constitute a \textit{finite} fraction of the entire space, despite the fact
that fractal support of any \textit{single-particle} wavefunction occupies
\textit{vanishing} fraction of the entire space. Therefore the global
spatial pattern of superconductivity is not a fractal but rather is
reminiscent of that of the \textit{multi-fractal metal} \cite{CueKra}.
The above estimate was done for a region near $T_{c}$. At temperatures much
below $T_{c}$, (\ref{Deltar}) should be replaced by the similar equation in
which the argument $\xi _{k}$ is replaced by $\sqrt{\xi _{k}^{2}+\Delta
_{k}^{2}}$. As one might expect, one ends up with the same estimates in
which $T_{c}$ gets replaced by $\Delta _{0}(0)$.
Higher moments of the distribution function $P[\tilde{\Delta}(\mathbf{r})]$
can be estimated using the algebra of multi-fractal states discussed in
section \ref{Fractality and correlations}, see Eq.(\ref{multi-ij}).
Straightforward generalization of Eq.(\ref{D12}) gives
\begin{equation}
\frac{\left( \overline{\tilde{\Delta}(\mathbf{r})}\right) ^{n}}{\overline{
\tilde{\Delta}(\mathbf{r}))^{n}}}\propto \left( T_{c}/E_{0}\right)
^{(1-d_{n}/d)(n-1)} \label{D13}
\end{equation
Therefore the moments of $\tilde{\Delta}(\mathbf{r})$ contain information
about all fractal dimensions $d_{n}$. This is because the fractality of
single-particle states manifests itself \textit{locally} at a scale smaller
than $L_{T}\sim (\nu _{0}T_{c})^{-1/d}$ , above this length the order
parameter and other properties of the superconductor become homogeneous,
similar to what happens in a multi-fractal metal \cite{CueKra} which has a
sparse fractal structure within the correlation length. Thus, this state
should be properly named\textit{\ fractal superconductor}.
\subsection{Low-temperature density of states. \label{Low-temperature
density of states}}
Local density of states (DoS)\ in the superconducting state is given by
\begin{equation}
\nu (\varepsilon ,\mathbf{r})=\frac{1}{2}\sum_{j}\left( 1+\frac{\xi _{j}}
\varepsilon }\right) \left[ \delta (\varepsilon -\varepsilon _{j})+\delta
(\varepsilon +\varepsilon _{j})\right] \psi _{j}^{2}(\mathbf{r})
\label{nu00}
\end{equation
where $\varepsilon _{j}=\sqrt{\xi _{j}^{2}+\Delta ^{2}(\xi _{j})}$ and
\Delta (\xi )$ is the solution of the gap equation (\ref{MMFA2}). We begin
with the calculation of the average DoS:
\begin{equation}
\nu (\varepsilon )=\nu _{0}\left\vert \frac{d\xi (\varepsilon )}
d\varepsilon }\right\vert \label{nu01}
\end{equation
where the function $\xi (\varepsilon )$ is defined as the (positive-valued)
inverse function for $\varepsilon (\xi )=\sqrt{\xi ^{2}+\Delta ^{2}(\xi )}$.
We plot in Fig.\ref{DoS01} the local DoS (at $T=0$) obtained numerically
using Eq.(\ref{nu00}), as well as the average DoS obtained from Eq.(\re
{nu01}) for $\gamma=0.57$; the usual BCS DoS is shown for comparison.
\begin{figure}[tbp]
\includegraphics[width=8cm]{nu-crit-10axx}
\caption{Local zero-temperature density of states at the mobility edge of
the 3D Anderson model, averaged over 5x5 square (blue dots); average $T=0$
superconductive DoS at the mobility edge, according to Eq.(\protect\ref{nu01
), shown by dashed red line; the BCS density of states with the same gap
value (full black line). }
\label{DoS01}
\end{figure}
Note that the terms $\xi _{j}/\varepsilon $ in the parenthesis in Eq.(\re
{nu00}) were irrelevant while calculating the average DoS because
contributions from $\xi =\pm \xi (\varepsilon )$ cancel out. There is no
such cancellation for the higher moments of $\nu (\varepsilon ,\mathbf{r})$.
In particular, these terms might lead to an asymmetry of local DoS: $\nu
(\varepsilon ,\mathbf{r})\neq \nu (-\varepsilon ,\mathbf{r})$. To quantify
this asymmetry we define the antisymmetric part of the DoS: $\nu
_{a}(\varepsilon ,\mathbf{r})=\frac{1}{2}(\nu (\varepsilon ,\mathbf{r})-\nu
(-\varepsilon ,\mathbf{r})$. Using then Eqs.(\ref{nu00},\ref{nu01}), we find
the variance
\begin{equation}
\overline{\nu _{a}^{2}(\varepsilon ,\mathbf{r})}=\frac{1}{2}\left( \nu
(\varepsilon )\frac{\xi (\varepsilon )}{\varepsilon }\right) ^{2}\left[
M(0)-M(2\xi (\varepsilon ))\right] \label{nu03}
\end{equation
where the function $M(\xi )$ is defined in Eq.(\ref{ij}). Since $M(0)\sim
\mathcal{V}^{\gamma }\gg M(2\xi )$, the variance (\ref{nu03}) diverges for
an infinite system, and one should consider the distribution function
W[\rho (\varepsilon ,\mathbf{r})]$ for the dimensionless variable
\begin{equation}
\rho (\varepsilon ,\mathbf{r})=\frac{\nu _{a}(\varepsilon ,\mathbf{r})}{\nu
(\varepsilon )}\frac{\sqrt{2}\varepsilon }{\xi (\varepsilon )} \label{nu04}
\end{equation
The distribution function $W[\rho (\varepsilon )]$ coincides with that of
the wavefunction's intensities $\mathcal{P}(\psi ^{2})$, see Eq.(2.33) of
the review~\cite{MirlinNewRep}. It is determined by the "singularity
spectrum" $f(\alpha )$; the shape of this function for the 3D Anderson model
is similar to the one found for the power-law banded matrix model~\cit
{PRBM,KrMut1997} with the parameter $b\approx 0.4$, see Fig.3 of paper\cit
{MirlinNewRep}.
It is clear from the definition (\ref{nu04}) that the normalized asymmetric
fluctuations $\nu _{a}(\varepsilon ,\mathbf{r})/\nu (\varepsilon )$ are
small for $\varepsilon \approx \Delta (0)$, where $\xi (\varepsilon
)\rightarrow 0$. However, for a generic value of the energy of the order of
\Delta (0)$, the asymmetry in the tunneling spectra is strong; we emphasize
that it occurs due to mesoscopic fluctuations which do not involve any
"regular" mechanism of particle-hole asymmetry.
\subsection{Superfluid density and critical current. \label{Superfluid
density and critical current}}
We define the superfluid density $\rho_s$ via the relation $\mathbf{j} =
-\rho_{s}\mathbf{A}/c$ assuming the transverse gauge with $\mathbf
\nabla\cdot A}=0$. To estimate $\rho_s(T)$ near $T_c$, one can use the
expression (\ref{FGL}) for the free energy functional, and make there the
replacement $\nabla \to \nabla - i(2e/\hbar c)\mathbf{A}$. Then taking the
double functional derivative $\delta^2 F_{GL}/\delta \mathbf{A}^2$ one
obtains:
\begin{equation}
\rho_s(T) = 2\left(\frac{2e}{\hbar}\right)^2\,\nu_0 T_c^2\, C\, \Psi^2(T)
\label{rhos01}
\end{equation}
where $\Psi(T)=\Psi_{\mathrm{MFA}}^2 = \frac{|a(T)|}{b} = 9.5 \left(1-\frac{
}{T_c}\right)$ and the coefficient $C \sim L^2_{T}$ is given by Eq.(\ref{CC
).
We also estimate the critical current $j_{c}$ in a standard way via the
Ginzburg-Landau functional (\ref{FGL}). By extrapolating the result of
Ginzburg-Landau theory to lower temperatures, one obtains
\begin{equation}
j_{c}=c\frac{2e}{\hbar }T_{c}(\nu _{0}T_{c})^{2/3} \label{Jc}
\end{equation}
Extrapolation (\ref{Jc}) of the Ginzburg-Landau result to $T=0$ might be
wrong by a factor of the order of unity, this factor for the fractal
superconductor might be different from the known factor for a conventional
"dirty-limit" superconductor.
The results Eqs.(\ref{rhos01}),(\ref{Jc}) do not contain any
\textquotedblleft fractal\textquotedblright\ specificity and coincide with
estimates of earlier works\cite{MaLee,KapitulnikKotliar1986} that used
scaling arguments. Yet, as we have demonstrated above, the web of
superconductivity is sparse and occupies only a small fraction $f\sim
\lambda \ll 1$ of space. The derivation of Eq.(\ref{CC}) demonstrates that
the fractality-independent result for the coefficient $C$ emerges because of
the compensation of the small parameter $\lambda ^{2}$ by the enhancement
factor $(E_{0}/T_{c})^{2\gamma }$. We interpret this cancellation as the
compensation of the small cross-section area of "superconductive filaments"
by a large current that they support.
\section{Superconductivity with a pseudogap. \label{Superconductivity with a
pseudogap}}
In this section we study the superconductivity formed against the background
of moderately localized single-electron eigenstates. We demonstrate here the
existence of the interesting regime where the local pairing (discussed above
in section \ref{Insulating state} in relation with the hard-gap insulator)
coexists with the long-range superconducting correlations. In this situation
the material demonstrates the "pseudo-gap" phenomenology characterized by
the insulating behavior of $R(T)$ in a significant temperature range above
the superconducting transition. We provide a semi-quantitative description
for the evolution of the pseudo-gap $\Delta _{P}$ and $T_{c}$ while $E_{F}$
moves across the mobility edge $E_{c}$. Surprisingly, we find that
T_{c}(E_{F})$ curve is non-monotonic, with a maximum reached at the
Fermi-level on the localized side of the Anderson transition. The most
important outcome of our analysis is the existence of a range of Fermi
energies for which superconductor is characterized by relatively high values
of $T_{c}$ and by a \textit{larger} insulating pseudo-gap $\Delta _{P}\gg
T_{c}$. In this regime the low energy properties such as the formation of
the superconductivity can be described by the truncated pseudo-spin model.
In this model the effective Hamiltonian is of the form (\ref{HamSpin}) while
the Hilbert space contains only the states in which each localized orbital
is either empty or populated by a pair of electrons. The effective
Hamiltonian (\ref{HamSpin}) does not contain $\Delta _{P}$ because the
states that differ in energy by $\Delta _{P}$ are absent from the truncated
Hilbert space. These states, however, appear when single particles are
excited, thus the single particle excitations spectrum contains both
correlations and local pairing effects. Below we discuss the region of
validity of the mean-field treatment of the pseudo-spin Hamiltonian (\re
{HamSpin}) and then turn to analysis of several measurable quantities which
may provide one with the \textquotedblleft proof of the
case\textquotedblright\ for realization of the pseudo-gap superconductivity
in real material: the temperature evolution of the local single-particle
density of states $\nu (\varepsilon ,T)$, the Andreev conductance spectra
G_{A}(V)$ at $T\ll T_{c}$, and the temperature dependence of the full
spectral weight $K(T)$.
\subsection{Transition temperature and insulating gap as functions of Fermi
energy. \label{Transition temperature}}
We demonstrated in section \ref{Comparison of Tc values} that the transition
temperature found by the virial expansion to the third order provides
results very close to the ones obtained with the modified MFA. These
calculations were performed for the critical case $E_{F}=E_{c}$ where the
function $M(\omega )$ is given by the power law (\ref{ij}). This coincidence
implies that the virial expansion to the third order provide the accurate
result for the transition temperature. Generally, both methods suffer from
the neglect of the off-diagonal matrix elements, these matrix elements mix
the odd and even fermionic sectors which might decrease the $T_{c}$ below
the MFA value by a factor of the order of unity (which would be consistent
with the $\mathrm{Gi}\sim 1$ obtained for the critical case in section \re
{Ginzburg parameters}). While there is no general reasons why the
off-diagonal matrix elements can be completely neglected in the metallic and
in the critical region, in the region of well localized single-particle
states the validity of the spin Hamiltonian (\ref{HamSpin}) is ensured by
the large value of the pseudogap $\Delta _{P}$. One thus expects that
pseudospin model provides very good description of the pseudogap state and a
semi-quantitative description of the crossover from the pseudogap state
(where it is exact) to the fractal superconductor (where its results might
differ by a factor of the order of unity). Because the superconductivity
suppression deep in the insulating regime is due to the decrease in the
effective number of neighbors of each pseudospin and this effect is missing
in the modified MFA we shall use the virial expansion applied to the spin
Hamiltonian (\ref{HamSpin}) to obtain the properties in both the crossover
and pseudogap regime.
We begin by applying this method to describe the evolution of $T_{c}$ as
Fermi-level crosses the mobility edge $E_{c}$ and gets into the region of
weakly localized states. In this regime, assuming that single electron
states are characterized by small participation ratio so that corresponding
\Delta _{P}<T_{c}$, we can use the simplified version of the virial approach
with the distribution function $\tanh \frac{E_{i}}{T}$ replaced by $\tanh
\frac{E_{i}}{2T}$, as it was done for critical states at $E_{F}=E_{C}$ in
Sec. V.
\begin{figure}[tbp]
\includegraphics[width=8cm]{Tc-Jij-Mij-20xx}
\caption{(Color online) Transition temperature as a function of the
Fermi-level position, determined by the virial expansion method. Blue dots
correspond to the pseudo-spin Hamiltonian Eq.(\protect\ref{HamSpin}) with
g_{\parallel }=0$, whereas other data were obtained after including the
Hartree-type terms proportional to $g_{\parallel }$ of repulsive (red
triangles and black squares) and attractive (green stars and violet
diamonds) sign (see text for details). All calculations were done at a fixed
coupling constant $g_{\perp }=g$.}
\label{TcVersusE}
\end{figure}
The numerical results for the minimal pseudo-spin Hamiltonian Eq.(\re
{HamSpin}) with $g_{\parallel }=0$ and coupling matrix $M_{ij}$ generated by
random 3D Anderson model Eq.(\ref{AM}) are shown in Fig.\ref{TcVersusE} by
blue points. They correspond to the fixed coupling constant $g=1.7$ in the
original model and, thus, \textit{dimensionless} coupling constant $\lambda
=g\nu _{0}$ varying as function of $E_{F}$. The energy dependence of density
of states is not negligible around $E_{c}$ as can be seen in Fig.\re
{FigDoS1}, as a result $\lambda (E_{F})$ drops considerably in the range of
Fermi-energies covered in Fig.~\ref{TcVersusE}. Nevertheless, the
calculation not only confirms our analytical prediction of the enhancement
of $T_{c}$ near the critical disorder made in section \ref{Modified
mean-field} but also shows that $T_{c}$ continues to increase in a
significant interval of Fermi energies corresponding to the localized
region. This surprising result can be traced back to the behavior of the
correlation function $M(\omega )=\mathcal{V}M_{ij}$ in the domain of weakly
localized states. Indeed, it was found previously (see the discussion at the
end of section VIII B of Ref.~\cite{CueKra}) that the effective decrease of
fractal dimension $d_{2}$, and, thus, the increase of $\gamma =1-d_{2}/d$
occurs while going further into the insulating domain. For small values of
\lambda $ it results in the increase of $T_{c}\propto \lambda ^{1/\gamma }$
evident in Fig.\ref{TcVersusE} in the range of $E_{c}<E<6.5$. Note that in
contrast to Fig.\ref{VirialTc} the Fig.\ref{TcVersusE} shows the data for
the fixed value of system size, $L=20$, with no finite-size scaling
adjustment. This is why the values of $T_{c}$ data for the critical states,
E=5.5$, differ slightly between these two data sets.
The generic pseudo-spin Hamiltonian Eq.(\ref{HamSpin}) contains also the
Hartree terms
\begin{equation}
H_{\mathrm{Hartree}}=-g_{\parallel }\sum_{ij}M_{ij}S_{i}^{z}S_{j}^{z}.
\label{HamZZ}
\end{equation
Moreover, for the short-range $\delta $-\textit{function} two-body
interaction, the coupling constant $g_{\parallel }$ is equal to the
Cooper-channel coupling constant $g_{\perp }$ and, thus, the terms
containing the spin-spin interaction in pseudo-spin Hamiltonian (\re
{HamSpin}) become isotropic. Because the full Hamiltonian (\ref{HamSpin})
remains highly anisotropic due to the presence of the single spin terms, the
effect of the $S_{i}^{+}S_{j}^{-}$ and $S_{i}^{z}S_{j}^{z}$ spin-spin
interactions is dramatically different: the former leads to the
superconducting instability at low temperatures, while the latter term by
itself has only weak effect of spin susceptibility for small $g_{\parallel
}\nu $. \ Because the transition temperature depends strongly on the spin
susceptibility, in the presence of the $S_{i}^{+}S_{j}^{-}$ the Hatree term
can have slightly more significant effect. Below we study the change in the
critical temperatures induced by the $S_{i}^{z}S_{j}^{z}$ interaction for a
generic ratio $\alpha =g_{\parallel }/g_{\perp }$ using the virial expansion
method.
The results are shown in Fig.~\ref{TcVersusE} by green and violet points for
$\alpha = +0.5$ and $\alpha = +1$, and by red and black points for $\alpha =
- 1$ and $\alpha = - 0.5$, respectively. We conclude that upon the account
for the Hartree terms, all the qualitative features of the original
T_c(E_F) $ behavior, including the enhancement of $T_{c}$ by disorder, are
preserved. Therefore below we stick to the simplest version of our model
with $\alpha=0$.
With a further shift of $E_{F}$ into the localized domain, the two effects
become important. First of them is the growth of the inverse participation
ratio $M_{j}$, and the related development of the parity gap $\Delta _{P}$
\thinspace , see Eq.(\ref{DeltaP}). In some range of $E_{F}$, as we will see
soon, the typical value of $\Delta _{P}$ \thinspace\ becomes larger than
T_{c}$, which means that the calculations of the virial coefficients should
be modified, taking into account \textit{exactly} the effect of "dilution"
by the odd states discussed in section \ref{Virial expansion}. The necessary
modification of the formalism is presented in the \ref{Virial expansion
including single-occupied states}. The dilution elimination by the
pseudo-gap leads to the increase of $T_{c}$ values (as compared to the data
for $\Delta _{P}=0$ shown in Fig.\ref{TcVersusE}, where the maximum effect
of dilution was assumed). However,\ as one can see in Fig.~\ref{TcDeltaP} in
the region where superconductivity exists this increase is not more than
30\%.
\begin{figure}[tbp]
\includegraphics[width=8cm]{Tc-AM-Gauss20c1xx}
\caption{(Color online) Virial expansion results for $T_{c}(E_{F})$
\thinspace\ (red squares) and typical pseudogap $\Delta _{P}$ \thinspace\
(black dots) as functions of $E_{F}$. The model with fixed value of the
attraction coupling constant $g=1.7$ was used; pairing susceptibilities were
calculated using equations derived in Appendix B.}
\label{TcDeltaP}
\end{figure}
The second effect, which takes over sufficiently far in the localized
region, is due to the drop of the effective coordination number $Z_{\mathrm
eff}}$ of the spin model defined by Eq.(\ref{HamSpin}). Indeed, the
effective number of states $k$ that are coupled to a given state $j$ can be
roughly estimated by $Z\sim \nu _{0}T_{c}L_{loc}^{3}$ (this estimate misses
the important logarithmic factor that we discuss below around Eq.(\ref{Zeff
) ). With decreasing of the localization length $L_{loc}$, the coordination
number $Z_{\mathrm{eff}}$ drops down, and eventually becomes less than
unity, which makes description in terms of average matrix elements $\mathcal
V}\overline{M_{ij}}=M(\omega )$ meaningless. When the Fermi energy is
increased further, $T_{c}$ starts to drop down sharply simply because most
individual pseudospins (with a possible exception of rare small spin
clusters) become essentially decoupled from each other.
\begin{figure}[tbp]
\includegraphics[width=8cm]{Tc-AM-Gauss20c1logxx}
\caption{(Color online) Virial expansion results for $T_{c}(E_{F})$ (red
dots), the typical pseudo-gap $\Delta _{P}$ (black squares) and the
corresponding level spacing $\protect\delta _{L}$ (green diamonds), as
functions of $E_{F}$ in the semi-logarithmic scale.}
\label{TcDeltaPdeltaL}
\end{figure}
Evolution of both $T_{c}$ and $\Delta _{P}$ with the Fermi-level position
are shown in Fig.\ref{TcDeltaP}, where we show the results obtained by the
virial expansion method, with a finite pseudo-gap $\Delta _{P}$ taken into
account. The most important feature seen in this plot is the existence of
some range of $E_{F}$ where superconductivity with an appreciable $T_{c}$
exists, but the parity gap $\Delta _{P}$ exceeds substantially the
superconductive gap $\Delta \approx 1.7T_{c}$. Below we concentrate
specifically on this range of parameters (approximately, it corresponds to
the range $E_{F}\in (7.5,8.5)$ for the 3D Anderson model studied here),
where the specific features of superconductivity developed on top of the
pseudo-gap are most pronounced. This is precisely the region of the \textit
pseudo-gap superconductivity} where the pseudo-spin model Eq.(\ref{HamSpin})
provides a quantitatively correct description of superconductivity.
Note that in this range of parameters the typical level spacing $\delta _{L}$
\thinspace\ inside the localization volume (estimated with the use of Eq.
\ref{d-L-IPR})) is much larger than both $T_{c}$ and $\Delta _{P}$, as it is
demonstrated in Fig.\ref{TcDeltaPdeltaL}. The very existence of a nonzero
T_{c}$ in this situation is unexpected. \footnote
In contrast, superconducting correlations in ultrasmall metall grain are
destroyed at $\delta \sim \Delta $, see discussion in section \re
{Ultra-small metallic grain}.} It is related to the fact (noticed in Ref.
\cite{CueKra} and discussed in section \ref{Matrix elements}) that the
localized states characterized by the localization length $L_{\mathrm{loc}}$
overlap effectively up to the distance $R({\omega })\gg L_{\mathrm{loc}}$
between them, where $\omega \ll \delta _{L}$ is the relevant energy window
(-T_{c},T_{c})$. In the present context we can put $\omega \sim T_{c}$ and
estimate the effective interaction range $R_{0}$ and coordination number $Z_
\mathrm{eff}}$ of the spin model (\ref{HamSpin}) using the definition in Eq.
\ref{R-X}) in subsection \ref{Matrix elements}, see also Eqs.(21,25) in the
paper~\cite{CueKra}. We obtain
\begin{equation}
Z_{\mathrm{eff}}=\nu _{0}\,\frac{4\pi }{3}R_{0}^{3}(T_{c})\cdot 2T_{c}
\label{Zeff}
\end{equation
where the function $R_{0}(\omega )$ should be determined numerically. Using
data shown in Fig.~\ref{TcDeltaPdeltaL} for the case of $E_{F}=8$ and similar data
for several other values of $E_{F}$ in the pseudogap range, where $T_{c}\ll
\delta _{L}$, we obtain the following values for the effective number of
neighbors in the energy window $(-T_{c},T_{c})$:
\begin{equation}
\begin{tabular}{|l|l|l|l|l|}
\hline
$E_{F}$ & 7.75 & 8.0 & 8.25 & 8.5\\ \hline
$T_{c}$ & 0.057 & 0.044 & 0.030 & 0.02 \\ \hline
$\delta _{L}$ & 1.0 & 1.9 & 3.5 & 6.5 \\ \hline
$Z_{\mathrm{eff}}$ & 0.08 & 0.09 & 0.09 & 0.07 \\ \hline
\end{tabular}
\label{ZeffTable}
\end{equation
Of course, the numerical coefficient in Eq.(\ref{Zeff}) is somewhat
arbitrary; the main conclusion from Eq.(\ref{Zeff}) is that (due to a
presence of the logarithmic enhancement in $R_{0}(\omega )$) the effective
number of neighbors remains nearly constant in a wide range of small
T_{c}/\delta _{L}\ll 1$. Together with a weak $T_{c}/\delta _{L}$ dependence
this explains why a significant $T_{c}$ is found even for $\delta
_{L}/T_{c}\sim 300$, see Fig.~\ref{TcDeltaPdeltaL}.
As we explain in more detail below, the possibility to find a pseudogaped
superconductive state with a numerically small effective coordination number
($Z_{\mathrm{eff}}\sim 0.1$) is natural. The qualitative reason is
that our definition of $Z_{\mathrm{eff}}$ implies average over all states.
Therefore, this average includes the states that overlap strongly with each
other and the ones which overlap little with any other states. The former
states form an infinite superconducting cluster while the latter do not contribute to
superconductivity and are largerly irrelevant. In the strongly disordered
regime the superconducting cluster occupies a small part of all states, so
the average is dominated by the states which overlap very little with each
other.
Upon a further increase of the Fermi energy into the localized region the
effective number of neighbors slowly decreases leading to a sharp drop in
T_{c}^{\mathrm{typ}}(E_{F})$, see Fig.~\ref{TcDeltaP}. Finally, when the
effective number of neighbors becomes less than some critical value $Z_{c}$,
the superconducting instability ceases to occur even at $T=0$. The details
of this quantum transition and the qualitative properties of the phase
formed at high disorder are discussed in \cite{IoffeMezard2009}. This work
solves the simplified model similar to (\ref{HamSpin}) in which spin are
located of a Bethe lattice with coordination number $Z^{\mathrm{Bethe}}$,
i.e. in this model $M_{ij}=1/Z^{\mathrm{Bethe}}$ for $Z^{\mathrm{Bethe}}$
neighbors and $0$ otherwise. \footnote
Notice that this definition of the number of neighbors differs a lot from
the (\ref{Zeff}) that counts the neighbors {\it only} in the narrow energy
interval of $T_{c}$.} The result of the solution \cite{IoffeMezard2009}
which is important for the present discussion is that critical fluctuations
become important only in the narrow vicinity of the quantum critical point
at which superconductivity disappears completely. In other words, the
critical fluctuations become important only in the systems in which
transition temperature is already strongly suppressed by disorder. In more
detail, the work \cite{IoffeMezard2009} distinguishes between the high
temperature phase where corrections to mean field are small and a low
temperature phase where mean field solution becomes qualitatively incorrect
due to the strong inhomogeneity of the formed ordered phase. The temperature
($T_{\mathrm{RSB}}$) that separates them corresponds to replica symmetry
breaking (RSB) in the formalism, it turns out to be numerically small even
for modest number of neighbors as illustrated in Fig. \ref{TRSB}. Viewed
differently, it means that the effective number of neighbors in this regime
is very low (see Fig. \ref{TRSB} insert).
As long as the transition temperature obtained in the mean field
approximation is above $T_{\mathrm{RSB}}$ shown in Fig. \ref{TRSB}, the mean
field equations are qualitatively correct. Thus, one can use mean-field
approximation for the Hamiltonian (\ref{HamSpin}) to get semi-quantitative
results for the transition temperature in a large interval of Fermi energies
at which $T_{c}(E_{F})$ decreases significantly from its maximum.
The solution of the Bethe lattice model~\cite{IoffeMezard2009} can be used
to estimate critical value of the effective coordination number $Z_{\mathrm
eff}}^{\mathrm{Bethe}}=\nu _{B}T_{c0}$ $Z_{\mathrm{RSB}}^{\mathrm{Bethe}}$
which corresponds to the value $Z_{\mathrm{RSB}}^{\mathrm{Bethe}}$ of the
Bethe-lattice coordination number where $T_{c}=T_{RSB}$ and standard
mean-field solution becomes qualitatively incorrect. The details of the
corresponding calculation will be presented in the extension~\cit
{FeigelIoffeMezard} of the paper~\cite{IoffeMezard2009}; the result, valid
in the limit $Z_{\mathrm{Bethe}}\gg 1$, is
\begin{equation}
Z_{\mathrm{eff}}^{\mathrm{Bethe}}=\lambda _{b}e^{1/2\lambda _{b}}\nu
_{B}T_{c0}=\frac{1}{2}\sqrt{\pi }e^{{-\mathrm{\mathbf{C}}}/2}\frac{\sqrt{\nu
_{B}T_{\mathrm{RSB}}}}{|\log (\nu _{B}T_{\mathrm{RSB}})|}\ll 1
\label{ZBethe}
\end{equation
where $\lambda _{b}=\nu _{B}g_{B}\ll 1$ is the dimensionless coupling
constant for the Bethe lattice model, with $\nu _{B}T_{c0}=\frac{4}{\pi }e^
\mathrm{\mathbf{C}}}e^{-1/\lambda _{b}}$, and $\mathrm{\mathbf{C}}=0.577..$
is the Euler number. Eq.(\ref{ZBethe}) demonstrates again, that typical
number of neighbors with energies within the energy stripe $\sim T_{c}$ can
be very small in the weak-coupling limit, contrary to naive expectations.
Physically, the reason for this is that in this regime a dilute
superconducting cluster is formed by a small number of sites. One expects
that similar phenomena should happen in physical systems. Although our
fractal model differs significantly from the Bethe lattice model due to
power-law dependence of the interaction strength $M(\omega )$, we expect
that qualitative conclusion about $Z_{\mathrm{eff}}^{\mathrm{min}}<1$
survives in the fractal model as well.
\begin{figure}[tph]
\includegraphics[width=8cm]{TRSB}
\caption{Temperature separating high and low temperature regimes of the
simplified pseudospin model (\protect\ref{HamSpin}) on Bethe lattice with
Z_{\mathbf{Bethe}}$ neighbors, for which $M_{ij}=1/Z_{\mathbf{Bethe}}$ and
density of states per site $\protect\nu _{B}$. In the high temperature
regime $T>T_{RSB}$ the mean field theory is semi-quantitatively correct, in
the low temperature regime it is qualitatively wrong; the value of $T_{RSB}$
does not depend upon coupling constant $g$ within the Bethe lattice model.
The inset shows the effective number of neighbors in this model defined by
Z_{\mathrm{eff}}^{\mathrm{Bethe}}=\protect\nu _{B}T_{c0}Z_{\mathrm{Bethe}}$
at which the mean field solution becomes qualitatively incorrect. }
\label{TRSB}
\end{figure}
The significant number of neighbors that persists up to numerically large
values of $\delta _{L}/T_{c}\sim 300$ implies that Ginzburg parameter
remains of the order of unity in this range. \footnote
This does not cotradict our findings in section \ref{Ginzburg parameters}
that Ginzburg parameter becomes $\mathrm{Gi}\sim O(1)$ already for the
fractal superconductor with the Fermi energy at the mobility edge. It only
implies that $\mathrm{Gi}$ grows very slowly and remains $O(1)$ upon a
further increase of the Fermi energy untill the the transition temperature
becomes small. The reason for it is the appearance of a large pseudogap
\Delta _{P}\gg T_{c}$ that suppresses the fluctuational processes that
involve off-diagonal terms $M_{ijkl}$.} In this regime the gap below $T_{c}$
can be semi-quantitatively approximated by the solution of the mean field
equation:
\begin{equation}
\Delta (\xi )=\frac{\lambda }{2}\int d\zeta \frac{M(\xi -\zeta )\Delta
(\zeta )}{\sqrt{\zeta ^{2}+\Delta ^{2}(\zeta )}}\tanh \frac{\sqrt{\zeta
^{2}+\Delta ^{2}(\zeta )}}{T} \label{MMFA5}
\end{equation
where we used the fact $\Delta _{P}\gg T_{c}$ to replace $2T\rightarrow T$
and\ $M(\omega )$ is given, for localized states, by interpolating formula,
Eq.(\ref{interp}).
\begin{figure}[tbp]
\includegraphics[width=8cm]{PTcGaussAllE2xx}
\caption{Distribution functions $P(T_{c})$ obtained by the virial expansion
for different Fermi energies. All values of $T_{c}$ were obtained in the
pseudo-spin approximation $\Delta _{P}\gg T_{c}$, which leads to some
overestimation of $T_{c}$'s at the Fermi energy $E_{F}=7.9$ where $\Delta
_{P}/T_{c}$ is not very large. The bi-modal character of the distribution
for $8.2<E_{F}<8.7$ indicates on the percolation character of the
superconductive transition.}
\label{TcDistribution}
\end{figure}
One of basic results of weak-coupling BCS theory is the universal relation
between low-temperature gap value and transition temperature, $\Delta
(0)=1.76T_{c}$. As was shown in section \ref{Pairing in the modified
mean-field}, the ratio $\Delta (0)/T_{c}$ stays about the same within MFA
for fractal superconductor with $\gamma $ exponent near $0.6$, corresponding
to 3D Anderson model, see Fig.\ref{DeltaGamma}b. The presence of strong pseudo-gap
$\Delta _{P}\gg T_{c}$ leads to the doubling of $T_{c}$ for the same value
of $\Delta (0)$; on the other hand, thermal fluctuations beyond MFA always
lead to suppression of $T_{c}$ with respect to $\Delta (0)$. We do not
attempt here an explicit calculation of the $\Delta (0)/T_{c}$ ratio in the
pseudo-gap region, which should take into account both the above effects;
roughly we may expect $\Delta (0)/T_{c}\sim 1.5-2$ in this region.
With a further shift of the Fermi energy deeper into the localized region (
E_{F}\geq 8.2$ for our model), a remarkable new feature appears in the
distribution of "transition temperatures" obtained from the truncated
criterion Eq.(\ref{virial3}) via the virial expansion. Namely, it acquires
two maxima as shown in Fig. \ref{TcDistribution}. One of them is located at
T_{c}=0$ and corresponds to an insulating phase, while another one
corresponds to a superconducting phase with a nonzero $T_{c}$. The magnitude
of this latter maximum decreases with increasing $E_{F}$, until it
disappears completely at about $E_{F}=8.7$.
Dramatic modification of the character of the distribution of $T_{c}$
implies that the qualitative change in the physics of the superconductive
transition. The appearance of two maxima in the distribution function means
that not-too-large systems available for numerical studies are {\it either}
superconducting or insulating. Larger systems and physical materials become
strongly inhomogeneous with some parts becoming superconducting and some
insulating at low temperatures. This observation is fully consistent with
the result of the analytical theory\cite{IoffeMezard2009} that predicts a
replica symmetry breaking transition upon the decrease of the effective
number of neighbors. This transition signals the absence of a local
self-averaging and a global behavior which is dominated by rare
superconducting paths. Thus, it is likely that for the physical, very large
system the transition in this energy interval occurs by a \textit{percolatio
} scenario.
The inhomogeneity of the superconductor in this regime is very different
from that of the "fractal" superconductivity that we discussed in the
previous section. The latter is \textit{macroscopically homogeneous} because
the strong inhomogeneity corresponding the fractal structure is seen only at
scales smaller than $L_{T}$. In this case, the distribution function of
transition temperatures \textit{for all macroscopic samples} with $L>L_{T}$
appears to be the Gaussian single-peak distribution with the width
decreasing to zero as the system size increases, typical of the
self-averaging quantities. In contrast the bimodal distribution observed in
numerical simulations and in analytic solution\cite{IoffeMezard2009} of the
pseudospin model signals of the macroscopic inhomogeneity and lack of self
averaging.
Finally we note that our treatment above assumed a second-order transition
to the superconducting phase, for instance, we have determined the
transition temperature as a temperature of Cooper instability. We cannot
exclude the first order transition to the superconducting phase although it
seems unlikely, especially in the regime of a large disorder.
\subsection{Tunneling conductance. \label{Tunneling conductance}}
Here we analyze in detail the tunneling conductance into the pseudo-gaped
state, at temperatures above and below the superconductive transition. We
compute the \textit{average} tunneling conductance in these two cases in
sections \ref{Tunneling Normal} and \ref{Tunneling Superconductor}. These
conductances can be probed by large tunneling junctions. Computation of
local tunneling conductance (as measured by scanning tunneling probe) is
more complicated, we are able to make only qualitative predictions for this
quantity in section \ref{Point contact tunneling}.
\subsubsection{Tunneling in a normal state. \label{Tunneling Normal}}
We start by discussing the temperature-dependent differential tunneling
conductance $G(V,T,\mathbf{r})$ in the presence of a pseudo-gap $\Delta _{P}$
\textit{above} the superconducting transition. In this case the total
tunneling current is given by
\begin{eqnarray} \label{IV5}
I(V) &=&\frac{G_{0}}{e\nu _{0}}\int \int d\varepsilon d\varepsilon _{1}\nu
_{0}(\varepsilon ,\mathbf{r})\left[ (1-f(\varepsilon ))f(\varepsilon
_{1})\delta (\varepsilon _{1}-\varepsilon +eV-\Delta _{P}^{(j)})\right.
\notag \\
&&\left. -(1-f(\varepsilon _{1}))f(\varepsilon )\delta (\varepsilon
_{1}-\varepsilon +eV+\Delta _{P}^{(j)})\right]
\end{eqnarray
where $G_{0}$ is the bare conductance of the tunnel junction, $\nu
_{0}(\varepsilon ,\mathbf{r})=\sum_{j}\psi _{j}^{2}(\mathbf{r})\delta
(\varepsilon -\tilde{\xi}_{j})$, and $\tilde{\xi}_{j}\equiv \xi _{j}-\Delta
_{P}^{(j)}$. Eq.(\ref{IV5}) was obtained by the following arguments. First,
we consider the tunneling from the probe to the sample, it is given by the
first term in (151). Here the $j$-th level in the sample is empty, the probe
level is full, which gives rise to the usual factor $(1-f(\varepsilon
))f(\varepsilon _{1})$ involving the Fermi distribution function
f(\varepsilon )$. In this case the energy conservation requires $\varepsilon
_{1}=\xi _{j}-eV=\varepsilon +\Delta _{P}^{(j)}-eV$. Second, we consider the
tunneling from the sample to the probe: the $j$-th level in the sample is
full, the probe's level is empty, the thermal distribution functions produce
the factor $f(\varepsilon )(1-f(\varepsilon _{1}))$; in this case the energy
of a pair of electrons on the $j$-th level contains the binding energy
-2\Delta _{P}^{(j)}$, so that the energy conservation reads: $\varepsilon
_{1}=\xi _{j}-2\Delta _{P}^{(j)}-eV=\varepsilon -\Delta _{P}^{(j)}-eV$.
Note that this calculation assumes the validity of Fermi Golden Rule for the
tunneling rate; it also neglects electron-electron interactions (apart from
the local terms leading to the parity gap $\Delta _{P}$). The simultaneous
validity of both these assumptions is not guaranteed when we consider
localized states with relatively large typical level spacing $\delta _{L}$;
we will discuss this more below in section \ref{Point contact tunneling}.
The differential conductance $G(V)=dI/dV$ corresponding to Eq.(\ref{IV5})
can be represented in the usual form
\begin{equation}
\frac{G(V,T,\mathbf{r})}{G_{0}}=\nu _{0}^{-1}\int d\varepsilon \nu
(\varepsilon ,\mathbf{r})\left( -\frac{\partial f(\varepsilon -eV)}{\partial
\varepsilon }\right) \label{G5}
\end{equation
where the density of states in the presence of a pseudo-gap (but above
T_{c} $) depends explicitly on temperature: {\small
\begin{eqnarray}
\nu _{n}(\varepsilon ,\mathbf{r}) &=&\sum_{j}\psi _{j}^{2}(\mathbf{r})\left[
\delta (\varepsilon -\tilde{\xi}_{j}-\Delta _{P}^{(j)})(1-f(\varepsilon
-\Delta _{P}^{(j)}))\right. \label{nu5} \\
&&\left. +\,\,\delta (\varepsilon -\tilde{\xi}_{j}+\Delta
_{P}^{(j)})f(\varepsilon +\Delta _{P}^{(j)})\right] \notag
\end{eqnarray
} Note that Eq.(\ref{nu5}) cannot be represented in terms of the
normal-state DoS, \thinspace \thinspace\ $\nu _{0}(\varepsilon ,\mathbf{r})
\thinspace\ , due to the correlations between the eigenfunction intensities
\psi _{j}^{2}(\mathbf{r})$ and the local pairing gaps $\Delta _{P}^{(j)}$.
The spatial average of the tunneling DoS can be obtained from Eq.(\ref{nu5})
by integration over $\mathbf{r}$ which replaces of $\mathcal{V}\psi _{j}^{2}
\mathbf{r})$ by unity in Eq.(\ref{nu5}).
The next (approximate) step in the simplification of this expression is to
replace the summation over eigenstates $j$ by averaging over distribution of
pairing gaps ${\ P}(\Delta _{P})d\Delta _{P}$, see Fig.~\ref{PDelta}. The
distribution function of the normalized gaps $y=\Delta _{P}/\Delta _{P}^
\mathrm{typ}}$ can be fit by analytical expression
\begin{equation}
P(y)=\frac{A}{y^{2}}\exp \left( -\frac{c}{y}-b_{1}y-b_{2}y^{2}\right) \quad
\label{PDeltafit}
\end{equation
where $c=1.1$, $b_{1}=0.16$, $b_{2}=0.03$ and $A=1.8$ for $E=8.0$
(coefficients $b_{1},b_{2}$ are not universal as can be seen from Fig.~\re
{FigDoS}).
The result for average density of states then reads:
\begin{equation}
\nu _{n}(\varepsilon )=\nu _{0}(\varepsilon )\int_{0}^{\infty }\frac
e^{-yz}+\cosh \frac{\varepsilon }{T}}{\cosh yz+\cosh \frac{\varepsilon }{T}
\cdot {\ P}(y)dy \label{nu50}
\end{equation
where $z=\Delta _{P}^{\mathrm{typ}}/T$ and $\nu _{0}(\varepsilon )$ is the
average (in general, energy-dependent) DoS in the normal state. In the
low-temperature limit Eq.(\ref{nu50}) reduces to Eq.(\ref{DoS}) from section
\ref{Insulating state}.
Evolution of the average tunneling DoS obtained from Eq.(\ref{nu5}) and its
approximation Eq.(\ref{nu50}) at different temperatures $T<\Delta _{P}^
\mathrm{typ}}$ for the 3D Gaussian Anderson model are shown in Fig.\re
{TDoS1}. The asymmetry with respect to the sign of $\varepsilon $ is due to
a non-negligible energy-dependence of the bare DoS $\nu _{0}(\varepsilon )$.
\begin{figure}[tbp]
\includegraphics[width=8cm]{Dosgap41dxx}
\caption{(Color online) Global tunneling density of states for the 3D
Gaussian Anderson model at $E_{F}=8$ and the system size $20^{3}$ for
different temperatures normalized to the typical value of the local gap
\Delta _{P}^{\mathrm{typ}}$. The data points stand for the all-numerical
evaluation of Eq.(\protect\ref{nu5}) whereas the lines were calculated by
means of Eqs.(\protect\ref{nu50}) and (\protect\ref{PDeltafit}).}
\label{TDoS1}
\end{figure}
The obtained results for the DoS are translated into the measurable
tunneling conductance by Eq.(\ref{G5}). The zero-bias conductances $G(0,T)$
obtained from the "exact" (Eq.(\ref{nu5})) and the approximate (Eq.(\re
{nu50})) expressions for the DoS are shown in Fig.~\ref{G6}. It is seen from
Figs.\ref{TDoS1} and \ref{G6} that the approximation (\ref{nu50}) works
reasonably well; below we will use this approximation in the further
analysis of the average tunneling conductance.
The ensemble-averaged tunneling conductance $G(V,T)$ as a function of
voltage is plotted in Fig.~\ref{G7} for several temperatures. Remarkably,
the curves we obtained demonstrate nearly exact crossing at $eV\approx
\Delta _{P}^{\mathrm{typ}}$. In other terms, the tunneling conductance at
this voltage is nearly $T$-independent. This unexpected feature provides a
simple way for experimental determination of $\Delta _{P}^{\mathrm{typ}}$.
Note that at very low temperatures $T\ll \Delta _{P}^{\mathrm{typ}}$ the
crossing point moves to somewhat lower voltages. Indeed, a simple analysis
of the integral in Eq.(\ref{nu50}) shows that in the $T\rightarrow 0$ limit
the derivative $d\nu (\varepsilon )/dT\propto dP(\varepsilon )/d\varepsilon $
vanishes at $\varepsilon =\varepsilon _{0}$ corresponding to the maximum of
the function $P(\varepsilon /\Delta _{P}^{\mathrm{typ}})$, i.e. $\varepsilon
_{0}\approx 0.5\Delta _{P}^{\mathrm{typ}}$, see Eq.(\ref{PDeltafit}) and
Fig.~\ref{PDelta}.
\begin{figure}[tbp]
\includegraphics[width=8cm]{GVT50xx}
\caption{(Color online) Zero-bias tunneling conductance as a function of
temperature below the typical pseudo-gap value; results obtained with exact
formula (\protect\ref{nu5}) are shown by black dots, whereas blue squares
correspond to the approximate formula (\protect\ref{nu50}).}
\label{G6}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=8cm]{GVT51xx}
\caption{(Color online) Ensemble-averaged tunneling conductance, Eq.(\protect
\ref{G5}), for several values of temperature below the typical local gap
\Delta _{P}^{\mathrm{typ}}$.}
\label{G7}
\end{figure}
The averaged tunneling conductance can be measured in the normal state by a
large-area tunnel junction contact. It is important that the temperatures
are not too low with respect to the typical $\Delta _{P}$ because as
temperature drops down, the bulk resistivity grows exponentially, leading to
development of a strong Coulomb gap (not taken into the account above) which
makes tunneling measurement unfeasible. The same problem becomes even more
severe for the STM measurements at low temperatures.
\subsubsection{Tunneling in a superconductor. \label{Tunneling
Superconductor}}
Provided that Fermi Golden Rule remains valid the tunneling current in the
superconductive state is given by the formula similar to Eq.(\ref{IV5}). We
only need to replace in Eq.(\ref{IV5}) the bare normal-state DoS $\nu
_{0}(\varepsilon ,\mathbf{r})$ by its superconductive counterpart given by
Eq.(\ref{nu00}). Then, repeating the derivation of Eq.(\ref{nu5}), we come
to the expression for the tunneling density of states in presence of both
pseudo-gap and the superconducting correlations:
\begin{eqnarray}
\nu _{sc}(\varepsilon ,\mathbf{r}) &=&\frac{1}{2}\sum_{j}\psi _{j}^{2}
\mathbf{r})\left\{ \left( 1+\frac{\tilde{\xi}_{j}}{\varepsilon -\Delta
_{P}^{(j)}}\right) \left[ \delta (\varepsilon -\varepsilon _{j}-\Delta
_{P}^{(j)})+\delta (\varepsilon +\varepsilon _{j}-\Delta _{P}^{(j)})\right]
\right. \notag \\
&&\times (1-f(\varepsilon -\Delta _{P}^{(j)}))+ \label{nu51} \\
&&\left. \left( 1+\frac{\tilde{\xi}_{j}}{\varepsilon +\Delta _{P}^{(j)}
\right) \left[ \delta (\varepsilon -\varepsilon _{j}+\Delta
_{P}^{(j)})+\delta (\varepsilon +\varepsilon _{j}+\Delta _{P}^{(j)})\right]
f(\varepsilon +\Delta _{P}^{(j)})\right\} \notag
\end{eqnarray
where $\varepsilon _{j}=\sqrt{\tilde{\xi}_{j}^{2}+\Delta ^{2}(\tilde{\xi
_{j})}$ and the function $\Delta (\xi )$ has to be determined by solving the
gap equation (\ref{MMFA5}). Below we use Eq.(\ref{nu51}) in order to derive
an expression for the \textit{averaged} tunneling conductance; situation
with the \textit{local} tunneling conductance is more complicated and cannot
be described this way, as explained in subsection \ref{Point contact
tunneling}.
The averaged density of states in the superconductive state with a
pseudo-gap, $\nu _{sc}(\varepsilon )$, can be obtained in the same way as it
was done above for the normal state, see Eq.(\ref{nu50}). The odd terms in
\tilde{\xi _{j}}$ proportional to $\tilde{\xi}_{j}/(\varepsilon \pm \Delta
_{P}^{j})$ in Eq.(\ref{nu51}) cancel out upon summation over $\tilde{\xi
_{j}=\pm |\tilde{\xi}_{j}|$, provided we neglected the energy-dependence of
typical $\Delta _{P}(E)$ inside the energy band of the width $\Delta E\sim
\Delta _{P}$. We get
\begin{equation}
\nu _{sc}(\varepsilon )=\int_{0}^{\infty }{\ P}(y)\left[ \frac{\nu
_{sc}^{(0)}(\varepsilon -yzT)}{e^{yz-\varepsilon /T}+1}+\frac{\nu
_{sc}^{(0)}(\varepsilon +yzT)}{e^{yz+\varepsilon /T}+1}\right] dy
\label{nu52}
\end{equation
where $z=\Delta _{P}^{typ}/T$ and $\nu _{sc}^{(0)}(\varepsilon )$ is the DoS
in the superconductive state without a pseudo-gap, defined by $\nu
_{sc}^{(0)}(\varepsilon )=\nu _{0}(\varepsilon )|d\xi (\varepsilon
)/d\varepsilon |$ (\ref{nu01}), similar to the one shown in Fig.~\ref{DoS01}
for the case of critical superconductor. Note that the function $\xi
(\varepsilon )$ that we need here differs slightly from the one shown in
Fig.~\ref{DoS01} due to the different form of the correlation function
M(\omega )$ entering modified MFA equation (\ref{MMFA5}) in the case of
pairing between localized states. However, the effect of this difference
reduces to an overall prefactor, as is seen from Fig.~\ref{Cw2}. We thus do
not expect significant difference between the solution $\Delta (\xi )$ and
the corresponding $\xi (\varepsilon )$ functions calculated within the
modified MFA approximation in the mildly localized region considered here
and for the "critical" case studied in section \ref{Cooper instability}.
Thus we use $d\xi (\varepsilon )/d\varepsilon |$ as it was found in section
\ref{Low-temperature density of states} but take into account a smooth
energy-dependent DoS pre-factor $\nu _{0}(\varepsilon )$, as it was done
above when computing Eq.(\ref{nu50}) and Fig.~\ref{TDoS1}.
Evolution of the $\nu _{sc}(\varepsilon )$ shapes as function of temperature
in the low-temperature range $T<0.5T_{c}$ is shown in Fig.\ref{nusc}
\begin{figure}[tbp]
\includegraphics[width=8cm]{test-nusc-BCSdxx}
\caption{(Color online) Ensemble-averaged tunneling DoS, Eq.(\protect\re
{nu52}), for several values of temperature much below the value of the
superconductive gap $\Delta _{0}\equiv \Delta _{0}(\protect\xi =0)$. The
latter is chosen to be $\Delta _{0}=0.5\Delta _{P}^{\mathrm{typ}}$, where
\Delta _{P}^{\mathrm{typ}}$ is the parity gap.}
\label{nusc}
\end{figure}
where we use the analytic interpolation formula Eq.(\ref{PDeltafit}) for the
distribution function ${P}(y)$, and the $T=0$ solution for the gap function
\Delta (\xi )$. The average tunneling conductance $G(V,T)$ obtained with
this DoS from Eq.(\ref{G5}) is shown in Fig. \ref{G52}.
\begin{figure}[tbp]
\includegraphics[width=8cm]{GVT52axx}
\caption{(Color online) Average tunneling conductance in the superconductive
state for several values of temperature much below the value of the
superconductive gap $\Delta _{0}=0.5\Delta _{P}^{\mathrm{typ}}$. }
\label{G52}
\end{figure}
The crossing point at this plot, i.e. the $T$-independent conductance at
some specific value of voltage, corresponds to $eV-\Delta _{0}\approx
0.5\Delta _{P}^{\mathrm{typ}}$. In zero energy limit the expression for the
density of states simplifies to
\begin{equation}
\nu _{sc}(0)=\int_{0}^{\infty }{\ P}(y)\frac{2\nu _{sc}^{(0)}(y)}{e^{y/T}+1
dy \label{nu53}
\end{equation
The temperature dependence of the corresponding zero bias tunneling
conductance $G(0,T)$ is shown in Fig.\ref{GVTs}.
\begin{figure}[tbp]
\includegraphics[width=8cm]{GVT52bxx}
\caption{Zero-biased conductance as function of temperature for the case of
\Delta _{0}=0.5\Delta _{P}^{\mathrm{typ}}$.}
\label{GVTs}
\end{figure}
The averaged DoS and the tunneling conductance presented in Figs.~\ref{G52},
\ref{GVTs} do not show any coherence peaks similar to the one presented in
Fig.~\ref{DoS01} for the critical superconductive state (section \re
{Low-temperature density of states}). This is due to the averaging of the
peak position $eV=\Delta _{0}+\Delta _{P}^{(j)}$ over a broad distribution
P(\Delta _{P})$ of the local pairing gaps $\Delta _{P}^{(j)}$. Note that the
non-monotonic behavior of the DoS curve in Fig. \ref{nusc} reflects the
(smeared) edges at $\varepsilon =\Delta _{P}^{\mathrm{typ}}\pm \Delta _{0}$
of the $\nu _{\mathrm{sc}}^{(0)}(\varepsilon )$ function shifted by the
paring gap $\Delta _{P}^{\mathrm{typ}}>\Delta _{0}$ and weighted by the
temperature-dependent distribution functions in Eq.(\ref{nu52}).
We emphasize again that the results obtained in this section give the
averaged tunneling conductance which can be measured with a large-area
junctions, such as the ones used in~\cite{Pratap}. They may differ
significantly from the local tunneling conductance measured by a small tip
and discussed in the following subsection.
\subsubsection{Point contact tunneling. \label{Point contact tunneling}}
\bigskip The equation (\ref{IV5}) can be easily generalized to describe
point conductance. One gets using the perturbation theory in the tunneling
amplitude, $t$ (which is equivalent to Fermi Golden Rule) that the most
general case the point conductance is given by equation (\ref{IV5}) where
\nu _{0}(\varepsilon ,\mathbf{r})$ is replaced by the exact single particle
Green function at the tunneling position, $\mathrm{Im}G_{R}(\epsilon ,r)$. In
a poor conductor in which single particle states are localized or almost
localized one expects that $\mathrm{Im}G_{R}(\epsilon ,r)$ shows a series of
peaks corresponding to the single particle states located in the vicinity of
the tunneling point. Applying this reasoning to the superconductor-insulator
transition in a situation when $E_{F}$ corresponds to a well developed
paring gap $\Delta _{P}$ and repeating the arguments of the previous
subsection one expects to observe the peaks in the tunneling conductance
with width $T$ and separation $\delta _{L}$.
\bigskip These arguments do not take into account the effect of the
collective modes of the superconductor which smear the peaks in the
tunneling conductance. In fact, the presence of the collective modes make
the observation of the peak structures in the density of states difficult.
Briefly, the experimental observation of tunneling is possible only in the
superconducting state, but in this state the effect of the collective modes
on the single particle density of states is large. We estimate this effect
within a simplified model, namely we use the pseudospin Hamiltonian (\re
{HamSpin}) to which we add the possibility of having single particle
excitation at the single site $0$:
\begin{equation}
H=\xi _{0}n+\Delta _{P}(1-n^{2})-g\left[ c_{\uparrow}^{\dagger }c_{\downarrow}^{\dagger
}\sum_{j}M_{0j}S_{j}^{-}+c_{\downarrow}c_{\uparrow}\sum_{j}M_{0j}S_{j}^{+}\right] +H_{PS}(S)
\label{H-PS-tun}
\end{equation
where $n=\sum_{a}c_{a}^{\dagger }c_{a}-1$ where $a=\uparrow ,\downarrow $
are two spin components, and the sum over $j$ runs over the neighbors of the
site $0.$ This model describes the process of electron tunneling in a single
localized state; in a more realistic case the electrons tunnel in few such
states with an amplitude proportional to $\psi _{\alpha }^{2}(r)$ where $r$
is a tunneling point. These processes are incoherent due to randomness of
local binding energies $\Delta _{P}^{(j)}$, so the physical tunneling
conductivity is the sum of the contributions from different eigenstates.
In a good BCS-like superconductor characterized by a large number of
neighbors one can replace pairing field $h_{\perp }=g$
\sum_{j}M_{0j}S_{j}^{\perp }$ by their average values $\left\langle h_{\perp
}\right\rangle =\Delta $. In this case the Green function of the electron at
site $0$ obeys Dyson equation
\begin{eqnarray}
G_{R}(\epsilon ) &=&\frac{1}{G_{0R}^{-1}(\epsilon )-\Delta
^{2}G_{0A}(-\epsilon )} \label{Dyson_BCS} \\
G_{0R}(\epsilon ) &=&\frac{1}{\epsilon -\xi _{0}-\Delta _{P}\mathrm{sgn}(\xi
_{0})+i0}.
\end{eqnarray
The resulting single electron Green function has a pole at $\epsilon _{0}
\sqrt{(\xi _{0}+\Delta _{P}\mathrm{sgn}(\xi _{0}))^{2}+\Delta ^{2}}$ which
shows the combined effect of the pseudogap and superconducting order
parameter on the electron. As one expects, in the case of a large number of
neighbors the electron density of states has a sharp peak at the pole energy
$\epsilon _{0}$ which is smeared only by thermal effects by $\delta \epsilon
\sim T$. As the disorder is increased the effective number of neighbors goes
down and fluctuations of the pairing field become relevant. This leads to
two physical effects. The first effect is that tunneling of the electrons
removes the spin from site $0$ which changes the effective spin Hamiltonian.
The new ground state spin wave function is orthogonal to the one realized
before the tunneling process. This leads to a suppression of the tunneling
of the electron with the energy exactly equal to $\epsilon _{0}$. The second
effect is that tunneling might be accompanied by the emission of the
pseudospin excitations. This leads to non-zero tunneling density of states
at energies larger than $\epsilon _{0}$. Both effects are controlled by the
same parameter, the effective number of neighbors which determines the
strength of the quantum fluctuations of the pairing field and both round the
peaks in the observed spectrum. In order to compute local tunneling
conductance one needs the full description of the spin fluctuations in the
regime of the small $Z_{eff}$ which is not presently available.
Below we estimate the second effect, i.e. the magnitude of the level
broadening in the framework of the perturbation theory. In perturbation
theory the emission of the pseudospin excitations is described by the
inclusion of the self energy part in the Dyson equation (\ref{Dyson_BCS}):
\begin{eqnarray}
G_{R}(\epsilon ) &=&\frac{1}{\widetilde{G}_{R}^{-1}(\epsilon )-\Delta ^{2
\widetilde{G}_{A}(-\epsilon )},\qquad \widetilde{G}_{R}(\epsilon )=\frac{1}
\epsilon -\xi _{0}-\Delta _{P}\mathrm{sgn}(\xi _{0})-\Sigma _{R}(\epsilon )},
\label{Sigma_R} \\
\mathrm{Im}\Sigma _{R}(\epsilon ) &=&g^{2}\sum_{j}M_{0j}^{2}\int_{0}^{\epsilon
}\frac{d\omega }{2\pi }\mathrm{Im}\widetilde{G}_{A}(\omega -\epsilon )\mathrm{Im
D_{R}(\omega )
\end{eqnarray
where $D(\omega )=$ $\left\langle \left\langle S^{\perp }S^{\perp
}\right\rangle \right\rangle _{\omega }$ is irreducible correlator of the
pairing field. Even in a weakly superconducting phase the pseudospin
excitations are delocalized and thus have a continuos spectrum. Here we
focus on the large energy scale $E\sim \Delta _{P}$ properties of the
spectrum, we do not discuss here the structure of tunneling conductance very
close to the edge, where coherence peaks are expected. At these large
energies, the density of collective modes is featureless, so we estimate the
$\mathrm{Im}D(\omega )=\nu _{B}\theta (\omega )$.
As it is clear from (\ref{Sigma_R}), the interaction with pseudospins lead
to the decay rate of the single particle excitations, $\Gamma_{el}$ , which
smears the peak structure of $\mathrm{Im}G(\epsilon )$ replacing it with the
threshold behavior. \ In the leading order of perturbation theory we replace
$\widetilde{G}(\omega -\epsilon )$ in (\ref{Sigma_R}) by $G_{0}(\omega
-\epsilon )$ and get for the level width $\Gamma _{el}=g^{2}\nu
_{B}\sum_{j}M_{0j}^{2}$.
Because the expression for the level width, $\Gamma _{el}$ contains the
square of the interaction constant, $g^{2}$, but only one sum over
neighboring sites, it is small in a good superconductor (where $gM\propto
1/Z $) but may become large in a weak superconductor close to
superconductor-insulator transition. We estimate it in the framework of the
Bethe lattice model for this transition discussed in section \ref{Transition
temperature}. In this model $M_{0j}=1/Z_{\mathrm{Bethe}}$, so we get $\Gamma
_{el}=g^{2}\nu _{B}/Z_{\mathrm{Bethe.}}$ In order to express it in terms of
the physical parameters, we use the equation (\ref{ZBethe}) that relates the
transition temperature to the effective number of neighbors $Z_{\mathrm
Bethe.}}$ at the onset of the weak superconductor regime. The onset of this
regime can be defined (and determined experimentally) by the beginning of
the sharp decrease of $T_{c}$ with the increase of disorder. We get:
\begin{equation}
\Gamma _{el}\sim (g\nu _{B})^{2}\sqrt{T_{c}E_{F}}\gg T_{c}\,.
\label{GammaEl}
\end{equation
The apparently large value of the level width in this regime implies that
distinct peaks become absent, being replaced by the threshold. For a
tunneling leading to a single state, $\alpha =0$, the threshold coincides
with the value of energy $\epsilon _{\alpha }\approx \left[ \xi _{\alpha
}^{2}+\Delta ^{2}\right] ^{1/2}+\Delta _{P}^{(\alpha )}$. This discussion
assumed that tunneling happens only in one localized state; in a more
realistic situation when tunneling amplitudes proportional to $\psi _{\alpha
}^{2}(r)$ are significant for a few states, the threshold energy is equal to
$\Delta $ plus a minimal value of $\min \Delta _{P}^{\alpha }$ available for
a contact tunneling process at a given point.
We conclude that the model Hamiltonian Eq.~(\ref{H-PS-tun}) results in the
threshold dependence of the tunneling conductance. The threshold
distribution (observed when scanning along the surface) is expected to be
similar to $P(\Delta _{P})$ distribution shown in Fig.~\ref{PDelta} whereas
the tunneling conductance $G(V)$ should be similar to the integral of that
distribution shown in Fig.~\ref{FigDoS}. These conclusions are in a
qualitative agreements with the data presented in~\cite{Sacepe2007,Sacepe2}.
The model ~(\ref{H-PS-tun}) neglects the off-diagonal matrix elements of the
form
\begin{equation*}
\sum_{j}c_{\alpha \uparrow }c_{j\downarrow }\sum_{k}M_{\alpha jkk}\langle
c_{k\uparrow }^{+}c_{k\downarrow }^{+}\rangle
\end{equation*
that are present in the full Hamiltonian and might be important for the
quantitative description of the data. Physically these terms lead to the
possibility of non-local process in which electron is converted into a hole
in the adjacent state and emits a collective mode quantum. The presence of a
sum over many states $j$ connected with a pre-selected state $\alpha $ leads
to a partial average over the gap distribution $P(\Delta _{P})$; this
suppresses the relative magnitude of spatial fluctuations of the local
thresholds. As a result, we expect that the threshold voltage is given by
\begin{equation}
eV_{\mathrm{th}}=\alpha _{\mathrm{th}}\overline{\Delta }_{P}+\Delta
\label{Vth}
\end{equation
with $\alpha _{\mathrm{th}}<1$ being numerical factor of order unity, with a
moderate spatial fluctuations.
The arguments leading to an estimate Eq.(\ref{GammaEl}) imply that peak
structure of the density of states might be only observed in a narrow
transient regime that corresponds to a well developed pseudogap combined
with a strong superconductor. It is not clear that this regime can be
realized in physical systems. In particular, it is very likely that
experiment \cite{Sacepe2007,Sacepe2} which reported the gap in the density
of states but no obvious peak structure were performed in a weakly
superconducting state close to the transition into the insulator. As
explained above, in this regime one expects that at a typical point the
spectrum of collective modes is continuos and featureless resulting in the
tunneling conductance characterized by a threshold that corresponds to the
minimal value of $\epsilon _{0}$. However, large inhomogeneity of the
ordered state (discussed in a different context in the section \re
{Transition temperature}) implies that at some locations the spectrum of
collective modes might develop large peaks leading to the peak structures in
the tunneling conductance.
Qualitatively, interaction of single electron with pseudospin collective
modes plays the same role as its interaction with other electrons in zero
bias anomaly and Fermi-edge singularity problem~\cite{Nozieres}. The
quantitative theory of these effects and in particular, the predictions for
the shape of the pseudogap require the full theory of the collective modes
appearing in the vicinity of superconductor-insulator transition, this is
beyond the approach developed in this paper.
\subsection{Andreev contact conductance at low temperatures. \label{Andreev
point contact}}
In the regime of a strong pseudo-gap $\Delta _{P}\geq \Delta _{0}$ a
peculiar situation occurs: the energy gap for a single-particle excitation
\Delta _{1}=\Delta _{P}+\Delta _{0}$ is \textit{larger} than the
two-particle excitation gap $2\Delta _{0}$. The crucial point here is that
the \textit{local} paring energy $\Delta _{P}$ drops out of the expression
for the two-particle excitation gap which is determined only by the \textit
collective} superconductive gap $\Delta _{0}$. Local paring energy becomes
important only when we have to break the pair to produce the single-particle
excitation. This allows to experimentally distinguish between the two types
of gaps: the paring gap $\Delta _{1}$ seen in the \textit{single-electron}
tunneling and the collective gap $\Delta _{0}$ which can be observed in the
experiments where the entire pair is transferred through the tunnel
junction. This type of tunneling experiments is called Andreev-contact
spectroscopy (see \cite{Deutscher} for the review of its application to
similar studies in high-$T_{c}$ cuprates). Since the probability to transfer
a pair is proportional to the square of the tunnel contact transparency
\mathcal{T}$, the latter should be not too small for the Andreev conductance
to be observable.
\begin{figure}[tbp]
\includegraphics[width=8cm]{AndreevConductivity}
\caption{(Color online) Andreev point-contact conductance (normalized by its
value at zero bias) is calculated within zero-temperature BTK approximation,
Eq.(\protect\ref{Gns}), for three different values of the contact
transparency, $\mathcal{T} = 0.2$ (blue line) , $0.4$ (red) and $0.6$
(yellow). Voltage $eV$ is normalized by the maximum value $\Delta_0$ of the
gap function $\Delta(\protect\xi)$. It is assumed that single-electron
tunnelling is fully suppressed by large gap $\Delta_P \gg \Delta_0$.}
\label{Andreev}
\end{figure}
Below we provide simplest estimates of Andreev conductance for the
pseudogapped fractal superconductor, based upon Blonder-Tinkham-Klapwijk
(BTK) approach~\cite{BTK}. We consider the case of moderately large
parameter of normal reflection $Z\geq 1$. Charge transport within BTK
approach is characterized by (energy-dependent) probabilities $A$, $B$, $C$
and $D$ of four processes which may occur when normal incident electron
approaches N-S boundary: $A(\varepsilon )$ is the probability of Andreev
reflection, $B(\varepsilon )$ corresponds to normal reflection,
C(\varepsilon )$ and $D(\varepsilon )$ correspond to two different processes
of single-electron transmission into superconductor. We are interested here
in the low-temperature limit $T\ll T_{c}\ll \Delta _{P}$ and relatively low
voltages $eV<\Delta _{P}$, thus the processes of electron transmission into
superconductor can be neglected and the full sub-pseudogap current can be
written as
\begin{equation*}
I_{NS}(V)\propto \int d\varepsilon \,\,2A(\varepsilon )[f(\varepsilon
-eV)-f(\varepsilon )]
\end{equation*
This expression is obtained from Eq.(17) of the BTK paper by the
substitution of the integrand $1+A-B\equiv 2A+C+D$ for $2A$, due to
vanishing of normal current. Therefore differential conductance is
proportional to the Andreev reflection probability $A(\varepsilon )$, which
we will write using BTK results, see Table II of Ref.~\cite{BTK}, but in
slightly different notations. To begin with, we introduce normal
transmission coefficient $\mathcal{T}=1/(1+Z^{2})$ and write NS differential
conductance in BTK approximation (at $T=0$) for usual superconductor as
\begin{eqnarray}
\frac{dI_{NS}}{dV} &=&G_{NS}(V) \label{Gns} \\
&=&2G_{0}\mathcal{T}\left( \frac{\Delta (eV)}{eV}\right) ^{2}\frac
n_{0}^{2}(eV)}{\mathcal{T}^{2}n_{0}^{2}(eV)+(2-\mathcal{T})^{2}+2n_{0}(eV)(2
\mathcal{T})\theta (eV-\Delta )} \notag
\end{eqnarray
where $G_{0}=2\mathcal{T}N_{ch}$ is the normal-state conductance of the
contact, $N_{ch}$ is the number of orbital conductance channels (in BTK
approach all channels are characterized by the same transmission coefficient
$\mathcal{T}$), and $n_{0}(\varepsilon )=\varepsilon /|\varepsilon ^{2}-
\Delta }^{2}|^{1/2}.$
In order to generalize Eq.(\ref{Gns}) for the case of fractal
superconductor, we consider subgap region $\varepsilon <\Delta _{0}$ and
region of real excitations $\varepsilon >\Delta _{0}$ separately. In the
subgap region we introduce function $n_{-}(\varepsilon )$ defined by
\begin{equation}
n_{-}(\varepsilon )=\frac{1}{\pi }\int d\xi \frac{\Delta (\xi )}
-\varepsilon ^{2}+\xi ^{2}+\Delta ^{2}(\xi )} \label{feps}
\end{equation
where $\Delta (\xi )$ is the solution to the gap equation (\ref{MMFA5}).
Then NS conductance in the subgap region is:
\begin{equation}
G_{NS}^{-}(V)=2G_{0}\mathcal{T}\frac{n_{-}^{2}(eV)}{(n_{-}^{2}(eV)-1
\mathcal{T}^{2}+(2-\mathcal{T})^{2}} \label{NSsub}
\end{equation
For the region above the gap one should introduce function $\hat{\Delta
(\varepsilon )\equiv \Delta (\xi (\varepsilon ))$, and replace in Eq.(\re
{Gns}) the function $n_{0}(\varepsilon )$ by the actual normalized density
of states $n_{+}(\varepsilon )=\nu (\varepsilon )/\nu _{0}$, defined by Eq.
\ref{nu01}):
\begin{equation}
G_{NS}^{+}(V)=2G_{0}\mathcal{T}\frac{(\hat{\Delta}(eV)/{eV
)^{2}\,\,n_{+}^{2}(eV)}{\mathcal{T}^{2}n_{+}^{2}(eV)+(2-\mathcal{T
)^{2}+2n_{+}(eV)(2-\mathcal{T})} \label{NSover}
\end{equation
The comment is in order here: we do not expect Eqs.(\ref{NSsub},\ref{NSover
) to be quantitatively exact for the problem considered, due to a number of
oversimplifying assumptions borrowed from the BTK approach (it was developed
for clean superconductors in a contact with clean metal). However,
comparison of BTK expression for $A(\varepsilon )$ with exact calculations
done for dirty NS junctions (see Appendix to the paper~\cite{Bezuglyi} and
references in that paper, and also paper~\cite{Golubov03} ) shows that the
key feature of the result (\ref{NSsub},\ref{NSover}), that is, the peak of
differential conductance at $eV=\Delta _{0}$, is reproduced by these more
adequate calculations; thus we hope it will survive for very strongly
disordered pseudogaped superconductor as well.
In Fig.~\ref{Andreev} we present the results of the computation of the $T=0$
differential Andreev conductivity based on Eqs.(\ref{NSsub},\ref{NSover})
for several values of the contact transparency. Note that we give here the
data for Andreev conductivity normalized to its value at zero voltage $G(0)
, which scales itself as $\mathcal{T}^{2}$. Therefore the plots presented in
Fig.\ref{Andreev} illustrate variation of the $G(V)$ peak shape, but do not
show the overall scale of $G(V)$ as function of $\mathcal{T}$. The gap
function $\hat{\Delta}(\varepsilon )$ was computed using the $T=0$ limit of
the solution to the modified MFA equation (\ref{MMFA5}) for the "critical"
M(\omega )$, Eq.(\ref{ij}). Voltage is normalized by the maximum value
\Delta _{0}$ of the gap function $\Delta (\xi )$.
Note that the plots in Fig.\ref{Andreev} illustrate the limit of a very
large paring gap $\Delta_{P} \gg \Delta_{0}$. In the case of parity gap
\Delta_P$ comparable to the superconductive gap $\Delta(0)$, the
point-contact conductance should contain two peaks: at $eV_1 = \Delta(0)$
due to Andreev processes, and at $eV_2 = \Delta_P + \Delta(0)$ due to
single-electron tunnelling. These two voltage values behave differently with
change of temperature and location of the point contact. The voltage $V_1$
is determined by the \textit{collective superconductive gap} $\Delta_{0}$
and, therefore, is position-independent but goes to zero as $T \to T_c$. In
contrast, $V_2$ fluctuates strongly between different contact locations due
to the local paring gap $\Delta_{P}$ but does not vanish as $T$ approaches
T_c$ from below. Recent experimental results~\cite{Dubouchet10} seem to
support the above predictions.
\subsection{Spectral weight of high-frequency conductivity and
superconducting density. \label{Spectral weight}}
The studies\cite{Basov} of high $T_{c}$ superconductors has shown that the
unconventional physics of these materials can be probed by the temperature
dependence of the spectral weight $K^{tot}(T)$ of the high-frequency
conductivity defined by
\begin{equation}
K^{tot}(T)=\frac{2}{\pi }\int_{0}^{\Omega _{max}}\Re \sigma (\omega
,T)d\omega +\rho _{s}(T)\equiv K(T)+\rho _{s}(T). \label{KT1}
\end{equation
Here $\sigma (\omega ,T)$ denotes the regular part of conductivity, whereas
the contribution $\rho _{s}=-c\delta \mathbf{j}/\delta \mathbf{A(\omega =0)}$
of superconductive part (proportional to $\delta (\omega )$) is explicitly
separated in the last term of Eq.(\ref{KT1}); the high-frequency cutoff
\Omega _{max}$ is chosen a several times larger than $T_{c}$.
In agreement with the data on conventional superconductors, BCS theory and
its strong coupling generalizations predict that $K^{tot}(T)$ does not
change as temperature decreases below the superconducting transition
temperature $T_{c}$: the appearance of $\rho _{s}(T)$ is compensated by the
decrease of the regular part. In contrast, ~for underdoped cuprates the
regular part $K(T)$ of the c-axis spectral weight changes below $T_{c}$ by
one half of $\rho _{s}(T)$ only, so that $K^{tot}(0)-K^{tot}(T_{c})=\frac{1}
2}\rho _{s}(0)$.\cite{Basov}
The observation~\cite{Basov} was explained by Ioffe and Millis~\cit
{IoffeMillis} as being caused by formation of a pseudo-gap that survives up
to a temperature $T_{PG}$ well above $T_{c}$, and at the same time by the
absence of the inter-layer coherence in the range $T_{c}<T<T_{PG}$ due to
strong quantum and thermal fluctuations. The weight non-conservation occurs
only for ~the c-axis conductivity, whereas the in-plane conductivity behaves
in a usual way in all cuprates.\cite{Basov} The theory~\cite{IoffeMillis}
attributes to the large phase fluctuations between weakly coupled layers and
a tunneling nature of c-axis transport, in contrast with the smooth phase
variations and the continuous electron motion in the planes.
We expect that pseudogapped superconductivity formed near
superconductor-insulator transition shows similar anomaly in the behavior of
the spectral weight $K_{tot}(T)$. In particular, we show that inside the
strong pseudo-gap region $T_{c}\ll \Delta _{P}$ the effect of the
K^{tot}(T) $ non-conservation is even stronger: its variation $\Delta
K^{tot}(T)$ is equal (within the mean field approximation) to $\rho _{s}(T)
, whereas the variations of the regular part of spectral weight $\Delta K(T)$
are smaller by $\sim 1/Z_{\mathrm{eff}}$ factor. In this derivation we shall
assume that pseudogap sets the largest scale in the problem, i.e. that the
upper frequency cutoff $\Omega _{max}$ satisfies $T_{c}\ll $ $\Omega
_{max}\ll \Delta _{P}$. This assumption allows us to neglect completely
single electron processes (which do not contribute for $\omega <\Delta _{P}
) and use pseudospin model for the computation of the conductivity in the
entire frequency range. Furthermore, because, as we shall show below the
conductivity in the pseudospin model decreases fast at frequencies above
T_{c}$, we can replace the upper cutoff $\Omega _{max}\rightarrow \infty $
when evaluating the spectral weight in this model.
In order to compute the conductivity in the pseudospin model we need to
include in the spin Hamiltonian (\ref{HamSpin}) the effect of the vector
potential $\mathbf{A}$:
\begin{equation}
H=\sum_{j}2\xi _{j}S_{j}^{z}-\frac{g}{2
\sum_{ij}M_{ij}(S_{i}^{+}S_{j}^{-}e^{i\phi _{ij}}+h.c.)\,, \label{HamSpinA}
\end{equation
where $\phi _{ij}=\frac{2e}{\hbar c}\int_{\mathbf{r}_{i}}^{\mathbf{r}_{j}
\mathbf{A}d\mathbf{r}$. Here $i,j$ denote localized single-particle states
(orbitals) with wavefunctions $\psi _{i,j}(\mathbf{r})$. The Hamiltonian in
the form (\ref{HamSpinA}) is applicable provided that the typical distance
R_{0}$ between centers of localization of the relevant eigenstates is much
longer than the localization length $L_{loc}$ (the same assumption is used
everywhere in this section, see Eq.(\ref{Zeff})). Then $i,j$ can also be
viewed as sites characterized by the position vectors $\mathbf{r}_{i}$,
\mathbf{r}_{j}$ of centers of localization of the corresponding orbitals.
The assumption $R_{0}\gg L_{\mathrm{loc}}$ implies that the theory is not
sensitive to an exact definition of "center of localization" inside the
localization radius. It is this assumption that allows to define the phase
\phi _{ij}$ in Eq.(\ref{HamSpinA}). Another assumption implied in (\re
{HamSpinA}) is that the magnetic field is weak enough so that its effect
reduces to the introduction of the phase factors $e^{i\phi _{ij}}$ and does
not change the matrix elements $M_{ij}$.
The electrical current across each link $(ij)$ that corresponds to the
Hamiltonian Eq.(\ref{HamSpinA}) is
\begin{equation}
\hat{\mathcal{I}}_{ij}=\frac{g}{2}M_{ij}\,i(S_{i}^{+}S_{j}^{-}e^{i\phi
_{ij}}-h.c.). \label{Noetter}
\end{equation
The response of the superconductor to the slowly spatially varying
electromagnetic field is fully characterized by $Q(\omega )=-c\delta
\mathcal{j}/\delta A_{k=0}(\omega )=Q_{0}+\widetilde{Q}(\omega )$ where we
separated the constant part $Q_{0}$ and the frequency dependent one
\widetilde{Q}(\omega )$ that satisfies $\widetilde{Q}(\omega \rightarrow
\infty )=0$. At high frequencies the conductivity is purely imaginary with
\mathrm{Im}\sigma (\omega )=$ $Q_{0}/\omega $, expressing $\mathrm{Im}\sigma
(\omega )$ through the real part by Kramers-Kronig relations one gets that
K^{tot}(T)=Q_{0}$.
The constant term $Q_{0}$ originates from the direct expansion of the
expression for the current (\ref{Noetter}) in $\phi _{ij}$
\begin{equation}
K^{tot}(T) = Q_{0}(T) =\frac{g}{2\mathcal{V}}\sum_{ij}\left( \frac{2e}{\hbar
}x_{ij}\right) ^{2}M_{ij}Q_{ij}^{(0)} \label{Q_0}
\end{equation
where $Q_{ij}^{(0)}=\left\langle S_{i}^{+}S_{j}^{-}+h.c.\right\rangle $.
The equation (\ref{Q_0}) implies that in the general case $Q_{0}$ cannot be
temperature independent because $\left\langle
S_{i}^{+}S_{j}^{-}+h.c.\right\rangle $ is generally temperature dependent.
Furthermore, one expects a strong temperature dependence to appear below
T_{c}$ where order parameter induces large spin-spin correlator. In the
leading order in $1/Z_{eff}$ one can replace $\left\langle
S_{i}^{+}S_{j}^{-}\right\rangle \approx \left\langle S_{i}^{+}\right\rangle
\left\langle S_{j}^{-}\right\rangle $ and conclude that the spectral weight
is constant above $T_{c}$ but acquires temperature dependence below $T_{c}$:
\begin{equation}
K^{tot}(T)=\frac{1}{4\mathcal{V}}\sum_{ij}gM_{ij}\left( \frac{2e}{\hbar
x_{ij}\right) ^{2}\frac{\Delta _{i}\Delta _{j}\tanh \beta \varepsilon
_{i}\tanh \beta \varepsilon _{j}}{\varepsilon _{i}\varepsilon _{j}}.
\label{K^total}
\end{equation
where $\varepsilon _{i}=\sqrt{\xi _{i}^{2}+\Delta _{i}^{2}}$.
We now compute the leading correction in $1/Z_{eff}$ to $\left\langle
S_{i}^{+}S_{j}^{-}\right\rangle $ \ above $T_{c}$ , we find that it does not
result in the spectral weight temperature dependence in this temperature
range. This result does not preclude that temperature dependence appears in
high orders in $1/Z_{eff}$ but it implies that the temperature dependence is
very small in this regime. Note that in this temperature range the
superconducting response is absent, so temperature independence of the full
weight implies temperature independence of its regular part. To compute the
leading contribution to $\left\langle S_{i}^{+}S_{j}^{-}\right\rangle $ it
is sufficient to diagonalize two spin problem with the Hamiltonian
\bigskip
\begin{equation}
H=\xi _{j}\sigma _{j}^{z}+\xi _{i}\sigma _{i}^{z}-\frac{g}{2
M_{ij}(S_{i}^{+}S_{j}^{-}+S_{j}^{+}S_{i}^{-})\,, \label{Two Spin H}
\end{equation}
which mixes states pairwise. \bigskip We get
\begin{equation*}
\left\langle S_{i}^{+}S_{j}^{-}+S_{j}^{+}S_{i}^{-}\right\rangle =\frac{g}{2
M_{ij}\left[ \frac{\tanh (\beta \xi _{i})-\tanh (\beta \xi _{j})}{4(\xi
_{i}-\xi _{j})}+\frac{\tanh (\beta \xi _{i})+\tanh (\beta \xi _{j})}{4(\xi
_{i}+\xi _{j})}\right]
\end{equation*}
Inserting this expression in (\ref{Q_0}) and averaging over distribution of
\xi $ , assuming that it is not correlated with the values of $M_{ij}$ and
x_{ij}$we get
\begin{eqnarray*}
Q_{0} &=&\sum_{ij}\overline{\left( \frac{2e}{\hbar }x_{ij}\right) ^{2}
\overline{\;\left( \frac{g}{2}M_{ij}\right) ^{2}}R \\
R &=&\;\overline{\frac{\tanh (\beta \xi _{i})-\tanh (\beta \xi _{j})}{2(\xi
_{i}-\xi _{j})}}
\end{eqnarray*
The temperature enters only in the last factor in this product,
differentiating it over the inverse temperature we fin
\begin{equation*}
\frac{dR}{d\beta }\;=\int d\xi _{1}d\xi _{2}\frac{\nu (\xi _{1})\nu (\xi
_{2})}{\xi _{1}-\xi _{2}}\left[ \frac{1}{\cosh ^{2}\beta \xi _{1}}-\frac{1}
\cosh ^{2}\beta \xi _{2}}\right]
\end{equation*
The integrand decreases exponentially fast for $\xi \gg T$ , neglecting the
contribution from these regions and assuming, as usual, that $\nu (\xi
\lesssim T)=\nu $ is constant at low energies we perform integration and get
the result announced earlier: $dR/d\beta =0$.
We now discuss the regular and superconducting contributions to the spectral
weight below $T_{c}$. The computation above demonstrates that leading terms
in $1/Z_{eff}$ involve only pairs of spins. This remains true below $T_{c}$
as well and allows us to compute easily the regular part of the spectral
weight in this temperature range. We begin by using the Kubo formula to
express it through the spin-spin correlator
\begin{equation}
K(T)=\frac{1}{\mathcal{V}}\sum_{ij}\left( \frac{2e}{\hbar }x_{ij}\right)
^{2}\left( \frac{g}{2}M_{ij}\right) ^{2}\int_{0}^{\infty }\frac{d\omega }
\omega }\mathrm{Re}R_{ij}(\omega ) \label{K_regular}
\end{equation}
where $R_{ij}(\omega )$ is current-current correlator
\begin{equation}
R_{ij}(\omega )=i\int_{0}^{\infty }dte^{i\omega t}\left\langle \left[ \hat{I
_{ij}(t),\hat{I}_{ij}(0)\right] \right\rangle \label{Rij}
\end{equation
evaluated in the absence of the external field: $\hat{I
_{ij}(t)=i(S_{i}^{+}S_{j}^{-}-S_{i}^{-}S_{j}^{+})_{t}$. The equation (\re
{K_regular}) can be simplified by using the Kramers-Kronig relation for the
correlator $R_{ij}(\omega )$:
\begin{equation}
K(T)=\frac{1}{\mathcal{V}}\sum_{ij}\left( \frac{ge}{\hbar
x_{ij}M_{ij}\right) ^{2}\mathrm{Im}R_{ij}(0) \label{KT}
\end{equation}
\bigskip\
\begin{eqnarray} \label{K_regular_below}
K(T &<&T_{c})=\frac{1}{8\mathcal{V}}\sum_{ij}\left( \frac{2e}{\hbar
x_{ij}\right) ^{2}\left( \frac{g}{2}M_{ij}\right) ^{2}\left[ \left( \frac
\xi _{i}}{\varepsilon _{i}}+\frac{\xi _{j}}{\varepsilon _{j}}\right) ^{2
\frac{\tanh (\varepsilon _{i}/T)-\tanh (\varepsilon _{j}/T)}{\varepsilon
_{i}-\varepsilon _{j}}\right. \notag \\
&&\left. +\left( \frac{\xi _{i}}{\varepsilon _{i}}-\frac{\xi _{j}}
\varepsilon _{j}}\right) ^{2}\frac{\tanh (\varepsilon _{i}/T)+\tanh
(\varepsilon _{j}/T)}{\varepsilon _{i}+\varepsilon _{j}}\right] .
\end{eqnarray
One can immediately see that the regular part (\ref{K_regular_below})
contains an extra factor $\sim \frac{gM_{ij}}{\varepsilon }$ compared to the
total spectral weight $K^{tot}(T)$ in (\ref{K^total}). Thus we conclude that
variation of the regular spectral weight below $T_c$ is smaller than total
spectral weight:
\begin{equation}
\frac{K(T_c) - K(T)}{K^{tot}(T)}\sim \frac{gM_{ij}}{T_{c}}\sim \frac{1}{Z_
\mathrm{eff}}}\, . \label{comparison}
\end{equation
The last estimate in Eq.(\ref{comparison}) may be considered as a practical
definition of $Z_{\mathrm{eff}}$. Indeed, in lattice spin models where each
spin is coupled to $Z_{\mathrm{eff}}$ other spins by a coupling constant $gM$
the transition temperature can be estimated from $gMZ_{\mathrm{eff}}\sim
T_{c}$. As we argued above (see Eqs.(\ref{Zeff},\ref{ZBethe}) and the
corresponding discussion in section \ref{Transition temperature}), the
effective coordination number $Z_{\mathrm{eff}}$ is not too small (and
superconductivity survives) in a wide region of localized single-particle
states where $\delta _{L}/T_{c}$ is larger than 1. Eq.(\ref{comparison})
implies that at $Z_{\mathrm{eff}} \gg 1$ the temperature dependence of the
total spectral weight at $T<T_{c}$ is almost entirely related with the
variation of the superconducting density:
\begin{equation*}
\delta K^{tot}=K^{tot}(T)-K^{tot}(T_{c})\approx \rho _{s}(T)
\end{equation*
which is thus described by Eq.(\ref{Q_0}).
\begin{figure}[tbp]
\includegraphics[width=8cm]{K-41b}
\caption{(Color online) Temperature dependence of regular spectral weight
K(T)$ and full spectral weight $K^{tot}(T)$ computed numerically (with 3D
Gaussian AM of lattice size $L=20$) in the region of pseudogap
superconductivity at $E_F=7.9$. We used Eqs.(\protect\ref{K_regular_below})
and (\protect\ref{K^total}), correspondingly, and employed approximate
expression for the gap function $\Delta(\protect\xi,T)$ in the factorized
form $\Delta(\protect\xi,T) = \Delta_c(\protect\xi)\protect\sqrt{1-T/T_c}$,
where $\Delta_c(\protect\xi)$ is the critical-point solution shown in Fig.
\protect\ref{DeltaTc}. Black stars stand for the simple averaging, whereas
blue circles correspond to typical averages. }
\label{FigSpectralWeight}
\end{figure}
We now compute $\rho _{s}(T)$ and $K^{tot}(T)$ as given by (\ref{K^total})
in the mean field approximation developed in this paper. We use modified MFA
equation for the gap function in the form (\ref{MMFA5}) and replace
x_{ij}^{2}$ under the sum by its average value $\frac{1}{2}R_{0}^{2}$
determined in Eq.(\ref{Zeff}) in (\ref{K^total}). The result is
\begin{equation}
\rho _{s}(T)\approx \frac{2\nu _{0}e^{2}R_{0}^{2}}{\hbar ^{2}
\int_{0}^{\infty }\frac{d\xi \Delta ^{2}(\xi )}{\sqrt{\xi ^{2}+\Delta
^{2}(\xi )}}\tanh \frac{\sqrt{\xi ^{2}+\Delta ^{2}(\xi )}}{T} \label{rhos2}
\end{equation
which can be further simplified at $T\rightarrow 0$:
\begin{equation}
\rho _{s}\approx \frac{2e^{2}R_{0}^{2}}{\hbar ^{2}}\int_{\Delta }^{\infty
\frac{\nu (\varepsilon )d\varepsilon }{\varepsilon }\hat{\Delta
^{2}(\varepsilon )\sim \frac{2\nu _{0}e^{2}R_{0}^{2}}{\hbar ^{2}}\Delta ^{2}
\label{rhos3}
\end{equation
This expression allows one to extract the effective interaction range $R_{0}
$ from the data on the superconducting density in the combination with the
gap $\Delta $ and DoS $\nu (\varepsilon )$ found from the Andreev
spectroscopy measurements (see section \ref{Andreev point contact}).
In the above discussion we assumed that effective number of neighbors $Z_
\mathrm{eff}}$ for the pseudospin model is large. It is not always the case
because, as discussed in the end of section~\ref{Transition temperature},
the superconducting state survives in the strongly spatially fluctuating
form in the regime $Z_{\mathrm{eff}}<1$. Qualitative picture of the spectral
weight temperature dependence in the pseudogap range of the Anderson model
formulated in section~\ref{Transition temperature} can be obtained from the
numerical evaluation of the regular part $K(T)$ given by (\re
{K_regular_below}) and the total weight $K^{tot}(T)$ given by (\ref{K^total
). The results are shown in Fig.\ref{FigSpectralWeight} for $E_{F}=7.9$. We
emphasize that the expressions (\ref{K^total},\ref{K_regular_below}) give
the leading orders in $1/Z_{\mathrm{eff}}$ expansion only, which is the
\textit{zero order} for Eq.(\ref{K^total}), and the \textit{first order} for
Eq.(\ref{K_regular_below}). As one sees in Fig.\ref{FigSpectralWeight}, both
$K(T)$ and $K^{tot}(T)$ change by a similar amounts, so in this example $Z_
\mathrm{eff}}\sim 1$. We expect that Fig.\ref{FigSpectralWeight} provides a
qualitatively good description of the spectral weight behavior in
pseudogapped superconductors.
\section{Summary of results and unsolved problems. \label{Summary of results
}
We presented a generalization of the BCS-like theory of superconductivity
that is appropriate for poor three-dimensional conductors in which Fermi
level is located in the vicinity of the mobility edge. The developed theory
neglects the effects of Coulomb repulsion. The reasons why this
approximation is appropriate to describe many disordered films were
discussed in section \ref{Experimental results}.
The model with BCS-like attraction and no Coulomb repulsion solved in the
bulk of the paper can be also realized by ultra-cold fermionic atomic gases
in optical traps~\cite{atoms,atoms2} with controlled disorder \cite{Aspect}.
The enhancement of $T_{c}$ by disorder predicted theoretically might turn
out to be a useful tool to observe BCS state in these systems because
without such enhancement $T_{c}$ is often too low to be reached
experimentally for the small values of the interaction constant $\lambda $.
Such experiments may be also important for the understanding of
superconductor-insulator transition in general.
The control parameter of the model is position of the Fermi level $E_{F}$
with respect to the mobility edge $E_{c}$. The major new (with respect to
the old works\cite{MaLee,KapitulnikKotliar1986,BulaSad}) ingredient of the
developed theory is the full account of the critical and weakly off-critical
wavefunction's fractality. We identified three qualitatively different
regimes: the hard-gap insulator realized at large disorder, the pseudogap
superconductor that appears when the phase correlations develop between well
localized Cooper pairs and the "fractal" superconductor that appears against
the background of fractal single particle states. Upon a further decrease of
disorder the fractal superconductor smoothly crosses over to a conventional
one.
For the insulating state, our main result is the computation of the
activation energy $T_{I}$ in the Arrhenius temperature dependence of
resistivity, $\ln R(T)\sim T_{I}/T$, at low temperatures. The hard-gap
character of this insulating state is due to the local attractive
electron-electron interaction which leads to formation of localized Cooper
pairs with nonzero binding energies $\Delta _{P}^{(j)}$, specific for each
j $-th localized orbital. The probability distribution of these energies is
similar to the distribution of inverse participation ratios for localized
eigenstates of the 3D Anderson model; it drops exponentially fast at low
energies, imitating a hard gap. The estimate for the activation energy
T_{I} $ as a function of $E_{F}-E_{c}$ is in a reasonable agreement with
experimental data~\cite{Shahar1992}.
The fractal superconducting state (realized for $E_{F}$ very close to $E_{c}
) is characterized by the following features:
\begin{enumerate}
\item {\ In the weak coupling limit, $\lambda \ll 1$, the transition
temperature, }${T}${$_{c}$, becomes a power law function of the interaction
constant, }$\lambda $. This leads to the{\ parametric enhancement of $T_{c}$
with respect to its value deep in the metal state. This conclusion is
different from previous works~\cite{MaLee,KapitulnikKotliar1986,BulaSad}
which assumed the validity of the \textquotedblleft Anderson
theorem\textquotedblright\ in the $E_{F}$ region very close to mobility
edge. The power law exponent in the $T_{c}(\lambda )\propto \lambda
^{3/(3-d_{2})}$ dependence is determined by the fractal dimension
d_{2}\approx 1.3$ of the critical eigenfunctions. }
\item {\ Strong local fluctuations of the pairing amplitude $\Delta (\mathbf
r})$ coexist with a unique and well-defined critical temperature $T_{c}$
below which a macroscopically coherent state appears.}
\item {\ The thermal Ginzburg parameter $\mathrm{Gi}$ of this transition is
a universal quantity of the order of unity which is independent of the
interaction constant $\lambda $, the same holds for the mesoscopic Ginzburg
number $\mathrm{Gi}_{d}$. This implies that thermal and mesoscopic
fluctuations may change the pre-factor in the power law $T_{c}(\lambda )$ by
a factor $\sim O(1)$ but cannot change the power-law itself.}
\item {\ Local single-particle density of states $\nu (\varepsilon ,\mathbf{
})$ fluctuates strongly in real space; these fluctuations lead to a strong
random asymmetry of the tunneling conductance, $G_{T}(V,\mathbf{r})\neq
G_{T}(-V,\mathbf{r})$. }
\item {\ The value of the superfluid response, Eq.(\ref{rhos01}), coincides
with result of ~\cite{MaLee} (Equation (4.1) in this paper), provided that
one inserts in it the correct values of $T_{c}$ and $\Delta _{0}$ given
above. }
\end{enumerate}
The pseudogap superconductor is predicted to occur when the Fermi energy
E_{F}$ is deep inside the localized band of single-particle states. Its
major feature is the presence of two distinct energy scales, both
originating from the Cooper attraction between weakly localized electrons:
the collective superconductive gap $\Delta (0)$ and the local binding energy
of a Cooper pair $\Delta _{P}$. The most unusual behavior is expected to
occur in the regime of a strong pseudo-gap $\Delta _{P}\gg (\Delta
(0),T_{c}) $. The very existence of superconducting correlations and a
nonzero $T_{c}$ in this regime is unexpected because it is characterized by
a typical level spacing $\delta _{L}=(\nu _{0}L_{loc}^{3})^{-1}$ which is
larger than $T_{c}$ and $\Delta (0)$: $\delta _{L}/T_{c}\propto (\Delta
_{P}/T_{c})^{3/d_{2}}$ (\ref{hard},\ref{D01}). The appearance of this regime
was not expected in previous studies\cit
{MaLee,KapitulnikKotliar1986,BulaSad} which concluded that the
superconducting state is stable only up to $\delta _{L}\leq T_{c}$. The
perseverance of superconducting coherence much deeper in the localized
region than was expected previously is the result of the enhancement of the
correlations between the wave-functions intensities \cite{CueKra}, which
occurs due to the Mott's mechanism of resonant mixing of localized states.
The key features of the pseudo-gap superconductor (in addition to the
features of fractal superconductor listed above) are the following: \newline
\begin{enumerate}
\item {\ Insulating behavior of the resistivity in a wide range of
temperatures above $T_{c}$. }
\item {\ Formation of almost hard gap without coherence peaks in the density
of states above $T_{c}$ with the gap value of the gap that fluctuates
significantly from point to point. }
\item {Growth of coherence peaks below a global $T_{c}$. The magnitude of
the coherence peaks (proportional to the local value of the superconducting
order parameter)\ fluctuates from point to point, these fluctuations become
very large close to the superconductor-insulator transition. }
\item {\ Two-peak feature in differential conductance at moderate
transmission probabilities as measured by the Andreev point-contact
spectroscopy below $T_{c}$. The lower peak voltage $V_{1}(T,\mathbf{r})$ is
expected to be $\mathbf{r}$ -independent, but vanishing as temperature
approaches $T_{c}$ from below. In contrast, the higher peak voltage $V_{2}(T
\mathbf{r})$ is expected to be almost $T$-independent, but strongly $\mathbf
r}$-dependent.}
\item {\ Strong temperature dependence of the total spectral weight
K^{tot}(T)$ of the high-frequency conductivity which shows with a loss of
major part of $K^{tot}(T)$ at temperatures $T\leq \Delta _{P}$, and its
re-appearance below $T_{c}$. The regular part $K(T)$ of the spectral weight
is smaller than the superconducting response $\rho _{s}$, see Eq.(\re
{comparison}). }
\end{enumerate}
We now briefly discuss possible extensions of the developed theory and open
questions. We begin with purely theoretical questions.
An obvious extension of the present theory would be a consistent account of
Coulomb interaction effects in fractal superconductors near the 3D mobility
edge. Here one should distinguish the effects of short range part of the
Coulomb repulsion that competes with the phonon attraction and the long
range part which might become very important in the insulator resulting in
the formation of the Cooper pair glass and qualitatively new physics.
Decreasing the dielectric constant in a real material would change it from
the fractal superconductor discussed in this paper to the material in which
superconductivity is suppressed by Coulomb interaction. Furthermore,
increasing only the short range part of the Coulomb interaction leads to the
'fermionic' mechanism while increase in the long range part leads to the
Coulomb blockade with completely different properties. The transition
between all these regimes are currently not understood, neither
theoretically nor experimentally. It seems likely that at least one of these
crossovers was observed as the 'region of poor scaling' in
\cite{Steiner2008}.
The character of the quantum critical point that separates the fractal
superconductivity and the insulator at $T=0$ deserves a further study. As
discussed in section \ref{Superconductivity with a pseudogap} on both sides
of this transition electrons are bound in the localized pairs, and the
transition itself consists in development of phase correlations, well
described in terms of the XY ordering of the Anderson pseudo-spins $\mathbf{
}_{j}$. This scenario is different from both "bosonic" and the "fermionic"
mechanisms of superconductor-insulator transition, so we suggest the name
"pseudo-spin" mechanism. The cavity approach to the study of such
transitions was developed recently in \cite{IoffeMezard2009}; the main
qualitative conclusion of this study is self-organized inhomogeneity of the
resulting superconductor in which phase correlations are dominated by the
rare, almost one-dimensional paths. This conclusion is supported by the
results of the virial expansion method shown in Fig. \ref{TcDistribution}
that point out to a percolation-like transition between the pseudo-gap
superconductor and an insulator. Another feature of this solution is a very
rapid decrease in the transition temperature beyond certain value of
disorder which might be seen experimentally as the apparent existence of the
lowest nonzero $T_{c}\approx 0.5\mathrm{K}$ found in amorphous InO$_{x}$
system~\cite{Sacepe2007}.
We now turn to experimental findings that lack (partially or entirely)
theoretical explanation.
As discussed in section \ref{Point contact tunneling} the quantitative
description of point contact tunneling requires the theory of collective
modes in the fractal superconducting state. In the absence of such theory
the spectacular, nearly rectangular shape of the \textit{local} tunneling
conductance $G(V)$ at low temperatures\cite{Sacepe2007} cannot be explained
quantitatively.
The theory of the hard-gap insulating state presented in section \re
{Insulating state} neglects the transport by Cooper pairs and takes into
account only single electron transport. It is therefore limited to the
region relatively far from the transition such as studied in Ref.\cit
{Shahar1992}. Theory of the incoherent transport by Cooper pairs close to
the transition was discussed in the recent papers \cit
{IoffeMezard2009,Muller2009}, the main conclusion of these studies is that
very close to the transition the behavior of Cooper pairs is controlled by
the many body mobility edge which becomes zero exactly at the transition.
Away from the transition the theory predicts activated behavior at the
lowest temperatures with the gap that increases fast away with disorder and
becomes infinite a short distance away from the transition signalling the
absence of incoherent transport by Cooper pairs~\cite{IoffeMezard2009}.
Recently the experimental paper~\cite{KowalOvadyahu2} reported the anomalous
size-dependence of the superconductor-insulator transition for extremely
wide range of sizes between 1 and 150 microns. In particular, the workers
observed a slow but significant dependence of the activation energy
T_{0}(L) $ on the system size $L$ (the distance between the metal contacts)
on the insulating side of the transition. In one case $T_{0}(L)$ decreased
from 13.5 K to 9.6 K while $L$ was changed from 145 to 12 microns at a fixed
width. Although the theory of the the incoherent transport by Cooper pairs
close to the transition \cite{IoffeMezard2009} predicts a strong size
dependence at mesoscopic scales due to inhomogeneities, the size dependence
at the huge scales observed experimentally are very difficult to explain.
The explanation of this effect seems currently to be beyond the reach of a
theory and presents a real challenge.
A different set of challenges is presented by the superconductor-insulator
transition induced by magnetic field for strongly disordered (almost
insulating) superconducting samples. As one would expect, the critical
field, $H_{c}$ (which corresponds to superconductor-insulator transition at
T=0$) is a strong function of disorder, for instance, it varies over two
orders of magnitude in the experiments\cite{Steiner2005}. Surprisingly,
there are indications that phenomenology of the transitions driven by
disorder and by magnetic field look differently~ \cit
{Shahar1992,Gantmakher1998,Shahar2004,Baturina2007a,Steiner2008,SacepeShahar09
. Namely, the critical point in disorder is characterized by the activated
behavior of the resistance, $R\sim \exp (T_{0}/T)$, with a large gap $T_{0}$
even very close to the superconductor-insulator transition. In contrast,
activated behavior, $R\sim \exp (T_{I}(H)/T)$, observed at fields above
critical $H_{c}$ is characterized by a small $T_{I}(H)$ which extrapolates
to zero at $H_{c}$ . In other words, the region in ($H,T$) plane where
transport is characterized by activated behavior with a small gap has a
peculiar wedge like shape as sketched in Figure \ref{PhaseDiagram}).
Generally, one expects that very close to the critical line the transport is
dominated by Cooper pairs; this expectation was experimentally confirmed for
ultrathin Bi films ~\cite{Valles2e}. In this case the experimental phase
diagram implies that the regime of Cooper pair dominated transport is narrow
in disorder but wide in field. Qualitatively this is likely to be due to the
fact that the effect of a small magnetic field is limited to the generation
of local phase differences which have a large effect of the superconducting
state but does not affect the incoherent transport of the pairs. However,
there is currently no consistent theory of this effect. On the experimental
side, it would be important to verify that the transport in this regime is
due to Cooper pairs for the InO and TiN films.
Furthermore, strongly disordered superconductors show puzzling behavior in
very large magnetic fields: their resistivity become temperature independent
and approaches the quantum limit, $h/e^{2}$. The field scale, $H_{P}$, at
which this behavior sets in is not sensitive to the disorder in contrast to
the critical field, $H_{c}.$ The second field scale appears first in the
fractal superconductor: it is the field that destroys local Cooper pairs and
suppresses the hard gap $\Delta _{P}$. In this regime, the parametric
difference between $H_{P}$ and $H_{c}$ is due to the low fractal dimension
d_{2}\approx 1.3$ of a single-particle wave-function that suppresses the
orbital effect of magnetic field on the pairing of two electrons localized
on this eigenstate, similarly to a very thin wire. In contrast, long-range
coherence requires the existence of large loops involving many localized
Cooper pairs. The typical size of these loops is larger than the
localization length $L_{loc}$,
leading to relatively small magnetic field scale $H_{c}\leq \Phi
_{0}/L_{loc}^{2}\ll H_{P}$. In the pseudogap superconductor the second field
scale is associated with the suppression of a large gap, $\Delta _{P}$,
responsible for Cooper pair formation whereas the critical field is
associated with the frustration of a weak pseudospin coupling. Thus, one
expects even larger difference between the field scales in this regime.
However, the universal value of the conductance in the regime $B\gg H_{P}$
is not expected theoretically and remains mysterious.
Nernst effect can be potentially a very sensitive probe of the nature of the
superconducting state.\cite{OngNernst} These experiments were performed
recently on InO$_{x}$ samples and show~\cite{Spathis} Nernst signal which
scales as $N(T)\propto T^{-n}$ , with the "Nernst exponent" $n\approx 7.5$
at low magnetic fields. Application of the conventional theory \cit
{LarkinVarlamovBook} of the superconducting fluctuations to the Nernst
effect in 2D superconductors~\cite{Serbin,Karen} gives Nernst signal that
scales as $1/T\ln T$ at high temperatures. This behavior was indeed observed
in conventional superconductor NbSi~\cite{AubinNernst}. Because Nernst
effect requires a motion of the electrical charge around the plaquette, the
temperature dependence is expected to be different in discrete systems where
motion around the minimal plaquette involve a large number of hops and each
hop implies an extra power of $1/T$ in the high temperature expansion~\cit
{Podolsky07}. For instance, because minimal plaquette on the hexagonal
lattice contains six sites, one expects that $N(T)\propto T^{-6}$ in this
case. The striking difference between NbSi and InO$_{x}$ behavior observed in~\cit
{Spathis} on very low $T_{c}$ (very disordered) samples indicates, in our
opinion, the importance of Cooper pair hoping between localized sites. This
behavior should get less pronounced further away from the transition, thus
we expect that slightly less disordered samples will show more
'conventional' exponent in Nernst effect.
A number of papers noted the apparent similarity between phenomenology of
disordered films of conventional superconductors discussed in this paper
and that of high $T_c$ oxides.\cite{Steiner2005,Steiner2004}
Very briefly, the transport and magnetic measurements show the formation
of pseudogap at temperatures much higher than superconducting $T_c$ in
underdoped materials,\cite{TimuskReview}
while STM measurements display highly inhomogeneous
order parameter (as measured by coherence peaks) combined with modestly
homogeneous tunneling gap.\cite{Kapitulnik2001,Davis2001,Davis2002,Yazdani2007}
Furthemore, a number of indirect evidences points out to the superconducting
nature of the pseudogap in these materials \cite{Corson1999,Wang2006,Yuli2009},
similar to the situation discussed in this paper.
The crucial difference between s-wave and d-wave pairing characterizing high
$T_c$ oxides is that, in contrast to s-wave pairing, modest elastic scattering
with mean free path of the order of superconducting coherence length suppresses
the d-wave superconductivity. Thus, localization of the wave functions is
incompatible with generic d-wave superconductivity. This conclusion might need
to be revised for the special case of superconductivity which is due to pairing
of electrons in a small area of the full Brillouin zone which might be the case
of cuprates \cite{Geshkenbein1997,Galitski2009}. In this case, the electrons
that are responsible for the pairing belong to two well separated patches on
the Fermi surface with each patch characterized by a small momentum $p_0$. The
density of states in these patches is large which makes scattering of these
electrons strong. Furthermore, because most impurities are located far from the
conducting copper oxide planes, it is likely that the elastic scattering, though
strong, does not mix different patches and thus does not inhibit d-wave pairing.
In this situation, single electron states inside each patch may become very similar
to the localized states discussed in this work making the developed theory
qualitatively correct. This would explain the main phenomenological features
mentioned above: formation of pseudogap far above $T_c$, highly inhomogeneous order
parameter observed in tunneling data, insulating behavior of LaSrCuO superconductors
in high magnetic fields, etc.
Finally, it is not clear why the phenomenology displayed by Josephson
junction networks\cite{FazioZant,Serret2002} was not observed in any
disordered film. In particular, Josephson junction networks close to
superconductor-insulator transition are characterized by the appearance of a
wide region of magnetic fields where resistivity has no temperature
dependence but varies by many orders of magnitude as a function of magnetic
field. This is in a sharp contrast with the films that show more or less
good crossing point of $R(B)$ isotherms. There is very little doubt that
Josephson networks are described by model Hamiltonian (\ref{H_JJ}), so one
has to conclude that it is not the appropriate model for most films. One
possible source of difference discussed above might be that many films are
characterized by a large value of the dielectric constant, while another
reason, relevant for the films with small dielectric constant, might be the
capacitance matrix is not dominated by nearest neighbors in these films.
Nearest neighbor capacitance matrix translates into the Coulomb interaction
that decreases only logarithmically with distance in the Josephson network
as opposed to $1/r$ expected in a film. The physical result might
be the formation of the Cooper pair glass in the Josephson network but not
in the disordered film or vice versa. If true, one expects that Josephson
networks with the Coulomb interaction screened by the ground plate might
display properties more similar to those of the films. Another difference
might be due to the fluxes produced by magnetic field are completely random
in a disordered films but are (almost) identical on the plaquettes of the
Josephson lattice. In this case one expects to see film behavior in
randomized Josephson network.
\bigskip We are grateful to B. L. Altshuler, T. I. Baturina, C. Chapelier,
T. Dubouchet, A. M. Finkelstein, V. F. Gantmakher, A. S. Ioselevich, I. V.
Lerner, A. Millis, A. D. Mirlin, M. Mueller, Z. Ovadyahu, V. V. Ryazanov, B.
Sacepe, V. Schmidt, D. Shahar, M. A. Skvortsov and V. I. Yudson for useful
discussions. This research was supported by Triangle de la physique 2007-36,
grants ANR-06-BLAN-0218, ARO 56446-PH-QC and DARPA HR0011-09-1-0009, by the
grant 07-02-00310 from Russian Foundation for Basic Research, and by the
program "Quantum physics of condensed matter" of Russian Academy of
Sciences. We thank the FEDER and the Spanish DGI for financial support
through Project No. FIS2007-62238. V.E.K acknowledges a hospitality of the
Institute for Nuclear Theory at the University of Washington and the DOE for
partial support during the completion of this work.
|
\section{Introduction}
At a rough conceptual level, living cells can be seen as devices that
convey the free energy derived from the breakdown of nutrients (mostly
sugars) into the chemical energy that fuels the production of all the
molecules required for survival and, when possible, growth and
reproduction. The complex intracellular machinery that underlies the
energy transduction is being increasingly unveiled at both the
biochemical (reactions) and the regulatory level (enzymes, their
corresponding genes, the genes' transcription factors, etc) through
the massive genomic information available for different organisms
\cite{gold}. In particular, the network of biochemical processes
collectively known as `metabolism' by which the nutrients are degraded
and the housekeeping molecules are manufactured has been mapped in
great detail for many bacteria and several eukaryotes
\cite{pals}. While the study of the structure and dynamics of single
metabolic pathways has a long standing history in biophysical
chemistry \cite{hs,beard}, the current wealth of data allows to
analyze the behavior of cellular reaction networks at the scale of the
whole genome. This is a crucial step both to shed light on the
emergence of a metabolic phenotype from the underlying genotype and to
formulate testable \cite{sau} predictions on a cell's adaptation and
response to perturbations (the key to multiple biotechnological
applications).
Ultimately, metabolic activity and capabilities are determined (or
limited) by various constraints of chemical (stoichiometric),
thermodynamic and regulatory origin. The uncertainty about their
details increases considerably as one passes from the stoichiometric
to the regulatory level. Current theoretical approaches therefore try
to infer the global organization of metabolism from simple schemes
that implement explicitly only the best known restrictions, making as
few assumptions as possible on the rest. Even so, it turns out that
some predictive and explanatory power on a cell's biochemical
functioning can be achieved.
We begin by presenting a short selective review of stoichiometric
models of metabolic networks as constraint-based systems. We then
focus on one specific problem, that of metabolite producibility,
characterizing the global metabolic output for the bacterium {\it
E.coli} in a specified growth medium. While a robust and
biologically significant production profile emerges, fluctuations in
nutrient usage or in the production level of key metabolites
occur. Such fluctuations can be associated to different patterns of
metabolic pathway activation. Finally we discuss some of the issues
where statistical mechanics tools may have a significant impact in the
near future. We shall limit biochemical details to a minimum and treat
the reaction system as a standard input-output network. The interested
reader is referred to e.g. \cite{hs} for a thorough introduction to
cellular metabolism.
\section{Flux analysis: a bird's eye view}
In essence (see Fig. \ref{cell}), a cell's metabolic network can be
seen as a set of interconnected chemical reactions coupled with a set
of transport processes.
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{cell}
\caption{\label{cell}Basic scheme of a cellular metabolic network. Ext
and Int denote respectively the exterior and the interior of the
cell. Reactions (resp. metabolites) are denoted by circles
(resp. squares). S1 and S2 are auxiliary fluxes supplying nutrients
n1 and n2 to the environment. T1, T2 and T3 denote membrane
transport reactions by which metabolites are taken in or expelled
from the cell. R is an intracellular reaction. A cellular
compartment Comp is also shown, together with the corresponding
transport reactions connecting its interior with the cytosol.}
\end{center}
\end{figure}
For bacteria, this includes the membrane transport mechanisms by which
nutrients are brought into the cell, and the intracellular reactions
by which they are degraded and new biochemical species are
produced. In organisms with compartmentalized cells (e.g. eukaryotes)
one should also account for the transport of metabolites into and out
of each compartment, i.e. for the cell's actual geometric
structure. The basic and most reliable information on the, say, $N$
reactions involving $M$ chemical species is encoded in the $M\times N$
{\it stoichiometric matrix} $\mathbf{\Xi}$, whose entry $\xi_i^\mu$
represents the stoichiometric coefficient with which species $\mu$
participates in reaction $i$. $\mathbf{\Xi}$ is a sparse, integer
matrix and a sign convention is usually adopted to discern products
($\xi_i^\mu>0$) from substrates ($\xi_i^\mu<0$). The stoichiometric
matrix often also contains a motivated assumption on reaction
directionality to account for the fact that, while all reactions are
in principle reversible, under physiological conditions some of them
may occur in one direction only. We shall treat physiologically
reversible reactions as two separate processes. In addition, we shall
always assume that $\mathbf{\Xi}$ also includes external supply fluxes
for the nutrients, i.e. auxiliary (free or fixed) reactions that
provide the nutrient to the environment. To have an idea, in such a
setting for the bacterium {\it E.coli} $N\simeq 1100$ and $M\simeq
700$, whereas $N\simeq 1500$ and $M\simeq 900$ for the unicellular
eukaryote {\it S.cerevisi\ae}.
If we denote by $\boldsymbol{\nu}\geq \boldsymbol{0}$ the
$N$-dimensional vector of reaction fluxes (the sign constraint arising
from our treatment of reversibility), the time evolution of the vector
$\mathbf{c}$ of metabolite concentrations is described simply by
$\dot{\mathbf{c}}=\mathbf{\Xi}\boldsymbol{\nu}$. Fluxes depend in
principle on the membrane transport mechanisms or on the enzyme
kinetics through various parameters like the rate constants
$\mathbf{k}$ and $\mathbf{c}$ itself and in turn, through them, on
temperature, activation energy, etc. If such details are known, then
functional dependencies
$\boldsymbol{\nu}=\boldsymbol{\nu}(\mathbf{c},\mathbf{k},\ldots)$ can
be chosen according to the corresponding kinetics
(e.g. Michaelis-Menten or Hill) and one can in principle solve the
dynamical system for the concentrations, provided all underlying
constants are known. Unfortunately, this is rarely the case in
genome-scale networks (see \cite{dyn} for a positive example). While
methods have been developed to partially overcome this limitation and
provide dynamical characterizations of metabolic activity (see
e.g. \cite{dfba,skm}), the standard modeling route assumes, on the
basis of the timescale separation between chemical processes and
genetic regulation, that metabolic networks operate in a non
equilibrium steady state where $\dot{\mathbf{c}}=\mathbf{0}$. Note
that flux vectors satisfying $\mathbf{\Xi}\boldsymbol{\nu}=\mathbf{0}$
describe states in which intracellular metabolites obey Kirchhoff-type
mass-balance conditions. If $N>M$, as it usually occurs in real
metabolic networks, such vectors form a set of dimension
$N-\text{rank}(\mathbf{\Xi})$ embedded in $\mathbb{R}_0^+$ that, in
absence of additional constraints, contains equivalent feasible flux
states of the network. In principle, this set should be explored
uniformly to extract the relevant biological
information. Unfortunately, Monte Carlo sampling becomes unaffordable
as soon as the dimension of the solution space exceeds a few tens (see
\cite{mc} for a feasible case; message-passing algorithms have been
recently proposed as a working alternative for microbial cells
\cite{mul}). In many cases it is however possible to select relevant
configurations by imposing the maximization of an objective function
normally represented as a linear combination of the fluxes with given
coefficients $\boldsymbol{\alpha}$,
i.e. $\boldsymbol{\alpha}\cdot\boldsymbol{\nu}$. This reduces the
problem to a linear optimization one, namely
\begin{equation}
\max_{\boldsymbol{0}\leq\boldsymbol{\nu}\leq\boldsymbol{\nu}_{max}} (\boldsymbol{\alpha}\cdot\boldsymbol{\nu})~~~\text{subject to}~~~\mathbf{\Xi}\boldsymbol{\nu}=\mathbf{0}
\end{equation}
The upper bounds on fluxes can be chosen to either model specific
extracellular conditions (e.g. a certain nutrient's supply or intake
cannot exceed a fixed level), to simulate known physiological
limitations (e.g. the rate of a certain reaction cannot exceed a fixed
level because of limited enzyme availability), or simply to impose
well-defined bounds on the solution space. This scenario is the basis
of various stoichiometry-based approaches such as Flux-Balance
Analysis (FBA) \cite{kau}. Clearly, the vector $\boldsymbol{\alpha}$
contains the crucial biological assumptions. In some cases, it is
grounded in experimental evidence. For instance, bacteria in rich
environments appear to evolve, under selective pressure, to optimize
their growth rate and reproduce as fast as possible \cite{ib}. A good
proxy for a bacterium's growth capability lies in its ability to
generate {\it biomass}, a combination of different metabolites in
precise stoichiometric proportions (including\footnote{See
\cite{biom2} for a list of metabolite abbreviations.} the 20
proteinogenic amino acids, the molecular energy carrier ATP, water,
key cofactors like nad, nadp and coenzyme-A etc.) that is used to
produce ADP, inorganic phosphate (PI), pyrophosphate (PPI) and water
\cite{biom,biom2}. This reaction represents the cell's use of
metabolic products in macromolecular processes (like building
proteins, membranes, etc.) that are not accounted for in
$\mathbf{\Xi}$. The maximization of biomass yield is in these cases a
widely used criterion implemented by adding the auxiliary biomass
reaction to $\mathbf{\Xi}$. In other cases, the choice may have
physiological justifications. The need for energetically-efficient
housekeeping e.g. in nutrient-limited conditions suggests that cells
optimize the ATP yield, which requires to maximize the total flux of
all ATP-producing reactions \cite{rama}. Other examples include the
minimization of glucose consumption (a proxy for efficient nutrient
usage \cite{oli}) or of the total flux of intracellular reactions (for
maximal enzymatic efficiency \cite{bona}). We refer the reader to
\cite{sys} for a comparison of the performance of several different
objective functions for predicting the fluxes of the main carbon
pathways in {\it E.coli} and to \cite{sri} for a study of objective
function selection criteria.
This kind of approach can be extended to biologically more complicated
situations like those induced by gene knock-outs that prevent the
execution of certain reactions. Based on experimental evidences, such
scenarios lead to consider different selection criteria, like the
minimization of the overall flux rearrangements with respect to the
wild type (by quadratic programming \cite{segre}) or the minimization
of rearrangements in the large-flux backbone with respect to the wild
type (by mixed linear-integer programming \cite{room}). The former
appears to be suited to capture the transient sub-optimal growth
states that a bacterium takes on immediately after the perturbation,
whereas the latter provides a better description of the states of fast
growth that the knock-out organism reaches on longer time scales under
selective pressure. However the dynamical structure of the response of
metabolic networks to perturbations is far from being satisfactorily
understood.
\section{Producibility and Von Neumann's problem}
An important problem related to flux analysis concerns the link
between the network's structure and its productive
capabilities. Following \cite{im}, a metabolite $\mu$ is said to be
{\it producible} from a given set of nutrients (to be specified) if
\begin{equation}\label{pr}
\exists~\boldsymbol{\nu}\geq\mathbf{0}~\text{such that}~~\mathbf{\Xi}\boldsymbol{\nu}\geq 0~\&~\left[\mathbf{\Xi}\boldsymbol{\nu}\right]_\mu>0
\end{equation}
(where
$\left[\mathbf{\Xi}\boldsymbol{\nu}\right]_\mu=\sum_{i=1}^N\nu_i\xi_i^\mu$)
that is, if at least one flux vector exists allowing for a net
production of $\mu$ irrespective of whether other metabolites are
being also produced (nutrient usage may evidently never exceed its
supply). Producible metabolites have the property that their
concentrations can increase in a stationary flux state with the sole
consumption of the nutrients, so that the cell is allowed to employ
them for purposes other than metabolic (e.g. to form proteins,
membranes, etc.). One would therefore expect that survival in a given
environment is given by the ability to produce the necessary
metabolites be they biomass components or else.
It is easy to understand that the possibility to actually produce a
metabolite may be limited by the emergence of conservation laws from
the stoichiometry. To clarify, note that $M$-dimensional vectors
$\mathbf{z}\geq \mathbf{0}$ satisfying
$\mathbf{z}^T\mathbf{\Xi}=\mathbf{0}$ (or, positive semidefinite
vectors from the left null-space of the stoichiometric matrix) define
linear combinations of metabolites such that the corresponding
weighted sums of their concentrations are constant over time
\cite{fam}. Such conserved moieties are abundant in real metabolic
networks \cite{moy}. Clearly, an accumulation of metabolites belonging
to one such pool, and hence their producibility, is ruled out by
simple stoichiometric reasons. The duality of producibility and
conservation can be exploited to identify (by linear programming) the
growth media allowing for the production of specified sets of
metabolites such as the biomass or slight modifications of it
\cite{belta}\footnote{An interesting extension of producibility is
{\it sustainability}, loosely defined (see \cite{oe}) as the
property of being producible using producible metabolites as
substrates besides nutrients.}.
In order to evaluate the robustness of the cellular production profile
(if any) emerging in given nutrient conditions one should study the
set
\begin{equation}\label{set}
\mathcal{V}=\{ \boldsymbol{\nu}~\text{such that}~\mathbf{\Xi}\boldsymbol{\nu}\geq 0 \}
\end{equation}
(we shall henceforth assume that $\sum_i\nu_i=N$). Statistical
sampling in this case turns out to be feasible, a possible route being
found in the work of J. Von Neumann. In order to present a rather
simplified version of his model \cite{jvn}, let us re-define the
stoichiometric matrix by separating the matrix $\mathbf{B}$ of input
coefficients from the matrix $\mathbf{A}$ of output coefficients, so
that $\mathbf{A}-\mathbf{B}=\mathbf{\Xi}$. It is simple to see
\cite{dm} that, given a constant $\rho>0$, a flux vector
$\boldsymbol{\nu}$ such that $\mathbf{A}\boldsymbol{\nu}\geq \rho
\mathbf{B}\boldsymbol{\nu}$ describes a network state in which every
species is being produced at a rate at least equal to $\rho$. In a
generic input-output system specified by matrices $\mathbf{A}$ and
$\mathbf{B}$, one can therefore ask what is the maximum value of
$\rho$ for which flux vectors satisfying
\begin{equation}\label{vn}
(\mathbf{A}-\rho\mathbf{B})\boldsymbol{\nu}\geq \mathbf{0}
\end{equation}
exist. This value (denoted here as $\rho^\star$) represents the
optimal productive performance allowed by the ways in which the
available processes combine the metabolites. Depending on whether
$\rho^\star$ is larger or smaller than 1, the system may have optimal
states that are expanding or contracting. Correspondingly, the flux
vector(s) satisfying the above set of constraints for
$\rho=\rho^\star$ are the optimal flux states of the system. When
$N,M\to\infty$ and $\mathbf{A}$ and $\mathbf{B}$ are taken to be
semi-positive but otherwise unstructured random matrices, one can
dissect the phase structure of the problem in detail by statistical
mechanics tools, specifically using the replica trick (in a fully
connected network where each reaction consumes and produces a finite
fraction of the possible metabolites \cite{dm}) or the cavity method
(in a finitely connected network where reactions use a finite number
of substrates to generate a finite number of products \cite{ran}). The
general lesson is that a critical value $n_c$ of the structural
parameter $n=N/M$ separates contracting from expanding regimes, with
enhanced dilution increasing (resp. decreasing) $\rho^\star$ in the
expanding (resp. contracting) phase. Moreover, in such cases a single
flux vector satisfies conditions (\ref{vn}) when $\rho=\rho^\star$. In
the biologically more sensible case in which $\mathbf{A}$ and
$\mathbf{B}$ are real stoichiometric matrices, the situation is
radically different. The mass balance conditions imposed by
stoichiometry or, equivalently, the existence of conserved metabolic
pools, imply $\rho_\star=1$ (and force a finite volume of solution in
a random model with realistically structured input and output matrices
\cite{epl}), so that the optimal solution space of (\ref{vn}) indeed
coincides with (\ref{set}). Exploiting the presence of $\rho$ it is
however possible to define a straightforward iterative algorithm that
samples $\mathcal{V}$ uniformly \cite{ran}. This allows for a complete
and extensive characterization of producible metabolites and of the
corresponding flux states in real cellular networks in a given
environment at a modest computational cost even for genome-scale
systems.
\section{Application to E.coli: production profiles and their fluctuations}
The volume (\ref{set}) generated by the stoichiometric matrix of the
bacterium {\it E.coli} has been studied in \cite{pnas}, revealing that
\begin{enumerate}
\item[(a)] the predicted ranges of variability of the fluxes in
$\mathcal{V}$ in specified extracellular conditions agree well with
the (limited) experimental data available on the reaction rates
inferred from $^{13}C$-based experiments in a similar nutrient
profile \cite{emm,sau};
\item[(b)] dynamically stiff variables, i.e. reactions with smaller
allowed ranges, tend to correspond, via the associated enzyme, to
{\it E.coli}'s phenomenologically essential genes, i.e. genes that
are both necessary for the organism's survival and highly conserved
across different bacterial species \cite{gerdes}.
\end{enumerate}
These results suggest that at least in some conditions metabolic
networks may operate close to their optimal productive capacity and
that the ``shape'' of $\mathcal{V}$ may contain useful information
relating the metabolic phenotype to the underlying genotype. The
natural question to ask now is whether a robust metabolite production
profile emerges and how it correlates with the physiologically defined
biomass (in suitable growth media). We therefore take a closer look at
the set of produced metabolites for {\it E.coli} in a minimal growth
medium, an environment with a tunable supply of a limited set of
nutrients formed by inorganic phosphate (PI), O$_2$, SO$_4$, CO$_2$,
K, NH$_3$ and a carbon source (glucose in this case) \cite{almaas}
that resembles closely the M9 medium widely used in the experimental
literature \cite{emm}. This network consists of 1057 reactions
involving 631 metabolites altogether (after pruning the trivial
producibility constraints, see \cite{pnas} for details).
For simplicity, we henceforth set
$\boldsymbol{\lambda}=\mathbf{\Xi}\boldsymbol{\nu}$ and label the flux
vectors from $\mathcal{V}$ as $\boldsymbol{\nu}_\alpha$,
$\boldsymbol{\nu}_\beta$, etc. assuming to have sampled $S$ such
configurations. We shall furthermore write
$\boldsymbol{\lambda}_\alpha=\mathbf{\Xi}\boldsymbol{\nu}_\alpha$.
The simplest way to check whether $\mathcal{V}$ defines a set of
consistently produced metabolites is by studying the empirical
correlation matrix $\mathbf{C}$ given by
\begin{equation}\label{C}
C^{\mu\mu'}=\frac{1}{S}\sum_{\alpha=1}^S \lambda^\mu_\alpha \lambda^{\mu'}_\alpha
\end{equation}
In the spirit of principal component analysis, $\mathbf{C}$'s
eigenvalue spectrum contains much information on the collective
properties of the system. In particular, it is possible to express the
underlying ``signals'' $\{\boldsymbol{\lambda}_\alpha\}$ as weighted
sums of the eigenvectors of $\mathbf{C}$. The presence of a large,
isolated eigenvalue $r_{max}$ indicates that $\lambda^\mu_\alpha\simeq
\sqrt{r_{max}} ~V_{max}^\mu \eta_\alpha$, where $\mathbf{V}_{max}$ is
the eigenvector corresponding to $r_{max}$ and $\eta_\alpha$ is a
unit-variance random variable. Hence $r_{max}$ and its corresponding
eigenvector may offer an effective, zero-order description of
$\mathcal{V}$ in terms of the emergent production profile. The
eigenvalue distribution for {\it E.coli} is shown in
Fig. \ref{bla}\footnote{To remedy the uncertainty about relative
fluctuations when $\lambda^\mu\ll 1$, one can set a threshold
$\epsilon$ below which $\lambda^\mu$ is ``effectively'' zero. For
the present case, values of $\epsilon\leq 0.01$ have been found to
provide the same qualitative picture.}.
\begin{figure}[!]
\begin{center}
\includegraphics[width=9cm]{eig-dis.jpg}
\includegraphics[width=13cm]{largesteig}
\caption{\label{bla}(Top) Histogram of the eigenvalues of the
empirical correlation matrix $\mathbf{C}$, Eq. (\ref{C}), for {\it
E.coli} in a minimal growth medium (see text). Inset: detail of
the ``network mode''. (Bottom) Components of the eigenvector of
$\mathbf{C}$ corresponding to the network mode, with some of the
producible metabolites explicitly indicated (see \cite{biom2} for a
list of metabolite abbreviations).}
\end{center}
\end{figure}
One clearly sees a giant eigenvalue separated by a roughly continuous
spectrum, well described by the Marcenko-Pastur law with a point mass
at $r=0$ \cite{bouc}. In metabolic terms, $r_{max}$ represents a
``network mode'', a production profile that is common (to a first
approximation) to all $\boldsymbol{\nu}\in\mathcal{V}$. The components
of the eigenvector corresponding to the network mode (Fig. \ref{bla},
bottom) point explicitly to some of the metabolites that are expected
to form the core of the metabolic production. These turn out to
include\footnote{See \cite{biom2} for a list of metabolite
abbreviations.}
\begin{itemize}
\item biomass components (e.g. 5mthf, atp, fad, h2o, nad, nadh, nadp, nadph)
\item biomass products (adp, h, pi, ppi)
\item end points of metabolic pathways (waste, e.g. ac, co2)
\end{itemize}
Such a profile is largely recovered by studying the average of
$\lambda^\mu$ over 500 flux configurations
$\boldsymbol{\nu}\in\mathcal{V}$ for each $\mu$, see Fig. \ref{uno}.
\begin{figure}
\begin{center}
\includegraphics[width=13cm]{meanprod}
\caption{\label{uno}Value of $\lambda^\mu$ averaged over 500 flux
states $\boldsymbol{\nu}\in\mathcal{V}$ versus $\mu$ (metabolite,
horizontal axis) for {\it E.coli} (631 metabolites) in a minimal
growth medium (see text). Arrows mark, from left to right, the
following metabolites (see \cite{biom2} for a list of metabolite
abbreviations): 5mthf, ac, ade, adp, amp, atp, co2, coa, fad, for,
h, h2o, hco3, k, leu-L, the nad group (nad, nadh, nadp, nadph), nh4,
pi, ppa, ppi, so4, succ and ura.}
\end{center}
\end{figure}
Focusing for clarity on the ones with the largest average
producibility we find multiple biomass components (5mthf, amp, atp,
coa, fad, h2o, leu-L, nad, nadh, nadp, nadph, ppi) and biomass
products (adp, h, pi and ppi), supporting a growth scenario; typical
waste products of metabolic activity (ac and co2, end point of
oxidative phosphorylation); and metabolites with clear metabolic roles
like formate (central for h production), hco3 and k (helping to
maintain a stable intracellular condition), succ (succinate, a key
node of the Krebs cycle). In addition to these, a host of producible
species appear, though with smaller $\lambda^\mu$, including all amino
acids (with the exception of leucine).
For most compounds, however, one finds
$[\mathbf{\Xi}\boldsymbol{\nu}]_\mu=0$ for each
$\boldsymbol{\nu}\in\mathcal{V}$. As said above,
$[\mathbf{\Xi}\boldsymbol{\nu}]_\mu=0$ if $\mu$ belongs to a conserved
metabolic pool. Interestingly, such moieties do not seem to exhaust
the list of unproducible metabolites, as metabolite producibility
turns out to be a significantly fluctuating property. To evaluate how
the production profile of each metabolite varies across solutions in
$\mathcal{V}$ one can calculate, for each $\mu$ and for each pair of
(distinct) flux vectors
$\boldsymbol{\nu}_\alpha,\boldsymbol{\nu}_\beta\in\mathcal{V}$, the
overlap
\begin{equation}
q^\mu_{\alpha\beta}=\frac{2 \lambda^\mu_\alpha \lambda^\mu_\beta}{(\lambda^\mu_\alpha)^2+(\lambda^\mu_\beta)^2}
\end{equation}
When the above quantity is averaged over pairs $\alpha\neq\beta$
(properly accounting for the case where at least one
$\lambda_\alpha^\mu\ll 1$), one obtains an index (we shall denote it
by $q^\mu$) that is closer to $1$ the smaller are the fluctuations in
$\lambda^\mu$ for metabolite $\mu$. The distribution of $q^\mu$ is
shown in Fig. \ref{disqmu}.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{untitled.jpg}
\caption{\label{disqmu}Distribution of $q^\mu=\langle
q^\mu_{\alpha\beta}\rangle$ (average over pairs of distinct
solutions from $\mathcal{V}$). The index e (extracellular) denotes
metabolites that also serve as nutrients. Arrow mark groups of
metabolites (as shown) falling in a certain range of $q$.}
\end{center}
\end{figure}
One sees that the production rate of many key metabolites including
co2, atp, amp, the nad group, etc. may oscillate considerably from
solution to solution in $\mathcal{V}$. This holds for intracellular
species (e.g. ATP or biomass components) as well as for extracellular
ones, so that even the amount of nutrients necessary for the cell to
maintain its productive capability in an environment where a specific
group of metabolites is present may have a large allowed range.
It is instructive to inspect two particular solutions out of those
sampled from $\mathcal{V}$, namely those with the largest and smaller
production of ATP, respectively. Denoting these respectively by
$\boldsymbol{\nu}_{max}$ and $\boldsymbol{\nu}_{min}$, in
Fig. \ref{due} we show, component by component, the quantities
$\Delta\boldsymbol{\lambda}=\boldsymbol{\lambda}_{max}-\boldsymbol{\lambda}_{min}$
and
$\Delta\boldsymbol{\nu}=\boldsymbol{\nu}_{max}-\boldsymbol{\nu}_{min}$.
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{maxmin}
\caption{\label{due}(Top) Component by component difference in
$\boldsymbol{\lambda}$ and (Bottom) in $\boldsymbol{\nu}$ between
the solution with the largest and smaller production rate of
ATP. The number of metabolites is 631, that of reactions
1057. Arrows in the top figure mark, from left to right, the
following metabolites: in the positive half, 3mob, 5aizc, 5mthf, ac,
ade, atp, co2, damp, dcyt, fad, for, gal, h, h2o, k(e), nad group,
nh4, pi(e), ppi, so3, so4(e) and urea; in the negative half 3mop,
5caiz, ade(e), adp, akg, co2(e), dhf, fadh2, for(e), h(e), lac,
pant-R, pi and urea(e).}
\end{center}
\end{figure}
Analyzing the production profiles it turns out that the state with
large ATP production is closely associated to the network mode,
i.e. it comes with a significant production of biomass components and
of metabolic waste like ac and co2. On the contrary states with
reduced ATP production are associated to considerably decreased co2
output and increased lactate output. This is consistent with an
increase of the flux through glycolisis, a central, anaerobic carbon
pathway with limited ATP productive efficiency that leads to the
generation of lactate. Large ATP output is instead achievable by
strengthening aerobic pathways that employ oxidative phosphorylation
with the concomitant production of CO$_2$. While a more careful
analysis is needed to map the exact locations of the flux
rearrangements, the overall difference in the organization of flux
between the two states is indeed significant. From a biological
perspective, these states are likely to lead to remarkably different
growth properties, suggesting that a sharper selection of production
profiles, and thus a better identification of metabolic objectives,
requires constraints that are still not included in the available
theories.
\section{Discussion}
Metabolic networks control, directly or indirectly, many of the most
basic tasks cells must accomplish, from the synthesis of amino acids,
to the maintenance of osmotic balance with the exterior, to the
response to environmental shifts. A possible key to improve our
understanding of their organization lies in our opinion in the
identification of the physical, stoichiometric, thermodynamic or
regulatory factors that intrinsically limit their productive potential
and determine the overall outcome of metabolic
activity. Constraint-based approaches provide a simple mathematical
frameworks where many of the emerging properties of metabolism can be
analyzed quantitatively. We have shown here that while the constraints
that are normally employed are sufficient to describe the main
physiological aspects of a cell's growth performance in a certain
medium, they may be still insufficient to infer precise metabolic
objective functions by which one could capture, e.g., objective shifts
under varying nutrient conditions. Otherwise it would be important to
understand whether the observed fluctuations have biological
counterparts. In addition to the problems presented here, the methods
of statistical mechanics developed for the analysis of
constraint-satisfaction problems may prove crucial to address many of
the important questions arising in this field, both for the structural
and the dynamical level \cite{rbc}.
\ack
This work was supported by the IIT (Italian Institute of Technology)
through the Seed Project DREAM.
\section*{References}
|
\section{Introduction}
\label{sec:1}
In financial markets, prices of stocks and commodities fluctuate over time
which then produce financial time series.
These time series are in fact of great interest both to practitioners and theoreticians
for making inferences and predictions.
Using modern day technologies, one can now obtain a vast amount of financial data
that record every transaction in financial markets which was not possible a couple of decades ago.
The analysis involved is also far more complicated.
With the tremendous amount of information obtained over the past decade,
researchers have now come to agree on several stylized facts about financial markets,
i.e., heavy tails (or fat tails in the terminology of finance) in asset return distributions,
absence of autocorrelations of asset returns,
volatility clustering, aggregational normality
and asymmetry between rises and falls~\cite{ref:sf1,ref:sf2,ref:sf3,ref:EPBOOK1,ref:EPBOOK2}.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure1}}
\caption{\label{fig:nasdaq:return}
The empirical data of the NASDAQ Composite index from February 8, 1971
through June 30, 2009. (a) shows the historical daily closing price
while (b) plots the daily returns during this period.}
\end{figure}
Figure~\ref{fig:nasdaq:return} (a) shows a plot of the historical daily closing values of
NASDAQ Composite index from February 8, 1971 through June 30, 2009
while figure~\ref{fig:nasdaq:return} (b) is its daily price returns during this period.
The price return $R_\tau(t)$ at time $t$ is defined as the difference
between the price $p(t)$ of a financial asset
(here it is the index value of NASDAQ) at time $t$
and its price a time $\tau$ before, $p(t-\tau)$, divided by $p(t-\tau)$,
\begin{equation}\label{eq:return}
R_\tau(t) = \frac{p(t) - p(t-\tau)}{p(t-\tau)} \,\, .
\end{equation}
Therefore, one can obtain the daily returns $R_{1}(t)$ by setting $\tau=1$ trading day and
these returns reflect the price fluctuations in this time series.
We will use daily returns to define fluctuations in a financial price series throughout this article.
As one can see in figure~\ref{fig:nasdaq:return} (b) that the daily returns are varying over time.
A naive thinking would be that these fluctuations are independent, identically distributed (iid)
variables generated by some random processes (i.e., random walks~\cite{ref:bachelier})
and therefore the probability density function of the returns should follow a Gaussian distribution.
However, it turns out that the empirical distributions of the returns are indeed heavy-tailed.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure2}}
\caption{\label{fig:nasdaq:pdf}
The probability density function of the normalized daily returns of the NASDAQ
index in figure~\ref{fig:nasdaq:return}.}
\end{figure}
In figure~\ref{fig:nasdaq:pdf},
we depict the probability density function of normalized daily returns of the NASDAQ index.
The normalized daily return is defined as $\left(R_1(t)-\mu_{R} \right)/\sigma_{R}$,
where $\mu_{R}$ and $\sigma_{R}$ denote the average and the standard deviation of $R_1(t)$.
One can clearly see that there are heavy tails at the two ends of the distribution.
For comparison, we also include a Gaussian fit with $\mu=0$ and $\sigma=1$.
This is one of the stylized facts that was discovered back in 1960s~\cite{ref:mandelbrot,ref:fama}.
Many studies have been carried out over the years on different financial time series
and the heavy tails in return distributions have always been observed.
There have been many suggestions on the form of the distributions
but no general consensus has been reached on the exact form of the tails so far.
We will not continue our discussion on this issue here
but refer our reader to the literature~\cite{ref:EPBOOK1,ref:EPBOOK2,ref:pdf1,ref:pdf2}.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure3}}
\caption{\label{fig:nasdaq:af}
The autocorrelation functions of the returns and its absolute value.}
\end{figure}
In addition to those heavy tails in return distributions,
large fluctuations in prices seem to lump together as well~\cite{ref:vc1,ref:vc2}.
If one examines the empirical time series shown in figure~\ref{fig:nasdaq:return},
it is easy to observe that large fluctuations in prices are more often
followed by large ones while small fluctuations are more likely
followed by small ones.
This stylized fact is known as volatility clustering~\cite{ref:vc3}.
In financial time series, it is not just that there are more large fluctuations
than pure random processes but also these large fluctuations tend to cluster together.
It is often suggested that a more quantitative way to view this property
is to look at the autocorrelations of the return series~\cite{ref:vc2}.
The autocorrelation function $C\left(x_t,x_{t+\tau}\right)$ is defined as
\begin{eqnarray}
C\left(x_t,x_{t+\tau}\right) \equiv
\frac{\left<\left(x_{t}-\left<x_{t}\right>\right) \left(x_{t+\tau}-\left<x_{t+\tau}\right>\right) \right>}
{\sqrt{\left<{x_t}^2\right> - \left<x_t\right>^2} \sqrt{\left<{x^2_{t+\tau}}\right> - \left<x_{t+\tau}\right>^2}} \,\, ,
\end{eqnarray}
where $\left<x\right>$ denotes the expectation value of the variable $x$.
While the returns themselves do not show the evidence of temporal correlations,
the absolute returns or their squares do display a positive,
pronounced slowly decaying autocorrelation which indeed exhibit power-law decay behaviour.
The autocorrelations of the absolute value or the square,
etc of the asset returns are often known as the nonlinear autocorrelations.
We will only consider the autocorrelation of the absolute returns
as an example of the nonlinear autocorrelation in this paper.
Figure~\ref{fig:nasdaq:af} are plots of the autocorrelation functions
of the returns and its absolute value for the time series shown in
figure~\ref{fig:nasdaq:return}.
It is easy to see that there is no correlation
among the returns since the autocorrelation function drops
to the noise level within a couple of days.
On the other hand, the autocorrelation function of the absolute returns,
i.e., the nonlinear autocorrelation does exhibit a much slower decay behaviour.
Researchers have fitted this with a power-law decay,
and it is not clear at this moment whether the slow decay should imply long time memory
of the financial time series~\cite{ref:vc3}.
However, one should also keep in mind that
if the time series do possess the properties of the long time memory
and the heavy-tailed distribution, many standard estimation procedures
(i.e., examining sample autocorrelations.) may fail to work~\cite{ref:vc3,ref:vc4,ref:vc5}.
Therefore, in order to have a more reliable measurement
of the volatility clustering, an alternative approach is also needed
while dealing with financial time series.
For instance, if only the clustering behaviour is concerned,
one can simply characterize this property by the concept of probability.
Table~\ref{tab:probability} is an example
which shows the probability of the occurrence of large and small
fluctuations following the occurrence of large or small fluctuations
on the previous day.
By large (small) fluctuations, we here choose them to be the
largest (smallest) 20\% of all the returns
and the remaining returns are denoted as the rest.
Therefore, each row in table~\ref{tab:probability} sums to unity.
It is easy to see that the probability that
there will be a large (small) return following
a large (small) one on the previous day is significantly higher
(larger than 20\% in this case) than that of a pure random process.
\begin{table}
\caption{\label{tab:probability}
The probability of the occurrence of large and small fluctuations following
the occurrence of large or small ones on the previous day (the first column).
The result here is for NASDAQ time series.}
\begin{center}\begin{tabular}{@{}ll|lll}
\hline
&20\% &Largest& Smallest& Rest\\
\hline
&Largest & 0.3947& 0.1156& 0.4897\\
&Smallest& 0.1265& 0.2401& 0.6334\\
&Rest & 0.1597& 0.2148& 0.6255\\
\end{tabular}\end{center}
\end{table}
A natural question to ask is whether the above stylized facts are indeed related to
each other and if so, is it possible for one to understand its origin.
In the following, we will give an attempt to answer the first question
which would hopefully shed light on searching for an answer to the second question.
This paper is organized as follows.
In section~\ref{sec:2}, we will give detailed analysis
of volatility clustering in financial time series.
In particular, we give arguments on what ingredient in financial time series
is responsible for reproducing the nonlinear autocorrelations of price
returns such as the one shown in figure~\ref{fig:nasdaq:af}.
We then introduce, in section~\ref{sec:3},
an index as a quantitative measure of volatility clustering in financial time series.
This would allow us to directly compare the degree of volatility clustering
across different financial time series.
The asymmetry between rises (gains) and falls (losses)
in the time series will be discussed in section~\ref{sec:4}.
Section~\ref{sec:5} will be the summary and discussion.
In this work,
we have carried out the analysis on seven different representative financial time series.
They include (i) NASDAQ Composite Index (NASDAQ), (ii) Standard \& Poor's 500 index (S\&P500),
(iii) Hang Seng Index (HSI), (iv) Microsoft stock price (MSFT),
(v) US Dollar/New Taiwan Dollar (USD/NTD),
(vi) Australian Dollar/New Taiwan Dollar (AUD/NTD)
and (vii) West Texas Intermediate (WTI).
While we use NASDAQ as an example throughout the paper,
we will include the results of other financial time series in the appendix.
\section{Volatility clustering and autocorrelation functions}
\label{sec:2}
We now begin our study by looking into the question of whether there
is a relationship among the heavy tails of return distributions,
volatility clustering and autocorrelation functions,
if the answer is yes, how they are related.
Let us begin by asking the following question:
Is it necessary for one to have a heavy-tailed distribution
in order for the nonlinear autocorrelation function to exhibit the slow decay?
To answer this question, let us now assume that the return distribution
follow a Gaussian distribution instead of the empirical distribution shown
in figure~\ref{fig:nasdaq:pdf}.
In this case, we assume the Gaussian distribution to have its mean and
standard deviation to be the same as the mean and the standard deviation
of the daily returns series in figure~\ref{fig:nasdaq:return}.
One can easily perform a simulation on this.
We now draw an equal number of returns from this Gaussian distribution
and call it the simulated data set.
After this is done, we sort both the empirical set and the simulated set
in the descending order of absolute returns.
We then substitute the values in the empirical data set by
the simulated data set one by one from the largest fluctuation to the smallest one
and calculate the nonlinear autocorrelation function of this rearranged Gaussian data.
The result is presented in figure~\ref{fig:simulation:af}.
For comparison, we also include the nonlinear autocorrelation of the empirical data
and the result from a pure Gaussian noise which is drawn from a Gaussian distribution
but without arranging the data according to the positions of empirical data set
like we do for the rearranged Gaussian data.
The pure Gaussian noise shows no temporal correlations as expected.
What is surprising is that the rearranged Gaussian returns
shows the same kind of slow decay behaviour as the empirical data set.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure4}}
\caption{\label{fig:simulation:af}
The nonlinear autocorrelation function of the empirical data and
of the rearranged Gaussian data.}
\end{figure}
On the other hand, if we randomize the temporal positions of the empirical returns,
namely, we reshuffle the original financial time series,
the result we obtain is always similar to the case of the pure Gaussian noise,
which means that there is no temporal correlation.
The above analysis therefore strongly suggests that
the heavy tails in return distributions are not responsible for
the slow decay behaviour of the nonlinear autocorrelation functions.
If the heavy tails in the distributions are not responsible for slow decay
in nonlinear autocorrelation functions,
what possible ingredients in the financial time series
would be responsible for such a slow decay behaviour.
We here try to provide an answer to this question.
Let us begin by looking at the clustering
of large price fluctuations in figure~\ref{fig:nasdaq:return}.
We begin by picking out the largest $p$\% (where $p$ is a constant)
fluctuations (whether they are positive or negative)
in the time series\footnote{
A similar treatment, the return interval approach~\cite{ref:RI1,ref:RI2,ref:RI3,ref:RI4},
is to pick the large fluctuations that are outside
$q$ standard deviations of the average value of the returns,
where $q$ is a pure number.}
and see whether their clustering behaviour would affect
the nonlinear autocorrelation function of the returns.
Since we are only interested in the clustering behaviour,
which in turn means the temporal positions but not the values of the
large fluctuations in the financial time series,
we can here simply use 1 to represent the largest $p$\% fluctuations and 0
for all the other smaller fluctuations.
In this way, we will have a sequence which contains only 0 and 1.
This will in turn make our analysis much easier to interpret.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure5}}
\caption{\label{fig:seq:af}
The nonlinear autocorrelation functions of the empirical data, the Gaussian data and
the sequences of 1s and 0s with $p=10$, 20 and 30.
$p$ here refers to the largest $p$\% fluctuations in the empirical returns
and are represented by 1s while the rest are represented by 0s.}
\end{figure}
Figure~\ref{fig:seq:af} shows the nonlinear autocorrelation
of figure~\ref{fig:nasdaq:return}
using 1 for the largest $p$\% fluctuations and 0 for the rest.
We here include the results for sequences with $p=10$, 20 and 30.
To facilitate our discussion, we also include both the nonlinear autocorrelation
for empirical data and the Gaussian noise for comparison.
One can see that all these sequences show similar slow decay behaviour
as the original empirical data set, though with smaller values.
This analysis thus shows that the positions of the large fluctuations
are essential for a slow-decaying nonlinear autocorrelation function.
Therefore, one can conclude that it is the clustering of large fluctuations
rather than the heavy tail in the return distribution
which should be responsible for the slow decay behaviour of nonlinear autocorrelation functions.
This fact has also been observed in the other financial time series
in our study and the results are presented in the appendix.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure6}}
\caption{\label{fig:USDNTD:af}
(a) The historical daily return series for the currency exchange rate USD/NTD (black)
and the series with the largest 20\% and smallest 20\% of the returns being swapped (red).
(b) The nonlinear autocorrelations of the original returns (line with open circles)
and the swapped returns (line with triangles).}
\end{figure}
Before we end this section, we would also like to make a further
study of the clustering of fluctuations in financial time series.
Instead of looking at the clustering of large fluctuations,
we now focus on the clustering of the small fluctuations in time series.
Since small fluctuations are smaller in value and
basically do not contribute to the nonlinear autocorrelation functions,
they are often left out in the discussion in the literature.
However, whether their temporal positions in a time series can have similar effects
as the large fluctuations is an interesting question that one can ask.
In figure~\ref{fig:USDNTD:af} (a), we plot the historical daily return time series
of the currency exchange rate USD/NTD from July 2, 2001 through June 30, 2009,
where the black line denotes the original empirical returns
while the red one represents the same set but with the largest 20\%
and smallest 20\% of the returns being swapped
\footnote{Our swapping procedure is described as follows.
We first swap the largest fluctuation (whether they are positive or negative)
with the smallest one in the time series
and then the second largest fluctuation with the second smallest one, and so on,
until the required percentage is achieved.}.
The nonlinear autocorrelations of the original empirical returns (line with open circles)
and the swapped returns (line with triangles) are drawn in figure~\ref{fig:USDNTD:af} (b).
We also include in this figure the autocorrelation function of the original returns for comparison.
One can see that although the line with triangles has values smaller than the original data set,
both lines have similar slow decay behaviour.
This in turn means that the clustering of small fluctuations in this returns series
has basically the same kind of feature as that of their large fluctuation counterparts.
On the other hand, as we swap the large and small fluctuations in the other
six financial time series that we have been investigating,
the nonlinear autocorrelation functions of the swapped returns series
show no sign of slow decay.
They basically drop very fast,
similar to the kind of Gaussian noise in figure~\ref{fig:simulation:af}.
This interesting fact will be discussed in more detail in the next section
as we introduce a clustering index to quantitatively study the
clustering behaviour of different financial time series.
The introduction of this index would then allow us to directly compare
the degree of clustering across different financial time series.
\section{Quantitative measurement of volatility clustering}
\label{sec:3}
As mentioned above, in order to discuss the volatility clustering in a
more quantitative way, it is better to introduce some parameters to
quantitatively measure the volatility clustering of different financial
time series that we can make comparison with.
We here introduce an index to quantify the volatility clustering in
the financial time series.
We begin by introducing a moving window with a certain window size
to scan through a given time series.
As an example, one can pick a window with size of $n$
(where $n$ is fixed throughout the scanning process) trading days.
Similar to what we have done in the previous section,
we can count the total number of trading days
that are among the largest $p$\% fluctuations in returns within this window
as we scan through the time series.
As we will see,
one can interpret this as the degree of volatility clustering of the largest $p$\%
fluctuations with respect to this particular window with size $n$.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure7}}
\caption{\label{fig:clustering:distribution}
The plot of the frequency distribution of the number of days with
largest 20\% fluctuations within a window of 10 trading days.}
\end{figure}
Figure~\ref{fig:clustering:distribution}
is an illustration of the clustering of the largest 20\% fluctuations
in figure~\ref{fig:nasdaq:return} with a window size of 10 trading days,
a span of two weeks in real daily life.
The statistics here is obtained by using the so called moving window method.
This means that we begin by putting the window on the first day
of the whole series and count the number of days among largest 20\% fluctuations
within this 10-day window.
This is the first step.
We then move on to the second day of the whole series and again count the number
of days among largest 20\% fluctuations within this next 10-day window, the second step.
We repeat the same procedure until we finish scanning through the whole time series.
The curve with full circles in figure~\ref{fig:clustering:distribution}
is a plot of the frequency distribution of the number of days
among the largest 20\% fluctuations within a 10-day period by using this moving window method.
To make it into a quantitative measure of the degree of clustering,
we need to compare it with a randomly generated time series for example, a Gaussian noise series.
The curve with open circles in figure~\ref{fig:clustering:distribution}
is the frequency distribution of the number of days
of the largest 20\% fluctuations within a 10-day period
from a simulated Gaussian noise series.
From figure~\ref{fig:clustering:distribution},
one can already visually tell the difference between these two curves.
To be more concise, we take the ratio of the standard deviation of the
number of days of the largest $p$\% fluctuations within the $n$-day window
between the empirical and the simulated data sets.
Mathematically, it is defined as $R_n \equiv \sigma_e/\sigma_{\rm G}$,
where $\sigma_e$ and $\sigma_{\rm G}$ are the standard deviation of the
number of days of the largest $p$\% fluctuations within an $n$-day period
for the empirical and simulated Gaussian data sets respectively.
The larger the ratio is, the larger the degree of clustering will be.
This result can be understood easily.
The average number of days of largest $p$\% fluctuations within
a window size of $n$ is equal to $p \times n/100$.
This is true irrespective of whether it is the empirical data set or the simulated one.
One can indeed see this for the simulated data set which has a peak near this value.
However, if the time series displays the phenomenon of
clustering of large fluctuations,
there will be a higher frequency of occurrence that the number of days
of the largest $p$\% fluctuations within this window
is much larger than the average value $p \times n/100$.
Similarly, there will also be a higher frequency of occurrence
that the number of days of the largest $p$\% fluctuations within this window
is much smaller than the average value $p \times n/100$.
This scenario will indeed be reflected in the value of the standard deviation
of the frequency distribution in figure~\ref{fig:clustering:distribution}.
Thus, one can simply take the ratio of the standard deviation
of the empirical and simulated data sets to get a quantitative measure
of the degree of clustering of the largest $p$\% fluctuations
of the financial time series that one is interested in.
The ratio or index $R_n$ that we introduce here can in fact be studied analytically.
It has both theoretical upper and lower bounds and the standard deviation of the
simulated Gaussian noise can also be calculated analytically.
Let us first derive the theoretical value of the standard deviation of the
simulated Gaussian noise.
Recall from above that the mean value of the
average number of days of the largest $p$\% fluctuations within
a $n$-day window is equal to $p \times n/100$.
For a total of $n$ days, the probability that there are $m$ days
with fluctuations among the largest $p$\% fluctuations can be written as,
\begin{eqnarray}
\frac{n!}{m!(n-m)!} P^m (1-P)^{n-m} \,\, ,
\end{eqnarray}
where $P$ denotes $p/100$.
We here convert the percentage into decimals for simplicity.
The standard deviation of the average number of days of the
largest $p$\% fluctuations within a $n$-day window is therefore equal to
\begin{eqnarray}
\sigma_{\rm G} = \left[\sum_{m=0}^n (m - P n)^2 P^m (1-P)^{n-m}\right]^{1/2} = \sqrt{nP(1-P)} \,\, ,
\end{eqnarray}
which is the familiar result in statistics for the standard deviation of a sequence
of $n$ random events with occurrence probability $P$.
The theoretical lower bound for the index corresponds to the case when the time series is
completely random, which is therefore equal to 1.
To get a theoretical upper limit of the standard deviation of
the average number of days of the largest $p$\% fluctuations within a $n$-day
period, we proceed as follows.
We look for the extreme case when all
the largest $p$\% fluctuations are ordered one after the other,
then followed by the rest of the data points
(one can of course reverse the order of the largest $p$\% fluctuations and the rest).
The first $p$\% of the data points will then be represented by 1
and the rest will be by 0, as what we have done in the above.
This is the case when we should have the largest possible degree of clustering.
If one plots this extreme case in figure~\ref{fig:clustering:distribution},
one will have two peaks in the frequency distribution function, one is at 0,
and the other is at $n$ (10 in the case in figure~\ref{fig:clustering:distribution}).
Let us now use a window of size $n$ and begin with the first data point,
which is a 1, and count the $n$ data points in this window,
all of which are 1s (assuming that the length
of the time series $N$ is much longer than the window size $n$).
Recall that we call this procedure to be step one.
We then let the window slide to the next data point, the second step, and so on.
As the moving window continues to move along the time series,
it will have moved $PN-n+1$ steps before it reaches the first 0.
We again have $P$ here to be equal to $p/100$ for simplicity.
As it continues to move along the time series,
the number of 1s will decrease
while the number of 0s will increase until the window consists of all 0s.
There are then $(1-P)N-n+1$ steps which has all 0s within the moving window.
For the whole time series, we have a total of $N-n+1$ steps so we have
to average over these steps.
It is now easy to calculate the standard deviation in this extreme case,
which is the square root of the expression in Eq. (\ref{eq:sigma}).
Recall that the average 1s within the moving window is $Pn$.
We then have
\begin{eqnarray}\label{eq:sigma}
&\frac{1}{N-n+1} \{ (PN-n)(n-P n)^2 + \left[ (1-P)N-n \right] (Pn)^2 + \sum_{m=0}^n (m-Pn)^2 \}\nonumber&\\
&=\frac{1}{N-n+1} \{ n^2(N-n-1)P(1-P) + \frac{n(n+1)(2n+1)}{6} - n^3 \left[ P^2+(1-P)^2 \right] \} \,.\;&
\end{eqnarray}
In the limit $PN$ and $(1-P)N >> n$, the right hand side of Eq. (\ref{eq:sigma}) reduces to $n^2 P(1-P)$.
Therefore, the theoretical limit of the standard deviation $\sigma_{\lim}$ as $N$ goes to infinity is
\begin{equation}
\sigma_{\lim} = \sqrt{n^2 P(1-P)} \,\, .
\end{equation}
The theoretical upper limit of $R_n$ is then equal to
\begin{equation}
R_{n}^{\lim} = \frac{\sigma_{\lim}}{\sigma_{\rm G}} = \sqrt{n} \,\, .
\end{equation}
Figure~\ref{fig:clustering_index} shows the value of the clustering index
for NASDAQ time series in figure~\ref{fig:nasdaq:return}
for various largest $p$\% fluctuations as a function of window size $n$.
The different curves represent the different largest $p$\% of the fluctuations in the time series.
We have included here the results for $p=5$, 10, 15 and 20.
Also included is the curve of the theoretical limit of the index.
The index values all start from unity when the window size $n$ corresponds to 1 trading day,
and gradually increase as the window size increases.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure8}}
\caption{\label{fig:clustering_index}
The clustering index, $R_n$, for the NASDAQ return series with
$p=5$ (solid), 10 (dash), 15 (dot) and 20 (dash dot).
The theoretical limit of the index is drawn as a thick line for comparison.}
\end{figure}
With the clustering index in hand, one can practically study the behaviour
of clustering of any sort of fluctuations in a financial time series.
Other than the largest $p$\% that we have looked into,
one can also look at the degree of clustering for small fluctuations.
To give the reader an idea of how one can use the index to study the properties
of financial time series,
we go back to a case which we considered in previous section.
Recall that we have studied a time series in which we swapped the largest $p$\%
and smallest $p$\% of the returns in the series,
as indicated in figure~\ref{fig:USDNTD:af}.
It turns out that the nonlinear autocorrelation function of the swapped data set
still exhibits similar slow decay behaviour.
On the other hand, we have analyzed the other six time series
that we consider in this paper and there is practically no such kind of slow
decay behaviour of the swapped data sets.
Using the index that we introduce here, the difference becomes clear.
Figure~\ref{fig:bottom20} shows the curves for the index value of the smallest 20\% returns
vs. window size in all the seven financial time series that we study in this work.
One can now see that the value of the index is rather small for each of the other
six financial time series when compared with the curve for USD/NTD.
This means that the clustering of the smallest 20\% returns of these other financial time series
indeed behave not much different from random sequences.
On the other hand, the clustering of the time series USD/NTD as shown
in figure~\ref{fig:USDNTD:af} is significantly larger
which in turn reflects the slow decay behaviour
of the swapped data set in figure~\ref{fig:USDNTD:af}.
A possible explanation for the anomaly of clustering behaviour shown in the USD/NTD time series
might be a result from the stronger interference of Taiwan Central Bank than other countries.
Since Taiwan is an export-oriented country, the government
will attempt to regulate more frequently the fluctuations in the daily closing value of USD/NTD.
Whether this is the main reason for such an anomaly is not clear to us for the moment.
The above example thus suggests that the index that we introduce here is a good indicator
to quantify the degree of clustering of fluctuations in financial time series.
In the next section, we will see that the index that we introduce here indeed contains
more information than people have previous observed in financial time series.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure9}}
\caption{\label{fig:bottom20}
The clustering index for smallest 20\% returns of the
NASDAQ, S\&P500, HSI, MSFT, USD/NTD, AUD/NTD and WTI series.
The theoretical limit of the index is drawn as a thick line for comparison.}
\end{figure}
\section{Rise/Fall asymmetry}
\label{sec:4}
There exists discussions in the literature~\cite{ref:as1,ref:as2,ref:as3,ref:as4}
about the asymmetry of asset returns
such as the skewness of the returns distribution in figure~\ref{fig:nasdaq:pdf}.
In the study of financial time series, one can for example,
ask whether there are more days that the returns are gains (rises) rather than losses (falls)
as some kind of asymmetry.
One can further ask whether the returns in gains would like to cluster more
or the other way round, how large the difference is,
and whether large fluctuations tend to cluster more, etc.
These can all be viewed as asymmetries in a financial time series.
With the index introduced in previous section,
one should hopefully be able to extract more information
on the asymmetries in financial time series
and to study these asymmetries on a more quantitative basis.
To study the asymmetries in financial time series, let us first give the definitions here.
In the case of the asymmetry between the largest/smallest returns,
we adopt the notation that the largest and smallest fluctuations
refer to the absolute returns as before.
We first obtain the clustering index for the largest and smallest $p$\% returns.
The asymmetry of largest/smallest returns $A_{\rm{ls}}$ is then defined as
\begin{eqnarray}
A_{\rm{ls}} = \frac{R_{\rm l} - R_{\rm s}}{R_{\rm l} + R_{\rm s}} \,\, ,
\end{eqnarray}
where $R_{\rm l}$ and $R_{\rm s}$ are the indices for the largest and smallest $p$\%
fluctuations respectively.
This asymmetry will give us an idea whether the large fluctuations
or the small fluctuations would like to cluster more
as we increase the size of the moving window.
From this definition, it is clear that $A_{\rm{ls}}$ is equal to zero
when the window size is equal to 1, since there are
an equal number of largest and smallest fluctuations.
In a similar fashion,
one can define the asymmetry between the largest positive and negative returns,
which we call $A_{+-}$ as follows
\begin{eqnarray}
A_{+-} = \frac{R_+ - R_-}{R_+ + R_-} \,\, ,
\end{eqnarray}
where $R_+$ and $R_-$ are the indices for the largest positive and negative returns
respectively.
We should remind our reader here that in the case of $A_{+-}$,
we first pick up the largest $p$\% fluctuations from the absolute returns
and then separate the fluctuations (returns) into positive and negative categories.
In this way, we can see the asymmetry between the large positive and negative
returns as well as their degree of clustering.
Notice that the asymmetries as defined above are bounded by 1 and -1.
Figure~\ref{fig:nasdaq:asymmetry} contains the plots of the asymmetries $A_{\rm{ls}}$ and $A_{+-}$
for $p$ = 15 and 20 for the NASDAQ time series in figure~\ref{fig:nasdaq:return}.
From the figure, it is easy to observe that the two curves for $A_{\rm{ls}}$ are always
positive, which means that the degree of clustering is more obvious
for large fluctuations than for small fluctuations in the NASDAQ times series.
On the other hand, the two curves for $A_{+-}$ are always below zero.
This reflects the fact that negative returns,
or big losses are likely to cluster together than big gains in the case of NASDAQ.
This is in agreement with some observations~\cite{ref:sf2,ref:as3}
indicating that there are more big losses rather than big gains in financial markets
since we have more big losses and these big losses are more likely to lump together.
We should remark here that the window size equals to 1
corresponds to the asymmetry of distribution of returns in figure~\ref{fig:nasdaq:pdf}.
In the case of NASDAQ, the asymmetry is negative.
There are however,
examples of financial time series that the asymmetry for the probability
density function is positive and they are included in the appendix below.
By increasing the size of the moving window, one can also study the asymmetry of
returns with respect to the clustering of large and small fluctuations.
Therefore, the use of the index to study asymmetries in financial time series
allows one to extract more information comparing to conventional methods.
\begin{figure}
\centering{\includegraphics[width=0.8\textwidth]{figure10}}
\caption{\label{fig:nasdaq:asymmetry}
The asymmetry $A_{\rm{ls}}$ and $A_{+-}$ with $p=15$ and 20 for NASDAQ return series.}
\end{figure}
The asymmetry between rises (gains) and falls (losses) in returns can also be observed
in the probability of the occurrence of large and small fluctuations following
the occurrence of large or small ones on the previous day,
which is shown in table~\ref{tab:probability:as}.
Unlike table~\ref{tab:probability},
we now separate the rise and fall of the fluctuations into separate categories.
Each row in this table sums to unity as before.
In this way, one can easily detect the difference.
\begin{table}
\caption{\label{tab:probability:as}
The probability of the occurrence of large and small rise/fall following
the occurrence of large or small rise/fall on the previous day (the first column).}
\begin{center}\begin{tabular}{@{}ll|lll}
\hline
&20\%&Largest (rise/fall)&Smallest (rise/fall)&Rest (rise/fall)\\
\hline
&Largest (rise) & 0.2054/0.1514& 0.0551/0.0724& 0.3319/0.1838\\
&Largest (fall) & 0.1856/0.2438& 0.0365/0.0681& 0.2122/0.2538\\
&Smallest (rise)& 0.0451/0.0573& 0.1437/0.1023& 0.3624/0.2892\\
&Smallest (fall)& 0.0790/0.0767& 0.1226/0.1100& 0.3471/0.2646\\
&Rest (rise) & 0.0742/0.0582& 0.1269/0.0934& 0.4475/0.1998\\
&Rest (fall) & 0.0737/0.1255& 0.1183/0.0888& 0.2888/0.3049\\
\end{tabular}\end{center}
\end{table}
\section{Summary and discussion}
\label{sec:5}
In this paper, we have made a detailed analysis of the stylized facts in financial time series.
We have found that the slow decay behaviour is directly related to the degree of clustering
of the large fluctuations (absolute returns) within the financial time series
while the heavy tails in return distributions do not seem to play a role here.
We have also introduced an index to quantitatively measure the
clustering behaviour of fluctuations in financial time series
and have given examples to demonstrate its advantages over the conventional methods.
This index has both theoretical lower and upper bounds.
It is equal to unity if the fluctuations are independent,
identically distributed within the financial time series.
On the other hand, its upper bound can also be analytically calculated
and in the limit when the time scale of the given series is much longer
than the window size $n$, the index $R_n$ is simply equal to $\sqrt{n}$.
With this index in hand,
one not only can study the asymmetry of the asset returns but also the effect of
clustering on the asymmetry properties in financial time series.
One can see that the larger fluctuations tend to cluster more than the smaller ones.
Similarly, big losses tend to lump together more severely than big gains.
These findings should be helpful to people who make investments in financial markets.
Indeed, the clustering index introduced might also be employed to investigate the
clustering behaviour of other nonlinear systems.
Whether the clustering of large fluctuations in financial time series is from
long time memory effect or other effects such as human psychology is not clear
to us at the moment.
Indeed, if only the clustering of large fluctuations is concerned,
one is able to find complex systems that are random in nature
but can possess the property of large degree of clustering.
It is also possible to find simple ways which can give both large degrees of clustering
and slow-decaying nonlinear autocorrelations in a simulated time series.
We will discuss these issues in a future publication~\cite{ref:LM}.
\section*{Acknowledgments}
This work was supported in part by the National Science Council of Taiwan
under grants NSC\#98-2120-M-001-002 and NSC\#97-2112-M-001-008-MY3.
|
\section{Perturbative Non-renormalizability and Feynman Path Integral}
\label{sec:one-loop}
In a quantum theory of gravity the coupling constant is dimensionful, $G \sim \mu^{2-d}$,
and within the standard perturbative treatment of radiative corrections one expects trouble
in four dimensions, based on purely on dimensional grounds.
The divergent one loop corrections are proportional to
$G \Lambda^{d-2} $ where $\Lambda$ is the ultraviolet cutoff, which
then leads to a bad high momentum behavior, with an effective running
Newton's constant
\begin{equation}
G(k^2) \, / \, G \, \sim \, 1 + c_1 (d) \; G \, k^{d-2} + \, O(G^2) \; .
\label{eq:g-pert}
\end{equation}
A more general argument for perturbative non-renormalizability
starts by considering the gravitational action with scalar curvature term $R$,
which involves two derivatives of the metric.
Then the graviton propagator in momentum space goes like $1/k^2$,
and the vertex functions like $k^2$.
In $d$ dimensions each loop integral with involve a momentum
integration $d^d k$, so that the superficial degree of divergence ${\cal D}$
of a Feynman diagram with $L$ loops is given by
\begin{equation}
{\cal D} = 2 + (d-2) \, L \; ,
\end{equation}
independent of the number of external lines.
One therefore concludes that for $d>2$ the degree of
divergence for Einstein gravity increases rapidly with loop order $L$, and
that the theory cannot be renormalized in naive perturbation theory.
A consequence of the lack of perturbative renormalizability is the fact
that new higher derivative counterterms arise to one-loop order [1]
\begin{equation}
\Delta {\cal L}_g = { \sqrt{g} \over 8 \pi^2 (d-4) }
\left (
{1 \over 120 } R^2 +
{ 7 \over 20 } R_{\mu\nu} R^{\mu\nu}
\right ) \; ,
\end{equation}
with even higher derivatives appearing at the next order.
One concludes that the standard approach based on a perturbative
expansion of the pure Einstein theory in four dimensions is not
convergent; in fact it is badly divergent.
A number of possible options have been proposed, the simplest of which
is to just add the above higher derivative terms to the original action.
The resulting extended theory is perturbatively renormalizabe to all orders,
but suffers potentially from unitarity problems.
But these cannot be satisfactorily addressed in perturbation theory,
as the theory is now asymptotically free in the
higher derivative couplings and presumably exhibits a non-trivial vacuum.
Alternatively, the gravity divergences can be cancelled by adding
new unobserved massless particles
and invoking supersymmety; in fact it has been claimed recently that $N=8$
supergravity might not be just renormalizable, but indeed finite to some
relatively high loop order.
The downside of this somewhat contrived approach is the artificial introduction
of a plethora of unobserved massless particles of spin $0,1/2,1,3/2$,
added to the original action in order to cancel the gravitational ultraviolet divergences.
Finally, string theory solves the problem of ultraviolet divergences by postulating the existence
of fundamental extended objects, thus in part bypassing the requirement of supersymmetry
and providing a natural cutoff for gravity, related to a fundamental string scale [2].
Nevertheless one important point that cannot be overlooked is the fact that in other
field theories,
which to some extent share with gravity the same set of ultraviolet problems
(the non-linear sigma model is the most notable one, and the best studied case),
the analogous result of Eq.~(\ref{eq:g-pert}) is in fact known to be {\it incorrect}.
It appears as an artifact of naive perturbation theory, which in four dimensions
does not converge, and seems to lead therefore to fundamentally incorrect answers.
The correct answer in these models is found instead either by expanding around the dimension
in which the theory {\it is} perturbatively renormalizable, or by solving it exactly in the large
$N$ limit and then computing $1/N$ corrections, or by solving it numerically on a lattice.
The generic new feature in these models is the existence of a non-trivial fixed point
of the renormalization group [3-7], which is inaccessible by perturbation theory
in four dimensions, and radically alters the ultraviolet behavior of the theory.
\footnote{
After QED, the second most accurate prediction of quantum field theory to date
is for a perturbatively non-renormalizable theory, the $O(N)$ non-linear $\sigma$-model
in three dimensions, whose field theoretic treatment based on a non-trivial
fixed point of the renormalization group, either
on the lattice or in the continuum, eventually provides detailed predictions
for scaling behavior and anomalous dimensions in the vicinity of the fixed point [5,8].
These have recently been verified experimentally to high accuracy in a sophisticated
space shuttle experiment [9] for critical superfuid Helium,
whose order parameter corresponds to $N=2$ in the non-linear $\sigma$-model.
}
The key question for gravity is therefore: are the ultraviolet problems just
an artifact of a naive application of perturbation theory in four dimensions,
as clearly happens in other perturbatively non-renormalizable theories that
also contain dimensionful couplings in four dimensions?
In the following I will limit my discussion to the approach
based on traditional quantum field theory methods and the renormalization
group, applied to the Einstein action with a cosmological term,
an avenue which in the end is intimately tied with the existence of a non-trivial
ultraviolet fixed point in $G$ in four dimensions.
The nature of such a fixed point was first discussed in detail by K. Wilson for scalar
and fermionic theories [3],
and the methods later applied to gravity in [6], where they were referred to as
asymptotic safety.
As discussed above, it is fair to say that so far this is the only approach known to work
in other not perturbatively renormalizable theories.
If non-perturbative effects play an important role in quantum gravity,
then one would expect the need for an improved formulation
of the quantum theory, which does not rely exclusively
on the framework of perturbation theory.
After all, the fluctuating quantum metric field $g_{\mu\nu}$ is dimensionless,
and carries therefore no natural scale.
For the somewhat simpler cases of a scalar field and non-Abelian gauge theories
a consistent
non-perturbative formulation based on the Feynman path integral
has been used for some time, and is by now well developed.
In a nutshell, the Feynman path integral formulation for quantum
gravitation can be expressed by the functional integral formula
\begin{equation}
Z = \int_ {\rm geometries } e^{ \, { i \over \hbar} I_{\rm geometry} } \;\; .
\end{equation}
Furthermore a bit of thought reveals that for gravity, to all orders
in the weak field expansion, there is really no difference of
substance between the Lorentzian (or pseudo-Riemannian) and the
Euclidean (or Riemannian) formulation, which can be mapped into
each other by analytic continuation. In the following therefore
the Euclidean formulation will be assumed, unless stated otherwise.
In function space one needs a metric
before one can define a volume element.
Therefore, following DeWitt, one first defines an invariant norm for metric deformations
\begin{equation}
\Vert \delta g \Vert^2 \, = \,
\int d^d x \; \delta g_{\mu \nu}(x) \,
G^{\mu \nu, \alpha \beta} \bigl [ g(x) \bigr ] \,
\delta g_{\alpha \beta}(x) \;\; ,
\end{equation}
with the supermetric $G$ given by the ultra-local
expression
\begin{equation}
G^{\mu \nu, \alpha \beta} \bigl [ g(x) \bigr ] \, = \,
\half \, \sqrt{g(x)} \, \left [ \,
g^{\mu \alpha}(x) \, g^{\nu \beta}(x) +
g^{\mu \beta}(x) \, g^{\nu \alpha}(x) - \lambda \,
g^{\mu \nu}(x) \, g^{\alpha \beta}(x) \, \right ]
\end{equation}
with $\lambda$ a real parameter, $\lambda \neq 2 / d $.
The DeWitt supermetric then defines a suitable functional volume element $\sqrt{G}$
in four dimensions,
\begin{equation}
\int [d \, g_{\mu\nu} ] \, = \,
\int \, \prod_x \, \prod_{\mu \ge \nu} \, d g_{\mu \nu} (x) \; .
\end{equation}
The Euclidean Feynman path integral for pure
Einstein gravity with a cosmological constant term is then written as
\begin{equation}
Z_{cont} \; = \; \int [ d \, g_{\mu\nu} ] \; \exp \, \left \{
- \lambda_0 \, \int d x \, \sqrt g \, + \,
{ 1 \over 16 \pi G } \int d x \, \sqrt g \, R \, \right \} \; .
\label{eq:zcont}
\end{equation}
An important aspect of this path integral is connected with
the global scaling properties of the action and the measure [10].
First one notices that in pure Einstein gravity with a bare cosmological
constant term
\begin{equation}
{\cal L} = - { 1 \over 16 \pi G_0} \, \sqrt{g} \, R \, + \, \lambda_0 \sqrt{g}
\end{equation}
one can rescale the metric by $ g_{\mu\nu} = \omega \, g_{\mu\nu}' $ with
$\omega$ a constant, giving
\begin{equation}
{\cal L} = - { 1 \over 16 \pi G_0} \, \omega^{d/2-1} \, \sqrt{g'} \, R'
\, + \, \lambda_0 \, \omega^{d/2} \, \sqrt{g'} \; .
\label{eq:rescale}
\end{equation}
This can then be interpreted as a rescaling of the two bare couplings
$ G_0 \rightarrow \omega^{-d/2+1} G_0 $,
$\lambda_0 \rightarrow \lambda_0 \, \omega^{d/2} $,
leaving the dimensionless combination $G_0^d \lambda_0^{d-2}$ unchanged.
Therefore only the latter combination has physical meaning in pure gravity.
In particular, one can always choose the scale $\omega = \lambda_0^{-2/d}$,
so as to adjust the volume term to have a unit coefficient.
The implication of this last result is that pure gravity only contains one
bare coupling $G_0$,
besides the ultraviolet cutoff $\Lambda$ needed to regulate the quantum theory.
\vskip 20pt
\section{Gravity in $2+\epsilon$ Dimensions and Non-Trivial UV Fixed Point}
\label{sec:phaseseps}
In two dimensions the gravitational coupling becomes dimensionless,
$G\sim \Lambda^{2-d}$,
and the theory appears perturbatively renormalizable.
In spite of the fact that the gravitational action reduces to a topological
invariant, it is meaningful to attempt to construct,
in analogy to what was suggested originally by Wilson for scalar field theories,
the theory perturbatively as a double series in $\epsilon=d-2$ and $G$.
The $2+\epsilon$ expansion for pure gravity then proceeds as follows [6,11].
First the gravitational part of the action
\begin{equation}
{\cal L} = - { \mu^\epsilon \over 16 \pi G} \, \sqrt{g} \, R \, + \, \lambda_0 \sqrt{g} \; ,
\label{eq:l-pure}
\end{equation}
with $G$ now dimensionless and $\mu$ an arbitrary momentum scale,
is expanded by setting
\begin{equation}
g_{\mu\nu} \, \rightarrow \, \bar g_{\mu\nu} = g_{\mu\nu} \, + \, h_{\mu\nu} \; ,
\end{equation}
where $g_{\mu\nu}$ is the classical background field, and $h_{\mu\nu}$
a small quantum fluctuation.
The quantity ${\cal L}$ in Eq.~(\ref{eq:l-pure}) is naturally identified with
the bare Lagrangian, and the scale $\mu$ with a microscopic ultraviolet
cutoff $\Lambda$, corresponding to the inverse lattice spacing in the lattice formulation.
After the quantum fluctuations in $h_{\mu\nu}$ are integrated out and
the cosmological constant term gets rescaled, one obtains the following
result for the renormalization group beta function for $G$:
with $N_S$ scalar fields and $N_F$ Majorana fermion fields the result to two loops reads [12]
\begin{equation}
\mu { \partial \over \partial \mu } G =
\beta (G) = \epsilon \, G \, - \, \beta_0 \, G^2
\, - \, \beta_1 \, G^3 \, + \, O( G^4, G^3 \epsilon , G^2 \epsilon^2 ) \; ,
\label{eq:beta-twoloop}
\end{equation}
with $\beta_0 = \twoth \, (25 - c) $ and
$ \beta_1 = { 20 \over 3 } \, (25 - c) $, and $c \equiv N_S+N_F/2$.
The physics of this result is contained in the fact that the gravitational
$\beta$-function determines the scale dependence
of Newton's constant $G$, and has the shape shown in Fig. 1.
\begin{center}
\epsfxsize=7cm
\epsfbox{beta-function.eps}
\end{center}
\noindent{\small Figure 1.
Renormalization group $\beta$-function
for gravity close to two dimensions.
The arrows indicate the coupling constant flow towards
increasingly larger distance scales.}
\medskip
\label{fig:beta-g-eps}
A closer examination of the result shows that
depending on whether one is on the right ($G>G_c$) or on the left
($G<G_c$) of the non-trivial ultraviolet fixed point at
\begin{equation}
G_c = { d - 2 \over \beta_0 } + O((d-2)^2 )
\end{equation}
(with $G_c$ positive, provided one has $c<25$)
the coupling will either flow to increasingly larger values of $G$,
or flow towards the Gaussian fixed point at $G=0$, respectively.
Furthermore the running of $G$ as a function of the sliding momentum
scale $\mu=k$ can be obtained by integrating
Eq.~(\ref{eq:beta-twoloop}), and one has to lowest order
\begin{equation}
G(k^2) \; \simeq \; G_c \, \left [
\, 1 \, \pm \, c_0 \, \left ( {m^2 \over k^2 } \right )^{(d-2)/2}
\, + \, \dots \right ] \; ,
\label{eq:grun-cont1}
\end{equation}
with $c_0$ a positive constant, and $m$ a new nonperturbative scale.
As in non-abelian gauge theories and QCD, this last quantity arises naturally as an
integration constant of the renormalization group equations.
The choice of $+$ or $-$ sign is then determined from whether one is
to the left (+), or to right (-) of $G_c$, in which case
the effective $G(k^2)$ decreases or, respectively, increases as one flows away
from the ultraviolet fixed point towards lower momenta, or larger distances.
Physically therefore the two solutions represent a gravitational screening ($G<G_c$), and a
gravitational anti-screening ($G>G_c$) situation [10].
Finally, at energies sufficiently high to become comparable
to the ultraviolet cutoff, the gravitational coupling $G$ eventually flows towards the
ultraviolet fixed point
$ G(k^2) \, \mathrel{\mathop\sim_{ k^2 \rightarrow \Lambda^2 }} \, G (\Lambda) $,
where $G(\Lambda)$ is the coupling at the cutoff scale $\Lambda$,
to be identified with the bare or lattice coupling.
One message therefore is that the quantum corrections
involves a new physical, renormalization group invariant scale $\xi=1/m$,
which cannot be fixed perturbatively and whose size determines the
scale for the new quantum effects.
In terms of the bare coupling $G(\Lambda)$, it is given by
\begin{equation}
\xi^{-1} = m = A_m \cdot \Lambda \,
\exp \left ( { - \int^{G(\Lambda)} \, {d G' \over \beta (G') } }
\right ) \; .
\label{eq:m-cont}
\end{equation}
The constant $A_m$ on the r.h.s. of Eq.~(\ref{eq:m-cont})
cannot be determined by perturbation theory;
it needs to be computed by non-perturbative lattice methods.
At the fixed point $G=G_c$ the theory is scale invariant by definition;
in statistical field theory language the fixed point corresponds to a phase transition.
In the vicinity of the fixed point one can write
\begin{equation}
\beta (G) \, \mathrel{\mathop\sim_{ G \rightarrow G_c }} \,
\beta' (G_c) \, (G-G_c) \, + \, O ((G-G_c)^2 ) \; .
\label{eq:beta-lin-g}
\end{equation}
If one defines the exponent $\nu$ by $ \beta ' (G_c) \, = \, - 1/ \nu $,
then from Eq.~(\ref{eq:m-cont}) one has by integration
\begin{equation}
m \, \mathrel{\mathop\sim_{G \rightarrow G_c }} \,
\Lambda \cdot A_m \, | \, G (\Lambda) - G_c |^{\nu} \;\; ,
\label{eq:m-cont1}
\end{equation}
with $\nu$ the correlation length exponent.
To two loops the results of [12] imply
\begin{equation}
\nu^{-1} = \epsilon \, + \, {15 \over 25 - c } \, \epsilon^2 \, + \, \dots
\label{eq:nueps}
\end{equation}
which gives, for pure gravity without matter ($c=0$) in four dimensions, to lowest order
the scaling exponent $\nu^{-1} = 2$, and $\nu^{-1} \approx 4.4 $ at the next order.
The key question raised by these $2 + \epsilon$ perturbative calculations is
therefore: what remains of the above phase transition in four dimensions,
how are the two phases of gravity characterized non-perturbatively,
and what is the value of the exponent $\nu$ determining
the running of $G$ in the vicinity of the fixed point
in four dimensions ?
To answer this question in a controlled way would seem to require the
introduction of a non-perturbative regulator, based on the lattice formulation
(since no other reliable non-perturbative regulator for field theories is
known to date).
\vskip 20pt
\section{Lattice Regularized Quantum Gravity}
\label{sec:lattice}
On the lattice the infinite number of degrees of freedom in the continuum
is restricted, by considering Riemannian spaces described
by only a finite number of variables, to the geodesic distances between
neighboring points.
Such spaces are taken to be flat almost everywhere, and referred to as
piecewise linear.
The elementary building blocks for $d$-dimensional space-time are then
simplices of dimension $d$.
A 0-simplex is a point, a 1-simplex is an edge, a 2-simplex is a triangle, a
3-simplex is a tetrahedron.
A $d$-simplex is a $d$-dimensional object with $d+1$ vertices and
$d(d+1)/2$ edges connecting them [13].
The geometry of the interior of a $d$-simplex is assumed to be flat,
and is therefore completely specified by the lengths of its $d(d+1)/2$ edges.
When focusing on one such $n$-simplex, it is convenient to label
the vertices by $0,1, 2, 3, \dots , n$ and denote the
square edge lengths by $l _ {01}^2 = l _ {10}^2$, ... , $l _ {0n}^2 $.
The simplex can then be spanned by the set of $n$ vectors
$e_1$, ... $e_n$ connecting the vertex $0$ to the
other vertices.
To the remaining edges within the simplex one then
assigns vectors $e_{ij} = e_i-e_j$ with $1 \le i < j \le n$.
Within each $n$-simplex one can define a metric
$ g_{ij} (s) \; = \; e_i \cdot e_j $, and then in terms of the edge lengths
$l_{ij} \, = \, | e_i - e_ j | $ the metric is given by
\begin{equation}
g_{ij} (s) \; = \; \half \,
\left ( l_{0i}^2 + l_{0j}^2 - l_{ij}^2 \right ) \; .
\label{eq:latmet}
\end{equation}
The volume of a general $n$-simplex can be found by the $n$-dimensional
generalization of the well-known formula for a tetrahedron, namely
\begin{equation}
V_n (s) \; = \; {1 \over n ! } \sqrt { \det g_{ij} (s) } \; .
\label{eq:vol-met}
\end{equation}
In a piecewise linear space curvature is detected by going around
elementary loops which are dual to a ($d-2$)-dimensional subspace.
From the dihedral angles $\theta (s,h)$
associated with the faces of the simplices meeting
at a given hinge $h$ one computes the deficit angle $\delta (h)$,
defined as [13]
\begin{equation}
\delta (h) \; = \; 2 \pi \, - \, \sum_{ s \supset h } \; \theta (s,h) \; ,
\label{eq:deficit}
\end{equation}
where the sum extends over all simplices $s$ meeting on $h$.
It then follows that the deficit angle $\delta$ is a measure
of the local curvature at $h$.
By considering rotation matrices around a hinge one can obtain
an expression for the Riemann tensor at the hinge $h$
\begin{equation}
R_{\mu\nu\lambda\sigma} (h) \; = \; {\delta (h) \over A_C (h) }
\, U_{\mu\nu} (h) \, U_{\lambda\sigma} (h)
\label{eq:riem-hinge}
\end{equation}
which is expected to be valid in the limit of small curvatures,
with $A_C (h) $ the area of the loop entangling the hinge $h$,
and $U$ a bivector describing the hinge's orientation.
From the expression for the Riemann tensor at a hinge
given in Eq.~(\ref{eq:riem-hinge}) one obtains by contraction the Ricci scalar
$ R (h) = 2 \, \delta (h) / A_C (h) $, and the
continuum expression $\sqrt{g} \, R$ is then obtained
by multiplication with the volume element $V (h) $ associated with
a hinge.
The curvature and cosmological constant terms then lead to the combined Regge lattice action
\begin{equation}
I_{\rm latt} (l^2) \; = \; \lambda_0 \sum_{\rm simplices \; s} \, V^{(d)}_s
\, - \, k \sum_{\rm hinges \; h} \, \delta_h \, V^{(d-2)}_h \; .
\label{eq:latac}
\end{equation}
One key aspect of this formulation is the local gauge invariance of the
lattice action, in analogy to the local gauge invariance of the Wilson
action for gauge theories.
Already on a flat 2-d lattice it is clear that one can move around a point on a surface,
keeping all the neighbors fixed, without violating the triangle inequalities, and
leaving local curvature invariants unchanged.
In $d$ dimensions this transformation has $d$ parameters and is an exact
invariance of the action.
When space is slightly curved, piecewise linear diffeomorphisms
can still be defined as the
set of local motions of points that leave the local contribution to the action,
the measure and the lattice analogs of continuum curvature invariants unchanged.
In the limit when the number of edges becomes very large one expects the full continuum diffeomorphism group to be recovered [14].
In order to write down a lattice path integral, one needs, besides the action,
a functional measure.
As the edge lengths $l_{ij}$ play the role of the continuum metric
$g_{\mu\nu}(x)$, one expects the discrete measure to involve an
integration over the squared edge lengths.
After choosing coordinates along the edges emanating from a vertex,
the relation between metric perturbations and squared edge
length variations for a given simplex based at 0 in $d$ dimensions is
from Eq.~(\ref{eq:latmet})
\begin{equation}
\delta g_{ij} (l^2) \; = \; \half \;
( \delta l_{0i}^2 + \delta l_{0j}^2 - \delta l_{ij}^2 ) \; .
\label{eq:latmet1}
\end{equation}
For one $d$-dimensional simplex labeled by $s$
the integration over the metric is thus equivalent to an
integration over the edge lengths, and one has the identity
\begin{equation}
\left ( {1 \over d ! } \sqrt { \det g_{ij}(s) } \right )
\prod_{ i \geq j } \, d g_{i j} (s) =
{\textstyle \left ( - { 1 \over 2 } \right ) \displaystyle}^{ d(d-1) \over 2 }
\left [ V_d (l^2) \right ] \!
\prod_{ k = 1 }^{ d(d+1)/2 } \! \! dl_{k}^2 \; .
\label{eq:simpmeas}
\end{equation}
Indeed there are $d(d+1)/2$ edges for each
simplex, just as there are $d(d+1)/2$ independent components for the metric
tensor in $d$ dimensions.
In addition, a certain set of simplicial inequalities need to be imposed on the edge lengths.
These represent conditions conditions on the edge lengths $l_{ij}$
such that the sites $i$ can be considered as vertices of a
$d$-simplex embedded in flat $d$-dimensional Euclidean space.
After summing over all simplices one derives
what is regarded as the lattice functional measure
representing the continuum DeWitt measure in four dimensions
\begin{equation}
\int [ d l^2] \; = \; \int_0^\infty \prod_{ ij } \, dl_{ij}^2
\; \Theta [ l_{ij}^2 ] \;\; .
\label{eq:lattmeas}
\end{equation}
Here $ \Theta [l_{ij}^2] $ is
a (step) function of the edge lengths, with the property
that it is equal to one whenever the triangle inequalities and their
higher dimensional analogs are satisfied and zero otherwise.
The lattice action of Eq.~(\ref{eq:latac}) for pure four-dimensional Euclidean
gravity then leads to the regularized lattice functional integral [7]
\begin{equation}
Z_{latt} \; = \; \int [ d \, l^2 ] \; \exp \left \{
- \lambda_0 \sum_h V_h \, + \, k \sum_h \delta_h A_h
\right \} \; ,
\label{eq:zlatt}
\end{equation}
where, as customary, the lattice ultraviolet cutoff is set equal to one
(i.e. all length scales are measured in units of the lattice cutoff).
Furthermore, $\lambda_0$ sets the overall scale, and can therefore be set
equal to one without any loss of generality, according to the scaling arguments
presented before.
The lattice partition function $Z_{latt}$ should then be compared to the
continuum Euclidean Feynman path integral for pure gravity of Eq.~(\ref{eq:zcont}).
In closing we note that what makes the Regge theory stand out compared to other
possible discretization of gravity is the fact that it is the {\it only} lattice theory known
to have the correct spectrum
of continuous excitations in the weak field limit, i.e. transverse traceless modes,
or, equivalently, helicity-two massless gravitons.
Indeed one of the simplest possible problems that can be treated in lattice quantum
gravity is the analysis of small fluctuations about a fixed flat simplicial background.
In this case one finds that the lattice graviton propagator in a
De Donder-like gauge is identical to the continuum expression [7].
\vskip 20 pt
\section{Strongly Coupled Gravity and Gravitational Wilson Loop}
\label{sec:strong}
As in non-abelian gauge theories, important information about the
non-perturbative ground state of the theory can be gained by
considering the strong coupling limit.
In lattice gravity an expansion can be performed for large $G$ or small $k=1/8 \pi G$,
and the resulting series is in general expected to be useful up to some $k=k_c$,
where $k_c$ is the lattice critical point,
at which point the partition function $Z$ eventually develops a singularity.
One starts from the lattice regularized
path integral with action Eq.~(\ref{eq:latac}) and
measure Eq.~(\ref{eq:lattmeas}).
The four-dimensional Euclidean lattice action usually contains
a cosmological constant and scalar curvature term as in Eq.~(\ref{eq:latac}),
\begin{equation}
I_{latt} \; = \; \lambda \, \sum_h V_h (l^2) \, - \,
k \sum_h \delta_h (l^2 ) \, A_h (l^2) \; .
\label{eq:ilatt1}
\end{equation}
The action only couples edges which belong either to
the same simplex or to a set of neighboring simplices, and can therefore
be considered {\it local} just like the continuum action;
it leads to the lattice partition function defined in Eq.~(\ref{eq:zlatt}).
When doing an expansion in the kinetic term
proportional to $k$, it is convenient to include the
$\lambda$-term in the measure, and
$Z_{latt}$ can then be expanded in powers of $k$,
\begin{equation}
Z_{latt}(k) \; = \; \int d \mu (l^2) \, \; e^{k \sum_h \delta_h \, A_h }
\; = \; \sum_{n=0}^{\infty} \, { 1 \over n!} \, k^n \,
\int d \mu (l^2) \, \left ( \sum_h \delta_h \, A_h \right )^n \;\; .
\label{eq:zlatt-k}
\end{equation}
$Z (k) \, = \, \sum_{n=0}^{\infty} a_n \, k^n $
is analytic at $k=0$, so this expansion is well defined up to the
nearest singularity in the complex $k$ plane.
In the gravity case the analogs of the gauge variables of Yang-Mills
theories are given by the connections, so it is natural when computing the
gravitational Wilson loop [15] to look for a first order formulation of Regge gravity [16].
For each neighboring pair of simplices $s,s+1$ one
can associate a Lorentz transformation $ R^{\mu}_{\;\; \nu} (s,s+1)$,
and one then might want to consider a near-planar closed loop $C$,
such as the one shown schematically in Fig.2.
Along a closed loop the overall rotation matrix is given by
\begin{equation}
R^{\mu}_{\;\; \nu} (C) \; = \;
\Bigl [ \prod_{s \, \subset C} R_{s,s+1} \Bigr ]^{\mu}_{\;\; \nu}
\end{equation}
\begin{center}
\epsfxsize=7cm
\epsfbox{wilson.eps}
\end{center}
\noindent{\small Figure 2.
Gravitational analog of the Wilson loop.
A vector is parallel-transported along the larger outer loop.
The enclosed minimal surface is tiled with parallel
transport polygons, here chosen to be triangles for illustrative
purposes.
For each link of the dual lattice, the elementary parallel transport
matrices ${\bf R}(s,s')$ are represented by arrows. }
\medskip
\label{fig:wilson}
In the quantum theory one is interested in averages
of the above product of rotations along a given path.
If the curvature of the manifold is small, then classically the expression
for the rotation matrix ${\bf R} (C )$ associated with a near-planar
loop can be re-written in terms of a surface
integral of the large-scale Riemann tensor, projected along the surface
area element bivector $A^{\alpha\beta} (C )$ associated with the orientation of
the loop,
\begin{equation}
R^{\mu}_{\;\; \nu} (C) \; \approx \;
\Bigl [ \, e^{\half \int_S
R^{\, \cdot}_{\;\; \cdot \, \alpha\beta} \, A^{\alpha\beta} ( C )}
\Bigr ]^{\mu}_{\;\; \nu} \;\; .
\label{eq:rotriem}
\end{equation}
Thus a direct calculation of the quantum Wilson loop could in principle
provide a way of determining the {\it effective} curvature on very
large distance scales, even in the case
where short distance fluctuations in the metric may be significant.
A detailed lattice calculation in the strong coupling limit then gives
the following result.
First one defines the lattice Wilson loop as
\begin{equation}
W(C) \; = \; < \; Tr[(U_C \; + \; \epsilon \; I_4) \;
R_1 \; R_2 \; ... \; ... \; R_n] \; > \; .
\label{eq:wloop-def2}
\end{equation}
where the $R_i$'s are the rotation matrices along the path and the factor
$(U_C + \epsilon I_4)$ contains some average direction bivector
$U_C$ for the loop, which is assumed to be close to planar.
Then for sufficiently strong coupling one can show that one naturally
obtains an area law, which here we express as
\begin{equation}
W ( C ) \, \simeq \, \exp ( - \, A_C / {\xi }^2 )
\label{eq:expdecay}
\end{equation}
where $\xi $ is the gravitational correlation length.
The appearance of $\xi$ follows from dimensional arguments, given that the
correlation length is the only relevant length scale in the vicinity of the
fixed point; the results can thus be considered analogous to the well-known
behavior for the Wilson loop in non-abelian gauge theories [17].
In the actual calculation the rapid decay of the quantum gravitational Wilson
loop as a function of the area can be seen as a general consequence of the
assumed disorder in the uncorrelated fluctuations of the parallel transport
matrices ${\bf R}(s,s')$ at large $G$.
A careful identification of (a suitable trace of) Eq.~(\ref{eq:rotriem})
with the expression in
Eq.~(\ref{eq:expdecay}), and in particular the comparison of the two
area-dependent terms, then yields the following estimate
for the macroscopic, large scale, average curvature in the large $G$ limit
\begin{equation}
{\bar R} \sim 1 / \xi^2 \; ,
\end{equation}
where $\xi$ is the quantity in Eq.~(\ref{eq:m-cont}).
An equivalent way of phrasing the last result is the suggestion that $1 / \xi^2$,
where $\xi$ is the renormalization group invariant
gravitational correlation length of Eq.~(\ref{eq:m-cont}), should
be identified, up to a constant of proportionality of order one,
with the observationally determined, large scale cosmological constant $\lambda$.
\vskip 20 pt
\section{Nonperturbative Gravity}
\label{sec:numerical}
The exact evaluation of the lattice functional integral
for quantum gravity by numerical methods
allows one, in principle, to investigate a regime
which is generally inaccessible by perturbation theory: where
the coupling $G$ is strong and quantum fluctuations in the metric
are large.
The hope is, in the end, to make contact with the analytic
results obtained in the $2+\epsilon$
expansion, and determine which scenarios are physically
realized in the lattice regularized model.
The main question one would therefore like to answer is whether
there is any indication that the non-trivial ultraviolet fixed
point scenario is realized in the lattice theory, in four dimensions.
This would imply,
as in the non-linear sigma model and similar models, the existence of at least
two physically distinct phases, and associated non-trivial scaling dimensions.
A clear physical characterization of the two gravitational phases would
also allow one, at least in principle, to decide which phase,
if any, could be realized in nature.
As discussed below, the lattice continuum limit
is taken in the vicinity of the fixed point,
so close to it is perhaps the physically most relevant regime.
At the next level one would hope to be able to establish
a quantitative connection with the continuum
perturbative results, such
as the $2+\epsilon$ expansion discussed earlier.
Since the lattice cutoff and the method of dimensional regularization
cut the theory off in the ultraviolet in rather different
ways, one needs to compare universal quantities
which are {\it cutoff-independent}.
An example is the critical exponent $\nu$, as well as any other
non-trivial scaling dimension that might arise.
One should note that within the $2+\epsilon$ expansion only {\it one}
such exponent appears, to {\it all} orders in the loop
expansion, as $ \nu^{-1} = - \beta ' (G_c) $.
Therefore one central issue in the four-dimensional lattice regularized theory
is the value of the universal scaling exponent $\nu$ [10,18]
[in Eqs.~(\ref{eq:grun-cont1}), (\ref{eq:m-cont1}) and (\ref{eq:nueps}) ] .
The starting point is again the lattice regularized
path integral with action as in Eq.~(\ref{eq:latac}) and
measure as in Eq.~(\ref{eq:lattmeas}).
Among the simplest quantum mechanical averages that one can compute
is one associated with the local curvature,
\begin{equation}
{\cal R} (k) \; \sim \;
{ < \int d x \, \sqrt{ g } \, R(x) >
\over < \int d x \, \sqrt{ g } > } \;\; .
\end{equation}
But the curvature associated with this quantity is
one that would be detected when parallel-transporting
vectors around very small infinitesimal loops.
Furthermore when computing correlations in quantum gravity new subtle
issues arise, due to the fact that the physical distance between any two points $x$ and $y$
\begin{equation}
d(x,y \, \vert \, g) \; = \; \min_{\xi} \; \int_{\tau(x)}^{\tau(y)} d \tau \;
\sqrt{ \textstyle g_{\mu\nu} ( \xi )
{d \xi^{\mu} \over d \tau} {d \xi^{\nu} \over d \tau} \displaystyle }
\end{equation}
is a fluctuating function of the background metric $g_{\mu\nu}(x)$.
Consequently physical correlations have to be defined at fixed
geodesic distance $d$, as in the following connected correlation
between observables $O$
\begin{equation}
< \int d x \, \int d y \, \sqrt{g} \, O(x) \; \sqrt{g} \, O(y) \;
\delta ( | x - y | - d ) >_c \; .
\end{equation}
Based on general arguments
one expects such correlations to either follow a power law decay
at short distances, or an exponential decay characterized by a correlation
length $\xi$ at larger distances
\begin{equation}
< \sqrt{g} \; O(x) \; \sqrt{g} \; O(y) \; \delta ( | x - y | -d ) >_c
\; \; \mathrel{\mathop\sim_{d \; \gg \; \xi }} \;\;
e^{-d / \xi } \;\;\;\; .
\label{eq:rr-exp}
\end{equation}
In practice such correlations at fixed geodesic distance
are difficult to compute numerically, and therefore not the best route
to study the critical properties of the theory.
But scaling arguments allow one to determine the scaling
behavior of correlation functions from critical exponents
characterizing the singular behavior of the free energy
$F (k) = - (1/V) \ln Z $ and of various local averages in the vicinity of the critical point.
In general a divergence of the correlation length $\xi$
\begin{equation}
\xi (k) \; \mathrel{\mathop\sim_{ k \rightarrow k_c}} \; A_\xi \;
| k_c - k | ^{ -\nu }
\label{eq:xi-k}
\end{equation}
signals the presence of a phase transition, and leads to the appearance
of a non-analyticty in the free energy $F(k)$.
One way to determine $\nu$ is from the curvature fluctuation, for which
one can show
\begin{equation}
\chi_{\cal R} (k) \; \mathrel{\mathop\sim_{ k \rightarrow k_c}} \;
A_{ \chi_{\cal R} } \; | k_c - k | ^{ -(2- d \nu) } \;\;\;\; .
\label{eq:chising}
\end{equation}
From such averages and fluctutations one can therefore, in principle, extract
the correlation length exponent $\nu$ of
Eq.~(\ref{eq:xi-k}), without having to compute an invariant correlation
function at fixed geodesic distance.
In general for the measure in Eq.~(\ref{eq:lattmeas})
one finds a well behaved ground state only for $k < k_c $ [10].
The system then resides in the `smooth' phase, with an effective
dimensionality close to four.
On the other hand, for $k > k_c$ the curvature becomes very large
and the lattice collapses locally into degenerate configurations
with very long, elongated simplices.
This last phenomenon is usually interpreted as a lattice remnant of the conformal
mode instability of Euclidean gravity.
There are a number of ways by which the critical exponents can
be determined to some accuracy from numerical simulations,and it is beyond
the scope of this short review to go into more details; one obtains eventually
in $d=4$ $ k_c \simeq 0.0636 $ and $ \nu \simeq 0.335 $, which suggests
\begin{equation}
\nu = 1/3
\end{equation}
for pure quantum gravity in four dimensions [18].
Note that at the critical point the gravitational coupling
is not weak, since $G_c \approx 0.626 $ in units of the ultraviolet
cutoff.
Fig. 3 shows a comparison of the critical exponent $\nu$ obtained
by three independent methods, namely the original lattice result in $d=2,3,4$ [18],
the recent $2+\epsilon$ expansion
(to one and two loops) [12],
and the even more recent renormalization group truncation method [19-21].
\begin{center}
\epsfxsize=8cm
\epsfbox{exp.eps}
\end{center}
\noindent{\small Figure 3.
Universal renormalization group scaling exponent $1/\nu$ of
Eq.~(\ref{eq:beta-lin-g}), computed in the
lattice theory of gravity (large dots for $d=2$, $d=3$ and $d=4$,
and continous interpolating line) [18].
For comparison the $2+\epsilon$ result is shown
(two lower curves, to one (lower) and two
(upper) loops) [12], as well as the recent Einstein-Hilbert truncation result
(smaller dots and connecting dashed line) [21].
The abscissa is, for convenience, a variable related to the space-time
dimension $d$ through $z=(d-2)/(d-1)$, which maps
$d=\infty$ (where it is known that $\nu=0$) to $z=1$.
Note that for a scalar theory one has $1/ \nu=2$ for $d \geq 4$ [3]. }
\medskip
\label{fig:exp}
\section{Renormalization Group, Lattice Continuum Limit and the Running of $G$}
\label{sec:contlim}
The lattice theory points to the existence
of a phase transition in pure quantum gravity, with a divergent
correlation length in the vicinity of the critical point,
as in Eq.~(\ref{eq:xi-k}), which can be re-written
in terms of the inverse correlation length $m \equiv 1 / \xi$
\begin{equation}
\xi^{-1} \, = \, m \, = \, \Lambda \, A_m \, | \, k_c \, - \, k \, |^{ \nu } \;\; .
\label{eq:m-latt}
\end{equation}
In the above expression the correct dimensions have been restored,
by inserting explicitly on the r.h.s. the ultraviolet
cutoff $\Lambda$.
Here $k$ and $k_c$ are dimensionless quantities,
corresponding to bare microscopic couplings at the
cutoff scale, $k \equiv k (\Lambda) \equiv 1/ 8 \pi G(\Lambda) $.
$A_m$ is related to $A_\xi$ in Eq.~(\ref{eq:xi-k}) by $A_m = A_\xi^{-1}$.
It is worth pointing out that the above expression for $m (k) $ is identical in structure
to the $2+\epsilon$ result for continuum gravity, Eq.~(\ref{eq:m-cont1}).
Then the lattice continuum limit corresponds to a large cutoff limit,
taken at {\it fixed} $m$ or $\xi$,
\begin{equation}
\Lambda \rightarrow \infty \; ,
\;\;\;\; k \rightarrow k_c \; ,
\;\;\;\; m \; {\rm fixed} \; ,
\end{equation}
which shows that the continuum limit is in fact reached in the
vicinity of the ultraviolet fixed point, $ k \rightarrow k_c$.
In practice, since the cutoff ultimately determines
the physical value of Newton's constant $G$, the cutoff $\Lambda$
cannot be taken to $\infty$, and it persists as a fundamental scale
in the theory.
A very large value will suffice, $\Lambda^{-1} \sim 10^{-33} cm$,
for which it will still be true that $ \xi \gg \Lambda^{-1}$, which
is all that is required for the continuum limit.
In order to discuss the renormalization group behavior
of the coupling in the lattice theory it is convenient
to re-write the result of Eq.~(\ref{eq:m-latt})
directly in terms of Newton's constant $G$ as
\begin{equation}
m \, = \, \Lambda \, \left ( { 1 \over c_0 } \right )^\nu \,
\left [ { G ( \Lambda ) \over G_c } - 1 \right ]^ \nu
\; ,
\label{eq:m-latt1}
\end{equation}
with the dimensionless constant $c_0$ related to
$A_m$ by $A_m = 1 / (c_0 k_c)^\nu $.
The above expression only involves the dimensionless
ratio $G(\Lambda)/G_c$, which is the only relevant
quantity here.
From the knowledge of the dimensionless constant
$A_m$ in Eq.~(\ref{eq:m-latt})
one can estimate from first principles the value of $c_0$ in
Eq.~(\ref{eq:m-latt1}) and later in Eq.~(\ref{eq:grun-k}).
Lattice results for the correlation functions at fixed geodesic distance
give a value for $A_m \approx 0.72 $ with a significant uncertainty,
which, when combined with the values $k_c \simeq 0.0636 $ and
$ \nu \simeq 0.335$ given above, gives
$c_0 = 1 /( k_c \, A_m^{1/\nu}) \simeq 42$.
Then the renormalization group invariance of
$m=\xi^{-1}$ requires that the running gravitational
coupling $G(\mu)$ varies in the vicinity of the
fixed point in accordance with the above equation, with
$\Lambda \rightarrow \mu$, where $\mu$ is an arbitrary momentum scale,
\begin{equation}
m \, = \, \mu \, \left ( { 1 \over c_0 } \right )^\nu \,
\left [ { G ( \mu ) \over G_c } - 1 \right ]^ \nu \; .
\label{eq:m-mu}
\end{equation}
The latter is equivalent to the renormalization group
invariance requirement
\begin{equation}
\mu \, { d \over d \, \mu } \, m ( \mu , G( \mu ) ) \, = \, 0
\label{eq:m-rg-latt}
\end{equation}
provided $G(\mu)$ is varied in a specific way.
Thus Eq.~(\ref{eq:m-rg-latt}) can be used to obtain
a Callan-Symanzik $\beta$-function for the coupling
$G(\mu)$ in units of the ultraviolet cutoff,
\begin{equation}
\mu \, { \partial \over \partial \, \mu } \, G ( \mu ) \; = \;
\beta ( G ( \mu ) ) \;\; ,
\label{eq:beta-g-mu}
\end{equation}
with $\beta (G)$ given in the vicinity of the non-trivial fixed point,
from Eq.~(\ref{eq:m-mu}), by
$ \beta (G ) \mathrel{\mathop\sim_{ G \rightarrow G_c}} - \, { 1 \over \nu } \, ( G- G_c ) $.
Or one can obtain the scale dependence of the gravitational
coupling directly from Eq.~(\ref{eq:m-mu}), which then gives
\begin{equation}
G( \mu ) \; = \;
G_c \left [ 1 \, + \, c_0 (m^2 / \mu^2 )^{1 / 2 \nu } \, + \,
O ( \, ( m^2 / k^2 )^{1 \over \nu} ) \, \right ]
\label{eq:grun-k}
\end{equation}
in the physical anti-screening phase.
Again, this last expression can be compared directly to the lowest order
$2+\epsilon$ result of Eq.~(\ref{eq:grun-cont1}).
The physical dimensions of $G$ can be restored, by multiplying the
above expression on both sides by the ultraviolet cutoff $\Lambda$,
if one so desires.
Physically the above lattice result implies
anti-screening: the gravitational coupling $G$ increases slowly with distance.
\vskip 20 pt
\section{Curvature Scales and Gravitational Condensate}
\label{sec:curvature}
The renormalization group running of $G (\mu)$ in Eq.~(\ref{eq:grun-k})
involves an invariant scale $\xi=1/m$.
At first it would seem that such a scale could take any value, including
a very small one based on the naive estimate $\xi \sim l_P$ - which would
then preclude any observable quantum effects in the foreseeable future.
But the results from the gravitational Wilson loop at strong coupling
would suggest otherwise, namely that the non-perturbative scale $\xi$
is in fact related to macroscopic {\it curvature}.
From astrophysical observation the average curvature is very small [22],
so one would conclude that $\xi$ has to be very
large and possibly macroscopic,
\begin{equation}
\lambda_{obs} \; \simeq \; { 1 \over \xi^2 }
\label{eq:xi_lambda}
\end{equation}
with $\lambda_{obs}$ the observed small but non-vanishing scaled cosmological constant.
A further indication that the identification of the observed cosmological
constant with a mass-like - and therefore renormalization group invariant - term
might make sense beyond the weak field limit can be seen, for example,
by comparing the structure of the three classical field equations
\begin{eqnarray}
R_{\mu\nu} \, - \, \half \, g_{\mu\nu} \, R \, + \, \lambda \, g_{\mu\nu} \;
& = & \; 8 \pi G \, T_{\mu\nu}
\nonumber \\
\partial^{\mu} F_{\mu\nu} \, + \, \mu^2 \, A_\nu \,
& = & \; 4 \pi e \, j_{\nu}
\nonumber \\
\partial^{\mu} \partial_{\mu} \, \phi \, + \, m^2 \, \phi \;
& = & \; {g \over 3!} \, \phi^3
\label{eq:masses}
\end{eqnarray}
for gravity, QED (made massive via the Higgs mechanism) and a self-interacting scalar
field, respectively.
Nevertheless it seems so far that the strongest
argument suggesting the identification of the scale $\xi$ with $\lambda$ is derived
from the calculation of the gravitational Wilson loop at strong [16].
This relationship, taken at face value, implies a very large, cosmological value
for $\xi \sim 10^{28} cm$, given the present bounds on $\lambda_{phys}$.
Thus a set of modified Einstein equations, incorporating the
quantum running of $G$, would read
\begin{equation}
R_{\mu\nu} \, - \, \half \, g_{\mu\nu} \, R \, + \, \lambda \, g_{\mu\nu}
\; = \; 8 \pi \, G(\mu) \, T_{\mu\nu}
\label{eq:field0}
\end{equation}
with $\lambda \simeq 1 / \xi^2 $, and $G(\mu)$ on the r.h.s.
scale-dependent, in accordance with
Eq.~ (\ref{eq:grun-k}).
The precise meaning of $G(\mu)$ in a covariant framework
is given below.
\vskip 20 pt
\section{Effective Covariant Field Equations }
\label{sec:effective}
The result of Eq.~(\ref{eq:grun-k}) implies a
running gravitational coupling in the vicinity of the
ultraviolet fixed point, with $m=1/\xi$, $c_0 > 0$ and $\nu \simeq 1/3$.
Since $\xi$ is expected to be very large,
the quantity $G_c$ in the above expression should now
be identified with the laboratory scale value
$ \sqrt{G_c} \sim 1.6 \times 10^{-33} cm$.
The effective interaction in real space is then obtained by Fourier transform,
but since the above expression is singular as $k^2 \rightarrow 0$,
the infrared divergence needs to be regulated, which can be achieved
by utilizing as the lower limit of momentum integration
$m=1/\xi$.
A properly infrared regulated version of the above would read
\begin{equation}
G(k^2) \; \simeq \; G_c \left [ \; 1 \,
+ \, c_0 \left ( { m^2 \over k^2 \, + \, m^2 } \right )^{1 \over 2 \nu} \,
+ \, \dots \; \right ] \; .
\label{eq:grun-k-reg}
\end{equation}
Then at very large distances $r \gg \xi$ the gravitational coupling is
expected to approach the finite value $G_\infty = ( 1 + c_0 + \dots ) \, G_c $.
The first step in analyzing the consequences of a running of $G$
is to re-write the expression for $G(k^2)$ in a coordinate-independent
way, for example by the use of a non-local Vilkovisky-type effective action.
Since in going from momentum to position space one
usually employs $k^2 \rightarrow - \Box$,
to obtain a quantum-mechanical running of the gravitational
coupling one has to make the replacement
$ G \;\; \rightarrow \;\; G( \Box ) $.
Therefore from Eq.~(\ref{eq:grun-k}) one obtains
\begin{equation}
G( \Box ) \, = \, G_c \left [ \; 1 \,
+ \, c_0 \left ( { 1 \over \xi^2 \Box } \right )^{1 / 2 \nu} \,
+ \, \dots \, \right ] \; ,
\label{eq:grun-box}
\end{equation}
and the running of $G$ is expected to lead to
a non-local gravitational action, for example of the form
\begin{equation}
I_{eff} \; = \; { 1 \over 16 \pi G } \int dx \sqrt{g} \,
\left [ 1 \, - \, c_0 \,
\left ( {1 \over \xi^2 \Box } \right )^{ 1 / 2 \nu} \,
+ \dots \right ] R \; \; .
\label{eq:ieff_sr}
\end{equation}
Due to the appearance of a fractional exponent,
the covariant operator appearing in the above expression
has to be suitably defined by analytic continuation.
The latter can be done, for example, by computing $\Box^n$ for positive
integer $n$ and then analytically continuing to $n \rightarrow -1/2\nu$.
Alternatively, had one not considered the action of Eq.~(\ref{eq:ieff_sr})
as a starting point for
constructing the effective theory, one would naturally be led
(as suggested by Eq.~(\ref{eq:grun-box}))
to consider instead the following effective field equation
\begin{equation}
R_{\mu\nu} \, - \, \half \, g_{\mu\nu} \, R \, + \, \lambda \, g_{\mu\nu}
\; = \; 8 \pi \, G ( \Box ) \, T_{\mu\nu} \; ,
\label{eq:field1}
\end{equation}
the argument again being the replacement
$G \, \rightarrow \, G(\Box) $ in the classical Einstein field equations.
Being manifestly covariant, these expressions at least satisfy some
of the requirements for a set of consistent field equations
incorporating the running of $G$.
The effective field equations of Eq.~(\ref{eq:field1}) can in fact be re-cast in a form
very similar to the classical field equations but
with a $ {\tilde T_{\mu\nu}} \, = \, \left [ G ( \Box ) / G \right ] \, T_{\mu\nu}$
defined as an effective, or gravitationally dressed, energy-momentum tensor.
Just like the ordinary Einstein gravity case,
in general ${\tilde T_{\mu\nu}}$ might not be covariantly conserved a priori,
$\nabla^\mu \, {\tilde T_{\mu\nu}} \, \neq \, 0 $, but ultimately the
consistency of the effective field equations demands that it
be exactly conserved, in consideration of the Bianchi identity satisfied
by the Riemann tensor.
In this picture therefore the running of $G$ can be viewed as contributing to a sort of
"vacuum fluid", introduced in order to to account for the gravitational vacuum polarization
contribution.
\vskip 20 pt
\section{Static Isotropic Solutions}
\label{sec:static}
One can show that the quantum correction due
to the running of $G$ can be described - at least in the non-relativistic
limit of Eq.~(\ref{eq:grun-k-reg}) when applied to Poisson's equation -
in terms of a vacuum energy density $\rho_m(r)$, distributed around
the static source of strength $M$ in accordance with
\begin{equation}
\rho_m (r) \; = \; { 1 \over 8 \pi } \, c_{\nu} \, c_0 \, M \, m^3 \,
( m \, r )^{ - {1 \over 2} (3 - {1 \over \nu}) }
\, K_{ {1 \over 2} ( 3 - {1 \over \nu} ) } ( m \, r )
\label{eq:rho_vac}
\end{equation}
and with $ c_{\nu} \; \equiv \; 2^{ {1 \over 2} (5 - {1 \over \nu}) } /
\sqrt{\pi} \, \Gamma( {1 \over 2 \, \nu} ) $,
and
\begin{equation}
4 \, \pi \, \int_0^\infty \, r^2 \, d r \, \rho_m (r) \; = \; c_0 \, M \;\; .
\label{eq:rho_vac2}
\end{equation}
More generally in the fully relativistic case,
after solving the covariant effective field equations
with $G(\Box)$ for $\nu =1/3$ one finds in Schwarzschild coordinates, and
in the limit $r \gg 2 M G $,
\begin{equation}
A^{-1} (r) \; = \; = \; B (r) \; = \; 1 \, - { 2 \, M \, G \over r } \, + \,
{4 \, c_0 \, M \, G \, m^3 \over 3 \, \pi } \, r^2 \, \ln \, ( m \, r )
\, + \, \dots
\label{eq:a_small_r3}
\end{equation}
The last expressions for $A(r)$ and $B(r)$ are therefore consistent with a
gradual slow increase in $G$ with distance, in accordance with the formula
\begin{equation}
G \; \rightarrow \; G(r) \; = \;
G \, \left ( 1 \, + \,
{ c_0 \over 3 \, \pi } \, m^3 \, r^3 \, \ln \, { 1 \over m^2 \, r^2 }
\, + \, \dots
\right )
\label{eq:g_small_r3}
\end{equation}
in the regime $r \gg 2 \, M \, G$.
The last result is in some ways reminiscent of the QED small-$r$ result
\begin{equation}
Q \; \rightarrow \; Q(r) \; = \; Q \, \left ( 1 \, + \,
{\alpha \over 3 \, \pi } \, \ln { 1 \over m^2 \, r^2 } \, + \, \dots
\right )
\label{eq:qed_s}
\end{equation}
In the gravity case, the correction vanishes as $r$ goes to zero: in this
limit one is probing the bare mass, unobstructed by its virtual graviton cloud.
In some ways the running $G$ term acts as a local cosmological
constant term, for which the
$r$ dependence of the vacuum solution for small $r$ is fixed by the nature
of the Schwarzschild solution with a cosmological constant term.
One could in fact wonder what these solutions might look like in $d$
dimensions, and after some straightforward calculations one finds
that in $d \ge 4 $ space-time dimensions a solution to the effective field equations
can only be found if in Eq.~(\ref{eq:grun-box}) $\nu=1/(d-1)$ exactly [23].
\vskip 20 pt
\section{Cosmological Solutions}
\label{sec:cosm}
A scale dependent Newton's constant is expected to lead to small modifications
of the standard cosmological solutions to the Einstein field equations.
Here I will summarize what modifications are
expected from the effective field equations on the basis of $G(\Box)$,
as given in Eq.~(\ref{eq:grun-box}), which itself originates in
Eqs.~(\ref{eq:grun-k-reg}) and (\ref{eq:grun-k}).
The starting point are the quantum effective field equations
of Eq.~(\ref{eq:field1}),
with $G(\Box)$ defined in Eq.~(\ref{eq:grun-box}).
In the Friedmann-Robertson-Walker (FRW) framework these are
applied to the standard homogeneous isotropic metric
\begin{equation}
d \tau^2 \; = \; dt^2 - a^2(t) \left \{ { dr^2 \over 1 - k\,r^2 }
+ r^2 \, \left( d\theta^2 + \sin^2 \theta \, d\varphi^2 \right) \right \} \; .
\end{equation}
It should be noted that there are in fact {\it two} related quantum contributions to the
effective covariant field equations.
The first one arises because of the presence of a non-vanishing
cosmological constant $\lambda \simeq 1 / \xi^2 $, caused by the
non-perturbative vacuum condensate of Eq.~(\ref{eq:xi_lambda}).
As in the case of standard FRW cosmology, this is expected to be
the dominant contributions at large times $t$, and gives an exponential
(for $\lambda>0$), or cyclic (for $\lambda < 0$) expansion of the scale factor.
The second contribution arises because of the explicit running of $G (\Box)$ in the
effective field equations.
The next step is therefore a systematic examination of the nature of
the solutions to the full effective field equations,
with $G ( \Box )$ involving the relevant covariant d'Alembertian operator
\begin{equation}
\Box \; = \; g^{\mu\nu} \nabla_\mu \nabla_\nu
\end{equation}
acting on second rank tensors as in the case of $T_{\mu\nu}$,
\begin{eqnarray}
\nabla_{\nu} T_{\alpha\beta} \, = \, \partial_\nu T_{\alpha\beta}
- \Gamma_{\alpha\nu}^{\lambda} T_{\lambda\beta}
- \Gamma_{\beta\nu}^{\lambda} T_{\alpha\lambda} \, \equiv \, I_{\nu\alpha\beta}
\nonumber
\end{eqnarray}
\begin{equation}
\nabla_{\mu} \left( \nabla_{\nu} T_{\alpha\beta} \right)
= \, \partial_\mu I_{\nu\alpha\beta}
- \Gamma_{\nu\mu}^{\lambda} I_{\lambda\alpha\beta}
- \Gamma_{\alpha\mu}^{\lambda} I_{\nu\lambda\beta}
- \Gamma_{\beta\mu}^{\lambda} I_{\nu\alpha\lambda} \; .
\end{equation}
To start the process, one assumes for example that $T_{\mu\nu}$ has a perfect fluid form,
for which one obtains the action of $\Box^n$ on $T_{\mu\nu}$, and
then analytically continues to negative fractional values of $n = -1/2 \nu $.
Even in the simplest case, with $G(\Box)$ acting on a {\it scalar} such as the trace
of the energy-momentum tensor $T^\lambda_{\; \lambda}$, one finds for the choice
$ \rho(t) = \rho_0 \, t^{\beta} $ and $ a(t) = r_0 \, t^{\alpha} $
the rather unwieldy expression
\begin{equation}
\Box^n \left[ - \rho (t) \right] \rightarrow 4^n \,
\left( - 1 \right)^{n+1} { \Gamma \left( {\beta \over 2} + 1 \right) \,
\Gamma \left( {\beta + 3 \, \alpha + 1 \over 2} \right) \over
\Gamma \left( {\beta \over 2} + 1 - n \right) \,
\Gamma \left( {\beta + 3\, \alpha + 1 \over 2} - n \right) } \, \rho_0 \, t^{\beta - 2 n} \; ,
\end{equation}
with integer $n$ then analytically continued to $n \rightarrow - {1 \over 2 \, \nu}$, with
$\nu=1/3$.
A more general calculation shows that a non-vanishing pressure contribution is generated in
the effective field equations, even if one initially assumes a pressureless fluid, $p(t)=0$.
After a somewhat lengthy computation one obtains for a universe filled with non-relativistic
matter ($p$=0) the following set of effective Friedmann equations, namely
\begin{eqnarray}
{ k \over a^2 (t) } \, + \,
{ \dot{a}^2 (t) \over a^2 (t) }
& \; = \; & { 8 \pi G(t) \over 3 } \, \rho (t) \, + \, { 1 \over 3 \, \xi^2 }
\nonumber \\
& \; = \; & { 8 \pi G \over 3 } \, \left [ \,
1 \, + \, c_t \, ( t / \xi )^{1 / \nu} \, + \, \dots \, \right ] \, \rho (t)
\, + \, \third \, \lambda
\label{eq:fried_tt}
\end{eqnarray}
for the $tt$ field equation, and
\begin{eqnarray}
{ k \over a^2 (t) } \, + \, { \dot{a}^2 (t) \over a^2 (t) }
\, + \, { 2 \, \ddot{a}(t) \over a(t) }
& \; = \; & - \, { 8 \pi G \over 3 } \, \left [ \, c_t \, ( t / \xi )^{1 / \nu}
\, + \, \dots \, \right ] \, \rho (t)
\, + \, \lambda
\label{eq:fried_rr}
\end{eqnarray}
for the $rr$ field equation.
In the above equations the running of $G$ appropriate for
the Robertson-Walker metric is given by
\begin{equation}
G(t) \; = \; G \, \left [ \; 1 \, + \, c_t \,
\left ( { t \over \xi } \right )^{1 / \nu} \, + \, \dots \, \right ] \; ,
\label{eq:grun_frw}
\end{equation}
with $c_t$ of the same order as $c_0$ of Eq.~(\ref{eq:grun-k}).
Note that it is the running of $G$ that induces an effective pressure term in the second
($rr$) equation, corresponding to the presence of a relativistic fluid
arising from the vacuum polarization contribution.
The second important feature of the new equations is an additional power-law
acceleration contribution, in addition to the standard one due to $\lambda$.
\vskip 20 pt
\section{Quantum Gravity and Cosmological Density Perturbations}
\label{sec:densa}
Besides the cosmic scale factor evolution and the static isotropic solutions
just discussed, the running of $G(\Box)$ also affects
the nature of matter density perturbations on very large scales.
In discussing these effects, it is customary to introduce a perturbed metric of
the form
\begin{equation}
{d\tau}^2 = {dt}^2 - a^2 \left( \delta_{ij} + h_{ij} \right) dx^i dx^j
\label{eq:pert-metric}
\end{equation}
with $a(t)$ the unperturbed scale factor and $ h_{ij} (\vec{x},t)$ a small
metric perturbation.
The next step is to determine the effects of the running of $G$ on the relevant
matter and metric perturbations, again by the use of the modified field equations.
For sufficiently small perturbations, one can expand $G(\Box)$ appearing in the
effective covariant field equations in powers of the metric perturbation $h_{ij} $ as
\begin{equation}
G(\Box) \; = \; G_0 \,
\left[
1 + \, { c_0 \over \xi^{1 / \nu} } \, \left( { 1 \over \Box^{(0)} } \right)^{1 / 2 \nu}
\, \left( 1 - {1 \over 2 \nu} \, {1 \over \Box^{(0)}} \, \Box^{(1)} (h) + \dots \right)
\right]
\label{eq:gbox_h}
\end{equation}
It is also customary to expand the density, pressure and metric trace perturbation
modes in spatial Fourier components
\begin{equation}
\delta \rho (\vec{x},t) = \delta \rho (t) \, e^{i \, \vec{k}\,\cdot \, \vec{x}}
\;\;\;\;
\delta p (\vec{x},t) = \delta p (t) \, e^{i \, \vec{k}\,\cdot \, \vec{x}}
\;\;\;\;
h (\vec{x},t) = h(t)\, e^{i \, \vec{k}\,\cdot \, \vec{x}} \; .
\end{equation}
Normally the Einstein field equations
$ R_{\mu\nu} - {1 \over 2} g_{\mu \nu} R + \lambda \, g_{\mu \nu} = 8 \pi G \, T_{\mu \nu} \, $
are given to first order in the small perturbations by
\begin{eqnarray}
{\dot{a} (t) \over a (t)}\, \dot{h} (t) & = & 8 \pi G \rho (t) \, \delta (t)
\nonumber \\
\ddot{h} (t) + 3 \, {\dot{a} (t) \over a (t)}\, \dot{h} (t)
& = & - 24 \pi \, G \, w \, \rho(t) \delta (t)
\end{eqnarray}
with $\delta (t) = \delta \rho (t) / \rho (t) $ and $w=0$ for non-relativistic matter,
yielding then a single equation for the trace of the metric perturbation $h(t)$,
\begin{equation}
\ddot{h} (t) + 2 \, {\dot{a} (t) \over a (t)} \dot{h} (t) \; = \;
- 8 \pi G ( 1 + 3\, w ) \rho(t) \delta (t) \; .
\end{equation}
Combined with the first order energy conservation
$ - {1 \over 2}\, \left( 1 + w \right)\, h (t) \; = \; \delta (t) $, this then gives
a single equation for the density contrast $\delta(t)$,
\begin{equation}
\ddot{\delta} (t) + 2 \, {\dot{a} \over a} \, \dot{\delta} (t) - 4 \pi \, G \, \rho(t) \, \delta(t) = 0 \; .
\end{equation}
In the case of a running $G(\Box)$ these equations need to be re-derived
from the effective covariant field equations, and lead to several
additional terms not present at the classical level [23].
In other words, the correct field equations for a running $G$ are not given simply by
a naive replacement $G \rightarrow G(t) $, which would lead to incorrect results,
and violate general covariance.
It is common practice at this point to write an equation for the density contrast
$\delta(a)$ as a function not of $t$, but of the scale factor $a(t)$, by utilizing the identities
\begin{equation}
\dot{f}(t) = a \, H \, {\partial f (a) \over \partial a}
\end{equation}
\begin{equation}
\ddot{f} (t) = {a}^2 \, {H}^2 \left( {\partial \ln H \over \partial a} + {1 \over a} \right) \,{\partial f (a) \over \partial a} + {a}^2 \, {H}^2 \, {{\partial}^2 f (a) \over \partial {a}^2}
\end{equation}
where $ f $ is any function of $ t $, and $ H \equiv \dot{a} (t) / a (t) $ is the Hubble constant.
This last quantity can then be obtained from the zero-th order $tt$ field equation
\begin{equation}
3 \, \left( { \dot{a} \over a } \right)^2 = 8 \, \pi \, G_0 \, \rho + \lambda
\end{equation}
re-written in terms of $H(a)$ as
\begin{equation}
H^2 (a) \equiv \left( {\dot{a} \over a} \right)^2
= \left( \dot{z} \over 1 + z \right)^2
= H_0^2 \left[ \Omega \, \left( 1 + z \right)^3 + \Omega_R \, \left( 1 + z \right)^2 + \Omega_{\lambda} \right]
\end{equation}
with $ a = {1 \over 1 + z} $ where $ z $ is a red shift, $ H_0 $ the Hubble constant evaluated today,
and $ \Omega $ the matter (baryonic and dark) density, $ \Omega_R $ the space curvature
contribution corresponding to a curvature $ k $ term, and $\Omega_{\lambda} $the dark energy part,
\begin{equation}
\Omega_{\lambda} \equiv {\lambda \over 3 \, H^2}
\;\;\;\;\;\;
\Omega \equiv { 8 \, \pi \, G_0 \, \rho \over 3 \, H^2 }
\;\;\;\;\;\; {\rm with } \;\;\;\;
\Omega + \Omega_{\lambda} = 1
\end{equation}
After introducing the parameter $\theta$ as the cosmological constant fraction
\begin{equation}
\theta \equiv \Omega_{\lambda} \, \left({8 \, \pi \, G \, \rho \over 3 \, H^2}\right)^{-1}
\; = \; { \Omega_{\lambda} \over \Omega}
\; = \; { 1 - \Omega \over \Omega }
\end{equation}
one then obtains an equation for the density contrast $\delta (a)$ in the normal
(i.e. non-running $G$) case
\begin{equation}
{\partial^2 \delta(a) \over \partial a^2 } +
\left[ {\partial \ln H(a) \over \partial a} +
{3 \over a } \right] \, {\partial \delta(a) \over \partial a} -
4\, \pi \, G_0 \,{1 \over a^2 H(a)^2}\, \rho (a) \, \delta(a) = 0
\end{equation}
with growing solution
\begin{equation}
\delta_0 (a) \; \sim \; a \cdot {} _2F_1 \left({1 \over 3}, 1; {11 \over 6}; - \, a^3 \, \theta \right)
\end{equation}
where $ _2F_1 $ is a hypergeometric function.
To determine the quantum correction to $\delta (a)$ originating from
$G(\Box)$ in Eq.~(\ref{eq:grun-box}), one sets
\begin{equation}
\delta(a) \; = \; \delta_0 (a) \left [ 1 + c_0 \, f(a) \right ] \; ,
\end{equation}
and then uses this linear Ansatz to find the form of $ f(a) $ to lowest order in $c_0$.
The correction is unambiguously determined from the field equations with a running
Netwon constant $ G(\Box) $, but here the running of $G$ (due to the choice of variables)
naturally takes on the form
\begin{equation}
G(a) = G_0 \left [ 1 + c_a \, \left( {a \over a_{0} } \right)^{\gamma} + \dots \right ]
\label{eq:grun-a}
\end{equation}
with a scale factor $a \approx a_0$ corresponding to a mode for which $k \approx \pi / \xi$
(thus $a_0$ is not necessarily identified with the scale factor "today").
Furthermore one has $ \gamma = 3 / 2 \, \nu $ with $ \nu = 1 / 3 $, as
determined from lattice gravity in four dimensions.
What then remains to be done is to compute the growth index
$ f \equiv \partial \ln \delta / \partial \ln a $, and from it
the growth index exponent $ \gamma $ defined through
$ f = \Omega^{\gamma} $ [24].
Ultimately one is interested in the value for this quantities in the vicinity
of a current matter fraction $\Omega \approx 0.25 $.
For a constant (i.e. not scale dependent) Newton's constant one has the
well know result $ f = 0.6028 $ and exponent $ \gamma = 0.5562 $.
An explicit calculation in the presence of a running $G$
and for a matter fraction $\Omega \approx 0.25$ gives,
to lowest linear order in the small quantum correction $c_0$ [25],
\begin{equation}
\gamma \; = \; 0.5562 - c_q \, c_a \; .
\end{equation}
Here $c_a$ is the coefficient of the quantum correction in the expression for $G(a)$,
and therefore fixed by the underlying lattice gravity calculations, and
$c_q$ an explicitly calculable numerical constant that comes out of the solution
of the full effective covariant field equations for $\delta(a)$.
The perturbed RW metric is well suited for discussing matter perturbations, but
occasionally one finds it more convenient to use a different metric parametrization,
such as the one derived from the conformal Newtonian (cN) gauge line element
\begin{equation}
d \tau^2 \; = \; a^2 (t) \left \{ (1+2 \, \psi ) \, dt^2 \; - \; (1-2 \, \phi ) \, \delta_{ij} \, dx^i dx^j \right \}
\label{eq:cn-gauge}
\end{equation}
with Conformal Newtonian potentials $\psi (\vec{x},t)$ and $\phi(\vec{x},t)$.
In this gauge, and in the absence of a $G(\Box)$, the unperturbed equations are
\begin{eqnarray}
\left ( { \dot{a} \over a } \right )^2 & = & { 8 \pi \over 3 } \, G \, a^2 \, \bar{\rho}
\nonumber \\
{ d \over d t } \left ( { \dot{a} \over a } \right ) & = &
- { 4 \pi \over 3 } \, G \, a^2 \, ( \bar{\rho} + 3 \bar{p} ) \; ,
\label{eq:cn_field_zeroth}
\end{eqnarray}
in the absence of spatial curvature ($k=0$).
In the presence of a running $G$ these again need to be modified, in accordance
with Eqs.~(\ref{eq:fried_rr}), (\ref{eq:fried_tt}) and (\ref{eq:grun_frw}).
A cosmological constant can be conveniently included in the $\bar{\rho}$ and $\bar{p}$,
with $\bar{\rho}_\lambda = \lambda/ 8 \pi G = - \bar{p}_\lambda $.
In this gauge scalar perturbations are characterized by Fourier modes
$\psi(\vec{k},t)$ and $\phi (\vec{k},t)$, and the first order Einstein field equations
in the absence of $G(\Box)$ read [26]
\begin{eqnarray}
k^2 \, \phi \, + \, 3 \, { \dot{a} \over a } \, \left ( \dot{\phi} \, + \, { \dot{a} \over a } \, \psi \right )
& = &
4 \pi \, G \, a^2 \, \delta T^0_{\;\; 0}
\nonumber \\
k^2 \, \left ( \dot{\phi} \, + \, { \dot{a} \over a } \, \psi \right )
& = &
4 \pi \, G \, a^2 \, ( \bar{\rho} + \bar{p} ) \, \theta
\nonumber \\
\ddot{\phi} \, + \, { \dot{a} \over a } \left ( 2 \dot{\phi} \, + \dot{\psi} \right )
+ \left ( 2 \, { \ddot{a} \over a } \, - \, { \dot{a}^2 \over a^2 } \right ) \, \psi
\, + \, { k^2 \over 3 } \, ( \phi \, - \, \psi )
& = &
{ 4 \pi \over 3 } \, G \, a^2 \, \delta T^i_{\;\; i}
\nonumber \\
k^2 \, ( \phi \, - \, \psi )
& = &
12 \pi \, G \, a^2 \, ( \bar{\rho} + \bar{p} ) \, \sigma \;\;\;\;
\label{eq:cn_field_pert}
\end{eqnarray}
where the perfect fluid energy-momentum tensor is given to linear order in the
perturbations
$\delta \rho = \rho - \bar{\rho} $ and $\delta p = p - \bar{p} $
by
\begin{eqnarray}
T^0_{\;\; 0} & = & - ( \bar{\rho} \, + \, \delta \rho )
\nonumber \\
T^0_{\;\; i} & = & ( \bar{\rho} \, + \, \bar{p} ) \, v_i \; = \; - T^i_{\;\; 0}
\nonumber \\
T^i_{\; j} & = & ( \bar{p} \, + \, \delta p ) \, \delta^i_{\; j} \, + \, \Sigma^i_{\; j}
\;\;\;\; \Sigma^i_{\; i}=0
\end{eqnarray}
and one has allowed for an anisotropic shear perturbation $\Sigma^i_{\; j}$ to
the perfect fluid form $T^i_{\; j}$.
The two quantities $\theta$ and $\sigma$ are commonly defined by
\begin{equation}
( \bar{\rho} \, + \, \bar{p} ) \, \theta \; \equiv \; i \, k^j \, \delta T^0_{\; j}
\;\;\;\;
( \bar{\rho} \, + \, \bar{p} ) \, \sigma \; \equiv \;
- ( \hat{k_i} \hat{k_j} - { 1 \over 3} \delta_{ij} ) \Sigma^i_{\; j}
\end{equation}
with $\Sigma^i_{\; j} \equiv T^i_{\, j} - \delta^i_{\; j} T^k_{\; k} /3 $ the traceless
component of $T^i_{\; j}$.
For a perfect fluid $\theta$ is the divergence of the fluid velocity, $\theta = i k^j v_j$,
with $v^j = d x^j / dt $ the small velocity of the fluid.
The field equations imply, by consistency, the covariant energy momentum conservation law
\begin{eqnarray}
\dot{\delta} & = & - (1+w) \, (\theta - 3 \dot{\phi} ) - 3 \, { \dot{a} \over a } \,
\left ( { \delta p \over \delta \rho } - w \right ) \delta
\nonumber \\
\dot{\theta} & = & - { \dot{a} \over a } \, ( 1 - 3 w ) \, \theta -
{ \dot{w} \over 1+w } \, \theta
+ { 1 \over 1+ w } { \delta p \over \delta \rho } \, k^2 \delta - k^2 \sigma + k^2 \psi
\end{eqnarray}
and relate the matter fields $\delta$, $\sigma$ and $\theta$ to the metric perturbations
$\phi$ and $\psi$.
where $\delta $ is the matter density contrast $\delta = \delta \rho / \rho $, and
$w$ is the equation of state parameter $w= p/\rho $.
In the presence of a $G(\Box)$ the above equations need to be re-derived and amended [25],
starting from the covariant field equations of Eq.~(\ref{eq:field1}) in the cN gauge
of Eq.~(\ref{eq:cn-gauge}), with zero-th order modified field equations as
in Eqs.~(\ref{eq:fried_tt}) and (\ref{eq:fried_rr}), using the expansion
for $G(\Box)$ given in Eq.~(\ref{eq:gbox_h}), but now in terms of the new cN
gauge potentials $\phi$ and $\psi$.
One key question is then the nature of the vacuum-polarization induced
anisotropic shear perturbation correction $\Sigma^i_{\; j}$
appearing in the covariant effective field equations analogous to
Eqs.~(\ref{eq:cn_field_pert}),
but derived with a $G(\Box)$.
In particular one would expect the quantum correction to the energy momentum
tensor appearing on the r.h.s. of Eq.~(\ref{eq:field1}) to contribute new terms
to the last of Eqs.~(\ref{eq:cn_field_pert}),
which could then account for a non-zero stress $\sigma$, and thus for a small
deviation from the classical result for a perfect fluid, $\phi = \psi$.
\vskip 20pt
{\bf Acknowledgements}
I wish to thank the organizers of the MG-12 conference, and in particular
Thibault d'Amour, Gabriele Veneziano and Remo Ruffini, for the opportunity
of attending the conference,
and numerous discussions relevant to the work presented here.
I have also benefitted from conversations with Alexey Vikhlinin regarding
the past and future observational constraints on the density contrast exponents from
the study of large galactic clusters.
I furthermore wish to express my gratitude to my collaborator Ruth M. Williams and
my student Reiko Toriumi, for contributing to the work described here.
Finally I thank Hermann Nicolai and the
Max Planck Institut f\" ur Gravitationsphysik (Albert-Einstein-Institut)
in Berlin for hospitality when part of the work
presented in this review was performed, and to the NSF for financial support.
\vskip 20pt
|
\section{Introduction}
Thirty years after the invention of the scanning tunneling microscope (STM) by Binnig and Rohrer\,\cite{1} the interest in this technique is still increasing. STM and scanning tunneling spectroscopy (STS) measurements on superconductors are one example of this emerging field. High-temperature superconductors have been successfully investigated by STM and STS\,\cite{2,3}. For the study of heavy-fermion superconductors with transitions temperatures of about 1\,K and below, STM and STS measurements at very low temperatures are required. A STM operating at such low temperatures would make it feasible to investigate properties of unconventional superconductors, e.\ g., the order parameter in UPt$_3$. To observe the full opening of the superconducting gap, the temperature has to be sufficiently below the critical temperature, e.\ g., $T_c\approx 550\,$mK for UPt$_3$\,\cite{4}.\\
In the last years, STMs for low temperatures down to $300\,$mK and moderate magnetic fields have become commercially available. However, for even lower temperatures and high magnetic fields $B\ge 10\,$T only a small number of home-built scan heads exist\,\cite{5}. The implementation of an STM into a dilution refrigerator is not straightforward, there are several requirements in order to successfully run the system, i.\ e., mechanical and electrical noise damping, a functional coarse approach for low temperatures, etc.\\
In this article we describe a home-built STM for low temperatures $T\approx 30\,$mK that works in a standard dilution refrigerator and in magnetic field up to $B=13\,$T. Calibration measurements on highly-ordered pyrolytic graphite (HOPG) and NbSe$_2$ at room temperature are presented. Calibration at low temperature and in magnetic fields has been performed on NbSe$_2$ in the Shubnikov phase. The measurements demonstrate the good spatial and energy resolution, respectively a high voltage resolution of the order of $1\,$K and stability of the instrument.
\section{Setup}
The low-temperature STM has to meet several stringent requirements. First of all the setup has to be rigid enough to provide mechanical stability and a good thermal contact to the mixing chamber. The choice of appropriate materials for temperature- and field-dependent measurements should take into account the thermal expansion and the magnetic properties. The setup should be electronically isolated from ground loops. Electromagnetic stray fields should be shielded and filtered from the leads.\\
In this section the general setup is discussed briefly. The STM is based on the design developed for scan heads used at higher temperatures\,\cite{6,7} but with a modified tip approach in order to allow for measurements in magnetic fields. The most important features of the design are its versatility and compactness which allows to fit it into the cryostat tail within a superconducting solenoid. The setup allows cycle times of about two to three days combined with a high mechanical stability. During refill with liquid helium, mechanical distortions occure, nevertheless the tip does not need to be retracted completely from its position close to the sample surface. This advantage compensates for the long warming and cooling time necessary for tip and sample changes.
\begin{figure}[ht]
\includegraphics[width=0.5\linewidth]{marz-fig-1.eps}
\caption{\label{fig:skizze} Sketch of the principal setup mounted to the mixing chamber. The illustration is not to scale. A copper plate supporting the copper-powder filters is mounted directly to the mixing chamber. An incised copper tube is used to ensure that the STM body, more precisely the sample position, is located in the homogeneous part of the magnetic field generated by a superconducting solenoid.}
\end{figure}
\subsection{General Assembly}
A standard $^3$He/$^4$He dilution refrigerator (GVL Leiden Cryogenics, Leiden, Netherlands) with a continuous 1\,K plate and a base temperature well below $T_{\rm{min}}\approx30\,$mK is used to host the home-built STM. The cryostat is equipped with a superconducting solenoid installed outside the vacuum chamber and designed to reach magnetic fields up to $B_{\rm{max}}=13\,$T. The general setup is sketched in figure~\ref{fig:skizze}. The STM body is made of highly resistive stainless steel (fig.~\ref{fig:bauteile} A) in order to reduce eddy currents, when high magnetic fields are applied. Furthermore it is supposed to be nonmagnetic, although a small magneto mechanical effect, e.\ g., due to magnetostriction can not be excluded, when high magnetic fields are applied.\\
The STM sample position is placed in the center of the magnet bore for having a homogeneous field. A heater made of a twisted-pair manganin wire allows heating of the whole scan head. It is wound around the stainless steel body and thermally anchored but electrically isolated with stycast epoxy (Stycast 2850FT Polymer). To avoid local hot spots, the heater is glued in a meander-like fashion to cover most of the surface. The temperature of the STM body is controlled by a Cernox bare chip thermometer calibrated in the temperature range from $T=30\,$mK to $T=1.5\,$K. The sensor is mounted inside an incised Cu screw into the STM body to allow good thermal contact between thermometer and STM body especially at low temperatures. The STM body itself is thermally anchored to the mixing chamber via an oxygen-free copper tube. In addition, a heat shield thermal anchored to the mixing chamber surrounds the STM body. It has been verified that a temperature gradient of less than $20\,$mK occurs between mixing chamber and STM body. As a test we recorded tunnel spectra of superconducting aluminum ($T_c=1.1\,$K) which exhibited a temperature dependence down to a STM body temperature of $\approx 100\,$mK. We observed a temperature dependent $dI/dV$ curve down to $200\,$mK, i.e., $T/T_c=0.18$. The fact that we did not observe a further sharpness of the BCS singularity at lower temperature is probably due to electronic noise. Thus we can conclude in this particular experiment (STM body temperature $\approx 100\,$mK) that the sample temperature is at least $200\,$mK. All temperatures given in this article correspond to the STM body temperature determined in the experiment from the calibrated sensor.
\subsection{STM}
\begin{figure*}[ht]
\includegraphics[angle=90,width=\linewidth]{marz-fig-2.eps}
\caption{\label{fig:bauteile} Photograph of the single parts of the scan head before assembly. (A) stainless steel body, (B) piezo tube, (C) sapphire roads, (D+E) sample holder (F) isolation and (G) plug for electrical connection of the piezo scan head, thermometers and heater.}
\end{figure*}
The STM consists of a scan-head unit with a commercial five-electrode piezo tube scanner (EBL \#2, Staveley NDT Technologies, East Hartford, USA) (fig.~\ref{fig:bauteile} B). The inner electrode of the piezo tube is used for the $z$ movement and the outer electrode is split into four parts, for $\pm\,x$ and $\pm\,y$ movement, respectively. The way of assembling the scan head from the parts of fig.~\ref{fig:bauteile} is displayed in fig.~\ref{fig:assembly}. Almost all parts are fixed with metallic screws to ensure good thermal coupling, only the piezo is glued with Torr Seal epoxy (Varian) to a macor\,\cite{macor} panel and the sapphire rods.\\
The coarse approach of the STM tip to the sample is done by a vertical slip-stick mechanism using the $z$ movement. The tip is attached to a non-magnetic CuBe slider which is supported by two sapphire rods (fig.~\ref{fig:bauteile} C).
\begin{figure}[ht]
\includegraphics[width=0.1\linewidth]{marz-fig-3.eps}
\caption{\label{fig:assembly} Sketch of the assembling of the parts from fig~\ref{fig:bauteile}. All parts are fixed with insulating screws, except for the piezo (B) and the sapphire roads (C) which are glued with Torr Seal and the 36-pin plug (mounted inside G) which is glued with stycast epoxy (Stycast 1266 Polymer).}
\end{figure}
The coarse approach of the tip, or rather the coarse approach of the slider, is based on inertia movement introduced by Pohl\,\cite{8}.
\begin{figure}[ht]
\includegraphics[width=0.4\linewidth]{marz-fig-4.eps}
\caption{\label{fig:schlitten} Sketch of the cross-section slider (parts B,C,D) mounted on the sapphire rods (A). The V/W shape leads to three lines at which the slider is touching each rod. A displacement of the sapphire rods due to horizontal forces is avoided by the V/W shape of the slider.}
\end{figure}
The cross-section of the slider is sketched in figure~\ref{fig:schlitten}. The upper part (fig.~\ref{fig:schlitten} B) consists of a CuBe cube with a cylindrical hole and three perpendicular drilling holes. One hole is used to fix the tip, another one is used for the connections of the current lead. The lower part (fig.~\ref{fig:schlitten} C) is connected through the middle hole to the upper part by a screw (fig.~\ref{fig:schlitten} D) and pre-loaded by a spring which allows to apply a variable pressure to the sapphire rods (fig.~\ref{fig:schlitten} A). This changes the friction between the slider and the sapphire rods and therefore permits the control of the speed, in particular the step width, of the coarse approach. The step width has to be large enough at room temperature to maintain a sufficient displacement at low temperatures. The V/W shape of the slider prevents a displacement of the sapphire rod due to horizontal forces.\\
The STM tips are prepared by cutting a $250\,\mathrm{\mu}$m Pt/Ir (80/20) wire in tension with a ceramic scissors under an angle of $45^\circ$. The typical length of the tip is about $2-3\,$mm. The tip is fixed with a CuBe screw to the slider, allowing a good thermal contact and high mechanical stability.\\
The sample is glued with silver conductive paint to a copper holder, which is fixed with two isolated bolts at the bottom of the main body (fig.~\ref{fig:bauteile} D). The copper holder with the sample is supported by stainless steel shoes (fig.~\ref{fig:bauteile} E) that are also isolated from the main body (fig.~\ref{fig:bauteile} F). To get rid of possible contamination, e.\ g., a water film on the sample surface, after a change of sample and tip at ambient conditions, the scan head is heated to $100^{\circ}\,$C, while the inner vacuum can is pumped. The outer parts, i.\ e., the magnet and the He dewar are cooled to liquid nitrogen temperature meanwhile. Thus, we can expect that nearly all water evaporates from the sample surface and condenses at the wall of the vacuum chamber. To control the temperature during heating a Pt-100 thermometer is attached on the reverse side of the sample holder (fig.~\ref{fig:bauteile} D). No further preparation is installed at present. However, sample space in the vacuum chamber allows for the installation of an in-situ cleavage station.\\
The electrical wiring for the piezo scan head, the thermometers and the heater is done via a 36-pin IC socket. The leads from the plug to the mixing chamber (including the copper-powder filters, see below) are made of polyimid-isolated $100-\mathrm{\mu}$m Cu wire. For the connection to the room temperature part of the cryostat manganin wires are used for the piezo leads to minimize the thermal conductance and to avoid a modification of the signal form due to large resistance changes. The thermometer leads from the mixing chamber to the 1\,K pot are made of superconducting NbTi wires.\\
The feedthroughs for current and voltage are made of gold-plated copper-shielded coaxial cable with stranded copper center wire (LakeShore CC-SC-100), which is connected via SMA plugs to the copper-powder filter placed at the mixing chamber. The connection to the upper part of the cryostat is done with stainless steel coaxial cables (LakeShore CC-SS-100).\\
The measurement of the capacitance between tip and sample is used to monitor and control the tip-sample distance during cooling the cryostat and at low temperatures. The maximum scanning area of the setup at room temperature is about $10\, \mathrm{\mu}$m$\times 10\, \mathrm{\mu}$m, at low temperatures it is reduced to $2\, \mathrm{\mu}$m$\times 2\, \mathrm{\mu}$m.
\subsection{Electronics}
The feedback loop, the high voltage amplification, and data acquisition electronics are commercially available units (ECS, Cambridge, UK). For STS measurements the output voltage is modulated with a frequency generator (Stanford Research Systems DS345, Sunnyvale). The current signal is deconvoluted with an analog lock-in amplifier (Ithaco 3916B, Ithaca, NY). For the current amplification, a current-voltage transformator with variable gain (Femto DHCL 200, Berlin) is used. The data acquisition is performed with a standard PC equipped with a digital signal processing (DSP) card.
\subsection{Electrical and Mechanical Filtering}
Electrical and mechanical filtering is the most important challenge for proper STM and STS performance. Main reasons for the electrical noise are ground loops and long leads, mechanical noise occurs mainly through vibrations of the pumping lines (in our case mostly of the $1\,$K pumping line and, to a lesser extant, the still pumping line) and through boiling of the cryogenic fluids. To reduce vibrational noise no liquid nitrogen shield was used, but super insulation foil in the outer vacuum can instead. Additionally, the He dewar was pressurized to $50-100\,$mbar above athmospheric pressure to reduce boil-off. By pumping on the He dewar no further advantage was gained.\\
To minimize electrical noise, the STM electronics is electrically isolated from all other electronics as well as from all pumping lines. This permits the use of the cryostat as the only ground in the whole system. All leads used for measurements and control are filtered by home-built copper-powder filters (skin-effect filters). These filters are mounted on top of the STM holder next to the mixing chamber. Their main purpose is to reduce high-frequency noise. The attenuation vs. frequency curve is shown in fig~\ref{fig:daempfung}.
\begin{figure}[ht]
\includegraphics[angle=90,width=0.7\linewidth]{marz-fig-5.eps}
\caption{\label{fig:daempfung} Damping behavior of the copper-powder filter. The measurement of the solid line was performed with a network analyzer in a frequency range $70\,$kHz$-20\,$GHz. The low-frequency data are determined by applying an ac signal on one end of the filter and measuring the response signal on the other end with an oscilloscope.}
\end{figure}
A further advantage of these filters is the good thermal coupling of the leads to the mixing chamber. The filters consist of $2\,$m of copper wire, copper powder (grain size $50\,\mathrm{\mu}$m) and a copper jacket that is attached very closely to the mixing chamber and therefore cools down the electronic system. Low-pass filters (cutoff frequency $\nu=2\,$kHz) are installed behind the high-voltage amplifier to suppress noise of the used electronics. An additional low-pass filter with a cutoff at $4\,$kHz is installed consecutively to the preamplifier.\\
The cryostat itself is hanging from a platform supported by three air cushions for damping the mechanical noise. Typical cut-off frequency of the low-pass system is $10\,$Hz. For additional damping of mechanical distortions (vibrations of pumping lines etc.) the scan head unit is fixed in an incised copper tube. The tube acts like a tuning fork, with a sharp resonance and damping of all other frequencies.
\section{Calibration at Room Temperature}
After assembling the instrument, first measurements were carried out at room temperature and at ambient pressure to calibrate the scanning unit. First, the calibration for large scan areas are performed on a commercially available calibration device ('NanoGrid', Schaefer Technologie GmbH, Langen). The grid is a $2 \,$mm$\,\times\,2\,$mm sized polymer chip with crossed lines at a relative distance of $d=160\,$nm. In order to make the surface conductive, AuPd is evaporation-deposited on top. From these measurements scaling factors of $30.0\,$nm/V for $\pm\,$x movement and $43.1\,$nm/V for $\pm\,$y movement are obtained.\\
\begin{figure}[ht]
\includegraphics[width=.7\linewidth]{marz-fig-6.eps}
\caption{\label{fig:hopg} Raw data of the atomic resolution on HOPG ($V_{bias}=76\,$mV, $I= 2\,$nA) at room temperature and ambient pressure. The lower graph shows the line profile marked in the scan. The well-known lattice constant was used to calibrate the scan head for room temperature and small scanning areas.}
\end{figure}
HOPG is used for the calibration of the piezo positioning elements, as is often done as a standard procedure for two reasons: First, it is relatively easy to get atomic resolution on HOPG, and second, the preparation of a clean surface is very simple. As HOPG has a layered structure, a fresh surface can be obtained by using adhesive tape to remove some of the top layers. The lattice parameters are well known from literature: the nearest neighbor distance is $0.142 \,$nm, whereas the in-plane lattice constant is $0.246 \,$nm\,\cite{9}. To achieve atomic resolution a higher spatial sensitivity for the scanning unit is needed. To increase the resolution at room temperature a voltage divider could be used optionally. The applied voltage to the scan piezo is divided by a factor of 10 to avoid the digitization artifacts. These steps appear in the control voltage of the piezo amplifier, which has a limited resolution of $12$ bit. Figure~\ref{fig:hopg} shows the raw data of the hexagonal atomic structure of the HOPG sample at room temperature at ambient conditions.
\section{Calibration at Low Temperature and in Magnetic Field}
\begin{figure}[ht]
\includegraphics[width=.5\linewidth]{marz-fig-7.eps}
\caption{\label{fig:nbse_2_tt-mark} Atomic resolution on NbSe$_2$ at $T \approx 50\,$mK, with $V=322.3\,$mV and $I=1.7\,$nA. The rombus indicates the unit cell of the hexagonal lattice. With this data the scan head was calibrated for low temperature and small scanning areas.}
\end{figure}
Besides HOPG, the layered type-II superconductor NbSe$_2$ with a critical temperature of $7.2\,$K was investigated. This material is a favorite object for STM and STS studies\,\cite{10,11,12,13} because it has a layered structure similar to HOPG and the sample preparation can be done in a similar way. Figure~\ref{fig:nbse_2_tt-mark} displays a topographic scan on NbSe$_2$ taken at $T \approx 50\,$mK. The in-plane lattice constant is $0.345\,$nm\,\cite{14}. From the comparison to the lattice constant obtained at room temperature and at low temperature the conversion factor between the room-temperature and the low-temperature calibration of the scan head can be calculated. This conversion factor is determined to be $\approx 0.2$ for small scanning areas.\\
\begin{figure*}[ht]
\includegraphics[angle=90,width=\linewidth]{marz-fig-8.eps}
\caption{\label{fig:0T-flussliniengitter} Measurements of the vortex lattice in NbSe$_2$, with $V=1.6\,$meV and $I=2\,$nA at $T\approx 100\,$mK for different fields $B=0$ (A), $B=0.2\,$T (B), $B=0.4\,$T (C) and $B=0.8\,$T (D). The bright spot in figure~A (arrow) shows a trapped flux line. The circles mark surface defects. They do not change the position with increasing field indicating that no major drift is observable.
}
\end{figure*}
In magnetic fields above the lower critical field, type-II superconductors like NbSe$_2$ enter the Shubnikov phase and magnetic flux can penetrate the sample in the form of quantized flux lines, regions where the superfluid density vanishes, i.\ e., regions with a normal conducting core. These flux lines are regularly arranged (Abrikosov lattice) in superconductors with low pinning and the change of the electronic structure between the superconductive to the normal-conductive density of states is directly observable with STM measurements. To observe the flux-line lattice of NbSe$_2$ a small magnetic field of $\le 1\,$ T has been applied. Field-dependent measurements of the local density of states in the constant current mode were performed at applied voltage $V\approx 1.6\,$mV. At this voltage the maximum of the coherence peak is observed in the tunneling spectra in zero field (see fig.~\ref{fig:spektroskopie}) and the maximal contrast between normal and superconducting regions is obtained.\\
In figure~\ref{fig:0T-flussliniengitter}, four STM images of the same surface region acquired in the constant current mode at $V\approx 1.6\,$mV and $T\approx 100\,$mK are shown for different applied magnetic fields. The image (A) is measured at zero field, but after a field sweep. A trapped flux line in the upper right-hand corner can be identified (arrow). Flux can still be present in the sample caused either by a remanent field of the superconducting magnet or by pinning of a flux line at a defect. Image (B) is taken at a magnetic field of $B=200\,$mT, the expected hexagonal structure of the Abrikosov lattice is clearly observed. Raising the magnetic field (C and D) further increases the density of vortices. For larger fields $B\ge 800\,$mT a noticeable drift at the lower parts of the scan area occurs (fig.~\ref{fig:0T-flussliniengitter} D), which possibly is due to heating of the setup by eddy currents while changing the field. However major shifts due to heating of the sample and/or the tip during field change can be excluded. As seen from the pictures (A-D) of figure~\ref{fig:0T-flussliniengitter}, the overall drift of the scanning area between different scans can be neglected, since characteristic surface defects remain at almost the same positions in every scan (white circles).\\
In order to demonstrate the performance of the STM in magnetic fields the lattice constant $a$ of the Abrikosov lattice has been extracted from the data. In figure~\ref{fig:gitter} the lattice constant $a$ is shown together with the expected field dependence\,\cite{15} $a=\left(4/3\right)^{1/4}\cdot\left(\Phi_0/B\right)^{1/2}$ where $\Phi_0=h/2e$, without any adjustable parameters. Therefore, imaging the vortex lattice allows a calibration of our scan head at low temperatures for large scanning areas, in addition to the calibration for small scanning ranges achieved by atomic resolution. The scaling factor with respect to the room-temperature motion of the scan head is $\approx 0.2$ in both cases.\\
\begin{figure}[ht]
\includegraphics[angle=90,width=.7\linewidth,clip=]{marz-fig-9.eps
\caption{\label{fig:gitter} Field dependence of the flux line lattice constant. The curve is calculated without any free parameter. The expected $a\propto (1/B)^{1/2}$ dependence is clearly seen in the measurements.}
\end{figure}
\section{Spectroscopy at Low Temperatures}
\begin{figure}[ht]
\includegraphics[angle=90,width=.7\linewidth,clip=]{marz-fig-10.eps}
\caption{\label{fig:spektroskopie} Differential conductance of a NbSe$_2$-PtIr tunnel contact at $T\approx 50\,$mK. The data are taken at one point on the surface averaged over $100$ measurements with $512$ steps each in this voltage span. The solid and dashed lines are calculated using the BTK theory, details are given in the text. The deviations between the fit are probably due to charge-density wave dynamics or to the gap anisotropy in NbSe$_2$.}
\end{figure}
While running the scan head in the STM mode gives insights into the topography of the sample surface, the STS mode directly yields the electronic density of states at the surface via the differential conductance vs. voltage curves. The characteristics of the superconducting density of states in NbSe$_2$ are expected to occur at $\Delta/e=1.75\cdot k_{\rm{B}}T_c/e=1.08\,$mV, i.\ e., at low DC bias. To resolve signatures at such low energies a high energy resolution, respectively voltage resolution $ \delta V < 1\,$meV is needed. For that purpose a switchable voltage divider (1:1, 1:10, 1:100, 1:1000) was installed to increase the resolution of the electronics. In order to record the tunneling curves with high resolution the tunneling voltage is modulated with an AC bias ($V_{AC}\approx 20\,\mathrm{\mu}$V) and the current signal is measured with a lock-in amplifier.\\ Figure~\ref{fig:spektroskopie} shows a typical measurement of the differential conductance of a NbSe$_2$-PtIr tunnel contact (circles) at $T=50\,$mK in zero field. At such low temperatures the spectrum shows pronounced conductance peaks due to the BCS singularity of the density of states\,\cite{15a}. The value of the energy gap is determined by fitting the spectra using the BTK theory\,\cite{16} for a large barrier parameter (tunnel limit). By fitting the positions at the maxima (solid line), the energy gap $\Delta=1.62\,$meV is obtained, whereas a fit to the width of zero conductance (dashed line) results in $\Delta=1.06\,$meV for the gap. These variations reflect a notable gap anisotropy in this material\,\cite{13} most probably caused by the presence of charge-density wave dynamics in NbSe$_2$ since our measurements are relatively slow. On the other hand a dynamic broadening due to charge density waves is also possible.\\
\begin{figure}[ht]
\includegraphics[width=.907\linewidth]{marz-fig-11.eps}
\caption{\label{fig:boundstate} Differential conductance of a NbSe$_2$-PtIr tunnel contact in the vortex state (left panel) at $T\approx 100\,$mK. The positions of the PtIr tip above the NbSe$_2$ sample are marked (right panel). Spectra 1 and 3 are measured in the superconducting regions outside the vortex, spectrum 2 is measured directly at the center of the vortex. Spectra are shifted along the y axis for clarity.}
\end{figure}
In an applied field $B=500\,$mT, line scans ('topographic mode') through a vortex were performed (not shown) in addition to the images of the vortex lattice. The spectra recorded along the line scan clearly show a somewhat broadened BCS-like density of states in the superconducting regions and an enhancement of the differential conductance at zero bias in the normal-conducting regions of the vortex center which can be attributed to Andreev bound states\,\cite{17}. This enhancement in the conductance has been reported earlier\,\cite{18,19} and is very sensitive to the position of the tip in respect to the vortex center.\\
As representative examples, three different spectra are shown in figure~\ref{fig:boundstate}. The data acquired at a particular point on the surface were averaged over $100$ measurements in the same voltage range of $|V|\le4\,$mV with $512$ steps each in this voltage span. Measurements (1) and (3) were performed away from a vortex and show the superconductive density of states. The spectrum at the vortex (2) clearly shows the enhancement at zero bias of the conductance due to the Andreev bound states. The observation of this state indicates the stable and reproducible positioning of the tip at the center of the vortex.
\section{Acknowledgments}
We thank R. Hoffmann, B. Pilawa, C. Debuschewitz and W. Wulfhekel for helpful discussions. Many thanks to C. P\'erez Le\'on for the help with the assembly of the electronics. We thank D. Rosenmann at the Argonne Nation Laboratory, Condensed Matter Division, for providing us the NbSe$_2$ sample.
|
\section{Introduction}\label{sec Introduction}
Since the proof of Fermat's Last Theorem \cite{Wiles 95},
\cite{Taylor Wiles 95} and the establishment of the full
Shimura-Taniyama-Weil conjecture \cite{BCDT 01}, many Diophantine
equations were solved using a strategy similar to that of the
proof of FLT. Amongst them are various so-called generalized
Fermat equations. These are Diophantine equations of the form
\begin{equation}\label{eqn generalized Fermat}
ax^p+by^q=cz^r, \quad x,y,z \in \mathbb{Z}, \quad xyz\not=0, \quad
\gcd(x,y,z)=1
\end{equation}
where $a,b,c$ are nonzero integers and $p,q,r$ are integers $\geq
2$. The nature of the solutions depends very much on the quantity
\[\chi(p,q,r):=\frac{1}{p}+\frac{1}{q}+\frac{1}{r}-1.\]
If $\chi(p,q,r)>0$, then there are either no solutions to
(\ref{eqn generalized Fermat}) or infinitely many. In the latter
case, there exist finitely many triples $(X,Y,Z)$ of binary forms
$X,Y,Z \in \mathbb{Q}[u,v]$ satisfying $aX^p+bY^q=cZ^r$ such that every
solution $(x,y,z)$ to (\ref{eqn generalized Fermat}) can be
obtained by specializing the variables $u,v$ to integers for one
of these triples; see \cite{Beukers 98}. If $\chi(p,q,r)=0$, then
the determination of the solutions to (\ref{eqn generalized
Fermat}) basically boils down to finding rational points on curves
of genus one. If $\chi(p,q,r)<0$, then there exists a curve $C$ of
genus $\geq 2$ and a covering $\phi: C \to \mathbb{P}^1$, both defined
over a number field $K$, such that for every solution $(x,y,z)$ to
(\ref{eqn generalized Fermat}) we have $[ax^p:cz^r] \in
\phi(C(K))$. Since by Faltings' theorem $C(K)$ is finite, there
are only finitely many solutions to (\ref{eqn generalized Fermat})
in this case; see \cite{Darmon Granville 95}.
Of special interest is the generalized Fermat equation with
$a=b=c=1$. In the case $\chi(p,q,r)>0$, all the (parameterized)
solutions are known; see \cite{Beukers 98} and \cite{Edwards 04}.
If $\chi(p,q,r)=0$, the only solution, up to sign and permutation,
is given by the Catalan solution $2^3+1^6=3^2$. If
$\chi(p,q,r)<0$, the only solutions known, up to sign and
permutation, are given by $2^3+1^q=3^2 \ (q \geq 7)$ and
\begin{multline*}
17^3+2^7=71^2, \quad 2213459^2+1414^3=65^7,\\
15312283^2+9262^3=113^7, \quad 76271^3+17^7=21063928^2;
\end{multline*}
\[1549034^2+33^8=15613^3,\quad 96222^3+43^8=30042907^2;\]
\[13^2+7^3=2^9;\]
\[7^2+2^5=3^4,\quad 11^4+3^5=122^2.\]
For a list containing many (families of) triples $(p,q,r)$ for
which (\ref{eqn generalized Fermat}) has been solved in this case,
we refer to \cite[Table 1]{PSS 07}. In this paper we will focus on
the special case $(p,q,r)=(2,2n,3)$ for $n \in \mathbb{Z}_{>0}$, i.e. we
are concerned with the Diophantine equation
\begin{equation}\label{eqn 2_2l_3}
x^2+y^{2n}=z^3, \quad x,y,z \in \mathbb{Z}, \quad xyz \not=0, \quad
\gcd(x,y,z)=1.
\end{equation}
\subsubsection*{Previously known results.}
For $n=1,2$ we have $\chi(2,2n,3)>0$, and in both cases there are
infinitely many solutions to (\ref{eqn 2_2l_3}). By factoring
$x^2+y^2=z^3$ over the Gaussian integers as $(x+iy)(x-iy)=z^3$ one
readily gets, that a solution to (\ref{eqn 2_2l_3}) for $n=1$
satisfies
\[(x,y,z)=(u(u^2-3 v^2), v (3 u^2-v^2),u^2+v^2)\]
for some $u,v \in \mathbb{Z}$ with $\gcd(u,v)=1$ and $uv\not=0$. By
demanding that $v (3u^2-v^2)$ is a square, one can obtain
parameterized solutions to (\ref{eqn 2_2l_3}) for $n=2$. This was
carried out by Zagier and reported in \cite{Beukers 98}; up to
sign there are $4$ parameterizations, $3$ of which have
coefficients in $\mathbb{Z}$. In \cite{Edwards 04} parameterized solutions
to (\ref{eqn 2_2l_3}) for $n=2$ were obtained by a different
method, the same parameterizations as in \cite{Beukers 98} were
found, except that the one with non-integer coefficients was
replaced by one with coefficients in $\mathbb{Z}$.
For $n=3$ there are no solutions to (\ref{eqn 2_2l_3}). This
follows readily from the well-known fact that the only rational
points on the elliptic curve given by $Y^2=X^3-1$ are
$(X,Y)=(1,0)$ and the point at infinity.
By demanding that $y$ is a square for the parameterized solutions
to (\ref{eqn 2_2l_3}) for $n=2$, one obtains genus $2$ curves such
that every solution to (\ref{eqn 2_2l_3}) for $n=4$ corresponds to
a rational point on one of these curves. This, together with the
determination of all rational points on these curves using
effective Chabauty methods, was carried out in \cite{Bruin 99}.
The result is that, up to sign, the only solution to (\ref{eqn
2_2l_3}) for $n=4$ is given by $1549034^2+33^8=15613^3$.
We see that it suffices to deal with $n$ a prime $> 3$. In
\cite{Chen 08}, the equation is studied using methods of modular
forms and Galois representations for approaching Diophantine
equations. In particular, an explicit criterion is given for
showing that (\ref{eqn 2_2l_3}) has no solution for a given prime
$n >7$. This criterion is verified for all primes $7 < n < 10^7$
except $n=31$, thereby showing that (\ref{eqn 2_2l_3}) has no
solutions for these values of $n$. The nonexistence of solutions
for $n=7$ follows from the work of \cite{PSS 07}, so the only
small values of $n$ left to deal with are $n=5,31$.
At this point we should mention that, for $m\in \mathbb{Z}_{>1}$, the
Diophantine equation
\[x^2+y^3=z^m, \quad x,y,z \in \mathbb{Z}, \quad xyz \not=0, \quad
\gcd(x,y,z)=1\] is much harder to deal with than (\ref{eqn
2_2l_3}) is (apart from some small values of $m$), even if we just
consider even $m$. This is because of the Catalan solution
$3^2+(-2)^3=1^m$. For $m \leq 5$ there are infinitely many
solutions. The parameterization for $m=2$ is again very easy to
obtain, for $m=3,4,5$ we refer to \cite{Edwards 04} ($m=3$ was
earlier done by Mordell and $m=4$ by Zagier). The case $m=6$ is
classical, it amounts to determining the rational points on the
elliptic curve given by $Y^2=X^3+1$. The cases $m=7,8,9,10$ are
solved in \cite{PSS 07}, \cite{Bruin 03}, \cite{Bruin 05},
\cite{Siksek} respectively.
\subsubsection*{New results.}
In this paper, we will extend the criterion mentioned above to all
primes $n>3$ and use it to show that there are no solutions for
$n=5$. This is basically done by showing that a Frey curve
associated to a (hypothetical) solution has irreducible mod-$n$
representation for $n=5,7$. By using extra local information
obtained from classical algebraic number theory, we obtain a
refined criterion, which is used to show that (\ref{eqn 2_2l_3})
has no solutions for $n=31$. So from a Diophantine point of view
our main result is the following.
\begin{theorem}\label{thm main}
If $n \in \{5,31\}$, then (\ref{eqn 2_2l_3}) has no solutions.
\end{theorem}
Although we focus on very specific equations, we feel that the
methods we use to overcome the earlier difficulties, could
definitely be used in other cases as well. Combining our main
result with the previously known results mentioned above, we
arrive at a description of all solutions for $n \leq 10^7$.
\begin{corollary}\label{cor main}
Let $n \in \mathbb{Z}_{>0}$ with $n \leq 10^7$. If $n \in \{1,2,4\}$, then
(\ref{eqn 2_2l_3}) has solutions, all of which are described
above. If $n \not\in \{1,2,4\}$, then (\ref{eqn 2_2l_3}) has no
solutions.
\end{corollary}
Finally, by applying a key idea from \cite{Chen Siksek 09}, we
solve (\ref{eqn 2_2l_3}) for infinitely many (prime) values of
$n$.
\begin{theorem}\label{thm n congruence}
If $n \in \mathbb{Z}_{>0}$ with $n \equiv -1 \pmod{6}$, then (\ref{eqn 2_2l_3})
has no solutions.
\end{theorem}
\section{A modular approach to $x^2+y^{2l}=z^3$}\label{sec A modular
approach}
Throughout this section, let $l$ denote a prime $>3$. We shall
explain how modular forms and Galois representations can be used
to study (\ref{eqn 2_2l_3}) for $n=l$. We want to stress that the
methods and results described in this section are not new and can
be found essentially in \cite{Chen 08}. For a general introduction
to using methods of modular forms and Galois representations for
approaching Diophantine equations, one could consult e.g. Chapter
2 of the author's Ph.D. thesis \cite{Dahmen 08}.
\begin{lemma}\label{lem descent}
Suppose that $(x,y,z)$ is a solution to (\ref{eqn 2_2l_3}) for
$n=l$. Then
\begin{equation}\label{eqn descent 1}
(x,y^l,z)=(u(u^2-3 v^2), v (3 u^2-v^2),u^2+v^2)
\end{equation}
for some $u,v \in \mathbb{Z}$ with $\gcd(u,v)=1,\ uv\not=0,\ 3|v$ and
$2|uv$. Furthermore,
\begin{equation}\label{eqn descent 2}
3v=r^l \quad \mathrm{and} \quad 3u^2-v^2=3s^l
\end{equation}
for some $r,s \in \mathbb{Z}$ with $\gcd(r,s)=1$ and $rs \not= 0$.
\end{lemma}
\begin{proof}
That (\ref{eqn descent 1}) holds for some nonzero coprime integers
$u,v$ follows immediately from the parameterization of solutions
to (\ref{eqn 2_2l_3}) for $n=1$, given in Section \ref{sec
Introduction}.
From $v(3 u^2-v^2)=y^l$ we see that up to primes dividing
$\gcd(v,3u^2-v^2)$, both $v$ and $3u^2-v^2$ must be $l$-th powers.
One easily checks that $\gcd(v,3u^2-v^2)$ equals either 1 or 3.
First suppose that $\gcd(v,3u^2-v^2)=1$. Then $v=r^l$ and
$3u^2-v^2=s^l$ for some nonzero $r,s \in \mathbb{Z}$. Furthermore, $r,s$
and $u$ are pairwise coprime and $(r^2)^l+s^l=3u^2$. However,
\cite[Theorem 1.1]{Bennett Skinner 04} tells us that for an
integer $n>3$ there are no nonzero (pairwise) coprime integers
$a,b,c$ satisfying $a^n+b^n=3c^2$, which contradicts
$\gcd(v,3u^2-v^2)=1$.
So suppose now that $\gcd(v,3u^2-v^2)=3$. Then $3 |v$ and of
course $3 \nmid u$, so $3||3u^2-v^2$. We get that $3v=r^l$ and
$3u^2-v^2=3s^l$ for some nonzero $r,s \in \mathbb{Z}$.
Finally, if both $u,v$ are odd, then $v (3u^2-v^2) \equiv 2
\pmod{4}$, but this contradicts that $v (3u^2-v^2)$ is an $l$-th
power.
\end{proof}
We have reduced (\ref{eqn 2_2l_3}) for $n=l$ to the equation $v (3
u^2-v^2)=y^l$ with $u,v,y$ as in the lemma above. This equation
can be approached using methods of modular forms and Galois
representations. In fact, one can use almost exactly the same
methods as used in \cite{Kraus 98} for studying the equation
$a^3+b^3=c^l$ in nonzero coprime integers $a,b,c$.
Suppose $(x,y,z)$ is a solution to (\ref{eqn 2_2l_3}) for $n=l$,
let $u,v,r,s$ be as in Lemma \ref{lem descent} and consider the
following Frey curve associated to this solution
\begin{equation}\label{eqn Frey Curve}
E: Y^2=\left\{
\begin{array}{ll}
X^3+2uX^2+\frac{v^2}{3}X & \mbox{if $u$ is even}; \\
X^3 \pm uX^2+\frac{v^2}{12}X,\ \pm u \equiv 1 \pmod{4} & \mbox{if
$u$ is odd}.
\end{array}
\right.
\end{equation}
Because $3|v$ and $2|v$ if $u$ is odd, we see that the model for
$E$ above actually has coefficients in $\mathbb{Z}$. Using Tate's
algorithm, or the in practice very handy to use \cite{Papadopoulos
93}, one finds that the model for $E$ is minimal at every prime
$p$, except at $p=2$ when $u$ is odd. Furthermore, the conductor
$N$ and the minimal discriminant $\Delta$ of $E$ satisfy (see also
\cite[Propositio 8]{Chen 08} and \cite[Lemma 2.1]{Bennett Skinner
04})
\begin{align*}
\Delta = & 2^{\alpha} \cdot 3^{-3} v^4(3u^2-v^2)=2^{\alpha} \cdot
3^{-6}
\left(r^4s\right)^l, \quad \alpha:=6 \mathrm{\ if\ } 2|u, \quad \alpha:=-12 \mathrm{\ if\ } 2\nmid u\\
N = & 2^{\beta} \cdot 3\, \mathrm{rad}_{\{2,3\}}(\Delta)=2^{\beta} \cdot
3\, \mathrm{rad}_{\{2,3\}}(rs), \quad \beta:=5 \mathrm{\ if\ } 2|u, \quad
\beta:=1 \mathrm{\ if\ } 2\nmid u,
\end{align*}
where for a finite set of primes $S$ and a nonzero $n \in \mathbb{Z}$,
$\mathrm{rad}_S(n)$ denotes the product of prime $p$ with $p|n$ and $p
\not\in S$.
\begin{remark}
From $x^2-z^3=-y^{2l}$, an obvious choice, up to quadratic twist,
for a Frey curve associated to the solution $(x,y,z)$ would be
\[E': Y^2=X^3-3zX+2x.\]
This model has (not necessarily minimal) discriminant $-2^6 \cdot
3^3(x^2-z^3)=2^6 \cdot 3^3 y^{2l}$, which has no primes $>3$ in
common with $c_4=2^4 \cdot 3^2 z$. Using (\ref{eqn descent 1}), we
can write $E'$ as
\[E': Y^2=X^3-3(u^2+v^2)X+2u(u^2-3 v^2),\]
from which we obtain that $E'$ has a rational $2$-torsion point
because the right-hand side factors as $(X+2 u) (X^2 - 2 u X + u^2
- 3 v^2 )$. The $2$-isogenous elliptic curve associates to the
$2$-torsion point $(X,Y)=(-2u,0)$ is simply, up to quadratic
twist, the Frey curve $E$ above. For both theory and practical
computation, it makes no difference whether $E$ or $E'$ is used,
except for establishing irreducibility of the mod-$5$
representation, where the computations actually become cleaner if
one uses $E'$ instead of $E$. The only reason for introducing $E$
above, is to follow \cite{Chen 08} more closely.
\end{remark}
For an elliptic curve $\mathfrak{E}$ over $\mathbb{Q}$, let $\rho_l^{\mathfrak{E}}:
\mathrm{Gal}(\overline{\mathbb{Q}}/\mathbb{Q}) \to \mathrm{Gl}_2(\mathbb{F}_l)$ denote the
standard $2$-dimensional mod-$l$ Galois representation associated
to $\mathfrak{E}$ (induced by the natural action of
$\mathrm{Gal}(\overline{\mathbb{Q}}/\mathbb{Q})$ on the $l$-torsion points $E[l]$
of $\mathfrak{E}$). Suppose that $\rho_l^E$ is irreducible. Then by
modularity \cite{BCDT 01} and level lowering \cite{Ribet 90} (note
that $\rho_l^E$ is finite at $l$), we obtain that $\rho_l^E$ is
modular of level $N_0:=2^{\beta} \cdot 3$, weight $2$ and trivial
character. This means that $\rho_l^E \simeq \rho_l^f$ for some
(normalized) newform $f \in S^2(\Gamma_0(N_0))$, where of course
$\rho_l^f$ denotes the standard $2$-dimensional mod-$l$ Galois
representation associated to $f$. If $u$ is odd, then $N_0=6$ and
since there are no newforms in $S^2(\Gamma_0(6))$ we have reached
a contradiction in this case. If $u$ is even, then $N_0=96$ and
$S^2(\Gamma_0(96))$ contains $2$ newforms, both rational and
quadratic twists over $\mathbb{Q}(i)$ of each other. So $\rho_l^E \simeq
\rho_l^f \simeq \rho_l^{E_0}$ for some elliptic curve $E_0$ over
$\mathbb{Q}$ with conductor $N_0=96$. As is well-known, we get from
$\rho_l^E \simeq \rho_l^{E_0}$ by comparing traces of Frobenius,
that for primes $p$
\[a_p(E_0) \equiv a_p(E) \pmod{l} \quad \mathrm{if\ } p \nmid N\]
\[a_p(E_0) \equiv a_p(E) (1+p) \equiv \pm (1+p) \pmod{l} \quad \mathrm{if\ } p | N \mathrm{\ and\ } p\nmid N_0=2^5 \cdot 3.\]
Note that we explicitly know $E_0$ up to isogeny and quadratic
twist. So if $E_0'$ is any elliptic curve over $\mathbb{Q}$ with conductor
$96$, e.g. given by $Y^2=X^3\pm X^2-2X$, then $a_p(E_0)^2 =
a_p(E_0')^2$ for all primes $p$. Therefore, for every prime $p$,
we can effectively compute the uniquely determined value
$a_p(E_0)^2$. We summarize some of the Diophantine information
obtained so far.
\begin{lemma}\label{lem Modular approach}
Suppose that $(x,y,z)$ is a solution to (\ref{eqn 2_2l_3}) for
$n=l$, let $u,v$ be as in Lemma \ref{lem descent}, let the
elliptic curve $E$ be given by (\ref{eqn Frey Curve}), let $E_0$
be any elliptic curve over $\mathbb{Q}$ with conductor $96$ and let $p>3$
be prime. If $\rho_l^E$ is irreducible and $a_p(E_0)^2 \not \equiv
(p+1)^2 \pmod{l}$, then $p\nmid v(3u^2-v^2)$ and $a_p(E_0)^2
\equiv a_p(E)^2 \pmod{l}$.
\end{lemma}
We see that in order to obtain a contradiction and conclude that
(\ref{eqn 2_2l_3}) has no solutions for $n=l$, it suffices (still
assuming the irreducibility of $\rho_l^E$) to find a prime $p>3$
such that $a_p(E_0)^2 \not\equiv (p+1)^2 \pmod{l}$ and $a_p(E_0)^2
\not\equiv a_p(E)^2 \pmod{l}$. A priori possible values of
$a_p(E)$ can be obtained by plugging all values of $u,v \pmod{p}$
with $p\nmid v(3u^2-v^2)$ into the equation for $E$. However, this
will never lead to a contradiction, since plugging the values
$u,v$ with $(|u|,|v|)=(2,3)$ into the equation for $E$, will give
us elliptic curves with conductor $96$. Now by reducing (\ref{eqn
descent 2}) modulo $p$, one obtains extra information on $u,v
\pmod{p}$ whenever one knows that the nonzero $l$-th powers modulo
$p$ are strictly contained in $\mathbb{F}_p^*$. This happens exactly when
$p \equiv 1 \pmod{l}$, say $p=kl+1$, in which case the nonzero
$l$-th powers modulo $p$ are given by
\[\mu_k(\mathbb{F}_p):=\set{\zeta \in \mathbb{F}_p}{\zeta^k=1}.\]
Heuristically speaking, the bigger $k$ is, the more possibilities
for $u,v \pmod{p}$ we expect, hence the more a priori
possibilities for $a_p(E)$ we expect, hence the less likely it is
to conclude that $a_p(E)^2 \not\equiv a_p(E_0)^2 \pmod{l}$. From a
computational point of view, it is desirable to only consider the
possibilities for $u/v \pmod{p}$, this indeed suffices, since $u/v
\pmod{p}$ determines the reduction of $E$ modulo $p$, denoted
$\tilde{E}(\mathbb{F}_p)$, up to quadratic twist and hence determines
$a_p(E)^2$ uniquely.
Continuing our more formal discussion, let indeed $k \in \mathbb{Z}_{>0}$ be
such that $p:=kl+1$ is prime (so $(p+1)^2 \equiv 4 \pmod{l}$) and
suppose that $a_p(E_0)^2 \not\equiv 4 \pmod{l}$. Then in
particular $p\nmid v(3u^2-v^2)=(rs)^l$. Define $U:=u/(3v)$. From
$3v=r^l$ and $3u^2-v^2=3s^l$, we get $U^2=1/(27)+(s/r^2)^l$. Since
$p \nmid rs$, we obtain for the reduction of $U$ modulo $p$,
denoted $\overline{U}$, that
\begin{equation}\label{eqn Skp}
\overline{U} \in S_{k,p}:=\set{\alpha \in
\mathbb{F}_p}{\alpha^2-\frac{1}{27}\in \mu_k(\mathbb{F}_p)}.
\end{equation}
For $\alpha \in S_{k,p}$ consider the elliptic curve over $\mathbb{F}_p$
given by
\begin{equation}\label{eqn Ealpha}
E_{\alpha}: Y^2=X^3+2\alpha X^2+\frac{1}{27}X.
\end{equation}
Note that $E_{\overline{U}}$ is a quadratic twist of
$\tilde{E}(\mathbb{F}_p)$, so that $a_p(E_{\overline{U}})^2 = a_p(E)^2$.
If now for all $\alpha \in S_{k,p}$ we have $a_p(E_{\alpha})^2
\not\equiv a_p(E_0)^2 \pmod{l}$, then $a_p(E)^2 \equiv
a_p(E_{\overline{U}})^2 \not\equiv a_p(E_0)^2 \pmod{l}$, a
contradiction which implies that (\ref{eqn 2_2l_3}) has no
solutions for $n=l$. Note that we have assumed $\rho_l^E$ to be
irreducible; one readily obtains (see Section \ref{sec
Irreducibility}) from \cite{Mazur 78} and \cite{Mazur Velu 72}
that this is actually the case if $l>7$. This proves the main
theorem of \cite{Chen 08}.
\begin{theorem}[{\cite[Theorem 1]{Chen 08}}]\label{thm Chen}
Let $l>7$ be prime and let $E_0$ be any elliptic curve over $\mathbb{Q}$
with conductor $96$. If there exists a $k \in \mathbb{Z}_{>0}$ such that
the following three conditions hold
\begin{enumerate}
\item $p:=kl+1$ is prime \item\label{item 2} $a_p(E_0)^2
\not\equiv 4 \pmod{l}$ \item\label{item 3} $a_p(E_{\alpha})^2
\not\equiv a_p(E_0)^2 \pmod{l}$ for all $\alpha \in S_{k,p}$,\\
where $S_{k,p}$ and $E_{\alpha}$ are given by (\ref{eqn Skp}) and
(\ref{eqn Ealpha}) respectively,
\end{enumerate}
then (\ref{eqn 2_2l_3}) has no solutions for $n=l$.
\end{theorem}
According to \cite{Chen 08}, it has been computationally verified
that for every prime $l$ with $7<l<10^7$ and $l\not=31$ there
exists a $k\in \mathbb{Z}_{>0}$ satisfying the three conditions of Theorem
\ref{thm Chen}, hence (\ref{eqn 2_2l_3}) has no solutions for
$n=l$ if $l$ equals one of these values.
We like to take this opportunity to point out a small omission in
the proof of \cite[Corollary 3]{Chen 08}. This corollary to
Theorem \ref{thm Chen} states that if $l>7$ is a Sophie Germain
prime, i.e. $p:=2l+1$ is prime, $\left(\frac{p}{7}\right)=1$ and
$\left(\frac{p}{13}\right)=(-1)^{(l+1)/2}$, then (\ref{eqn
2_2l_3}) has no solutions for $n=l$. As pointed out in \cite{Chen
08}, the conditions for the Legendre symbols are equivalent (by
quadratic reciprocity) to $S_{2,p}$ being empty, in which case
condition \ref{item 3} of Theorem \ref{thm Chen} trivially holds.
However, in order to use Theorem \ref{thm Chen} to deduce that
(\ref{eqn 2_2l_3}) has no solutions, we need of course prove that
condition \ref{item 2} also holds. This can be done as follows.
Note that both isogeny classes of elliptic curves over $\mathbb{Q}$ with
conductor $96$ contain an elliptic curve with rational torsion
group of order $4$. So $4|p+1-a_p(E_0)$, and hence $4|a_p(E_0)$.
This implies that if $a_p(E_0) \equiv \pm 2 \pmod{l}$, then
$|a_p(E_0) \mp 2| \geq 2l$. That this last inequality cannot hold,
follows directly from the Hasse bound $|a_p(E_0)| \leq 2 \sqrt{p}$
together with $l>3$, which completes the proof of the corollary.
Note that the corollary and its proof are analogously to work in
\cite{Kraus 98}, namely Corollaire 3.2 and its proof.
\section{Refinements}\label{sec Refinements}
In this section we shall prove our main result, Theorem \ref{thm
main}.
\subsection{Irreducibility.}\label{sec Irreducibility}
As pointed out in \cite{Chen 08}, since $E$ has at least one odd
prime of multiplicative reduction and a rational point of order
$2$, the irreducibility of $\rho_l^E$ for primes $l>7$ follows
from \cite[Corollary 4.4]{Mazur 78} and \cite{Mazur Velu 72}. In
fact, one does not need the first mentioned property of $E$, but
only that $E$ has a rational point of order $2$ to arrive at the
desired conclusion, since it is well-known that the modular curves
$X_0(2l)$ for primes $l>7$ have no noncuspidal rational points.
Irreducibility of $\rho_l^E$ for the remaining primes $l>3$ can
also be obtained.
\begin{theorem}\label{thm Irreducibility}
Let $l>3$ be prime and consider the elliptic curve $E$ given by
(\ref{eqn Frey Curve}), where $u,v \in \mathbb{Z}$ with
$v(3u^2-v^2)\not=0$. Then $\rho_l^E$ is irreducible.
\end{theorem}
\begin{proof}
The only cases left to deal with are $l=7$ and $l=5$.
The modular curve $X_0(14)$ is of genus one, it has $2$
noncuspidal rational points, which correspond to elliptic curves
with $j$-invariant $j_{14}$ equal to $-3^3 \cdot 5^3$ or $3^3
\cdot 5^3\cdot 17^3$; this readily follows from \cite[Chapter
5]{Ligozat}. Denote the $j$-invariant of $E$ by $j(u,v)$. For both
values of $j_{14}$ it is completely straightforward to check that
the equation $j(u,v)=j_{14}$ has no solutions with $u,v \in \mathbb{Z}$
and $(3u^2-v^2)v\not=0$. This proves that $\rho_l^E$ is
irreducible for $l=7$.
Since the modular curve $X_0(10)$ is of genus zero and has a
rational cusp (which is nonsingular of course), it has infinitely
many noncuspidal rational points. So we have to work a little
harder to obtain the irreducibility of $\rho_l^E$ for $l=5$. Up to
quadratic twist, the Frey curve $E$ is $2$-isogenous to
\[E': Y^2=X^3-6uX^2+3(3u^2-v^2)X\]
(replacing $X$ by $X+2u$, gives us back the model for $E'$ given
in Section \ref{sec A modular approach}). So $\rho_5^E$ is
irreducible if and only if $\rho_5^{E'}$ is irreducible. The
$j$-invariant $j'(u,v)$ of $E'$ (which is a twist of the $j$-map
from $X(2)$ to $X(1)$) is given by
\begin{eqnarray*}
j'(u,v) & = & \frac{1728(u^2 + v^2)^3}{v^2(3u^2 - v^2)^2}\\
& = & \frac{1728 u^2 (u^2 - 3v^2)^2}{v^2(3 u^2 - v^2)^2}+1728.
\end{eqnarray*}
The $j$-map from the modular curve $X_0(5)$ to $X(1)$, denoted
$j_5$, is given by
\begin{eqnarray*}
j_5(s,t) & = & \frac{(t^2+10st+5s^2)^3}{s^5 t} \\
& = & \frac{(t^2+4st-s^2)^2(t^2+22st+125s^2)}{s^5t}+1728.
\end{eqnarray*}
By comparing $j-1728$, we see that every $[u:v] \in \mathbb{P}^1(\mathbb{Q})$
such that $\rho_5^{E'}$ is irreducible gives rise to a rational
point on the curve $C$ in $\mathbb{P}^1 \times \mathbb{P}^1$ give by
\[
C: \frac{1728 u^2 (u^2 - 3v^2)^2}{v^2(3 u^2 -
v^2)^2}=\frac{(t^2+4st-s^2)^2(t^2+22st+125s^2)}{s^5t}.
\]
Letting
\[
X:=\frac{t}{s},\quad Y:=\frac{u(u^2-3v^2)}{v(3u^2-v^2)} \cdot
\frac{st}{t^2+4st-s^2}
\]
defines a covering by $C$ of the elliptic curve over $\mathbb{Q}$ given by
\[1728Y^2=(X^2+22X+125)X.\]
By checking that this elliptic curve has rank $0$ and torsion
subgroup of order $2$, we get that the only rational points on $C$
are those with $[s:t]=[1:0]$ or $[s:t]=[0:1]$. For these values of
$[s:t]$ we have $j_5(s,t)=\infty$, so they correspond to cusps on
$X_0(5)$. We conclude that $\rho_l^{E'}$, and hence $\rho_l^E$, is
irreducible for $l=5$. This completes the proof.
\end{proof}
This irreducibility result obviously leads to the following
strengthening of Theorem \ref{thm Chen}.
\begin{theorem}\label{thm Chen refined l>3}
Theorem \ref{thm Chen} holds true with the condition $l>7$
replaced by $l>3$.
\end{theorem}
\subsubsection{$l=5$.}
In order to show that (\ref{eqn 2_2l_3}) has no solutions for
$n=l:=5$, it suffices by Theorem \ref{thm Chen refined l>3} to
show that there exists a $k \in \mathbb{Z}_{>0}$ satisfying the three
conditions of Theorem \ref{thm Chen}. Let $k:=2$, then
$p:=kl+1=11$ is prime, $a_p(E_0)^2 \equiv 1 \not\equiv 4 \pmod{l}$
and $S_{2,p}$ is empty because $1/27\pm 1$ are not squares in
$\mathbb{F}_p$. So the conditions are satisfied for $k=2$ and we conclude
that this proves Theorem \ref{thm main} in case $n=l=5$.
\begin{remark}
By our irreducibility result, Theorem \ref{thm Irreducibility},
the corollary mentioned at the end of Section \ref{sec A modular
approach} also holds with the condition $l>7$ replaced by $l>3$
(the case $l=7$ is in fact trivial, because $7$ is not a Sophie
Germain prime). So for obtaining Theorem \ref{thm main} in case
$n=l:=5$ it sufficed to check that $\left(\frac{p}{7}\right)=1$
and $\left(\frac{p}{13}\right)=(-1)^{(l+1)/2}$ for $p=11$.
\end{remark}
\subsection{Using more local information.}
For $l:=31$ there is no $k \in \mathbb{Z}_{>0}$ known which satisfies the
three conditions of Theorem \ref{thm Chen} (it seems in fact very
unlikely that such a $k$ exists), so that we cannot use this
theorem to prove the nonexistence of solutions of (\ref{eqn
2_2l_3}) for $n=l=31$. This is not too surprising, since, loosely
speaking, $l=31$ is very far from being a Sophie Germain prime in
the sense that the smallest $k \in \mathbb{Z}_{>0}$ such that $kl+1$ is
prime, which is $k=10$, is not so small compared to $l=31$.
Now let $l>3$ be prime, suppose that $(x,y,z)$ is a solution to
(\ref{eqn 2_2l_3}) for $n=l$ and let $u,v,r,s$ be as in Lemma
\ref{lem descent}. The idea is to factor the left-hand side of
$3u^2-v^2=3s^l$ over the ring of integers $R:=\mathbb{Z}[\sqrt{3}]$ in
order to obtain more local information on $u,v$, which, together
with Lemma \ref{lem Modular approach} should lead to a
contradiction. We claim that
\begin{equation}\label{eqn descent 3}
3v=r^l, \quad \sqrt{3}u-v=\sqrt{3}x_1^l \epsilon \quad
\mathrm{and} \quad \sqrt{3}u+v=\sqrt{3}x_2^l \epsilon^{-1}
\end{equation}
for some nonzero $x_1,x_2 \in R$ and unit $\epsilon \in R^*$. Note
that $R$ has class number one. Using $\gcd(u,v)=1$, $2|uv$ and
$3|v$ we readily get $(\sqrt{3}u-v,\sqrt{3}u+v)=(\sqrt{3})$ and
$\sqrt{3}||\sqrt{3}u \pm v$. From
$(\sqrt{3}u-v)(\sqrt{3}u+v)=3s^l$ the claim now follows.
As before, let $E$ be given by (\ref{eqn Frey Curve}), let $E_0$
be any elliptic curves over $\mathbb{Q}$ with conductor $96$, let $k \in
\mathbb{Z}_{>0}$ be such that $p:=kl+1$ is prime and suppose that $a_p(E_0)^2
\not\equiv 4 \pmod{l}$. Note that by Theorem \ref{thm
Irreducibility} we get that $\rho_l^E$ is irreducible. Again, by
Lemma \ref{lem Modular approach} we get $p\nmid
v(3u^2-v^2)=(rs)^l$ and we see that if we can show that
$a_p(E_0)^2 \not\equiv a_p(E)^2 \pmod{l}$, then we reach a
contradiction which shows that (\ref{eqn 2_2l_3}) has no solutions
for $n=l$. Now suppose furthermore that $p$ splits in $R$. Denote
by $\Ideal{p}$ any of the 2 primes of $R$ lying above $p$, for any
$x\in R$ denote by $\overline{x}$ the reduction of $x$ modulo
$\Ideal{p}$ in $R/\Ideal{p} \simeq \mathbb{F}_p$ and set
$r_3:=\overline{\sqrt{3}}$. By reducing (\ref{eqn descent 3})
modulo $\Ideal{p}$ we get
\[
3\overline{v}=\zeta_0, \quad r_3 \overline{u}-\overline{v}=r_3
\zeta_1 \overline{\epsilon}, \quad \mathrm{and} \quad r_3
\overline{u}+\overline{v}=r_3 \zeta_2 \overline{\epsilon}^{-1}
\]
for some $\zeta_0,\zeta_1,\zeta_2 \in \mu_k(\mathbb{F}_p)$. Let
$U:=u/(3v)$ as before, set $\zeta_1':=\zeta_1/\zeta_0,
\zeta_2':=\zeta_2/\zeta_0$ and divide by $3 r_3 \overline{v}=r_3
\zeta_0$ to obtain
\begin{equation}\label{eqn U bar}
\overline{U}-\frac{1}{3r_3}=\zeta_1'\overline{\epsilon},\quad
\overline{U}+\frac{1}{3r_3}=\zeta_2'\overline{\epsilon}^{-1}.
\end{equation}
By Dirichlet's unit theorem $R^*= \langle -1,\epsilon_f \rangle$
for some fundamental unit $\epsilon_f \in R$ (we can take for
example $\epsilon_f=2+\sqrt{3}$). So $\zeta_1'\overline{\epsilon},
\zeta_2'\overline{\epsilon}^{-1}$ belong to the subgroup of
$\mathbb{F}_p^*$ generated by $\mu_k(\mathbb{F}_p)$ and $\overline{\epsilon_f}$,
which we denote by $G_{k,p}$. If $\overline{\epsilon_f} \not \in
\mu_k(\mathbb{F}_p)$ one easily obtains that $G_{k,p}=\mathbb{F}_p^*$. Together
with (\ref{eqn U bar}) this only gives us back the original
information $\overline{U} \in S_{k,p}$. If however
$\overline{\epsilon_f} \in \mu_k(\mathbb{F}_p)$, then of course
$G_{k,p}=\mu_k(\mathbb{F}_p)$ and we get
\begin{equation}\label{eqn Skp bis}
\overline{U} \in S_{k,p}':=(\mu_k(\mathbb{F}_p) +\frac{1}{3r_3}) \cap
(\mu_k(\mathbb{F}_p) -\frac{1}{3r_3}).
\end{equation}
This might be much stronger information, because possibly (and
heuristically speaking, for large $k$ even very likely) $S_{k,p}'$
is strictly contained in $S_{k,p}$. So suppose
$\overline{\epsilon_f}^k=1$. Since
$a_p(E)^2=a_p(E_{\overline{U}})^2$, we reach a contradiction if
for all $\alpha \in S_{k,p}'$ we have $a_p(E_{\alpha})^2
\not\equiv a_p(E_0)^2 \pmod{l}$. We arrive at the following.
\begin{theorem}\label{thm sqrt 3 Kraus}
Let $l>3$ be prime and let $E_0$ be any elliptic curve over $\mathbb{Q}$
with conductor $96$. If there exists a $k \in \mathbb{Z}_{>0}$ such that
the following five conditions hold
\begin{enumerate}
\item $p:=kl+1$ is prime \item $p$ splits in $\mathbb{Z}[\sqrt{3}]$
\item\label{item fundamental unit}
$p|\mathrm{Norm}_{\mathbb{Q}(\sqrt{3})/\mathbb{Q}}\left((2+\sqrt{3})^k-1\right)$ \item
$a_p(E_0)^2 \not\equiv 4 \pmod{l}$ \item $a_p(E_{\alpha})^2
\not\equiv a_p(E_0)^2 \pmod{l}$ for all $\alpha \in S_{k,p}'$,\\
where $S_{k,p}'$ and $E_{\alpha}$ are given by (\ref{eqn Skp bis})
and (\ref{eqn Ealpha}) respectively,
\end{enumerate}
then (\ref{eqn 2_2l_3}) has no solutions for $n=l$.
\end{theorem}
\begin{remark}
Suppose that $k \in \mathbb{Z}_{>0}$ satisfies the first two conditions of
Theorem \ref{thm Chen}. Heuristically speaking, by considering the
expected size of $S_{k,p}$, it seems highly unlikely that if $k
\gtrsim l$, the last condition is satisfied. Now condition
\ref{item fundamental unit} of Theorem \ref{thm sqrt 3 Kraus} is
very restrictive. But if $k \in \mathbb{Z}_{>0}$ does satisfy the first four
conditions of Theorem \ref{thm sqrt 3 Kraus}, then the expected
size of $S'_{k,p}$ is much smaller then that of $S_{k,p}$ and it
only becomes highly unlikely that the last condition is satisfied
when $k \gtrsim l^2$. This is of course all very rough, but the
main idea about the benefits of Theorem \ref{thm sqrt 3 Kraus}
are hopefully clear.
\end{remark}
\subsubsection{$l=7$.}
Let $l:=7$. There are no problems with irreducibility. However, a
large computer search did not reveal a $k \in \mathbb{Z}_{>0}$ satisfying the
three conditions of Theorem \ref{thm Chen} or the five conditions
of Theorem \ref{thm sqrt 3 Kraus} (and we believe that it is very
unlikely that such a $k \in \mathbb{Z}_{>0}$ exists).
Although the modular methods can give us many congruence relations
for a hypothetical solution $(x,y,z)$ of (\ref{eqn 2_2l_3}) with
$n=l=7$, such as $2|x$ and $29|y$, we are not able to show that
(\ref{eqn 2_2l_3}) has no solutions for $n=l=7$ along these lines.
Anyway, in \cite{PSS 07} all finitely many nonzero coprime $a,b,c
\in \mathbb{Z}$ satisfying $a^2+b^3=c^7$ are determined and one readily
checks that for none of the solutions $-c$ is a square. So we
conclude that (\ref{eqn 2_2l_3}) has no solutions for $n=l=7$, as
mentioned before.
\subsubsection{$l=31$.}
Now let $l:=31$, $k:=718$ and $p:=kl+1=22259$. Then one quickly
verifies that $p$ is a prime that splits in $\mathbb{Z}[\sqrt{3}]$ and
that $\overline{\epsilon_f}^k=1$. We calculate $a_p(E_0)=\pm 140$,
so $a_p(E_0)^2 \equiv 8 \not\equiv 4 \pmod{l}$. Finally, the set
$S_{k,p}'$ is easily determined explicitly. All the elements
$\alpha \in S_{k,p}'$ are given in the first column of Table
\ref{Table U ap} together with the corresponding values of
$a_p(E_{\alpha})^2 \pmod{l}$.
\begin{table}[ht]
\begin{center}
\begin{tabular}{c|c}
$\pm \alpha$ & $a_p(E_{\alpha})^2 \pmod{l}$\\
\hline
127 & 28\\
1852 & 19\\
2818 & 1\\
3146 & 10\\
3615 & 9\\
3764 & 16\\
4419 & 18\\
5889 & 25\\
7994 & 20\\
8058 & 0\\
8330 & 19\\
10171 & 18\\
10561 & 5\\
\end{tabular}
\caption{elements of $S_{k,p}'$ with corresponding values of
$a_p(E_{\alpha})^2 \pmod{l}$}\label{Table U ap}
\end{center}
\end{table}
Recall that $a_p(E_0)^2 \equiv 8 \pmod{l}$, so from the second
column of Table \ref{Table U ap} we see that $a_p(E_{\alpha})^2
\not\equiv a_p(E_0)^2 \pmod{l}$ for all $\alpha \in S_{k,p}'$. By
Theorem \ref{thm sqrt 3 Kraus} we can now conclude that (\ref{eqn
2_2l_3}) has no solutions for $n=l=31$. This concludes the proof
of Theorem \ref{thm main}.
\begin{remark}
For $l=31$ the smallest $k \in \mathbb{Z}_{>0}$ satisfying the five conditions
of Theorem \ref{thm sqrt 3 Kraus} is $k=718$. Another $k \in \mathbb{Z}_{>0}$
satisfying these conditions is $k=2542$, and it seems very likely
that there are no other such $k$.
\end{remark}
\section{Extra information from quadratic reciprocity}
In this section we shall prove Theorem \ref{thm n congruence}.
In \cite[Section 4]{Chen Siksek 09}, quadratic reciprocity (over
$\mathbb{Q}$) is used to obtain that there are no nonzero coprime integers
$a,b,c$ satisfying $a^3+b^3=c^n$ for $n \in \mathbb{Z}_{>0}$ with $n\equiv 51,
103 \mathrm{\ or\ } 105 \pmod{106}$. The same method can be
applied to (\ref{eqn 2_2l_3}). The key information obtained is the
following.
\begin{lemma}\label{lem quadratic reciprocity}
Let $l>3$ be prime, suppose that $(x,y,z)$ is a solution to
(\ref{eqn 2_2l_3}) for $n=l$ and let $r,s$ be as in Lemma \ref{lem
descent}. Then $s-r^2$ is a square modulo $7$.
\end{lemma}
\begin{proof}
Let $u,v$ be as in Lemma \ref{lem descent}. Using (\ref{eqn
descent 2}) we get
\begin{equation}\label{eqn quadratic}
s-r^2|3(s^l-r^{2l})=3u^2-28v^2.
\end{equation}
Suppose that $q$ is an odd prime that divides $s-r^2$. Then
$(3u)^2 \equiv 3 \cdot 7 (2v)^2 \pmod{q}.$ From $3v=r^l$ and
$\gcd(r,s)=1$ we obtain $q\nmid 2v$, so $3 \cdot 7$ is a square
modulo $q$, i.e. $\left(\frac{3\cdot7}{q}\right) \in \{0,1\}$. By
quadratic reciprocity we obtain $\left(\frac{q}{3\cdot7}\right)
\in \{0,1\}$. We have $\left(\frac{-1}{3\cdot7}\right)=1$. However
$\left(\frac{2}{3\cdot7}\right)=-1$, but we claim that
$\mathrm{ord}_2(s-r^2)$ is even, so that we get
$\left(\frac{s-r^2}{3\cdot7}\right) \in \{0,1\}$. We claim
furthermore that $\left(\frac{s-r^2}{3}\right)=1$, from which it
now follows that $s-r^2$ is a square modulo $7$. It remain to
prove our two claims. By Theorem \ref{thm Irreducibility} and the
discussion in Section \ref{sec A modular approach}, we know that
$2|u$, so $2\nmid v$. Suppose first that $2||u$, then
$3(u/2)^2-7v^2 \equiv 4 \pmod{8}$, hence
$\mathrm{ord}_2(s^l-r^{2l})=\mathrm{ord}_2(3u^2-28v^2)=4$. Next suppose that
$4|u$, then we immediately get
$\mathrm{ord}_2(s^l-r^{2l})=\mathrm{ord}(3u^2-28v^2)=2$. So in any case,
$\mathrm{ord}_2(s^l-r^{2l})$ is even. From the fact that $r,s$ are both
odd and by counting terms in $(s^l-r^{2l})/(s-r^2)$, we get that
the latter quantity is odd. We conclude that
$\mathrm{ord}_2(s-r^2)=\mathrm{ord}_2(s^l-r^{2l})$ is even, which proves our first
claim. For our second claim, note that $3|r, 3\nmid s$ and $l$ is
odd, so it remains to prove that $s^l$ is a square modulo $3$.
From $3|v$ we get $3|v^2/3$, so from (\ref{eqn descent 2}) we get
$s^l=u^2-v^2/3\equiv u^2 \pmod{3}$. This proves our second claim,
thereby finishing the proof of the lemma.
\end{proof}
From Lemma \ref{lem Modular approach} we obtain some information
on $(s/r^2)^l$ modulo $7$ in a straightforward way.
\begin{lemma}\label{lem straightforward}
Let $l>3$ be prime, suppose that $(x,y,z)$ is a solution to
(\ref{eqn 2_2l_3}) for $n=l$ and let $r,s$ be as in Lemma \ref{lem
descent}. Then $7 \nmid r$ and $(s/r^2)^l \equiv 2 \pmod{7}$
\end{lemma}
\begin{proof}
We shall use Lemma \ref{lem Modular approach} for $p=7$, so let
$u,v,E,E_0$ be as in that lemma. We calculate $a_7(E_0)=\pm 4$, so
$a_7(E_0) \not\equiv \pm 8 \pmod{l}$. Together with the
irreducibility of $\rho_l^E$ form Theorem \ref{thm
Irreducibility}, we get that $7 \nmid v(3u^2-v^2)=(rs)^l$ and
$a_7(E_0)^2 \equiv a_7(E)^2 \pmod{l}$. Let $U:=u/(3v)$ as before
and $\overline{U}$ the reduction of $U$ modulo $7$. Then
$a_7(E_0)^2 \equiv a_7(E_{\overline{U}})^2 \pmod{l}$, where
$E_{\alpha}$ denotes the elliptic curve over $\mathbb{F}_7$ given by
(\ref{eqn Ealpha}) for $\alpha \in \mathbb{F}_7$. We compute that for
$\alpha \not= \pm 1$ we have $a_7(E_0)^2-a_7(E_{\alpha})^2 \in
\{12,16\}$ (for $\alpha=\pm 1$ we have in fact
$a_7(E_0)^2=a_7(E_{\alpha})^2$). We see that if $\alpha \not= \pm
1$, then $a_7(E_0)^2 \not\equiv a_7(E_{\alpha})^2 \pmod{l}$.
Together with $a_7(E_0)^2 \equiv a_7(E_{\overline{U}})^2 \pmod{l}$
we obtain that $\overline{U}=\pm 1$. From (\ref{eqn descent 2}) we
get that
\[\left(\frac{s}{r^2}\right)^l=\frac{3u^2-v^2}{27v^2}=U^2-\frac{1}{27}.\]
Reducing this modulo $7$ and using $\overline{U}=\pm 1$, we arrive
at $(s/r^2)^l \equiv 2 \pmod{7}$.
\end{proof}
Combining the two lemmas above readily leads to the following.
\begin{theorem}\label{thm congruence l}
Let $l$ be prime with $l \equiv -1 \pmod{6}$. Then (\ref{eqn
2_2l_3}) has no solutions for $n=l$.
\end{theorem}
\begin{proof}
Suppose that $(x,y,z)$ is a solution to (\ref{eqn 2_2l_3}) for
$n=l$ and let $r,s$ be as in Lemma \ref{lem descent}. From Lemma
\ref{lem straightforward} and $l\equiv -1 \pmod{6}$, we get
$(s/r^2)^l \equiv (s/r^2)^{-1} \equiv 2 \pmod{7}$. This gives
$s/r^2-1 \equiv 2^{-1}-1 \equiv 3 \pmod{7}$. From Lemma \ref{lem
quadratic reciprocity} we get that $s/r^2-1$ is a square modulo
$7$. However, $3$ is not a square modulo $7$, a contradiction
which proves the theorem.
\end{proof}
Theorem \ref{thm n congruence} now follows by noting that if $n
\equiv -1 \pmod{6}$, then $n$ is divisible by a prime $l \equiv -1
\pmod{6}$.
\begin{remark}
If $l \equiv 1 \pmod{6}$, then Lemma \ref{lem straightforward}
leads to $s/r^2-1 \equiv 1 \pmod{7}$, and since $1$ is a square in
$\mathbb{F}_7$ we do not get a contradiction together with Lemma \ref{lem
quadratic reciprocity}. Furthermore, a similar relation as
(\ref{eqn quadratic}) is given by
$s+r^2|3(s^l+r^{2l})=3u^2+26v^2$, however, similar arguments as
above do not work in this case also.
Finally, in \cite{Chen Siksek 09} not only quadratic reciprocity
over $\mathbb{Q}$ is applied to obtain results about the generalized
Fermat equation $a^3+b^3=c^n$, but also quadratic reciprocity over
(other) number fields. Applying similar methods in our situation
is not so straightforward. This is because $a^3+b^3$ splits into
linear factors in an imaginary quadratic number field, whose ring
of integers has finite unit group, whereas $v(3u^2-v^2)$ splits
into linear factors in a real quadratic number field, whose ring
of integers has infinite unit group. Note that we already dealt
with this infinite unit group in one particular situation, namely
when we solved (\ref{eqn 2_2l_3}) for $n=31$. Trying to approach
(\ref{eqn 2_2l_3}) using quadratic reciprocity over number fields
might be interesting future research.
\end{remark}
\section*{Acknowledgements}
This article contains material from Chapter 3 of the author's
Ph.D. thesis \cite{Dahmen 08}. The author would like to thank
Frits Beukers for many useful discussions related to this
material.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
General Relativity (GR) is the most widely accepted gravity theory proposed by Einstein in 1916, and it has been tested in several field strength regimes being on of the most successful and accurate theories in physics \cite{Will}. The field equations can be obtained using a variational principle, from the well know Einstein-Hilbert action \cite{Misner}-\cite{Weinberg}. The methodology leads to a boundary contribution which is usually dropped out \cite{Carroll},\cite{Hawking}, setting null fluxes through Gauss-Stokes theorem. It can be done by imposing that the variation of the metric and its first derivative vanishes in the boundary \cite{Wald}. These conditions can be relaxed whether a boundary term is introduced, called the Gibbons-York-Hawking boundary term \cite{Hawking1},\cite{Hawking2}. With this boundary term is necessary only to fix the variation of the metric in the boundary. There are some references \cite{Wald}-\cite{Padmanabhan} where this boundary term is shown explicitly. \\\\
However, GR is not the only relativistic theory of gravity. In the last decades several generalizations of Einstein field equations have been proposed \cite{Schmidt}-\cite{Querella}. Within these extended theories of gravity nowadays a subclass, known as $f(R)$ theories, are an alternative for classical problems, as the accelerated expansion of the universe, instead of Dark Energy and Quintessence models \cite{Nojiri}-\cite{Capozziello2}. $f(R)$ theories of gravity are basically extensions of the Einstein-Hilbert action with an arbitrary function of the Ricci scalar $R$ \cite{Faraoni}-\cite{Sotiriou}. The field equations were founded in \cite{Buchdahl}, and including boundary terms in fourth order gravity in \cite{Barth}-\cite{Odintsov2}. The Gibbons-York-Hawking like term in $f(R)$ gravity was explored in \cite{Madsen}, with an augmented variational principle in \cite{Fatibene},\cite{Francaviglia}, and using a scalar-tensor framework in \cite{Casadio}-\cite{Dyer}. Here we obtain the field equations from a metric $f(R)$ action with boundary terms, using only variational principles. We get a well constrained mathematical problem setting $\delta g_{\alpha\beta} = 0$ and $\delta R = 0$ in the boundary.
\section{General Relativity: The Einstein-Hilbert action with the Gibbons-York-Hawking boundary term}
We consider the space-time as a pair $(\mathcal{M},g)$ with $\mathcal{M}$ a four-dimensional manifold and $g$ a metric on $\mathcal{M}$. GR is based on the Einstein's Field equations (without cosmological constant and geometrical units $c=1$), which gives the form of the metric $g_{\alpha\beta}$ on the manifold $\mathcal{M}$:
\begin{equation}\label{campo1}
R_{\alpha\beta} - \frac{1}{2}R g_{\alpha\beta} = \kappa T_{\alpha\beta},
\end{equation}
where $R_{\alpha\beta}=R_{\alpha\eta\beta}^{\eta}$ is the Ricci tensor, $R=R^{\alpha\beta}R_{\alpha\beta}$ the Ricci scalar, and $T_{\alpha\beta}$ the stress-energy tensor, with $\kappa=8\pi G$, and sign convention $(-,+,+,+)$. The Riemann tensor is given by:
\begin{equation}
R_{\beta\gamma\delta}^{\alpha} = \partial_{\gamma}\Gamma_{\delta\beta}^{\alpha} - \partial_{\delta}\Gamma_{\gamma\beta}^{\alpha} + \Gamma_{\gamma\sigma}^{\alpha} \Gamma_{\delta\beta}^{\sigma} - \Gamma_{\sigma\delta}^{\alpha}\Gamma_{\gamma\beta}^{\sigma},
\end{equation}
in terms of the connections $\Gamma_{\beta\gamma}^{\alpha}$. The Einstein field equations can be recovered by using the variational principle $\delta S = 0$, with $S$ expressing the total action. In terms of Einstein-Hilbert action $S_{EH}$, Gibbons-York-Hawking boundary term $S_{GYH}$ and the action associated with all the matter fields $S_{M}$, the total action can be written by \cite{Poisson}:
\begin{equation}
S = \frac{1}{2\kappa}\bigl(S_{EH} + S_{GYH}\bigr) + S_{M},
\end{equation}
where
\begin{equation}\label{eh1}
S_{EH} = \int_{\mathcal{V}} d^4x\, \sqrt{-g}R,
\end{equation}
\begin{equation}
S_{GYH} = 2\oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|}K,
\end{equation}
here $\mathcal{V}$ is a hypervolume on $\mathcal{M}$, $\partial \mathcal{V}$ its boundary, $h$ the determinant of the induced metric, $K$ is the trace of the extrinsic curvature of the boundary $\partial \mathcal{V}$, and $\varepsilon$ is equal to $+1$ if $\partial \mathcal{V}$ is timelike and $-1$ if $\partial \mathcal{V}$ is spacelike (it is assumed that $\partial \mathcal{V}$ is nowhere null). Coordinates $x^{\alpha}$ are used for the finite region $\mathcal{V}$ and $y^{\alpha}$ for the boundary $\partial \mathcal{V}$. Now we will obtain the Einstein field equations varying the action with respect to $g^{\alpha\beta}$. We fixed the variation with the condition \cite{Wald},\cite{Poisson}
\begin{equation}\label{frontera}
\delta g_{\alpha\beta}\biggl|_{\partial \mathcal{V}} =0,
\end{equation}
i.e., the variation of the metric tensor vanishes in the boundary $\partial \mathcal{V}$. We use the results \cite{Poisson},\cite{Carroll}
\begin{equation}\label{varmet1}
\delta g_{\alpha\beta} = -g_{\alpha\mu}g_{\beta\nu}\delta g^{\mu\nu}, \qquad \delta g^{\alpha\beta} = -g^{\alpha\mu}g^{\beta\nu}\delta g_{\mu\nu},
\end{equation}
\begin{equation}
\delta \sqrt{-g} = -\frac{1}{2}\sqrt{-g} g_{\alpha\beta}\delta g^{\alpha\beta},
\end{equation}
\begin{equation}
\delta R_{\beta\gamma\delta}^{\alpha} = \nabla_{\gamma}(\delta\Gamma_{\delta\beta}^{\alpha}) - \nabla_{\delta}(\delta\Gamma_{\gamma\beta}^{\alpha}),
\end{equation}
\begin{equation}\label{palatini}
\delta R_{\alpha\beta} = \nabla_{\gamma}(\delta\Gamma_{\beta\alpha}^{\gamma}) - \nabla_{\beta}(\delta\Gamma_{\gamma\alpha}^{\gamma}).
\end{equation}
We give a detailed review for the variation principles in GR following \cite{Wald},\cite{Poisson} and \cite{Carroll},. The variation of the Einstein-Hilbert term gives
\begin{equation}\label{varaccion1}
\delta S_{EH} = \int_{\mathcal{V}} d^4x \, \bigl(R\delta\sqrt{-g} + \sqrt{-g}\, \delta R\bigr).
\end{equation}
Now with $R = g^{\alpha\beta}R_{\alpha\beta}$, we have that the variation of the Ricci scalar is
\begin{equation}
\delta R = \delta g^{\alpha\beta}R_{\alpha\beta} + g^{\alpha\beta}\delta R_{\alpha\beta}.
\end{equation}
using the Palatini's identity (\ref{palatini}) we can write \cite{Carroll}:
\begin{align}
\delta R &= \delta g^{\alpha\beta}R_{\alpha\beta} + g^{\alpha\beta}\bigl(\nabla_{\gamma}(\delta\Gamma_{\beta\alpha}^{\gamma}) - \nabla_{\beta}(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr),\nonumber \\
&= \delta g^{\alpha\beta}R_{\alpha\beta} + \nabla_{\sigma
}\bigl(g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr),
\end{align}
where we have used the metric compatibility $\nabla_{\gamma}g_{\alpha\beta}\equiv 0$ and relabeled some dummy indices. Inserting this results for the variations in expression (\ref{varaccion1}) we have:
\begin{align}
\delta S_{EH} &= \int_{\mathcal{V}} d^4x \, \bigl(R\delta\sqrt{-g} + \sqrt{-g}\, \delta R\bigr),\nonumber \\
&= \int_{\mathcal{V}} d^4x \, \biggl(-\frac{1}{2}Rg_{\alpha\beta}\sqrt{-g}\, \delta g^{\alpha\beta} + R_{\alpha\beta}\sqrt{-g}\delta g^{\alpha\beta} + \sqrt{-g}\nabla_{\sigma}\bigl(g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr)\biggr),\nonumber \\
&= \int_{\mathcal{V}} d^4x \, \sqrt{-g} \biggl(R_{\alpha\beta}-\frac{1}{2}Rg_{\alpha\beta}\biggr)\delta g^{\alpha\beta} + \int_{\mathcal{V}} d^4x\sqrt{-g} \nabla_{\sigma}\bigl(g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr).
\end{align}
Denoting the divergence term with $\delta S_{B}$,
\begin{equation}
\delta S_B = \int_{\mathcal{V}} d^4x \, \sqrt{-g}\, \nabla_{\sigma}\bigl(g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr),
\end{equation}
we define
\begin{equation}\label{V}
V^{\sigma} = g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma}),
\end{equation}
then the boundary term can be written as
\begin{equation}\label{boundary}
\delta S_B = \int_{\mathcal{V}} d^4x \, \sqrt{-g}\, \nabla_{\sigma}V^{\sigma}.
\end{equation}
Using Gauss-Stokes theorem \cite{Poisson},\cite{Carroll}:
\begin{equation}\label{gauss}
\int_{\mathcal{V}} d^{n}x\, \sqrt{|g|}\nabla_{\mu}A^{\mu} = \oint_{\partial \mathcal{V}}d^{n-1}y\, \varepsilon\sqrt{|h|}n_{\mu}A^{\mu},
\end{equation}
where $n_{\mu}$ is the unit normal to $\partial \mathcal{V}$. Using this we can write (\ref{boundary}) in the following boundary term
\begin{equation}
\delta S_B = \oint_{\partial \mathcal{V}} d^{3}y\, \varepsilon \sqrt{|h|}n_{\sigma}V^{\sigma},
\end{equation}
with $V^{\sigma}$ given in (\ref{V}). The variation $\delta \Gamma_{\beta\alpha}^{\sigma}$ is obtained by using that $\Gamma_{\beta\alpha}^{\sigma}$ is the Christoffel symbol $\bigl\{_{\beta\alpha}^{\sigma}\bigr\}$:
\begin{equation}\label{simbolo}
\Gamma_{\beta\gamma}^{\alpha} \equiv \Bigl\{_{\beta\gamma}^{\alpha}\Bigr\} = \frac{1}{2}g^{\alpha\sigma}\bigl[\partial_{\beta}g_{\sigma\gamma} + \partial_{\gamma}g_{\sigma\beta} - \partial_{\sigma}g_{\beta\gamma}\bigr],
\end{equation}
getting
\begin{align}\label{varsim1}
\delta \Gamma_{\beta\alpha}^{\sigma} &= \delta \biggl(\frac{1}{2}g^{\sigma\gamma}\bigl[\partial_{\beta}g_{\gamma\alpha} + \partial_{\alpha}g_{\gamma\beta} - \partial_{\gamma}g_{\beta\alpha}\bigr]\biggr),\nonumber \\
&= \frac{1}{2}\delta g^{\sigma\gamma}\bigl[\partial_{\beta}g_{\gamma\alpha} + \partial_{\alpha}g_{\gamma\beta} - \partial_{\gamma}g_{\beta\alpha}\bigr] + \frac{1}{2}g^{\sigma\gamma}\bigl[\partial_{\beta}(\delta g_{\gamma\alpha}) + \partial_{\alpha}(\delta g_{\gamma\beta}) - \partial_{\gamma}(\delta g_{\beta\alpha})\bigr].
\end{align}
From the boundary conditions $\delta g_{\alpha\beta} = \delta g^{\alpha\beta}=0$ the variation (\ref{varsim1}) gives:
\begin{equation}
\delta \Gamma_{\beta\alpha}^{\sigma}\Bigl|_{\partial \mathcal{V}} = \frac{1}{2}g^{\sigma\gamma}\bigl[\partial_{\beta}(\delta g_{\gamma\alpha}) + \partial_{\alpha}(\delta g_{\gamma\beta}) - \partial_{\gamma}(\delta g_{\beta\alpha})\bigr],
\end{equation}
and
\begin{equation}
V^{\mu}\Bigl|_{\partial \mathcal{V}} = g^{\alpha\beta}\biggl[\frac{1}{2}g^{\mu\gamma}\bigl[\partial_{\beta}(\delta g_{\gamma\alpha}) + \partial_{\alpha}(\delta g_{\gamma\beta}) - \partial_{\gamma}(\delta g_{\beta\alpha})\bigr]\biggr] - g^{\alpha\mu}\biggl[\frac{1}{2}g^{\nu\gamma}\partial_{\alpha}(\delta g_{\nu\gamma}) \biggr],
\end{equation}
we can write
\begin{align}
V_{\sigma}\Bigl|_{\partial \mathcal{V}} = g_{\sigma\mu}V^{\mu}\Bigl|_{\partial \mathcal{V}} &= g_{\sigma\mu}g^{\alpha\beta}\biggl[\frac{1}{2}g^{\mu\gamma}\bigl[\partial_{\beta}(\delta g_{\gamma\alpha}) + \partial_{\alpha}(\delta g_{\gamma\beta}) - \partial_{\gamma}(\delta g_{\beta\alpha})\bigr]\biggr] - g_{\sigma\mu}g^{\alpha\mu}\biggl[\frac{1}{2}g^{\nu\gamma}\partial_{\alpha}(\delta g_{\nu\gamma}) \biggr], \nonumber \\
&= \frac{1}{2}\delta_{\sigma}^{\gamma}g^{\alpha\beta}\bigl[\partial_{\beta}(\delta g_{\gamma\alpha}) + \partial_{\alpha}(\delta g_{\gamma\beta}) - \partial_{\gamma}(\delta g_{\beta\alpha})\bigr] - \frac{1}{2}\delta_{\sigma}^{\alpha}g^{\nu\gamma}\bigl[\partial_{\alpha}(\delta g_{\nu\gamma}) \bigr],\nonumber \\
&= g^{\alpha\beta}\bigl[\partial_{\beta}(\delta g_{\sigma\alpha}) - \partial_{\sigma}(\delta g_{\beta\alpha})\bigr].
\end{align}
We now evaluate the term $n^{\sigma}V_{\sigma}\bigl|_{\partial \mathcal{V}}$ by using for this that
\begin{equation}
g^{\alpha\beta} = h^{\alpha\beta} + \varepsilon n^{\alpha}n^{\beta},
\end{equation}
then
\begin{align}
n^{\sigma}V_{\sigma}\Bigl|_{\partial \mathcal{V}} &= n^{\sigma}(h^{\alpha\beta}+\varepsilon n^{\alpha}n^{\beta})[\partial_{\beta}(\delta g_{\sigma\alpha}) - \partial_{\sigma}(\delta g_{\beta\alpha})], \nonumber \\
&= n^{\sigma}h^{\alpha\beta}[\partial_{\beta}(\delta g_{\sigma\alpha}) - \partial_{\sigma}(\delta g_{\beta\alpha})],
\end{align}
where we use the antisymmetric part of $\varepsilon n^{\alpha}n^{\beta}$ with $\varepsilon = n^{\mu}n_{\mu}=\pm 1$. To the fact $\delta g_{\alpha\beta}=0$ in the boundary we have $h^{\alpha\beta}\partial_{\beta}(\delta g_{\sigma\alpha})=0$ \cite{Poisson}. Finally we get
\begin{equation}
n^{\sigma}V_{\sigma}\Bigl|_{\partial \mathcal{V}} = -n^{\sigma}h^{\alpha\beta}\partial_{\sigma}(\delta g_{\beta\alpha}).
\end{equation}
Thus the variation of the Einstein-Hilbert term is:
\begin{equation}
\delta S_{EH} = \int_{\mathcal{V}} d^4x \, \sqrt{-g} \biggl(R_{\alpha\beta}-\frac{1}{2}Rg_{\alpha\beta}\biggr)\delta g^{\alpha\beta}-\oint_{\partial \mathcal{V}} d^{3}y\, \varepsilon \sqrt{|h|} h^{\alpha\beta}\partial_{\sigma}(\delta g_{\beta\alpha})n^{\sigma}.
\end{equation}
Now we consider the variation of the Gibbons-York-Hawking boundary term:\\
\begin{equation}
\delta S_{GYH} = 2\oint_{\partial \mathcal{V}} d^3y\, \varepsilon\sqrt{|h|}\delta K.
\end{equation}
Using the definition of the trace of extrinsic curvature \cite{Poisson}:
\begin{align}
K &= \nabla_{\alpha}n^{\alpha}, \nonumber \\
&= g^{\alpha\beta}\nabla_{\beta}n_{\alpha}, \nonumber \\
&= (h^{\alpha\beta}+\varepsilon n^{\alpha}n^{\beta})\nabla_{\beta}n_{\alpha}, \nonumber \\
&= h^{\alpha\beta}\nabla_{\beta}n_{\alpha}, \nonumber \\
&= h^{\alpha\beta}(\partial_{\beta}n_{\alpha}-\Gamma_{\beta\alpha}^{\gamma}n_{\gamma}),
\end{align}
the variation is
\begin{align}\label{deltaK}
\delta K &= -h^{\alpha\beta}\delta\Gamma_{\beta\alpha}^{\gamma}n_{\gamma}, \nonumber \\
&= -\frac{1}{2}h^{\alpha\beta}g^{\sigma\gamma}\bigl[\partial_{\beta}(\delta g_{\sigma\alpha}) + \partial_{\alpha}(\delta g_{\sigma\beta}) - \partial_{\sigma}(\delta g_{\beta\alpha})\bigr]n_{\gamma}, \nonumber \\
&= -\frac{1}{2}h^{\alpha\beta}\bigl[\partial_{\beta}(\delta g_{\sigma\alpha}) + \partial_{\alpha}(\delta g_{\sigma\beta}) - \partial_{\sigma}(\delta g_{\beta\alpha})\bigr]n^{\sigma}, \nonumber \\
&= \frac{1}{2}h^{\alpha\beta}\partial_{\sigma}(\delta g_{\beta\alpha})n^{\sigma}.
\end{align}
This comes from the variation $\delta\Gamma_{\beta\alpha}^{\gamma}$ evaluated in the boundary, and the fact that $h^{\alpha\beta}\partial_{\beta}(\delta g_{\sigma\alpha})=0$, $h^{\alpha\beta}\partial_{\alpha}(\delta g_{\sigma\beta})=0$. Then we have for the variation of the Gibbons-York-Hawking boundary term:
\begin{equation}
\delta S_{GYH} = \oint_{\partial \mathcal{V}} d^3y\, \varepsilon\sqrt{|h|}h^{\alpha\beta}\partial_{\sigma}(\delta g_{\beta\alpha})n^{\sigma}.
\end{equation}
We see that this term exactly cancel the boundary contribution of the Einstein-Hilbert term. Now, if we have a matter action defined by:
\begin{equation}\label{matteraction}
S_M = \int_{\mathcal{V}} d^4x\, \sqrt{-g} \mathcal{L}_M[g_{\alpha\beta},\psi],
\end{equation}
where $\psi$ denotes the matter fields. The variation of this action takes the form:
\begin{align}
\delta S_M &= \int_{\mathcal{V}} d^4x\, \delta(\sqrt{-g} \mathcal{L}_M),\nonumber \\
&= \int_{\mathcal{V}} d^4x\, \biggl(\frac{\partial \mathcal{L}_M}{\partial g^{\alpha\beta}}\delta g^{\alpha\beta}\sqrt{-g} + \mathcal{L}_M\delta\sqrt{-g}\biggr),\nonumber \\
&= \int_{\mathcal{V}} d^4x\,\sqrt{-g} \biggl(\frac{\partial \mathcal{L}_M}{\partial g^{\alpha\beta}} -\frac{1}{2} \mathcal{L}_Mg_{\alpha\beta}\biggr)\delta g^{\alpha\beta},
\end{align}
as usual, defining the stress-energy tensor by:
\begin{equation}\label{tensem}
T_{\alpha\beta} \equiv -2\frac{\partial \mathcal{L}_M}{\partial g^{\alpha\beta}} + \mathcal{L}_Mg_{\alpha\beta} = -\frac{2}{\sqrt{-g}}\frac{\delta S_M}{\delta g^{\alpha\beta}},
\end{equation}
then:
\begin{equation}\label{variacionener}
\delta S_M = -\frac{1}{2}\int_{\mathcal{V}} d^4x\,\sqrt{-g} T_{\alpha\beta}\delta g^{\alpha\beta},
\end{equation}
imposing the total variations remains invariant with respect to $\delta g^{\alpha\beta}$. Finally the equations are writing as:
\begin{equation}\label{campo1}
\frac{1}{\sqrt{-g}}\frac{\delta S}{\delta g^{\alpha\beta}} = 0, \Longrightarrow R_{\alpha\beta} - \frac{1}{2}R g_{\alpha\beta} = \kappa T_{\alpha\beta} ,
\end{equation}
which corresponds to Einstein field equations in geometric units $c=1$.
\section{Field equations in $f(R)$ gravity}
As we mentioned above the modified theories of gravity have been studied in order to explain among the accelerated expansion of the universe. One of these theories is the modified $f(R)$ gravity which consists in add additional higher order terms of the Ricci scalar in the Einstein-Hilbert action \cite{Nojiri},\cite{Sami},\cite{Sotiriou}. There are three versions of $f(R)$ gravity: Metric formalism, Palatini formalism and metric-affine formalism \cite{Sotiriou}. Here we focus only in the metric formalism; for a detailed deduction of field equations in the Palatini and the metric-affine formalism see \cite{Sotiriou3},\cite{Sotiriou4}. Again, we consider the space-time as a pair $(\mathcal{M},g)$ with $\mathcal{M}$ a four-dimensional manifold and $g_{\alpha\beta}$ a metric on $\mathcal{M}$. Now the lagrangian is an arbitrary function of the Ricci scalar $\mathcal{L}[g_{\alpha\beta}] = f(R)$, the relation of the Ricci scalar and the metric tensor is given assuming a Levi-Civita connection of the manifold. i.e. a Christoffel symbol. This lagrangian was presented in \cite{Fatibene} using augmented variational principles. The general action can be written as \cite{Dyer}:
\begin{equation}
S_{mod} = \frac{1}{2\kappa}\bigl(S_{met} + S'_{GYH}\bigr) + S_{M},
\end{equation}
with the bulk term
\begin{equation} \label{accion1}
S_{met} = \int_{\mathcal{V}} d^4x\, \sqrt{-g}f(R),
\end{equation}\\
and the Gibbons-York-Hawking like boundary term \cite{Madsen},\cite{Dyer}
\begin{equation}
S'_{GYH} = 2\oint_{\partial \mathcal{V}} d^3y\, \varepsilon\sqrt{|h|} f'(R)K,
\end{equation}
with $f'(R) = df(R)/dR$. Again, $S_M$ represents the action associated with all the matter fields (\ref{matteraction}). We fixed the variation to the condition
\begin{equation}\label{frontera2}
\delta g_{\alpha\beta}\biggl|_{\partial \mathcal{V}} =0.
\end{equation}
First, the variation of the bulk term is:
\begin{equation}\label{varaccion2}
\delta S_{met} = \int_{\mathcal{V}} d^4x \, \bigl(f(R)\delta\sqrt{-g} + \sqrt{-g}\, \delta f(R)\bigr),
\end{equation}
and the functional derivative of the $f(R)$ term can be written as
\begin{equation}
\delta f(R) = f'(R) \delta R.
\end{equation}
Using the expression for the variation of the Ricci scalar:
\begin{equation}
\delta R = \delta g^{\alpha\beta}R_{\alpha\beta} + \nabla_{\sigma}\bigl(g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr),
\end{equation}
where the variation of the term $g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})$ is given in \ref{Appendix A}.
With this result the variation of the Ricci scalar becomes
\begin{align}
\delta R &= \delta g^{\alpha\beta}R_{\alpha\beta} + \nabla_{\sigma
}\bigl(g^{\alpha\beta}(\delta\Gamma_{\beta\alpha}^{\sigma}) - g^{\alpha\sigma}
(\delta\Gamma_{\alpha\gamma}^{\gamma})\bigr),\nonumber \\
&= \delta g^{\alpha\beta}R_{\alpha\beta} + g_{\mu\nu}\nabla_{\sigma}\nabla^{\sigma}(\delta g^{\mu\nu}) - \nabla_{\sigma}\nabla_{\gamma}(\delta g^{\sigma\gamma}),\nonumber \\
&= \delta g^{\alpha\beta}R_{\alpha\beta} + g_{\alpha\beta}\square(\delta g^{\alpha\beta}) - \nabla_{\alpha}\nabla_{\beta}(\delta g^{\alpha\beta}).
\end{align}
Here we define $\square \equiv\nabla_{\sigma}\nabla^{\sigma}$ and relabeled some indices. Putting the previous results together in the variation of the modified action (\ref{varaccion2}):
\begin{align}\label{varaccionf}
\delta S_{met} &= \int_{\mathcal{V}} d^4x \, \bigl(f(R)\delta\sqrt{-g} + \sqrt{-g}\, f'(R) \delta R\bigr), \nonumber \\
&= \int_{\mathcal{V}} d^4x \, \biggl(-f(R)\frac{1}{2}\sqrt{-g}\, g_{\alpha\beta}\delta g^{\alpha\beta} + f'(R)\sqrt{-g}\Bigl(\delta g^{\alpha\beta}R_{\alpha\beta} + g_{\alpha\beta}\square(\delta g^{\alpha\beta}) - \nabla_{\alpha}\nabla_{\beta}(\delta g^{\alpha\beta})\Bigr)\biggr),\nonumber \\\
&= \int_{\mathcal{V}} d^4x \, \sqrt{-g} \biggl(f'(R)\Bigl(\delta g^{\alpha\beta}R_{\alpha\beta} + g_{\alpha\beta}\square(\delta g^{\alpha\beta}) - \nabla_{\alpha}\nabla_{\beta}(\delta g^{\alpha\beta})\Bigr)-f(R)\frac{1}{2}\, g_{\alpha\beta}\delta g^{\alpha\beta}\biggr).
\end{align}
Now we will consider the next integrals:
\begin{equation}\label{integrals}
\int_{\mathcal{V}} d^4x \, \sqrt{-g} f'(R)g_{\alpha\beta}\square(\delta g^{\alpha\beta}), \qquad \int_{\mathcal{V}} d^4x \, \sqrt{-g} f'(R)\nabla_{\alpha}\nabla_{\beta}(\delta g^{\alpha\beta}).
\end{equation}
We shall see that these integrals can be expressed differently performing integration by parts. For this we define the next quantities:
\begin{equation}\label{M}
M_{\tau} = f'(R)g_{\alpha\beta}\nabla_{\tau}(\delta g^{\alpha\beta}) - \delta g^{\alpha\beta} g_{\alpha\beta}\nabla_{\tau}(f'(R)),
\end{equation}
and
\begin{equation}\label{N}
N^{\sigma} = f'(R)\nabla_{\gamma}(\delta g^{\sigma\gamma}) - \delta g^{\sigma\gamma}\nabla_{\gamma}(f'(R)).
\end{equation}
The combination $g^{\sigma\tau}M_{\tau} + N^{\sigma}$ is
\begin{equation}
g^{\sigma\tau}M_{\tau} + N^{\sigma} = f'(R)g_{\alpha\beta}\nabla^{\sigma}(\delta g^{\alpha\beta}) - \delta g^{\alpha\beta} g_{\alpha\beta}\nabla^{\sigma}(f'(R)) +
f'(R)\nabla_{\gamma}(\delta g^{\sigma\gamma}) - \delta g^{\sigma\gamma}\nabla_{\gamma}(f'(R)),
\end{equation}
in the particular case $f(R) = R$, the previous combination reduces to the expression (\ref{V}) with equation (\ref{vargamma}). The quantities $M_{\tau}$ and $N^{\sigma}$ allow us to write the variation of the bulk term (\ref{varaccionf}) in the following way (for details see \ref{Appendix B}):
\begin{multline}\label{varf1}
\delta S_{met} = \int_{\mathcal{V}} d^4x \, \sqrt{-g} \biggl(f'(R)R_{\alpha\beta} + g_{\alpha\beta}\square f'(R) - \nabla_{\alpha}\nabla_{\beta}f'(R)-f(R)\frac{1}{2}\, g_{\alpha\beta}\biggr)\delta g^{\alpha\beta}\\ + \oint_{\partial \mathcal{V}} d^{3}y\, \varepsilon\sqrt{|h|}n^{\tau}M_{\tau} + \oint_{\partial \mathcal{V}} d^{3}y\, \varepsilon\sqrt{|h|}n_{\sigma}N^{\sigma}.
\end{multline}
In the next section we will work out with the boundary contribution from (\ref{varf1}), and show how this terms cancel with the variations of the $S'_{GYH}$ action.
\subsection{Boundary terms in $f(R)$ gravity}
We express the quantities $M_{\tau}$ and $N^{\sigma}$ calculated in the boundary $\partial \mathcal{V}$. Is convenient to express them in function of the variations $\delta g_{\alpha\beta}$. Using the equation (\ref{varmet1}) in (\ref{M}) and (\ref{N}) yields :
\begin{equation}\label{M1}
M_{\tau} = -f'(R)g^{\alpha\beta}\nabla_{\tau}(\delta g_{\alpha\beta}) + g^{\alpha\beta}\delta g_{\alpha\beta} \nabla_{\tau}(f'(R)),
\end{equation}
and
\begin{equation}\label{N1}
N^{\sigma} = -f'(R)g^{\sigma\mu}g^{\gamma\nu}\nabla_{\gamma}(\delta g_{\mu\nu}) + g^{\sigma\mu}g^{\gamma\nu}\delta g_{\mu\nu}\nabla_{\gamma}(f'(R)).
\end{equation}
To evaluate this quantities in the boundary we use the fact that $\delta g_{\alpha\beta}|_{\partial \mathcal{V}}=\delta g^{\alpha\beta}|_{\partial \mathcal{V}}=0$, then the only terms not vanishing are the derivatives of $\delta g_{\alpha\beta}$ in the covariant derivatives. Hence we have
\begin{equation}\label{M2}
M_{\tau}\biggl|_{\partial \mathcal{V}} =-f'(R)g^{\alpha\beta}\partial_{\tau}(\delta g_{\alpha\beta}),
\end{equation}
and
\begin{equation}\label{M1}
N^{\sigma}\biggl|_{\partial \mathcal{V}} = -f'(R)g^{\sigma\mu}g^{\gamma\nu}\partial_{\gamma}(\delta g_{\mu\nu}),
\end{equation}
We now compute $n^{\tau}M_{\tau}\bigl|_{\partial \mathcal{V}}$ and $n_{\sigma}N^{\sigma}\bigl|_{\partial \mathcal{V}}$ which are the terms in the boundary integrals (\ref{varf1})
\begin{align}
n^{\tau}M_{\tau}\biggl|_{\partial \mathcal{V}} &= -f'(R)n^{\tau}(\varepsilon n^{\alpha}n^{\beta}+h^{\alpha\beta})\partial_{\tau}(\delta g_{\alpha\beta}), \nonumber\\
&= -f'(R)n^{\sigma}h^{\alpha\beta}\partial_{\sigma}(\delta g_{\alpha\beta}),
\end{align}
where we rename the dummy index $\tau$. In the other hand
\begin{align}
n_{\sigma}N^{\sigma}\biggl|_{\partial \mathcal{V}} &= -f'(R)n_{\sigma}(h^{\sigma\mu}+\varepsilon n^{\sigma}n^{\mu})(h^{\gamma\nu}+\varepsilon n^{\gamma}n^{\nu})\partial_{\gamma}(\delta g_{\mu\nu}), \nonumber\\
&= -f'(R)n^{\mu}(h^{\gamma\nu}+\varepsilon n^{\gamma}n^{\nu})\partial_{\gamma}(\delta g_{\mu\nu}), \nonumber\\
&= -f'(R)n^{\mu}h^{\gamma\nu}\partial_{\gamma}(\delta g_{\mu\nu}) \nonumber\\
&= 0,
\end{align}
where we have used that $n_{\sigma}h^{\sigma\mu}=0$, $\varepsilon^2=1$ and the fact that de tangential derivative $h^{\gamma\nu}\partial_{\gamma}(\delta g_{\mu\nu})$ vanishes. With this results the variation of the action $S_{met}$ becomes:
\begin{multline}\label{varf2}
\delta S_{met} = \int_{\mathcal{V}} d^4x \, \sqrt{-g} \biggl(f'(R)R_{\alpha\beta} + g_{\alpha\beta}\square f'(R) - \nabla_{\alpha}\nabla_{\beta}f'(R)-f(R)\frac{1}{2}\, g_{\alpha\beta}\biggr)\delta g^{\alpha\beta}\\ -\oint_{\partial \mathcal{V}} d^{3}y\, \varepsilon\sqrt{|h|}f'(R)n^{\sigma}h^{\alpha\beta}\partial_{\sigma}(\delta g_{\alpha\beta}).
\end{multline}
We proceed with the boundary term $S'_{GYH}$ in the total action. The variation of this term gives
\begin{align}
\delta S'_{GYH} &= 2\oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|}\bigl(\delta f'(R)K + f'(R)\delta K\bigr), \nonumber \\
&= 2\oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|}\bigl(f''(R)\delta R\, K + f'(R)\delta K\bigr).
\end{align}
Using the expression for the variation of $K$, equation (\ref{deltaK}), we can write\\
\begin{align}\label{varfront}
\delta S'_{GYH} &= 2\oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|}\biggl(f''(R)\delta R\, K + \frac{1}{2}f'(R)h^{\alpha\beta}\partial_{\sigma}(\delta g_{\beta\alpha})n^{\sigma}\biggr), \nonumber \\
&= 2\oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|}f''(R)\delta R\, K + \oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|}f'(R)h^{\alpha\beta}\partial_{\sigma}(\delta g_{\beta\alpha})n^{\sigma}.
\end{align}
We see that the second term in (\ref{varfront}) cancels the boundary term in the variation (\ref{varf2}), and in addition we need to impose $\delta R = 0$ in the boundary. Similar argument is given in \cite{Dyer}.\\\\
Finally, with the variation of the matter action, given in (\ref{variacionener}), the total variation of the action of modified $f(R)$ gravity is:\\
\begin{multline}
\delta S_{mod} = \frac{1}{2\kappa}\int_{\mathcal{V}} d^4x \, \sqrt{-g} \biggl(f'(R)R_{\alpha\beta} + g_{\alpha\beta}\square f'(R) - \nabla_{\alpha}\nabla_{\beta}f'(R)-\frac{1}{2}f(R)\, g_{\alpha\beta}\biggr)\delta g^{\alpha\beta}\\ -\frac{1}{2}\int_{\mathcal{V}} d^4x\,\sqrt{-g} T_{\alpha\beta}\delta g^{\alpha\beta}.
\end{multline}
Imposing that this variation becomes stationary we have:\\
\begin{equation}\label{variacionfinal}
\frac{1}{\sqrt{-g}}\frac{\delta S_{mod}}{\delta g^{\alpha\beta}} = 0 \Longrightarrow f'(R)R_{\alpha\beta} + g_{\alpha\beta}\square f'(R) - \nabla_{\alpha}\nabla_{\beta}f'(R)-\frac{1}{2}f(R)\, g_{\alpha\beta} = \kappa T_{\alpha\beta},
\end{equation}\\
which are the field equations in the metric formalism of $f(R)$ gravity.
\section{Conclusions}
We have obtained the field equations in the metric formalism of $f(R)$ gravity by using the direct results from variational principles.
The modified action in the metric formalism of $f(R)$ gravity plus a Gibbons-York-Hawking like boundary term must be written as:
\begin{equation}
S_{mod} = \frac{1}{2\kappa}\biggl[\int_{\mathcal{V}} d^4x\, \sqrt{-g}\Bigl(f(R) + 2\kappa\mathcal{L}_M[g_{\alpha\beta},\psi]\Bigl)+ 2\oint_{\partial \mathcal{V}} d^3y \, \varepsilon\sqrt{|h|} f'(R)K\biggr],
\end{equation}
with $f'(R) = df(R)/dR$ and $\mathcal{L}_M$ the lagrangian associated with all the matter fields. From the quantities $M_{\sigma}$ and $N^{\sigma}$, defined in (\ref{M}) and (\ref{N}) respectively, we recovered GR plus Gibbons-York-Hawking boundary term in the particular case $f(R)=R$. We see that including the boundary term, we have a well behaved mathematical problem setting both, $\delta g_{\alpha\beta}=0$ and $\delta R = 0$ in $\partial \mathcal{V}$.\\\\\\\\\\
\textbf{\large{Acknowledgements}:} The authors are grateful with the Observatorio Astron\'omico Nacional, Bogot\'a, Colombia, where this paper was carried out. A. Guarnizo acknowledges the financial support by the Programa de Becas para Estudiantes Sobresalientes de Posgrado, Universidad Nacional de Colombia.
\newpage
|
\section{Introduction}
After the breakthrough paper by Alon and Milman \cite{AM} in 1985,
it became apparent that regular graphs, whose \DEF{spectral gap}
(i.e.~the difference between the largest and the second largest eigenvalue)
is large, posses some extraordinary properties, like unusually fast expansion
and resemblance to random graphs. This led to the definition of
\DEF{Ramanujan graphs}. These are $d$-regular graphs whose second largest
eigenvalue does not exceed the value $2\sqrt{d-1}$.
Lubotzky, Phillips, and Sarnak \cite{LPS},
and independently Margulis \cite{Ma}, were the first to show that Ramanujan graphs
exist. Their constructions are based on number theory and work when
the degree $d$ is equal to $p+1$ for some prime $p$. Later, several new
constructions were discovered, showing that Ramanujan graphs exist for every
degree $d\ge3$ which is of the form $p^k+1$ for some integer $k\ge1$ and some
prime $p$, see Morgenstern \cite{Morg}.
It is not immediately clear why the special choice of $2\sqrt{d-1}$ is taken
when defining Ramanujan graphs. One reason is that this is the spectral radius
of the infinite $d$-regular tree, which is the universal cover for all
$d$-regular graphs. Another reason is the following result
of Alon and Boppana (see~\cite{Al}) which shows that this is the smallest
number that makes sense.
\begin{theorem}[Alon-Boppana]
\label{thm:AB}
For every $d\ge2$ and every $\varepsilon>0$, there are only finitely many
$d$-regular graphs whose second largest eigenvalue is at most
$2\sqrt{d-1}-\varepsilon$.
\end{theorem}
Alternative proofs of Theorem \ref{thm:AB} were given by Friedman \cite{Fr}
and by Nilli \cite{Ni1}, who has recently further simplified her arguments
in \cite{Ni2}. Actually, the proofs in \cite{Fr,Ni2} imply a stronger version
of Theorem \ref{thm:AB} by making the same conclusion for more
eigenvalues than just the second largest one.
This strengthening is attributed to Serre \cite{Se} (see also \cite{DSV,FL,HLW}),
who stated the following quantitative version of the Alon-Boppana Theorem:
\begin{theorem}[Serre]
\label{thm:Serre}
For every positive integer $d$ and every $\varepsilon>0$, there exists a constant
$c=c(d,\varepsilon)$ such that every $d$-regular graph of order $n$
has at least $cn$ eigenvalues that are larger than $2\sqrt{d-1}-\varepsilon$.
\end{theorem}
In this paper we give a multipartite generalization of the Alon-Boppana
Theorem, see Theorem \ref{thm:main}. The Ramanujan value $2\sqrt{d-1}$
is replaced by the spectral radius of the universal covering
tree of the multipartite parameters (cf.\ Section~\ref{sect:2} for definitions).
Our proof has similarities with
Nilli's proof \cite{Ni2}, and seems to be even simpler if restricted to the
special case of $d$-regular graphs. The main step is
based on the interlacing theorem and is entirely elementary.
Our proofs work also in the case when graphs are not regular.
In the simplest, monopartite case, our result extends the Alon-Boppana-Serre result
to non-regular graphs of minimum degree $d$ and bounded maximum degree.
A strengthening of this form has been obtained previously by Hoory \cite{Ho}.
In the next simplest two-partite case it is shown that for every $\varepsilon>0$ and any positive
integers $d_1,d_2,d$, every $n$-vertex graph of maximum degree
at most $d$, whose vertex set is the union of (not necessarily disjoint) subsets
$V_1,V_2$, such that every vertex in $V_i$ has at least $d_i$
neighbors in $V_{3-i}$ for $i=1,2$, has $\Omega_\varepsilon(n)$ eigenvalues that are larger than $\sqrt{d_1-1}+\sqrt{d_2-1}-\varepsilon$.
After submission of this paper, S. Cioab\v a informed us about some related
work. Greenberg \cite{Gr} obtained a generalized version of the Serre theorem in a similar form as ours, but only claiming that there are eigenvalues whose absolute value is larger than $2\sqrt{d-1}-\varepsilon$. Cioab\v a \cite{Ci}
improved Greenberg's work to the same form as given in Theorem \ref{thm:Serre}. Greenberg's result also appears in \cite{LN}.
In the last section we tailor the proofs to obtain a strengthening
of the Alon-Boppana-Serre theorem by showing that the lower
bound $2\sqrt{d-1}-\varepsilon$ can be replaced by $2\sqrt{d-1} + \delta$ for
some $\delta>0$ if graphs have bounded universal girth (see Section \ref{sect:6}
for the definition). On the other side of the spectrum, if the \DEF{odd girth}
(i.e.\ the length of a shortest odd cycle) is large, then we obtain
an Alon-Boppana-Serre type theorem for the negative eigenvalues.
If $G$ is a (finite) graph, we denote by $\l_i=\l_i(G)$ the $i$th largest eigenvalue
of the adjacency matrix $A(G)$ of $G$, respecting multiplicities.
The largest eigenvalue of $G$, $\rho(G)=\l_1(G)$, is also referred to
as the \DEF{spectral radius} of $G$. It follows from the Perron-Frobenius
theorem (see, e.g.~\cite{HJ}) that $\rho(G)$ is an eigenvalue of $G$ that
has an eigenvector $x$ whose coordinates are all non-negative. Moreover,
if $G$ is connected, then $x$ is strictly positive.
If $r\ge 1$ is an integer, a set $S$ of vertices of a graph $G$ is said to be
\DEF{$r$-apart} if any two vertices in $S$ are at distance at least $r+1$ in $G$.
We denote by $\alpha_r(G)$ the maximum cardinality of a vertex set in $G$ that
is $r$-apart. Note that $\alpha_1(G)$ is the usual independence number of the
graph.
Let $G$ be a graph, $v\in V(G)$, and let $r$ be an integer.
We denote by $G_r(v)$ the induced subgraph of $G$ on vertices that are
at distance at most $r$ from $v$. The subgraph $G_r(v)$ is called the
\DEF{$r$-ball} around $v$ in $G$.
We allow infinite graphs, but they will always be locally finite.
In particular, the $r$-ball around any vertex of a graph $G$ is always
finite.
\section{Universal covers and subcovers}
\label{sect:2}
Let $D=[d_{ij}]_{i,j=1}^t$ be a square matrix of order $t\ge1$, whose entries
$d_{ij}$ are non-negative integers. For $i=1,\dots,t$, we define the
\DEF{$i^{th}$ degree} in $D$ as the integer $d_i = \sum_{j=1}^t d_{ij}$.
Suppose that $D$ further satisfies the following conditions:
\begin{itemize}
\item[\bf (D1)] If $d_{ij}=0$, then also $d_{ji}=0$.
\item[\bf (D2)] The graph of $D$ is connected, i.e., for every
$i,k\in \{1,\dots,t\}$ there are integers $m_1,m_2,\dots,m_s$ in $\{1,\dots,t\}$,
where $m_1=i$, $m_s=k$, and $d_{m_j m_{j+1}} > 0$ for $j=1,\dots,s-1$.
\item[\bf (D3)] For every sequence of (distinct) integers $m_1,m_2,\dots,m_s$ in
$\{1,\dots,t\}$, we have
$$
d_{m_1 m_2}d_{m_2 m_3}\cdots d_{m_{s-1}m_s}d_{m_s m_1} =
d_{m_1 m_s}d_{m_s m_{s-1}}\cdots d_{m_3 m_2}d_{m_2 m_1}.
$$
\end{itemize}
Such a matrix is called a \DEF{$t$-partite degree matrix}.
Let $D=[d_{ij}]$ be a $t$-partite degree matrix. If a graph $G$ admits a
partition of its vertex set into $t$ classes, $V(G)=U_1\cup\cdots\cup U_t$,
such that every vertex in $U_i$ has precisely $d_{ij}$ neighbors in $U_j$,
for all $i,j=1,\dots,t$, then we say that $D$ is a \DEF{$t$-partite degree matrix}
for $G$. The corresponding partition $U_1\cup\cdots\cup U_t$ is said to be an
\DEF{equitable partition} for $D$; see, e.g.~\cite{GR}.
\begin{lemma}
\label{lem:universal cover}
Let\/ $D$ be a $t$-partite degree matrix.
{\rm (a)} There exists a finite graph $G$ whose degree matrix is $D$.
{\rm (b)} There exists a tree $T_D$ whose degree matrix is $D$.
The tree $T_D$ is determined up to isomorphism.
\end{lemma}
\begin{proof}
(a) First we remark that the condition (D3) implies that the set of equalities
$n_i d_{ij} = n_j d_{ji}$, $i,j\in \{1,\dots,t\}$, has a positive solution
$n_1,\dots,n_t$. Since all $d_{ij}$ are integers, there is a solution whose
values $n_i$ ($i=1,\dots,t$) are positive integers. To obtain a graph $G$,
we take vertex sets $U_i$ of cardinalities $n_i$ for $i=1,\dots,t$, and
join $U_i$ and $U_j$ so that the edges between them form
a $(d_{ij},d_{ji})$-biregular bipartite graph. Then it is clear that
$D$ is a $t$-partite degree matrix for $G$.
(b) To get $T_D$, we just take what is known as the universal cover of
the graph $G$ obtained in part (a).
We add a short proof of existence of $T_D$ that does not use
the property (D3) which is needed in (a).
Let us first assume that $d_i\ge2$ for $i=1,\dots,t$. This case is intuitively
clear and we leave the details of the proof for the reader. Note that $T_D$
is always infinite in this case.
The rest of the proof is by induction on $t$. We may assume that $d_t\le1$.
If $t=1$, then $T_D$ is either a single vertex (if $d_1=0$) or an edge
(if $d_1=1$). If $t>1$, then (D2) implies that $d_t=1$ and the non-zero element
in row $t$ of $D$ is not $d_{tt}$. Thus, there is a unique $j<t$ such that
$d_{tj}=1$. Let $p=d_{jt}$. By (D1), we conclude that $p>0$ and all other
elements in the column $t$ of $D$ are zero. Let $D'$ be the submatrix of $D$
obtained by deleting the last row and the last column. Since this operation
acts like removing a vertex of degree 1 from a graph, $D'$ still satisfies
(D1)--(D2), and hence we can apply the induction hypothesis to find the tree
$T_{D'}$. Finally, we obtain $T_D$ by adding, to each vertex in $V_j$,
$p$ pendant edges. All new vertices are of degree 1 and form the class $V_t$
in $T_D$.
\end{proof}
The tree $T_D$ is called the \DEF{universal cover} of the multipartite
degree matrix $D$. Let $V_1\cup\cdots\cup V_t$ be the corresponding equitable
partition of $V(T_D)$. If $G$ is any graph whose $t$-partite degree matrix is
$D$, there is a covering projection $\pi_G : T_D\to G$ which maps vertices in
$V_i$ onto the $i$th class of the equitable partition of $V(G)$.
Covering projections, universal covers and equitable partitions are regularly
used in algebraic graph theory. In the sequel we shall introduce a weaker notion,
distinguished by the prefix `sub', in which only those properties that are
important for our main results will be preserved.
A degree matrix $D$ is said to be a \DEF{$t$-partite subdegree matrix} for
a graph $G$ if there is a graph homomorphism $\pi^D_G: T_D\to G$ which is
\DEF{locally $1$-$1$}, i.e., for each vertex $v\in V(T_D)$,
$\pi^D_G$ maps edges incident with $v$ injectively to the edges incident
with $\pi^D_G(v)$. The homomorphism $\pi^D_G: T_D\to G$ is called
a \DEF{subuniversal projection} and the tree $T_D$ is a \DEF{subuniversal cover}
of~$G$.
If $\pi^D_G: T_D\to G$ is a subuniversal projection, let
$U_i = \pi^D_G(V_i)\subseteq V(G)$, $i=1,\dots,t$. Then it is easy to see
that the (not necessarily disjoint) vertex-sets $U_1,\dots,U_t$ satisfy
the following condition: Every vertex in $U_i$ has at least $d_{ij}$ neighbors
in $U_j$, for all $i,j=1,\dots,t$.
This gives a necessary condition for existence of a subuniversal projection.
Unfortunately, this condition is not sufficient. But if we ask that
every vertex in $U_i$ has at least $d_{ij}+1$ neighbors in $U_j$, for all
$i,j=1,\dots,t$, then the existence of a subuniversal projection to
$G$ is easily verified.
\begin{theorem}
\label{thm:ball and tree}
Suppose that\/ $D$ is a subdegree matrix for a (possibly infinite) graph $G$,
and let\/ $T_D$ be the corresponding subuniversal cover.
Let $v\in V(G)$ and let $s\in V(T_D)$ be a vertex that is mapped to $v$ via
a subuniversal projection $\pi^D_G$. Then for every $r\ge0$, the spectral radius
of the $r$-ball in $G$ is at least as large as the spectral radius of
the corresponding $r$-ball in $T_D$, $\rho(G_r(v))\ge\rho(T_{D,r}(s))$.
\end{theorem}
\begin{proof}
The spectral radius of a connected graph $H$ can be expressed as
\begin{equation}
\rho(H) = \limsup_{q\to \infty} (w_{2q}(H,u))^{1/(2q)},
\label{eq:walks1}
\end{equation}
where $w_{2q}(H,u)$ denotes the number of closed walks of length $2q$ in $H$
starting at the vertex $u\in V(H)$.
Every closed walk in $T_{D,r}(s)$
starting at $s$ is projected by $\pi^D_G$ to a closed walk in $G_r(v)$
starting at $v$. The projection of these walks is 1-1, since $\pi^D_G$ is
locally 1-1. Hence,
\begin{equation}
w_{2q}(G_r(v),v) \ge w_{2q}(T_{D,r}(s),s).
\label{eq:walks2}
\end{equation}
This inequality in combination with (\ref{eq:walks1}) implies that
$\rho(G_r(v))\ge\rho(T_{D,r}(s))$.
\end{proof}
\section{The spectral radius of infinite trees}
If $G$ is a connected infinite (locally finite) graph, we define its
spectral radius $\rho(G)$ as
\begin{equation}
\rho(G) = lim_{r\to\infty} \rho(G_r(v))
\label{eq:rho}
\end{equation}
where $v$ is any vertex of $G$. It is easy to see that the limit
exists (it may be infinite if the degrees of $G$ have no finite upper bound)
and that it is independent of the choice of $v$. The spectral radius of
infinite graphs defined above coincides with the notion obtained through the
spectral theory of linear operators in Hilbert spaces; we refer to
\cite{MW} for an overview.
The monotonicity property of the spectral radius of finite graphs implies that
for every connected finite graph $H$ and any proper subgraph $H'$ of $H$,
we have $\rho(H') < \rho(H)$.
Since $G$ is connected, infinite, and locally finite,
$G_r(v) \ne G_{r+1}(v) \ne G$ for every $r\ge0$.
This implies that
$$
\rho(G_r(v)) < \rho(G_{r+1}(v)) < \rho(G).
$$
Let us remark that the spectral radius of an infinite $d$-regular
tree is equal to $2\sqrt{d-1}$, the value that appears in the definition of
Ramanujan graphs. This was proved by Kesten \cite{Ke}, see also
Dynkin and Malyutov \cite{DM}, Cartier \cite{Ca}, and Woess \cite{Wo}.
We will use the spectral radius of universal cover trees introduced
in the previous section to replace the Ramanujan bound $2\sqrt{d-1}$
with the corresponding bound suitable for our multipartite generalization.
In the special case when the graph is the infinite $d$-regular tree,
which shall be denoted by $\TT_d$, it is easy to determine the precise rate
of convergence in (\ref{eq:rho}).
\begin{theorem}
\label{thm:convergence}
For every integer $d\ge 2$, we have
$$\rho(\TT_{d,r}) > 2\sqrt{d-1}\,\Bigl(1 - \frac{\pi^2}{r^2} + O(r^{-3})\Bigr).$$
\end{theorem}
\begin{proof}
Let $w_q(G,v_0)$ denote the number of closed walks of length $q$. It will be convenient to consider the subtree $\TT'_d$ of $\TT_d$ which is equal to the connected component containing the vertex $v_0$ of the subgraph obtained after deleting an edge of $\TT_d$ incident with $v_0$. The vertex $v_0$ has degree $d-1$ in $\TT'_d$, while all other vertices still have degree $d$.
The tree $\TT'_d$ has a natural projection onto the one-way-infinite path $P_\infty$ (whose vertices we denote by the non-negative integers $0, 1, 2, \dots$) such that all vertices at distance $i$ from $v_0$ are mapped onto the vertex $i$ in $P_\infty$. Every closed walk (based at $v_0$) of length $2q$ in $\TT'_d$ is projected onto a closed walk in $P_\infty$ based at the vertex $0$. Moreover, the $r$-ball $\TT'_{d,r}(v_0)$
in $\TT'_d$ is projected onto the path $P_{r+1} \subset P_\infty$ on vertices
$0,1,\dots,r$.
Whenever we walk away from $v_0$ in $\TT'_d$, we have $d-1$ choices to do so. This implies that
\begin{equation}
w_{2q}(\TT'_{d,r},v_0) = (d-1)^q \, w_{2q}(P_{r+1},0).
\label{eq:walks T_d,r}
\end{equation}
When $q\to \infty$, the quantities raised to the power $1/(2q)$ tend to the spectral radii of the corresponding graphs, and we conclude that
$\rho(\TT'_{d,r})=\sqrt{d-1}\,\rho(P_{r+1}) = 2\sqrt{d-1}\,\cos(\frac{\pi}{r+2}) = 2\sqrt{d-1}\,\bigl(1 - \tfrac{\pi^2}{r^2} + O(r^{-3})\bigr)$.
Since $\TT'_{d,r}$ is a proper finite subgraph of $\TT_{d,r}$, this implies the (strict) inequality of the theorem.
\end{proof}
The rate of convergence is likely the same for more general universal covers of finite graphs. We propose the following conjecture.
\begin{conjecture}
\label{conj:convergence}
For every multipartite degree matrix $D$, there exists a constant $c=c(D)$
such that for every $s\in V(T_D)$, we have
$$\rho(T_{D,r}(s)) \ge \rho(T_D) - cr^{-2}.$$
\end{conjecture}
\section{Multipartite Ramanujan graphs}
In this section we introduce a generalized notion of Ramanujan graphs.
The following lemma shows that we cannot simply compare $\l_2(G)$ with
$\rho(T_D)$ as is the case for $d$-regular graphs.
\begin{lemma}
\label{lem:contains eigenvalues}
If\/ $D$ is a $t$-partite degree matrix, then all eigenvalues of\/ $D$ are
real and their algebraic multiplicity is equal to their geometric multiplicity.
If\/ $D$ is a multipartite degree matrix of a finite graph\/ $G$, then every
eigenvalue of $D$ is also an eigenvalue of\/ $G$. Moreover, $\rho(G)=\rho(D)$.
\end{lemma}
\begin{proof}
Let $n_1,\dots,n_t$ be a positive solution of the system
$n_i d_{ij} = n_j d_{ji}$, $i,j\in \{1,\dots,t\}$, which was shown to exist in
the proof of Lemma \ref{lem:universal cover}(a). If $R$ is the diagonal matrix
of order $t$ whose entry $R_{ii}$ is equal to $n_i^{1/2}$ ($i=1,\dots,t$),
then $RDR^{-1}$ is a symmetric matrix. This implies the first part of the lemma.
To verify the second part, let $\l$ be an eigenvalue of $D$, and
let $y=(y_i\mid i=1,\dots,t)$ be an eigenvector
for $\l$. Let $V_1\cup\cdots\cup V_t$ be the partition of $V(G)$ corresponding
to the degree matrix $D$. If we set $x_v = y_i$ for every $v\in V_i$, then it is
easy to see that $x=(x_v\mid v\in V(G))$ is an eigenvector of the adjacency
matrix of $G$ for the eigenvalue $\l$.
To prove the last claim, observe that the eigenvalue $\rho(D)$ has a positive eigenvector by the Perron-Frobenius Theorem. Its lift in $G$ is a positive
eigenvector of $G$ for the eigenvalue $\rho(D)$. Again, by applying
the Perron-Frobenius Theorem, we conclude that this eigenvector corresponds to the
largest eigenvalue of $G$.
\end{proof}
Let $D$ be a $t$-partite degree matrix. Let $k$ be the largest integer such
that $\l_k(D) \ge \rho(T_D)$. Note that $k$ exists since $\l_1(D) \ge \rho(T_D)$.
We say that a finite graph $G$ with degree matrix (resp.~subdegree matrix) $D$
is \DEF{$D$-Ramanujan} (resp.~\DEF{$D^+$-Ramanujan}) if
$\lambda_{k+1}(G)\le \rho(T_D)$. We believe that there is an abundance of
generalized Ramanujan graphs and propose the following conjectures
(in which we assume that the minimum degree of $D$ is at least 2).
\begin{conjecture}
\label{conj:exist D-Ramanujan2}
If there exists a $D$-Ramanujan graph for a multipartite degree matrix $D$,
then there exist infinitely many $D$-Ramanujan graphs.
\end{conjecture}
\begin{conjecture}
\label{conj:exist D-Ramanujan+}
If there exists a $D^+$-Ramanujan graph for a multipartite degree matrix $D$,
then there exist infinitely many $D^+$-Ramanujan graphs.
\end{conjecture}
\begin{conjecture}
\label{conj:exist D-Ramanujan}
If\/ $D$ is a degree matrix of order $t\ge2$, and $\l_2(D) < \rho(T_D)$,
then there exist infinitely many $D$-Ramanujan graphs.
\end{conjecture}
One cannot exclude the possibility that there exist infinitely many $D$-Ramanujan graphs for every degree matrix $D$, but our knowledge is too limited at this point to propose this as a conjecture.
\section{A generalized Alon-Boppana-Serre theorem}
\begin{theorem}
\label{thm:main}
Let\/ $D$ be a multipartite degree matrix, and let\/ $\rho_D = \rho(T_D)$.
{\rm (a)}
For every $\varepsilon>0$, there exists an integer $r=r(D,\varepsilon)$ such that
for every integer $k\ge1$ and for every graph $G$, if $D$ is a subdegree matrix
of $G$ and $\alpha_{2r+1}(G)\ge k$, then
$\l_k(G) \ge \rho_D - \varepsilon$.
{\rm (b)}
For $\varepsilon>0$ and every positive integer $\Delta$, there exists a constant
$c=c(D,\Delta,\varepsilon)>0$ such that every graph $G$ of order $n$, of maximum
degree at most $\Delta$ and with subdegree matrix $D$, has
at least $cn$ eigenvalues that are larger than $\rho_D-\varepsilon$.
\end{theorem}
\begin{proof}
(a) Let $r=r(D,\varepsilon)$ be the smallest integer such that
$\rho(T_{D,r})\ge \rho_D - \varepsilon$, and let $k$ and $G$ be as specified.
Since $\alpha_{2r+1}(G) \ge k$, there are vertices $v_1,\dots,v_k$ that are
$(2r+2)$-apart. The $r$-balls $G_r(v_1),\dots,G_r(v_k)$ around
these vertices are not only pairwise disjoint, but also form an induced subgraph
of $G$. By the eigenvalue interlacing property for induced subgraphs, we know that
$$
\l_k(G) \ge \l_k(G_r(v_1)\cup\cdots\cup G_r(v_k)) \ge
\min \{ \rho(G_r(v_i))\mid 1\le i\le k \}.
$$
By Theorem \ref{thm:ball and tree} and by our choice of $r$, we have
$$
\rho(G_r(v_i)) \ge \rho(T_{D,r}(s_i)) \ge \rho_D - \varepsilon.
$$
This completes the proof of (a).
(b) This part follows from (a). It is just to be noted that
any $(2r+1)$-ball in $G$ contains at most
$B = \tfrac{\Delta}{\Delta-2}(\Delta-1)^{2r+1}$ vertices.
Thus, $\alpha_{2r+1}(G)\ge n/B$, and hence part (a) applies
with $c=B^{-1}$.
\end{proof}
It is worth mentioning that the condition involving $\alpha_{2r+1}(G)$ in
Theorem \ref{thm:main}(a) is necessary if we only assume that $D$ is a
subdegree matrix. Simple examples showing this are provided by the
family of all complete graphs $K_n$ whose second largest eigenvalue
is always equal to $-1$, or by the family of all complete bipartite
graphs $K_{m,n}$ whose second eigenvalue is 0.
For the special case when $D=[d]$, Theorem \ref{thm:convergence} gives the precise
description for the values $r(d,\varepsilon)$ and $c(d,\Delta,\varepsilon)$
in Theorem \ref{thm:main}. By Theorem~\ref{thm:convergence},
\begin{equation}
r = r(d,\varepsilon) = \pi\biggl(\frac{2\sqrt{d-1}}{\varepsilon}\, \biggr)^{1/2}\,
\bigl(1 + O\bigl(d^{-1/4}\varepsilon^{1/2}\bigr) \bigr)
\nonumbe
\end{equation}
and
\begin{equation}
c(d,\Delta,\varepsilon) = \tfrac{\Delta}{\Delta-2}(\Delta-1)^{-(2r+1)}
\nonumbe
\end{equation}
will do the job.
As an example, let us consider the following special case.
The bipartite degree matrix
\begin{equation}
D = \left[\begin{matrix}
0 & d_1 \\
d_2 & 0
\end{matrix}\right]
\label{eq:D}
\end{equation}
involves, in particular, all bipartite graphs with bipartition
$V= A\cup B$, whose degrees in $A$ are at least $d_1$ and whose degrees in
$B$ are at least $d_2$. The spectral radius of $T_D$ is (cf.~\cite{MW})
$$
\rho(T_D) = \sqrt{d_1-1} + \sqrt{d_2-1}\,.
$$
Thus, only finitely many bipartite $(d_1,d_2)$-biregular graphs have their
$k$th eigenvalue ($k\ge2$) smaller than $\rho(T_D) - \varepsilon$.
Theorem \ref{thm:main} suggests the following strengthening, which we will prove
directly by using Theorem \ref{thm:ball and tree}.
\begin{corollary}
\label{cor:2partite}
Let $d_1\le d_2\le d$ be positive integers, and let ${\cal G}_{d_1,d_2}^d$ be
the set of all graphs whose maximum vertex degree is at most $d$ and whose
vertex set is the union of (not necessarily disjoint) subsets
$U_1,U_2$, such that every vertex in $U_i$ has at least $d_i$
neighbors in $U_{3-i}$ for $i=1,2$.
For every $\varepsilon>0$, every $n$-vertex graph $G\in {\cal G}_{d_1,d_2}^d$
has\/ $\Omega_\varepsilon(n)$ eigenvalues larger than
$\sqrt{d_1-1} + \sqrt{d_2-1} - \varepsilon$.
\end{corollary}
\begin{proof}
We claim that there exists a subuniversal projection $\pi^D_{G}$, where $D$ is
the degree matrix given in (\ref{eq:D}). The tree $T_D$ is $(d_1,d_2)$-biregular.
We map a vertex $v_0$ of degree $d_1$ in $T_D$ onto the vertex $u\in U_1$.
After fixing $v_0$, we extend the mapping to a locally 1-1
homomorphism in a greedy fashion (by taking the breadth-first search order
of vertices of $T_D$ starting at $v_0$) so that vertices of degree $d_i$ are mapped
to $U_i$, $i=1,2$.
Let $v_1$ be a neighbor of $v_0$ in $T_D$ and let $v = \pi_G^D(v_1)\in U_2$.
Theorem \ref{thm:ball and tree} shows that
for large enough $r=r(d_1,d_2,\varepsilon)$,
\begin{eqnarray}
\rho(G_r(u)) &\ge& \rho(T_{D,r}(v_0))\ \ge\ \rho(T_D)-\varepsilon,
\label{eq:reachable}\\
\rho(G_r(v)) &\ge& \rho(T_{D,r}(v_1))\ \ge\ \rho(T_D)-\varepsilon.
\label{eq:reachable2}
\end{eqnarray}
Since the maximum degree of $G$ is bounded by $d$, the $(2r+1)$-balls in $G$ have
bounded number of vertices, say at most $B$. Therefore,
$\alpha_{2r+1}(G)\ge n/B$, and so there are at least this many pairwise
non-adjacent induced $r$-balls around vertices in $G$. As before, the eigenvalue
interlacing theorem and (\ref{eq:reachable})--(\ref{eq:reachable2})
imply that linearly many eigenvalues
of $G$ are larger than
\medskip
\hskip 2.3cm
$\rho(T_D)-\varepsilon = \sqrt{d_1-1} + \sqrt{d_2-1} - \varepsilon$.
\end{proof}
\section{Global girth and Ramanujan graphs}
\label{sect:6}
All known Ramanujan graphs are Cayley graphs and their girth increases with their
order. We shall use the method of this paper to explain why the girth cannot be
bounded. Actually, we shall prove that a small girth condition implies that
$d$-regular graphs are ``far from being Ramanujan;''
see Theorem \ref{thm:girth} below.
Let $G$ be a graph. A closed walk $v_1v_2\dots v_kv_1$ is
\DEF{retracting-free} if $v_{i-1}\ne v_{i+1}$ for $i=1,\dots,k$
(where $v_0=v_k$ and $v_{k+1}=v_1$).
It is easy to see that if $G$ is a finite graph with minimum degree at least 2, then
for every vertex $v$ of $G$ there exists a retracting-free closed walk through $v$.
Let $g(v)$ be the length of a shortest retracting-free closed walk through $v$.
The \DEF{universal girth} of $G$, denoted by $m(G)$, is the smallest integer
$k$ such that every vertex in $G$ has a
retracting-free closed walk of length $k$. Let us observe that $m(G)$ is at most
the least common multiple of the values $g(v)$, $v\in V(G)$.
Also, if $G$ is vertex-transitive, then $m(G)$ is equal to the girth of~$G$.
Let $\XX_{d,g}$ be the graph obtained from the $(d-2)g$-regular tree $T$ by
expanding each vertex $v\in V(T)$ into the cycle $C_v$ of length $g$, such that each
vertex of $C_v$ is incident with $d-2$ of the edges of $T$ incident with $v$.
See Figure \ref{fig:1} showing the case of $d=4$ and $g=4$.
The graph $\XX_{d,g}$ is the Cayley graph of the free product of $d-2$ copies
of $\ZZ_2$ and one copy of $\ZZ_g$ (with the natural generating set).
\begin{figure}[htb]
\epsfxsize=6.4truecm
\slika{AlonBoppana1.eps}
\caption{Expanding a vertex into a cycle}
\label{fig:1}
\end{figure}
Paschke \cite{Pa} determined the spectral radius of $\XX_{d,g}$:
\begin{theorem}[Paschke]
\label{thm:Paschke}
For $d\ge3$ and\/ $g\ge3$, the graph\/ $\XX_{d,g}$ has spectral radius
$$
\min_{s>0} \,(d-2)\, \phi\biggl(\frac{1+\cosh sg}{\sinh sg\sinh s}\biggr)
+ 2\cosh s > 2\sqrt{d-1},
$$
where $\phi(t) = \frac{\sqrt{1+t^2} - 1}{t}$.
\end{theorem}
Paschke~\cite{Pa} used this result to provide a non-trivial lower bound on
the spectral radius of infinite vertex-transitive graphs of the
given girth~$g$. He showed that a vertex transitive $d$-regular graph
containing a $g$-cycle has spectral radius at least $\rho(\XX_{d,g}$).
The formula in Theorem \ref{thm:Paschke} gives a lower bound of the form
$$
2\sqrt{d-1} + \frac{2(d-2)}{(d-1)^{(g+1)/2}}\, h(d,g),
$$
where $h$ is a function such that such that for every $g\ge 3$,
$\lim_{d\to\infty} h(d,g)=1$, and for every $d\ge3$,
$\lim_{g\to\infty} h(d,g)=1$.
Now, we strengthen the Alon-Boppana-Serre theorem by showing that the lower bound
$2\sqrt{d-1}-\varepsilon$ can be replaced by
$2\sqrt{d-1} + \delta$ for some $\delta>0$ if graphs have bounded universal girth.
\begin{theorem}
\label{thm:girth}
For every $\Delta\ge d\ge3$ and every $g\ge3$, there exist $\delta>0$ and $c>0$
such that every $n$-vertex graph $G$ with minimum degree at least $d$, maximum
degree at most $\Delta$ and
universal girth at most $g$ has at least $\lceil cn\rceil$ eigenvalues that
are larger than $2\sqrt{d-1} + \delta$.
\end{theorem}
\begin{proof}
The proof follows the same pattern as the proof of Theorem \ref{thm:main}, except that we use the graph $\XX_{d,m}$, where $m=m(G)\le g$ is the universal girth of $G$,
playing the role of the universal cover $\TT_d$.
Here, we have to take $r$ large enough so that
$\rho((\XX_{d,m})_r) \ge 2\sqrt{d-1} + \delta$.
Such an $r$ exists because of (\ref{eq:rho}) and since
$\rho(\XX_{d,m}) > \rho(\TT_d) = 2\sqrt{d-1}$.
\end{proof}
It is straightforward to generalize the proof of Theorem \ref{thm:girth} to
the setting of degree matrices. What we need is just an analogue of the Paschke
theorem. However, we do not intend to dig into the details in this note.
\section{The other side of the spectrum}
As shown in the previous section, small universal girth yields improved
lower bounds on large eigenvalues, so Ramanujan graphs must have growing
girth. On the other hand, large girth has some further cosequences.
In particular, it shows that the negative eigenvalues satisfy the
Alon-Boppana-Serre property as well.
Let us first formulate the monopartite version for the negative eigenvalues.
It involves the notion of the \DEF{odd girth} of the graph, meaning the length
of a shortest cycle of odd length in the graph. (If $G$ is bipartite, then the
odd girth is $\infty$.) This result was obtained earlier by Friedman \cite{Fr}
and Nilli \cite{Ni2}; it also appears in Ciaob\v a \cite{Ci} (with a slightly weaker estimate of $\varepsilon$).
\begin{theorem}
\label{thm:ABTnegative1}
For every $\Delta\ge d\ge 2$ and $g\ge 3$, there exists a positive constant
$c=c(d,\Delta,g)>0$ such that every graph $G$ of order $n$, of minimum degree $d$,
maximum degree at most $\Delta$, and with odd girth at least $g$ has
at least $cn$ eigenvalues that are larger than
$2\sqrt{d-1}\bigl(1-\varepsilon\bigr)$
and has at least\/ $cn$ eigenvalues that are smaller than
$-2\sqrt{d-1}\bigl(1-\varepsilon\bigr)$,
where $\varepsilon = \bigl(\frac{2\pi}{g}\bigr)^2 + O(g^{-3})$.
\end{theorem}
\begin{proof}
(Sketch) The proof is essentially the same as the proof of Theorem \ref{thm:main}(b),
where we take $r=\lfloor\tfrac{1}{2}g\rfloor-1$ and apply the estimate of
Theorem \ref{thm:convergence}. The assumption that the odd girth is more than $2r+1$
shows that the $r$-balls in $G$ contain no cycles of odd length. In particular, they are bipartite and hence their spectrum is symmetric with respect to 0.
Thus, knowing that the spectral radius $\l$ is large, we conclude
that the smallest eigenvalue $-\l$ is large in absolute value.
Now, we can use the interlacing theorem for the smallest eigenvalues of $G$
compared to the eigenvalues of the induced subgraph of $G$ consisting of
disjoint $r$-balls around $\lceil cn \rceil$ vertices that are $(2r+2)$-apart.
\end{proof}
The generalized version of Theorem \ref{thm:ABTnegative1} holds as well.
The proof is the same, except that we do not provide
an explicit estimate on $\varepsilon$ in terms of the odd girth.
\begin{theorem}
\label{thm:ABTnegative2}
Let\/ $D$ be a multipartite degree matrix, and let\/ $\rho_D = \rho(T_D)$.
For every $\varepsilon>0$ and every positive integer $\Delta$, there exists an integer
$g=g(D,\varepsilon)$ and a positive constant
$c=c(D,\Delta,\varepsilon)>0$ such that every graph $G$ of order $n$, of maximum
degree at most $\Delta$, with subdegree matrix $D$, and with odd girth at least $g$ has
at least $cn$ eigenvalues that are smaller than $-\rho_D+\varepsilon$.
\end{theorem}
|
\section{Proofs of Lemmas}
\subsection{Proof of Lemma~\ref{lmm:prune-range}}
Consider the example of Figure~\ref{fig_proof}. To prove the lemma by contradiction, assume the
following equation holds:
\[
D(o_q(t),o_i(t)) \le e \label{eqn:1}
\]
Since $l'_q$ is a line segment (with actual tolerance
$\delta(l'_q)$) in the simplified trajectory $o'_q$, there exists a
location $a_q$ on $l'_q$ such that $D(a_q,o_q(t)) \le \delta(l'_q)$.
Similarly, there exists a location $a_i$ on $l'_i$ such that
$D(a_i,o_i(t)) \le \delta(l'_i)$. Due to the triangular inequality,
\[
D(a_q,a_i) \le D(a_q,o_q(t)) + D(o_q(t),o_i(t)) + D(o_i(t),a_i) \label{eqn:2}
\]
Combining the inequalities, we obtain:
\begin{eqnarray}
D(a_q,a_i) \le \delta(l'_q) + e + \delta(l'_i) \label{eqn:3}
\end{eqnarray}
On the other hand, $a_q$ ($a_i$) is a location on line segment
$l'_q$ ($l'_i$). Hence, equation (\ref{eqn:4}) holds
\begin{eqnarray}
D_{LL}(l'_q,l'_i) \le D(a_q,a_i) \label{eqn:4}
\end{eqnarray}
From the last two inequalities (\ref{eqn:3}) and (\ref{eqn:4}), we get:
\begin{eqnarray}
D_{LL}(l'_q,l'_i) \le e + \delta(l'_q) + \delta(l'_i) \label{eqn:5}
\end{eqnarray}
Therefore, the resulting contradiction of (\ref{eqn:5}) proves Lemma~\ref{lmm:prune-range}.
\subsection{Proof of Lemma~\ref{lmm:prune-group-range}}
Note that for all $l'_i \in S$, we have $\delta_{max}(S) \ge \delta(l'_i)$ and \\*
$D_{min}(\mathcal{B}(l'_q),\mathcal{B}(S)) \le D_{LL}(l'_q,l'_i)$. If the following equation
satisfies:
\[
D_{min}(\mathcal{B}(l'_q),\mathcal{B}(S)) > e + \delta(l'_q) + \delta_{max}(S)
\]
then, the next equation must also hold:
\[
D_{LL}(l'_q,l'_i) > e + \delta(l'_q) + \delta(l'_i) \label{eqn:8}
\]
The rest of this proof follows directly from Lemma~\ref{lmm:prune-range}.
\subsection{Proof of Lemma~\ref{lmm:prune-star-range}}
Since $l'_q$ is a line segment (with actual tolerance $\delta(l'_q)$) in the simplified trajectory
$o'_q$ , the location $l'_q(t)$ meets:
\[
D(l'_q(t),o_q(t)) \le \delta(l'_q) \label{eqn:9}
\]
Similarly, the location $l'_i(t)$ satisfies:
\[
D(l'_i(t),o_i(t)) \le \delta(l'_i) \label{eqn:10}
\]
In addition, we have:
\[
D_{*}(l'_q,l'_i) \le D(l'_q(t),l'_i(t)) \label{eqn:11}
\]
The logic of the remainder of the proof is the same as in the proof of
Lemma~\ref{lmm:prune-range}.
\pagebreak
\section{Additional Experiments}
\subsection{MC vs. CMC}\label{app:mc2}
In this experiment, we intend to demonstrate empirically that methods for the discovery of moving
clusters cannot be used to compute convoys directly (see Section~\ref{sec:trjclust}). Specifically,
we study the discovery accuracies of convoys by a solution for moving cluster (MC2). MC2 reports
results of the convoy query if the portion of common objects in any two consecutive clusters $c_1$
and $c_2$ is not below a given threshold parameter $\theta$, i.e., $\frac{ |c_1 \cap c_2| }{ |c_1
\cup c_2| } \ge \theta$.
Let $R_m$ be a result set of convoys discovered by MC2 and $R_c$ be another set obtained by CMC (or
CuTS). We measure the proportions of false positives in Figure \ref{exp_mc2}(\emph{a}) by verifying
whether each convoy $v \in R_m$ satisfies the query condition with respect to $m$, $k$, and $e$
using the results of CMC (i.e., $(\frac{|R_m - R_c|}{|R_m|})\times 100$). Likewise, false negatives
in Figure \ref{exp_mc2}(\emph{b}) are computed by $(\frac{|R_c - R_m|}{|R_c|})\times 100$).
\begin{figure}[hbt]
\center
\includegraphics[width=8.5cm]{pics/exp/mc2.eps}
\caption{Discovery Quality of the MC method for Convoys}
\label{exp_mc2}
\end{figure}
In fact, MC2 reported bigger numbers of convoys than what CMC does because MC2 does not have the
lifetime constraint $k$. This feature was especially obvious for the Cattle dataset that is larger
than the others. As a result, the proportions of actual convoys in the result set were very low,
and the numbers of false positives were very high in Figure \ref{exp_mc2}(\emph{a}). For the other
datasets, false positives went up as the $\theta$ value grew since the number of convoys reported
by MC2 also increased. Let $\theta_{c_1c_2}$ be a ratio of common objects between two snapshot
clusters $c_1$ and $c_2$. Assume that there are four consecutive snapshot clusters $c_1, c_2, c_3,$
and $c_4$, and $\theta_{c_1c_2}=1.0, \theta_{c_2c_3}=0.8, \theta_{c_3c_4}=1.0$. If we set the value
of $\theta$ to be equal to or smaller than 0.8, one moving cluster having all the snapshot clusters
will be reported (say $MC_{c_1c_2c_3c_4}$). In contrast, when $\theta > 0.8$, MC2 will discover two
moving clusters $MC_{c_1c_2}$ and $MC_{c_3c_4}$. Therefore, a higher $\theta$ value may produce a
larger number of moving clusters as convoy results.
Even though MC2 returns many convoys, the result set did not necessarily contain all actual
convoys. We investigate this aspect by computing false negatives in Figure \ref{exp_mc2}(\emph{b}).
In general, the number of false negatives increases as the $\theta$ value increases because the
number of convoys discovered by MC2 also increases.
Note that if many actual convoys exist for different parameter settings, the proportions of both
false positives and false negatives may increase considerably. Therefore, the use of moving cluster
methods for convoy discovery is ineffective and unreliable.
\section{Coherent Moving Cluster (CMC)} \label{sec:cmc}
A simple technique for computing a convoy is to first perform (density--connected) clustering on
the objects at each time and then to extract their common objects in an attempt to form convoys.
This approach is similar to the methods for discovering moving clusters \cite{st-clustering1}.
However, those are unable to discover the exact convoy results, as explained next:
\pagebreak
\begin{itemize}
\setlength{\parskip}{-2pt} \item Let $c_t$ and $c_{t+1}$ be (snapshot) clusters at times $t$ and
$t+1$. These clusters belong to the same \emph{moving cluster} if
they share at least the fraction $\theta$ objects ($|c_t \cap
c_{t+1}|/|c_t \cup c_{t+1}| \geq \theta$), where $\theta$ is a
user-specified threshold value between 0 and 1.
The problem of applying moving cluster methods for convoy discovery is that no absolute $\theta$ value exists that can be used to compute the exact
convoy results---either false hits may be found, or actual convoys
may remain undiscovered, as explained in Section~\ref{sec:trjclust}.
\setlength{\parskip}{-2pt} \item A moving cluster can be formed as long as two snapshot clusters
have at least $\theta$ overlap, even for only two consecutive
time. The lifetime ($k$) constraint does not apply to moving
clusters, but is essential for a convoy.
\setlength{\parskip}{-2pt} \item As pointed in the previous section, a trajectory may have some
missing time points due to irregular location sampling (e.g., $o_3$ at $t=2$ in Figure
~\ref{fig_cmc}(\emph{a})). In this case, we cannot measure the density--connection for all objects
involved over those missing times.
\end{itemize}
In order to solve the above problems for convoy discovery, we extend the moving cluster method into
our \emph{Coherent Moving Cluster} algorithm (CMC). First, we generate virtual locations for the
missing time points. If any trajectory has a location at time $t_i$, but another does not during
its time interval, we apply linear interpolation to create the virtual points at $t_i$. Second, to
accommodate the lifetime ($k$) constraint, we require each candidate convoy to have (at least) $k$
clusters $c_t, c_{t+1}, \cdots, c_{t+k-1}$ during consecutive time points. Third, we test the
condition $|c_t \cap c_{t+1} \cap \cdots \cap c_{t+k-1}| \geq m$, to determine whether sufficiently
many common objects are shared. If all conditions are satisfied, the candidate is reported as an
actual convoy.
We proceed to illustrate algorithm CMC using Figure~\ref{fig_cmc},
with the parameters $m=2$ and $k=3$. Let $c_t^i$ be the $i$-th
snapshot cluster at time $t$. Clusters at time $t$ are obtained by
applying a snapshot density clustering algorithm (e.g., DBSCAN
\cite{s-clustering3}) on the objects' locations at time $t$.
\begin{figure}[hbt]
\centering
\includegraphics[width=8.2cm]{pics/fig_cmc.eps}
\caption{Query Processing of CMC, $m=2$}
\label{fig_cmc}
\end{figure}
Table~\ref{tab:steps-cmc} illustrates the execution steps of the algorithm. At time $t_1$, we
obtain a cluster $c_1^1$ (with objects $o_1$, $o_2$, and $o_3$) and consider it a convoy candidate
$v_1$. At time $t_2$, we retrieve a cluster $c_2^1$, which is then compared with $v_1$. Since
$c_2^1$ and $v_1$ have $m=2$ common objects, we compute their intersection and update candidate
$v_1$. At time $t_3$, we discover two clusters $c_3^1$ and $c_3^2$. Since $c_3^1$ shares no
objects with $v_1$, we consider $c_3^1$ as another convoy candidate $v_2$. As $c_3^2$ shares $m=2$
common objects with $v_1$, we update $v_1$ to be its intersection with $c_3^2$. Eventually, $v_1$
is reported as a convoy because it contains $m=2$ common objects from clusters during $k=3$
consecutive time points.
\begin{table} [hbt]
\centering
\begin{tabular}{|c|c|c|}
\hline
\textbf{Timestamp} & \textbf{Clusters} & \textbf{Candidate set $V$} \\ \hline
$t_1$ & $c_1^1$ & $v_1=c_1^1$ \\ \hline
$t_2$ & $c_2^1$ & $v_1=c_1^1 \cap c_2^1$ \\ \hline
$t_3$ & $c_3^1$, $c_3^2$ & $v_1=c_1^1 \cap c_2^1 \cap c_3^2$, $v_2=c_3^1$ \\ \hline
\end{tabular}
\caption{Execution Steps of CMC}\label{tab:steps-cmc}
\end{table}
Algorithm~\ref{alg:cmc} presents the pseudocode for the CMC algorithm. The algorithm takes as
inputs a set of object trajectories $O$ and convoy query parameters $m$, $k$, and $e$.
We use $V$ to represent the set of convoy candidates. We then perform processing for each time
point (in ascending order).
The set $V_{next}$ introduced in Line~3 is used to store candidates produced at the current time
$t$. Then, we consider only objects $o \in O$ whose time intervals cover time $t$, i.e., $t \in
o.\tau$. Their locations $o(t)$ are inserted into the set $O_t$. If any object $o \in O_t$ has a
missing location at $t$, a virtual point is computed and then inserted.
Next, we apply DBSCAN on $O_t$ to obtain a set $C$ of clusters (Line~7).
The clusters in $C$ are compared to existing candidates in $V$. If they share at least $m$ common
objects (Line~11), the current objects of the candidate $v$ are replaced by the common objects
between $c$ and $v$ and are then inserted into the set $V_{next}$ (Lines~13--15). At the same time,
we increment the lifetime of the candidate (Lines~14). Each candidate with its lifetime (at least)
$k$ is reported as a convoy (Lines~17--18).
Clusters (in $C$) having insufficient intersections with existing candidates are inserted as new
candidates into $V_{next}$ (Lines~19--23). Then all candidates in $V_{next}$ are copied to $V$ so
that they are used for further processing in the next iteration.
\begin{algorithm}[!h]
\small \caption{\bf CMC (Set of object trajectories $O$, Integer $m$, Integer $k$, Distance
threshold $e$)}
\begin{algorithmic}[1]
\STATE $V \leftarrow \emptyset$
\FOR {each time $t$ (in ascending order)}
\STATE $V_{next} \leftarrow \emptyset$
\STATE $O_t \leftarrow \{ o(t) \; | \; o \in O \; \wedge \; t \in o.\tau \}$
\IF { $O_t$.size $< m $}
\STATE skip this iteration
\ENDIF
\STATE $C \leftarrow$ DBSCAN($O_t, e, m$)
\FOR {each convoy candidate $v \in V$}
\STATE $v.$assigned $\leftarrow$ false
\FOR {each snapshot cluster $c \in C$}
\IF { $|c \cap v| \geq m $}
\STATE $v$.assigned $\leftarrow$ true
\STATE $v \leftarrow c \cap v$
\STATE $v$.endTime $ \leftarrow t$
\STATE $V_{next} \leftarrow V_{next} \cup v$
\STATE $c$.assigned $\leftarrow$ true
\ENDIF
\ENDFOR
\IF {$v$.assigned = false and $v$.lifetime$\ge k$}
\STATE $V_{result} \leftarrow V_{result} \cup v$
\ENDIF
\ENDFOR
\FOR {each $c \in C$}
\IF {$c$.assigned = false}
\STATE $c$.startTime $\leftarrow t$
\STATE $c$.endTime $\leftarrow t$
\STATE $V_{next} \leftarrow V_{next} \cup c$
\ENDIF
\ENDFOR
\STATE $V \leftarrow V_{next}$;
\ENDFOR
\RETURN $V_{result}$
\end{algorithmic}
\label{alg:cmc}
\end{algorithm}
\section{Conclusion} \label{sec:conc}
Discovering convoys in trajectory data is a challenging problem, and existing solutions to related
problems are ineffective at finding convoys. This study formally defines a \emph{convoy query}
using density-based notions, and it proposes four algorithms for computing the convoy query. Our
main algorithms (CuTS, CuTS+, and CuTS*) use line simplification methods as the foundation for a
filtering step that effectively reduces the amounts of data that need further processing. In order
to ensure that the filters do not eliminate convoys, we bound the errors of the discovery
processing over the simplified trajectories. Through our experimental results with real datasets,
we found that CuTS* showes the best performance. CuTS+ also performes well when the given
trajectories have a small number of objects and long histories.
\small
\section*{Acknowledgment}
National ICT Australia is funded by the Australian Government's Backing Australia's Ability
initiative, in part through the Australian Research Council (ARC). This work is supported by grant
DP0663272 from ARC.
\bibliographystyle{abbrv}
\section{Convoy Discovery Using Trajectory Simplification (CuTS)} \label{sec:cuts}
The CMC algorithm incurs high computational cost because it generates virtual locations for all
missing time points and performs expensive clustering at every time. In this section, we apply the
filter-and-refinement paradigm with the purpose of reducing the overall computational cost. For the
filter step, we simplify the original trajectories and apply clustering on the simplified
trajectories to obtain convoy candidates. The goal is to retrieve a superset of the actual convoys
efficiently. In the refinement step, we consider each candidate convoy in turn. In particular, we
perform clustering on the original trajectories of the objects involved to determine whether the
convoy indeed qualifies. The resulting CuTS algorithm is guaranteed to return correct convoy
results.
\subsection{Simplifying Trajectories} \label{sec:simple}
Given a trajectory represented as a polyline $o= \langle p_1, p_2,\cdots, p_T \rangle$, and a
tolerance $\delta$, the goal of {\em trajectory simplification} is to derive another polyline $o'$
such that $o'$ has fewer points and deviates from $o$ by at most $\delta$. We say that $o'$ is a
{\em simplified trajectory} of $o$ with respect to $\delta$.
We apply the Douglas-Peucker algorithm (DP), as discussed in Section~\ref{sec:trjsimple}, to
simplify a trajectory.
Initially, DP composes the line $\overline {p_1p_T}$ and finds the point $p_i \in o$ farthest from
the line. If the distance $D_{PL}(p_i,\overline{p_1p_T}) \leq \delta$, segment $\overline {p_1p_T}$
is reported as the simplified trajectory $o'$. Otherwise, DP recursively processes the
sub-trajectories $\langle p_1, \cdots, p_i \rangle$ and $\langle p_i,\cdots, p_T \rangle$,
reporting the con\-ca\-tenation of their simplified trajectories as the simplified trajectory~$o'$.
\begin{figure}[hbt]
\centering
\includegraphics[width=7.5cm]{pics/fig_simple.eps}
\caption{Trajectory Simplification}
\label{fig_simple}
\end{figure}
Figure~\ref{fig_simple}(\emph{a}) illustrates the application of DP on the trajectories in
Figure~\ref{fig_convoy}. For $o_1$ trajectory, we first construct the virtual line between its end
points. Since the distance between the farthest point (i.e., $p_1$) and the virtual line exceeds
$\delta$, point $p_1$ will be kept in $o_1$'s corresponding simplified trajectory $o'_1$. Regarding
$o_2$, the distance of the furthest point (i.e., $p_2$) from the virtual line is below $\delta$;
thus, all intermediate points are removed from $o_2$'s simplified trajectory.
Figure~\ref{fig_simple}(\emph{b}) visualizes the simplified trajectories. Notice that each point in
a simplified trajectory corresponds to a point in the original trajectory and is associated with a
time value.
\vspace{0.2cm} \noindent \textbf{Measuring actual tolerances of simplified trajectories :}
We observe that an actual tolerance smaller than $\delta$ may exist so that the simplified
trajectory is valid. In the example of Figure~\ref{fig_simple}(\emph{a}), the actual tolerance of
$o'_2$ is determined by the distance between $p_2$ and the virtual line. We formally define the
actual tolerance as follows:
\begin{mydefinition} \label{def:act-tol}
\textbf{(Actual Tolerance)} Let $l'$ be a line segment in the
simplified trajectory $o'$, whose original trajectory is $o$.
The actual tolerance $\delta(l')$ of $l'$ is defined as:
$max_{t \in l'.\tau} \; D_{PL}(o(t),l')$.
The actual tolerance $\delta(o')$ of $o'$ is defined as the maximum
$\delta(l')$ value over all its line segments.
\end{mydefinition}
The actual tolerance of each line segment $l'$ of $o'$ can be computed
easily by examining the locations of $o$ during the corresponding time
interval $l'.\tau$. In addition, the derivation of these tolerance
values can be seamlessly integrated into the DP algorithm so that the
original trajectory $o$ needs not be examined again.
The actual tolerances are valuable in the sense that they can be exploited to tighten the distance
computation for simplified trajectories, as we will show in the next section.
\subsection{Distance Bounds for Range Search} \label{sec:errorbound}
A simplified trajectory $o'$ may contain many omitted locations in comparison to its original
trajectory $o$. Thus, it is not possible to perform (density--connected) clustering at individual
time. If we generate virtual positions for the omitted points as done in CMC, there is no use for
the trajectory simplification.
The main challenge becomes one of performing clustering on the line segments of simplified
trajectories so that each snapshot cluster (on the original trajectories) is captured by a cluster
of line segments (from the simplified trajectories).
In density-based clustering techniques (e.g., DBSCAN), the core operation is $e$-neighborhood
search, i.e., to find objects within distance $e$ of a given object, at a fixed time $t$. We
proceed to develop the implementation of this core operation in the context of line segments. Let a
line segment $l'_q$ be given; our goal is then to retrieve all line segments $l'_i$ whose original
trajectory $o_i$ can possibly satisfy the condition $D(o_q(t),o_i(t)) \le e$ for some time point
$t$. This way, all qualifying convoy candidates are guaranteed to be found in the filter step.
Let $o'_q$ and $o'_i$ be simplified trajectories of the original trajectories $o_q$ and $o_i$. At a
given time $t$, the locations of $o_q$ and $o_i$ are $o_q(t)$ and $o_i(t)$. Observe that the
endpoints of line segments in $o'_q$ are timestamped. Let $l'_q$ be a line segment in $o'_q$ such
that its time interval $l'_q.\tau$ covers $t$. Similarly, we use $l'_i$ to denote the line segment
in $o'_i$ satisfying $t \in l'_i.\tau$. Figure~\ref{fig_proof} shows an example of two line
segments $l'_q$ and $l'_i$.
\begin{figure}[hbt]
\centering
\includegraphics[width=5.5cm]{pics/fig_proof.eps}
\caption{Trajectory Segments with Time Intervals Covering $t$}
\label{fig_proof}
\end{figure}
Lemma~\ref{lmm:prune-range} establishes the relationship between distances in the original
trajectories and those in the simplified trajectories.
\begin{lemma} \label{lmm:prune-range} Let $o'_q$ ($o'_i$) be the
simplified trajectory of original trajectory $o_q$ ($o_i$). Given a
time $t$, let $l'_q$ ($l'_i$) be the line segment in $o'_q$ ($o'_i$)
with a time interval that covers $t$.
If $D_{LL}(l'_q,l'_i) > e + \delta(l'_q) + \delta(l'_i)$ then
$D(o_q(t),o_i(t)) > e$.
\end{lemma}
Lemma~\ref{lmm:prune-range} allows us to prune line segments $l'_i$ during the range search of the
given line segment $l'_q$. Figure~\ref{fig_search} illustrates the extended range for search over
simplified line segments with error bounds. In Figure~\ref{fig_search}(\emph{a}), half of the
points on the original trajectory are omitted (i.e., a 50\% reduction) with the given $\delta$
value. To enable correct discovery processing over the simplified trajectories (dotted lines), we
enlarge the search space as shown in Figure~\ref{fig_search}(\emph{b}).
\begin{figure}[hbt]
\centering
\includegraphics[width=6.6cm]{pics/fig_search.eps}
\caption{Range Search with Error Bounds}
\label{fig_search}
\end{figure}
Notice that we still need to scan all $l'_i$ whose time intervals intersect with that of $l'_q$.
For example, the time interval [$t_3$,$t_7$] of the second line segment of $o'_q$ in
Figure~\ref{fig_part}(\emph{a}) intersects all of $o'_i$'s line segments. To obtain better
performance, we intend to prune a subset $S$ of line segments fast. During the range search of the
given line segment $l'_q$, Lemma~\ref{lmm:prune-group-range}, next, enables us to prune an
non-qualifying $S$ before examining its line segments. The proofs of Lemma~\ref{lmm:prune-range}
and Lemma~\ref{lmm:prune-group-range} are provided in the appendix.
\begin{lemma} \label{lmm:prune-group-range} Let $S$ be a subset of
line segments $l'_i$ (from simplified trajectories). Let $\mathcal{B}(S)$ be the minimum bounding
box of all segments in $S$, $S.\tau = \bigcup_{l'_i \in S} l'_i.\tau$, and $\delta_{max}(S) =
\max_{l'_i \in S} \delta(l'_i)$. Let line segment $l'_q$ have a time interval that intersects with
that of $S$, i.e., $S.\tau \cap l'_q.\tau \ne \emptyset$.
If $D_{min}(\mathcal{B}(l'_q),\mathcal{B}(S)) > e + \delta(l'_q) +
\delta_{max}(S)$ then \\
$D(o_q(t),o_i(t)) > e$ holds for all $l'_i \in S$.
\end{lemma}
We proceed to outline how to perform range search for $l'_q$ in multiple steps by gradually
tightening the condition: First, we retrieve a set of line segments $S$ whose time intervals
overlap with that of $l'_q$. We then apply Lemma~\ref{lmm:prune-group-range} to prune
non-qualifying line segments in $S$ at an early stage. Next, for each remaining line segment in
$S$, we discard non-qualifying line segments by applying Lemma~\ref{lmm:prune-range}. Any surviving
line segment is included in the $e$-neighborhood of the line segment $l'_q$. Using this multi-step
range search for line segments, we are able to perform density--connected clustering of line
segments efficiently.
\begin{figure}[hbt]
\centering
\includegraphics[width=8.5cm]{pics/fig_partition.eps}
\caption{Measure of $\omega(o'_q,o'_i)$ and Time Partitioning}
\label{fig_part}
\end{figure}
\noindent \textbf{Extension for trajectories :}
So far, we have addressed range search only for line segments. In fact, it is feasible to
generalize the search to apply to an entire trajectory. And by applying clustering on trajectories
directly, we further reduce the cost of the filter step. As we will see in the next section, the
technique below is applicable to sub-trajectories as well, enabling us to control the granularity
of the filter step.
We aim to retrieve all simplified trajectories $o'_i$ whose original trajectories $o_i$ possibly
satisfy the condition $D(o_q(t),o_i(t)) \le e$ for some time $t$. In case $o'_q$ and $o'_i$ have
disjoint time intervals (i.e., $o'_q.\tau \cap o'_i.\tau = \emptyset$), they cannot belong to the
same convoy. Otherwise, we define their $\omega$ value as follows:
$\omega(o'_q,o'_i)= \min \{ D_{LL}(l'_q,l'_i) - \delta(l'_q) - \delta(l'_i) \; | \; l'_i \in o'_i,
\; l'_q \in o'_q, $
\\*
$l'_q.\tau \cap l'_i.\tau \ne \emptyset\}$.
Figure~\ref{fig_part}(\emph{a}) shows an example of computing the $\omega(o'_q,o'_i)$ value between
two simplified trajectories $o'_q$ and $o'_i$. Line segments with shared time interval are linked
by dotted lines, contributing a term in the value of $\omega(o'_q,o'_i)$.
If $\omega(o'_q,o'_i) > e$, no time $t$ exists such that
$D(o_q(t),o_i(t)) \le e$. Otherwise, their locations in the original
trajectories may be within distance $e$ for some time $t$.
\subsection{The CuTS Algorithm} \label{sec:adaptive}
We first present a general overview of the CuTS (\underline{C}onvoy Discovery \underline{u}sing
\underline{T}rajectory \underline{S}implification) algorithm, then illustrate aspects of the
algorithm with examples, and finally present the details of the algorithm.
In the filter step, we first apply simplification (with tolerance $\delta$) to the original
trajectories in order to obtain their simplified trajectories. We then partition the time domain
(with each partition covering $\lambda$ time points) and assign the line segments of each $o'_i$ to
qualifying partitions. Next, we perform clustering on those line segments. Clusters across adjacent
partitions with common objects are used to form convoy candidates. In the refinement step, we
perform clustering of the original trajectories of the objects in each convoy candidate. The total
computational cost of the CuTS algorithm is the sum of the simplification, the clustering, and the
refinement costs. Our experiments in Section~\ref{sec:exp} suggest that the simplification and
refinement costs are very low in practice.
To understand the filter step of CuTS better, consider Figure~\ref{fig_part}(\emph{b}) where the
time domain is divided into equal-length ($\lambda=4$) partitions $\mathcal{T}_1$ and
$\mathcal{T}_2$ with time intervals $[t_1,t_4]$ and $[t_4,t_7]$, respectively.
The time partition $\mathcal{T}_1$ contains the following line
segments: $l^1_1$ and $l^2_1$ of $o'_1$, $l^1_2$ of $o'_2$, and
$l^1_3$ and $l^2_3$ of $o'_3$.
Note that the line segment $l^2_3$ will be inserted into both $\mathcal{T}_1$ and $\mathcal{T}_2$
to avoid any possible false dismissal when we compute the value of $\omega(o'_q,o'_i)$ in
Figure~\ref{fig_part}(\emph{a}).
\vspace{0.2cm} \noindent \textbf{Algorithm description.}
Algorithm~\ref{alg:CuTS_filter} presents the pseudocode of CuTS's filter step. In addition to the
convoy query parameters $m$, $k$, and $e$, two internal parameters $\delta$ (tolerance for
trajectory simplification) and $\lambda$ (the length of each partition) also need to be specified.
Those parameter values are relevant to the performance only (e.g., execution time) and do not
affect the correctness. Guidelines for choosing their values will be presented in
Section~\ref{sec:param}.
\begin{algorithm}[hbt]
\small
\caption{\bf CuTS\_Filter (Object set $O$, Integer $m$, Integer $k$,
Distance threshold $e$)}
\begin{algorithmic}[1]
\STATE $\delta \leftarrow$ ComputeDelta($O$, $e$)
\FOR {each trajectory $o_i \in O$}
\STATE $o'_i \leftarrow$ Douglas-Peucker($o_i, \delta$)
\ENDFOR
\STATE $\lambda \leftarrow$ ComputeLambda($O$, $k$, $(\sum_i |o_i| ) / (\sum_i |o'_i| )$)
\STATE $V \leftarrow \emptyset$
\STATE divide the time domain into $\lambda$-length disjoint partitions
\FOR {each time partition $\mathcal{T}_z$ (in ascending order)}
\STATE $V_{next} \leftarrow \emptyset$
\FOR {each $o'_i$ satisfying $o'_i.\tau \cap \mathcal{T}_z.\tau \ne \emptyset$}
\STATE insert $l^j_i \in o'_i$ (intersecting time interval of $\mathcal{T}_z$) into $\mathcal{G}$
\ENDFOR
\STATE $C \leftarrow$ TRAJ-DBSCAN($\mathcal{G}, e, m$)
\FOR {each convoy candidate $v \in V$}
\STATE $v.$assigned $\leftarrow$ false
\FOR {each cluster $c \in C$}
\IF {$|c \cap v| \geq m$}
\STATE $v$.assigned $\leftarrow$ true
\STATE $v' \leftarrow c \cap v$
\STATE $v'$.lifetime $\leftarrow v$.lifetime + $\lambda$
\STATE $V_{next} \leftarrow V_{next} \cup v'$
\STATE $c$.assigned $\leftarrow$ true
\ENDIF
\ENDFOR
\IF {$v$.assigned = false and $v$.lifetime$\ge k$}
\STATE $V_{cand} \leftarrow V_{cand} \cup v$
\ENDIF
\ENDFOR
\FOR {each $c \in C$}
\IF {$c$.assigned = false}
\STATE $c$.lifetime $\leftarrow \lambda$
\STATE $V_{next} \leftarrow V_{next} \cup c$
\ENDIF
\ENDFOR
\STATE $V \leftarrow V_{next}$
\ENDFOR
\RETURN $V_{cand}$
\end{algorithmic}
\label{alg:CuTS_filter}
\end{algorithm}
Lines~2--3 of the algorithm perform trajectory simplification for all objects. Next, the time
domain is partitioned, each partition holding $\lambda$ consecutive time points.
Time partitions are then processed iteratively in ascending order of their time. Let the current
loop consider the time partition $\mathcal{T}_z$. The algorithm builds a polyline (i.e., a sequence
of line segments) from a simplified trajectory $o'_i$, which contains the line segments of $o'_i$
whose time intervals intersect to $\mathcal{T}_z$. It then stores all the polylines from each
simplified trajectory into a data structure $\mathcal{G}$. Next, density clustering is performed
for the sub-trajectories in $\mathcal{G}$ (see Line~11).
The set $V$ keeps track of the convoy candidates found in previous iterations, whereas the set
$V_{next}$ stores new candidates found in the current iteration. For Lines~12--20, each cluster $c
\in C$ (found in the current iteration) is joined with those in $V$, as long as their intersections
have at least $m$ objects. Also, candidate convoys with lifetime above $k$ are inserted into the
candidate set $V_{cand}$. Clusters that cannot join with previous convoy candidates are then
considered as new candidates (Lines~23--26).
Finally, Algorithm~\ref{alg:CuTS_verify} contains the pseudocode of the refinement step of the CuTS
algorithm. Suppose that $v$ is the convoy candidate in the candidate set $V$ that is currently
being examined. We first determine the time interval $[t_{start}, t_{end}]$ for $v$ and then
identify the set $O'$ of the original trajectories whose line segments appear in $v$. Finally, we
apply CMC for trajectories in $O'$, considering only time points in the interval $[t_{start},
t_{end}]$.
\begin{algorithm}[hbt]
\small
\caption{\bf CuTS\_Refinement (Candidate set $V_{cand}$, Object set
$O$, Integer $m$, Integer $k$, Distance threshold $e$)}
\begin{algorithmic}[1]
\FOR {each $v \in V_{cand}$}
\STATE $t_{start} \leftarrow$ start time of $v$
\STATE $t_{end} \leftarrow$ end time of $v$
\STATE $O' \leftarrow \{ o_i \in O \; | \; l^j_i.\tau \in v \cap l^j_i.\tau \in o_i \}$
\STATE call CMC($O', m, k, e$) with the time interval $[t_{start},t_{end}]$
\ENDFOR
\end{algorithmic}
\label{alg:CuTS_verify}
\end{algorithm}
\section{Experiments} \label{sec:exp}
In this experimental study, we first compare the discovery efficiency between CMC, which is an
adaption of a moving-clustering algorithm (MC2) \cite{st-clustering1} for our convoy discovery
problem, and the CuTS family (CuTS, CuTS+, and CuTS*). We then analyze the performance of each
method of the CuTS family while varying the settings of their key parameters.
We implemented the above algorithms in the C++ language on a Windows Server 2003 operating system.
The experiments were performed using an Intel Xeon CPU 2.50 GHz system with 16GB of main memory.
\subsection{Dataset and Parameter Setting}
For studying the performance of our methods in a real-world setting, we used several real datasets
that were obtained from vehicles and animals. Due to the different object types, their trajectories
have distinct characteristics, such as the frequency of location sampling and data distributions.
The details of each dataset are described as follows:
\vspace{0.2cm}
\vspace{0.1cm} \noindent \textbf{Truck}: We obtained 276 trajectories of 50 trucks moving in the
Athens metropolitan area in Greece \cite{rtreeportal}. The trucks were carrying concrete to several
construction sites for 33 days while their locations were measured. To be able to find more
convoys, we regarded each trajectory as a distinct truck's trajectory and removed the day
information from the data. Thus, the dataset became 276 trucks' movements on the same day.
\vspace{0.1cm} \noindent \textbf{Cattle}: To reduce a major cost for cattle producers, a virtual
fencing project in CSIRO, Australia studied managing herds of cattle with virtual boundaries. We
obtained 13 cattle's movements for several hours from the project. Their locations were provided by
GPS-enabled ear--tags every second. A distinguishable aspect of this dataset is its very large
number of timestamps.
\vspace{0.1cm} \noindent \textbf{Car}: Normal travel patterns of over 500 private cars were
analyzed for building reasonable road pricing schemes in Copenhagen, Denmark. We obtained 183 cars'
trajectories during one week \cite{daisy}. Trajectories in this dataset had very different lengths.
\vspace{0.1cm} \noindent \textbf{Taxi}: The GPS logs of 500 taxis in Beijing, China were recorded
during a day and studied in Institute of Software, Chinese Academy of Sciences. The locations of
the trajectories were sampled irregularly. For example, some taxis reported their locations every
three minutes, while some did it once in several minutes.
\vspace{0.2cm}
In our experiments, we defined a convoy as containing at least 3 objects (except Cattle due to the
small number of objects) that travel closely for 3 minutes (i.e., $m=3$ and $k=180$). We also
adjusted the values of neighborhood range $e$ to be able to find 1 to 100 convoys for each dataset.
To perform convoy discovery using our main methods (CuTS, CuTS+, and CuTS*), we still need to
determine two key parameters, namely the tolerance value ($\delta$) for trajectory simplification
and the length of time partition ($\lambda$). These parameter values were computed by our
guidelines that will be discussed in Section~\ref{sec:param}.
Table~\ref{tab:data} provides (i) detailed information of each dataset,
(ii) the settings of the parameters to be used throughout our experiments, and (iii) the number
of convoys discovered by our proposed methods with the parameters.
\begin{table}[hbt]
\centering
\small
\begin{tabular}{|c|c|c|c|c|}
\hline & {\bf Truck} & {\bf Cattle} & {\bf Car} & {\bf Taxi} \\
\hline number of objects ($N$) & 267 & 13 & 183 & 500 \\
\hline time domain length ($T$) & 10586 & 175636 & 8757 & 965 \\
\hline average trajectory length & 224 & 175636 & 451 & 82 \\
\hline data size (points) & 59894 & 2283268 & 82590 & 41144 \\
\hline\hline number of convoy objects ($m$) & 3 & 2 & 3 & 3 \\
\hline convoy lifetime ($k$) & 180 & 180 & 180 & 180 \\
\hline neighborhood range ($e$) & 8 & 300 & 80 & 40 \\
\hline simplification tolerance ($\delta$)& 5.9 & 274.2 & 63.4 & 31.5 \\
\hline time partition length ($\lambda$)& 4 & 36 & 24 & 4 \\
\hline\hline number of convoys discovered & 91 & 47 & 15 & 4 \\
\hline
\end{tabular}
\caption{Settings for Experiments}\label{tab:data}
\end{table}
\subsection{CMC vs. The CuTS Family}\label{exp:cmc}
First, we compared the efficiency of CMC versus the CuTS family. Over all the datasets, the CuTS
family was 3.9 times (at least) to 33.1 times (at most) faster than CMC, as seen in
Figure~\ref{exp_cmc}, and especially CuTS* had the highest efficiency. The performance differences
were more obvious in the Car and the Taxi datasets though their data sizes (total number of points)
were less than 10\% of Cattle's data size. Since those two datasets had many numbers of missing
points and different lifetimes of each trajectory, CMC incurred extra computational cost to make
virtual points for those missing times to measure density-connection correctly (see
Section~\ref{sec:cmc}). It also caused a considerable growth of the actual data size for the
discovery processing. Notice that our main methods, the CuTS family, can perform the discovery
without any extra processing regardless of the number of missing points.
\begin{figure}[hbt]
\center
\includegraphics[width=8.5cm]{pics/exp/cmc.eps}
\caption{Comparisons of Query Processing Time}
\label{exp_cmc}
\end{figure}
In Figure~\ref{exp_anal}, we report on the elapsed times of each method of the CuTS family for the
Cattle and Taxi datasets (magnified views of the results in Figure~\ref{exp_cmc}). For brevity, we
show the two most distinctive results only. In the results for the Cattle dataset, the
simplification cost dominates for all the methods. In general, convoy processing is more sensitive
to the number of objects $N$ than to the number of timestamps $T$ since the clustering method
(DBSCAN) has $O(N^2)$ computational cost ($O(N \cdot \log N)$ with a spatial index). The Cattle
dataset has only 13 objects, and the cost of each clustering is very low though it is performed $T$
times. As a result, the total discovery times are more influenced by the simplification process
than the filtering and refinement steps.
\vspace{0.2cm}
\begin{figure}[!h]
\center
\includegraphics[width=8.5cm]{pics/exp/anal.eps}
\caption{Analysis of Query Processing Cost}
\label{exp_anal}
\end{figure}
The reason why each simplification method has different efficiency will be studied in
Section~\ref{exp:cuts}. Recall that, although the CuTS family needs much time for trajectory
simplification on the Cattle dataset, their total discovery times are still much lower than those
of CMC in the previous experiments.
Another interesting observation found with the Cattle data is that CuTS+ has not only faster
trajectory simplification, but also lower refinement cost. This is because DP+ as used in CuTS+ has
not only higher efficiency of simplification, but also tighter error bounds than DP as used by
CuTS, as described in Section~\ref{sec:cuts+}.
Compared to the Cattle data, trajectory simplification had very low computational cost on the Taxi
dataset. As the Taxi dataset has a short $T$ but a larger $N$, the clustering cost dominates the
discovery time. In addition, since the number of convoy candidates was small for this data (will be
shown in the next experiments), only little refinement was necessary.
For the other two datasets, the composition of computational time was about 70\%-80\% for filtering
(around 5\%-15\% for trajectory simplification) and 20\%-30\% for refinement. Therefore, it is very
reasonable to `invest' some time in trajectory simplification.
We also studied the effect of using the actual tolerance for the range search of clustering. When
we perform the trajectory simplification, we use the tolerance value $\delta$, named the
\emph{global tolerance} here. The key process of the simplification is to remove intermediate
points whose distances from the virtual line linking two end points of the original trajectory do
not exceed $\delta$. Any distance of those removed points (i.e., \emph{actual tolerance}) is always
smaller than or equal to the global tolerance (see Section~\ref{sec:simple}). The actual tolerance
is useful for range search since the search area should be reduced.
\begin{figure}[hbt]
\center
\includegraphics[width=8.5cm]{pics/exp/act_tol.eps}
\caption{Effect of Actual Tolerance}
\label{exp_acttol}
\end{figure}
Figure~\ref{exp_acttol}(\emph{a}) demonstrates the filtering power of the global and actual
tolerances for CuTS*. We omit the results for CuTS and CuTS+ because they are similar. As shown,
the number of candidates after the filtering step decreases considerably when we use the actual
tolerance. The advantage of the improved filtering by the actual tolerance is reflected in the
efficiency of convoy discovery as shown in Figure~\ref{exp_acttol}(\emph{b}). Yet, the effect is
relatively small on the Truck and the Taxi datasets. This is because some candidates that do not
need much computation for the refinement step are pruned when using the actual tolerance. We
present a more precise way of measuring the filter's effectiveness in the following section.
\subsection{CuTS vs. CuTS+ vs. CuTS*} \label{exp:cuts}
We have already discussed different techniques for trajectory simplification. The difference
between the original Douglas-Peucker algorithm (DP) and its temporal extension DP* was covered in
Section~\ref{sec:trjsimple}. We also developed a DP variant, named DP+, in Section \ref{sec:cuts+}.
It is of interest to compare the performance of those methods.
Figure \ref{exp_dp}(\emph{a}) illustrates the differences of their reduction power for the Cattle
dataset. We skip the results for the other datasets because they show similar trends. With the same
values of tolerance, DP shows higher reduction rates than does DP*. This is natural since DP* uses
the time-ratio distance to approximate points, which is always equal to or greater than the
perpendicular distance of DP (see Section~\ref{sec:cuts*}). Furthermore, the vertex reduction of
DP+ is lower than that of DP. This is because DP+ does not preserve the shapes of the original
trajectories well when compared to DP. This aspect was explained in Section~\ref{sec:cuts+}.
\begin{figure}[hbtt]
\center
\includegraphics[width=8.5cm]{pics/exp/dp.eps}
\caption{Comparison of Trajectory Simplification Methods}
\label{exp_dp}
\end{figure}
\vspace{0.2cm}
In Figure \ref{exp_dp}(\emph{b}), DP+ exhibits the fastest elapsed time among the methods because
of its more effective division process. An interesting observation of the figure is that the
efficiency of all the methods grows as reduction ratios increase. Recall that all the methods
utilize the divide-and-conquer paradigm, which divides an input trajectory until no point exceeds a
given $\delta$. With a larger value of $\delta$, their division processes are likely to meet the
`end' quicker. For this reason, DP* also performs slower than the other methods (lower reduction
power than the others).
Next, we compare the discovery effectiveness and efficiency for the CuTS family. Given very large
values of $e$ and $\delta$, the CuTS family may produce one candidate containing all actual results
after the filter step and then the candidate may be divided into a large number of real convoys
through the refinement step. Thus, we cannot use the count of false positives as a measure of the
filters' effectiveness for our study.
Instead, we calculate \emph{refinement unit} that represents the computational cost of candidates
for the refinement step, which reflects the filtering power of each method effectively.
Specifically, the clustering cost of the convoy objects in each candidate is computed and then
multiplied by the candidate's lifetime. As mentioned earlier, the computational cost of clustering
is either $O(N^2)$ without index or $O(N \cdot \log N)$ with a spatial index. To clarify the
differences of each filter method, we considered the clustering without index support in our
experiments. For example, if a convoy candidate has 3 objects and its lifetime is 2, the refinement
unit is $3^2 \times 2 = 18$. Next, we aggregate each candidate's unit to obtain the total
refinement unit.
\begin{figure}[hbt]
\center
\includegraphics[width=8.5cm]{pics/exp/tol.eps}
\caption{Effect of Simplification Tolerance ($\delta$)}
\label{exp_tol}
\end{figure}
Figure~\ref{exp_tol} demonstrates the filtering power and the total discovery times for the CuTS
family when varying $\delta$. We omit the results for the Truck and Cattle datasets, but those two
datasets will be used in the next experiments. As expected, CuTS* has the lowest refinement unit
for both datasets, which yields the highest efficiency as well. In addition, CuTS+ has a better
filtering effectiveness than does CuTS. As discussed in Section~\ref{sec:cuts+}, the actual
tolerances obtained by DP+ of CuTS+ are always smaller or equal to those obtained by DP of CuTS. As
a result, the search range for clustering is reduced, and the filtering power grows in the figure.
\pagebreak
Another observation found in Figure~\ref{exp_tol}, for all members of the CuTS family is that both
the filters' effectiveness and the discovery efficiency decrease as the tolerance value increases
as the $\delta$ values affect not only the result of trajectory simplification, but also that of
range search for clustering.
Although the total elapsed times of the Car data grow steadily with increasing $\delta$, those of
the Taxi data stay almost constant or increase only very slightly. This is because the enlargement
of the search range is not sufficient to find more actual convoys with respect to the given
parameters. From this point, we can infer that the trajectories of the Taxi dataset are distributed
relatively uniformly, and thus the number of taxis traveling together within a given (reasonable)
distance is low.
\begin{figure}[hbt]
\center
\includegraphics[width=8.5cm]{pics/exp/lambda.eps}
\caption{Effect of Time Partitioning ($\lambda$)}
\label{exp_lambda}
\end{figure}
Lastly, we study how the size of the time partition $\lambda$ affects the results of the convoy
discovery. In fact, a large value of $\lambda$ yields an ineffective filtering step, whereas more
times of clustering are performed with a small value of $\lambda$ (see Section~\ref{sec:adaptive}).
In the Truck dataset of Figure~\ref{exp_lambda}, CuTS* shows better performance than the other
methods regardless of the $\lambda$ value. Also, both the effectiveness of the filters and the
efficiency of the discovery process decrease when $\lambda> 10$ for this dataset, for all methods
of the CuTS family.
On the other hand, the discovery efficiency of the CuTS family declines over the Cattle dataset
when $\lambda < 30$, although their refinement unit increases steadily in the same range of
$\lambda$. This implies that an appropriate $\lambda$ value is influenced by not only the filter's
effectiveness, but also another fact, possibly the length of trajectories since the average size of
Cattle's trajectories is very large.
Another interesting observation found in the Cattle dataset is that CuTS+ has similar efficiency to
CuTS*, and it is even faster for $\lambda \geq 50$. As seen in Figure~\ref{exp_anal}, trajectory
simplification is the key part of the total discovery time on this dataset. Therefore, faster
trajectory simplification (i.e., DP+) plays a more important role in the discovery efficiency in
this case.
\subsection{Parameter Determination of CuTS} \label{sec:param}
Proper values of $\delta$ and $\lambda$ may be difficult to find in some applications since they
are dependent on the data characteristics. In this section, we provide guidelines for determining
settings for these parameters. Note that the parameters do not affect the correctness of discovery
results, but only affect execution times.
\vspace{0.2cm} \noindent \textbf{Tolerance for trajectory simplification ($\delta$)} : It is
obvious that a larger value of $\delta$ for DP of CuTS achieves a higher reduction result of
trajectory simplification. On the other hand, a large $\delta$ value is also used for the range
search of clustering in the CuTS algorithm; hence, the filter step of CuTS may not be tight enough
to prune many unnecessary candidate objects. In this tradeoff, our goal is to find a value
satisfying the following conditions : (i) the original trajectories become well simplified, and
(ii) the distance bounds are sufficiently tight, implying an effective filter process.
As the first step, we perform the original DP algorithm over a trajectory with $\delta=0$. In each
step of the division process (see details in Sections~\ref{sec:trjsimple} and \ref{sec:simple}), we
store the actual tolerance values in ascending order. Since $\delta=0$, the process continues until
all intermediate points of the original trajectory are tested.
In the next step, we find the largest variance between two adjacent tolerances stored, and then
select the smaller one of those two tolerances. For example, assume that the DP method with
$\delta=0$ results in the 10 actual tolerance values $\delta_1,\delta_2,\cdots,\delta_{10}$ in
Figure~\ref{fig_param}(\emph{a}) through the first step. The difference in the tolerance values is
the largest between $\delta_5$ and $\delta_6$. We then select $\delta_5$ as a tolerance value
$\delta_s$. This selection is performed as long as $\delta_i < e$ (the dark gray bars in
Figure~\ref{fig_param}(\emph{a})). From our experimental studies, we found out that the filtering
power of the CuTS family decreases considerably on some datasets when we pick $\delta_i> e$.
Lastly, we perform the above steps for a sufficient time (e.g., 10\% of $N$) and average the
$\delta_s$ values selected to obtain a final $\delta$ for the processing of trajectory
simplification.
The idea behind this method is to find a relatively small $\delta$ value that achieves a reasonable
reduction through simplification. In the figure, if we pick $\delta_{10}$ and apply it to the
trajectory simplification, the reduction ratio will be nearly 100\%. Likewise, the use of
$\delta_5$ for the simplification is able to yield around a 50\% reduction although it does not
necessarily follow the same division processes with $\delta=0$ as the first step. If we pick
$\delta_6$ instead, it may bring (approximately) 60\% of trajectory reduction, which is slightly
higher than 50\%. However, the value of $\delta_6$ is much bigger than $\delta_5$, and the
effectiveness of range search can decrease dramatically.
\begin{figure}[hbt]
\center
\includegraphics[width=8.5cm]{pics/fig_param.eps}
\caption{Value Selection of $\delta$ and $\lambda$}
\label{fig_param}
\end{figure}
\noindent \textbf{Length of time partition ($\lambda$)} : In Section~\ref{sec:adaptive}, we
discussed about dividing the time domain $T$ into time partitions for discovery processing, each of
which has length $\lambda$. If a time partition $\mathcal{T}_i$ has a large value for $\lambda$,
many line segments of a simplified trajectory within $\mathcal{T}_i$ form a long polyline. Thus,
the distances among those polylines become small, and many objects are likely to form a cluster
together, leading to ineffective filtering results. In contrast, a small value of $\lambda$
involves many computationally expensive clustering processes ($T/\lambda$ times).
Suppose that $o'_1$ and $o'_2$ in Figure~\ref{fig_param}(\emph{b}) are simplified trajectories. In
the figure, one clustering with $\lambda_1$ is obviously more efficient than two processes with
$\lambda_0$ because both cases have the same minimum distance between $o'_1$ and $o'_2$. From this
example, we can infer the value of $\lambda_1$ by computing $\frac{|o'|}{|o|} \times o.\tau$, where
$|o|$ ($|o'|$) is the number of points in the trajectory $o$ ($o'$) and $o.\tau$ is the time
interval of $o$.
In practice, however, there may be some time points that one (simplified) trajectory has, but
others do not have, such as $p'_2$ on $o'_3$ in Figure~\ref{fig_param}(\emph{c}). Using the
$\lambda_1$ for this case should not keep the filter's `good' effectiveness, and we need to lower
the $\lambda_1$ value. We can roughly estimate the probability that such case occurs by looking at
how densely a trajectory exits in the time space $T$. Notice that each trajectory may have a
different length ($o.\tau$) and may appear and disappear at any arbitrary time points in $T$. Thus,
the density of the trajectory is obtained by $o.\tau / T$. Finally, the probability that an object
has an intermediate time point within $\lambda_1$ is $(\lambda_1 - 2) \times o.\tau / T$. Together,
we obtain $\lambda = \lambda_1 - (\lambda_1 - 2) \times o.\tau / T$, rewriting $\lambda = o.\tau
\times ( \frac{|o'|}{|o|} \times ( 1 - \frac{o.\tau}{T} ) + \frac{2}{T})$.
So far, we have considered the computation of $\lambda$ for a single object. To obtain an overall
value of $\lambda$, we perform the above computation for all objects and average the values. Note
that all the statistics for this $\lambda$ computation can be easily gathered when a dataset is
loaded into the system (or one scan for disk-based implementations).
Although this method does not capture the distribution of a dataset precisely, the value of
$\lambda$ is quickly obtained and brings reasonable efficiency of the CuTS family.
\section{Extensions of CuTS} \label{sec:ext}
In this section, we introduce two enhancements of CuTS. One accelerates the process of trajectory
simplification and brings higher efficiency. The other shortens the search range for clustering by
considering temporal information of trajectories, reducing the number of candidates after the
filter step of CuTS.
\subsection{Faster Trajectory Simplification - CuTS+} \label{sec:cuts+}
The Douglas-Peucker algorithm (DP) utilizes the divide-conquer technique (see
Section~\ref{sec:trjsimple}). It is well-known that techniques built on the divide-conquer paradigm
show the best performance if a given input is divided into two sub-inputs equally in each division
step. Inspired by this, we modify the original DP algorithm for speeding up the simplification
process, obtaining DP+.
Specifically, DP+ selects the closest point to the middle of a given trajectory among the points
exceeding tolerance value $\delta$ at each approximation step. Figure \ref{fig_dp+}(\emph{a})
demonstrates an original trajectory having seven points, which has two intermediate points $p_4$
and $p_6$ whose distances from $\overline{p_1p_7}$ are greater than the given $\delta$ value (the
gray area in the figure). The DP method selects the point having the largest distance (i.e.,
$\delta_6$); hence, the result of this division step will be as shown in Figure
\ref{fig_dp+}(\emph{b}).
\begin{figure}[!h]
\centering
\includegraphics[width=8.5cm]{pics/fig_dp+.eps}
\caption{Comparison between DP and DP+}
\label{fig_dp+}
\end{figure}
In contrast, our DP+ method picks the point $p_4$ that is the closest to the middle point of ${p_1,
p_2, \cdots, p_7}$ among intermediate points exceeding $\delta$ (i.e., $p_4$ and $p_6$). This
technique divides $\overline{p_1p_7}$ into two sub-trajectories $\overline{p_1p_4}$ and
$\overline{p_4p_7}$, which have similar numbers of points (Figure \ref{fig_dp+}(\emph{c})).
Therefore, the whole process of trajectory simplification is expected to be more efficient.
\pagebreak
Compared with DP, DP+ may have lower simplification power. In fact, each division process of DP+
does not preserve the shape of a given original trajectory well; hence, the next division process
may not be as effective as that of DP. For example, in Figure~\ref{fig_dp+}(\emph{c}), $p_6$ will
be kept using DP+ because $D_{PL}(p_6,\overline {p_4 p_7}) > \delta$, and then the simplified
trajectory will be $p_1,p_4,p_6,p_7$, whereas $p_1,p_6,p_7$ will be the result of DP in
Figure~\ref{fig_dp+}(\emph{b}).
In spite of the lower reduction, DP+ can enhance the discovery processing of CuTS in two areas.
First, note that we are interested in efficient discovery of convoys in this study. As long as the
search distances are bounded, faster simplification of trajectories can play a more important role
in finding convoys. Second, the actual tolerances obtained by DP+ are always smaller or equal to
those obtained by DP (e.g., $\delta_4 < \delta_6$ in the example). This tightens the error bounds
of range search for clustering, leading a more effective filter step.
We extend CuTS to CuTS+, which is built on the DP+ simplification method. All other discovery
processes of CuTS+ are the same as those of CuTS.
\subsection{Temporal Extension - CuTS*} \label{sec:cuts*}
Recall that CuTS applies trajectory simplification (DP) on original trajectories in the filter
step. However, as we will see shortly, intermediate locations on simplified line segments cannot be
associated with fixed timestamps. Consequently, the bounds on distances between line segments may
not be tight, the result being that overly many convoy candidates can be produced in the filter
step. This may yield a more expensive refinement step.
In this section, we extend CuTS to CuTS* by considering temporal aspects for both the trajectory
simplification and the distance measure on simplified trajectories. This enables us to tighten
distance bounds between simplified trajectories, improves the effectiveness of the filter step.
\vspace{0.3cm} \noindent \textbf{Comparison between DP and DP*}: We discussed the differences
between the two trajectory simplification techniques DP \cite{DP} and DP* \cite{st-simple1} in
Section~\ref{sec:trjsimple}. In Figure~\ref{fig_stsimpleFT}(\emph{b}), DP* translates the time
ratio of $p_2$ between $p_1$ and $p_3$ into a location $p'_2$ on the line segment
$\overline{p_1p_3}$. Since $p_2$ exceeds the $\delta$ range of $p'_2$, the point $p_2$ is kept in
the simplified trajectory $o'_1$, which is different from DP.
From the example, we can see that DP* has a lower vertex reduction ratio for trajectories.
Nevertheless, DP* permits us to derive tighter distance measures between trajectory segments,
improving the overall effectiveness of the filter step.
\begin{figure}[hbt]
\centering
\includegraphics[width=7.2cm]{pics/fig_dist3d}
\caption{Different Distance Measures of Trajectory Segments}
\label{fig_dist3d}
\end{figure}
\pagebreak
Figure~\ref{fig_dist3d}(\emph{a}) shows two simplified line segments $l'_1$ and $l'_2$, obtained
from DP. Here, $l'_1$ has the endpoints $p'_1$ and $p'_4$, corresponding to its locations at times
$t_1$ and $t_4$. Similarly, $l'_2$ has endpoints $b'_3$ and $b'_5$, corresponding to its locations
at times $t_3$ and $t_5$. The shortest distance between $l'_1$ and $l'_2$ is given by
$D_{LL}(l'_1,l'_2)$.
Figure~\ref{fig_dist3d}(\emph{b}) contains simplified line segments from DP*. Since DP* captures
the time ratio in the simplified line segment, we are able to derive the locations $l'_1(3)$ on
$l'_1$ and $l'_2(4)$ on $l'_2$. Let $l'_p=\{p_u,p_v\}$ be a simplified line segment having a time
interval $l'_p.\tau=[u,v]$. The location of $l'_p$ at a time $t \in [u,v]$ is defined as:
$$l'_p(t) = p_u + \frac{t-u}{v-u} (p_v-p_u)$$
Note that the terms $l'_p(t)$, $p_u$, and $(p_v-p_u)$ are 2D vectors representing locations.
Before defining $D_*(l'_1,l'_2)$ formally, we need to introduce the time of the \emph{Closest Point
of Approach}, called the CPA time ($t_{CPA}$) \cite{st-join6}. This is the time when the distance
between two dynamic objects is the shortest, considering their velocities. Let $l'_q=\{q_w,q_x\}$
be another simplified line segment during $l'_q.\tau=[w,x]$. The CPA time of $l'_p$ and $l'_q$ is
computed by :
$$t_{CPA}= \frac{-(p_u - q_w) \cdot (l'_p(t) - l'_q(t))}{|l'_p(t) - l'_q(t)|^2}$$
where, $l'_q(t)$, $q_w$, and $(q_w-q_x)$ are also location vectors.
Observe that the common interval of $l'_1$ and $l'_2$ is $[t_3, t_4]$ (gray area in
Figure~\ref{fig_dist3d}(\emph{b})). The {\em tightened} shortest distance $D_*(l'_1,l'_2)$ between
them is computed as :
$$D_*(l'_1,l'_2)= D(l'_1(t_{CPA}),l'_2(t_{CPA}))\hspace{0.5cm} t_{CPA} \in (l'_1.\tau \cap l'_2.\tau)$$
When their time intervals do not intersect, i.e., $l'_1.\tau \cap l'_2.\tau = \emptyset$, their
distance is set to $\infty$ .
Clearly, $D_*(l'_1,l'_2)$ is longer than $D_{LL}(l'_1,l'_2)$; hence, the line segments in
Figure~\ref{fig_dist3d}(\emph{b}) have a lower probability of forming a cluster together than do
those in Figure~\ref{fig_dist3d}(\emph{a}). These tightened distance bounds improve the
effectiveness of the filter step.
\vspace{0.2cm} \noindent \textbf{Distance bounds for DP* simplified line segments}:
Using the notations from Lemma~\ref{lmm:prune-range}, we derive the counterpart that uses the
tightened distance $D_*$ between line segments (as opposed to the distance $D_{LL}$).
Lemma~\ref{lmm:prune-star-range} establishes the relationship between distances in original
trajectories and those in simplified trajectories (obtained by DP*). The proof is provided in the
appendix.
\begin{lemma} \label{lmm:prune-star-range} Suppose that $o'_q$
($o'_i$) is the simplified trajectory (from DP*) of the original
trajectory $o_q$ ($o_i$). Given a time $t$, let $l'_q$ ($l'_i$)
be the line segment in $o'_q$ ($o'_i$) with time interval covering
$t$.
If $D_{*}(l'_q,l'_i) > e + \delta(l'_q) + \delta(l'_i)$ then
$D(o_q(t),o_i(t)) > e$.
\end{lemma}
\noindent \textbf{CuTS* algorithm for convoy discovery}:
We develop an enhanced algorithm, called CuTS*, to exploit the above tightened distance bounds for
query processing. Two components of CuTS need to be replaced. First, CuTS* applies DP* for the
trajectory simplification. Second, during density clustering in the filter step,
Lemma~\ref{lmm:prune-star-range} is utilized in the range search operations (as opposed to
Lemma~\ref{lmm:prune-range}). The above modifications improve the effectiveness of the filter step
in CuTS*. The following table summarizes the key components of CuTS and its extensions.
\begin{table}[hbt]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline {\bf Method} & {\bf CuTS} & {\bf CuTS+} & {\bf CuTS*} \\
\hline simplification & DP \cite{DP} & DP+ [Section~\ref{sec:cuts+}] & DP* \cite{st-simple1} \\
\hline distance function & $D_{LL}$ & $D_{LL}$ & $D_*$ \\
\hline
\end{tabular}
\end{table}
\section{Introduction} \label{sec:intro}
Although the mobile Internet is still in its infancy, very large
volumes of position data from moving objects are already being
accumulated. For example, Inrix, Inc. based in Kirkland, WA receive
real-time GPS probe data from more than 650,000 commercial fleet,
delivery vehicles, and taxis~\cite{inrix}.
As the mobile Internet continues to proliferate and as congestion becomes increasingly widespread
across the globe, the volumes of position data being accumulated are likely to soar. Such data may
be used for many purposes, including travel-time prediction, re-routing, and the identification of
ride-sharing opportunities.
This paper addresses one particular challenge to do with the
extraction of meaningful and useful information from such position
data in an efficient manner.
The movement of an object is given by a continuous curve in the $(\mbox{space}, \mbox{time})$
domain, termed a {\em trajectory}. The past trajectory of an object is typically approximated based
on a collection of time-stamped positions, e.g., obtained from a GPS device. As an example,
Figure~\ref{fig:lossyflock}(\emph{a}) depicts the trajectories of four objects $o_1$, $o_2$, $o_3$,
and $o_4$ in $(x,y,t)$ space.
Given a collection of trajectories, it is of interest to discover
groups of objects that travel together for more than some minimum
duration of time.
A number of applications may be envisioned. The identification of delivery trucks with coherent
trajectory patterns may be used for throughput planning. The discovery of common routes among
commuters may be used for the scheduling of collective transport. The identification of cars that
follow the same routes at the same time may be used for the organization of carpooling, which may
reduce congestion, pollution, and CO$_2$ emissions.
\begin{figure}[hbt]
\centering
\includegraphics[width=6.0cm]{pics/fig_lossyflock.eps}
\caption{\textit{Lossy-flock} Problem}
\label{fig:lossyflock}
\end{figure}
The discovery of so-called \emph{flocks}
\cite{st-flock4,st-flock3,st-flock2} has received some attention. A
flock is a group of objects that move together within a disc of some
user-specified size.
On the one hand, the chosen disk size has a substantial effect on the
results of the discovery process. On the other hand, the selection of
a proper disc size turns out to be difficult, as situations can occur
where objects that intuitively belong together or do not belong
together are not quite within any disk of the given size or are within
such a disk. And for some data sets, no single appropriate disc size
may exist that works well for all parts of the $(\mbox{space},
\mbox{time})$ domain.
In Figure~\ref{fig:lossyflock}(\emph{a}), all objects travel together in a natural group. However,
as shown in Figure~\ref{fig:lossyflock}(\emph{b}), object $o_4$ does not enter the disc and is not
discovered as a member of the flock. A key reason why this \emph{lossy-flock} problem occurs is
that what constitutes a flock is very sensitive to the user-specified disc size, which is
independent of the data distribution.
In addition, the use of a circular shape may not always be appropriate. For example, suppose that
two different groups of cars move across a river and each group has a long linear form along roads.
A sufficient disc size for capturing one group may also capture the other group as one flock.
Ideally, no particular shape should be fixed apriori.
To avoid rigid restrictions on the sizes and shapes of the trajectory patterns to be discovered, we
propose the concept of \emph{convoy} that is able to capture generic trajectory pattern of any
shape and any extent. This concept employs the notion of density connection \cite{s-clustering3},
which enables the formulation of arbitrary shapes of groups.
Given a set of trajectories $O$, an integer $m$, a distance value $e$, and a lifetime $k$, a
\emph{convoy query} retrieves all groups of objects, i.e., \emph{convoys}, each of which has at
least $m$ objects so that these objects are so-called density--connected with respect to distance
$e$ during $k$ consecutive time points. Intuitively, two objects in a group are density--connected
if a sequence of objects exists that connects the two objects and the distance between consecutive
objects does not exceed $e$. (The formal definition is given in Section~\ref{sec:prob}.)
Each group of objects in the result of a convoy query is associated
with the time intervals during which the objects in the group traveled
together.
The efficient discovery of convoys in a large trajectory database is
a challenging problem. Convoy queries compute sets of objects and
are more expensive to process than spatio-temporal
joins~\cite{st-join4}, which compute pairs of objects.
Past studies on the retrieval of similar trajectories generally use
distance functions that consider the distances between pairs of
trajectories across all of time~\cite{EDR,LCSS,DTW}. In contrast, we
consider distances during relatively short durations of time.
Other relevant work concerns the clustering of moving objects
\cite{st-clustering4,st-clustering1,st-clustering2}. In these works, a moving cluster exists if a
shared set of objects exists across adjacent time, but objects may join and leave a cluster during
the cluster's lifetime. Hence, moving clusters carry different semantics and do not necessarily
qualify as convoys.
Jeung et. al. first proposed the convoy query and outlined preliminary techniques for convoy
discovery \cite{convoy08}. In this paper, we extend the work, which develops more advanced
algorithms and analyzes each discovery method in real world settings. Specifically, we introduce
four effective and efficient algorithms for answering the convoy query. The first method adopts the
solution for moving cluster discovery to our convoy problem. The second method, called CuTS
(\emph{\underline{C}onvoy Discovery \underline{u}sing \underline{T}rajectory
\underline{S}implification}), employs the filter-refinement framework --- a set of candidate
convoys are retrieved in the filter step, and then they are further processed in the refinement
step to produce the actual convoys. In the filter step, we apply line simplification techniques
\cite{DP} on the trajectories to reduce their sizes; hence, it becomes very efficient to search for
convoys over simplified trajectories. We establish distance bounds between simplified trajectories,
in order to ensure that no actual convoy is missing from the candidate convoy set. The third method
(CuTS+) accelerates the process of trajectory simplification of CuTS to increase the efficiency of
the filter step even further. The last method, named CuTS*, is an advanced version of CuTS that
enhances the effectiveness of the filter step by introducing tighter distance bounds for simplified
trajectories.
The main novelties of this paper are summarized as follows:
\begin{itemize}
\setlength{\parskip}{-3pt} \item
Our filter step operates on trajectories processed by line simplification techniques; this is
different from most related works that employ spatial approximation (e.g., bounding boxes) in the
filter step. The rationale is that conventional methods using bounding boxes introduce substantial
empty space, rendering them undesirable for the processing of trajectory data.
\setlength{\parskip}{-3pt} \item
To guarantee correct convoy discovery, we establish distance bounds for range search over
simplified trajectories. In contrast, the distance bounds studied elsewhere \cite{st-simple2} are
applicable only to specific query types, not to the convoy problem.
\setlength{\parskip}{-3pt} \item
We study various trajectory simplification techniques in conjunction with different query
processing mechanisms. In addition, we show how to tighten the distance bounds.
\setlength{\parskip}{-3pt} \item
We present comprehensive experimental results using several real trajectory data sets, and we
explain the advantages and disadvantages of each proposed method.
\end{itemize}
The remainder of this paper is organized as follows: In Section~\ref{sec:related}, we discuss
previous methods related to the convoy query. We formulate the focal problem of this paper in
Section~\ref{sec:prob}. A modified method of moving cluster for the convoy discovery is shown in
Section~\ref{sec:cmc}. We propose more efficient methods based on trajectory simplification in
Sections~\ref{sec:cuts} and \ref{sec:ext}. Section~\ref{sec:exp} reports the results of
experimental performance comparisons, followed by conclusions in Section~\ref{sec:conc}.
\section{Problem Definition} \label{sec:prob}
This section formalizes the convoy problem. We start with the definitions of distances for points,
line segments, and bounding boxes :
\begin{mydefinition}
\textbf{(Distance Functions)}\\
\vspace{-0.4cm}
\begin{itemize}
\setlength{\parskip}{-2pt} \item Given two points $p_u$ and $p_v$,
$D(p_u,p_v)$ is defined as the Euclidean distance between $p_u$ and
$p_v$.
\setlength{\parskip}{-2pt} \item Given a point $p$ and a line segment $l$, $D_{PL}(p,l)$ is defined
as the shortest (Euclidean) distance between $p$ and any point on
$l$.
\setlength{\parskip}{-2pt} \item Given two line segments $l_u$ and $l_v$, $D_{LL}(l_u,l_v)$ is
defined as the shortest (Euclidean) distance between any two points
on $l_u$ and $l_v$, respectively.
\setlength{\parskip}{-2pt} \item With $\mathcal{B}_u$ and $\mathcal{B}_v$ being boxes then
$D_{min}(\mathcal{B}_u,\mathcal{B}_v)$ is defined as the minimum
distance between any pair of points belonging to each of the two
boxes.
\end{itemize}
\end{mydefinition}
The boxes introduced in the definition will be used for the bounding of line segments.
Next, the time domain is defined as the ordered set $\{t_1, t_2, \cdots, t_T\}$, where $t_j$ is a
time point and $T$ is the total number of time points.
In our problem setting, we consider a practical trajectory database model. We assume each
trajectory may have a different length from others and may also appear or disappear at any time in
$T$. In addition, each location of a trajectory can be sampled either regularly (e.g., every
second) or irregularly (i.e, some missing time points from $T$ may exist between two consecutive
time points of the trajectory).
The trajectory of an object $o$ is represented by a polyline that is given as a sequence of
timestamped locations $o=\langle p_a,p_{a+1}, \cdots, p_b \rangle$, where $p_j = (x_j,y_j,t_j)$
indicates the location of $o$ at time $t_j$, with $t_a$ being the start time and $t_b$ being the
end time. The time interval of $o$ is $o.\tau = [t_a,t_b]$. A shorthand notation is to use $o(t_j)$
for referring to the location of $o$ at time $t_j$ (i.e., location $p_j$).
Figure~\ref{fig_convoy} illustrates the polylines representing the
trajectories of three objects $o_1, o_2$, and $o_3$, during the time
interval from $t_1$ to $t_4$.
\begin{figure}[hbt]
\centering
\includegraphics[width=5.7cm]{pics/fig_convoy.eps}
\caption{An Example of a Convoy}
\label{fig_convoy}
\end{figure}
As a precursor to defining the convoy query, we need to understand
the notion of density connection~\cite{s-clustering3}. Given a
distance threshold $e$ and a set of points $S$, the $e$-neighborhood
of a point $p$ is given as $\mathit{NH}_e(p) = \{q \in S \; | \;
D(p, q) \leq e$\}.
Then, given a distance threshold $e$ and an integer $m$, a point $p$ is {\em directly
density--reachable} from a point $q$ if $p \in \mathit{NH}_e(q)$ and $|\mathit{NH}_e(q)| \geq m$.
A point $p$ is said to be {\em density--reachable} from a point $q$ with respect to $e$ and $m$ if
there exists a chain of points $p_1$, $p_2$, ..., $p_n$ in set $S$ such that $p_1 = q$, $p_n = p$,
and $p_{i+1}$ is directly density--reachable from $p_i$.
\begin{mydefinition} \label{def:dcon} \textbf{(Density--Connected)}
Given a set of points $S$, a point $p \in S$ is density--connected to
a point $q \in S$ with respect to $e$ and $m$ if there exists a
point $x \in S$ such that both $p$ and $q$ are density--reachable
from $x$.
\end{mydefinition}
The definition of density--connection permits us to capture a group of ``connected'' points with
arbitrary shape and extent, and thus to overcome the the lossy-flock problem shown in Figure
\ref{fig:lossyflock}. By considering density--connected objects for consecutive time points, we
define the {\em convoy query} as follows:
\begin{mydefinition}
\textbf{(Convoy Query)} Given a set of trajectories of $N$ objects,
a distance threshold $e$, an integer $m$, and an integer lifetime
$k$, the convoy query returns all possible groups of objects, so
that each group consists of a (maximal) set of density-connected
objects with respect to $e$ and $m$ during at least $k$ consecutive
time points.
\end{mydefinition}
Consider the convoy query with the parameters $m=2$ and $k=3$ issued over the trajectories in
Figure~\ref{fig_convoy}. $\langle{o_2,o_3},[t_1, t_3]\rangle$ is the result, meaning that $o_2$ and
$o_3$ belong to the same convoy during consecutive time points from $t_1$ to $t_3$.
Table~\ref{tab:summary-notation} offers the notations introduced in this section and to be used
throughout the paper.
\begin{table} [hbt]
\centering
\small
\begin{tabular}{|c|c|}
\hline
\textbf{Symbol} & \textbf{Meaning} \\ \hline
$p$ & Point/location (in the spatial domain) \\ \hline
$t$ & Time point \\ \hline
$o_i$ & Original trajectory of an object \\ \hline
$o_i(t)$ & Location of $o_i$ at time $t$ \\ \hline
$o'_i$ & Simplified trajectory (of $o_i$) \\ \hline
$l'_i$ & Line segment of $o'_i$ \\ \hline
$o'_i.\tau$ & Time interval of $o'_i$ \\ \hline
$l'_i.\tau$ & Time interval of $l'_i$ \\ \hline
$D(p_u,p_v)$ & Euclidean distance between points \\ \hline
$D_{PL}(p,l)$ & The shortest distance from point to line segment \\ \hline
$D_{LL}(l_u,l_v)$ & The shortest distance between line segments \\ \hline
$\mathcal{B}(l)$ & The minimum bounding box of $l$ \\ \hline
$D_{min}(\mathcal{B}_u,\mathcal{B}_v)$ & The minimum distance between two boxes \\ \hline
\end{tabular}
\caption{Summary of Notation}\label{tab:summary-notation}
\end{table}
\section{Related Work} \label{sec:related}
We first review existing work on trajectory clustering and, then cover
trajectory simplification, which is an important aspect of our
techniques for convoy discovery. We end by considering spatio-temporal
joins and distance measures for trajectories.
\subsection{Clustering over Trajectories} \label{sec:trjclust}
Given a set of points, the goal of {\em spatial clustering} is to
form clusters (i.e., groups) such that (i) points within
the same cluster are close to each other, and (ii) points from
different clusters are far apart.
In the context of trajectories, the locations of trajectories can be clustered at chosen time
points. Consider the trajectories in Figure~\ref{fig_lossymc}(\emph{a}). We first obtain a cluster
$c_1$ at time $t=1$, then a cluster $c_2$ at $t=2$, and eventually a cluster $c_3$ at $t=3$.
Kalnis et al. propose the notion of a {\em moving cluster} \cite{st-clustering1}, which is a
sequence of spatial clusters appearing during consecutive time points, such that the portion of
common objects in any two consecutive clusters is not below a given threshold parameter $\theta$,
i.e., $\frac{ |c_t \cap c_{t+1}| }{ |c_t \cup
c_{t+1}| } \ge \theta$, where $c_t$ denotes a cluster at time $t$.
There is a significant difference between a convoy and a moving
cluster.
For instance, in Figure~\ref{fig_lossymc}(\emph{a}), $o_2,o_3$, and $o_4$ form a convoy with 3
objects during 3 consecutive time points. On the other hand, if we set $\theta=1$ (i.e., require
100\% overlapping clusters), the overlap between $c_1$ and $c_2$ is only $\frac{3}{4}$, and the
above objects will not be discovered as a moving cluster. Next, in
Figure~\ref{fig_lossymc}(\emph{b}), if we set $\theta = \frac{1}{2}$ then $c_1$, $c_2$, and $c_3$
become a moving cluster. However, this is not a convoy.
\begin{figure}[hbt]
\centering
\includegraphics[width=6.5cm]{pics/fig_lossymc.eps}
\caption{Convoys Versus Moving Clusters}
\label{fig_lossymc}
\end{figure}
Spiliopoulou et al.~\cite{st-clustering9} study transitions in moving clusters (e.g., disappearance
and splitting) between consecutive time points. As transitions are based on the consideration of
common objects at consecutive time points, their techniques do not support convoy discovery either.
Next, Li et al.~\cite{st-clustering2} study the notion of \emph{moving micro cluster}, which is a
group of objects that are not only close to one another at the current time, but are also expected
to move together in the near future.
Recently, Lee et al.~\cite{TRACLUS} have proposed to partition trajectories into line segments and
build groups of close segments. This proposal does not consider the temporal aspects of the
trajectories. As a result, some objects can belong to the same group even though they have never
traveled close together (at the same time). Most recently, Jensen et al.~\cite{st-clustering4} have
proposed techniques for maintaining clusters of moving objects. They consider the clustering of the
current and near-future positions, while we consider past trajectories.
As mentioned earlier, several slightly different notions of a \emph{flock}
\cite{st-flock3,st-flock2} relate to that of a convoy. The notion most relevant to our study
defines a flock as a group of at least $m$ objects staying together within a circular region of
radius $e$ during a specified time interval \cite{st-flock4,st-flock3}. Al-Naymat et
al.~\cite{st-flock4} apply random projection to reduce the dimensionality of the data and thus
obtain better performance. Gudmundsson et al.~\cite{st-flock3} propose approximation techniques and
exploit an index to accelerate the computation of flock. It is also shown that the discovery of the
longest-duration flock is an NP-hard problem. It is worth noticing that these studies exhibit the
lossy-flock problem identified in Section~\ref{sec:intro}.
\subsection{Trajectory Simplification} \label{sec:trjsimple}
A trajectory is often represented as a {\em polyline}, which is a sequence of connected line
segments. Line simplification techniques have been proposed to simplify polylines according to some
user-specified resolution \cite{DP,DP-speed}.
The Douglas-Peucker algorithm (DP) \cite{DP} is a well-known and efficient method among the line
simplification techniques. Given a polyline specified by a sequence of $T$ points $\langle p_1,
p_2, \cdots, p_T \rangle$ and a distance threshold $\delta$, the goal is to derive a new polyline
with fewer points while deviating from the original polyline by at most $\delta$. The DP algorithm
initially constructs the line segment $\overline {p_1 p_T}$. It then identifies the point $p_i$
farthest from the line. If this point's (perpendicular) distance to the line is within $\delta$
then DP returns $\overline {p_1 p_T}$ and terminates. Otherwise, the line is decomposed at $p_i$,
and DP is applied recursively to the sub-polylines $\langle p_1, p_2, \cdots, p_i \rangle$ and
$\langle p_i,\cdots, p_T \rangle$.
As the worst-case time complexity of this algorithm is $O(T^2)$,
Hershberger et al.~\cite{DP-speed} show a faster version of this method with time complexity of
$O(T \cdot \log T)$. However, it is assumed that an object's trajectory cannot intersect itself,
which is not a valid assumption for the data we consider.
The DP technique deals with line simplification only in the spatial domain, ignoring the time
domain of the trajectories. Consider the example in Figure~\ref{fig_stsimpleFT}(\emph{a}).
Since the distance from $p_2$ to $\overline {p_1p_3}$ is within $\delta$, the DP algorithm omits
$p_2$ and simply returns $\overline {p_1p_3}$. Similarly, $q_2$ is also omitted and the polygon is
simplified to $\overline {q_1q_3}$.
In contrast, Meratnia et al.~\cite{st-simple1} take into account the temporal aspects in line
simplification. Figure~\ref{fig_stsimpleFT}(\emph{b}) exemplifies the working procedure of their
algorithm (say, DP*). First, DP* derives the point $p'_2$ on the line $\overline {p_1p_3}$ by
calculating the ratio of $p_2$'s time between $t$=1 of $p_1$ and $t$=3 of $p_3$. Then, it measures
the distance $D(p_2,p'_2)$ between $p_2$ and $p'_2$, instead of the perpendicular distance from
$p_2$ to $\overline {p_1p_3}$. Since $D(p_2,p'_2)> \delta$, $p_2$ is still kept after the
simplification, while it was removed by using DP in Figure~\ref{fig_stsimpleFT}(\emph{a}).
\begin{figure}[hbt]
\centering
\includegraphics[width=8.5cm]{pics/fig_dp.eps}
\caption{Comparison of Different Trajectory Simplifications}
\label{fig_stsimpleFT}
\end{figure}
\subsection{Distance Measures and Joins} \label{sec:stjoin}
A basic way of measuring the distance between two trajectories used in the literature is to compute
the sum of their Euclidean distances over time points. Such a distance measure may not be able to
capture the inherent distance between trajectories because it does not take into account particular
features of trajectories (e.g., noise, time distortion). Thus, it is important to devise a distance
function that ``understands'' the characteristics of trajectories.
A well-known approach is Dynamic Time Warping (DTW) \cite{DTW}, which applies dynamic programming
for aligning two trajectories in such a way that their overall distance is minimized.
More recent proposals for trajectory distance functions include
Longest Common Subsequence (LCSS)~\cite{LCSS}, Edit Distance on Real
Sequence (EDR)~\cite{EDR}, and Edit distance with Real Penalty
(ERP)~\cite{ERP}.
Lee et al.~\cite{TRACLUS} point out that the above distance
measures capture the global similarity between two trajectories, but
not their local similarity during a short time interval. Thus, these
measures cannot be applied in a simple manner for convoy discovery.
Given two data sets $P_1$ and $P_2$, spatio-temporal joins find pairs
of elements from the two sets that satisfy predicates with both
spatial and temporal attributes~\cite{st-join2}. The {\em close-pair
join} reports all object pairs ($o_1$, $o_2$) from $P_1 \times P_2$
with distance $D_{\tau}(o_1, o_2)\leq e$ within a time interval $\tau$
being bounded by a user-specified distance $e$. Plane-sweep techniques
\cite{st-join6,st-join5} have been proposed for evaluating
spatio-temporal joins. Like the close-pair join, the {\em trajectory
join} \cite{st-join4} aims at retrieving all pairs of
similar trajectories between two datasets. Bakalov et al.~\cite{st-join4} represent trajectories as
sequences of symbols and apply sliding window techniques to measure the symbolic distance between
possible pairs. These studies consider pairs of objects, whereas we consider sets of objects.
|
\section{Introduction}
Recently, the realization of ultracold polar molecular gases has been
regarded as one of the most promising research directions in the field of
atomic and molecular physics \cite{Doyle04,Summary}. Ultracold polar
molecules, with their long-range and anisotropic dipole-dipole interactions
\cite{Santos00,Syi00}, have attracted much attention in a variety of
research areas, such as quantum information science \cite{DeMille02}-\cite%
{Yelin06} and precision measurement \cite{Kozlov95}-\cite{DeMille08}.
There are two typical routes to achieve quantum degenerate gases of
molecules. One is through the direct cooling of molecules, which is hard to
achieve due to the complex internal levels of molecules \cite{Doyle98}. The
alternative one is to couple a pair of degenerate atoms by photoassociation
(PA) \cite{Jones06} or Feshbach resonance (FR) \cite{Thorsten06}. However
the diatomic molecule formed by a PA or FR process is usually loosely bound
and energetically unstable. They have to be adiabatically transferred into a
tightly bound ground state via a stimulated Raman adiabatic passage (STIRAP)
\cite{Bergmann98}. The success of the STIRAP is based on a coherent
population trapping (CPT) state, which is accomplished by a pair of pulses
in a counterintuitive sequence. During the adiabatic transfer, the system
can follow the superposition between initial and final states, preventing
any incoherent losses involving the middle unstable levels. Thereby, the
high phase-space density of the initial gas can be coherently preserved.
Currently, intensive experimental efforts to obtain quantum degenerate gases
of molecules have been made by combining FR with STIRAP, which serves as an
effective way to produce molecules in lower vibrational levels \cite%
{Winkler07}-\cite{Aikawa09}. However, due to the strong vibrational
quenching, a severe particle loss appears near the FR threshold. One way to
solve this problem is to apply the optical lattice technique \cite%
{Winkler07,Lang08}, in which inelastic collisions between molecules are well
suppressed by preparing one single molecule per lattice site. Alternatively,
all-optical transfer of molecules toward quantum degeneracy using a
"two-color PA" method has been demonstrated experimentally \cite{Wynar00}-%
\cite{Sage05}, where the excited molecules are moved down by a coherent dump
field, instead of by spontaneous decay \cite{Kerman04}-\cite{Deiglmayr08}.
However, one common bottleneck with PAs is the small free-bound
Franck-Condon factor (FCF), which requires an intense PA power to achieve an
efficient adiabatic transfer \cite{Vardi97,Vardi99}. To date, the most
promising way to overcome this PA weakness is by the FR-assisted PA scheme
proposed first by Verhaar \textit{et. al. }\cite{Abeelen98,Courteille98} and
verified by many groups later on \cite{Tolra03}-\cite{Kuznetsova09}. In
terms of these studies, if a Feshbach quasi-bound state is adjusted close to
the continuum, the atomic scattering wavefunction, acquiring some
bound-state properties, becomes more localized. This gives rise to a
dramatic enhancement of the free-bound FCF. As a result, the PA intensity
required for a given atom-molecule transfer efficiency can be greatly
reduced, compared with the case without the assistance of FR.
In the present work, we propose an all-optical scheme to achieve a high
transfer efficiency of atoms into molecules with a low PA power. For the
purpose, we consider a "$R$-type" atom-molecule conversion model (see figure %
\ref{model} in solid arrows) through a photoassociative STIRAP procedure.
Such a model is similar to the "$R$-transfer" suggested by Nikolov \textit{%
et. al. }\cite{Nikolov00} as well as to the work by Band and Julienne \cite%
{Band95}. In their works, molecules with an upper high-lying state are
generated first through a step-wise PA excitation from free atoms (e.g. $%
\left\vert 0_{1,2}\right\rangle \rightarrow \left\vert m\right\rangle
\rightarrow \left\vert e\right\rangle $ in figure \ref{model}), followed by
a radiative decay to populate a series of ground manifolds (e.g. $\left\vert
e\right\rangle \rightarrow \left\vert g\right\rangle $ in figure \ref{model}%
). In this paper, we apply a coherent dump field to make the transition from
state $\left\vert e\right\rangle $ to state $\left\vert g\right\rangle $,
instead of by spontaneous emission. This leads to an accessible
atom-molecule adiabatic passage between the initial ($\left\vert
0_{1,2}\right\rangle $) and target states ($\left\vert g\right\rangle $).
Compared with a conventional "$\Lambda $-type" model (see figure \ref{model},
dash-dotted arrows), the CPT state supported in the "$R$-type" scheme has
been perturbed by a newly embedded state $\left\vert m\right\rangle $, which
is absent in previous STIRAPs. As we will show, state $\left\vert
m\right\rangle $ can help to reduce the power in PA field, whose stability
properties will play a significant role in the molecular production. Under a
simple numerical comparison, we have identified that the PA power required
in the "$\Lambda $-type" model must be much higher than that in the "$R$%
-type" model for achieving the same final efficiency.
This paper is organized as follows. In section 2, after briefly reviewing a
similar idea of FR-induced STIRAP, we come up with our photoassociative
STIRAP model and develop the underlying mean-field equations for the studies
of a quasi-CPT description. In section 3, a generalized adiabatic theorem
involving all the Bogoliubov collective modes is introduced to evaluate the
adiabatic condition and quasi-CPT lifetime in our scheme. In section 4,
numerical simulations for the cases described in section 2 and 3 are
implemented by using practical parameters. The laser profiles applied in the
calculations are optimized according to adiabatic condition (section 4.1)
and other relevant assumptions (section 4.2). Finally, a summary is given in
section 5.
\section{Model and dark state theory}
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=2.45in, width=2.5851in
]%
{model.eps}%
\caption{(color online) Schematic diagram: "$R$-type" transfer in
solid arrows comprising the transitons of $\left\vert
0_{1,2}\right\rangle \rightarrow \left\vert m\right\rangle
\rightarrow\left\vert e\right\rangle \rightarrow \left\vert
g\right\rangle $ with the corresponding coupling fields
$\Omega_{PA}$, $\Omega_{p}$ and $\Omega_{d}$, respectively;
"$\Lambda$-type" transfer in dash-dotted arrows with $\left\vert
0_{1,2}\right\rangle \rightarrow\left\vert e\right\rangle
\rightarrow\left\vert g\right\rangle $ transitions characterized by
$\Omega_{PA}^{\prime}$ and $\Omega_{d}$. All the
other parameters are described in the text.}%
\label{model}%
\end{center}
\end{figure}
Before moving to concrete illustrations of the model, we briefly review the
idea of the FR-induced STIRAP method \cite{Kokkelmans01,Mackie02}. In a
typical FR-induced STIRAP, a number of colliding atoms undergo a strong
association into quasi-bound molecules when a magnetic field is swept close
to or across the FR, characterized by coupling strength $\alpha $ and
binding energy $\varepsilon $. Subsequently, these quasi-bound molecules are
further transferred into stable molecules via a STIRAP. Clearly, here $%
\alpha $ and $\varepsilon $ correspond to $\Omega _{PA}$ and $\delta _{m}$
(see figure \ref{model}) in our scheme, respectively. A major advantage of
our scheme is that $\Omega _{PA}$ and $\delta _{m}$ can be manipulated more
conveniently than $\alpha $ and $\varepsilon $ over the time scales. Because
the latter quantities are highly dependent on the atomic intrinsic
properties, especially the coupling strength $\alpha $, which is fixed by the
hyperfine interaction and is hence independent of time, $\varepsilon $ is
experimentally tunable via an external magnetic field, while the former
quantities, being controllable by optical means, are easily selected.
Turning to our scheme, as depicted in figure \ref{model}, we study a "$R$%
-type" five-level atom-polar-molecule formation. Two species of free atoms
prepared in $\left\vert 0_{1}\right\rangle $ and $\left\vert
0_{2}\right\rangle $ states are first coupled into molecules in an
intermediate high-lying state $\left\vert m\right\rangle $ with Rabi
frequency $\Omega _{PA}$ and detuning $\delta_{m}$. Simultaneously, a pair
of pump-dump lasers are applied to move these loosely bound molecules in $%
\left\vert m\right\rangle $ down to the lowest molecular ground state $%
\left\vert g\right\rangle $, where $\Omega_{p}$, $\Omega_{d}$ stand for
coupling strengths and $\delta_{e}$, $\delta_{g}$ for the two- and
three-photon detunings. This scheme has several attractive properties.
Firstly, the presence of state $\left\vert m\right\rangle $ brings one extra
bound-bound transition from state $\left\vert m\right\rangle $ to state $%
\left\vert e\right\rangle $; hence it becomes easier for the PA field to
associate atoms into the $\left\vert m\right\rangle $, rather than a higher $%
\left\vert e\right\rangle $ state. Secondly, transitions $\left\vert
m\right\rangle \rightarrow\left\vert e\right\rangle $ and $\left\vert
e\right\rangle \rightarrow\left\vert g\right\rangle $ are preferred because
of the favorable bound-bound FCFs. Meanwhile, the $\left\vert
0_{1,2}\right\rangle \rightarrow\left\vert m\right\rangle $ transition is
also accessible by an optimal control of the PA field in the time domain.
As usual, we start our discussions with a set of coupled Gross-Pitaevskii's
equations. In the mean-field treatment, where every quantum field operator $%
\hat{\Psi}_{i}$ has been replaced by its normalized amplitude $\psi_{i}$
\cite{Heinzen00}, this yields%
\begin{eqnarray}
i\dot{\psi}_{0_{1}} &
=-\frac{\Omega_{PA}}{2}\psi_{0_{2}}^{\ast}\psi
_{m},\label{dy1}\\
i\dot{\psi}_{0_{2}} &
=-\frac{\Omega_{PA}}{2}\psi_{0_{1}}^{\ast}\psi
_{m},\label{dy2}\\
i\dot{\psi}_{m} & =-\left( \delta_{m}+i\gamma_{m}\right) \psi_{m}%
-\frac{\Omega_{PA}}{2}\psi_{0_{1}}\psi_{0_{2}}-\frac{\Omega_{p}}{2}\psi
_{e},\label{dy3}\\
i\dot{\psi}_{e} & =-\left( \delta_{e}+i\gamma_{e}\right) \psi_{e}%
-\frac{\Omega_{p}}{2}\psi_{m}-\frac{\Omega_{d}}{2}\psi_{g},\label{dy4}\\
i\dot{\psi}_{g} & =-\delta_{g}\psi_{g}-\frac{\Omega_{d}}{2}\psi
_{e},\label{dy5}%
\end{eqnarray}
where $\gamma_{i}$ (i=m,g) is introduced phenomenologically to
describe the spontaneous decay of the $\left\vert i\right\rangle $ state
to other undetected states, and it is possible to find a relatively
stable $\left\vert m\right\rangle $ state with its decay rate
$\gamma_{m}\ll\gamma_{e}$ \cite{Band95,Napolitano94}. The initial
and target states are assumed to be sufficiently stable with
$\gamma_{0_{1(2)},g}\equiv0$. For an easy analysis without loss of
the main physics, inter- and intra-species collisions have been ignored
under typical parameters \cite{Kuznetsova08}. After a global gauge
transformation, we can safely consider all the Rabi frequencies to
be real positive values without loss of generality.
A CPT state is always expected to move all the population into a
target state as long as the adiabatic condition holds. In order to
derive the corresponding adiabatic parameter, we first search for
the CPT distributions for the
following assumptions:%
\begin{equation}
\psi_{0_{1,2}}=\phi_{0}e^{-i\mu
t},\psi_{e}=0,\psi_{m,g}=\phi_{m,g}e^{-2i\mu
t}. \label{ansatz}%
\end{equation}
Here $\phi_{i}$ is a steady-state amplitude, we consider $\phi_{0_{1}}%
=\phi_{0_{2}}=\phi_{0}$ for a balanced system, and $\mu$ is the
atomic chemical potential. By ignoring all the decays and inserting
equation (\ref{ansatz}) into equations (\ref{dy1})-(\ref{dy5}) with
particle number conservation:
$2(\phi_{0}^{2}+\phi_{m}^{2}+\phi_{g}^{2})=1$, a generalized
three-photon
resonance is given by%
\begin{equation}
\delta_{g\pm}=-2\mu_{\pm}=\frac{-\Omega_{PA}^{2}/2}{\delta_{m}\pm\sqrt
{\delta_{m}^{2}+\Omega_{PA}^{2}\left( 3+\chi^{2}\right) /2}},
\label{delta_g}%
\end{equation}
leading to the following CPT descriptions with $\phi_{e}=0$:%
\begin{eqnarray}
\phi_{0} & =\sqrt{\frac{1}{2}-\phi_{m}^{2}\left( 1+\chi^{2}\right) },\label{CPT1}\\
\phi_{m} & =-\frac{\bar{\Omega}_{PA}/2}{1+\sqrt{1+\bar{\Omega}_{PA}%
^{2}\left( 3+\chi^{2}\right) /2}},\label{CPT2}\\
\phi_{g} & =-\chi\phi_{m}. \label{CPT3}%
\end{eqnarray}
where $\bar{\Omega}_{PA}=\Omega_{PA}/\delta_{m}$, $\chi=\Omega_{p}/\Omega_{d}%
$. In equation (\ref{delta_g}), the choice of $\delta_{g}$ is
determined by $\delta_{m}$. If $\delta_{m}>0$,
$\delta_{g}=\delta_{g+}$ and $\mu=\mu_{+}$, whereas if $\delta_{m}<0$,
$\delta_{g}=\delta_{g-}$ and $\mu=\mu_{-}$. From equations
(\ref{CPT1})-(\ref{CPT3}), we note that when $\left\vert
\bar{\Omega }_{PA}\right\vert $ and $\chi$ both change from 0 to
large positive values, population initially prepared in states
$\left\vert 0_{1,2}\right\rangle $ will be gradually converted into
molecules in state $\left\vert g\right\rangle $ under three-photon
resonance [equation (\ref{delta_g})]. Also it is worth emphasizing
that such a CPT state has been perturbed since
$\phi_{m}\neq0$, and is called a "quasi-CPT" state. In the limit
of $\bar{\Omega}_{PA}\ll1$, population in state $\left\vert
m\right\rangle $ is virtually empty, we find that a complete
transfer is still possible as long as $\chi$ varies from 0 to
$\infty$. In other words, the change by $\bar{\Omega}_{PA}$ has been
accomplished by varying $\Omega_{p}$; thus the existence of state
$\left\vert m\right\rangle $ is\ quite helpful for a relatively
small $\bar{\Omega}_{PA}$ value.
Actually, in the dynamics, if we use a strong PA laser to trigger
the $\left\vert 0_{1,2}\right\rangle \rightarrow\left\vert
m\right\rangle $ transition, particle accumulations in state
$\left\vert m\right\rangle $ will inevitably arise. Therefore,
in order to avoid a considerable loss from state $\left\vert
m\right\rangle $, pulse durations in STIRAP must be much shorter
than $\left\vert m\right\rangle $ state's lifetime. On the other
hand, if we deeply reduce the $\Omega_{PA}$ value, the population in the
$\left\vert m\right\rangle $ state will greatly be suppressed;
meanwhile, a large fraction of atoms are left in the continuum,
unpaired, because of a poor atom-molecule coupling strength. This
conflict can be generalized to the properties of a quasi-CPT state,
in which case one may prefer the use of moderate PA power.
Results in equations (\ref{CPT1})-(\ref{CPT3}) are for the case of
$\delta_{m}\neq0$. If $\delta_{m}=0$, i.e. the PA laser is exactly
resonant with the free-bound transition, then equation
(\ref{delta_g}) is reduced to $\delta_{g\pm}=\pm
\Omega_{PA}/(6+2\chi^{2})^{1/2}$\ with the following CPT solutions:%
\begin{equation}
\phi_{0}^{2}=2\phi_{m}^{2}=2\phi_{g}^{2}/\chi^{2}=\left(
3+\chi^{2}\right)
^{-1}. \label{CPT_re}%
\end{equation}
Equation (\ref{CPT_re}) shows a constant population ratio
between states $\left\vert 0_{1,2}\right\rangle $ and $\left\vert
m\right\rangle $, i.e. $\phi_{0}^{2}/\phi_{m}^{2}=2$. This equality
contrasts with the standard CPT evolution, especially at t=0,
which implies a poor transfer efficiency at $\delta_{m}=0$. As a
result, a nonzero $\delta_{m}$ value is favored in our
consideration.
\section{Adiabatic Theorem}
To derive the adiabatic parameter for the quasi-CPT state, we adopt
a standard linearized approach as in \cite{Pu07,Jing08} by
adding a small fluctuation $\delta\psi_{i}$ to the
instantaneous steady-state solution $\phi_{i}$,%
\begin{equation}
\psi _{0_{1,2}}=\left( \phi _{0}+\delta \psi _{0_{1,2}}\right)
q\left( t\right) ,\psi _{e}=\delta \psi _{e}q^{2}\left( t\right)
,\psi _{m,g}=\left(
\phi _{m,g}+\delta \psi _{m,g}\right) q^{2}\left( t\right) \label{linearizeda}%
\end{equation}
where $q\left( t\right) =\exp[-\int_{0}^{t}\mu\left(
t^{\prime}\right) dt^{\prime}]$, and $\mu(t)$ is a time-dependent
chemical potential given by $\mu\left( t\right) =\mu_{+\left(
-\right) }$ (see equation (\ref{delta_g})). Substituting equation
(\ref{linearizeda}) into the mean-field equations
(\ref{dy1})-(\ref{dy5}) with the help of CPT descriptions, we
eventually arrive at a set of linearized equations for the vector
$\mathbf{\delta}\mathbf{\psi=}\left[ \delta
\psi_{0_{1}},\delta\psi_{0_{2}},\delta\psi_{m},\delta\psi_{e},\delta\psi
_{g}\right] ^{T}$ with its conjugate vector
$\mathbf{\delta}\mathbf{\psi
}^{\ast}$%
\begin{equation}
\mathbf{\delta\dot{\Psi}=-}i\mathbf{M\delta\Psi-\Gamma\delta\Psi-\dot{\Phi}},
\label{linearized}%
\end{equation}
where
\[
\mathbf{M=}\left(
\begin{array}
[c]{cc}%
\mathbf{A} & \mathbf{B}\\
-\mathbf{B} & -\mathbf{A}%
\end{array}
\right) ,\mathbf{\Gamma=}\left(
\begin{array}
[c]{cc}%
\mathbf{\gamma} & 0\\
0 & \mathbf{\gamma}%
\end{array}
\right) ,
\]
and
\begin{eqnarray}
\mathbf{A} & =-\frac{1}{2}\left(
\begin{array}
[c]{ccccc}%
2\mu & 0 & \Omega_{PA}\phi_{0} & 0 & 0\\
0 & 2\mu & \Omega_{PA}\phi_{0} & 0 & 0\\
\Omega_{PA}\phi_{0} & \Omega_{PA}\phi_{0} & 2\left(
\delta_{m}+2\mu\right) &
\Omega_{p} & 0\\
0 & 0 & \Omega_{p} & 0 & \Omega_{d}\\
0 & 0 & 0 & \Omega_{d} & 0
\end{array}
\right) ,\\
\mathbf{B} & =-\frac{\Omega_{PA}\phi_{m}}{2}\left(
\begin{array}
[c]{ccccc}%
0 & 1 & 0 & 0 & 0\\
1 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0\\
0 & 0 & 0 & 0 & 0
\end{array}
\right).
\end{eqnarray}
In equation (\ref{linearized}), some notations are
$\mathbf{\delta\Psi=}\left[
\mathbf{\delta}\mathbf{\psi,\delta}\mathbf{\psi}^{\ast}\right]
^{T}$, $\mathbf{\dot{\Phi}=}\left[
\mathbf{\dot{\phi},\dot{\phi}}\right] ^{T}$ with
$\mathbf{\dot{\phi}=}\left[
\dot{\phi}_{0},\dot{\phi}_{0},\gamma_{m}\phi
_{m}+\dot{\phi}_{m},0,\dot{\phi}_{g}\right] ^{T}$.
$\mathbf{\gamma}$ is a $5\times5$ matrix with
$\gamma_{33}=\gamma_{m}$ and $\gamma_{44}=\gamma_{e}$ being the only
nonzero elements. In addition, we have assumed detunings
$\delta_{e,g}=-2\mu(t)$ to be chirped \cite{Koenig04}.
Furthermore, we introduce a generalized Bogoliubov-de Gennes (BdG)
equation for
matrix $\mathbf{M}$,%
\begin{equation}
\mathbf{M}\left( t\right) \mathbf{w}_{i}\left( t\right)
=\omega_{i}\left(
t\right) \mathbf{w}_{i}\left( t\right) , \label{BdG}%
\end{equation}
where $\omega_{i}$ and $\mathbf{w}_{i}=\left[ \mathbf{u}_{i},\mathbf{v}%
_{i}\right] ^{T}$ are the well-defined $i$th eigenenergy and
eigenvector,
respectively. $\mathbf{u}_{i}$ and $\mathbf{v}_{i}$ contain familiar
Bogoliubov
$u$-$v$ parameters for each species%
\begin{equation}
\mathbf{u}\left( \mathbf{v}\right) _{i}=[u\left( v\right) _{i0_{1}%
},u\left( v\right) _{i0_{2}},u\left( v\right) _{im},u\left(
v\right)
_{ie},u\left( v\right) _{ig}]^{T}. \label{uv}%
\end{equation}
From the BdG equation, taking into account the special structure of matrix \textbf{M},
one can show the quantities $\omega_{i}^{2}$ are the eigenenergies of the matrix $(\mathbf{A}%
+\mathbf{B})(\mathbf{A}-\mathbf{B})$, which can be obtained from the
following
equation:%
\begin{equation}
\left( \omega_{i}^{2}\right) ^{2}\left( \left(
\omega_{i}^{2}\right)
^{3}-a_{1}\left( \omega_{i}^{2}\right) ^{2}+a_{2}\left( \omega_{i}%
^{2}\right) -a_{3}\right) =0, \label{cubic}%
\end{equation}
where the coefficients $a_{i}$ are given as%
\begin{eqnarray}
a_{1} & =\left( \delta_{m}+2\mu\right) \left(
\delta_{m}+6\mu\right)
+\frac{\Omega_{p}^{2}+\Omega_{d}^{2}}{2},\\
a_{2} & =\frac{\left( \delta_{m}+2\mu\right) }{2}\left( 2\Omega_{p}%
^{2}\mu+\Omega_{d}^{2}\left( \delta_{m}+6\mu\right) \right)
+\frac{\left(
\Omega_{p}^{2}+\Omega_{d}^{2}\right) ^{2}}{16},\\
a_{3} & =\frac{\left( \delta_{m}+2\mu\right)
\Omega_{d}^{2}}{16}\left( 4\Omega_{p}^{2}\mu+\Omega_{d}^{2}\left(
\delta_{m}+6\mu\right) \right) .
\end{eqnarray}
Eigenenergies implied in equation (\ref{cubic}) comprise a doublet
0 mode $\omega_{0,1}=0$ and three pairs of excited modes $\left(
\omega _{j},-\omega_{j}^{\ast}\right) $ (j=2,3,4). We find that
$\omega_{j}$ is real and has to be determined by biorthonormal
relations for its corresponding eigenvector $\mathbf{w}_{j}$.
Detailed elucidations of biorthonormality have been published elsewhere
\cite{Ling07}. In addition, we realize that the dynamical
instability is impossible here due to the absence of collisions.
To accomplish the goal of deriving the adiabatic theorem, we have to
expand an arbitrary vector $\mathbf{\delta\Psi}$ in the
dressed-state picture with a complete set of eigenvectors. By
solving the BdG equation with 0 eigenenergies, we are able to
obtain $\mathbf{w}_{0}$ and $\mathbf{w}_{1}$ (unnormalized) explicitly using
Gram-Schmidt orthogonalization,
\begin{eqnarray}
\mathbf{w}_{0} & =\left( -1,1,0,0,0,1,-1,0,0,0\right) ^{T},\label{w0}\\
\mathbf{w}_{1} & =\left( \frac{\phi_{0}}{2\phi_{m}},\frac{\phi_{0}}%
{2\phi_{m}},1,0,-\chi,\frac{-\phi_{0}}{2\phi_{m}},\frac{-\phi_{0}}{2\phi_{m}%
},-1,0,\chi\right) ^{T}. \label{w1}%
\end{eqnarray}
Here $\mathbf{w}_{0}$, being a real dark state, is entirely
decoupled with other eigenmodes, because its source term
($\mathbf{w}_{0}^{T}\mathbf{\dot {\Phi}}$) and inter-coupling term
($\mathbf{w}_{0}^{T}\mathbf{\Gamma \mathbf{\eta}_{+}w}_{j}$) both
vanish (see equation (\ref{cj}) below for detailed notations),
whereas, $\mathbf{w}_{1}$ is most likely to be triggered through its
nonzero inter-coupling strength ($\mathbf{w}_{1}^{T}\mathbf{\Gamma
\mathbf{\eta}_{+}w}_{j}$), which can be roughly estimated by its decay rate%
\begin{equation}
\mathbf{w}_{1}^{T}\mathbf{\Gamma\eta}_{+}\mathbf{Q}=\frac{\gamma_{m}}{\phi
_{0}^{2}/2\phi_{m}^{2}+2(1+\chi^{2})}\equiv1/\tau_{cpt}, \label{cpt_lifetime}%
\end{equation}
where $\mathbf{Q}$ is a newly introduced vector complementary to
$\mathbf{w}_{1}$ with a well-defined normalization,%
\begin{equation}
\mathbf{w}_{1}^{T}\mathbf{\eta}_{+}\mathbf{Q}=1,\label{normal}%
\end{equation}
through the definition of%
\begin{equation}
\mathbf{MQ=w}_{1}/v.\label{v_d}%
\end{equation}
Here, $v$ is a coefficient to be determined, and $\mathbf{\eta}_{+}$(and
$\mathbf{\eta}_{-}$ below) are given in \cite{Ling07}.
Combining equation (\ref{normal}) with (\ref{v_d}), we find the
vector $\mathbf{Q}$ takes a
special form: $\mathbf{Q}=\mathbf{[}q_{0},q_{0},q_{m},q_{e},q_{g},q_{0}%
,q_{0},q_{m},q_{e},q_{g}\mathbf{]}^{T}$. Detailed expressions for
$q_{i}$ and $v$ are presented in the appendix.
The CPT lifetime $\tau_{cpt}$ defined in equation
(\ref{cpt_lifetime}) is clearly inversely-proportional to
$\gamma_{m}$ and $\phi_{m}$, which agrees with our intuitions. In
other words, the presence of state $\left\vert m\right\rangle $
actually gives rise to a finite lifetime for the quasi-CPT state.
Any pulse duration used in the system has to be much shorter than
$\tau_{cpt}$; otherwise, a big particle loss from state $\left\vert
m\right\rangle $ is unavoidable. One effective way to achieve a long
$\tau_{cpt}$ is to search for a relatively stable $\left\vert
m\right\rangle $ state with a small $\gamma_{m}$ value. Other
excited eigenenergies and eigenvectors are too complicated to
list here, but they can be conveniently derived from equation
(\ref{BdG}) with (\ref{cubic}).
Since other inter-coupling strengths for $\mathbf{w}_{1}$ are also
proportional to $\gamma_{m}$ as in equation (\ref{cpt_lifetime}) and
$\gamma_{m}$ is considered to be much smaller than $\gamma_{e}$, we
shall safely ignore the contributions from $\mathbf{w}_{1}$ and
expand $\mathbf{\delta\Psi}$ in the parameter space with the help of
three excited eigenmodes $\mathbf{w}_{j}$
(j=2,3,4) only, taking the form of%
\begin{equation}
\mathbf{\delta\Psi=}\sum_{j=2}^{4}\left( c_{j}\mathbf{\eta}_{+}\mathbf{w}%
_{j}-c_{j}^{\ast}\mathbf{\eta}_{-}\mathbf{w}_{j}^{\ast}\right). \label{DPHI}%
\end{equation}
Through inserting equation (\ref{DPHI}) into (\ref{linearized}), and
with the help of biorthonormality relations, finally, we obtain a
set of coupling
equations for $c_{j}\left( t\right) $,%
\begin{equation}
\dot{c}_{j}+i\omega_{j}c_{j}+\mathbf{\mathbf{w}}_{j}^{T}\mathbf{\Gamma
\delta\Psi}=\mathbf{-\mathbf{w}}_{j}^{T}\mathbf{\dot{\Phi}} \label{cj}%
\end{equation}
Terms like $\mathbf{\dot{w}}_{j}^{T}\mathbf{\delta\Psi}$ have been
eliminated in equation (\ref{cj}) because the eigenvector
$\mathbf{\mathbf{w}}_{j}$ changes very slowly in the adiabatic
limit. Generally speaking, if a system is said to operate in an
adiabatic regime, population in any excited mode (nonzero
eigen-mode) remains small. Hence, we shall introduce a typical
adiabatic parameter definition
\begin{equation}
r\left( t\right) =\sqrt{\frac{\left\vert c_{2}\right\vert
^{2}+\left\vert
c_{3}\right\vert ^{2}+\left\vert c_{4}\right\vert ^{2}}{3}}\ll1\label{r_value}%
\end{equation}
A reduction in the $r$-value means an increase in the adiabaticity;
in general, it can be accomplished by a longer pulse or a stronger
pump field. In the adiabatic regime, if a system evolves in a
CPT state, an entire population conversion is achievable. However,
the CPT lifetime implied in our model places a limitation for both
the pulse duration and PA intensity, leading to a slightly larger
$r$-value. This point will be discussed in section 4.1 toward the
goal of obtaining an optimal pulse duration and PA intensity for an
efficient transfer.
In equation (\ref{cj}), since $\dot{c}_{j}$ can be ignored
adiabatically, we further rewrite it as a series of linearized
coupling
equations:%
\begin{equation}
\left(
\begin{array}
[c]{cc}%
\mathbf{F}+i\mathbf{\omega} & \mathbf{G}\\
\mathbf{G} & \mathbf{F}-i\mathbf{\omega}%
\end{array}
\right) \left(
\begin{array}
[c]{c}%
\mathbf{c}\\
\mathbf{c}^{\ast}%
\end{array}
\right) =-\left(
\begin{array}
[c]{c}%
\dot{\Phi}_{w}\\
\dot{\Phi}_{w}%
\end{array}
\right) \label{coupled}%
\end{equation}
where%
\[
\mathbf{F=}\left(
\begin{array}
[c]{ccc}%
f_{22} & f_{23} & f_{24}\\
f_{23} & f_{33} & f_{34}\\
f_{24} & f_{34} & f_{44}%
\end{array}
\right) ,\mathbf{G=}\left(
\begin{array}
[c]{ccc}%
0 & g_{23} & g_{24}\\
-g_{23} & 0 & g_{34}\\
-g_{24} & -g_{34} & 0
\end{array}
\right) ,
\]
with the definitions of $\mathbf{c=}\left[ c_{2},c_{3},c_{4}\right]
^{T}$,
$\dot{\Phi}_{w}=[\mathbf{\mathbf{w}}_{2}^{T}\mathbf{\dot{\Phi},\mathbf{w}}%
_{3}^{T}\mathbf{\dot{\Phi},\mathbf{w}}_{4}^{T}\mathbf{\dot{\Phi}]}^{T}$,
$\mathbf{\omega}=\omega_{j}\mathbf{D}$ (j=2,3,4), $\mathbf{D}$ is a
$3\times3$ unit matrix, and
\begin{eqnarray}
f_{ij} & =\gamma_{m}\left( u_{im}u_{jm}-v_{im}v_{jm}\right)
+\gamma
_{e}\left( u_{ie}u_{je}-v_{ie}v_{je}\right) ,\label{f}\\
g_{ij} & =\gamma_{m}\left( v_{im}u_{jm}-u_{im}v_{jm}\right)
+\gamma
_{e}\left( v_{ie}u_{je}-u_{ie}v_{je}\right) ,\label{g}%
\end{eqnarray}
We solve $c_{j}(c_{j}^{\ast})$ values from equations (\ref{coupled})
numerically and insert them into equation (\ref{r_value}), a
time-dependent $r$-function is ultimately accessible. It needs to
be noted that all the $u(v)$s in equation
(\ref{f}) and (\ref{g}) have been normalized according to biorthogonality,%
\begin{equation}
\mathbf{w}_{i}^{T}\mathbf{\eta}_{+}\mathbf{w}_{j}=\delta_{ij},\mathbf{w}%
_{i}^{T}\mathbf{\eta}_{-}\mathbf{w}_{j}=0. \label{bio}%
\end{equation}
\section{Numerical Analysis}
In the following numerical calculations, we intended to achieve
a highly-efficient ground molecular production under an optimization
of all the optical fields, including $\Omega_{PA},$ $\Omega_{p}$ and
$\Omega_{d}$. From CPT descriptions [equations
(\ref{CPT1}-\ref{CPT3})], we adopt a common pair of
counterintuitive pump-dump pulses for $\left\vert m\right\rangle
$-$\left\vert
g\right\rangle $ transition with the same width $T$%
\begin{equation}
\Omega_{p,d}=\frac{\Omega_{p,d}^{0}}{2}\left( 1\pm\tanh\left(
\frac
{t-t_{p,d}}{T}\right) \right) \label{Rabi}%
\end{equation}
where $\Omega_{p,d}^{0}$, $t_{p,d}$ are for the peak Rabi
frequencies and central positions respectively. Based on equations
(\ref{CPT1})-(\ref{CPT3}), the PA Rabi frequency $\Omega_{PA}$,
which must start from 0, is considered to share the same profile as
$\Omega_{p}$ except for a different peak amplitude
$\Omega_{PA}^{0}$. Here, the detuning $\delta_{m}$ is fixed at a
finite value for simplicity.
\subsection{Optimal PA pulse}
In what follows, we seek to gain from the $r$-value in equation
(\ref{r_value}) insights into the parameters, especially for an
appropriate PA amplitude $\Omega_{PA}^{0}$ and a pulse duration $T$.
As we already understand, applying a longer pulse or a more intense PA
laser will lead to a lower $r$-value. If a system's adiabaticity
($r$-value) is kept in a low level, which means the system will
operate within the adiabatic regime, any excited modes are
greatly suppressed. In a pure-CPT environment, adiabaticity indeed
becomes a sufficient criterion for a complete transfer. However, in
our scheme, we observe in addition to adiabaticity, a long CPT lifetime
is another significant criterion for an efficient transfer.
In a dynamical process, the $r$-value obtained from equation
(\ref{r_value}) varies with time. We find that $r$-value estimated
at $t_{s}$ which is defined by
$\phi_{0}^{2}(t_{s})=2\phi_{g}^{2}(t_{s})$ turns out to be a good
estimate of the degree of adiabaticity. Thus,
$r_{s}$ and $\tau_{cpt}$ values displayed in figure \ref{r_dy}(a), (b) are both evaluated at $t=t_{s}$.%
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=3.5434in, width=2.8532in
]%
{7_06_09_1.eps}%
\caption{(a) $r$-value at $t=t_{s}$ versus pulse duration $T$ under
different PA amplitudes: from top to bottom $\Omega_{PA}^{0}=10^{5}$
s$^{-1}$, $4\times10^{5}$ s$^{-1}$, $10^{6}$ s$^{-1}$ and $10^{7}$
s$^{-1}$, respectively. (b) $\tau_{cpt}$ estimated at $t=t_{s}$ as a
function of $\Omega_{PA}^{0}$. The four circles (from left to right)
denote $\tau_{cpt}$=17.8 ms, 2.2 ms, 0.97 ms, 257 $\mu s$ with
respect to the corresponding $\Omega_{PA}^{0}$s shown in (a). (c)
Final molecular production $\eta =2|\psi_{g}(\infty)|^{2}$ versus pulse
duration $T$ under the same four cases as
in (a). The other parameters are described in the text.}%
\label{r_dy}%
\end{center}
\end{figure}
Figure \ref{r_dy}(a) and (c) present the variations of adiabaticity
$r(t_{s})$ and final efficiency $\eta$(=$\left\vert
2\psi_{g}(\infty)\right\vert ^{2}$) as a function of pulse width $T$,
respectively. As plotted in figure 2(a), either a longer pulse (from left to
right) or a stronger PA amplitude (from the top to the bottom) leads to an
improved adiabaticity. Furthermore, when $\Omega_{PA}^{0}$ is very
weak, such as 10$^{5}$ s$^{-1}$ (in solid), the $r$-value is around 1.0,
which cannot well satisfy the adiabatic condition $r\ll1$. Although
at this time, the CPT lifetime in figure \ref{r_dy}(b) is long
enough (more than 15 ms) to support a longer pulse duration, a large
part of atoms will be left in the continuum, resulting in poor
molecular production, which is no more than 30\% (see the solid curve in
(c)). On the other hand, if we use an intense laser,
$\Omega_{PA}^{0}=10^{7}$ s$^{-1}$, then the adiabaticity reduces
into the 0.01 level, whereas simultaneously, the CPT lifetime is only
around 250$\mu s$, leading to a dramatic reduction in $\eta$ as $T$
increases (see the dash-dotted curve in figure 2(c)), because with a longer $T$
value, a number of molecules decay spontaneously due to
$\gamma_{m}$. Obviously, if $T<20\mu s$, a relatively
higher $\eta$ value (%
$>$%
50\%) is still attainable.
In addition, we study two moderate cases with the PA amplitudes:
$\Omega _{PA}^{0}=4\times10^{5}$ s$^{-1}$ (in dashed) and
$\Omega_{PA}^{0}=10^{6}$ s$^{-1}$ (in dotted). No impressive
differences are observable in adiabaticity according to figure
\ref{r_dy}(a), where both are around the 0.1 level. Meanwhile,
the $\tau_{cpt}$ values represented in figure \ref{r_dy}(b) are both close to
1 ms, which do offer more space for a tunable $T$ value. Final
efficiencies in figure \ref{r_dy}(c) clearly exhibit a $T$-dependent
feature, while staying at a highly efficient level compared with two
former cases.
In light of the above discussions, we conclude that the adiabaticity
indeed serves as a useful tool to select favorable parameters.
Meanwhile, it is equivalently important to take the CPT
lifetime into
consideration. Here, we prefer to use $T=30$ $\mu s$, $\Omega_{PA}^{0}%
=4\times10^{5}$ s$^{-1}$.
\subsection{Optimal pump-dump pulse sequence}
The goal of this subsection is to design the optimal pump-dump
two-pulse sequence to maximize the yield of molecules.
Clearly, there are five individual parameters to be determined: $t_{p}$, $t_{d}$,
$\Omega _{p}^{0}$, $\Omega _{d}^{0}$ and $\delta_{m}$. Such a
five-parameter variation is difficult to carry out. However, from
the CPT descriptions, we guess that the population
dynamics are most likely to be affected by the ratio $\chi=\Omega_{p}%
/\Omega_{d}$ instead of the $\Omega_{p}$ and $\Omega_{d}$ values. Therefore, we
introduce two new variables, which are $dt=t_{d}-t_{p}$ for pulse
delay and $\chi^{0}=\Omega _{p}^{0}/\Omega_{d}^{0}$ for peak
amplitude ratio. In combination with the one-photon detuning
$\delta_{m}$, there are three effective
quantities to be optimized.%
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=3.2127in, width=2.5087in
]%
{6_30_09_02.eps}%
\caption{Contour plot of final transfer efficiencies under different
sets of (a) $\left[ \delta_{m},\chi^{0}\right] $ where $dt=105\mu
s$ and (b)
$\left[ \delta_{m},dt\right] $, where $\chi^{0}=6$. Here, $\Omega_{d}%
^{0}=10^{7}$ s$^{-1}$, $t_{p}=75$ $\mu s$. Lighter areas correspond
to high efficiencies. Pure white areas denote efficiencies more than
$80\%$. The other
parameters are listed in the text.}%
\label{optimized}%
\end{center}
\end{figure}
Figure \ref{optimized}(a) and (b) show the contour plots of the
final molecular productions for sets of $\left[
\delta_{m},\chi^{0}\right] $ and $\left[ \delta_{m},dt\right] $,
respectively, where the lighter areas correspond to higher efficiencies.
Especially, pure white regimes are for $\eta\geq80\%$. These two
mappings have several attractive features. Firstly, a symmetric
pattern along the $\delta_{m}$ direction is explicitly
observable, which
can be ascribed to the existence of a three-photon resonance [equation (\ref{delta_g}%
)]. Regardless of whether $\delta_{m}$ is positive or negative, either
$\delta _{g+}=-2\mu_{+}$ or $\delta_{g-}=-2\mu_{-}$ will be
satisfied. In other words, a double-resonant condition must hold on
both sides of $\delta_{m}$, leading to a symmetric double-peak
pattern. Similar patterns have been demonstrated by the
Autler-Townes splitting effect \cite{Aulter55,Bauer09}, which
usually takes place when an optical field is detuned close to an
exact transition frequency. To be more understandable, if we
artificially add a small perturbation to a resonance, the
double-peak profile will be correspondingly shifted. Since this
shift employs no improvement in the molecule production, we
will leave this point for future interested readers.
Secondly, if we fix $\left\vert \delta_{m}\right\vert $ around
$0.05\times 10^{7}$ s$^{-1}$ and gradually increase the values of
$\chi^{0}$ and $dt$, the final transfer efficiencies express similar
variations. Seen in figure \ref{optimized}(a), based on
$\Omega_{d}^{0}=10^{7}$s$^{-1}$, $dt=105\mu s$, if $\chi^{0}$
changes from 1.0 to 6.0, a dramatic enhancement for $\eta$ is
explicit. When further increasing $\chi^{0}$ up to 10.0, $\eta$ values
will be slowly decreasing. A similar trend with $\eta$ as the pulse
delay $dt$ varies
is depicted in figure \ref{optimized}(b) where $\chi^{0}=6$, $\Omega_{d}%
^{0}=10^{7}$s$^{-1}$. When $\delta_{m}=0$, efficiencies are very
poor, which agrees with our CPT predictions equation (\ref{CPT_re})
in section 2. Finally, we find that the base value of
$\Omega_{d}^{0}$ offers few contributions to the transfer. If
$\Omega_{d}^{0}$ is set as $2\times 10^{7}$s$^{-1}$, we will
obtain a much analogous contour plot to figure
\ref{optimized}(a)(not shown).
A brief conclusion for the sections 4.1 and 4.2 is that we are
provided with rich ways to select relevant parameters for
optimal atom-molecule conversion.
\subsection{Population dynamics}
In the following, we consider a concrete example in our "$R$-type"
scheme using the parameters based on our previous discussions.
Optimal parameters are given
by $T=30$ $\mu s$, $\Omega_{PA}^{0}=4\times10^{5}$ s$^{-1}$, $\Omega_{p}%
^{0}=6\times10^{7}$ s$^{-1}$, $\Omega_{d}^{0}=10^{7}$ s$^{-1}$,
$t_{p}=75$ $\mu s$, $t_{d}=180$ $\mu s$, $\left\vert
\delta_{m}\right\vert =5.4\times 10^{5}$ s$^{-1}$,
$\gamma_{m}=3\times10^{4}$ s$^{-1}$ and $\gamma_{e}=10^{7}$
s$^{-1}$. Numerical results are plotted in figure \ref{dy_case}. By
directly integrating the mean-field dynamic equations
(\ref{dy1})-(\ref{dy5}), we produce a population dynamics which
contains all the field amplitudes in figure \ref{dy_case}(b).
Observably, more than 85\% of the atoms $\psi_{0_{1,2}}^{2}$ convert
into ground-state molecules $\psi_{g}^{2}$. Compared with the CPT
dynamics shown in figure \ref{dy_case}(a), a good agreement is
clearly seen, except for a slightly lower $2\psi_{g}^{2}$ coming from
spontaneous decays. In particular, we need to mention that
the $\phi_{m}^{2}(\psi_{m}^{2})$ amplitude (dotted) has been deeply
suppressed, with a maximum value smaller than 0.02.
Figure \ref{dy_case}(c) represents the adiabaticity defined in
equation (\ref{r_value}) as time changes (in solid). By solving
equations (\ref{coupled}) numerically, a complete $r$-value is able
to be determined from the $c_{j}$ values. Three excited eigenenergies
obtained from equation (\ref{cubic}) are displayed in figure
\ref{dy_case}(d) and the inset, where $\omega_{3}$ is smaller than
$\omega_{2,4}$ by orders of magnitude. In the dressed state picture,
$\omega_{j}$ stands for the energy of the $j$th eigenstate, and generally
speaking, a higher-energy eigenstate is usually more difficult to
populate than a lower one. Thereby, in deriving the adiabaticity, we
shall safely neglect the contributions from $\omega_{2,4}$ and
$\mathbf{w}_{2,4}$, simplifying equations
(\ref{coupled}) with $\omega_{3},\mathbf{w}_{3}$ only, yielding%
\begin{equation}
i\omega_{3}c_{3}+f_{33}c_{3}=-\mathbf{\mathbf{w}}_{3}^{T}\mathbf{\dot{\Phi}},
\label{c3}%
\end{equation}
which leads to a reduced assessment on adiabaticity:
$r_{a}=\left\vert
\mathbf{\mathbf{w}}_{3}^{T}\mathbf{\dot{\Phi}/(}i\omega_{3}+f_{33})\right\vert
/\sqrt{3}$. Clearly, $r_{a}$ matches with $r$ in figure \ref{dy_case}(c) perfectly.%
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=4.0271in, width=3.0577in
]%
{6_30_09_03.eps}%
\caption{(color online) (a) CPT dynamics, (b) population dynamics,
(c) time-dependent adiabaticity values, (d) three excited
eigen-energies as time changes where $\omega_{3}$ is at least two
orders smaller than $\omega_{2,4}$
(shown in the inset).}%
\label{dy_case}%
\end{center}
\end{figure}
One critical concern in our scheme is the stability of state
$\left\vert m\right\rangle $, which indeed plays a vital role in
determining the final transfer efficiency. In our calculations, we
use the $\left\vert m\right\rangle $ state lifetime to be
$\tau_{m}=1/(2\pi\gamma_{m})=5.3\mu s$, which is comparable with
the earlier work of Napolitano \textit{et al} \cite{Napolitano94}.
Figure
\ref{lifetime} displays how the final efficiency $\eta$(=2$\psi_{g}^{2}%
(\infty)$) varies as a function of $\tau_{m}$. Clearly, if
$\tau_{m}$ is smaller than $1\mu s$, $\eta$ drops rapidly as
$\tau_{m}$ becomes shorter. However, if we are able to find a more
stable intermediate state, with a lifetime longer than 10 $\mu s$,
the corresponding transfer efficiency reaches as high as 90\%. The
arrow shown in figure \ref{lifetime} points to the
$\tau_{m}$ value used in our paper.%
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=2.1649in, width=2.8636in
]%
{7_04_09_2.eps}%
\caption{The final transfer efficiency $\eta$ as a function of the
intermediate state lifetime $\tau_{m}$. An arrow corresponds to the lifetime we have used.}%
\label{lifetime}%
\end{center}
\end{figure}
Finally, it is meaningful for a numeric estimate of the feasibility of our
scheme by taking the KRb molecule as a possible candidate in experiment. Based
on the predictions in \cite{Aikawa09,Beuc06}, it is experimentally possible
to find out an appropriate $\left\vert m\right\rangle $ state with a
relative long lifetime, e.g. $3^{1}\Sigma ^{+}$. There is a transition
dipole moment of $\sim $ 1 ea$_{0}$ for the state $3^{1}\Sigma ^{+}$, which
corresponds to the lifetime of several $\mu s$. In addition, the $1^{1}\Pi $
is a good candidate for the high-lying $\left\vert e\right\rangle $ state
because of its purely singlet character and favorable transition dipole
moment associated with the lowest singlet state $1^{1}\Sigma ^{+}$ ($%
\left\vert g\right\rangle $ state). A rough estimation of the PA power,
adopting the parameters of the free-bound FCF $I_{FCF}\sim $10$^{-14}$ m$%
^{3/2}$ for KRb \cite{Drummond02}-\cite{Dulieu04} and an initial atomic
density $n_{0}=10^{20}$m$^{-3}$, gives rise to a PA laser intensity of $%
I_{PA}=2c\varepsilon _{0}(\hbar \Omega _{PA}^{0(el)}/\mu )^{2}\approx 512.7$
W/cm$^{2}$ for our "R-type" scheme, where $c$ is the light velocity, $%
\varepsilon _{0}$ is vacuum permittivity, $\mu $ is the dipole moment, $%
\Omega _{PA}^{0(el)}$ is the electronic Rabi frequency defined by $\Omega
_{PA}^{0(el)}=\Omega _{PA}^{(0)}/(I_{FCF}\sqrt{n_{0}})$ \cite{Drummond02}.
However, in a standard "$\Lambda $-type" system (see figure \ref{model}),
the absence of state $\left\vert m\right\rangle $ requires a more intense PA
field to stimulate particles into the highly excited state $\left\vert
e\right\rangle $. As a result, to achieve the same production rate of
molecules as in our "R-type" case, the required PA power must be $\Omega
_{PA}^{\prime 0}\sim 6.4\times 10^{6}$ s$^{-1}$. The corresponding PA
intensity is $I_{PA}^{\prime }\sim 1.31\times 10^{5}$ W/cm$^{2}$, giving
other parameters the same as in the "R-type" case. Evidently, the above
numeric estimate shows a more than 250 times power reduction in our "R-type"
approach.
\section{Summary}
Although the magnetic FR-assisted PA technique has been considered as the
most promising way for the purpose of overcoming the PA weakness, the
primary drawback in such a scheme is the strong inelastic-collisional loss
of Feshbach molecules, especially when the magnetic field is tuned near the
resonant point. To eliminate the bottleneck in the magnetic FR-assisted PA
technique, we work out a robust all-optical atom-molecule conversion model
through a "$R$-type" photoassociative STIRAP, where an intermediate state $%
\left\vert m\right\rangle $ is introduced to form a quasi-CPT state. In
terms of the detailed adiabatic theorem, we show that the quasi-CPT state
can lead to a higher atom-molecule transfer efficiency with a lower PA laser
power, compared to the normal CPT state in the conventional all-optical "$%
\Lambda $-type" two-color PA configuration.
The key reason for the lowered
power of PA laser is due to the existence of an intermediate state $%
\left\vert m\right\rangle $. In this case, it is easier to photoassociate
free atoms into this low-lying $\left\vert m\right\rangle $ state, rather
than a high-lying $\left\vert e\right\rangle $ state. Since molecules in
state $\left\vert m\right\rangle $ are unstable, the subsequent STIRAP
transfer from the intermediate state $\left\vert m\right\rangle $ to the
final state $\left\vert g\right\rangle $ must be rapid enough to avoid the
loss of molecules from the $\left\vert m\right\rangle $ state, here
characterized by a finite CPT lifetime. In addition, we also show that the
lifetime of state $\left\vert m\right\rangle $ will significantly affect the
final transfer efficiency. A specific estimation to illustrate the
feasibility of our approach is performed. Finally we want to emphasize that
the scheme proposed here is the first one to overcome the inefficiency of PA
with only all-optical fields involved. This may open up new opportunities
for experimental endeavors to create polar molecular condensates directly
from ultracold atoms. A more careful treatment taking into account nonlinear
collisions will be left for future explorations.
\section*{Acknowledgments}
This work is supported by the National Natural Science Foundation of China
under Grant No. 10588402, the National Basic Research Program of China (973
Program) under Grant No. 2006CB921104, the Program of Shanghai Subject Chief
Scientist under Grant No. 08XD14017, and the Shanghai Leading Academic
Discipline Project under Grant No. B480 (W.Z.), the National Natural Science
Foundation of China under Grant No. 10974057 and No. 10874045, Shanghai
Pujiang Program under Grant No. 08PJ1405000 (L.Z.).
|
\section{Introduction}\label{sec:intro}
Recent advances in the structure theory of locally compact groups
bring new stimulus to the investigation of the class of locally compact groups
which are non-discrete,
topologically simple and compactly generated. The hope for progress in
this direction is based on
the one hand on a fairly satisfactory understanding of the connected case,
which is completely described thanks to the solution to Hilbert fifth
problem, and on the other
hand to recent progress on the theory of totally disconnected
locally compact groups which tends to show that there is a sharp
contrast between the structure of discrete and non-discrete groups. This
has been notably illustrated by George Willis {\em et.\@ al.\@} in a series of works
starting with \cite{Willis94} and by Marc Burger and Shahar Mozes in their
thorough study of non-discrete automorphism groups of locally finite trees
\cite{Burger-Mozes1}. It was moreover shown in
\cite{Caprace-Monod-monolith} that, under some
natural restrictions which exclude the existence of infinite
discrete quotients, a compactly generated locally compact group
decomposes into finitely many pieces (namely subquotients) which are
all compact, or compactly generated abelian, or compactly
generated and topologically simple. This provides additional
motivation to focus on simple groups in this category.
Let us denote by $\mS$ the class of non-discrete totally disconnected
locally compact groups which are topologically simple. The compactly
generated groups belonging to $\mS$ and known to us at the time of writing
fall into the following classes\footnote{Observe that the examples in this
list fall into countably many isomorphism classes; we do not know whether $\mS$
contains uncountably many pairwise non-isomorphic compactly generated
groups. In the discrete case, it is known since \cite{Camm} that there
are uncountably many isomorphism classes of finitely generated simple
groups.}
:
\begin{itemize}
\item semi-simple linear algebraic groups over local fields (including
groups of mixed type in the sense of Tits);
\item complete Kac--Moody groups over finite fields;
\item tree-automorphism groups and their avatars.
\end{itemize}
The groups from this last class include the tree-automorphism groups
satisfying a simplicity criterion established by Tits~\cite{Tits:trees}
and recalled in Section~\ref{sec:Tits} below. This criterion shall be
referred to as \textbf{Tits' independence property}. The aforementioned
avatars include automorphism groups of some CAT(0) cube complexes
\cite{Haglund-Paulin} and Neretin's group of tree-spheromorphisms
\cite{Neretin}, whose simplicity is proved in \cite{Kapoudjian}.
\medskip
In this paper we focus on compactly generated locally compact groups which
act on trees and satisfy Tits' independence property. Following the
spirit of the work by Burger--Mozes~\cite{Burger-Mozes1}, the central
theme of our work is to investigate to what extent the global structure of
these groups is determined by their local structure. A first sample of this
phenomenon is provided by the following (see also Theorem~\ref{thm:LocPrim} below),
where $E(v)$ denotes the set of edges emanating from the vertex $v$.
\begin{theoremintro}\label{thmi:LocPrim}
Let $T$ be a tree all of whose vertices have valence~$\geq 3$ and $G \in \mS$ be a compactly generated closed subgroup of $\Aut(T)$ which does not stabilise any proper non-empty subtree and which satisfies Tits' independence property. Then the following conditions are equivalent.
\begin{enumerate}[\rm (i)]
\item Every proper open subgroup of $G$ is compact.
\item For every vertex $v \in V(T)$, the induced action of $G_v$ on $E(v)$ is primitive; in particular $G$ is edge-transitive.
\end{enumerate}
\end{theoremintro}
It turns out that, in the above setting, the $G_v$-action on $E(v)$ cannot
be cyclic (see Lemma~\ref{lem:LocCyclic} below). Moreover, the hypotheses of thickness of $T$ and of minimality of the $G$-action can be relaxed, see Theorem~\ref{thm:LocPrim} below.
\begin{remark}
It was shown by Burger--Mozes \cite[Theorem~4.2]{Burger-Mozes2} that a closed subgroup $H \leq \Aut(T)$ which is locally $\infty$-transitive (or equivalently, a closed subgroup acting doubly transitively on the boundary $\partial T$, see \cite[Lemma~3.1.1]{Burger-Mozes1}) enjoys the \textbf{Howe--Moore property}. It is well known that in a locally compact group satisfying the Howe--Moore property, any proper open subgroup is compact. This is consistent with Theorem~\ref{thmi:LocPrim} since a locally $\infty$-transitive action is clearly locally $2$-transitive, and hence locally primitive. We do not know of any \emph{local} characterization of the Howe--Moore property for closed subgroups of $\Aut(T)$ (acting cocompactly on $T$); in particular, we do not know whether the condition that the action be locally $\infty$-transitive is necessary.
\end{remark}
Theorem~\ref{thmi:LocPrim} relates the global structure of $G$ to the
structure of its \emph{maximal} compact subgroups. In the spirit of a
general theory inspired by the classical case of Lie groups, it would be
even more desirable to relate the structure of $G$ to arbitrarily small
compact open subgroups. In order to address this issue, we consider the
group $\mathscr L(G)$ of \textbf{germs of automorphisms} of $G$. By
definition, this group consists of isomorphisms between compact open
subgroups of $G$, modulo the equivalence relation which identifies
isomorphisms between pairs of compact open subgroups which coincide on
respective open subgroups. Alternatively the group $\mathscr L(G)$ can be
defined as the group of \textbf{abstract commensurators} of any compact
open subgroup of $G$. The group $\mathscr L(G)$ is defined for any totally
disconnected locally compact group; it is trivial when $G$ is discrete.
Since every identity neighbourhood of $G$ contains some compact open
subgroup, the group $\mathscr L(G)$ is an invariant of $G$ determined by
its local structure.
The kernel of the canonical homomorphism $\comm : G \to \mathscr L(G)$ is precisely
the {\bf quasi-centre} $\QZ(G)$ of $G$. If the quasi-centre is closed, then
$\mathscr L(G)$ carries a natural group topology,
which is again totally disconnected and locally compact and such that the map $G \to \mathscr L(G)$ is continuous. If in addition $\QZ(G)$ is discrete, then the groups $G$ and $\mathscr L(G)$ are \textbf{locally isomorphic}, \emph{i.e.} they contain isomorphic compact open subgroups. In particular $\mathscr L(\mathscr L(G)) \cong \mathscr L(G)$. We refer to Section~\ref{sec:Germ} below for more details.
It is a natural question to ask to what extent the
group $G$ can be recovered from its local structure.
Following \cite{BarneaErshovWeigel}, we
say that a group $G \in \mS$ is \textbf{rigid} if
any isomorphism between two compact open subgroups of $G$ extends to a unique
automorphism of $G$.
An equivalent way to state this, is to say that
the canonical homomorphism $\Aut(G) \to \mathscr L(G)$ is an isomorphism.
If $G$ is compactly generated, then the group $G \cong \comm(G)$ can be identified with the intersection of
all non-trivial closed normal subgroups of $\acomm(G)$ (see Proposition~\ref{prop:AdmissibleRigid} below);
in this sense a rigid compactly generated group is thus determined by its local structure.
On the other hand, it is clear that two groups containing isomorphic compact
open subgroups have isomorphic groups of germs of automorphisms. In order to address this issue, we shall
say that a group $G \in \mS$ is \textbf{Lie-reminiscent} if for any topologically simple group $H \in
\mS$ which is locally isomorphic to $G$, we have $H \cong G$. Using the ideas developed in
\cite{BarneaErshovWeigel}, one can show that a compactly generated group $G \in \mS$
which is Lie-reminiscent is necessarily rigid (see Proposition~\ref{prop:AdmissibleRigid} below). The examples of simple groups provided by Theorems~\ref{thmi:L(G)} and~ \ref{thmi:NewSimpleGroups} below show that the converse fails, even for compactly generated groups in the class $\mS$.
The following result, proved in Theorem~\ref{thm:NonRigid} below, shows in
conjunction with Proposition~\ref{prop:AdmissibleRigid} that any compactly generated simple tree-automorphism group
satisfying Tits' independence property is not Lie-reminiscent.
\begin{theoremintro}\label{thmi:NonRigid}
Let $T$ be a locally finite tree and $G \in \mS$ be a
compactly generated closed subgroup of $\Aut(T)$ satisfying Tits' independence property.
Then $G$ is not rigid.
\end{theoremintro}
The fact that all these groups are not Lie-reminiscent provides a negative answer
to a question of George Willis \cite[Problem 4.3]{Willis07}. Another example of a non-rigid
compactly generated group belonging to $\mS$ was constructed in \cite{BarneaErshovWeigel}.
Our next goal is to give an explicit description of the group of germs of
automorphisms for some groups $G \in \mS$ satisfying Tits' independence
condition. The groups we shall focus on are the \textbf{universal groups
with prescribed local action} defined by
Burger--Mozes~\cite{Burger-Mozes1}. The precise definition of these groups
is recalled in Section~\ref{sec:BM} below.
\begin{theoremintro}\label{thmi:L(G)}
Let $d >1$ and $F \leq \Sym(d+1)$ be a doubly transitive finite permutation group
and $G = U(F)^+$ be the universal simple group acting on the regular tree
of degree $d+1$ locally like $F$. Then the following conditions are equivalent:
\begin{enumerate}[\rm (i)]
\item $\mathscr L(G)$ is compactly generated.
\item $G$ is locally isomorphic to some compactly generated rigid group $H \in \mS$.
\item $\norma_{\Sym(d)}(F_0) = F_0$, where $F_0$ denotes a point stabiliser in $F$.
\end{enumerate}
If these conditions hold, then the commutator subgroup $[\mathscr L(G), \mathscr L(G)]$ is open,
abstractly simple, and has index~$1$ or $2$ in $\mathscr L(G)$.
\end{theoremintro}
Condition (iii) of Theorem~\ref{thmi:L(G)} clearly excludes the alternating groups or the sharply $2$-transitive groups. Although we did not try to be fully exhaustive here, it seems however that the proportion of doubly transitive finite permutation groups which satisfy that condition is rather large (see \cite{Cameron} for a list of all finite $2$-transitive groups).
Theorem~\ref{thmi:L(G)} will be deduced from a detailed study of the group of abstract commensurators of self-replicating profinite wreath branch groups which was largely inspired by a reading of \cite{BarneaErshovWeigel}. As a consequence of this study, we shall establish the following theorem, which highlights an infinite family of compactly generated locally compact groups which are simple and rigid, but not Lie-reminiscent. We recall that a locally compact group is called \textbf{locally elliptic} if every compact subset is contained in a compact subgroup. We refer to \cite[\S4]{Brown} for a precise definition of the Higman--Thompson group $F_{d, k}$ appearing below. For any $k>0$, the group $F_{2, k}$ is isomorphic to Thompson's group $F$.
\begin{theoremintro}\label{thmi:NewSimpleGroups}
Let $d > 1$, $D \leq \Sym(d)$ be transitive and $W = W(D)$ be the profinite branch group defined as the infinitely iterated wreath product of $D$ with itself. Then for every $k>0$, there is a locally compact group $M = M(D, k)$ which is topologically simple and rigid, and which contains the direct product $W^k$ of $k$ copies of $W$ as a compact open subgroup. Moreover:
\begin{enumerate}[\rm (i)]
\item $M$ is uniquely determined up to isomorphism.
\item $M$ is compactly generated if and only if $\norma_{\Sym(d)}(D) = D$.
\item $[\acomm(M): \comm(M)] \leq 2$.
\item $\acomm(M) = F_{d, k} \cdot A_k$, where $F_{d, k}$ is a copy of the Higman--Thompson group embedded as a discrete subgroup, and $A_k$ is a non-compact locally elliptic open subgroup such that $F_{d, k} \cap A_k = 1$.
\end{enumerate}
\end{theoremintro}
The connection between Theorems~\ref{thmi:L(G)} and~\ref{thmi:NewSimpleGroups} is the following:
Let $F$, $F_0$ and $G$ be as in Theorem~\ref{thmi:L(G)}, and define $M = M(F_0, 2)$ as in Theorem~\ref{thmi:NewSimpleGroups}.
Then the simple group $[\acomm(G), \acomm(G)]$ appearing in Theorem~\ref{thmi:L(G)} coincides with the group $M = M(F_0, 2)$.
Theorem~\ref{thmi:L(G)} applies notably to the full symmetric group $F = \mbox{$\Sym(d+1)$}$; in
that case $\mathscr L(G)$ coincides with {Neretin's group of
spheromorphisms} introduced in \cite{Neretin}. Equivalently, the Neretin group of the regular tree of degree~$d+1$ is isomorphic to the group $M(\Sym(d), 2)$ appearing in Theorem~\ref{thmi:NewSimpleGroups}. We can thus summarize our results about Neretin's group of spheromorphisms as follows; as mentioned above, the simplicity statement is originally due to Ch.~Kapoudjian~\cite{Kapoudjian}.
\begin{corintro}
Let $T$ be the regular tree of degree $d+1$, and let $G$ be Neretin's group of spheromorphisms of $T$.
Then:
\begin{enumerate}[\rm (i)]
\item
$G = \mathscr L(\Aut(T))$. In particular $\mathscr L(G) \cong G$, i.e.\@ $G$ is hyperrigid in the sense of \cite{BarneaErshovWeigel}.
\item
$G$ is compactly generated, abstractly simple and rigid, but not Lie-reminiscent.
\end{enumerate}
\end{corintro}
\begin{remark*}
After completing this paper, we learned that Thomas Weigel has generalized Neretin's group
in a different direction, namely by considering more general classes of rooted trees
(but always taking the full group of almost automorphisms of such a rooted tree).
This provides in particular other examples giving a negative answer to George Willis' question
from \cite[Problem 4.3]{Willis07}; see \cite{Weigel}.
\end{remark*}
\subsection*{Acknowledgements}
We warmly thank Marc Burger for stimulating discussions and for suggesting the term \emph{germ of automorphism}.
We are grateful to Thomas Weigel for helpful comments on an earlier draft of this paper.
Furthermore, two anonymous referees did a wonderful job and provided us with a long list of very useful comments,
which allowed us to improve the exposition of the paper.
A part of this was written up while the first-named author was visiting the Hausdorff Institute for Mathematics in Bonn;
he thanks the institute for its hospitality.
\section{Preliminaries}\label{sec:prel}
Let $T$ be a locally finite tree.
We will use the notation $V(T)$ for the set of vertices of $T$, and $E(T)$ for the edge set.
Given $v \in V(T)$, we denote by $E(v)$ the set of edges containing $v$.
Recall that $\Aut(T)$ comes equipped with a natural topology, namely the {\bf permutation topology}, which is defined
by declaring a subgroup $U \leq \Aut(T)$ to be open if and only if $U$ contains the pointwise stabiliser
of some finite subset $S \subset T$.
This topology coincides with the topology of pointwise convergence, as well as with the compact-open topology
(when $V(T)$ is endowed with the discrete topology).
It is Hausdorff, totally disconnected and locally compact;
it is discrete if and only if there is a finite subset $S \subset T$ the pointwise stabiliser of which is trivial.
We first point out that the simple locally compact groups acting properly on $T$ can naturally be viewed as closed subgroups of $\Aut(T)$.
\begin{lemma}\label{lem:metr}
Assume that $G \in \mS$ is a group admitting a continuous and proper action on $T$.
Then $G$ is homeomorphic to its image in $\Aut(T)$, which is closed and metrisable.
\end{lemma}
\begin{proof}
Since the action is proper and $T$ is locally finite, vertex stabilisers are compact and
the desired result follows from a standard compactness argument,
recalling that $G$ is necessarily metrisable and separable (see \cite{KakutaniKodaira} or~\cite[Ch.~II, Th.~8.7]{HewittRoss}).
\end{proof}
From now on, we will only consider closed subgroups of $\Aut(T)$.
We will collect some interesting properties for compactly generated closed subgroups of $\Aut(T)$ contained in $\mS$.
To begin, we give a useful criterion (due to Bass and Lubotzky) for when a closed subgroup of $\Aut(T)$
is compactly generated.
\begin{defn}
Assume that $G$ is a (closed) subgroup of $\Aut(T)$, such that
$G$ stabilises a subtree $T' \leq T$ but no subtree of $T'$.
Then we say that $T'$ is a {\bf minimal invariant subtree} of $T$ for the $G$-action.
\end{defn}
\begin{remark}\label{re:invT}
\begin{compactenum}[(i)]
\item
Assume that $G$ is a subgroup of $\Aut(T)$
such that $G$ does not fix an end of $T$.
Then there exists a minimal $G$-invariant subtree of $T$.
If in addition $G$ fixes at most one vertex of $T$,
then this minimal invariant subtree is unique.
See \cite[Corollaire~3.5]{Tits:trees}.
\item
If $T'$ is a minimal $G$-invariant subtree of $T$,
then either $T'$ is infinite and has no endpoints (an endpoint is a vertex with valency $1$),
or $T'$ consists of a single vertex or a single edge.
\end{compactenum}
\end{remark}
The following fact is well-known; we include a detailed proof for the reader's convenience.
\begin{lemma}\label{le:ccpt}
Assume that $G$ is a closed subgroup of $\Aut(T)$,
and that $T'$ is a minimal $G$-invariant subtree of $T$.
Then $G$ is compactly generated if and only if its action on $T'$ is cocompact.
In particular, if $G$ acts edge-transitively on $T$, then $G$ is compactly generated.
\end{lemma}
\begin{proof}
Observe that the action of $G$ on any locally finite graph $\Gamma$ is cocompact
if and only if $G$ has finitely many orbits on $\Gamma$.
(Indeed, a compact subset of $\Gamma$ contains only finitely many vertices and edges.)
So assume first that the action of $G$ on $T'$ is cocompact;
then $G$ has finitely many orbits on $T'$.
Let $C$ be the convex hull of a finite fundamental domain for the action of $G$ on $T'$;
then $C$ is a finite subtree of $T'$.
Fix some $v \in C$, and let $H$ be the set of elements of $G$ mapping $v$ into $C$.
Then $H$ is the union of a finite number of cosets of the compact open subgroup $G_v \leq G$,
and hence $H$ is a compact subset of $G$.
Now let $\{ v_1,\dots,v_n \}$ be the sets of vertices adjacent to some vertex of $C$,
but not contained in $C$.
Then for each $v_i$, there is an element $g_i \in G$ mapping $v_i$ into $C$.
Let $S = \{ g_1,\dots,g_n \}$, and fix some $v \in C$;
then for each $g \in G$, there is some $s \in \langle S \rangle$ mapping
$v^g$ into $C$ (this follows by induction on the distance from $v^g$ to $C$).
But then $gs$ maps $v$ into $C$, and hence $gs \in H$.
We conclude that the compact set $H \cup S$ generates $G$.
Conversely, assume that $G = \langle S \rangle$ for some compact set $S \subseteq G$.
Let $v \in V(T')$ be arbitrary;
then $U := G_v$ is a compact open subgroup of $G$.
Then $S$ is covered by the open sets $Ug$, where $g$ runs through the elements of~$S$,
and since $S$ is compact, there is a finite subcover $S \subseteq \{ U g_i \mid 1 \leq i \leq k \}$.
It follows that $G$ is generated by the compact set $S' = U \cup \{ g_1, \dots, g_k \}$.
For each $i=1,\dots,k$, let $v_i = v \cdot g_i$, and let $F$ be the (finite) convex hull of
$v, v_1, \dots, v_k$;
then $F \cup F \cdot s$ is connected, for each $s \in S'$.
Now let $T''$ be the union of all $G$-translates of $F$.
Since $G = \langle S' \rangle$, it follows that $T''$ is connected,
and hence it is a subtree of $T'$;
it is clearly $G$-invariant, and hence $T'' = T'$ by the minimality of $T'$.
Since $F$ is finite, this shows that $G$ has only finitely many orbits on $T'$,
and we conclude that $G$ acts cocompactly on $T'$.
\end{proof}
\begin{defn}
Recall that an element $g \in \Aut(T)$ acting without inversion is called {\bf elliptic} if it fixes some point,
and it is called {\bf hyperbolic} otherwise.
Every hyperbolic element $g \in \Aut(T)$ stabilises a geodesic line, called the {\bf axis} of $g$, on which $g$
acts by a non-trivial translation.
\end{defn}
Observe that if $G \in \mS$ is a subgroup of $\Aut(T)$, then $G$ acts without inversion on $T$
(else it would contain an index $2$ subgroup).
\begin{lemma}\label{lem:basic}
Assume that $G \in \mS$ is a closed subgroup of $\Aut(T)$ that is compactly generated.
Then:
\begin{enumerate}[\rm (i)]
\item
Every compact subgroup of $G$ is contained in a vertex stabiliser (which is compact and open).
\item
The group $G$ does not fix an end of $T$.
\item
Every ascending chain of compact open subgroups of $G$ stabilises;
in particular, every compact subgroup of $G$ is contained in a maximal one.
\item
Any open subgroup $H \leq G$ which is not compact contains a hyperbolic element.
\item
If an open subgroup $H \leq G$ fixes an end of $T$, then either $H$ is compact, or $H$ stabilises a geodesic line of $T$.
\item
For every open subgroup $H \leq G$, there exists a minimal $H$-invariant subtree of $T$.
\end{enumerate}
Moreover, if the action of $G$ on $T$ is edge-transitive, then:
\begin{enumerate}[\rm (i)]\setcounter{enumi}{6}
\item
The vertex stabilisers are precisely the maximal compact subgroups of $G$;
they fall into two conjugacy classes, which correspond to the bipartition of $T$.
\item
Any tree $T'$ admitting a continuous proper edge-transitive
action of $G$ is isomorphic to $T$.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{compactenum}[\rm (i)]
\item
Any bounded subset of $T$ admits a canonical centre, hence compact
subgroups all have a fixed point. Since $G$ acts without inversion,
a subgroup which fixes a point of $T$ necessarily fixes a vertex.
\item
Suppose that $G$ fixes an end $\xi \in \partial(T)$.
Let $\chi_\xi \colon G \to \ZZ$ be the corresponding Busemann character;
then for any $g \in G$, $\chi_\xi(g) = 0$ precisely when $g$ is elliptic.
Since $G$ is simple, the image of $\chi_\xi$ is trivial, hence $G$ consists of elliptic elements only.
Therefore $G$ stabilises each horoball centred at $\xi$.
Since the family of all these horoballs is nested,
it follows that $G$ is the union of an ascending chain of compact subgroups.
Since $G$ is compactly generated, however, this chain is stabilising, and hence $G$ is compact.
The desired statement follows since a simple profinite group is necessarily finite, hence it cannot belong to $\mathcal S$.
\item
By (ii) and Remark~\ref{re:invT}(i), there exists a minimal $G$-invariant subtree $T'$ of $T$;
since $G$ is simple, its action on $T'$ is faithful, so there is no loss of generality in
assuming $T' = T$.
By Lemma~\ref{le:ccpt}, the action of $G$ on $T$ is now cocompact, \emph{i.e.}\@ $G$ has only finitely many
orbits on $V(T)$.
In particular, there are only finitely many conjugacy classes of vertex stabilisers in $G$.
Every compact open subgroup has finite positive volume;
since $G$ is simple, it is unimodular, and hence the Haar measure is conjugacy invariant.
This implies that the volume of any vertex stabiliser is bounded above by a constant;
since every compact subgroup fixes a vertex by (i),
it follows that the volume of any compact subgroup is bounded,
and hence any ascending chain of compact open subgroups stabilises.
\item
It is well known that if a group acts on a tree in such a way that every element has a fixed point,
then the whole group has a fixed point or a fixed end;
see \cite[Proposition 3.4]{Tits:trees} or \cite[Exercise~2, p.~66]{Serre}.
Assume that $H \leq G$ is open and contains only elliptic elements.
If $H$ has a fixed point in $T$, then $H$ is compact and we are done.
Otherwise $H$ fixes an end $\xi \in \partial T$ and since every element of $H$ is elliptic,
it follows that $H$ stabilises each horoball centred at $\xi$,
and hence $H$ is the union of an ascending chain of compact subgroups.
Since $H$ is open and since every compact subgroup is contained in some compact open subgroup by (i),
this implies that $G$ contains an ascending chain of compact open subgroups, the union of which is $H$.
By (iii), this chain is stationary, which implies that $H$ is compact as desired.
\item
Let $\xi \in \partial(T)$ be an end fixed by $H$, and assume that $H$ is non-compact.
Let $\chi_\xi \colon H \to \ZZ$ be the corresponding Busemann character, and let $H_0 = \ker(\chi_\xi)$.
Then $H_0$ is open, and since it stabilises all horoballs centred at $\xi$, it is a union of compact subgroups;
hence $H_0$ is a compact open normal subgroup of $H$ by (iii).
It follows that $H$ acts on the fixed tree $T^{H_0}$ of $H_0$.
Since $H$ is non-compact, it contains a hyperbolic element by (iv),
and hence the image of $\chi_\xi$ is non-trivial.
Let $h \in H$ be such that $\chi_\xi(h)$ generates $\chi_\xi(H)$;
then $H = \langle H_0, h \rangle$.
Let $\lambda$ be the axis of the hyperbolic element $h$;
then $\lambda \subseteq T^{H_0}$ because $\lambda$ is the smallest subtree invariant under $h$.
We conclude that $H$ stabilises the axis of the hyperbolic element $h$.
\item
If $H$ does not fix an end of $T$, then the result follows from Remark~\ref{re:invT}(i).
So assume that $H$ fixes an end of $T$.
Then by (v), either $H$ is compact, in which case $H$ fixes a vertex;
or $H$ stabilises a geodesic line $\lambda$ of $T$,
so in particular any $H$-invariant subtree contains $\lambda$.
The intersection of all $H$-invariant subtrees of $T$ is thus non-empty, and clearly minimal.
\item
Assume that some vertex stabiliser $G_v$ is not maximal compact, say $G_v \leq H$ with $H$ compact;
then by (i), there is another vertex $w$ such that $G_v \leq G_w$, and
by considering the path from $w$ to $v$, we may assume that $w$ is a neighbour of $v$.
This would imply that every element of $G$ fixing $v$ would fix the edge $\{v,w\}$,
but this contradicts the edge-transitivity.
Since $G$ acts edge-transitively and without inversion, it has precisely two orbits on the vertices of $T$;
the second claim follows.
\item
Since the $G$-action is edge-transitive, the tree $T$ is biregular.
It suffices to show that the two valencies of the two classes of vertices are uniquely determined by $G$.
Denoting by $\mathcal U_1$ and $\mathcal U_2$ the two conjugacy classes of maximal compact subgroups of $G$ provided by~(vii),
one verifies easily that these two degrees coincide with
\[ \min \{ |U_1: U_1 \cap U_2|, \; U_1 \in \mathcal U_1, \; U_2 \in \mathcal U_2\} \phantom{\,.} \]
and
\[ \min \{ |U_2: U_1 \cap U_2|, \; U_1 \in \mathcal U_1, \; U_2 \in \mathcal U_2\} \,, \]
respectively.
\qedhere
\end{compactenum}
\end{proof}
\section{Groups with Tits' independence property}\label{sec:Tits}
\subsection{Tits' independence property}\label{ss:Tits}
The following remarkable theorem of Tits \cite{Tits:trees} shows that if $G$ is in some sense
large enough, then $G$ has a (usually rather big) abstractly simple subgroup.
We first make our definition of ``large enough'' precise.
\begin{defn}
Let $G \leq \Aut(T)$, let $C$ be a (finite or infinite) chain of $T$, and let $F$ be the pointwise
stabilizer of $C$ in $G$.
For each vertex $v$ of $T$, we denote by $\pi(v)$ the vertex of $C$ closest to $v$.
The vertex sets $\pi^{-1}(c)$ with $c \in V(C)$ are all invariant under $F$; let $F_c$ denote
the permutation group obtained by restricting the action of $F$ to $\pi^{-1}(c)$.
Then there is a natural homomorphism
\[ \varphi \colon F \to \prod_{c \in V(C)} F_c \,. \]
We say that $G$ satisfies {\bf Tits' independence property}
if and only if $\varphi$ is an isomorphism.
\end{defn}
\begin{remark}
If $G \leq \Aut(T)$ is a {\em closed} subgroup, then $G$ satisfies Tits' independence property
if and only if for every edge $e \in E(T)$, the pointwise edge stabiliser $G_{(e)}$ can be decomposed as
$G_{(e)} = G_{(h_1)} G_{(h_2)}$, where $h_1$ and $h_2$ are the rooted half-trees emanating from the
endpoints of $e$, \emph{i.e.}\@ $E(T)$ is the disjoint union of $E(h_1)$, $e$ and $E(h_2)$.
See, for example, \cite[Lemma~10]{Am}.
\end{remark}
\begin{theorem}[\cite{Tits:trees}]\label{th:tits}
Let $T$ be a simplicial tree, $G$ a subgroup of $\Aut(T)$, and $G^+$ the subgroup of $G$ generated
by the edge-stabilizers.
Assume that $G$ does not stabilize a proper subtree, and that $G$ does not fix an end of $T$.
If $G$ satisfies Tits' independence property, then $G^+$ is simple (or trivial).
\end{theorem}
\subsection{Going from $G$ to $G^+$}
As we will see, the group $G^+$ can be characterised topologically as the \textbf{monolith} of $G$, \emph{i.e.} the unique minimal closed normal subgroup of $G$ or, equivalently, the intersection of all non-trivial closed normal subgroups of $G$. We first recall the definition of the quasi-centre, which was introduced by Burger and Mozes in \cite{Burger-Mozes1}.
\begin{defn}\label{def:QZ}
The {\bf quasi-centre} $\QZ(G)$ of $G$ is
the characteristic (but not necessarily closed) subgroup
consisting of all those elements admitting an open centraliser.
(This group is sometimes called the {\em virtual centre} $\mathrm{VZ}(G)$ by some authors.)
\end{defn}
The following elementary but important fact is well known, and we will use it frequently without explicitly mentioning it.
\begin{lemma}
If $N \unlhd G$ is a discrete normal subgroup, then $N \leq \QZ(G)$.
In particular, if $\QZ(G) = 1$, then every non-trivial normal subgroup is non-discrete.
\end{lemma}
\begin{proof}
Let $N \unlhd G$ be a discrete normal subgroup.
Let $g \in N$ be arbitrary; then $g^G$ is a discrete conjugacy class.
This implies that $g$ has an open centraliser, and hence $g \in \QZ(G)$.
\end{proof}
\begin{prop}
Let $G$ be a closed subgroup of $\Aut(T)$ satisfying Tits' independence property,
and assume that $G$ does not stabilize a proper subtree, and that $G$ does not fix an end of $T$.
Let $G^+$ be the subgroup generated by its edge stabilisers.
Then
\begin{compactenum}[\rm (i)]
\item
$\QZ(G) = 1$.
\item
$G^+$ is the unique minimal closed normal subgroup of $G$.
\end{compactenum}
\end{prop}
\begin{proof}
\begin{compactenum}[(i)]
\item
Let $g \in \QZ(G)$, and let $v$ be an arbitrary vertex of $T$.
Then by definition, there exists a finite subset $S \subset T$ such that $g$ centralises the pointwise stabiliser $G_{(S)}$,
and we may assume that $S$ contains the vertex $v$ (by replacing $S$ by $S \cup \{ v \}$).
Let $\tilde{S}$ be the fixtree of $G_{(S)}$; in particular, $G_{(S)} = G_{(\tilde{S})}$.
Then $g$ stabilises $\tilde{S}$. By Tits' independence property the group $ G_{(\tilde{S})}$ splits
as a direct product of many factors, and each factor naturally corresponds to a vertex of $\tilde S$
contained in at least one edge not in $\tilde S$.
Since $g$ centralises $G_{(\tilde{S})}$,
it normalises each of these factors and, hence, it acts trivially on the vertices of $\tilde{S}$
contained in at least one edge not in $\tilde S$; but then it acts trivially on all the vertices of $\tilde{S}$.
In particular, $g$ fixes $v$, and since $v$ was arbitrary, $g$ is trivial, as desired.
\item
Let $1 \neq N$ be a closed normal subgroup of $G$;
we will show that $N$ contains $G^+$.
If $N \cap G^+ = 1$, then $N$ would be a discrete normal subgroup of $G$.
This is impossible, however, since $\QZ(G) = 1$ by (i).
Hence $N \cap G^+ = G^+$ since $G^+$ is simple by Theorem~\ref{th:tits},
and this proves our claim.
\qedhere
\end{compactenum}
\end{proof}
We point out that $G^+$ need not be compactly generated, even if the group $G$ is.
\begin{example}
Let $T$ be the Cayley tree of the free group $F_2$ associated with a basis $\{a,b \}$. We view $T$ as a bi-coloured graph (one colour per generator). Let $G \leq \Aut(T)$ be the full automorphism group of this coloured graph. Then $G$ is closed; it is furthermore cocompact since it contains $F_2$.
Let $G^+$ denote the (open) subgroup of $G$ generated by the pointwise edge-stabilisers.
Then $G^+$ is simple but not compactly generated.
\end{example}
\begin{proof}
The fact that $G^+$ is simple follows from Theorem~\ref{th:tits}.
Let $V = F_2$ be the vertex-set of $T$.
Let $D_\infty = \langle a', b' \mid (a')^2 = (b')^2 = 1 \rangle$ denote the infinite dihedral group.
Let also $f \colon F_2 \to D_\infty$ be the canonical homomorphism defined by $f(a)= a'$ and $f(b)=b'$.
We now define a map
\[ \delta \colon V \times V \to D_\infty \colon (v, w) \mapsto f(w^{-1}v) \,. \]
Observe that $\delta(v,w)$ only depends on the colours of the edges of the unique path from $v$ to $w$;
in particular, $\delta$ is $G$-equivariant. The crucial observation is now the following: for each $g \in G$ which fixes pointwise some edge of $T$, we have $\delta(v, g.v) = 1$ for all $v \in V$. Since $G^+$ is generated by edge-stabilizers, it readily follows that $\delta(v, g.v) = 1$ for all $g \in G^+$ and $v \in V$.
Now pick a base vertex $v_0$ and consider an `infinite staircase' consisting of a bi-infinite geodesic line $(\dots, v_{-1}, v_0, v_1, \dots)$ such that
$\delta(v_0, v_{2n}) = (a' b')^n$ for every integer $n$.
It follows from the preceding discussion that no element of $G^+$ can map $v_0$ to $v_i$ for any $i \neq 0$;
in fact, this bi-infinite geodesic line is a fundamental domain for the $G^+$-action on $T$.
In particular, we conclude that the $G^+$-action is not cocompact.
Since $G^+$ acts minimally on $T$, Lemma~\ref{le:ccpt} implies that $G^+$ is not compactly generated;
on the other hand, $G$ does act cocompactly on $T$ and hence the same lemma implies that $G$ is compactly generated.
\end{proof}
\subsection{Open subgroups}
It turns out that for simple groups with Tits' independence property, we have some control over the open subgroups.
\begin{prop}\label{prop:Titsopen}
Let $G \in \mS$ be a closed subgroup of $\Aut(T)$ satisfying Tits' independence property,
and let $H$ be an open subgroup of $G$.
If $T'$ is a minimal $H$-invariant subtree of $T$ (which always exists by Lemma~\textup{\ref{lem:basic}(vi)}),
then $H$ contains the full edge-stabiliser $G_e$ for each edge $e \in E(T')$.
\end{prop}
\begin{proof}
If $H$ is compact, then it is contained in a vertex stabiliser $G_v$, and hence
every minimal $H$-invariant subtree of $T$ is reduced to a single vertex;
in this case, the statement is empty.
So assume that $H$ is not compact.
Then by Lemma~\ref{lem:basic}(iv), there is some hyperbolic element $\gamma \in H$;
let $L$ be the axis of $\gamma$.
Since $H$ is open, on the other hand, there is some finite subtree $S$ of $T$ such that the pointwise
stabiliser $G_{(S)}$ of $S$ is contained in $H$.
We claim that $L \subseteq T'$.
Indeed, let $v \in V(T')$ be arbitrary, and let $z$ be the vertex of $L$ closest to $v$;
then $\gamma$ maps $z$ to another vertex $z\gamma$ of $L$, and the path between $v$ and $v\gamma$
contains the path between $z$ and $z\gamma$, which contains at least one edge $e$ of $L$.
Since $v, v\gamma \in V(T')$, we conclude that $e \in E(T')$, and hence the $\langle \gamma \rangle$-orbit
of $e$ is contained in $E(T')$; since the convex hull of this orbit is precisely $L$, we have $L \subseteq T'$ as claimed.
Next, suppose that $e$ is an edge of $T$, and let $T_1$ and $T_2$ be two subtrees of~$T$
not containing $e$, and lying at different sides of~$e$.
We claim that if the pointwise stabilisers $G_{(T_1)}$ and $G_{(T_2)}$ are both contained in~$H$,
then so is the edge stabiliser $G_e$.
Indeed, let $T'_i$ be the unique half-tree rooted at one of the endpoints of $e$ containing $T_i$ for $i \in \{ 1,2 \}$.
Then $G_{(T'_i)} \leq G_{(T_i)} \leq H$ for each $i$, and hence, by Tits' independence property,
\[ G_e = \bigl\langle G_{(T'_1)} \cup G_{(T'_2)} \bigr\rangle \leq H \]
as claimed.
In particular, if $e'$ and $e''$ are two edges of $T$ such that
$G_{e'}$ and $G_{e''}$ are contained in $H$, then for every edge $e$ lying
on the geodesic from $e'$ to $e''$ we have $G_e \leq H$ as well.
Now let $e \in E(L)$ be arbitrary.
By applying appropriately high positive or negative powers of $\gamma$ to $S$, we get subtrees $S_1$ and $S_2$
not containing $e$ and lying at different sides of $e$.
Since $G_{(S_i)} \leq \langle G_{(S)}, \gamma \rangle \leq H$ for both $i=1,2$,
we can apply the previous paragraph to conclude that $G_e \leq H$.
Since $T'$ is a minimal $H$-invariant subtree of $T$ containing the edge $e$, the convex hull
of the $H$-orbit of $e$ (which is clearly $H$-invariant) has to coincide with $T'$ itself.
So let $f \in E(T')$ be arbitrary; then there exist elements $\alpha_1,\alpha_2 \in H$ such that
$f$ lies on the geodesic from $e\alpha_1$ to $e\alpha_2$.
Since $G_{e\alpha_i} = G_e^{\alpha_i} \leq H$, we conclude that $G_f \leq H$ as well.
\end{proof}
We say that a locally compact group has {\bf few open subgroups} if every proper open subgroup
is compact.
The large edge-transitive tree automorphism groups with few open subgroups can be characterised in terms of their local action.
\begin{theorem}\label{thm:LocPrim}
Let $T$ be a tree and $G \in \mS$ be a compactly generated non-compact closed subgroup of $\Aut(T)$ satisfying Tits' independence property. Then the following conditions are equivalent.
\begin{enumerate}[\rm (i)]
\item Every proper open subgroup of $G$ is compact.
\item $G$ splits as an amalgamated free product $G \cong A *_C B$, where $A$ and $B$ are maximal compact open subgroups and $C = A \cap B$ is a maximal subgroup of both $A$ and $B$. Moreover the Bass--Serre tree associated to this amalgam admits a $G$-equivariant embedding in $T$ which is isometric up to a scaling factor.
\item $G$ splits as an amalgamated free product $G \cong A *_C B$, where $A$ and $B$ are maximal compact open subgroups, $C = A \cap B$ and the $A$-action on $A/C$ (resp. the $B$-action on $B/C$) is primitive and non-cyclic. Moreover the Bass--Serre tree associated to this amalgam admits a $G$-equivariant embedding in $T$ which is isometric up to a scaling factor.
\end{enumerate}
\end{theorem}
A tree is called \textbf{thick} if the valence of every vertex is at least three. The following immediate corollary is nothing but a reformulation of Theorem~\ref{thmi:LocPrim} from the introduction.
\begin{cor}\label{cor:LocPrim}
Let $T$ be a thick tree and $G \in \mS$ be a compactly generated closed subgroup of $\Aut(T)$ acting minimally on $T$ and satisfying Tits' independence property. Then the following conditions are equivalent.
\begin{enumerate}[\rm (i)]
\item Every proper open subgroup of $G$ is compact.
\item For every vertex $v \in V(T)$, the induced action of $G_v$ on $E(v)$ is primitive; in particular $G$ is edge-transitive.
\item For every vertex $v \in V(T)$,
the induced action of $G_v$ on $E(v)$ is primitive and non-cyclic; in particular $G$ is edge-transitive.
\end{enumerate}
\end{cor}
We shall need a subsidiary observation, which clarifies the equivalence between (ii) and (iii) in Theorem~\ref{thm:LocPrim} and Corollary~\ref{cor:LocPrim}.
\begin{lemma}\label{lem:LocCyclic}
Let $G\leq \Aut(T)$ be a closed subgroup satisfying Tits' independence property, acting edge-transitively on $T$. If $G \in \mS$, then there is no vertex $v \in V(T)$ such that the $G_v$-action on $E(v)$ is free.
\end{lemma}
\begin{proof}
Since $G$ is edge-transitive, it follows that $G_v$ is transitive on $E(v)$ for all $v \in V(T)$. Let now $v_0 \in V(T)$ be such that the $G_{v_0}$-action on $E(v_0)$ is free. Let $v_0^\perp$ denote the subset of $V(T)$ consisting of all vertices adjacent to, but distinct from, $v_0$. Let $G_0$ denote the subgroup of $G$ generated by $\{G_w \; | \; w \in v_0^\perp\}$. Then $G_0$ is an open subgroup of $G$ and moreover we have $E(v_0) \cdot G_0 = E(T)$. Notice moreover that, in view of the assumption that $G_{v_0}$ acts freely on $E(v_0)$, it follows from Bass--Serre theory that $E(v_0)$ is a strict fundamental domain for the $G_0$-action on $E(T)$ and that $G_0$ is the free product of the groups $\{G_w \; | \; w \in v_0^\perp\}$ amalgamated over their intersection $\bigcap_{ w \in v_0^\perp} G_w $.
Since $G_0$ acts cocompactly on $T$, we infer that $T$ is a minimal invariant subtree for $G_0$. In view of Proposition~\ref{prop:Titsopen}, it follows that $G_e$ is contained in $G_0$ for every edge $e \in E(T)$. Since $G$ is simple, it is generated by its edge-stabilisers, and we infer that $G = G_0$. Since $G$ acts edge-transitively while $G_0$ has exactly $| E(v_0)|$ orbits on $E(T)$, we infer that the star $E(v_0)$ is in fact reduced to a single edge of $T$. Therefore $v_0$ is an endpoint of $T$. In particular $T$ is bounded and hence, $G$ is finite. This is absurd since the class $\mS$ does not contain any discrete group.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:LocPrim}]
Observe that by Proposition~\ref{prop:Titsopen}, each open subgroup $H$ of $G$ has a minimal invariant subtree, and such a subtree reduces to a single vertex if and only if $H$ is compact. There is no loss of generality in assuming that $G$ acts minimally on $T$. Since $G$ is compactly generated, we deduce from Lemma~\ref{le:ccpt} that the action of $G$ on $T$ is cocompact.
\begin{compactitem}
\item[(i) $\Rightarrow$ (ii)]
Bass--Serre theory provides us with a graph of group decomposition of $G$ over a finite graph $\Gamma$. The corresponding vertex- and edge-groups are nothing but vertex- and edge-stabilisers for the $G$-action on $T$. In particular they are compact open subgroups of $G$.
Collapsing the aforementioned graph of group decomposition to a single edge, we obtain a presentation of $G$ as a non-trivial amalgamated free product or as an HNN-extension. Since $G$ is simple, it does not map onto $\mathbf Z$ and can therefore not be an HNN-extension. Thus we have a decomposition $G = A *_C B$ with $C= A \cap B$. Since $A$ and $B$ are generated by vertex-groups of the initial decomposition, they are open subgroups of $G$. They must therefore be compact since $G$ has few open subgroups.
Let $\tilde T$ be the Bass--Serre tree associated with the decomposition $G = A *_C B$, and let $\tilde x, \tilde y \in V(\tilde T)$ denote the vertices respectively fixed by $A$ and $B$. Applying Lemma~\ref{lem:basic} to $\tilde T$, we deduce that $A$ and $B$ are maximal compact subgroups. In particular $A$ and $B$ are vertex stabilisers for the $G$-action on $T$.
\medskip
We now show that the $G$-action on $\tilde T$ is locally primitive. To this end, suppose that there is a vertex $v \in V(\tilde T)$ for which the local action of $G_v$ on $E(v)$ is not primitive. Fix an edge $e = \{ v, w \} \in E(\tilde T)$. Since $G$ is edge-transitive, it follows that the $G_v$-action on $E(v)$ is transitive. However it is not primitive by assumption; we deduce that there is a subgroup $H_1 \leq G_v$ containing $G_e$ properly but which does not act transitively on $E(v)$. In particular $H_1 \neq G_v$.
Set $H_2 = G_w$ and consider the open subgroup $H = \langle H_1, H_2 \rangle$. In fact, we have $H = H_1 *_{G_e} H_2$ and the set $S = e \cdot H$ is a minimal $H$-invariant subtree of $\tilde T$, which is a Bass--Serre tree corresponding to the above decomposition of $H$. Since $H_1$ is not transitive on $E(v)$, we infer that $S$ is a proper subtree of $\tilde T$. In particular $H$ is properly contained in $G$. By construction the tree $S$ is unbounded, so that $H$ is non-compact. Thus we have shown that $G$ contains a proper non-compact open subgroup, which is absurd. This confirms that $G$ acts locally primitively on $\tilde T$ as claimed.
\medskip
It only remains to exhibit a $G$-equivariant embedding $\tilde T \to T$. To this end, we pick vertices $x, y$ of $T$ with $A = G_x$ and $B = G_y$ and such that $d(x, y)$ is minimal with respect to this property. Since $[x, y] \cdot G$ is a connected and $G$-invariant subset of $T$, it coincides with $T$ by minimality of the $G$-action. Since the $G$-action on $\tilde T$ is locally primitive, it easily follows that $A, B$ and $C$ are the only compact subgroups of $G$ which contain $C$ as a subgroup. Now for every element $h \in G$ such that $zh \in [x, y]$ for some $z \in ]x, y[$, the group $C = G_x \cap G_y$ is contained in the compact subgroup $G_z \cap h\inv G_z h$. This implies that $C = G_z$ and, in particular, that $z$ fixes $[x, y]$ pointwise. In other words, this shows that $[x, y] $ is a strict fundamental domain for the $G$-action on $T$. On the other hand it is clear that the edge $[\tilde x, \tilde y]$ is a strict fundamental domain for the $G$-action on $T$. Therefore, the assignments $\tilde x \mapsto x$ and $\tilde y \mapsto y$ extend to a $G$-equivariant map $\tilde T \to T$ which is isometric up to multiplying the metric on $\tilde T$ by a factor equal to $d(x, y)$, the presence of the latter factor accounting for the possibility that the segment $[x, y]$ be divided into several edges by some vertices of valence two.
\item[(ii) $\Rightarrow$ (iii)]
Immediate from Lemma~\ref{lem:LocCyclic}.
\item[(iii) $\Rightarrow$ (i)]
We may and shall assume without loss of generality that $T$ coincides with the Bass--Serre tree of the given amalgam decomposition of $G$.
Assume that for every vertex $v \in V(T)$, the induced action of $G_v$ on $E(v)$ is primitive and non-cyclic. For every pair of edges $e_1, e_2$ containing $v$, we have that both $G_{e_1}$ and $G_{e_2}$ are two different maximal subgroups of $G_v$, which are non-trivial since the $G_v$-action on $E(v)$ is not cyclic. We deduce that $G_v = \langle G_{e_1}, G_{e_2} \rangle$.
Now let $H$ be an arbitrary open subgroup of $G$, and suppose that $H$ is not compact. Then there is a minimal invariant subtree $S$ for $H$, which has no endpoints.
By Proposition~\ref{prop:Titsopen}(ii), $H$ contains every edge stabiliser $G_e$ with $e \in E(S)$.
Since $G$ is generated by the edge stabilisers and $H \neq G$, the equality $H=G$ will follow provided we show that $S=T$.
Since $S$ has no endpoint, it follows that for every vertex $v \in V(S)$, the star $E(v)$ contains at least two edges $e_1, e_2$ in $S$. By the previous paragraph, we have $G_v = \langle G_{e_1}, G_{e_2} \rangle \leq H$. Since $G_v$ acts transitively on $E(v)$, we deduce that for every $v \in V(S)$ we have $S_v = E(v)$. Clearly this implies that $S = T$, as desired.
\qedhere
\end{compactitem}
\end{proof}
\section{Groups with a prescribed local action}
\subsection{Burger-Mozes' universal group $U(F)$}\label{sec:BM}
We will now focus on a family of examples constructed by M.~Burger and Sh.~Mozes~\cite{Burger-Mozes1}.
Let $d>2$ be a positive integer, let $F \leq \Sym(d)$ be a permutation group on the set $\mathbf d = \{1, \dots, d\}$ and let $T$ be the regular tree of degree $d$. Pick a colouring $i \colon E(T) \to \mathbf d$ of the edge-set of $T$ by the elements of $\mathbf d$ such that its restriction to the star $E(v)$ around every vertex $v$ is a bijection. It is clear that this colouring is unique up to an automorphism of $T$. Let now $U(F)$ be the automorphism group defined by
\[ U(F) = \{g \in \Aut(T) \mid i \circ g \circ (i |_{E(v)})^{-1} \in F \text{ for all } v \in V(T) \} \,. \]
Let $U(F)^+$ denote the subgroup generated by the pointwise edge-stabilisers. One shows that $U(F)^+$ is edge-transitive if and only if $F$ is transitive and generated by its point stabilisers. In that case $U(F)^+$ has index two in $U(F)$ and it follows from Theorem~\ref{th:tits} that $U(F)^+$ is simple. Furthermore, this assumption also ensures that the group $G = U(F)^+$ acts locally as~$F$; in other words for each vertex $v \in V(T)$ the $G_v$-action on the star $E(v)$ is isomorphic to the $F$-action on $\mathbf d$. It is shown in~\cite[\S3.2]{Burger-Mozes1} that every vertex-transitive subgroup of $\Aut(T)$ whose vertex stabilisers act locally like $F$ (on the star of the fixed vertex) is conjugate to a subgroup of $U(F)$.
\begin{quote}\em
We will assume from now on that $F$ is transitive and generated by its point stabilisers.
\end{quote}
\subsection{Open subgroups of $U(F)^+$}
We retain the notation and assumptions of Section~\ref{sec:BM}. Let us moreover denote by $G$ the simple group $U(F)^+$.
\begin{prop}\label{pr:openU(F)}
We have the following.
\begin{enumerate}[\rm (i)]
\item
The group $F$ is primitive if and only if every proper open subgroup of $G$ is compact.
\item
Suppose that $F$ is imprimitive, with maximal blocks of imprimitivity of cardinality~$k$,
and assume moreover that $F$ acts regularly on each such block.
Then for each edge $e \in E(T)$, the quotient $\norma_G(G_e)/G_e$ is virtually free.
Moreover, if $k \geq 3$, then $G$ possesses open subgroups which are not compactly generated.
\end{enumerate}
\end{prop}
\begin{proof}
\begin{compactenum}[\rm (i)]
\item
This follows immediately from Theorem~\ref{thm:LocPrim}.
\item
We shall use the following definition. For each subset of colours $\mathbf{b} \subseteq \mathbf{d}$, we define a {\bf $\mathbf{b}$-tree}
to be a subtree $S \subseteq T$ which only uses colours from $\mathbf{b}$, and which is maximal
with respect to this property, \emph{i.e.}\@ for each vertex $v \in V(S)$ we have $i(S \cap E(v)) = \mathbf{b}$.
Let now $e$ be an arbitrary edge of $T$, and let $\mathbf{b} \subseteq \mathbf{d}$ the unique block of imprimitivity
containing the colour of $e$.
Denote the $\mathbf{b}$-tree containing $e$ by $S$;
then $S$ is a regular tree of degree $|\mathbf{b}| = k$.
By the regularity condition imposed on $F$, every element of $G_e$ fixes the tree $S$ elementwise,
and in fact $G_e = G_f$ for every $f \in E(S)$.
On the other hand, the maximality of the blocks implies that a point stabiliser $F_a$ does not fix any point
outside the block containing $a$, and this implies that $G_e \neq G_f$ for every $f \not\in E(S)$.
We conclude that $\norma_G(G_e)$ is equal to the global stabiliser $G_S$ of the subtree $S$, and clearly
$\norma_G(G_e)$ acts edge-transitively on $S$.
Hence the discrete group $\norma_G(G_e) / G_e$ acts transitively and properly (\emph{i.e.}\@ with finite vertex stabilisers) on $S$, and
Bass--Serre theory implies that $\norma_G(G_e) / G_e$ is virtually free.
Moreover, it is virtually abelian free if and only if the tree $S$ is a line or a point, \emph{i.e.}\@ if $k < 3$. In particular, if $k\geq 3$, then $\norma_G(G_e) / G_e$ contains subgroups that are not finitely generated.
Lifting such a group back to a subgroup of $\norma_G(G_e)$ provides us with an open subgroup of $G$ which is
not compactly generated.
\qedhere
\end{compactenum}
\end{proof}
\begin{remark}
A topological group is called \textbf{Noetherian} if it satisfies the ascending chain condition on open subgroups. A locally compact group is Noetherian if and only if every open subgroup is compactly generated. In particular, it follows from Proposition~\ref{pr:openU(F)} that compactly generated elements of $\mS$ need not be Noetherian.
\end{remark}
\subsection{Local structure}\label{sec:LocalStr}
By the {\bf local structure} of $G$, we mean the properties shared by all compact open subgroups of $G$; such properties are invariant up to commensurability.
The structure of vertex-stabilisers in $G$ may be described in terms of the finite group $F$.
More precisely, the vertex-stabilisers have the structure of an infinitely iterated wreath product of finite groups:
\begin{prop}[{\cite[Section~3.2]{Burger-Mozes1}}]\label{pr:wr}
Let $F \leq \Sym(d)$.
The maximal compact subgroup $U(F)^+_v$ of $U(F)^+$
is obtained as the projective limit $\varprojlim A_n$ of the
groups
\[ \begin{cases}
A_0 = F \,; \\
A_n = F_a \wr A_{n-1} & \text{for } n \geq 1,
\end{cases} \]
where $F_a \leq F$ denotes a point-stabiliser. Similarly, for an edge stabiliser $U(F)^+_e$ we have
$U(F)^+_e \cong \varprojlim D_n \times \varprojlim D_n \cong \varprojlim (D_n \times D_n)$ with
\[ \begin{cases}
D_1 = F_a \,; \\
D_n = F_a \wr D_{n-1} = D_{n-1} \wr F_a & \text{for } n \geq 2 \,.
\end{cases} \]
\end{prop}
We warn the reader that all wreath products are considered with their imprimitive wreath product action,
and that the point stabilisers $F_a$ are considered with their action on $(d-1)$ elements
(even if this action has fixed points). We shall come back to the profinite group $D = \varprojlim D_n$ in Section~\ref{sec:branch} below.
We will now investigate the extent to which the group $F$ is determined by the local or global structure of $G$.
\begin{theorem}\label{th:global}
\begin{compactenum}[\rm (i)]
\item
Let $F, F' \leq \Sym(d)$ be two groups satisfying the conditions of Section~\textup{\ref{sec:BM}},
and assume that $U(F)^+ \cong U(F')^+$.
Then $F \cong F'$ as permutation groups, \emph{i.e.}\@ $F$ and $F'$ induce equivalent permutation representations on $\mathbf{d}$.
\item
Let $G' \in \mathcal{S}$ be a group acting continuously, properly and edge-transitively on some tree $T$,
and assume that $G' \cong G = U(F)^+$ for some group $F \leq \Sym(d)$ satisfying the conditions of Section~\textup{\ref{sec:BM}}.
Then $G$ and $G'$ have equivalent actions on the tree $T$ of degree $d = |\mathbf{d}|$.
\end{compactenum}
\end{theorem}
\begin{proof}
Let $F \leq \Sym(d)$ be a group satisfying the conditions of Section~\ref{sec:BM}.
We will show that we can recover $F$ from the group $G = U(F)^+$ alone, and that the local action
around every vertex of the tree is given by the action of $F$ on $\mathbf{d}$;
this will simultaneously prove (i) and (ii).
By Lemma~\ref{lem:basic}(viii), we know that the number $d$, which is the degree of the tree $T$ on which $G$ acts,
is uniquely determined by $G$.
Let $K$ be an arbitrary maximal compact subgroup of $G$, and let $L$ be another
maximal compact subgroup of $G$ minimizing the index $[K : K \cap L]$.
Define
\[ C = \bigcap_{g \in K} g\inv (K \cap L) g. \]
We claim that $K/C \cong F$ and that the conjugation action of $K/C$ on
the set of conjugates of $K \cap L$ in $K$
is equivalent to the permutation action of $F$ on~$\mathbf{d}$.
By Lemma~\ref{lem:basic}(vii), $K$ and $L$ are two (different)
vertex stabilisers, say $K = G_v$ and $L = G_w$ with $v,w \in V(T)$.
Let $z$ be the unique neighbour of $v$ lying on the unique path from $v$ to $w$;
then $G_v \cap G_w \leq G_v \cap G_z$, so by the minimality of the index
$[K : K \cap L]$, we have $K \cap L = G_v \cap G_z$.
Since $K = G_v$ acts transitively on the star $E(v)$, we see that
$C$ is equal to the pointwise stabiliser of $E(v)$, \emph{i.e.}\@ it is the kernel
of the action of $K$ on $E(v)$.
We conclude that $K/C \cong F$, and the action of $K/C$ on $E(v)$ is precisely
given by the action of $F$ on $\mathbf{d}$.
\end{proof}
We can summarise the relations between the structure of $F$ (as a permutation group), $U(F)^+$ (as a topological group),
and $U(F)^+_v$ (as the commensurability class of the profinite group $U(F)^+_v$) in the following diagram.
\[ \bfig
\node a(800,400)[U(F)^+_v]
\node b(0,800)[F]
\node c(0,0)[U(F)^+]
\arrow|l|/@<2pt>/[c`b;\text{Thm.~\ref{th:global}}]
\arrow|r|/@<2pt>/[b`c;\text{def.}]
\arrow|l|/@<2pt>/[a`b;\text{\bf\normalsize ?}]
\arrow|r|/@<2pt>/[b`a;\text{\ Prop.~\ref{pr:wr}}]
\arrow|l|/@<-2pt>/[a`c;\text{\bf\normalsize ?}]
\arrow|r|/@<-2pt>/[c`a;\text{\quad \ cpt.\@ open subgps.}]
\efig \]
The two remaining question marks correspond precisely to George Willis's question~\cite[Problem~4.3]{Willis07}.
It turns out that already for this class of groups, the answer to this question is negative.
More precisely, we will show the existence of two non-isomorphic groups $U(F)^+ \not\cong U(F')^+$ with isomorphic
edge stabilisers $U(F)^+_e \cong U(F')^+_{e}$.
\begin{prop}
Assume that $F$ and $F'$ are two non-isomorphic subgroups of $\Sym(d)$ with equal point stabilisers $F_a = F'_a$
acting on $d-1$ elements.
Then $U(F)^+_e \cong U(F')^+_e$ but $U(F)^+ \not\cong U(F')^+$,
\emph{i.e.}\@, the groups $U(F)^+$ and $U(F')^+$ are locally isomorphic but not isomorphic.
\end{prop}
\begin{proof}
Since $F_a$ and $F'_a$ are isomorphic permutation groups, Proposition~\ref{pr:wr} implies that $U(F)^+_e \cong U(F')^+_e$.
On the other hand, $F \not\cong F'$, and it follows from Theorem~\ref{th:global}(i) that $U(F)^+ \not\cong U(F')^+$.
\end{proof}
We now point out that we can indeed find two non-isomorphic subgroups $F, F' \leq \Sym(d)$ satisfying the conditions of
Section~\ref{sec:BM} (\emph{i.e.}\@ transitive and generated by their point stabilisers) that have equal point stabilisers.
The following example is the smallest possible (in terms of the permutation degree $d$).
\begin{example}
Let $d=8$, let $F = \PSL(2,7) \leq \Sym(8)$ with its natural action on the projective line $\GF(7) \cup \{ \infty \}$,
and let $F' = \AGammaL(1,8) \cong \AGL(1,8) \rtimes \langle \sigma \rangle \leq \Sym(8)$
with its natural action on the affine line $\GF(8)$,
where $\sigma$ is the generator of $\Gal\bigl(\GF(8)/\GF(2)\bigr) \cong C_3$.
Then $F$ and $F'$ are two non-isomorphic doubly transitive permutation groups, and they have isomorphic point stabilizers
$F_\infty \cong F'_0 \cong C_7 \rtimes C_3$ with an equivalent permutation action.
Indeed, the equivalence of the permutation representation is induced by the bijection
\[ \beta \colon \GF(7) \to \GF(8)^\times \colon n \mapsto \zeta^n \,, \]
where $\zeta$ is a generator of the multiplicative group $\GF(8)^\times$.
We conclude that the groups $U(F)^+$ and $U(F')^+$ are two non-isomorphic groups
with isomorphic edge stabilizers and commensurable vertex stabilizers.
\end{example}
\section{Rigidity for simple t.d.l.c.\@ groups}
\subsection{Germs of automorphisms}\label{sec:Germ}
We recall some elements of terminology which were already defined in the introduction.
Let $\mS$ denote the class of non-discrete topologically
simple totally disconnected locally compact groups. We say that two
groups $G_1, G_2 \in \mS$ are {\bf locally isomorphic} if
they contain isomorphic compact open subgroups. Moreover, a group $G
\in \mS$ will be called {\bf Lie-reminiscent} if any $H \in
\mS$ locally isomorphic to $G$ is in fact (globally)
isomorphic to $G$.
Any two compact open subgroups of a totally disconnected locally compact group $G$ are
commensurable. In particular, the commensurability class of any
compact open subgroup of $G$ depends only on its local structure,
in the sense that it can be reconstructed from any identity
neighbourhood. This commensurability class determines an algebraic object, namely the group $\acomm(G)$ of \textbf{germs of automorphisms}. This is defined as the quotient of the set of all isomorphisms $f: U \to V$ between compact open subgroups of $G$, divided by the equivalence relation which identifies two isomorphisms $f_1: U_1 \to V_1$ and $f_2: U_2 \to V_2$ if they coincide on some open subgroup of $U_1 \cap U_2$.
Notice that if $G$ is compact, hence profinite, then $\acomm(G)$ is nothing but the group of {\bf abstract commensurators} of $G$, which we denote by $\mathrm{Comm}(G)$. In fact, for any totally disconnected locally compact group $G$, we have $\acomm(G) =\acomm(U)$ for any compact open subgroup $U \leq G$. In particular $\acomm(G)$
depends only on the commensurability class of compact open subgroups of $G$.
Since $G$ commensurates its compact open subgroups,
there is a canonical homomorphism
\[ \comm \colon G \to \acomm(G) \colon g \mapsto \comm(g).\]
Following \cite{BarneaErshovWeigel}, we endow $\acomm(G)$ with the \textbf{strong topology}, which is defined as the finest group topology which makes the homomorphism $\comm$ continuous. In this way $\acomm(G)$ is a topological group, which need not be Hausdorff in general.
\begin{prop}\label{prop:Germ:basic}
Let $G$ be a totally disconnected locally compact group.
\begin{enumerate}[\rm (i)]
\item The kernel of $\comm : G \to \acomm(G)$ is the quasi-centre $\QZ(G)$ of $G$.
\item If $\QZ(G)$ is closed, then $\acomm(G)$ is totally disconnected and locally compact.
\item If $\QZ(G)$ is discrete, then $G$ is locally isomorphic to $\acomm(G)$.
\end{enumerate}
\end{prop}
\begin{proof}
\begin{compactitem}
\item[(i)] Immediate from the definition of $\QZ(G)$, see Definition~\ref{def:QZ}.
\item[(ii)] By definition of the strong topology, the group $\comm(G)$ is open in $\acomm(G)$. By (i) it is isomorphic to $G/\QZ(G)$, which is totally disconnected and locally compact provided $\QZ(G)$ is closed. In that case, the profinite identity neighbourhoods of $G/\QZ(G)$ are also identity neighbourhoods of $\acomm(G)$. The desired result follows.
\item[(iii)] If $\QZ(G)$ is discrete, then there is some compact open subgroup $U \leq G$ such that $U \cap \QZ(G) = 1$. Thus $G$ and $G/\QZ(G)$ are locally isomorphic and the result follows from (i). \qedhere
\end{compactitem}
\end{proof}
We emphasize that the quasi-centre need not be closed in general, even for simple groups. In fact, the examples of non-compactly generated simple groups
constructed by G.~Willis in \cite[\S3]{Willis07} have a dense
quasi-centre. However, the situation is more favorable in the case of compactly generated topologically simple groups, as illustrated by the following result due to Barnea--Ershov--Weigel (see also \cite[Prop.~4.3]{Caprace-Monod-monolith}).
\begin{prop}[{\cite[Theorem~4.8]{BarneaErshovWeigel}}]\label{prop:QZ}
Any compactly generated totally disconnected locally compact group with dense quasi-centre admits a basis of identity neighbourhoods consisting of compact open {normal} subgroups. In particular, if $G \in \mS$ and if $G$ is compactly generated, then $\QZ(G) = 1$.
\end{prop}
Let $\Aut(G)$ denote the group of homeomorphic automorphisms of $G$.
Clearly such automorphisms also commensurate compact open subgroups
of $G$, so that we get a canonical homomorphism
\[ \kappa \colon \Aut(G) \to \acomm(G) \,; \]
note that $\comm$ factors through $\kappa$ via the map $G \to \Inn(G) \leq \Aut(G)$.
Following the terminology introduced in
\cite{BarneaErshovWeigel}, we say that a simple group $G \in
\mS$ is {\bf rigid} if $\kappa$ is bijective. Equivalently, this means that every germ of automorphism of $G$ extends to a unique global automorphism.
It is useful to have several equivalent ways to express rigidity at our disposal.
\begin{prop}[\cite{BarneaErshovWeigel}]\label{pr:rigid}
Let $G \in \mS$ such that $\QZ(G)=1$.
Then the following conditions are equivalent:
\begin{enumerate}[\rm (a)]
\item
$G$ is rigid, \emph{i.e.}\@ $\kappa$ is bijective.
\item
$\kappa$ is surjective.
\item
$\comm(G)$ is a normal subgroup of $\acomm(G)$.
\item
Any isomorphism between a pair of compact open subgroups of $G$ can be extended to a unique automorphism of $G$.
\end{enumerate}
\end{prop}
\begin{proof}
Condition (d) is precisely the definition of rigidity as in \cite{BarneaErshovWeigel}, and hence
\cite[Proposition~3.7]{BarneaErshovWeigel} shows that (a), (c) and (d) are equivalent.
Of course (a) implies (b).
Assume finally that (b) holds; then $\comm(G) = \kappa(\Inn(G)) \unlhd \kappa(\Aut(G)) = \acomm(G)$,
so (c) holds.
\end{proof}
Our next goal is to clarify the relation between the notion of rigidity and of Lie-reminiscence for the elements of $\mS$. We shall need the following two lemmas.
\begin{lemma}\label{lem:QZopen}
Let $G \in \mS$ be compactly generated. Then every open subgroup of $\acomm(G)$ has trivial quasi-centre. In particular $\acomm(\acomm(G)) = \acomm(G)$.
\end{lemma}
\begin{proof}
Since $G \in \mS$ is compactly generated, we have $\QZ(G) = 1$ and hence $\QZ(U)=\Comm(U)=1$ for any compact open subgroup $U \leq G$. By \cite[Prop.~3.2(c)]{BarneaErshovWeigel}, it follows that $\QZ(\acomm(G)) = \QZ(\Comm(U)) = 1$. The first assertion follows since $\QZ(\acomm(G))$ contains the quasi-centre of any open subgroup of $\acomm(G)$.
Since $\acomm(G)$ depends only on the commensurability class of the compact open subgroups of $G$, the second statement now follows from Proposition~\ref{prop:Germ:basic}(iii).
\end{proof}
\begin{lemma}\label{le:GH}
Let $G,H \in \mS$ be two locally isomorphic compactly generated groups.
Then $\langle \comm(G), \comm(H) \rangle \leq \acomm(G) = \acomm(H)$ is
topologically simple.
\end{lemma}
\begin{proof}
Since $G$ and $H$ are locally isomorphic, we have $\acomm(G) = \acomm(H) = \acomm$;
recall that $\comm(G)$ and $\comm(H)$ are open subgroups of $\acomm$.
Now let $N \neq 1$ be a closed normal subgroup of $\langle \comm(G), \comm(H) \rangle$.
If $N \cap \comm(G)$ were trivial, then $N$ would be a discrete subgroup of $\langle \comm(G), \comm(H) \rangle$, which is impossible by Lemma~\ref{lem:QZopen}.
Since $\comm(G) \cong G$ is topologically simple, this implies $\comm(G) \leq N$, and similarly
$\comm(H) \leq N$.
We conclude that $N = \langle \comm(G), \comm(H) \rangle$ as claimed.
\end{proof}
\begin{cor}
Let $G, H \in \mS$ be compactly generated. If $G$ and $H$ are
locally isomorphic, then there is a compactly generated group $S \in
\mS$ in which both $G$ and $H$ embed as open subgroups.
\end{cor}
\begin{proof}
This immediately follows from Lemma~\ref{le:GH} since $\comm$ is injective because $\QZ(G)$ and $\QZ(H)$ are trivial, see Proposition~\ref{prop:QZ}.
\end{proof}
In view of Proposition~\ref{prop:QZ}, the following result applies notably to any compactly generated group in $\mS$.
\begin{prop}\label{prop:AdmissibleRigid}
Let $G \in \mS$ be such that $\QZ(G)= 1$. Then:
\begin{enumerate}[\rm (i)]
\item Up to isomorphism, there is a unique rigid simple group $\tilde G \in \mS$ which is locally isomorphic to $G$; it can be defined as the intersection of all non-trivial closed normal subgroups of $\acomm(G)$.
\item If $G$ is Lie-reminiscent, then it is rigid.
\item If $G$ is rigid, then any compactly generated group $H \in \mS$ locally
isomorphic to $G$ is isomorphic to some open subgroup of $G$.
\end{enumerate}
\end{prop}
It is important to keep in mind that the canonical rigid simple group $\tilde G$ need not be compactly generated, even if $G$ is so. This is notably illustrated by Theorem~\ref{thmi:L(G)} from the introduction.
\begin{proof}[Proof of Proposition~\ref{prop:AdmissibleRigid}]
\begin{compactenum}[(i)]
\item
Let $\tilde G \leq \acomm(G)$ denote the intersection of all non-trivial closed normal subgroups of $\acomm(G)$.
Notice that any such normal subgroup is non-discrete because $\QZ(\acomm(G)) = 1$ by Lemma~\ref{lem:QZopen},
and hence meets the open subgroup $\comm(G)$ non-trivially.
Since $G \in \mS$, it follows that every non-trivial closed normal subgroup of $\acomm(G)$ contains $\comm(G)$ and is open;
in particular $\tilde G$ contains $\comm(G)$.
We next show that $\tilde G$ belongs to $\mS$ or, in other words, that it is topologically simple. So let $N$ be a closed normal subgroup of $\tilde G$. As before, Lemma~\ref{lem:QZopen} ensures that $N$ is non-discrete. Since $\tilde G$ contains the topologically simple group $\comm(G)$ as an open subgroup, we deduce that the intersection $N \cap \comm(G)$ is non-trivial. But $\comm(G)$ being topologically simple, we infer that $\comm(G) \subseteq N$. In other words, every closed normal subgroup of $\tilde G$ contains $\comm(G)$;
hence the same is true for the intersection $M$ of all these closed normal subgroups. Thus $M$ is an open characteristic subgroup of $\tilde G$. In particular, it is a closed normal subgroup of $\acomm(G)$. It then follows from the definition that $M = \tilde G$,
whence $\tilde G$ is topologically simple as desired.
The fact that $\tilde G$ is rigid follows from Proposition~\ref{pr:rigid} since $\tilde G$ is normal in $\acomm(G) = \acomm(\tilde G)$ by construction.
Let now $O \in \mS$ be any rigid group locally isomorphic to $G$. Since $O$ is topologically simple, its quasi-centre is either trivial or dense. Since $\QZ(G) =1$, every compact open subgroup of $G$ has trivial quasi-centre. The quasi-centre of $O$ can therefore not be dense. In particular $O$ embeds in $\acomm(O) \cong \acomm(G)$, see Proposition~\ref{prop:Germ:basic}. Let us identify $O$ with its image in $\acomm(G)$. Proposition~\ref{pr:rigid} guarantees that $O$ is open and normal; it therefore contains $\tilde G$. But $\tilde G$ is normal in $\acomm(G)$, and hence also in $O$. Since $O$ is simple, we infer that $O = \tilde G$, as desired.
\item
Let $\tilde G \in \mS$ be the rigid group provided by (i). Since $G$ and $\tilde G$ are locally isomorphic, we infer that if $G$ is Lie-reminiscent, then it must be isomorphic to $\tilde G$. Hence $G$ is rigid, as desired.
\item
Assume now that $G$ is rigid. Then $\comm(G)$ is normal in $\acomm(G)$ and any compactly generated group $H \in \mS$ locally
isomorphic to $G$ embeds as an open subgroup in $\acomm(G)$
via the injection $\comm \colon H \to \acomm(H) = \acomm(G)$.
Therefore the open subgroups $\comm(H)$ and $\comm(G)$ of $\acomm(G)$ meet
non-trivially, and since $\comm(G)$ is normal while $H$ is
topologically simple, we obtain $\comm(H) \subseteq \comm(G)$ as desired.
\qedhere
\end{compactenum}
\end{proof}
\begin{remark}
It is interesting to point out that the above proof implies moreover that, if $G$ is \emph{abstractly} simple, then so is $\tilde G$.
\end{remark}
\begin{remark}
The group $\tilde G$ is called the \textbf{open normal core} of $\acomm(G)$ in \cite{BarneaErshovWeigel}, and it is denoted by $\mathrm{Onc}(\Comm(U)_S)$, where $U$ is a compact open subgroup of $G$. It is defined in \emph{loc.~cit.} as the intersection of all open normal subgroups of $\acomm(G) = \Comm(U)$; the above proposition shows that these two definitions coincide in our setting.
\end{remark}
\subsection{A local property of compactly generated rigid simple groups}
As mentioned above, the rigid group $\tilde G$ appearing in Proposition~\ref{prop:AdmissibleRigid}(i) need not be compactly generated in general, even if $G$ is so. We will study this question in detail for Burger--Mozes universal groups in order to prove Theorem~\ref{thmi:L(G)} from the introduction. In that the study, the following general fact will be helpful; it is implicitly contained in Section~8 from \cite{BarneaErshovWeigel}.
\begin{prop}\label{prop:Aut}
Let $G \in \mS$ be compactly generated and rigid. Then for any compact open subgroup $U \leq G$ and any characteristic open subgroup $V \leq U$, the automorphism group $\Aut(U)$ embeds as a subgroup of $\Aut(V)$ whose index is at most countable.
\end{prop}
\begin{proof}
Since $V$ is characteristic in $U$, there is a map $\Aut(U) \to \Aut(V)$. Since $G$ has trivial quasi-centre (see Proposition~\ref{prop:QZ}), so does $U$ and hence the above map is injective. All we need to show is that its image has countable index.
We shall identify the groups $G$, $\Aut(U)$ and $\Aut(V)$ with their canonical images in $\mathscr L(G)$. Since $G$ is compactly generated, every open subgroup has countable index. In particular the normaliser of $U$ has countable index in $G$, and hence $U$ has countably many conjugates in $G$.
Since $G$ is normal in $\mathscr L(G)$ (see Proposition~\ref{pr:rigid}), every conjugate of $U$ in $\mathscr L(G)$ is contained in $G$. Lemma~\ref{lem:QZopen} guarantees that the normaliser of $U$ in $\mathscr L(G)$ is nothing but $\Aut(U)$. Therefore, we infer that $\Aut(U)$ has countable index in $\mathscr L(G)$. The latter index being clearly an upper bound for the index of $\Aut(U)$ in $\Aut(V)$, the desired result follows.
\end{proof}
\subsection{Non-rigidity of tree-automorphism groups satisfying Tits' independence property}
\begin{theorem}\label{thm:NonRigid}
Let $T$ be a locally finite tree and $G \leq \Aut(T)$ be a non-trivial compactly generated simple closed subgroup satisfying Tits' independence property. Then $G$ is not rigid, and in particular it is not Lie-reminis\-cent.
\end{theorem}
\begin{proof}
Upon replacing $T$ by a minimal $G$-invariant subtree, there is no loss of generality in assuming the $G$-action to be minimal. In particular it is cocompact by Lemma~\ref{le:ccpt}.
We need to consider a certainly family of sub-trees of $T$, associated to each pair $(v, A)$ consisting of a vertex $v \in V(T)$ and a set $A \subseteq E(v)$ of edges containing $v$. To such a pair $(v, A)$, we associate the subtree
$$h(v, A) \subseteq T$$
whose vertex set is
$$\{v\} \cup \{w \in V(T) \; | \; [v, w] \cap A \neq \varnothing\}.$$
In other words, $h(v, A)$ is the subtree containing $v$ whose vertices different from $v$ are separated from $v$ by an edge in $A$.
In particular, $h(v, E(v)) = T$ and $h(v, \varnothing) = \{v\}$. Similarly, if $A =\{e\}$ consists of a single edge, then $h(v, \{e\})$ is the union of $e$ with the half-tree determined by $e$ and not containing $v$.
Since $G$ acts cocompactly on $T$, it has finitely many orbits of vertices and edges and, hence, finitely many orbits of subtrees of the form $h(v, A)$ as above.
We denote these orbits by $\Omega^1, \dots, \Omega^k$.
Given $h \in \Omega^i$, we denote by $G_{(h)}$ the pointwise stabiliser of the subtree $h$.
Fix a base vertex $v \in T$.
Given $n>0$, we denote by $U_n \leq G$ the pointwise stabiliser of the ball of radius $n$ around $v$.
Let $T_n$ be the fixtree of $U_n$, and let $T_n^{[0]}$ be the \emph{thick part} of $T_n$, i.e.\@ the subset consisting of those vertices of $T_n$ all of whose neighbours in $T$ also belong to $T_n$. In particular, the $(n-1)$\nobreakdash-ball around $v$ is contained in $T_n^{[0]}$. Notice that $T_n^{[0]}$ need not be connected \emph{a priori}. Moreover $T_n \setminus T_n^{[0]}$ is necessarily non-empty, since otherwise we would have $T_n = T_n^{[0]} = T$ and $G$ would be discrete, hence trivial.
In fact, we claim more precisely that $T_n$ coincides with the convex hull of $T_n \setminus T_n^{[0]} $. Indeed, consider an edge $e \in E(T_n)$, and let $h^+$ and $h^-$ be the two half-trees of $T$ determined by $e$. Then we claim that $h^+$ is not entirely contained in $T_n$.
Indeed, suppose that $h^+$ were contained in $T_n$; then $U_n$ would fix $h^+$ pointwise.
But then the pointwise stabilizer $G_{(h^+)}$ would be open; this would force the pointwise stabilizer $G_{(h^-)}$ to be finite, since $G_{(e)} \cong G_{(h^+)} \times G_{(h^-)}$ by Tits' independence property. Since $G$ has trivial quasi-centre by Proposition~\ref{prop:QZ}, we deduce that $G_{(h^-)}$ is trivial.
It is easy to see that this would imply that $G$ itself is trivial, which is absurd. Thus $T_n$ does not contain $h^+$. This is equivalent to saying that $h^+$ contains some vertex $v^+ \in T_n \setminus T_n^{[0]}$. Similarly, one shows that $h^-$ contains some vertex $v^- \in T_n \setminus T_n^{[0]}$. Thus $e \subset [v^+, v^-]$. Since $e$ was an arbitrary edge of $T_n$, this confirms the claim that $T_n$ is the convex hull of $T_n \setminus T_n^{[0]}$.
\medskip
For each vertex $v \in T_n \setminus T_n^{[0]}$, the intersection $A = E(v) \cap E(T_n)$ is a proper subset of $E(v)$. In particular $h\big(v, E(v) \cap E(T_n)\big)$ is a proper subtree of $T$ containing $T_n$. We define
$$H_n = \bigg\{ h\big(v, E(v) \cap E(T_n)\big) \; | \; v \in T_n \setminus T_n^{[0]}\bigg\}.$$
as the collection of all subtrees of that form. Moreover, for each $i$ and each $n$, we set
$\Omega_n^i = \Omega^i \cap H_n$. Notice that $\Omega_n^i$ can be finite or infinite.
Now Tits' independence property implies that
\[ U_n \cong \prod_{h \in H_n} G_{(h)} = \prod_{i=1}^k \Bigl( \prod_{h \in \Omega_n^i} G_{(h)} \Bigr) . \]
Set $N_n = \norma_G(U_n)$; then $U_n \leq G_v \leq N_n$.
Moreover, let
\begin{multline*}
A_n = \{ \varphi \in \Aut(G) \mid \varphi(U_n) = U_n, \text{ and for all } i \in \{ 1,\dots,k \} \\[-.2ex]
\text{ and all } h \in \Omega_n^i, \text{ we have } \varphi(G_{(h)}) = G_{(h')} \text{ for some } h' \in \Omega_n^i \} .
\end{multline*}
Since $G$ is simple it injects in $\Aut(G)$ and it will be convenient to abuse notation and identify $G$ with its image in $\Aut(G)$.
Modulo this convention, we notice that $N_n$ is contained in $A_n$: Indeed, the group $N_n$ acts on the fixtree $T_n$. Therefore, it preserves $T_n^{[0]}$ and, hence, it permutes the elements of $\Omega_n^i$ for all $i \in \{1, \dots, k\}$. Thus $N_n \leq A_n$. Since on the other hand, the definition of $A_n$ implies that $A_n \cap G$ normalizes $U_n$, we conclude that $N_n = A_n \cap G$, which implies in particular that $N_n$ is a normal subgroup of $A_n$.
By definition we have a canonical homomorphism
\[ \pi_n^i : A_n \to \Sym(\Omega_n^i) \]
for each $i = 1, \dots, k$. The product of these defines a homomorphism
\[ \pi_n : A_n \to \prod_{i=1}^k \Sym(\Omega_n^i). \]
Observe that $\ker(\pi_n) \cap N_n $ acts trivially on the set $T_n \setminus T_n^{[0]}$. Since $T_n$ is the convex hull of $T_n \setminus T_n^{[0]}$ by the above, we deduce that $\ker(\pi_n) \cap N_n$ acts trivially on $T_n$, and is thus contained in $U_n$. This shows that $\ker(\pi_n) \cap N_n = U_n $ for all $n$. Moreover, since $U_n \neq G_v$ for all sufficiently large $n$,
this implies that $\pi_n(N_n)$ is non-trivial for all sufficiently large $n$.
\smallskip
Suppose now for a contradiction that $G$ is rigid. Then every abstract commensurator of $U_n$ extends to an automorphism of $G$. In particular any automorphism of $U_n$ permuting isomorphic factors extends to an element of $A_n$. This implies that the above map $\pi_n^i$ is surjective for all $i$ and $n$.
Let $i$ and $n$ be such that $\Omega_n^i$ is infinite (and hence countably infinite), and assume that $\pi_n^i(N_n) \neq 1$. Recall that $G$ is second countable since it is metrisable and compactly generated. In particular $N_n$ is second countable, and the discrete image $\pi^i_n(N_n)$ is therefore at most countable. In particular it is a countable normal subgroup of the uncountable group $\Sym(\Omega_n^i)$.
By the Baer--Schreier--Ulam theorem \cite{Baer,SchreierUlam},
$\Sym(\Omega_n^i)$ has only two proper non-trivial normal subgroups,
namely the subgroup of all finitary permutations, and the subgroup of alternating finitary permutations;
both normal subgroups are locally finite and infinite. We deduce that $\pi_n^i(N_n)$ is locally finite and, hence, that $N_n$ is \textbf{locally elliptic},
\emph{i.e.}\@ every compact subset is contained in a compact subgroup. Since $N_n $ is open in $G$, Lemma~\ref{lem:basic}(iv) guarantees that this can only be true if $N_n$ is compact;
but then $\pi_n^i(N_n)$ is finite, so $\pi_n^i(N_n)$ is trivial after all.
We infer that $\pi_n^i(N_n) = \{1\}$ for all $i$ and $n$ such that $\Omega_n^i$ is infinite;
in particular $\pi_n(N_n) $ is finite for all $n$.
Similarly, if the group $\pi_n^i(N_n)$ is non-trivial and $|\Omega_n^i | \geq 5$, then it coincides with either $\Sym(\Omega_n^i)$ or $\Alt(\Omega_n^i)$.
\medskip
Our next goal is to show that $\pi_n^i(G_v)$ is either $\Sym(\Omega_n^i)$ or $\Alt(\Omega_n^i)$ for some fixed $i$ and infinitely many values of $n$. To this end, we first notice that $N_n$ is compact, since $\ker(\pi_n) \cap N_n = U_n$ is compact and since $\pi_n(N_n) $ is finite. Thus for each $n$ there is some vertex $v_n$ such that $N_n \leq G_{v_n}$, and hence $G_v \leq N_n \leq G_{v_n}$. Since $G$ is unimodular (because it is simple) and acts cocompactly on $T$, we deduce that
the values $[G_{v_n} : G_{v}]$ are bounded, and hence
\[ s = \sup_n [N_n : G_v] < \infty. \]
Recall that, for $m \geq 5$, the only non-trivial subgroup of $\Sym(m)$ of index $< m$ is $\Alt(m)$, and that $\Alt(m)$ has no non-trivial subgroup of index~$<m$. It follows that for sufficiently large $n$, if $|\Omega_n^i| > \max\{5, s\}$ and $\pi_n^i(N_n)$ is non-trivial, then
$\pi_n^i(G_v)$ also coincides with either $\Sym(\Omega_n^i)$ or $\Alt(\Omega_n^i)$.
On the other hand, we have
\[ |\pi_n(G_v)| = |G_v / U_n| \leq |\pi_n(N_n)| , \]
which shows that $ |\pi_n(N_n)|$ tends to infinity with $n$. In particular, the subset $I \subset \{1, \dots, k\}$ consisting of those $i$ such that $| \pi^i_n(N_n) | > \min\{60, s!\}$ for infinitely many values of $n$, is non-empty. In view of the conclusion of the preceding paragraph, we deduce that if $i \in I$, then $\pi^i_n(G_v) = \Sym(\Omega_n^i)$ or $\Alt(\Omega_n^i)$ for infinitely many values of $n$, as desired.
\medskip
The final contradiction is now obtained as follows. Since $G_v$ preserves each sphere around $v$, it follows that for each $i$ and $n$ with $\pi^i_n(G_v) = \Sym(\Omega_n^i)$ or $\Alt(\Omega_n^i)$, there is some $m \geq n$ such that the permutation representation $\pi_n^i : G_v \to \Sym(\Omega_n^i)$ is a sub-representation of the permutation action of $G_v$ on the $m$-sphere around $v$. The latter action is imprimitive, with minimal blocks of imprimitivity of size at most $d-1$, where $d$ is the maximal valence of a vertex of $T$;
in particular, the number of blocks of the sphere of radius $m$ tends to infinity with $m$.
On the other hand, the number of sets $\Omega_n^i$ is bounded by $k$, for each $n$.
Therefore, the pigeonhole principle implies that there is some $i \in I$ for which $\pi_n^i(G_v)$ is an imprimitive subgroup of $\Sym(\Omega_n^i)$ for all $n$, whose order tends to infinity with $n$.
This finally contradicts the fact that for this $i$ and some $n$ large enough, we have $\pi^i_n(G_v) = \Sym(\Omega_n^i)$ or $\Alt(\Omega_n^i)$.
\end{proof}
\section{Commensurators of self-replicating wreath branch groups}\label{sec:Comm}
\subsection{The vocabulary of branch groups and rooted trees}
We recall basic notions from the theory of branch groups.
\begin{defn}
A tree $\mathcal T$ is called a {\bf regular rooted tree} of degree $d$ (with \textbf{root} $r$), if $r$ is a vertex
of valency $d$, and every other vertex of the tree $\mathcal{T}$ has valency $d+1$.
In particular, every automorphism of $\mathcal{T}$ fixes $r$, and the group $\Aut(\mathcal T)$ is a compact, and hence profinite, group.
A vertex $v \in V(\mathcal{T})$ is called at {\bf level $n$} if it has (graph-theoretical) distance $n$ from the root $r$.
Given a vertex $v$ of level $n$, we denote by $\mathcal T_v$ the subtree of $\mathcal T$ consisting of those vertices of level~$\geq n$ which are separated from the root by~$v$.
Given a group $W$ acting on a rooted tree $\mathcal T$,
we denote by $\st_W(n)$ the pointwise stabiliser of the sphere of radius $n$ around the root.
Given a vertex $v$ of $\mathcal T$ of level $n$, the \textbf{restricted stabiliser} of $v$ in $W$,
denoted by $\rist_W(v)$,
is defined as the subgroup of $\st_W(n)$ acting trivially on $\mathcal T_w$ for every vertex $w$ of level $n$ different from $v$.
We say that $W$ is {\bf level-transitive} if it acts transitively on the set of vertices at level $n$, for each $n$.
(All these definitions can be extended to more general rooted trees, and are particularly useful for spherically
homogeneous trees, but we will mostly need the rooted regular trees in what follows.)
\end{defn}
\begin{defn}
A profinite group $W$ is called a {\bf branch group}, if there is a rooted tree $\mathcal{T}$
and an embedding of $W$ into $\Aut(\mathcal{T})$ as a closed subgroup, such that $W$ is level-transitive on $\mathcal{T}$, and such that for each $n \geq 1$ the subgroup generated by the restricted stabilisers of all vertices at level~$n$ is of finite index (and hence open) in $W$. We say that $W$ is \textbf{saturated}, if $\st_{W}(n)$ is a characteristic subgroup of $W$ for all $n \geq 0$.
\end{defn}
\begin{remark}
\begin{compactenum}[(i)]
\item If $W$ is a branch group, then it has trivial quasi-centre (this follows \emph{e.g.} from \cite[Theorem~2(c)]{Grigorchuk}).
\item
Some authors use the terminology {\em rigid stabiliser} instead of restricted stabiliser,
and they denote this by $\mathrm{rist}_W(v)$ instead of $\rist_W(v)$.
\item
It is possible to define branch groups without explicitly referring to the associated rooted tree;
see, for example, \cite[Section~5]{Grigorchuk}.
\end{compactenum}
\end{remark}
\subsection{Self-replicating wreath branch groups and their auto\-mor\-phisms}\label{sec:branch}
In the rest of the paper, we shall focus on a special class of profinite branch groups which we now introduce.
Let $d >1$ and $D \leq \Sym(d)$ be a finite transitive permutation group.
Set $D_1 = D$ and $D_{n+1} = D_1 \wr D_n$ for all $n \geq 1$, where the wreath product is taken with its imprimitive wreath product action
(which is the natural action of $D_n$ on the $d^n$ leaves of the corresponding finite rooted tree).
The family $(D_n)_{n >0}$ naturally constitutes a projective system of finite groups. We denote by $W(D) = \varprojlim D_n$ the corresponding profinite group and call it the \textbf{wreath branch group} determined by $D$; see also Proposition~\ref{pr:wr} above.
We shall view $W(D)$ as a closed subgroup of the automorphism group of the regular rooted tree $\mathcal T$ of degree $d$ on which $W(D)$ acts continuously and faithfully. Since $D$ is transitive, it follows that $W(D)$ acts on $\mathcal T$ as a branch group. Notice that for all $v \in V(\mathcal T)$, there is an isomorphism $\mathcal T_v \to \mathcal T$ which conjugates $\rist_{W(D)}(v) $ onto $ W(D)$; in particular $\rist_{W(D)}(v) \cong W(D)$. We shall refer to this property by saying that $W(D)$ is \textbf{self-replicating}.
(Although this is not relevant to our purposes, we point out that, up to isomorphism, the group $W(D)$ is the unique closed branch subgroup $W \leq \Aut(\mathcal T)$ such that $W / \st_W(1) \cong D$ and that $W$ is self-replicating.)
\begin{lemma}\label{lem:JustInfinite}
$W(D)$ is just-infinite if and only if $D$ is perfect.
\end{lemma}
\begin{proof}
Immediate from \cite[Theorem~4]{Grigorchuk}.
\end{proof}
Although we shall not need it in this paper, we mention the following striking characterisation, due to M.~Burger and Sh.~Mozes (see \cite{Moz}).
\begin{theorem}[Burger--Mozes]
The profinite group $W(D)$ is topologically finitely generated if and only if $D \leq \Sym(d)$ is perfect and fixed-point-free.
\end{theorem}
\medskip
Automorphisms of wreath branch groups have been studied in~\cite{Lavreniuk99}. We denote by $W^k$ the direct product of $k$ copies of $W$.
\begin{lemma}\label{lem:Aut(W)}
The group $W = W(D)$ is saturated and for any $k>0$, we have $\Aut(W^k) = \Aut(W) \wr \Sym(k)$. Moreover we have
\[ \Aut(W) = \norma_{\Aut(\mathcal T)}(W) \]
and
\[ \Out(W) \cong \prod_{n>0} \norma_{\Sym(d)}(D)/D. \]
In fact $\Aut(W)$ is contained in $W(\norma_{\Sym(d)}(D))$ and coincides with the projective limit $\varprojlim A_n$, where the sequence of groups $A_n \leq \Sym(d^n)$ is defined inductively by
$A_1 = \norma_{\Sym(d)}(D)$ and $A_{n+1} = B_{n} \rtimes A_n \leq A_1 \wr A_n$, and $B_{n}$ is the subgroup of $\big(\norma_{\Sym(d)}(D)\big)^{d^n}$ consisting of those $d^n$-tuples $(\alpha_1, \dots, \alpha_{d^n})$ such that $\alpha_i \equiv \alpha_j \mod D$ for all $i, j \in \{1, \dots, d^n\}$.
\end{lemma}
\begin{proof}
The main result from \cite{Lavreniuk99} guarantees that $W(D)$ is saturated and that $\Aut(W) = \norma_{\Aut(\mathcal T)}(W)$. This is also established in \cite[Theorem~8.2]{LN02}, where it is moreover proved that $\Aut(W^k) = \Aut(W) \wr \Sym(k)$.
Set $A = \norma_{\Aut(\mathcal T)}(W) \leq \Aut(\mathcal T)$, $A_n = A/\st_A(n)$ and $W_n = W/\st_W(n)$ for all $n >0$. Clearly $A \cong \varprojlim A_n$. We view $A_n$ and $W_n$ as subgroups of $\Aut(\mathcal T_n)$, where $\mathcal T_n$ denotes the finite rooted tree consisting of the truncation of $\mathcal T$ at level~$n$.
We have $A_n = \norma_{\Aut(\mathcal T_n)}(W_n)$. In particular $A_1 \cong \norma_{\Sym(d)}(D)$ and $A_{n+1}$ is contained in $A_1 \wr A_n = \big( \bigoplus_{i=1}^{d^n} A_1\big) \rtimes A_n$. Referring to the latter semi-direct decomposition, we shall describe the elements of $A_{n+1}$ as $(d^n+1)$-tuples of the form $\big( (\alpha_1, \dots, \alpha_{d^n}), w\big)$, where $\alpha_i \in A_1$ and $w \in A_n$.
Consider the elements $\alpha=\big( (\alpha_1, \dots, \alpha_{d^n}), 1\big)$ and $g = \big( (1, \dots, 1), w\big)$ of $A_{n+1}$, where $\alpha_i \in A_1$ and $w \in W_n$. Since $W_{n+1}$ is normal in $A_{n+1}$ and contains $g$, we deduce that the commutator $[\alpha, g]$ belongs to $W_{n+1}$. In particular we deduce that $\alpha_i\inv \alpha_{i\cdot w} \in D$ for all $i$. Since $W_n$ is a transitive subgroup of $\Sym(d^n)$, we deduce that $\alpha_i \equiv \alpha_j \mod D$ for all $i$ and $j$. This confirms the claim that $A_{n+1} \cong B_n \rtimes A_n$.
Consider the subgroup $C_n = D^{d^n} \leq B_n$. Notice that $C_n$ is normal in $A_{n+1}$. Moreover the quotient $A_{n+1}/ C_n$ is isomorphic to the direct product $B_n/C_n \times A_n \cong \norma_{\Sym(d)}(D)/D \times A_n$. Since $C_n$ is contained in $W_{n+1}$, it follows inductively that $A_{n+1}/W_{n+1} \cong \prod_{i=1}^{n+1} \norma_{\Sym(d)}(D)/D$. Therefore, we have $\Out(W) \cong \prod_{n>0} \norma_{\Sym(d)}(D)/D$ as desired.
\end{proof}
\subsection{Abstract commensurators and Higman--Thompson groups}
We shall need to consider the so called Higman--Thompson groups, an excellent introduction to which can be consulted in \cite[\S4]{Brown}. (Numerous other excellent references are available concerning this subject, see notably \cite[\S3]{Rover} and references cited there). For each $d > 1$ and $k>0$, there is a Higman--Thompson group $V_{d,k}$ whose definition will be recalled below. Remark that K.~Brown \cite{Brown} denotes this group by $G_{d,k}$, but we prefer the notation $V_{d,k}$ in order to avoid confusion with (subgroups of) our group $G$. The groups $V_{d, k}$ were first introduced by G.~Higman in \cite{Higman} as generalisations of a group~$V$ earlier introduced by R.~Thompson. It is shown in \cite[Theorem~5.4]{Higman} that the derived group of $V_{d, k}$ is simple and finitely presented, and has index one or two according as $d$ is even or odd (see also \cite[Theorems~4.16 and~4.17]{Brown}).
For the sake of completeness, we briefly describe how the groups $ V_{d, k}$ can be defined. To this end, we denote by $\mathcal T_d$ the regular rooted tree of degree $d$ and consider the graph $\mathcal T_{d, k}$ which is the unique rooted tree having $k$ vertices of level~$1$ and such that $(\mathcal T_{d, k})_v \cong \mathcal T_d$ for all vertices $v$ of level~$1$. In a similar way as in Section~\ref{sec:BM}, we fix a colouring
$$ i : E(\mathcal T_{d, k}) \to \{1, \dots, d\}$$
of the edge-set of $\mathcal T_{d, k}$ such that, for each vertex $v$ of level~$n>0$, the restriction of $i$ to the set $D(v)$ is bijective, where $D(v)$ is the set of those edges joining $v$ to one of its $d$ neighbours of level~$n+1$. An \textbf{almost automorphism} of the tree $\mathcal T_{d, k}$ is a triple of the form $(A, B, \varphi)$, where $A$ and $B$ are finite subtrees of $\mathcal T_{d, k}$ containing the root such that $| \bd A | = | \bd B |$, and $\varphi$ is an isomorphism of forests $\mathcal T_{d, k} \setminus A \to \mathcal T_{d, k} \setminus B$. The \textbf{group of almost automorphisms} of $\mathcal T_{d, k}$, denoted by $\mathrm{AAut}(\mathcal T_{d, k})$, is the quotient of the set of all almost automorphisms by the relation which identifies two almost automorphisms $(A, B, \varphi)$ and $(A', B', \varphi')$ if there exists some finite subtree $A''$ containing $A \cup A'$ and such that $\varphi$ and $\varphi'$ coincide on $\mathcal T_{d, k} \setminus A''$. It is easy to verify that $\mathrm{AAut}(\mathcal T_{d, k})$ is indeed a group.
An element of $\mathrm{AAut}(\mathcal T_{d, k})$ is called \textbf{rooted} if it can be represented by an almost automorphism $(A, B, \varphi)$ such that for each vertex $v$ of $\mathcal T_{d, k} \setminus A$, the composed map
$$ i |_{D(\varphi(v))} \circ \varphi \circ \big( i |_{D(v)} \big)\inv $$
is the identity permutation of $\{1, \dots, d\}$. The \textbf{Higman--Thompson group} $V_{d, k}$ is defined as the subgroup of
$\mathrm{AAut}(\mathcal T_{d, k})$ consisting of the rooted elements. Its isomorphism type is independent of the choice of the colouring~$i$.
Following K.~Brown, we shall also consider the torsion-free subgroup $F_{d,k} \leq V_{d, k}$ consisting of the so-called \textbf{order-preserving} elements (see \cite[\S4]{Brown}). In order to define it, we first remark that a rooted almost automorphism $(A, B, \varphi)$ is uniquely determined by the bijection $\bd A \to \bd B$ of the leaf sets of $A$ and $B$ that it defines. Now we fix a \textbf{planar embedding} of the graph $\mathcal T_{d, k} $ in the plane $\RR^2$, \emph{i.e.} an embedding where edges are represented by segments and two edges intersect in a point if and only if this point represents a common vertex.
Moreover, we choose this embedding in such a way that the root coincides with the origin and that all vertices of level $n$ lie on the line $\{(x, y) \in \RR^2 \mid y = n\}$. Once this planar embedding is fixed, the leaf set $\bd A $ of each finite subtree $A \subset \mathcal T_{d, k}$ inherits a well-defined ordering, say from left to right. A rooted almost automorphism $(A, B, \varphi)$ is then called \textbf{order-preserving} if the induced bijection $\bd A \to \bd B$ preserves that ordering. The subgroup $F_{d, k} \leq V_{d, k}$ consists by definition of those elements which can be represented by an order-preserving rooted almost automorphism. One verifies easily that these form indeed a subgroup. The group which is commonly known as \textbf{Thompson's group $F$} is nothing but $F_{2, 1}$, while Thompson's group $V$ is $V_{2, 1}$. It can be shown that $F_{d, k}$ is isomorphic to $F_{d, 1}$ for any $k>0$, see \cite[Proposition~4.1]{Brown}, but we will not need this fact here.
Notice that every automorphism of $\mathcal T_{d, k} $ defines a unique element of the group $\AAut(\mathcal T_{d, k})$. Thus there is an embedding $\Aut(\mathcal T_{d, k} ) \to \AAut(\mathcal T_{d, k})$; its image intersects the subgroup $F_{d, k}$ trivially.
Let us finish this discussion by a final comment concerning $V_{d, k}$. We have mentioned above that $V_{d, k}$ is finitely generated, that it is simple if $d$ is even and possesses a simple subgroup of index~$2$ when $d$ is odd. This quotient of order~$2$ can easily be understood from the above definition. Indeed, we have pointed out that a rooted almost automorphism $(A, B, \varphi)$ is uniquely determined by a bijection $\partial A \to \partial B$. We can post-compose this bijection with the unique order-preserving bijection $\bd B \to \bd A$. In this way, we see that $(A, B, \varphi)$ is uniquely determined by a permutation of $\bd A$. This permutation is either even or odd. One sees that when $d$ is even, we can always enlarge $A$ and $B$ (by ``unfolding a leaf'' to the next level) so as to represent a rooted almost automorphism by an even permutation, while when $d$ is odd, the even or odd character of this permutation is independent of the chosen representative.
\bigskip
Now we return to the setting of the preceding paragraph and consider the wreath branch group $W = W(D)$ associated to a finite permutation group $D \leq \Sym(d)$ as before. For $k \geq 1$, we denote by $W^k$ the direct product of $k$ copies of $W$. We assume throughout that $D$ is transitive. Since $W$ is self-replicating, we have $\st_W(n) \cong W^{d^n}$ for all $n \geq 0$. In particular $\st_W(1)$ is isomorphic to the direct product of $d$ copies of $W$. Thus it follows from \cite[Theorem~1.1]{Rover} that the group of abstract commensurators $\Comm(W) = \acomm(W)$ contains a copy of the Higman--Thompson group $V_{d, 1}$. More generally, we have the following.
\begin{lemma}\label{lem:HT}
For each $k \geq 1$, the action of the Higman--Thompson group $V_{d, k}$ by almost automorphisms on $\mathcal T_{d, k}$ defines an embedding $V_{d, k} \to \acomm(W^k)$.
\end{lemma}
\begin{proof}
By definition, the group $W^k$ can be viewed as a closed subgroup of $\Aut(\mathcal T_{d, k})$ fixing pointwise the $k$ vertices of level~$1$. Moreover, for each vertex $v$ different from the root, we have $\rist_{W^k}(v) \cong W$. Therefore, every rooted automorphism of $T_{d, k}$ indeed defines a germ of automorphism of $W^k$, and it follows from the definitions that this induces an embedding $V_{d, k} \to \acomm(W^k)$.
\end{proof}
We need to consider some specific open subgroups of $\acomm(W^k)$, which we shall now describe. For all $k >0$ and $n \geq 0$, we set
\[ A_{k, n+1} = \Aut(W^{kd^n}) = \Aut(W) \wr \Sym(k d^n). \]
The latter equality follows from Lemma~\ref{lem:Aut(W)}. Thus $A_{1, 1} = \Aut(W)$. This lemma also guarantees that $W$ is saturated and, more generally, that the wreath branch group $W \wr \Sym(k d^n)$ is saturated. This implies moreover that $\Aut(W)$ embeds in $\Aut(\st_W(1) ) = \Aut(W^d) = \Aut(W) \wr \Sym(d)$. More generally, we have
\begin{align*}
A_{k, n} &= \Aut(W^{kd^{n-1}}) \\
&= \Aut(W) \wr \Sym(k d^{n-1}) \\
&\leq \big(\Aut(W) \wr \Sym(d) \big) \wr \Sym(k d^{n-1}) \\
&\leq \Aut(W) \wr \Sym(k d^{n})\\
&= A_{k, n+1} \,.
\end{align*}
Thus we have natural inclusions $A_{k,1} \leq A_{k,2} \leq \dotsm$ for all $k >0$, and the inductive limit
\[ A_k = \bigcup_n A_{k,n} \]
is thus a group. Notice that $A_k$ embeds as a subgroup in $\acomm(W^k)$; we shall identify it with its image.
In particular $A_k$ inherits the strong topology from $\acomm(W^k)$.
Remark moreover that $A_k$ is open in $\acomm(W^k)$ since it contains $\comm(W^k)$.
In a similar way as above, we define $A_{k, n+1}^+ = \Aut(W) \wr \Alt(k d^n)$, which is a subgroup of index two in $A_{k, n+1}$. By restricting the inclusion $A_{k, n} \leq A_{k, n+1}$, we have $A_{k, n}^+ \leq A_{k, n+1}$. We also define an open subgroup $A_k^+ \leq A_k$ by setting $A_k^+ = \bigcup_n A_{k,n}^+$.
\medskip
It will also be useful to introduce the following analogous subgroups, namely
\[ O_{k, n+1} = W \wr \Sym(k d^n) \hspace{.5cm} \text{and} \hspace{.5cm} O^+_{k, n+1} = W \wr \Alt(k d^n)\]
for all $k >0$ and $n \geq 0$, and
\[ O_{k} = \bigcup_n O_{k, n+1} \hspace{.5cm} \text{and} \hspace{.5cm} O^+_{k} = \bigcup_n O^+_{k, n+1} \leq O_k.\]
Thus $O_k$ and $O_k^+$ are open subgroups of $ \acomm(W^k)$, and are respectively contained in $A_k$ and $A_k^+$.
\begin{remark}\label{rem:A_k^+}
It is important to notice that $A_{k, n}^+ \leq A_{k, n+1}^+$ if and only if $D \leq \Alt(d)$. This implies that $[A_k : A_k^+] = 2$ if $D \leq \Alt(d)$, and $A_k^+ = A_k$ otherwise. Similarly, we have $[O_k : O_k^+] = 2$ if $D \leq \Alt(d)$, and $O_k^+ = O_k$ otherwise.
\end{remark}
\begin{lemma}\label{lem:A_k}
We have the following.
\begin{enumerate}[\rm (i)]
\item The groups $A_k$ and $A_k^+$ (resp. $O_k$ and $O_k^+$) are non-compact and locally elliptic.
\item $A_k^+$ (resp. $O_k^+$) is topologically simple.
\end{enumerate}
\end{lemma}
\begin{proof}
We shall prove the statements for $A_k$ and $A_k^+$; the arguments for $O_k$ and $O_k^+$ are similar.
\begin{compactenum}[(i)]
\item By definition $A_k \leq \acomm(W^k)$ is a union of an infinite ascending chain of compact open subgroups of $\acomm(W^k)$. The desired assertion is thus clear.
\item
Let $N \leq A_k^+$ be a non-trivial closed normal subgroup. Since $W^k$ has trivial quasi-centre, so does $\acomm(W^k)$ by \cite[Prop.~3.2(c)]{BarneaErshovWeigel}. Since the quasi-centre of a group contains the quasi-centre of any open subgroup, we infer that $A_k^+$ has also a trivial quasi-centre. Thus $N$ is non-discrete. It therefore meets $\comm(W^k)$ non-trivially. In particular, we deduce that, for $n$ large enough, the intersection $N \cap A_{k, n+1}^+$ maps onto a non-trivial normal subgroup of $\Alt(kd^n)$. Therefore, it maps onto $\Alt(kd^n)$. It follows that the same holds for $N \cap A_{k, m+1}^+$ for all $m \geq n$. This implies in particular that $N \cap A_{k, n+1}^+$ is dense in $A_{k, n+1}^+$. Since $N$ is closed, we deduce that $\bigcup_n A_{k,n}^+$ is contained in $N$, and the desired simplicity assertion follows. A similar argument shows that any non-trivial closed normal subgroup of $A_k$ contains $A_k^+$.\qedhere
\end{compactenum}
\end{proof}
The following result describes the structure of the group of abstract commensurators of a self-replicating wreath branch group. We identify $V_{d, k}$ and $F_{d, k}$ with their images in $\acomm(W^k)$, see Lemma~\ref{lem:HT}.
\begin{theorem}\label{thm:CommBranch:Iwasawa}
Let $k > 0$. Then
$\acomm(W^k)$ is generated by its subgroups $V_{d,k}$ and $ A_{k,2} $, it is contained in the group $ \AAut(\mathcal T_{d, k})$ of almost automorphisms of the tree $\mathcal T_{d, k}$ and, moreover, we have the decomposition
\[ \acomm(W^k) = F_{d, k} \cdot A_k, \]
where $F_{d, k} \cap A_k = 1$. In particular $F_{d, k}$ is a discrete subgroup of infinite covolume, and $A_k$ is a maximal locally elliptic subgroup of $\acomm(W^k)$. Moreover the closure of $V_{d, k}$ coincides with the open subgroup
$F_{d, k} \cdot O_k \leq \acomm(W^k) $.
\end{theorem}
We postpone the proof until the end of the section and first proceed to describe some consequences.
\begin{cor}\label{cor:M(D, k)}
Let $d >1$, let $D \leq \Sym(d)$ be transitive and set $W = W(D)$. Let moreover $k>0$ and let $M = M(D, k)$ denote the intersection of all non-trivial closed normal subgroups of $\acomm(W^k)$. We have the following.
\begin{enumerate}[\rm (i)]
\item $M$ is a rigid group belonging to the class $\mS$.
\item $M$ is open in $\acomm(W^k)$ and contains $\comm(W^k)$.
\item $[\acomm(W^k) : M] \leq 2$.
\item $[\acomm(W^k) : M] = 2$ if and only if $d$ is odd and $D \leq \Alt(d)$. In that case we have $M = F_{d, k} \cdot A_k^+$.
\end{enumerate}
\end{cor}
\begin{proof}
\begin{compactenum}[(i)]
\item
By Lemma~\ref{lem:A_k}, the group $A_k^+$ is topologically simple and contains $W^k$ as a compact open subgroup. Thus $\acomm(W^k) = \acomm(A_k^+)$ and, since $W^k$ has trivial quasi-centre, so does $A_k^+$. Therefore the assertion~(i) and the fact that $M$ is open follow Proposition~\ref{prop:AdmissibleRigid}.
\item[(ii)] Clear since $\comm(W^k) \leq A_k^+ \leq M$.
\item[(iii)] By definition $M$ is normal in $\acomm(W^k)$. Set $V = V_{d, k} $ and identify it with its image in $\acomm(W^k)$, see Lemma~\ref{lem:HT}.
Since $V$ is dense in $F_{d, k} \cdot O_k \leq \acomm(W^k)$ by Theorem~\ref{thm:CommBranch:Iwasawa} and since $O_k$ is open, it follows that $V \cap O_k$ is dense in $O_k$. In particular $M \cap V$ contains an infinite normal subgroup of $V$.
Since the derived group $[V, V]$ is simple (see, for example, \cite[Theorem~4.16]{Brown}), it follows that $M$ contains $[V, V]$.
Since $[V, V]$ contains $F_{d, k}$, we deduce from Theorem~\ref{thm:CommBranch:Iwasawa} that $\acomm(W^k)/M \cong A_k/A_k \cap M$. Lemma~\ref{lem:A_k} guarantees that $M$ contains $A_k^+$ and the desired assertion thus follows since $[A_k, A_k^+] \leq 2$, see Remark~\ref{rem:A_k^+}.
\item[(iv)]
If $d$ is even, then $V$ is simple. Since it is dense in $F_{d, k} \cdot O_k \leq \acomm(W^k)$ and since $M$ is open and normal, we deduce that $M $ contains $F_{d,k} \cdot O_k$. Moreover $M$ contains $A_k^+$ by (iii), and since $A_k = O_k \cdot A_k^+$, we infer that $M= \acomm(W^k)$, as desired.
Assume now that $d$ is odd and that $D \leq \Sym(d)$ contains an odd permutation. Since $V \cap W^k$ is dense in $W^k$, the existence of this permutation guarantees that $V \cap W^k$ contains an element $\sigma$ which is not annihilated in $V/ [V, V]$ (see the discussion on $V/[V, V]$ at the beginning of the present subsection above). Since $\acomm(W^k)/M$ embeds in the abelianization $V/[V, V]$, we have $\acomm(W^k) = \la \sigma \ra \cdot M$. By construction $\sigma \in M$, whence $M =\acomm(W^k)$, as desired.
\medskip
Assume conversely that $d$ is odd and that $D \leq \Alt(d)$. Then $A_k$ contains an open subgroup $A_k^+$ of index two which is topologically simple, see Lemma~\ref{lem:A_k} and Remark~\ref{rem:A_k^+}.
By Theorem~\ref{thm:CommBranch:Iwasawa}, we have $\acomm(W^k) = F_{d, k} \cdot A_k$. Moreover $F_{d, k}$ is contained in the derived group $[V, V]$. Therefore, we infer that the set $F_{d, k} \cdot A^+_k$ is a proper subset of $\acomm(W^k)$ which contains $[V, V]$. Notice that $F_{d, k} \cdot A^+_k$ is open in $\acomm(W^k)$ since $A^+_k$ is so. Moreover its complement is also open, since it coincides with the translate $F_{d, k} \cdot A^+_k \cdot \sigma$ for any element $\sigma \in A_k \setminus A^+_k$. In particular $F_{d, k} \cdot A^+_k$ is closed. Thus we have highlighted a proper closed subset of $\acomm(W^k)$ which contains the subgroup generated by $[V, V] \cup A^+_k$. Since $\acomm(W^k)$ is generated by $V \cup A^+_k$, we deduce that
\[ M = \la [V, V] \cup A_k^+ \ra = F_{d, k} \cdot A^+_k \]
and, hence, that $M$ has index two in $\acomm(W^k)$, as desired.\qedhere
\end{compactenum}
\end{proof}
The following result is a straightforward adaptation of Theorem~4.16 from \cite{BarneaErshovWeigel}.
\begin{cor}\label{cor:EmbeddingSimple}
Assume that the finite group $D$ is perfect, and let $W = W(D)$. Then $W^k$ embeds as a compact open subgroup in some compactly generated topologically simple group $G \in \mathcal S$.
\end{cor}
\begin{proof}
Let $G \leq \acomm(W^k)$ be the closed subgroup generated by $[V_{d, k}, V_{d, k}]$. Thus $G$ is generated by the compact open subgroup $W^k$ and the finitely generated group $[V_{d, k}, V_{d, k}]$. In particular it is compactly generated. By Lemma~\ref{lem:JustInfinite}, the group $W$ is just-infinite. Therefore, so is $W \wr \Alt(k)$, which embeds as a compact open subgroup in $G$. As in the proof of \cite[Theorem~4.16]{BarneaErshovWeigel}, one deduces that any non-trivial closed normal subgroup of $G$ is open. Since $V_{d, k} \cap W^k$ is dense in $W^k$, it follows that any open subgroup of $G$ contains an infinite subgroup of $V_{d, k}$. The result now follows from the fact that $[V_{d, k}, V_{d, k}]$ is simple and of index at most two in $V_{d,k}$.
\end{proof}
Before undertaking the proof of Theorem~\ref{thm:CommBranch:Iwasawa}, we need some additional notation and terminology. Let $\tilde W = W \wr \Sym(k)$ and let $\tilde{\mathcal T} = \mathcal T_{d, k}$ denote the rooted tree whose root has valence $k$ and such that $\tilde{\mathcal T}_v $ is isomorphic to $\mathcal T = \mathcal T_d$ for each vertex $v$ distinct from the root. The group $\tilde W$ naturally acts on $\tilde{\mathcal T}$ by automorphisms and $\tilde W$ is branch. Moreover $\acomm(W^k) \cong \acomm(\tilde W)$.
A pair of vertices of $\tilde{\mathcal T}$ is called \textbf{independent} if no geodesic ray emanating from the root contains both. More generally, a subset of $V(\tilde{\mathcal T})$ consisting of pairwise independent vertices is called independent. A \textbf{leaf set} is an independent subset of $V(\tilde{\mathcal T})$ which is maximal among independent sets. Given $n \geq 0$, we denote by $V_n(\tilde{\mathcal T})$ (resp. $V_{\geq n}(\tilde{\mathcal T})$) the set of vertices of level $n$ (resp. at least $n$). Thus $V_n(\tilde{\mathcal T})$ is a leaf set.
We shall need the following.
\begin{lemma}\label{lem:Rover}
Let $n \geq 0$, $B \leq \tilde W$ be an open subgroup and $\alpha: \st_{\tilde W}(n) \to B$ be an isomorphism.
(In particular $\alpha$ defines some element of $ \acomm(\tilde W)$.)
\begin{enumerate}[\rm (i)]
\item There is a map $\Phi : V(\tilde{\mathcal T}) \to V(\tilde{\mathcal T})$ such that for each vertex $v \in V_{\geq n}(\tilde{\mathcal T})$, we have $\alpha(\rist_{\tilde W}(v)) = \rist_{\tilde W}(\Phi(v))$.
\item $\Phi$ maps every leaf set $L \subset V_{\geq n}(\tilde{\mathcal T})$ to a leaf set, and the restriction $\Phi : L \to \Phi(L)$ is injective (and hence bijective if $L$ is finite).
\end{enumerate}
\end{lemma}
\begin{proof}
We start by collecting a few preliminary observations. By \cite[Theorem~1.2]{Rover}, the group $\acomm({\tilde W})$ is isomorphic to the relative commensurator of ${\tilde W}$ in the homeomorphism group of the boundary $\partial \tilde{\mathcal T}$. Since the isomorphism $\alpha$ defines a unique element of $\acomm({\tilde W})$, we can view it as an element of $\mathrm{Homeo}(\bd \tilde{\mathcal T})$.
Given any subset $O \subset \bd \tilde{\mathcal T}$ containing at least two points, there is a unique vertex $v$ of maximal possible level such that $O \subset \bd \tilde{\mathcal T}_v$. We denote this vertex by $\mathfrak v(O)$. On the other hand, notice that the action of ${\tilde W}$ on $\partial \tilde{\mathcal T}$ is continuous, and every closed subgroup of ${\tilde W}$ has closed orbits since it is compact. Therefore the restricted stabiliser $\rist_{\tilde W}(v)$ of a vertex $v$ has a unique orbit that consists of more than one point.
We denote this orbit by $\mathfrak o(v)$. We have $\mathfrak v( \mathfrak o(v)) = v$.
\begin{compactenum}[(i)]
\item
Assume that $v$ has level at least $n$. This allows one to consider
\[ \Phi(v) = \mathfrak v(\alpha(\mathfrak o(v))) \,; \]
recall that $\alpha$ is regarded as a homeomorphism of $\partial\tilde{\mathcal T}$.
It is clear from the definition that $\alpha \rist_{\tilde W}(v) \alpha\inv \leq \rist_{\tilde W}(\Phi(v))$. Notice that $\mathfrak o(\Phi(v)) = \alpha(\mathfrak o(v))$. Therefore $\mathfrak o(v)$ is the unique orbit of $\alpha\inv \rist_{\tilde W}(\Phi(v)) \alpha$ that consists of more than one point.
We infer that $\alpha\inv \rist_{\tilde W}(\Phi(v)) \alpha \leq \rist_{\tilde W}(v)$, whence $\alpha \rist_{\tilde W}(v) \alpha\inv = \rist_{\tilde W}(\Phi(v))$ as desired.
\item
A pair of vertices $\{v, v'\} \subset V(\tilde{\mathcal T})$ is independent if and only if $\mathfrak o(v)$ and $\mathfrak o(v')$ are disjoint. Moreover a finite independent subset $L \subset V(\tilde{\mathcal T})$ is a leaf set if and only if $\bd \tilde{\mathcal T} = \bigcup_{v \in L} \mathfrak o(v)$. It readily follows that $\Phi$ maps independent sets to independent sets, and leaf sets to leaf sets. Finally, given a leaf set $L \subset V_{\geq n}(\tilde{\mathcal T})$, then $\{\Phi(v), \Phi(v')\}$ is an independent pair for all $v \neq v' \in L$. In particular the restriction of $\Phi$ to $L$ is injective.
\qedhere
\end{compactenum}
\end{proof}
Notice that the lemma implies in particular that every germ of automorphism of $W^k$ acts by almost automophism on $\tilde{\mathcal T} = \mathcal T_{d, k}$.
\begin{proof}[Proof of Theorem~\ref{thm:CommBranch:Iwasawa}]
Every abstract commensurator of $\tilde W$ admits a representative of the form $\alpha: \st_{\tilde W}(n) \to B$ for some $n \geq 0$ and some open subgroup $B \leq \tilde W$. We now invoke Lemma~\ref{lem:Rover} to the leaf set $L = V_n(\tilde{\mathcal T})$. In view of the definition of the embeddings $F_{d, k} \subset V_{d,k} \subset \acomm(W^k)$ (see Lemma~\ref{lem:HT}), we deduce that there exists an element $g \in F_{d, k}$ such that $\alpha g$ normalises $ \st_{\tilde W}(n)$. Since $\Aut(\st_{\tilde W}(n))$ is contained in $A_k$ (see Lemma~\ref{lem:Aut(W)}) we deduce the equality
\[ \acomm(W^k) = \acomm(\tilde W) = A_k\cdot F_{d, k} = F_{d, k} \cdot A_k, \]
as well as the fact that the group $\acomm(W^k)$ consists entirely of almost automorphisms of the tree $\mathcal T_{d, k}$.
Given $v \in V_{\geq 2}(\tilde{\mathcal T})$, there exist $\alpha \in V_{d, k} $ such that $\Phi(v) $ has level exactly two, where $\Phi$ is defined by Lemma~\ref{lem:Rover}. Moreover, for all $n$ and each vertex $v$ of level~$n$, we have seen that $A_{k,n} = \Aut(\rist_{\tilde W}(v)) \wr \Sym(k d^{n-1})$. The equality $\acomm(\tilde W) = \la V_{d, k}, A_{k,2} \ra$ now follows from the fact that $\Sym(k d^{n-1})$ is in fact contained in $V_{d, k}$.
That $F_{d,k} \cap A_k = 1$ follows from the definition of $F_{d,k}$.
We next consider the group $V \cap A_k$. It consists of those elements of $V$ which can be represented by rooted almost automorphisms of the form $(A, B, \varphi)$, where $A$ and $B$ are both balls centred at the root of $\mathcal T_{d, k}$. It is then not difficult to notice that $V \cap O_k$ is dense in $O_k$. Since $F_{d, k}$ is contained in $V_{d, k}$, we deduce from the equality $\acomm(W^k) = F_{d, k} \cdot A_k$ that $V_{d, k}$ is dense in $F_{d,k} \cdot O_k \leq \acomm(W^k)$, as desired.
Finally, combining the triviality of $F_{d,k} \cap A_k$ with the fact that $A_k$ is open, we deduce that the covolume of $F_{d,k}$ in $\acomm(W^k)$ is bounded below by the volume of $A_k$, which is infinite since $A_k$ is not compact (see Lemma~\ref{lem:A_k}). Since $F_{d, k}$ is discrete and torsion-free, no subgroup of $\acomm(W^k)$ containing $A_k$ properly can be locally elliptic.
\end{proof}
\subsection{Embeddings in compactly generated rigid simple groups}
We have seen in Corollary~\ref{cor:EmbeddingSimple} that if $D$ is perfect, then the wreath branch group $W(D)$ embeds as a compact open subgroup in some compactly generated simple group. Similarly, if $D$ coincides with a point stabiliser in some finite transitive permutation group $F \leq \Sym(d+1)$, then $W(D)^2$ embeds as an edge-stabiliser in the Burger--Mozes simple group $U(F)^+$.
Our next goal is to determine when $W(D)^k$ embeds in a compactly generated simple group which is \emph{rigid}. Recall that a rigid simple group $M = M(D, k)$ containing $W(D)^k$ as a compact open subgroup was defined in Corollary~\ref{cor:M(D, k)}. By Proposition~\ref{prop:AdmissibleRigid}(i), it is, up to isomorphism, the unique topologically simple locally compact group satisfying this condition.
\begin{theorem}\label{thm:CommBranch}
Let $D \leq \Sym(d)$ be transitive, let $W=W(D)$ and let $M \in \mS$ be a rigid simple group containing a compact open subgroup isomorphic to $W^k$ for some $k>0$. Then the following conditions are equivalent.
\begin{enumerate}[\rm (i)]
\item $M$ is compactly generated.
\item $\mathscr L(M)$ is compactly generated.
\item $\mathscr L(M)$ is compactly generated and $[\mathscr L(M): M]\leq 2$.
\item $\mathscr L(M) = V_{d, k} \cdot W^k$.
\item $\Out(W) = 1$.
\item $\norma_{\Sym(d)}(D) =D$.
\end{enumerate}
\end{theorem}
\begin{proof
\begin{asparaitem}\itemsep1ex
\item[(v) $\Leftrightarrow$ (vi)]
Immediate from Lemma~\ref{lem:Aut(W)}.
\item[(iii) $\Rightarrow$ (i) and (ii)]
Obvious.
\item[(i) $\Rightarrow$ (v) ]
By Lemma~\ref{lem:Aut(W)}, the group $W$ is saturated and, hence, the subgroup $\st_W(1)$ is characteristic in $W$. Moreover,
$\big(\st_W(1)\big)^k$ is characteristic in $W^k$.
Also, $\Out(W)$ is either trivial or uncountable.
It is straightforward to deduce from the description of the groups $\Aut(W)$ and $\Aut(W^k)$ given in Lemma~\ref{lem:Aut(W)} that if $\Out(W)$ is non-trivial, then the image of $\Aut(W^k)$ in $\Aut\big( (\st_W(1))^k\big) $ is of uncountable index. In view of Proposition~\ref{prop:Aut}, the hypothesis that $M$ is rigid and compactly generated guarantees that this index is at most countable. Therefore $\Out(W)$ must be trivial, as desired.
\item[(ii) $\Rightarrow$ (v)]
Assume that $\Out(W)$ is not trivial. Then it is uncountable, and so is $\Out(W^k)$, see Lemma~\ref{lem:Aut(W)}. This implies that the index of $W^k$ in $\mathscr L(M)$ is uncountable. In particular $\mathscr L(M)$ cannot be $\sigma$-compact, and \emph{a fortiori} not compactly generated.
\item[(v) $\Rightarrow$ (iv)]
Since $A_{k, n} \cap V_{d, k}$ maps onto $\Sym(kd^n)$ for all $n>0$,
we deduce from Theorem~\ref{thm:CommBranch:Iwasawa} that $\mathscr L(M) = \Gamma \cdot W^k$, where $\Gamma = V_{d, k}$.
The claim follows.
\item[(iv) $\Rightarrow$ (iii)]
Clearly the assumption implies that $\mathscr L(M)$ is compactly generated.
The inequality $[\mathscr L(M):M] \leq 2$ follows from Corollary~\ref{cor:M(D, k)}.\qedhere
\end{asparaitem}
\end{proof}
We conclude this section with the proof of the last two theorems announced in the introduction.
\begin{proof}[Proof of Theorem~\ref{thmi:L(G)}]
We deduce from Proposition~\ref{prop:AdmissibleRigid} that $G$ embeds as an open subgroup in some rigid simple group $\tilde G \in \mS$. Let $D$ denote the point-stabiliser $F_0$ viewed as a subgroup of $\Sym(d)$. Set $W =W(D)$ and recall that every edge-stabiliser in $G$ is isomorphic to $W^2 = W \times W$.
It follows from these observations that the hypotheses of Theorem~\ref{thm:CommBranch} are satisfied with $k=2$. All the desired statements now follow, except the fact that $H = [\mathscr L(G), \mathscr L(G)] $ is abstractly simple. The passage from topological simplicity to abstract simplicity goes as follows.
First recall that $\comm(G)$ is open in $\acomm(G)$. Moreover $G$ is abstractly simple (see Theorem~\ref{th:tits}) and is thus contained in $H$. This implies that $H$ is itself open. Let now $N$ be an abstract normal subgroup of $H$. The intersection $G \cap N$ is either trivial or equal to $G$. Since $G$ is open, the former case implies that $N$ is discrete while the latter case implies that $N$ is open. In either case $N$ is closed. The result follows.
\end{proof}
We finally prove Theorem~\ref{thmi:NewSimpleGroups}, or rather the following more detailed statement.
\begin{theorem}\label{thm:NewSimpleGroups}
Let $d > 1$, $D \leq \Sym(d)$ be transitive and $W = W(D)$ be the profinite branch group defined as the infinitely iterated wreath product of $D$ with itself. Then for every $k>0$, there is a locally compact group $M = M(D, k)$ which is topologically simple and rigid, and which contains the direct product $W^k$ of $k$ copies of $W$ as a compact open subgroup. Moreover:
\begin{enumerate}[\rm (i)]
\item $M$ is uniquely determined up to isomorphism.
\item $M$ is compactly generated if and only if $\norma_{\Sym(d)}(D) = D$.
\item $[\acomm(M): \comm(M)] \leq 2$ and $[\acomm(M): \comm(M)] = 2$ if and only if $d$ is odd and $D \leq \Alt(d)$.
\item $\acomm(M) = F_{d, k} \cdot A_k$, where $F_{d, k}$ is a copy of the Higman--Thompson group embedded as a discrete subgroup, and $A_k$ is a non-compact locally elliptic open subgroup such that $F_{d, k} \cap A_k = 1$. Moreover $A_k$ is a maximal locally elliptic subgroup of $\acomm(M)$. Furthermore $F_{d, k}$ is contained in $\comm(M)$ and $A_k$ possesses an open subgroup $A_k^+$ of index at most two which is topologically simple and equally contained in $\comm(M)$.
\item If $k \equiv k' \mod (d-1)$, then $M(D, k) \cong M(D, k')$.
\item If $D' \leq \Sym(d')$ contains a simple subgroup which does not embed as a subgroup of $D$, then $M(D, k) \not \cong M(D', k')$ for all $k, k' >0$.
\end{enumerate}
\end{theorem}
\begin{proof}
The existence of a rigid $M = M(D,k) \in \mS$ follows from Corollary~\ref{cor:M(D, k)}. Assertion~(i) follows from Proposition~\ref{prop:AdmissibleRigid}, Assertion~(ii) is contained in Theorem~\ref{thm:CommBranch} while Assertions~(iii) and~(iv) follow from Theorem~\ref{thm:CommBranch:Iwasawa} and~Corollary~\ref{cor:M(D, k)}.
\medskip
In order to prove Assertions~(v) and~(vi), first notice that groups of the form $M(D, k)$ are all rigid. In particular two of them are isomorphic if and only if they are locally isomorphic. Assertion~(v) is now straightforward to establish. For Assertion~(vi), it suffices to show that if $D'$ contains a simple subgroup $S$ which does not embed as a subgroup of $D$, then $W(D)^k$ and $W(D')^{k'}$ are not locally isomorphic. It is not difficult to show that every identity neighbourhood of $W(D')^{k'}$ contains a finite subgroup isomorphic $D'$. On the other hand, one shows that no finite quotient of $W(D)^k$ contains a subgroup isomorphic to $S$, which implies that the profinite group $W(D)^k$ itself does not contain any copy of $S$. Thus $W(D)^k$ and $W(D')^{k'}$ are not locally isomorphic, as claimed.
\end{proof}
\begin{remark}
None of the compactly generated simple groups appearing in Theorem~\ref{thm:NewSimpleGroups} (or in Corollary~\ref{cor:EmbeddingSimple}) admits any continuous, proper and \emph{cocompact} action on any locally finite tree, or on any locally finite CAT(0) cell complex. Indeed, the Higman--Thompson group $V_{d, k}$, as well as its derived group $[V_{d, k}, V_{d, k}]$, contains a copy of every finite group. This implies that if $V_{d, k}$ acts on a CAT(0) cell complex $X$, then the size of links of vertices in $X$ is unbounded, thereby preventing the action of any group containing $V_{d, k}$ from being cocompact.
\end{remark}
|
\section{Introduction}
Two appealing ideas, both due to R. Penrose, provide a different perspective to our understanding of physical fields. The first of these is to try to build up space-time and quantum mechanics from combinatorial principles. One way to attempt this is from so-called spin networks and Penrose argues how $3$-dimensional space arises from systems with large angular momentum. A spin network is a graph whose vertices have degree $3$ (the number of edges incident with each vertex is $3$) and whose edges are labeled by an integer which represents twice the angular momentum \cite{Pe-4}. The second idea is to consider twistor space, the space of null geodesics, as the more basic object from which space-time points should be derived \cite{Pe-3}. Twistor diagrams can be considered as a natural adaptation of the combinatorial perspective to the twistor program \cite{Pe-3}. Our principal aim in this article is to give an alternative way in which combinatorial structures arise from the twistorial construction of fields.
One of the basic objects of twistor theory, a shear-free ray congruence, can be viewed alternatively as a semi-conformal complex valued mapping which evolves in time \cite{Ba-We-1, Ba-Wo-1, Ba-Wo-2}. This latter object, which we shall consider as a physical field, is perfectly suited to be defined on a finite graph (or network). In this context, we shall refer to the function as \emph{holomorphic}, since in the plane, a semi-conformal mapping is either holomorphic or anti-holomorphic. What is fascinating is that only certain graphs support a holomorphic function. If the order of the graph is sufficiently small, computer programs can be used to generate such functions.
To a graph $\Gamma$ endowed with a holomorphic function $\phi : V(\Gamma ) \rightarrow \CC$, where $V(\Gamma )$ is the set of vertices of $\Gamma$, we can associate its twistor dual $L_{\Gamma}$, whose vertices are the edges of the original graph, sometimes called the \emph{line-graph}, as well as a function $\psi : V(L_{\Gamma}) \rightarrow \CC$. In the spirit of twistor theory, where light rays are considered to be the fundamental objects, we consider the graph $L_{\Gamma}$ as the basic object from which physical fields and space-time points should be deduced. Indeed, a vertex of $\Gamma$ arises as a complete subgraph in $L_{\Gamma}$ upon which $\sum \psi^2$ vanishes. An outline of the paper is as follows.
We first of all explain how a shear-free ray congruence on $4$-dimensional Minkowski space can be viewed as an evolving family of complex-valued semi-conformal mappings on $3$-dimensional space-like slices. This is the basis of our generalization to graphs.
In Section \ref{sec:graphs}, we discuss finite graphs. In particular, we recall the notion of holomorphic mapping between graphs and introduce the concept of holomorphic function on a graph. Holomorphic mappings are then characterized by the property that they preserve holomorphic functions (Proposition \ref{prop:holo}). The properties of holomorphic functions are discussed in relation to quantum graphs, spin networks and orthographic projection.
A holomorphic function on a graph is equivalent to an isotropic $1$-form which vanishes around closed cycles. On a regular graph of degree $3$ oriented by colour, we show how, from an isotropic $1$-form, we can recover a spinor field defined on the vertices, which corresponds to the spinor field defining a shear-free ray congruence on space-time. Finally, in Section \ref{sec:line-graph}, we discuss the twistor correspondence between a graph and its line-graph.
\section{Shear-free ray congruences on Minkowski space} \label{sec:SFR}
The Penrose twistor correspondance associates to a light ray in Minkowsi space $\MM^4$, a point in a $5$-dimensional CR-submanifold $\Nn^5$ of $\CP^3$ \cite{Pe-1}. We can obtain $\Nn^5$ as follows.
We first of all compactify $\MM^4$ by adding a light cone at infinity to obtain the manifold $\ov{\MM}^4$ diffeomorphic to $S^1 \times S^3$.
The Hopf fibration $\pi : \CP^3 \rightarrow S^4$ is the map given by
$$
\pi ([z_1, z_2, z_3, z_4]) = [z_1 + z_2 \jj , z_3 + z_4 \jj ] \in \HP^1
$$
where we use homogeneous coordinates $[z_1, z_2, z_3, z_4]$ for points of $\CP^3$ and where $\HP^1$ is the quaternionic projective space. On identifying $\HP^1$ with $S^4$ and letting $S^3$ be the equatorial $3$-sphere given by $\Re [z_1 + z_2 \jj , z_3 + z_4 z\jj ] = 0$, we see that $\pi ([z_1, z_2, z_3, z_4]) \in S^3$ if and only if
$$
z_1 \ov{z_3} + z_2 \ov{z_4} + \ov{z_1} z_3 + \ov{z_2}z_4 = 0\,.
$$
We then define $\Nn^5 = \pi^{-1}(S^3)$. Note that $S^3$ is parallizable and so the bundle $\Nn^5$ is diffeomorphic with $S^3 \times S^2$. There is now a natural identification between $\pi^{-1}(S^3)$ and the unit tangent bundle $T^1S^3$ to $S^3$. If we consider $S^3$ as the compactified slice $t = 0$ in $\ov{\MM}^4$, then a point $(x,v)$ of $\Nn^5 \cong S^3 \times S^2$ gives the light ray passing through $x$ with direction $v$.
The picture can be unified by introducing the flag manifold ${\bf F}_{12}$ of pairs $(\ell , \Pi )$ consisting of resp. $1$- and $2$-dimensional subspaces of $\CC^4$ with $\ell \subset \Pi$ and considering the double fibration:
$$
\begin{array}{ccccc}
& & {\bf F}_{12} & & \\
& \swarrow & & \searrow & \\
\CP^3 & & & & G_2(\CC^4) \\
\cup & & & & \cup \\
\Nn^5 & & & & \ov{\MM}^4
\end{array}
$$
where $G_2(\CC^4)$ is the Grassmannian of complex $2$-dimensional subspaces of $\CC^4$ and where the left projection is given by $(\ell , \Pi ) \mapsto \ell$ and the right by $(\ell , \Pi ) \mapsto \Pi$. A point $\ell$ of $\CP^3$ determines a plane in $G_2(\CC^4)$, called an $\alpha$-plane, given by all the $\Pi$ containing $\ell$. This plane may or may not intersect $\ov{\MM}^4$; if it does it does so in a null geodesic. The points of $\CP^3$ which gives light rays are precisely the points of $\Nn^5$.
In order to describe a shear-free ray congruence (SFR), it is useful to have the notion of conformal foliation. We formulate this in terms of a semi-conformal mapping, which will be the fundamental object we discuss later in the context of graphs.
A Lipschitz map $\phi : (M^m, g) \rightarrow (N^n, h)$ between Riemannian manifolds is said to be \emph{semi-conformal} if, at each point $x\in M$ where $\phi$ is differentiable (dense by Radmacher's Theorem), the derivative $\dd \phi_x : T_xM\rightarrow T_{\phi (x)}N$ is either the zero map or is conformal and surjective on the complement of $\ker \dd\phi_x$ (called the \emph{horizontal distribution}). Thus, there exists a number $\lambda (x)$ (defined almost everywhere), called the \emph{dilation}, such that $\lambda (x)^2 g(X,Y)$ $ = $ $\phi^*h(X,Y)$, for all $X,Y\in (\ker \dd\phi_x)^{\perp}$. If $\phi$ is of class $C^1$, then we have a useful characterisation in local coordinates, given by
$$
g^{ij}\phi_i^{\alpha}\phi_j^{\beta} = \lambda^2 h^{\alpha\beta}\,,
$$
where $(x^i), (y^{\alpha})$ are coordinates on $M, N$, respectively and $\phi^{\alpha}_i = \partial (y^{\alpha}\circ \phi ) / \partial x^i$. The fibres of a smooth submersive semi-conformal map determine a conformal foliation, see \cite{Va} and conversely, with respect to a local foliated chart, we may put a conformal structure on the leaf space with respect to which the projection is a semi-conformal map. We then have the identity:
$$
\left(\Ll_Ug\right)(X,Y) = - 2 U(\ln \lambda )\,g(X,Y),
$$
for $U$ tangent and $X,Y$ orthogonal to the foliation. This latter equation can be taken to be the characterisation of a conformal foliation. Specifically, a foliation is called \emph{conformal} if there is a function $a = a(U)$ which depends only on $U$, such that
$$
\left(\Ll_Ug\right)(X,Y) = a(U) \,g(X,Y),
$$
for $U$ tangent and $X,Y$ orthogonal to the foliation. The relation between $a$ and the dilation $\lambda$ can now be deduced by calculating the mean curvature of the horizontal distribution, see, for example \cite{Ba}.
A shear-free ray congruence on a region $A \subset \MM^4$ is a foliation by null-geodesics which is without shear. That is, if $W$ represents the tangent vector field to the congruence, then at a point $x\in A$, the metric complement $W^{\bot}$ is $3$ dimensional and contains $W$ itself; if we take a $2$-dimensional spacelike complement $S$ in $W^{\bot}$, then for the congruence to be \emph{shear-free}, we require Lie transport of vectors in $S$ along $W$ to be conformal. This property is independent of the choice of $S$. By the Kerr Theorem, locally a shear-free ray congruence is defined by the intersection of $\Nn^5$ with a complex analytic surface $\Ss$ \cite{Pe-Ri-2}. In general the congruence of light rays defined by $\Nn^5 \cap \Ss$ will be multivalued with singularities. Solutions of the zero rest-mass field equations
$$
\nabla^{AA'}\phi_{AB\ldots L} = 0
$$
are then given by considering a function $f(z_1, z_2, z_3, z_4)$ homogeneous of degree $- n - 2$ and taking a contour integral in an appropriate way. This is the basis of the Penrose transform, which is an integral transform from sheaf cohomology in the twistor space into the space of massless fields, see \cite{Ea, Pe-3, Wa-We} for details.
In \cite{Ba-We-1}, the equations for an SFR are reformulated in such a way that will enable us to adapt them to the context of graphs. Specifically, if $W$ is tangent to a future pointing congruence of null curves on a region $A \subset \MM^4$, then at each point $(t,x) \in \MM^4$, we can decompose $W$ into its timelike and spacelike components: $W = \partial_t + U$, where $U$ is a unit tangent to the slice $\RR^3_t = \{ (t, x_1, x_2, x_3) \in \MM^4: t \ {\rm const}\ \}$. Then $W$ is tangent to an SFR if and only if
\begin{equation} \label{SFR}
\left\{ \begin{array}{lrcl}
{\rm (i)} & \frac{\partial U}{\partial t} & = & - \nabla^{\RR^3_t}_UU \\
{\rm (ii)} & 0 & = & (\Ll_Ug) (X + \ii Y,X + \ii Y) \,,
\end{array} \right.
\end{equation}
where $\{ X, Y, U \}$ is an orthonormal basis tangent to $\RR^3_t$ at each point and $g$ is the standard Euclidean metric on $\RR^3_t$. Indeed, the unit direction field $U$ can be represented by a spinor field $[\mu^A]\in \CP^1$ and then (\ref{SFR}) is equivalent to the usual spinor representation of an SFR:
$$
\mu^A\mu^B\nabla_{AA'}\mu_B = 0\,.
$$
Note that (\ref{SFR})(i) is equivalent to the geodesic condition $\nabla^{\MM^4}_WW = 0$, whereas (\ref{SFR})(ii) is equivalent to the property that $U$ be tangent to a conformal foliation on each slice $\RR^3_t$. Furthermore, one can show that if (\ref{SFR})(i) is satisfied everywhere and (\ref{SFR})(ii) on an initial slice $\RR^3_0$, then (\ref{SFR})(ii) is satisfied for all $t$ \cite{Ba-We-1}.
If we locally integrate the vector field $U$, so that for each $t$ it is tangent to the fibres of a semi-conformal mapping $\phi = \phi_t: B_t \rightarrow \CC$ ($B_t$ open in $\RR^3_t$), the above equations are equivalent to the pair \cite{Ba-We-1}:
\begin{equation} \label{evol-scm}
\left\{ \begin{array}{lrcl}
{\rm (i)} & d\left( \frac{\partial \phi }{\partial t} \right) (U) & = & -\tau (\phi ) \\
{\rm (ii)} & 0 & = & g(\mbox{\rm grad}\, \phi , \mbox{\rm grad}\, \phi ) \,,
\end{array} \right.
\end{equation}
where $\mbox{\rm grad}\, \phi$ is the (complex) gradient with respect to the metric $g$ on $\RR^3_t$.
In fact one can easily check that \eqref{evol-scm} is invariant under the replacement of $\phi_t$ by $\psi_t = \zeta_t \circ \phi_t$, where $\zeta_t$ is an arbitrary conformal transformation of a domain of the complex plane; this is precisely the gauge freedom one requires in the choice of $\phi_t$.
\section{Holomorphic functions on a graph} \label{sec:graphs}
A \emph{finite graph $\Gamma$ of order $n$} is a set $V$ of cardinality $n$ endowed with a binary relation $\sim$. For $x,y \in V$, if $x\sim y$ we will say that $x$ and $y$ are \emph{neighbours}, or are \emph{joined by an edge} and we will represent this diagrammatically by drawing a line segement between $x$ and $y$. We suppose in what follows that the relation $\sim$ is symmetric, so that edges are not directed, although most of our discussion also applies to directed graphs. We prefer to use the term \emph{directed}, rather than the more usual \emph{oriented}; the latter term being reserved for a notion of orientation of a (undirected) graph, rather akin to orientation of a manifold, which we will define later. We do not allow the relation $\sim$ to be reflexive, so that the graph $\Gamma$ does not contain loops, neither do we allow multiple edges, although once more, the discussion can be adapted to this more general situation. We can represent the edges as a subset $E$ of the formal symmetric product $V\odot V$ and so express the graph $\Gamma$ as the pair $\Gamma = (V,E)$. It will often be convenient to represent an edge $(x,y)\in E$ using the notation $\ov{xy}$, or, if we impose a direction on the edge, by $\vec{xy}$. We say that the edge $\ov{xy}$ is \emph{incident with the vertex $x$} (and also with $y$).
It is our aim to represent fields purely in terms of the combinatorial properies of graphs and as far as possible to dispense with notions of (semi-) Riemannian geometry. However, a natural generalisation of our theory is to endow each edge with a real number, called its \emph{length} and to consider what are called \emph{metric graphs}. One can even go further, and suppose that an angle is defined between edges incident with a given vertex, however, this now becomes an approximation of (semi-)Riemannian geometry and would defeat our purpose of developing a purely \emph{combinatorial} theory.
Many notions of Riemannian geometry translate into combinatorial properties of graphs. A useful references is the book by Chung \cite{Ch}, which uses slightly different conventions. We outline below, those notions which are essential to our development.
Given a graph $\Gamma = (V, E)$, to each $x \in V$, we define its \emph{degree} $m(x)$ to be the number of edges incident with $x$. A graph is called \emph{regular} if $m(x) = m$ is constant for each vertex. We define the \emph{tangent space at $x\in V$}, to be the set $T_x\Gamma := \{ \vec{xy}: \ov{xy}\in E\}$. That is, each element of $T_x\Gamma$ is a \emph{directed} edge, with base point $x$ and end point $y\sim x$. Given a function $\phi : V \rightarrow \RR^N$ with values in a Euclidean space and a vector $X = \vec{xy} \in T_x\Gamma$, we define its \emph{directional derivative in the direction} $X$ to be the number
$$
d\phi_x(X) = \phi (y) - \phi (x)\,.
$$
Note that we could extend the notion of tangent space to include all linear combinations of edges $\vec{xy}$, $y\sim x$, to obtain a vector space, but we prefer to use a discrete concept for the tangent space. If $\omega : T_x\Gamma \rightarrow \RR^N$, then we define its \emph{co-derivative at $x$} to be the quantity
$$
\dd^* \omega (x) = -\frac{1}{m(x)}\sum_{y\sim x} \omega (\vec{xy})\,.
$$
If for each $x\in V$ we have given a map $\omega = \omega_x : T_x\Gamma \rightarrow \RR^N$, then provided $\omega (\vec{xy}) = - \omega (\vec{yx})$, we will refer to $\omega$ as an $\RR^N$-valued $1$-form. In particular, if $f : V \rightarrow \RR$ is a function, then $\dd f$ is a $1$-form and we have
\begin{eqnarray*}
\dd^*\dd f (x) & = & - \frac{1}{m(x)}\sum_{y\sim x} (f (y) - f (x)) \\
& = & f (x) - \frac{1}{m(x)}\sum_{y \sim x} f (y) \\
& = & \Delta f (x)\,,
\end{eqnarray*}
where we define the \emph{Laplacian of $f$} to be the quantity
$$
\Delta f (x) := f (x) - \frac{1}{m(x)} \sum_{y\sim x} f (y)\,.
$$
Note that our sign convention for the Laplacian is such that its eigenvalues are positive.
The notion of semi-conformal mapping between graphs was introduced by H. Urakawa in 2000 \cite{Ur-1, Ur-2}. More recently, these have been called \emph{holomorphic mappings} by M. Baker and S. Norine in their development of Riemann surface theory in the context of finite graphs \cite{Ba-No-1, Ba-No-2}. Motivated by our Proposition \ref{prop:holo} below, we shall also refer to these as holomorphic mappings between graphs.
Let $\Gamma_1 = (V_1,E_1)$ and $\Gamma_2 = (V_2, E_2)$ be two (not necessarily finite) graphs. Then a mapping $\phi : V_1 \rightarrow V_2$ between the vertices is defined to be a \emph{mapping of graphs}, if, whenever $x\sim y$ $(x,y\in V_1$) we have, \emph{either} $\phi (x) = \phi (y)$, \emph{or} $\phi (x) \sim \phi (y)$. In this case we will write: $\phi : \Gamma_1 \rightarrow \Gamma_2$.
\begin{definition} \label{def:semi-conf} Let $\phi : \Gamma_1 = (V_1, E_1) \rightarrow \Gamma_2 = (V_2, E_2)$ be a mapping of graphs. Then we say that $\phi$ is {\rm holomorphic at} $x\in V_1$ if, on setting $z = \phi (x)$, for all $z'\sim z$, the number
$$
\lambda (x, z'):= \sharp \{ x'\sim x : \phi (x') = z'\}\,,
$$
is well-defined and depends only on $x$ (i.e. it is independent of the choice of $z'$) in which case we write $\lambda (x) = \lambda (x, z')$. We say that $\phi$ is {\rm holomorphic} if it is holomorphic at every point. In this case, if $x\in V_1$ is such that $\phi (y) = \phi (x)$ for all $y\sim x$, we set $\lambda (x) = 0$ and so obtain a well-defined function $\lambda : V_1 \rightarrow \NN$, called the {\rm dilation} of $\phi$.
\end{definition}
The above definition can easily be extended to mappings of metric graphs, where now $\Gamma_1$ and $\Gamma_2$ are endowed length functions $\ell_1, \ell_2$ defined on the edges $E_1, E_2$, respectively \cite{An}. The dilation is then replaced by the function
$$
\lambda (x) = \ell_2(\ov{\phi (x)z'})\sum_{\stackrel{x'\sim x}{\phi (x') = z'}}\frac{1}{\ell_1(\ov{xx'})}\,.
$$
An \emph{automorphism} of a graph $\Gamma = (V,E)$ is a bijective mapping $\phi : V \rightarrow V$ such that $x\sim y$ if and only if $\phi (x) \sim \phi (y)$. It follows that an automorphism is holomorphic with dilation identically equal to $1$. We interpret such a mapping as the analogue of an \emph{isometry} in the setting of smooth manifolds. Thus a semi-conformal map generalizes this notion.
Given a graph $\Gamma = (V,E)$ and a vertex $x\in V$, then a function $f : V \rightarrow \RR$ is harmonic at $x$ if $\Delta f(x) = 0$ -- we will call such a function a \emph{local harmonic function}. In \cite{Ur-1, Ur-2} it is shown that a mapping between graphs pulls back local harmonic functions to local harmonic functions if and only if it is semi-conformal. This concept is the discrete analogue of a \emph{harmonic morphism} \cite{Ba-Wo-3}.
We now introduce one of the fundamental objects of our study, namely a \emph{holomorphic function} on a graph.
\begin{definition} Let $\Gamma = (V,E)$ be a (not necessarily finite) graph, then a function $\phi : V\rightarrow \CC$ is called {\rm holomorphic at} $x\in V$ if
$$
\sum_{y\sim x} (\dd \phi (\ov{xy}))^2 = \sum_{y\sim x} (\phi (y) - \phi (x))^2 = 0\,.
$$
We say that $\phi : \Gamma \rightarrow \CC$ is {\rm holomorphic} if it is holomorphic at every vertex $x\in V$.
\end{definition}
The notion is a natural adaptation of that of a semi-conformal mapping $\phi : M^m \rightarrow \CC$ from a Riemannian $m$-manifold into the complex plane, as discussed in Section \ref{sec:SFR}. For, $\phi : U \subset \RR^2 \rightarrow \CC$ is semi-conformal if and only if
$$
\left( \frac{\partial \phi}{\partial x}\right)^2 + \left( \frac{\partial \phi}{\partial y}\right)^2 = 4 \frac{\partial \phi}{\partial z}\frac{\partial \phi}{\partial \ov{z}} = 0\,.
$$
That is, if and only if $\phi$ is holomorphic or anti-holomorphic. But on a graph, we do not a priori have a notion of orientation, which in the plane is precisely what distinguishes holomorphic from anti-holomorphic, which justifies the above definition. However, we do sacrifice \emph{linearity} in the equation for holomorphicity, which is an essential ingredient in the study by Baker and Norine who develop their theory using harmonic functions.
\medskip
\begin{center}
\setlength{\unitlength}{0.254mm}
\begin{picture}(244,92)(15,-156)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-120)(120,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-80)(240,-100)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(240,-100)(160,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-140)(40,-120)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-80)(200,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(200,-80)(240,-100)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-140)(80,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-140)(40,-120)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(240,-100)(210,-130)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(210,-130)(160,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-120)(70,-90)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(70,-90)(120,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(200,-80){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(240,-100){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(210,-130){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(160,-140){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(80,-140){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(40,-120){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(70,-90){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(120,-80){\ellipse{4}{4}}
\special{color rgb 0 0 0}\put(215,-141){\shortstack{$0$}}
\special{color rgb 0 0 0}\put(245,-101){\shortstack{$\ii$}}
\special{color rgb 0 0 0}\put(195,-76){\shortstack{$1+\ii$}}
\special{color rgb 0 0 0}\put(110,-76){\shortstack{$1+2\ii$}}
\special{color rgb 0 0 0}\put(53,-84){\shortstack{$2+2\ii$}}
\special{color rgb 0 0 0}\put(11,-114){\shortstack{$2+\ii$}}
\special{color rgb 0 0 0}\put(70,-156){\shortstack{$1+\ii$}}
\special{color rgb 0 0 0}\put(160,-156){\shortstack{$1$}}
\special{color rgb 0 0 0}
\end{picture}
\end{center}
\medskip
\begin{center}
{\small Figure 1: \emph{Example of a finite graph endowed with a holomorphic function}}
\end{center}
\medskip
We now prove an analogue in the context of holomorphic functions, of a theorem of Urakawa \cite{Ur-1, Ur-2}, that holomorphic (or semi-conformal) mappings between graphs are characterized as those mappings which preserve harmonic functions.
\begin{proposition} \label{prop:holo}
Let $\phi : \Gamma_1 = (V_1, E_1) \rightarrow (V_2, E_2)$ be a mapping between graphs. Then $\phi$ is holomorphic if and only if it preserves local holomorphic functions, that is, if $f: V_2 \rightarrow \CC$ is holomorphic at $\phi (x)$ $(x\in V_1$), then $f\circ \phi$ is holomorphic at $x$. In particular, if $\phi : \Gamma_1 \rightarrow \Gamma_2$ is holomorphic, then $f\circ \phi$ is also holomorphic for every holomorphic function $f : V_2 \rightarrow \CC$.
\end{proposition}
\noindent \emph{Proof}: Suppose that $\phi : \Gamma_1 \rightarrow \Gamma_2$ is holomorphic and let $f: V_2 \rightarrow \CC$ be holomorphic at $y \in V_2$. Consider the function $f \circ \phi$. We show that it is holomorphic at each point $x$ with $\phi (x) = y$. Now
\begin{eqnarray*}
\sum_{x'\sim x}\Big( (f\circ \phi )(x') - (f \circ \phi )(x)\Big)^2 & = & \sum_{x'\sim x}(f(\phi (x')) - f(y))^2 \\
& = & \lambda (x) \sum_{y'\sim y} (f(y')- f(y))^2 = 0 \,,
\end{eqnarray*}
by the holomorphicity of $\phi$.
Conversely, suppose that $\phi : \Gamma_1 \rightarrow \Gamma_2$ preserves local holomorphic functions. Let $y \in V_2$ and let $x \in \phi^{-1}(y) \in V_1$. If there is only one vertex $y_1 \sim y$, then the condition of holomorphicity at $x$ is trivially satisfied, so we may suppose there are at least two distinct vertices joined by an edge to $y$. Let $y_1, y_2 \sim y$. We want to show that $\lambda (x, y_1) = \lambda (x, y_2)$. Consider the function $f$ holomorphic at $y$ given by $f(y) = 0$, $f(y_1) = \ii$, $f(y_2) = 1$ and $f(y') = 0$ for all $y'\sim y$ with $y' \neq y_1, y_2$. By hypothesis, $f\circ \phi$ is holomorphic at $x$, so that, if $x_1, \ldots , x_r \sim x$ satisfy $\phi (x_1) = \cdots = \phi (x_r) = y_1$ and $x_{r+1}, \ldots , x_{r+s} \sim x$ satisfy $\phi (x_{r+1}) = \cdots = \phi (x_{r+s}) = y_2$, then
$$
\sum_{x'\sim x} \Big((f\circ \phi )(x') - (f\circ \phi )(x)\Big)^2 = - r + s\,,
$$
which must vanish, so that $r = s$ and $\lambda (x, y_1 ) = \lambda (x, y_2)$. Since $y_1, y_2 \sim y$ are arbitrarily chosen, we conclude that $\phi$ is holomorphic.
\hfill q.e.d.
\medskip
We will consider a pair $(\Gamma , \phi )$, of a graph together with a holomorphic function $\phi : \Gamma \rightarrow \CC$, as a (static) field. Later on, we will consider how to introduce a dynamic into the field. It may be appropriate in the context of quantum field theory to view $\phi$ as a probability amplitude defined at each vertex. Note that if $\phi : \Gamma \rightarrow \CC$ is a holomorphic function, then so is $c\phi + a$ for any complex constants $a,c \in \CC$.
A holomorphic function can be viewed as a special case of a more general object, which we refer to as an isotropic $1$-form.
\begin{definition} \label{def:isotropic-form} Let $\omega$ be a $1$-form defined on a graph $\Gamma = (V,E)$. Then we call $\omega$ {\rm isotropic} if
$$
\sum_{y\sim x}(\omega (\ov{xy}))^2 = 0\,,
$$
at each vertex $x\in V$.
\end{definition}
Then the derivative $\dd\phi$ of a holomorphic function is an isotropic $1$-form. Conversely, we require an integrability condition on $\omega$ in order that it be the derivative of a function. This amounts to the requirement that $\sum_k\omega (e_k)$ should vanish around any cycle $\{ e_k\}_k$ (a cycle being a sequence of directed edges $\{ e_1, e_2, \ldots , e_r\}$ such that the point of arrival of $e_k$ is the start point of $e_{k+1}$ with $e_{r+1}$ then being identified with $e_1$). For if this is the case, then we define $\phi$ at a fixed vertex $x_0$, say to take the value $\phi_0$ and then set $\phi (y) = \phi_0 + \omega (\vec{xy})$ for $y\sim x$. Continuation of this process to all vertices is well-defined on account of the cycle condition.
A quantum graph is a metric graph, such that each edge supports a solution to the $1$-dimensional Schr\"odinger equation with a compatibility condition at each vertex, see \cite{Gn-Sm} and the references cited therein. We can view the pair $(\Gamma , \phi )$ of a graph endowed with a holomorphic function as a similar structure, where we replace a solution to the $1$-dimensional Schr\"odinger equation on an edge $\vec{xy}$ by the amplitude $\phi (y) - \phi (x)$. The compatibility condition at each vertex becomes $\sum_{y\sim x} (\phi (y) - \phi (x))^2 = 0$.
A spin network, in its more recent formulation, consists of a graph where each edge has a label which corresponds to a representation of a particular group. To each vertex is associated an intertwiner which relates these different representations. The original spin networks of Penrose consist of regular graphs with each vertex having degree $3$ and with associated group $\SU (2)$ \cite{Ro-Sm}. Note that the character of an irreducible representation is an algebraic integer, that is, it is the root of some monic equation. We do not know if there may be a deeper connection between spin networks and pairs $(\Gamma , \omega )$, where $\omega$ is an isotropic $1$-form with the different values $\omega (\vec{xy})$ corresponding to characters of representations satisfying polynomial identities at each vertex.
Another interesting construction is the following. Given $n$ complex numbers $z_1, z_2, \ldots , z_n$ satisfying $\sum_{k = 1}^n z_k{}^2 = 0$, then one can construct an $n$-dimensional cube in $\RR^n$ such that there exists an orthogonal projection from $\RR^n$ onto $\CC$ which maps one vertex $v_0\in \RR^n$ to $0\in \CC$ and its neighbouring vertices $v_1, v_2, \ldots , v_n \in \RR^n$ to the points $z_1, z_2, \ldots , z_n$. Conversely, given any orthogonal projection $\pi :\RR^n \rightarrow \CC$, then the complex numbers $z_k = \pi (v_k-v_0)$ satisfy $\sum z_k{}^2 = 0$. This property is known under the name of \emph{Gauss' fundamental theorem of axonometry} \cite{Ga} and when $n = 3$, the projection of the vertices is known as \emph{orthographic projection}. For example, the three dimensional cube supports the holomorphic function indicated in Figure 2.
\medskip
\begin{center}
\setlength{\unitlength}{0.254mm}
\begin{picture}(180,167)(25,-176)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-40)(80,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-80)(160,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-60)(120,-20)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-20)(40,-40)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-80)(80,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-60)(160,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-140)(80,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-40)(40,-120)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-120)(80,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-140)(120,-100)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-100)(120,-75)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-60)(120,-20)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-120)(75,-110)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-100)(95,-105)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(160,-60){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(160,-140){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(120,-100){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(120,-20){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(40,-40){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(80,-80){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(40,-120){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(80,-160){\ellipse{4}{4}}
\special{color rgb 0 0 0}\put(165,-156){\shortstack{$-1$}}
\special{color rgb 0 0 0}\put(80,-176){\shortstack{$0$}}
\special{color rgb 0 0 0}\put(25,-126){\shortstack{$1$}}
\special{color rgb 0 0 0}\put(130,-101){\shortstack{$0$}}
\special{color rgb 0 0 0}\put(170,-66){\shortstack{$-1+\sqrt{2}\ii$}}
\special{color rgb 0 0 0}\put(130,-21){\shortstack{$\sqrt{2}\ii$}}
\special{color rgb 0 0 0}\put(80,-71){\shortstack{$\sqrt{2}\ii$}}
\special{color rgb 0 0 0}\put(24,-29){\shortstack{$1+\sqrt{2}\ii$}}
\special{color rgb 0 0 0}
\end{picture}
\end{center}
\medskip
\begin{center}
{\small Figure 2: \emph{The $1$-skeleton of the cube endowed with a holomorphic function}}
\end{center}
\medskip
The projection of the vertices of other regular polyhedra, satisfy other polynomial equations. For example, the equation $(z_1 + \cdots z_n)^2 - (n+1)(z_1{}^2 + \cdots + z_n{}^2) = 0$ is satisfied by the orthogonal projections $z_1, \ldots , z_n$ of the vertices of a regular tetrahedron \cite{Ea-Pe}.
The case of the cube shows how we can see an $n$-dimensional space arising from a regular graph $(\Gamma , \phi )$ with common vertex degree $n$ endowed with a holomorphic function. Specifically, at each vertex $x$, the complex numbers $\phi (y) - \phi (x)$ $(y\sim x)$ generate a cube in $\RR^n$.
On some infinite graphs, the construction of a holomorphic function can be easily achieved. For example, let $\Gamma$ be the integer lattice in $\RR^N$, with edges joining vertices whose components differ by $1$ in a single entry. Then given any complex valued function $g_0$ defined on the set $\{ (x_1, x_2, \ldots , x_{N-1}, 0)\in \ZZ^N\}$ and another one $g_1$ defined on $\{ (x_1, x_2, \ldots , x_{N-1}, 1)\in \ZZ^N\}$, we can now construct a holomorphic function $\phi$ by extension. Explicitly, $\phi (x_1, x_2, \ldots , x_{N-1}, 2)$ is obtained by solving the equation
\begin{eqnarray*}
\sum_{k = 1}^{N-1} \left\{ \big( g_1 (x_1, \ldots x_k - 1, \ldots , x_{N-1}, 1) - g_1 (x_1, \ldots x_k, \ldots , x_{N-1}, 1)\big)^2\right\} \\
+ \big( g_0 (x_1, \ldots , x_{N-1}, 0) - (g_1 (x_1, \ldots , x_{N-1}, 1)\big)^2 \qquad \\ + \big(\phi (x_1, \ldots , x_{N-1}, 2) - (g_1 (x_1, \ldots , x_{N-1}, 1)\big)^2 = 0
\end{eqnarray*}
for $\phi (x_1, \ldots , x_{N-1}, 2)$, and so on. In general, at each step there will be two solutions and so infinitely many branches will be defined on $\RR^N$. We can view such holomorphic functions as solving an intitial value problem: given a function $g$ and its normal derivative on a hypersurface $S$, find a holomorphic function $\phi$ which coincides with $g$ and has the same normal derivative on $S$. However, finding \emph{finite} graphs which support a holomorphic function seems much harder and at present, using a computer, we can only test examples with a small number of vertices. For example, MAPLE fails to find a holomorphic function on the $1$-skeleton of the dodecahedron in a reasonable time, however, it does show the existence of isotropic $1$-forms.
We now wish to show how, given a graph endowed with a holomorphic function, we can recover a spinor field on the graph. Let $\Gamma = (V,E)$ be a regular graph with common vertex degree $m$. An \emph{orientation on} $\Gamma$ is a colouring of the edges of the graph with the numbers $1, 2, \ldots , m$. By a \emph{colouring}, we mean an assignment of a number $k \in \{ 1, 2, \ldots , m\}$ to each edge so that no two edges incident with the same vertex have the same colour. For example, the $1$-skeleton of the cube, above, is coloured as follows:
\medskip
\begin{center}
\setlength{\unitlength}{0.254mm}
\begin{picture}(168,152)(20,-166)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-40)(80,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-80)(160,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-60)(120,-20)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-20)(40,-40)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-80)(80,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-40)(40,-120)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(40,-120)(80,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-160)(160,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-140)(160,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(160,-140)(120,-100)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-100)(120,-75)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-60)(120,-20)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(120,-100)(95,-105)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(75,-110)(40,-120)
\special{color rgb 0 0 0}\put(125,-166){\shortstack{$1$}}
\special{color rgb 0 0 0}\put(170,-106){\shortstack{$2$}}
\special{color rgb 0 0 0}\put(85,-126){\shortstack{$2$}}
\special{color rgb 0 0 0}\put(65,-26){\shortstack{$3$}}
\special{color rgb 0 0 0}\put(100,-71){\shortstack{$3$}}
\special{color rgb 0 0 0}\put(20,-86){\shortstack{$2$}}
\special{color rgb 0 0 0}\put(45,-151){\shortstack{$3$}}
\special{color rgb 0 0 0}\put(65,-61){\shortstack{$1$}}
\special{color rgb 0 0 0}\put(145,-41){\shortstack{$1$}}
\special{color rgb 0 0 0}
\end{picture}
\medskip
\begin{center}
{\small Figure 3: \emph{The $1$-skeleton of the cube with an orientation giving colouring}}
\end{center}
\medskip
\end{center}
Let $\Gamma = (V,E)$ be a regular graph of degree $3$ which is oriented by the colours $\{ 1,2,3\}$. Suppose further, that $\Gamma$ is endowed with an isotropic $1$-form $\omega$. Then, given a vertex $x\in V$, we can associate to $x$ a triple of complex numbers $\xi (x) = (\xi_1, \xi_2, \xi_3)$, where $x\sim y_1,y_2,y_3$, $\xi_k = \omega (\vec{xy_k})$ and we suppose the edge $\ov{xy_k}$ has colour $k$ $(k = 1,2,3)$. Since $\xi_1{}^2 + \xi_2{}^2 + \xi_3{}^2 = 0$, the symmetric matrix
$$
(\Omega^{AB}):= \left( \begin{array}{cc} - \xi_2 - \xi_3\ii & \xi_3 \\ \xi_3 & \xi_2 - \xi_3\ii
\end{array}
\right) \quad (A,B \in \{ 0,1\} )
$$
has determinant zero and so can be written in the form $\Omega^{AB} = \mu^A\mu^B$, for some spinor $(\mu^A) \in \CC^2$ (defined up to sign). We therefore have a spinor field $\mu^A$ on $\Gamma$ that provides the analogue of the spinor field on $\RR^3$ which generates an SFR in Minkowski space, as described in Section \ref{sec:SFR}.
We can proceed further and construct the analogue of the vector field $U$ (tangent to the associated conformal foliation in the smooth case) at each vertex. In fact, $\mu = \mu^0/\mu^1 = - (\xi_2 + \ii \xi_3)/\xi_1$ represents the direction of $U$ in the chart given by stereographic projection, so that
$$
U = \frac{1}{|\xi_1|^2 + |\xi_2 + \ii \xi_3|^2}\left( |\xi_2 + \ii \xi_3|^2 - | \xi_1|^2, - \ov{\xi_1}(\xi_2 + \ii \xi_3)\right)\,.
$$
It is now possible to consider the discrete analogue of equation (\ref{evol-scm}):
$$
\dd \left( \frac{\partial\phi_n}{\partial n}\right) (U) = - \Delta \phi_n\,,
$$
for a family of complex-valued functions $\{ \phi_n\}$ parametrized by the natural numbers, equivalently:
\begin{equation} \label{evol-scm-discrete}
\dd \phi_{n+1}(U) = - \Delta \phi_n\,.
\end{equation}
However, care needs to be taken in the choice of sign of $U$ when applying this equation, since our construction has essentially only found a \emph{non-oriented} direction $U$ at each vertex. In the case when $\phi_n$ is a given \emph{holomorphic} function, we can ask whether (\ref{evol-scm-discrete}) determines successive functions $\phi_{n+1}$ which are also holomorphic. We do not have a general result to this effect, but it does turn out to be the case for the graph consisting of the $1$-skeleton of the cube. The following table constructs the successive holomorphic function, which is unique up to addition of a constant.
\medskip
\begin{center}
{\scriptsize
\begin{tabular}{|c|c|c|c|c|c|} \hline
Vertex & $\xi$ & $U$ & $\dd \phi_{n+1}(U)$ & $\Delta \phi_n$ & $\phi_{n+1}$ \\ \hline
$1$ & $(1, \sqrt{2}\ii , 1)$ & $\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(2) - \phi_{n+1}(7))$ & $- \frac{\sqrt{2}}{3}(\sqrt{2} + \ii )$ & $\frac{2\sqrt{2}}{3}$ \\
$2$ & $(-1, \sqrt{2}\ii , -1)$ & $\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(1) - \phi_{n+1}(8))$ & $ \frac{\sqrt{2}}{3}(\sqrt{2} - \ii )$ & $\frac{2(\sqrt{2}+\ii )}{3}$ \\
$3$ & $(1, -\sqrt{2}\ii , 1)$ & $\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(5) - \phi_{n+1}(4))$ & $- \frac{\sqrt{2}}{3}(\sqrt{2} - \ii )$ & $\frac{2(\sqrt{2}+ \ii )}{3}$ \\
$4$ & $-(1, \sqrt{2}\ii , 1)$ & $-\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(3) - \phi_{n+1}(6))$ & $ \frac{\sqrt{2}}{3}(\sqrt{2} + \ii )$ & $\frac{2\sqrt{2}}{3}$ \\
$5$ & $-(1, \sqrt{2}\ii , 1)$ & $-\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(3) - \phi_{n+1}(6))$ & $ \frac{\sqrt{2}}{3}(\sqrt{2} + \ii )$ & $\frac{2\ii}{3}$ \\
$6$ & $(1, -\sqrt{2}\ii , 1)$ & $-\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(5) - \phi_{n+1}(4))$ & $ -\frac{\sqrt{2}}{3}(\sqrt{2} - \ii )$ & $0$ \\
$7$ & $(-1, \sqrt{2}\ii , -1)$ & $-\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(1) - \phi_{n+1}(8))$ & $ \frac{\sqrt{2}}{3}(\sqrt{2} - \ii )$ & $0$ \\
$8$ & $(1, \sqrt{2}\ii , 1)$ & $-\frac{1}{\sqrt{2}}(1,0,-1)$ & $\frac{1}{\sqrt{2}}(\phi_{n+1}(2) - \phi_{n+1}(7))$ & $ -\frac{\sqrt{2}}{3}(\sqrt{2} + \ii )$ & $\frac{2\ii}{3}$ \\ \\ \hline
\end{tabular}
}
\end{center}
\medskip
\section{The twistor correspondence between graphs} \label{sec:line-graph}
Twistor space, as first introduced by R. Penrose \cite{Pe-1}, is the space whose points correspond to light rays in Minkowski space. More precisely, there is a $5$-real dimensional CR-submanifold of $\CP^3$ whose points are the light rays. In order to complete the picture it is necessary to compactify and to complexify $\MM^4$ to the complex Grassmannian $G_2(\CC^4)$ of complex $2$-planes through the origin in $\CC^4$. Via the twistor double fibration, a point of $\CP^3$ now determines an $\alpha$-plane in $G_2(\CC^4)$, which, if it intersects the real space $\MM^4$, does so in a null geodesic (see, for example \cite{Wa-We}.
On the other hand, associated to the three-dimensional space forms is their \emph{mini-twistor space}: the space of all geodesics. For example, the mini-twistor space of $\RR^3$ is the complex surface given by the tangent bundle to the $2$-sphere: $TS^2$; each line in $\RR^3$ being defined by its direction $u \in S^2$ and its displacement from the origin $c \in T_uS^2$ ($c$ is the unique vector starting at the origin which hits the line at right angles) (see, \cite{Ba-Wo-3}). In view of these correspondences, it is very natural to define the twistor dual of a graph to be the graph whose vertices are the edges of the original graph, where two vertices are connected if and only if the corresponding edges in the original graph are incident. This dual graph is a well-known classical concept called the line-graph.
Precisely, given a graph $\Gamma = (V,E)$, then the \emph{line-graph} or \emph{twistor dual of} $\Gamma$ is the graph $L_{\Gamma } = (E,T)$, where, for $X,Y \in E$, we have $X\sim Y$ if and only if $X$ and $Y$ are incident in $\Gamma$. The only connected graph that is isomorphic to its line-graph is a cyclic graph and H. Whitney showed that, with the exception of the graphs $K_3$ (the complete graph on three vertices) and $K_{1,3}$ (the bipartite graph with edges joining one vertex to three other unconnected vertices)), any two connected graphs with isomorphic line graphs are isomorphic \cite{Wh}. Not every graph arises as the line-graph of a graph, specifically, there are nine classified graphs, such that provided a given graph $L$ doesn't contain one of them as a subgraph, then $L = L_{\Gamma}$ is the line-graph of some graph $\Gamma$ \cite{Va-Wi, Be}. As an example, Figure 4 shows the line-graph of the graph of Figure 1.
\medskip
\begin{center}
\setlength{\unitlength}{0.254mm}
\begin{picture}(194,124)(43,-162)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-60)(60,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(60,-140)(200,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(200,-140)(220,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(220,-60)(80,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(130,-40)(175,-40)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(175,-40)(220,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(130,-40)(80,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-60)(55,-85)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(55,-85)(45,-120)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(45,-120)(60,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(60,-140)(100,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(100,-160)(150,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(150,-160)(200,-140)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(200,-140)(225,-115)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(225,-115)(235,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(235,-80)(220,-60)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(200,-140){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(225,-115){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(235,-80){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(220,-60){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(175,-40){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(130,-40){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(80,-60){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(55,-85){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(45,-120){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(60,-140){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(100,-160){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\special{sh 0.3}\put(150,-160){\ellipse{4}{4}}
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(150,-160)(175,-145)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(225,-115)(210,-125)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(200,-130)(190,-135)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(235,-80)(215,-115)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(220,-60)(225,-90)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(175,-40)(200,-55)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(225,-75)(235,-80)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(130,-40)(185,-50)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(80,-60)(145,-45)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(55,-85)(75,-70)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(110,-50)(130,-40)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(45,-120)(65,-85)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(60,-140)(55,-110)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(100,-160)(75,-145)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(45,-120)(55,-130)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(60,-140)(150,-160)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(100,-160)(120,-155)
\special{color rgb 0 0 0}\allinethickness{0.254mm}\path(135,-150)(165,-145)
\special{color rgb 0 0 0}
\end{picture}
\end{center}
\medskip
\begin{center}
{\small Figure 4: \emph{The line-graph of the graph of Figure 1}}
\end{center}
\medskip
We can now pursue the twistor correspondence, so that a vertex of a graph $\Gamma$ corresponds to a complete subgraph of the line graph $L_{\Gamma}$. This latter object is then the discrete analogue of the complex projective line corresponding to all the light rays passing through a given point. If now $\Gamma$ is endowed with an isotropic $1$-form $\omega : T\Gamma \rightarrow \CC$, then, on giving each edge a direction, we can define a corresponding \emph{dual} function $\psi : V(L_{\Gamma}) \rightarrow \CC$, by $\psi (X) = \omega (\vec{xy})$, where $X = \ov{xy}$ has direction $\vec{xy}$. It follows that if $x \in V(\Gamma )$ and $C_x$ is the complete subgraph of $L_{\Gamma}$ corresponding to $x$, then
\begin{equation} \label{psi}
\sum_{X\in C_x} \psi (X)^2 = 0\,.
\end{equation}
Note that this latter condition is independent of the choice of direction given to each edge in $\Gamma$.
Conversely, given a graph $L$ which is the line graph of a graph $\Gamma$, and a function $\psi : V(L) \rightarrow \CC$ satisfying $\eqref{psi}$ for each complete subgraph $C_x$ corresponding to a vertex $x \in V(\Gamma )$, then on giving each edge in $\Gamma$ a direction, we can define an isotropic $1$-form on $\Gamma$. If further $\Gamma$ is regular of degree three and oriented by colour, as described in the previous section, we then have a spinor field $\mu^A$ on $\Gamma$ giving the analogue of an SFR. This provides a discrete analogue of the Kerr Theorem, which associates to a complex analytic surface in $\CP^3$, a shear-free ray congruence in Minkowski space.
|
\section{Introduction}
In recent years, systems of dynamical nodes interconnected through a complex
network have attracted a good deal of attention \cite{S01}. Biological and
chemical networks, neural networks, social and economic networks \cite{Jac08}%
, the power grid, the Internet and the World Wide Web \cite{DM03} are
examples of the wide range of applications that motivate this interest (see
also \cite{New03}, \cite{BLMCH06} and references therein). Several modeling
approaches can be found in the literature \cite{DM03}, \cite{WS98}, \cite%
{BA99}. In this paper, we focus our attention on the so-called small-world phenomenon and a model proposed by Newman and Strogatz to replicate this phenomenon.
Once the network is modeled, one is usually interested in two types of
problems. The first involves \emph{structural properties }of the model. The
second involves the performance of \emph{dynamical processes} run on those
networks. In the latter direction, the performance of random walks \cite%
{Lov93}, Markov processes \cite{Bre01}, gossip algorithms \cite{BGPS06},
consensus in a network of agents \cite{OFM07}, \cite{JLM02}, or
synchronization of oscillators \cite{S03}, \cite{PC98}, are very well
reported in the literature. These dynamical processes are mostly studied in
the traditional context of deterministic networks of relatively small size
and/or regular structure. Even though many noteworthy results have been
achieved for large-scale probabilistic networks \cite{LYCC04}--\cite{SBG07},
there is substantial reliance on numerical simulations.
The \emph{eigenvalue spectrum }of an undirected graph contains a great deal
of information about structural and dynamical properties \cite{C97}. In
particular, we focus our attention on the spectrum of the \emph{%
(combinatorial) Laplacian matrix }uniquely associated with an undirected graph \cite{Big93}.
This spectrum contains useful information about, for example, the number of
spanning trees, or the stability of synchronization of a network of
oscillators. We analyze the low-order moments of the Kirchhoff matrix
spectrum corresponding to small-world networks.
The paper is organized as follows. In Section
II, we review the master stability function approach. In Section III, we
derive closed-form expressions for the low-order moments of the Laplacian
eigenvalue distribution associated with a probabilistic small-world network.
Our expressions are valid for networks of asymptotically large size. Section
IV applies our results to the problem of synchronization of a probabilistic
small-world network of oscillators. The numerical results in this section
corroborate our predictions.
\section{Synchronization of Nonlinear Oscillators}
In this section we review the master-stability-function (MSF) approach,
proposed by Pecora and Carrol in \cite{PC98}, to study local stability of
synchronization in networks of nonlinear oscillators. Using this approach,
we reduce the problem of studying local stability of synchronization to the
algebraic problem of studying the spectral support of the Laplacian matrix
of the network. First, we introduce some needed graph-theoretical background.
\subsection{Spectral Graph Theory Background}
In the case of a network with symmetrical connections, undirected graphs
provide a proper description of the network topology. An undirected graph $G$
consists of a set of $N$ nodes or vertices, denoted by $V=\left\{
v_{1},...,v_{n}\right\} $, and a set of edges $E$, where $E\in V\times V$.
In our case, $\left( v_{i},v_{j}\right) \in E$ implies $\left(
v_{j},v_{i}\right) \in E,$ and this pair corresponds to a single edge with
no direction; the vertices $v_{i}$ and $v_{j}$ are called \emph{adjacent}
vertices (denoted by $v_{i}\sim v_{j}$) and are \emph{incident} to the edge $%
\left( v_{i},v_{j}\right) $. We only consider simple graphs (i.e.,
undirected graphs that have no self-loops, so $v_{i}\neq v_{j}$ for an edge $%
\left( v_{i},v_{j}\right) $, and no more than one edge between any two
different vertices). A \emph{walk }on $G$ of length $k$ from $v_{0}$ to $%
v_{k}$ is an ordered set of vertices $\left( v_{0},v_{1},...,v_{k}\right) $
such that $\left( v_{i},v_{i+1}\right) \in E,$ for $i=0,1,...,k-1$; if $\nu
_{k}=\nu _{0}$ the walk is said to be \emph{closed}.
The \emph{degree} $d_{i}$ of a vertex $v_{i}$ is the number of edges
incident to it. The \emph{degree sequence} of $G$ is the list of degrees,
usually given in non-increasing order. The \emph{clustering coefficient},
introduced in \cite{WS98}, is a measure of the number of triangles in a
given graph, where a triangle is defined by the set of edges $\left\{ \left(
i,j\right) ,\left( j,k\right) ,\left( k,i\right) \right\} $ such that $i\sim
j\sim k\sim i$. Specifically, we define clustering as the total number of
triangles in a graph, $T\left( G\right) ,$ divided by the number of
triangles in a complete (all-to-all) graph with $N$ vertices, i.e., the
coefficient is equal to $T\left( G\right) \left/ \binom{N}{3}\right. .$
It is often convenient to represent graphs via matrices. There are several
choices for such a representation. For example, the \emph{adjacency matrix }%
of an undirected graph $G,$ denoted by $A(G)=[a_{ij}]$, is defined
entry-wise by $a_{ij}=1$ if nodes $i$ and $j$ are adjacent, and $a_{ij}=0$
otherwise. (Note that $a_{ii}=0$ for simple graphs.) Notice also that the
degree $d_{i}$ can be written as $d_{i}=\sum_{j=1}^{N}a_{ij}$. We can
arrange the degrees on the diagonal of a diagonal matrix to yield the \emph{%
degree matrix}, $D=diag\left( d_{i}\right) $. The \emph{Laplacian matrix }%
(also called Kirchhoff matrix, or combinatorial Laplacian matrix) is defined
in terms of the degree and adjacency matrices as $L(G)=D(G)-A(G).$ For
undirected graphs, $L(G)$ is a symmetric positive semidefinite matrix \cite%
{Big93}. Consequently, it has a full set of $N$ real and orthogonal
eigenvectors with real non-negative eigenvalues. Since all rows of $L$ sum
to zero, it always admits a trivial eigenvalue $\lambda _{1}=0$, with
corresponding eigenvector $\mathbf{v}_{1}=\left( 1,1,...,1\right) ^{T}$.
The moments of the Laplacian eigenvalue spectrum are central to our paper.
Denote the eigenvalues of our $N\times N$ symmetric Laplacian matrix $L(G)$
by $0=\lambda _{1}\left( G\right) \leq ...\leq \lambda _{N}\left( G\right) $%
. The \emph{empirical spectral density (ESD)} of $L(G)$ is defined as%
\begin{equation*}
\mu _{G}\left( \lambda \right) =\frac{1}{N}\sum_{i=1}^{N}\delta \left(
\lambda -\lambda _{i}\right) ,
\end{equation*}%
where $\delta (\cdot )$ is the Dirac delta function. The $k$\emph{-}th order
moment of the ESD of $L(G)$ is defined as:
\begin{equation*}
q_{k}(G)=\frac{1}{N}\sum_{i=1}^{N}\lambda _{i}(G)^{k}
\end{equation*}%
(which is also called the $k$-th \emph{order spectral moment\footnote{%
Given that our interest is in networks of growing size (i.e., number of
nodes $N$), a more explicit notation for $\mu $ and $q_{k}$ would perhaps
have been $\mu ^{\left( N\right) }$ and $q_{k}^{\left( k\right) }$. However,
for notational simplicity, we shall omit reference to $N$ in there and other
quantities in this paper.}}).
In the following subsection, we illustrate how a network of identical
nonlinear oscillators synchronizes whenever the Laplacian spectrum is
contained in a certain region on the real line. This \emph{region of
synchronization }is exclusively defined by the dynamics of each isolated
oscillator and the type of coupling \cite{PC98}, \cite{LC05}. This
simplifies the problem of synchronization to the problem of locating the
Laplacian eigenvalue spectrum.
\subsection{\label{Sync region}Synchronization as a Spectral Graph Problem}
Several techniques have been proposed to analyze the synchronization of
coupled identical oscillators. In \cite{W01}, well-known results in control
theory, such as the passivity criterion, the circle criterion, and a result
on observer design are used to derive synchronization criteria for an array
of identical nonlinear systems. In \cite{SW04}, the authors use contraction
theory to derive sufficient conditions for global synchronization in a
network of nonlinear oscillators. We pay special attention to the
master-stability-function (MSF) approach, \cite{PC98}. This approach
provides us with a criterion for local stability of synchronization based on
the numerical computation of Lyapunov exponents. Even though quite different
in nature, the mentioned techniques emphasize the key role played by the
graph eigenvalue spectrum.
In this paper we consider a time-invariant network of $N$ identical
oscillators, one located at each node, linked with `diffusive' coupling. The
state equations modeling the dynamics of the network are
\begin{equation}
\mathbf{\dot{x}}_{i}=\mathbf{f}\left( \mathbf{x}_{i}\right) +\gamma
\sum_{j=1}^{N}a_{ij}\Gamma \left( \mathbf{x}_{j}-\mathbf{x}_{i}\right) ,%
\text{ }i=1,...,N \label{network dynamics}
\end{equation}%
\newline
where $\mathbf{x}_{i}$ represents an $n$-dimensional state vector
corresponding to the $i$-th oscillator. The nonlinear function $\mathbf{f}%
\left( \cdot \right) $ describes the (identical) dynamics of the isolated
nodes. The positive scalar $\gamma $ can be interpreted as a global coupling
strength parameter. The $n\times n$ matrix $\Gamma $ represents how states
in neighboring oscillators couple linearly, and $a_{ij}$ are the entries of
the adjacency matrix. By simple algebraic manipulations, one can write down
Eq. (\ref{network dynamics}) in terms of the Laplacian entries, $L(G)=\left[
l_{ij}\right] $, as
\begin{equation}
\mathbf{\dot{x}}_{i}=\mathbf{f}\left( \mathbf{x}_{i}\right) -\gamma
\sum_{j=1}^{N}l_{ij}\Gamma \mathbf{x}_{j}\text{, for }i=1,...,N.
\label{Network Dynamics}
\end{equation}
We say that the network of oscillators is at a synchronous equilibrium if $%
\mathbf{x}_{1}(t)=\mathbf{x}_{2}(t)=...=\mathbf{x}_{N}(t)=\mathbf{\phi }%
\left( t\right) $, where $\mathbf{\phi }\left( t\right) $ represents a
solution for $\mathbf{\dot{x}}=\mathbf{f}\left( \mathbf{x}\right) $. In \cite%
{PC98}, the authors studied the local stability of the synchronous
equilibrium. Specifically, they considered a sufficiently small
perturbation, denoted by $\mathbf{\varepsilon }_{i}(t)$, from the
synchronous equilibrium, i.e.,
\begin{equation*}
\mathbf{x}_{i}(t)=\mathbf{\phi }\left( t\right) +\mathbf{\varepsilon }%
_{i}(t).
\end{equation*}%
After appropriate linearization, one can derive the following equations to
approximately describe the evolution of the perturbations:
\begin{equation}
\mathbf{\dot{\varepsilon}}_{i}=\mathbf{Df}\left( t\right) \,\mathbf{%
\varepsilon }_{i}(t)-\gamma \sum_{j=1}^{n}l_{i,j}\Gamma \,\mathbf{%
\varepsilon }_{j}(t)\text{, for }i=1,...,N. \label{Variational Equation}
\end{equation}%
where $\mathbf{Df}\left( t\right) $ is the Jacobian of $\mathbf{f}\left(
\cdot \right) $ evaluated along the trajectory $\mathbf{\phi }\left(
t\right) $. This Jacobian is an $n\times n$ matrix with time-variant
entries. Following the methodology introduced in \cite{PC98}, Eq. (\ref%
{Variational Equation}) can be similarity transformed into a set of linear
time-variant (LTV) ODEs of the form:
\begin{equation}
\mathbf{\dot{\xi}}_{i}\mathbf{=}\left[ \mathbf{Df}\left( t\right) +(\gamma
\lambda _{i}\left( G\right) )~\Gamma \right] \mathbf{\xi }_{i},\text{ for }%
i=1,...,N, \label{Linear Time-Periodic}
\end{equation}%
where $\{\lambda _{i}\left( G\right) \}_{1\leq i\leq N}$ is the set of
eigenvalues of $L\left( G\right) $. Based on the stability analysis
presented in \cite{PC98}, the network of oscillators in (\ref{network
dynamics}) presents a locally stable synchronous equilibrium if the
corresponding maximal nontrivial Lyapunov exponents of (\ref{Linear
Time-Periodic}) is negative for $i=2,...,N$.
Inspired in Eq. (\ref{Linear Time-Periodic}), Pecora and Carroll studied in
\cite{PC98} the stability of the following parametric LTV-ODE in the
parameter $\sigma $:
\begin{equation}
\mathbf{\dot{\xi}=}\left[ \mathbf{Df}\left( t\right) +\sigma \Gamma \right]
\mathbf{\xi ,} \label{LTP MSF}
\end{equation}%
where $\mathbf{Df}\left( t\right) $ is the linear time-variant Jacobian in
Eq. (\ref{Variational Equation}). The master stability function (MSF),
denoted by $F\left( \sigma \right) $, is defined as the value of the maximal
nontrivial Lyapunov exponent of (\ref{LTP MSF}) as a function of $\sigma $.
Note that $F\left( \sigma \right) $ depends exclusively on $\mathbf{f}\left(
\cdot \right) $ and $\Gamma $, and is independent of the coupling topology,
i.e., independent of $L\left( G\right) $. The region of synchronization is,
therefore, defined by the range of $\sigma >0$ for which $F\left( \sigma
\right) <0$. For a broad class of systems, the MSF is negative in the
interval $\sigma \in \left[ 0,\sigma _{\max }\right] \equiv S$ (although
more generic stability sets are also possible, we assume, for simplicity,
this is the case in subsequent derivations). In order to achieve
synchronization, the set of scaled nontrivial Laplacian eigenvalues, $%
\{\gamma \lambda _{i}\}_{2\leq i\leq N}$, must be located inside the region
of synchronization, $S$. This condition is equivalent to: $\gamma \lambda
_{2}>0$ and $\gamma \lambda _{N}<\sigma _{\max }$.
We illustrate how to use of the above methodology in the following example:
\bigskip
\begin{example}
\label{Rossler example}Study the stability of synchronization of a ring of 6
coupled R\"{o}ssler oscillators \cite{MMZ04}. The dynamics of each
oscillator is described by the following system of three nonlinear
differential equations:
\begin{eqnarray*}
\dot{x}_{i} &=&-\left( y_{i}+z_{i}\right) , \\
\dot{y}_{i} &=&x_{i}+a\,y_{i}, \\
\dot{z}_{i} &=&b+z_{i}\left( x_{i}-c\right) .
\end{eqnarray*}
\end{example}
The adjacency entries, $a_{ij}$, of a ring graph of six nodes are $a_{i,j}=1$
if $j\in \{(i+1)\text{ mod }6,(i-1)\text{ mod }6\}$, for $i=1,2,...,6$, and $%
a_{ij}=0$ otherwise. The dynamics of this ring of oscillators are defined
by:
\begin{equation}
\left[
\begin{array}{c}
\dot{x}_{i} \\
\dot{y}_{i} \\
\dot{z}_{i}%
\end{array}%
\right] =\left[
\begin{array}{c}
-\left( y_{i}+z_{i}\right) \\
x_{i}+a\,y_{i} \\
b+z_{i}\left( x_{i}-c\right)%
\end{array}%
\right] +\gamma \sum_{j\in R\left( i\right) }\left[
\begin{array}{c}
x_{j}-x_{i} \\
0 \\
0%
\end{array}%
\right] \label{Network of Rosslers}
\end{equation}%
where \ we have chosen to connect the oscillators through their $x_{i}$
states exclusively. Our choice is reflected in the structure of the $3\times
3$ matrix, $\Gamma $, inside the summation in Eqn. (\ref{Network of Rosslers}%
).
\bigskip
Numerical simulations of an isolated R\"{o}ssler oscillator unveil the
existence of a periodic trajectory with period $T=5.749$ when the parameters
in Eqn. (\ref{Network of Rosslers}) take the values $a=0.2,\,b=0.2,$ and $%
c=2.5$ (see Fig. 1). We denote this periodic trajectory by $\mathbf{\phi }%
\left( t\right) =\left[ \phi _{x}\left( t\right) ,\phi _{y}\left( t\right)
,\phi _{z}\left( t\right) \right] $. In our specific case, the LTP
differential equation (\ref{LTP MSF}) takes the following form:
\begin{equation}
\mathbf{\dot{\xi}}=\left( \left[
\begin{array}{ccc}
0 & -1 & -1 \\
1 & a & 0 \\
\phi _{z}\left( t\right) & 0 & c%
\end{array}%
\right] +\sigma \left[
\begin{array}{ccc}
1 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & 0%
\end{array}%
\right] \right) \mathbf{\xi ,} \label{LTP ODE Rossler}
\end{equation}%
where the leftmost matrix in the above equation represents the Jacobian of
the isolated R\"{o}ssler evaluated along the periodic trajectory $\mathbf{%
\phi }\left( t\right) ,$ and the rightmost matrix represents $\Gamma $.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Fig1.eps}
\caption{Periodic trajectory with period $T=5.749$ in a R\"{o}ssler oscillator when the parameters
in Eqn. (\ref{Network of Rosslers}) take the values $a=0.2,\,b=0.2,$ and $%
c=2.5$. }
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{Fig2.eps}
\caption{Numerical values of the maximum Floquet exponent of
Eqn. (\ref{LTP ODE Rossler}) for $\sigma \in \left[ 0,15\right] $,
discretizing at intervals of length $0.2$.}
\end{figure}
In Fig. 2, we plot the numerical values of the maximum Floquet exponent of
Eqn. (\ref{LTP ODE Rossler}) for $\sigma \in \left[ 0,15\right] $,
discretizing at intervals of length $0.2$. This plot shows the range in
which the maximal Floquet exponent is negative. This range of stability is $%
S=\left( 0,\sigma ^{\ast }\right) $, for $\sigma ^{\ast }\approx 4.7$. The
MSF criterion introduced in \cite{PC98} states that the synchronous
equilibrium is locally stable if the set of values $\{\gamma ~\lambda
_{i}\left( G\right) \}_{i=2,...,n}$ lies inside the stability range, $S$.
For the case of a 6-ring configuration, the eigenvalues of $L\left( G\right)
$ are $\left\{ {0,1,1,3,3,4}\right\} $, so the set $\{\gamma \lambda
_{i}\}_{i=2,...,n}$ is $\left\{ \gamma ,\gamma ,3\gamma ,3\gamma ,4\gamma
\right\} .$ Therefore, we achieve stability for $\gamma \in (0,\sigma ^{\ast
}/\lambda _{n}\left( G\right) )$, where in our case $\sigma ^{\ast }/\lambda
_{n}\left( G\right) \approx 1.175$.
We now illustrate this result with several numerical simulations. First, we
plot in Fig. 3.a the temporal evolution of the $x_{i}$ states of the 6-ring
when $\gamma =1.0$. Observe how, since $\gamma \in \left( 0,1.175\right) $,
we achieve asymptotic synchronization. On the other hand, if we choose $%
\gamma =1.3\notin \left( 0,1.175\right) $, the time evolution of the set of
oscillators does not converge to a common trajectory (see Fig. 3.b);
instead, the even and odd nodes settle into two different trajectories.
\begin{figure*}
\centering
\includegraphics[width=0.95\linewidth]{Fig3.eps}
\caption{In Fig. a, we plot the temporal evolution of the $x_{i}$ states of the 6-ring
when $\gamma =1.0$. In Fig.b, we plot the time evolution of the set of
oscillators for $\gamma =1.3\notin \left( 0,1.175\right)$.}
\end{figure*}
In the next subsection, we propose an approach to estimating the support of
the eigenvalue distribution of large-scale probabilistic networks from
low-order spectral moments. This allows us to predict synchronization in a
large-scale Chung-Lu network.
\section{Spectral Analysis of Small-World Networks}
In this section we study the Laplacian eigenvalue spectrum of a variant of
Watts-Strogatz small-world network \cite{WS98}. After describing the model,
we use algebraic graph theory to compute explicit expressions for the
Laplacian moments of a small-world network as a function of its parameters.
Our derivations are based on a probabilistic analysis of the expected
spectral moments of the Laplacian for asymptotically large small-world
networks.
\subsection{\label{SW model}Small-World Probabilistic Model}
We consider a one-dimensional lattice of $N$ vertices, $\left\{
v_{1},...,v_{N}\right\} $, with periodic boundary conditions, i.e., on a
ring, and connect each vertex $v_{i}$ to its $2k$ closest neighbors, i.e., $%
v_{i}$ is connected to the set of nodes $\left\{ v_{j}:j\in \left[ \left(
i-k\right) \text{mod}N,\left( i+k\right) \text{mod}N\right] \right\} $.
Then, instead of rewiring a fraction of the edges in the regular lattice as
proposed by Watts and Strogatz \cite{WS98}, we add some random `shortcuts'
to the one-dimensional lattice. These shortcuts are added by independently
assigning edges between each pair of nodes $\left( i,j\right) ,$ $1\leq
i<j\leq N$ with probability $p$. The resulting small-world graph is
intermediate between a regular lattice (achieved for $p=0$) and a classical
random graph (achieved for $p=1$). In general, \emph{small-world networks}
share properties with both the regular grid and the classical random graph
for $0<p<1$. In particular, they show the following apparently contradictory
features:
\emph{(i)} most nodes are not neighbors of one another (such as in a regular
grid), and
\emph{(ii)} most nodes can be reached from every other node by a small
number of steps (such as in a random graph).
An interesting property observed in this model was the following: for small
probability of rewiring, $p\ll 1$, the number of triangles in the network is
nearly the same as that of the regular lattice, but the average
shortest-path length is close to that of classical random graphs. In the
rest of the paper we assume we are in the range of $p$ in which this
property holds, in particular, we will prescribe $p$ to be \thinspace $r/N$,
for a given parameter $r$.
In the coming sections we shall study spectral properties of the Laplacian
matrix associated to the above small-world model. In our derivations we will
need the probabilistic distribution for the degrees. It is well known that,
for asymptotically large graphs, the degree distribution of a classical
random graph with average degree $r$ is a Poisson distribution with rate $r$%
. Hence, the degree distribution of the above small-world network is%
\begin{equation}
\Pr (d_{i}=d)=\left\{
\begin{array}{ll}
0, & \text{for }d<2k, \\
\frac{r^{d-2k}e^{-r}}{(d-2k)!}, & \text{for }d\geq 2k,%
\end{array}%
\right. \label{Shifted Poisson}
\end{equation}%
which corresponds to a Poisson with parameter $r$ `shifted' $2k$ units. The
Poisson distribution is shifted to take into account the degree of the
regular $2k$-neighbors ring superposed to the random shortcuts.
Furthermore, it is well known that the clustering coefficient (or,
equivalently, the number of triangles) of the regular $2k$-neighbors rings
is very lightly perturbed by the addition of random shortcuts for $p=r/N$.
In particular, one can prove the following result:%
\begin{equation}
\mathbb{E}[T]=(1+o(1))\,\frac{1}{3}N\binom{2k}{2}, \label{Triangles SW}
\end{equation}%
where the dominant term, $\frac{1}{3}N\binom{2k}{2}$, corresponds to the
exact number of triangles in a $2k$-neighbors ring with $N$ nodes.
In the following section, we shall derive explicit expressions for the first
low-order spectral moments of the Laplacian matrix associated with the
small-world model herein described. Even though our analysis is far from
complete, in that only low-order moments are provided, valuable information
regarding spectral properties can be retrieved from our results.
\subsection{Algebraic Analysis of Spectral Moments}
In this section we deduce closed-form expressions for the first three
moments of the Laplacian spectrum of any simple graph $G$. First, we express
the spectral moments as a trace using the following identity:
\begin{equation}
q_{k}\left( G\right) =\frac{1}{N}\sum_{i=1}^{N}\lambda _{i}\left( G\right)
^{k}=\frac{1}{N}\,\text{tr}\,L\left( G\right) ^{k}.
\end{equation}%
This identity is derived from the fact that trace is conserved under
diagonalization (in general, under any similarity transformation). In the
case of the first spectral moment, we obtain
\begin{equation*}
q_{1}=\frac{1}{N}\text{tr}\left( D-A\right) =\frac{1}{N}\sum_{i=1}^{N}d_{i}.
\end{equation*}%
where $\overline{d}$ is the average degree of the graph. For analytical and
numerical reasons, we define the normalized Kirchhoff moment as
\begin{equation}
\overline{q}_{k}=\frac{1}{N}\sum_{i=1}^{N}\left( \lambda _{i}/\overline{d}%
\right) ^{k}=\frac{1}{N\,\overline{d}^{k}}\,\text{tr}\left( D-A\right) ^{k}.
\label{normalized moment}
\end{equation}
The fact that $D$ and $A$ do not commute forecloses the possibility of using
Newton's binomial expansion on $\left( D-A\right) ^{k}$. On the other hand,
the trace operator allows us to cyclically permute multiplicative chains of
matrices. For example, tr$\,\left( AAD\right) =$tr$\,\left( ADA\right) =$tr$%
\,\left( DAA\right) $. Thus, for words of length $k\leq 3$, one can
cyclically arrange all binary words in the expansion of (\ref{normalized
moment}) into the standard binomial expression:
\begin{equation}
\overline{q}_{k}=\sum_{\alpha =0}^{k}\binom{k}{\alpha }\frac{(-1)^{\alpha }}{%
\overline{d}^{k}N}\text{tr\thinspace }\left( A^{\alpha }D^{k-\alpha }\right)
\text{,\quad \quad for }k\leq 3.
\end{equation}
Also, we can make use of the identity tr$\left( A^{\alpha }D^{k-\alpha
}\right) =\sum_{i=1}^{N}\left( A^{\alpha }\right) _{ii}d_{i}^{k-\alpha }$ to
write
\begin{equation}
\overline{q}_{k}=\sum_{\alpha =0}^{k}\sum_{i=1}^{N}\binom{k}{\alpha }\frac{%
(-1)^{\alpha }}{\overline{d}^{k}N}d_{i}^{k-\alpha }\left( A^{\alpha }\right)
_{ii}\text{,\quad \quad for }k\leq 3. \label{newton-like}
\end{equation}
Note that this expression is not valid for $k\geq 4$. For example, for $k=4$%
, we have that tr$\left( AADD\right) \neq $tr$\left( DADA\right) .$
We now analyze each summand in expression (\ref{newton-like}) from a
graph-theoretical point of view. Specifically, we find a closed-form
solution for each term tr$\left( A^{i}D^{j}\right) $, for all pairs $1\leq
i+j\leq 3$, as a function of the degree sequence and the number of triangles
in the network. In our analysis, we make use of the following results from
\cite{Big93}:
\bigskip
\begin{lemma}
The number of closed walks of length $\alpha $ in a graph $G$, joining node $%
i$ to itself, is given by the $i$-th diagonal entry of the matrix $A^{\alpha
}$.
\end{lemma}
\bigskip
\begin{corollary}
Let $G$ be a simple graph. Denote by $t_{i}$ the number of triangles
touching node $i$. Then,
\begin{equation}
\left( A\right) _{ii}=0,\text{ }\left( A^{2}\right) _{ii}=d_{i},\text{ and }%
\left( A^{3}\right) _{ii}=2\,t_{i}. \label{algebraic graph}
\end{equation}
\end{corollary}
\bigskip
After substituting (\ref{algebraic graph}) into (\ref{newton-like}), and
straightforward algebraic simplifications, we obtain the following exact
expression for the low-order normalized spectral moments of a given
Kirchhoff matrix $K$:%
\begin{equation}
\overline{q}_{k}=\left\{
\begin{array}{lc}
1, & \text{for }k=1, \\
\frac{1}{N\,\overline{d}^{2}}\left(
\sum_{i=1}^{N}d_{i}^{2}+\sum_{i=1}^{N}d_{i}\right) , & \text{for }k=2, \\
\frac{1}{N\,\overline{d}^{3}}\left[ \left(
\sum_{i=1}^{N}\,d_{i}^{3}+3\sum_{i=1}^{N}d_{i}^{2}\right) -6\,T\right] , &
\text{for }k=3,%
\end{array}%
\right. \label{low order moments}
\end{equation}%
where $T=\frac{1}{3}\sum_{i=1}^{N}t_{i}$ is the total number of triangles%
\footnote{%
A triangle is defined by a set of (undirected) edges $\left\{ \left(
i,j\right) ,\left( j,k\right) ,\left( k,i\right) \right\} $ such that $i\sim
j\sim k\sim i$.} in the network.
It is worth noting how our spectral results are written in terms of two
widely reported measurements, \cite{New03}: the \emph{degree sequence} and
the \emph{clustering coefficient }(which provides us with the total number
of triangles.) This allows us to compute low-order spectral moments of many
real-world networks without performing an explicit eigenvalue decomposition.
\bigskip
\subsection{\label{SW moments}Probabilistic Analysis of Spectral Moments}
In this section, we use Eq. (\ref{low order moments}) to compute the first
three expected Laplacian moments of the small-world model under
consideration. The expected moments can be computed if we had explicit
expressions for the moments of the degrees, $\mathbb{E}[d_{i}]$, $\mathbb{E}%
[d_{i}^{2}]$, \ and $\mathbb{E}[d_{i}^{3}]$, and the expected number of
triangles, $\mathbb{E}[T]$. Since we know the degree distribution (\ref%
{Shifted Poisson}) for this model, the moments of the degrees can be
computed to be:%
\begin{eqnarray}
\mathbb{E}[d_{i}] &=&r+2k, \label{Degrees SW} \\
\mathbb{E}[d_{i}^{2}] &=&r^{2}+\left( 1+4k\right) \,r+4k^{2}, \notag \\
\mathbb{E}[d_{i}^{3}] &=&r^{3}+15\left( 3+6k\right) \,r^{2}+\left(
1+6k+12k^{2}\right) ~r-8k^{3}. \notag
\end{eqnarray}%
We can therefore substitute the expressions (\ref{Triangles SW}) and (\ref%
{Degrees SW}) in Eq. (\ref{low order moments}) in order to derive the
following expressions for the (non-normalized) expected Laplacian moments
for $N\rightarrow \infty $:%
\begin{eqnarray}
\mathbb{E}[q_{1}] &\rightarrow &\,r+2k, \label{Moments SW} \\
\mathbb{E}[q_{2}] &\rightarrow &\,r^{2}+\left( 4k+2\right) \,r+4k^{2}+2k,
\notag \\
\mathbb{E}[q_{3}] &\rightarrow &r^{3}+\left( 6k+6\right) \,r^{2}+\left(
12k^{2}+18k+4\right) \,r \notag \\
&&+8k^{2}+8k^{3}+2k. \notag
\end{eqnarray}
In the following table we compare the numerical values of the Laplacian
moments corresponding to one random realization of the model under
consideration with the analytical predictions in (\ref{Moments SW}). In
particular, we compute the moments for a network of $N=512$ nodes with
parameters $p=r/N=4/N$ and $k=3.$ It is important to point out that the indicated numerical values are obtained for one realization only, with no
benefit from averaging.%
\begin{equation*}
\begin{tabular}{llll}
\hline
Moment order & 1$^{st}$ & 2$^{nd}$ & 3$^{rd}$ \\ \hline
Numerical realization & 10.14 & 116.96 & 1,467.6 \\
Analytical expectations & 10 & 114 & 1,431 \\
Relative error & 1.38\% & 2.53\% & 2.49\% \\ \hline
\end{tabular}%
\end{equation*}
In the next subsection, we use an approach introduced in \cite{P07}\ to
estimate the support of the eigenvalue distribution using the first three
spectral moments. In coming sections, we shall use this technique to predict
whether the Laplacian spectrum lies in the region of synchronization.
\subsection{\label{Triangular Fitting}Piecewise-Linear Reconstruction of the
Laplacian Spectrum}
Our approach, described more fully in \cite{P07}, approximates the spectral
distribution with a triangular function that exactly preserves the first
three moments. We define a triangular distribution $t\left( \lambda \right) $
based on a set of abscissae $x_{1}\leq x_{2}\leq x_{3}$ as%
\begin{equation*}
t\left( \lambda \right) :=\left\{
\begin{array}{ll}
\frac{h}{x_{2}-x_{1}}\left( \lambda -x_{1}\right) , & \text{for }\lambda \in %
\left[ x_{1},x_{2}\right) , \\
\frac{h}{\left( x_{2}-x_{3}\right) }\left( \lambda -x_{3}\right) , & \text{%
for }\lambda \in \left[ x_{2},x_{3}\right] , \\
0, & \text{otherwise.}%
\end{array}%
\right.
\end{equation*}%
where $h=2/\left( x_{3}-x_{1}\right) $. The first three moments of this
distribution, as a function of the abscissae, are given by%
\begin{eqnarray}
M_{1} &=&\frac{1}{3}\left( x_{1}+x_{2}+x_{3}\right) , \label{Sym Moments} \\
M_{2} &=&\frac{1}{6}\left(
x_{1}^{2}+x_{2}^{2}+x_{3}^{2}+x_{1}x_{2}+x_{1}x_{3}+x_{2}x_{3}\right) ,
\notag \\
M_{3} &=&\frac{1}{10}\left(
x_{1}^{3}+x_{1}^{2}x_{2}+x_{1}^{2}x_{3}+x_{2}^{3}+x_{2}^{2}x_{1}\right.
\notag \\
&&+\left.
x_{2}^{2}x_{3}+x_{3}^{3}+x_{3}^{2}x_{1}+x_{3}^{2}x_{2}+x_{1}x_{2}x_{3}%
\right) . \notag
\end{eqnarray}%
Our task is to find the set of values $\left\{ x_{1},x_{2},x_{3}\right\} $
in order to fit a given set of moments $\left\{ M_{1},M_{2},M_{3}\right\} $.
The resulting system of algebraic equations is amenable to analysis, based
on the observation that the moments are symmetric polynomials\footnote{%
A symmetric polynomial on variables $\left( x_{1},x_{2},x_{3}\right) $ is a
polynomial that is unchanged under any permutation of its variables.}.
Following the methodology in \cite{P07}, we can find the abscissae $\left\{
x_{1},x_{2},x_{3}\right\} $ as roots of the polynomial:
\begin{equation}
x^{3}-\Pi _{1}x^{2}+\Pi _{2}x-\Pi _{3}=0, \label{Polynomial Roots}
\end{equation}%
where
\begin{eqnarray}
\Pi _{1} &=&3\,M_{1}, \label{ESD from moments} \\
\Pi _{2} &=&9\,M_{1}^{2}-6\,M_{2}, \notag \\
\Pi _{3} &=&27\,M_{1}^{3}-36\,M_{1}M_{2}+10\,M_{3}. \notag
\end{eqnarray}
The following example illustrates how this technique provides a reasonable
estimation of the Laplacian spectrum for small-world Networks.
\begin{example}
\label{Triangle example}Estimate the spectral support of the small-world
model described in Subsection \ref{SW model} for parameters $N=512,$ $p=4/N$
and $k=3$. In subsection \ref{SW moments} we computed the expected spectral
moments of this particular network to be $\left\{
M_{1}=10,M_{2}=114,M_{3}=1,431\right\} $. Thus, we apply the above technique
with these particular values of the moments to compute the following set of
abscissae for the triangular reconstruction $\left\{
x_{1}=1.577,x_{2}=8.662,x_{3}=19.76\right\} $. In Fig. 4 we compare the
triangular function that fits the expected spectral moments with the
histogram of the eigenvalues of one random realization of the Laplacian
matrix. We also observe that any random realization of the eigenvalue
histograms of the Laplacian is remarkably close to each other. Although a
complete proof of this phenomenon is beyond the scope of this paper, one can
easily proof using the law of large numbers that the distribution of
spectral moments in (\ref{low order moments}) concentrate around their mean
values.
\end{example}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{Fig4.eps}
\caption{Comparison between the histogram of the eigenvalues of one random realization of the Laplacian
matrix of a small-world model with parameters $N=512,$ $p=4/N$ and $k=3$, and the
triangular function that fits the expected spectral moments.}
\end{figure}
We observe that the above estimation is valid for a large range in the
values of the parameters. For example, in Fig. 5, we compare the values of
the triangular abscissae $x_{1}$ and $x_{3}$ with the extreme points of the
Laplacian spectral support, $\lambda _{2}$ and $\lambda _{n}$, for a
small-world network with $N=512$ nodes, $k=3$, and $p$ in the range of
values $\left[ 1/N:0.01/N:10/N\right] .$ It is important to point out that,
in this case too, the numerical values for the eigenvalues are obtained for
one realization only, with no benefit from averaging. In the next section,
we propose a methodology which uses results presented in previous sections
to predict the local stability of the synchronous state in a small-world
network of oscillators.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{Fig5.eps}
\caption{Comparison between the values of
the triangular abscissae $x_{1}$ and $x_{3}$ with the extreme points of the
Laplacian spectral support, $\lambda _{2}$ and $\lambda _{n}$, for a
small-world network with $N=512$ nodes, $k=3$, and $p$ in the range of
values $\left[ 1/N:0.01/N:10/N\right] .$}
\end{figure}
\section{Analytical Estimation of Synchronization}
In this section we use the expressions in (\ref{Moments SW}) and the
triangular reconstruction in the above subsection to predict synchronization
in a large small-world network of coupled nonlinear oscillators.
Specifically, we study a network of coupled R\"{o}ssler oscillators, as
those in Example \ref{Rossler example}. We build our prediction based on the
following steps:
\begin{enumerate}
\item Determine the \emph{region of synchronization} following the technique
presented in Subsection \ref{Sync region}. As illustrated in Example \ref%
{Rossler example}, the \emph{set of scaled eigenvalues} $\{\gamma \lambda
_{i}^{\left( K\right) }\}_{i=2,...,N}$ must lie in a certain region of
stability, $S$, to achieve synchronization (in our example $S=\left(
0,\sigma ^{\ast }\approx 4.7\right) $).
\item Compute the \emph{expected spectral moments} of the Laplacian
eigenvalue spectrum for a given set of parameters using the set of Eqns. in (%
\ref{Moments SW}).
\item Estimate the \emph{support of the Laplacian eigenvalue spectrum,} $%
\{\lambda _{i}^{\left( K\right) }\}_{i=2,...,N}$,\ using the methodology
presented in Subsection \ref{Triangular Fitting}. From Example \ref{Triangle
example}, we have that $s_{l}=1.57$ and $s_{u}=19.76$ are good estimates of
the lower and upper extremes of the spectral support, respectively.
\item Compare the region of stability in \textbf{Step 1} with the estimation
of the spectral support in \textbf{Step 3}, i.e., $\left( 1.57~\gamma
,19.76~\gamma \right) $.
\end{enumerate}
Following the above steps, one can easily verify that our estimated spectral
support, $\left( 1.57~\gamma ,19.76~\gamma \right) $, lies inside the region
of stability, $\left( 0,\sigma ^{\ast }\approx 4.7\right) $, for $0<\gamma
<4.7/19.76\approx 0.238$. Therefore, the small-world network of $512$
coupled R\"{o}ssler oscillators is predicted to synchronize whenever the
global coupling strength satisfies $\gamma \in \left( 0,0.238\right) $.
\subsection{Numerical Results}
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{Fig6.eps}
\caption{We plot the dynamics of the $x$-states for 512 R\"{o}ssler oscillators (as the one
described in Example \ref{Rossler example}) interconnected through the
Small--World network with $p=4/N$ and $k=3$, in Fig.a. In Fig.b, we observe a clear
exponential convergence of the errors towards zero.}
\end{figure*}
In this section we present numerical simulations supporting our conclusions.
We consider a set of identical $512$ R\"{o}ssler oscillators (as the one
described in Example \ref{Rossler example}) interconnected through the
Small--World network defined in Example \label{Triangle example copy(1)} ($%
p=4/N$ and $k=3$). Using the methodology proposed above, we have predicted
that the synchronous state of this system is locally stable if the coupling
parameter $\gamma $ lies in the interval $\left( 0,0.238\right) $. We run
several simulations with the dynamics of the oscillators presenting
different values of the global coupling strength $\gamma $. For each
coupling strength, we present two plots: \emph{(i)} the evolution of
the\thinspace $512$ $x$-states of the R\"{o}ssler oscillators in the time
interval $0\leq t\leq 40$, and \emph{(ii)} the evolution of $x_{i}\left(
t\right) -\bar{x}\left( t\right) $ for all $i$, where $\bar{x}\left(
t\right) =\frac{1}{N}\sum_{i}x_{i}\left( t\right) $. Since our stability
results are local, we have to carefully choose the initial states for the
network of oscillators. For our particular choice of parameters, the
(isolated) R\"{o}ssler oscillator presents a stable limit cycle. For our
simulations, we have chosen as initial condition for each oscillator in the
network a randomly perturbed version of a particular point of this stable
limit cycle. This particular point is $\mathbf{s}_{0}=\left(
3.5119,-3.5332,0.2006\right) $. We have chosen the perturbed initial state
for the $i$-th oscillator to be $\mathbf{s}_{0}+\mathbf{e}_{i}$, where $%
\mathbf{e}_{i}$ is a uniformly distributed random variable in the $3$%
-dimensional cube $\left[ -2,2\right] ^{3}$, and $\mathbf{e}_{i}$ is
independent of $\mathbf{e}_{j}$ for $i\neq j$.I
In our first simulation, we use a coupling strength $\gamma =0.1\in \left(
0,0.238\right) $; thus, we predict the synchronous state to be locally
stable. Fig. 6 (a) and (b) represents the dynamics $x$-states for the 512
oscillators in the small-world network. In this case, we observe a clear
exponential convergence of the errors to zero. In the second simulation, we
choose $\gamma =0.3\notin \left( 0,0.238\right) $; thus, we predict the
synchronous state to be unstable. In fact, we observe in Figs. 7.a and
7.b how synchronization is clearly not achieved.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{Fig7.eps}
\caption{We plot the dynamics of the $x$-states for the 512 R\"{o}ssler oscillators for $\gamma =0.3\notin \left( 0,0.238\right) $ in Fig.a. We observe in Fig.b. how the errors do not converge to zero.}
\end{figure*}
\section{Conclusions and Future Research}
In this paper, we have studied the eigenvalue distribution of the Laplacian
matrix of a large-scale small-world networks. We have focused our attention
on the low-order moments of the spectral distribution. We have derived
explicit expressions of these moments as functions of the parameters in the
small-world model. We have then applied our results to the problem of
synchronization of a network of nonlinear oscillators. Using our
expressions, we have studied the local stability of the synchronous state in
a large-scale small-world network of oscillators. Our approach is based on
performing a triangular reconstruction matching the first three moments of
the unknown spectral measure. Our numerical results match our predictions
with high accuracy. Several questions remain open. The most obvious
extension would be to derive expressions for higher-order moments of the
Kirchhoff spectrum. A more detailed reconstruction of the spectral measure
can be done based on more moments.
\section{ACKNOWLEDGMENTS}
The first author gratefully acknowledges George C. Verghese and Vincent Blondel for their comments and suggestions on this work.
|
\section{Introduction}
\IEEEPARstart{C}{ontrolling} and optimizing the switching of thin magnetic films in a variety of spintronic devices has been of great interest since the discovery of giant magnetoresistance (GMR) \cite{ref:Binasch}-\cite{ref:Baibich}. Many ways of control have been demonstrated such as spin-torque \cite{ref:Katine}, toggle \cite{ref:Engel} and heat assisted switching \cite{ref:Beech} to mention a few. Today, the search for new applications has pushed the need even further. In thermally assisted MRAM \cite{ref:Prejbeanu} the information is stored by exchange coupling between an antiferromagnetic (AFM) and a ferromagnetic (FM) film. Switching of the FM film is controlled by heating the two films above the N\'eel temperature of the AFM and then cooling down in the presence of an external magnetic field.
Recently a new device based on a novel way to control switching was proposed. By exchange coupling two FM films through a weakly FM alloy the coupling is controlled by changes in temperature. At room temperature the alloy is weakly FM and the two films are exchange coupled through the alloy. At a temperature higher than the Curie temperature the alloy is paramagnetic and the two strongly FM films decouple. Provided the films have separate switching fields their moments can then be aligned either parallel or antiparallel. Either external or internal heating can be used. By using a weakly FM alloy of Ni and Cu, we demonstrate this new way to thermally control the switching of a magnetic film and how it can be used with a GMR read out layer.
\section{Experimental Details}
The complete device structure is shown in Fig 1. All films were deposited on thermally oxidized Si substrates using magnetron sputtering at a base pressure better than $5\cdot10^{-8}$ Torr. The bottom Ni$_{80}$Fe$_{20}$ (NiFe) film was used as an underlayer to promote the growth of AFM Ir$_{20}$Mn$_{80}$ (IrMn) \cite{ref:Berkowitz}. The IrMn was used to exchange bias a Co$_{90}$Fe$_{10}$ (CoFe) reference layer in a CoFe/Cu/NiFe GMR read out layer. A Ni-Cu alloy was cosputtered from Ni and Cu targets such that the Curie temperature of the alloy was just above room temperature. On top of the alloy an exchange biased NiFe layer was deposited. Finally a Ta capping layer was added to prevent oxidation. The argon pressure during sputtering was kept at 3 mTorr. To induce exchange bias at the IrMn/FM interfaces the whole deposition process was performed in a magnetic field of 350 Oe.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{FIG1.pdf}
\caption{The complete device structure deposited on a thermally oxidized Si substrate by magnetron sputtering. During measurements the device is placed in an external magnetic field, $H_{A}=50$ Oe, directed opposite to the AFM pinning direction.}
\label{fig_stack}
\end{figure}
To characterize the samples the temperature was raised using a thin-film heater, during which time current in plane (CIP) GMR measurements were performed. The temperature was controlled through a feed back loop using a type-T thermocouple in close contact with the sample.
Using photolithography the stacks were patterned into strips 50 $\mu$m wide and 1 mm long and then bonded at the edges with aluminum wires. During measurements the current was swept between $\pm$50 mA corresponding to a maximum current density of $10^{6}$A/cm$^{2}$. The resistance of the strips were obtained using four point measurements. While sweeping the current a constant field of 50 Oe was applied directed opposite to the AFM pinning direction.
\section{Results and Discussion}
At temperatures higher than room temperature the Ni-Cu alloy becomes paramagnetic and the magnetic moment of the NiFe read out layer is free to rotate in the external field (see Fig. 1). At temperatures closer to room temperature the alloy is ferromagnetic and effectively exchange couples to the read out layer. Because of this exchange coupling the magnetic moment of the NiFe read out layer will turn and point in the opposite direction to the external field. By varying the temperature, either externally or internally, we can thus control the direction of the NiFe magnetic moment in the read out tri-layer.
\subsection{Antiferromagnetic Pinning}
IrMn was chosen for the AFM pinning because of its high blocking temperature and thermal stability \cite{ref:Samant}. To reach blocking temperatures above 150$^{\circ}$C the films must be 10 nm or thicker \cite{ref:Devasahayam}. However, as the thickness increases the exchange bias decreases. It is important that the bottom CoFe reference layer stays exchange biased when the Ni-Cu alloy goes through the FM to paramagnetic phase transition. For all samples measured the bottom AFM had a thickness of 15 nm and the top AFM a thickness of 12 nm.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{FIG2.pdf}
\caption{Current in plane giant magnetoresistance versus applied field for a sample at (a) room temperature and (b) 100$^{\circ}$C heated with an external heater. The grayed out area indicates the region in which the external field should be applied during device operation.}
\label{fig_heater}
\end{figure}
Fig. 2 shows the GMR versus applied magnetic field for a typical sample (a) at room temperature and (b) at 100$^{\circ}$C. The temperature was controlled with an external heater. At room temperature the shift of the hysteresis loop, $H_{E}$, along the magnetic field axis for the CoFe reference layer is 260 Oe and decreases with increasing temperature. At 100$^{\circ}$C $H_{E}$ has decreased to 130 Oe but is still well separated from the NiFe read out layer which has a loop shift of 20 Oe. By cycling the temperature to look for changes in $H_{E}$ after each completed cycle, it was concluded that neither of the AFM layers reach their blocking temperature. This procedure was repeated for all samples before patterning.
From the data in Fig. 2 it can be seen that the external magnetic field, $H_{A}$, should be in the range 70 to 30 Oe (grey area) applied opposite to the AFM pinning direction. If $H_{A}$ is too large the NiFe moment switches already at room temperature. If $H_{A}$ is too small nothing will happen to the read out layer as the NiFe layer is coupled to the CoFe reference layer --- indicated by the 20 Oe loop shift at 100$^{\circ}$C. This is probably due to N\'eel coupling through the 3.5 nm thick Cu spacer \cite{ref:Kools}. However, any residual exchange coupling through the Ni-Cu spacer cannot be completely ruled out \cite{ref:Hernando}.
\subsection{Ni-Cu Alloying}
The Curie temperature of the Ni-Cu alloy decreases linearly with decreased Ni concentration and reaches 0 K at a concentration of 44 at.\% Ni \cite{ref:Hicks}. In order to get the Curie temperature just above room temperature the concentration of Ni should be between 70-80 at.\% \cite{ref:Dutta}-\cite{ref:Sousa}. By varying the deposition rates of Ni and Cu during co-sputtering different Ni-Cu alloys were obtained. To find the optimal alloy composition the FM to paramagnetic phase transition of the alloys were characterized using a magnetometer with a built in heater. At 72 at.\% Ni the alloy is FM at room temperature and can be used to couple two FM films. At 100$^{\circ}$C the phase transition is completed and the two FM films decouple. In order to completely separate two FM films at high temperatures the thickness of the Ni-Cu alloy has to be in the range 20-30 nm. A possible explanation for this is that the alloy is not completely homogenous after co-sputtering at room temperature but contain regions with higher Curie temperature. If the Ni-Cu spacer is very thin these regions could extend to the alloy interfaces and couple the two FM films. An extreme case would be if the alloy contained clusters of pure Ni with a Curie temperature of 354$^{\circ}$C. Another possible explanation is that the two FM films are coupled by exchange interactions through the spacer, even when the spacer is in the paramagnetic phase. It has been indicated that exchange can propagate through paramagnetic regions on length scales of several nanometers \cite{ref:Hernando}.
\subsection{Thermoelectrically Controlled Switching}
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{FIG3.pdf}
\caption{Resistance versus current in plane for a 50 $\mu$m wide and 1 mm long strip. The circled part shows the region where the read out layer goes from parallel to antiparallel. Inset shows only the GMR transition with the thermal background subtracted.}
\label{fig_IV}
\end{figure}
Fig. 3 shows how the switching of the NiFe read out layer is controlled by current flowing through a 50 $\mu$m wide and 1 mm long strip. The intrinsic heating caused by the current is used to increase the temperature of the strip. At first the resistance changes linearly with temperature, which scales with the square of the heating current. However, at 38 mA the NiFe read out layer looses the exchange coupling and starts to rotate in the external magnetic field. Note that the read out layer is situated close to the center of the stack, and therefore not affected by the self-field of the current through the strip. The nonlinear resistance change continues until a current of 44 mA is reached. At higher currents the two FM films in the read out layer are completely antiparallel and the resistance increases linearly with temperature again. When the current is decreased the sample cools down and the exchange coupling through the Ni-Cu alloy comes back. No degradation of this behavior when cycling the current can be observed. The inset to Fig. 3 shows the change in resistance related to the FM to paramagnetic phase transition only. The thermal background, which scales with the square of the current, has been subtracted.
The change in current required to rotate the magnetization 180$^{\circ}$ depends on a number of factors. First there is the temperature change, $\Delta T$, required for the transition between the parallel and the antiparallel state. By looking at CIP GMR versus field curves for different temperatures (similar to Fig. 2) it can be concluded that $\Delta T$ for the sample in Fig. 3 is 35$^{\circ}$C. Less current would be required if a FM alloy with a sharper phase transition, $\Delta T<35^{\circ}$C, could be found. This would result in lower power consumption and faster switching.
The small hysteresis seen i Fig. 3 is caused by the free layer coercivity. If the read out layer produced the same GMR signal but had a film with higher coercivity, this hysteresis would be more significant and a larger change in current would be required to switch the magnetization back and forth.
Another way to decrease the required current is to increase the change in resistance between the parallel and antiparallel state. Then less current would be needed in order to achieve the same $\Delta T$. In this experiment the device is operated with current flowing in the plane of the film. However, if the device would be operated with current flowing perpendicular to the plane an increase of the magnetoresistance signal with roughly a factor of two is expected \cite{ref:Bass}. Different materials in the read out layer could also give a better signal. In a test to increase the signal the NiFe read out layer was replaced by a CoFe/NiFe/CoFe tri-layer. Fig. 4 shows the resulting minor loop at 100$^{\circ}$C compared to a sample with only a NiFe layer. As expected, the signal increases to almost 3\% but due to the inherent crystalline anisotropy of CoFe the coercivity also increases.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{FIG4.pdf}
\caption{Current in plane giant magnetoresistance versus applied field for a sample with a CoFe/NiFe/CoFe tri-layer as the free layer together with a sample having only a NiFe free layer to show the difference in signal and coercivity. The measurement was done at a temperature of 100$^{\circ}$C. }
\label{fig_read_out}
\end{figure}
The best way to increase the signal and decrease the required current would be to use tunneling magnetoresistance (TMR) instead of GMR for read out \cite{ref:Julliere}. A signal increase with a factor of 10--100 is expected \cite{ref:Yuasa}-\cite{ref:Parkin}. However, because of the rapid drop in TMR with applied voltage the resistance area product of the tunnel barrier must be relatively low in order to achieve the power densities required for heating. For heating a magnetic stack containing a tunnel barrier typical current densities of $10^{6}$A/cm$^{2}$ are used \cite{ref:Prejbeanu}. Assuming the device operates at 300 mV the required resistance area product would be 30 $\Omega \mu $m$^{2}$.
Before demonstrating the above thermoelectrically controlled switching in a device, the switching speed should be considered. For large samples the switching speed is limited by the thermal time constant. In order to measure the time constant the sample was heated to a temperature above room temperature. The heater was then turned off and the resistance change versus time measured. As expected for a 50 $\mu$m wide and 1 mm long strip, a very large time constant of 30s was recorded. If fast switching is required the size of the device would have to be decreased. For submicrometer sized junctions typical time constants are in the order of nanoseconds \cite{ref:Kerekes}. Since the device is operated by changes in current any series inductance could also limit the switching speed and has to be taken into account.
\section{Conclusion}
We have successfully demonstrated thermoelectrically controlled switching of a magnetic film element. The switching is well controlled by external or internal heating. In order to completely reverse the magnetization of the film a temperature increase of 35$^{\circ}$C is needed. No degradation upon thermal cycling is observed. For optimal performance the device should be small in size, have a read out layer with large magnetoresistance and minimal free layer coercivity. This new way to thermoelectrically control magnetic switching in magnetic multilayers is promising for applications such as high frequency oscillators \cite{ref:Kadigrobov}.
\section*{Acknowledgment}
This work was supported by EU-FP7-FET-STELE.
|
\section{Introduction}
A sufficient condition, called {\em nonvanishing determinants}, is shown in~\cite{EKPKL06} for a full-rate (\mbox{$n^2$} input symbols in an \mbox{$n\times n$} transmission matrix) space-time block code achieving the optimal diversity-multiplexing gains tradeoff~\cite{ZT03}. Full-rate space-time block codes with nonvanishing determinants have been extensively designed with cyclic division algebras, e.g.,~\cite{EKPKL06,GX09,KR05,LC09}. For these designs, smaller pairwise error probabilities of maximum likelihood detections require larger normalized diversity products, which can be obtained by choosing integer non-norm elements with smaller absolute values~\cite{GX09,LC09}. All known methods have constructed \mbox{$1+\mathbf i$} and \mbox{$2+\mathbf i$} to be integer non-norm elements with the smallest absolute values over quadrature amplitude modulations (QAM) for the number of transmit antennas \mbox{$n$}: \mbox{$\{n:5\leq n\leq 40,8\nmid n\}$} and \mbox{$\{n:5\leq n\leq 40,8\mid n\}$}, respectively~\cite{LC09}.
Via explicit constructions, this paper proves that \mbox{$1+\mathbf i$} is an integer non-norm element with the smallest absolute value over QAM for every \mbox{$n\geq 5$}. Section~\ref{prelim} introduces some fundamental definitions in algebraic number theory and discusses their properties, which are helpful in deriving the new constructional procedure in Section~\ref{method}. Section~\ref{numerical} confirms numerically the improvement in normalized diversity products by adopting \mbox{$1+\mathbf i$} as a non-norm element instead of \mbox{$2+\mathbf i$}. Finally, conclusions are drawn in Section~\ref{conc}.
Throughout this article, \mbox{$\mathbb Q$} and \mbox{$\mathbb Z$} mean the field consisting of all rational numbers and the ring consisting of all integers, respectively. For a field \mbox{$\mathbb E$}, the sets of all algebraic integers and nonzero elements therein are denoted by \mbox{$\mathcal O_\mathbb E$} and \mbox{$\mathbb E^*$}, respectively. Moreover, \mbox{$\mathbb E_\mathfrak p$} represents the completion of \mbox{$\mathbb E$} with valuation corresponding to a nonzero prime ideal \mbox{$\mathfrak p$} of \mbox{$\mathcal O_\mathbb E$}. \mbox{$\zeta_m$} is a primitive \mbox{$m$-th} root of unity.
\section{Preliminary Knowledge in Algebraic Number Theory}
\label{prelim}
The following two paragraphs briefly mention some useful tools in ramification theory~\cite[Ch. I, Sec. 6]{J96}.
Let \mbox{$\mathbb E$} be a number field and \mbox{$\mathbb F$} be an abelian extension over \mbox{$\mathbb E$} with degree \mbox{$n$}. A nonzero prime ideal \mbox{$\mathfrak P$} of \mbox{$\mathcal O_\mathbb F$} is said to {\em lie over} another nonzero prime ideal \mbox{$\mathfrak p$} of \mbox{$\mathcal O_\mathbb E$}, written as \mbox{$\mathfrak P\mid \mathfrak p$}, if \mbox{$\mathfrak P\cap \mathcal O_\mathbb E=\mathfrak p$}. Each \mbox{$\mathfrak p\mathcal O_\mathbb F$} with \mbox{$\mathfrak p$} a nonzero prime ideal of \mbox{$\mathcal O_\mathbb E$} has the unique (up to a reindexing) factorization in \mbox{$\mathcal O_\mathbb F$}: \mbox{$\mathfrak p\mathcal O_\mathbb F=(\mathfrak P_1\cdots\mathfrak P_g)^{e}$} where \mbox{$\mathfrak P_1,\ldots,\mathfrak P_g$} are distinct nonzero prime ideals of \mbox{$\mathcal O_\mathbb F$} with \mbox{$\mathfrak P_1,\ldots,\mathfrak P_g\mid\mathfrak p$}, and \mbox{$e$} is a positive integer known as the {\em ramification index} of each \mbox{$\mathfrak P_1,\ldots,\mathfrak P_g$} over \mbox{$\mathfrak p$}. Since \mbox{$\mathcal O_\mathbb E$} and \mbox{$\mathcal O_\mathbb F$} are both Dedekind domains, \mbox{$\mathfrak p$} is a maximal ideal in \mbox{$\mathcal O_\mathbb E$} and likewise for those \mbox{$\mathfrak P_1,\ldots,\mathfrak P_g$} in \mbox{$\mathcal O_\mathbb F$}, implying that \mbox{$\mathcal O_\mathbb E/\mathfrak p$} and \mbox{$\mathcal O_\mathbb F/\mathfrak P_1,\ldots,\mathcal O_\mathbb F/\mathfrak P_g$} are all fields. In fact, each \mbox{$\mathcal O_\mathbb F/\mathfrak P_i$} is a field extension of \mbox{$\mathcal O_\mathbb E/\mathfrak p$} with the same degree \mbox{$f=[\mathcal O_\mathbb F/\mathfrak P_i:\mathcal O_\mathbb E/\mathfrak p]$}, called the {\em residue class degree} of \mbox{$\mathfrak P_i$} over \mbox{$\mathfrak p$}, such that
\begin{equation}
\label{sum}
efg=n.
\end{equation}
Of particular interests are the two extreme cases: \mbox{$f=n$} or \mbox{$e=n$}. In either one,~(\ref{sum}) automatically acknowledges that there is exactly one nonzero prime ideal of \mbox{$\mathcal O_\mathbb F$} lying over \mbox{$\mathfrak p$}, i.e., \mbox{$g=1$}. If \mbox{$f=n$} then \mbox{$\mathfrak p$} is said to be {\em inert} in \mbox{$\mathbb F/\mathbb E$}; if \mbox{$e=n$} then \mbox{$\mathfrak p$} is said to be {\em totally ramified} in \mbox{$\mathbb F/\mathbb E$}.
Suppose further that \mbox{$\mathbb L/\mathbb E$} is another abelian extension where \mbox{$\mathbb F/\mathbb E$} is a sub-extension. Let \mbox{$\mathfrak p$}, \mbox{$\mathfrak P$}, and \mbox{$\wp$} be nonzero prime ideals of \mbox{$\mathcal O_\mathbb E$}, \mbox{$\mathcal O_\mathbb F$}, and \mbox{$\mathcal O_\mathbb L$}, respectively, with \mbox{$\mathfrak P\mid \mathfrak p$} and \mbox{$\wp\mid \mathfrak P$}. The respective ramification indices \mbox{$e$}, \mbox{$e'$}, and \mbox{$e''$} of \mbox{$\wp$} over \mbox{$\mathfrak p$}, \mbox{$\wp$} over \mbox{$\mathfrak P$}, and \mbox{$\mathfrak P$} over \mbox{$\mathfrak p$}, and the respective residue class degrees \mbox{$f$}, \mbox{$f'$}, and \mbox{$f''$} of \mbox{$\wp$} over \mbox{$\mathfrak p$}, \mbox{$\wp$} over \mbox{$\mathfrak P$}, and \mbox{$\mathfrak P$} over \mbox{$\mathfrak p$} satisfy
\begin{equation}
\label{tower}
e=e'e'' \mbox{ and } f=f'f''.
\end{equation}
Together~(\ref{sum}) with~(\ref{tower}), if a nonzero prime ideal \mbox{$\mathfrak p$} of \mbox{$\mathcal O_\mathbb E$} is inert (respectively, totally ramified) in \mbox{$\mathbb L/\mathbb E$}, then it is also inert (respectively, totally ramified) in every sub-extension \mbox{$\mathbb F/\mathbb E$} of \mbox{$\mathbb L/\mathbb E$}.
Let's focus on the situation where \mbox{$\mathbb F/\mathbb E$} is a cyclic extension. Recall that the global norm \mbox{$N_{\mathbb F/\mathbb E}(\gamma)$} of \mbox{$\gamma\in\mathbb F$} is $$N_{\mathbb F/\mathbb E}(\gamma)=\prod_{\sigma\in\mathrm{Gal}(\mathbb F/\mathbb E)}\sigma(\gamma)\in \mathbb E.$$ For every \mbox{$\gamma\in\mathbb E^*$}, \mbox{$N_{\mathbb F/\mathbb E}(\gamma)=\gamma^n$} implies \mbox{$\gamma^n\in N_{\mathbb F/\mathbb E}(\mathbb F^*)$}. Hence, the order of \mbox{$\gamma$} modulo \mbox{$N_{\mathbb F/\mathbb E}(\mathbb F^*)$} always divides \mbox{$n$}. An element \mbox{$\gamma\in\mathbb E^*$} is called a {\em non-norm element} of \mbox{$\mathbb F/\mathbb E$} if the order of \mbox{$\gamma$} modulo \mbox{$N_{\mathbb F/\mathbb E}(\mathbb F^*)$} is \mbox{$n$}. For a pair of nonzero prime ideals \mbox{$\mathfrak p$} of \mbox{$\mathcal O_\mathbb E$} and \mbox{$\mathfrak P$} of \mbox{$\mathcal O_\mathbb F$} with \mbox{$\mathfrak P\mid\mathfrak p$}, \mbox{$\mathbb F_\mathfrak P/\mathbb E_\mathfrak p$} is a Galois extension with \mbox{$\mathrm{Gal}(\mathbb F_\mathfrak P/\mathbb E_\mathfrak p)$} isomorphic to a subgroup of \mbox{$\mathrm{Gal}(\mathbb F/\mathbb E)$}~\cite[Ch. III, Thm. 1.2]{J96}. Thus \mbox{$\mathbb F_\mathfrak P/\mathbb E_\mathfrak p$} is a cyclic extension, too. The local norm \mbox{$N_{\mathbb F_\mathfrak P/\mathbb E_\mathfrak p}(\gamma)$} of \mbox{$\gamma\in\mathbb F\subseteq\mathbb F_\mathfrak P$} is $$N_{\mathbb F_\mathfrak P/\mathbb E_\mathfrak p}(\gamma)= \prod_{\sigma\in\mathrm{Gal}(\mathbb F_\mathfrak P/\mathbb E_\mathfrak p)}\sigma(\gamma)\in \mathbb E_\mathfrak p.$$ Lemma~\ref{hasse} connects global and local norms.
\begin{lem}[{\cite[Ch. V, Thm. 4.6]{J96}}]
\label{hasse}
For each \mbox{$\gamma \in\mathbb E^*$}, if \mbox{$\gamma\not\in N_{\mathbb F_\mathfrak P/\mathbb E_\mathfrak p}(\mathbb F_\mathfrak P^*)$} for some pair of nonzero prime ideals \mbox{$\mathfrak p$} of \mbox{$\mathcal O_\mathbb E$} and \mbox{$\mathfrak P$} of \mbox{$\mathcal O_\mathbb F$} with \mbox{$\mathfrak P\mid \mathfrak p$}, then \mbox{$\gamma\not\in N_{\mathbb F/\mathbb E}(\mathbb F^*)$}.
\end{lem}
Lemma~\ref{comp} shows that it suffices to consider those cyclic extensions with degrees equal to prime powers and a pre-specified non-norm element.
\begin{lem}
\label{comp}
Given two cyclic extensions \mbox{$\mathbb F/\mathbb E$} and \mbox{$\mathbb K/\mathbb E$} with degrees \mbox{$n_1$} and \mbox{$n_2$}, respectively, if \mbox{$\gcd(n_1,n_2)=1$} and \mbox{$\gamma\in \mathbb E^*$} is a non-norm element of both extensions, then \mbox{$\mathbb F\mathbb K/\mathbb E$} is a cyclic extension with degree \mbox{$n_1n_2$} and \mbox{$\gamma$} a non-norm element.
\end{lem}
\begin{proof}
From \mbox{$\gcd(n_1,n_2)=1$}, \mbox{$\mathbb F\mathbb K/\mathbb E$} is a cyclic extension with degree \mbox{$n_1n_2$}. Write \mbox{$\mathbb L=\mathbb F\mathbb K$}. Suppose, on the contrary, that \mbox{$\gamma$} is not a non-norm element of \mbox{$\mathbb L/\mathbb E$}; i.e., the order of \mbox{$\gamma$} modulo \mbox{$N_{\mathbb L/\mathbb E}(\mathbb L^*)$} is a proper divisor of \mbox{$n_1n_2$}. Then there exists a prime number \mbox{$q\mid n_1n_2$} such that $$\gamma^{n_1n_2/q}\in N_{\mathbb L/\mathbb E}(\mathbb L^*).$$ Assume that \mbox{$q\mid n_1$} without loss of generality. By the transitivity of norm~\cite[Ch. I, Cor. 5.4]{J96}, $$\gamma^{n_1n_2/q}\in N_{\mathbb F/\mathbb E}(N_{\mathbb L/\mathbb F}(\mathbb L^*))\subseteq N_{\mathbb F/\mathbb E}(\mathbb F^*).$$ In other words, the order of \mbox{$\gamma$} modulo \mbox{$N_{\mathbb F/\mathbb E}(\mathbb F^*)$} has to divide \mbox{$n_1n_2/q$}. It also divides \mbox{$[\mathbb F:\mathbb E]=n_1$}. The order of \mbox{$\gamma$} modulo \mbox{$N_{\mathbb F/\mathbb E}(\mathbb F^*)$} must be a divisor of \mbox{$\gcd(n_1n_2/q,n_1)=n_1/q$}, a contradiction to \mbox{$\gamma$} being a non-norm element of \mbox{$\mathbb F/\mathbb E$}.
\end{proof}
If a nonzero prime ideal \mbox{$\mathfrak p$} of \mbox{$\mathcal O_\mathbb E$} is totally ramified in \mbox{$\mathbb F/\mathbb E$}, and the extension degree \mbox{$[\mathbb F:\mathbb E]$} is not divisible by the characteristic of field \mbox{$\mathcal O_\mathbb E/\mathfrak p$}, then \mbox{$\mathfrak p$} is said to be {\em totally and tamely ramified} in \mbox{$\mathbb F/\mathbb E$}. Some later serviceable instances are illustrated in Example~\ref{ttr}.
\begin{example}
\label{ttr}
Let \mbox{$p$} be a prime number. Then \mbox{$[\mathbb Q(\zeta_p):\mathbb Q]=p-1$} and \mbox{$p$} is totally ramified in \mbox{$\mathbb Q(\zeta_p)/\mathbb Q$}~\cite[Ch. I, Thm. 10.1]{J96}. Since the characteristic of \mbox{$\mathbb Z/p\mathbb Z$} is \mbox{$p$}, \mbox{$p$} is totally and tamely ramified in \mbox{$\mathbb Q(\zeta_p)/\mathbb Q$}.
If further \mbox{$p\equiv 1\pmod{4}$} then there exist two positive integers \mbox{$a$} and \mbox{$b$} such that \mbox{$p=a^2+b^2$}, leading to exactly two distinct nonzero prime ideals of \mbox{$\mathcal O_{\mathbb Q(\mathbf i)}=\mathbb Z[\mathbf i]$} lying over \mbox{$p$}: \mbox{$(a+b\mathbf i)\mathbb Z[\mathbf i]$} and \mbox{$(a-b\mathbf i)\mathbb Z[\mathbf i]$}. Let \mbox{$\mathfrak p$} be either of them. By~(\ref{sum}), \mbox{$\mathfrak p$} has ramification index \mbox{$1$} over \mbox{$p$}. Fix a nonzero prime ideal \mbox{$\wp$} of \mbox{$\mathcal O_{\mathbb Q(\zeta_p,\mathbf i)}$} with \mbox{$\wp\mid \mathfrak p$} and let \mbox{$\mathfrak P=\wp\cap\mathcal O_{\mathbb Q(\zeta_p)}$} accordingly. Then \mbox{$\mathfrak P\mid p$}. Since \mbox{$p$} is totally ramified in \mbox{$\mathbb Q(\zeta_p)/\mathbb Q$}, \mbox{$\mathfrak P$} has ramification index \mbox{$p-1$} over \mbox{$p$}. By~(\ref{tower}), the tower of field extensions \mbox{$\mathbb Q\subset \mathbb Q(\zeta_p)\subset \mathbb Q(\zeta_p,\mathbf i)$} gives that the ramification index of \mbox{$\wp$} over \mbox{$p$} is at least \mbox{$p-1$}. Since \mbox{$\mathfrak p$} has ramification index \mbox{$1$} over \mbox{$p$}, again by~(\ref{tower}) the tower of field extensions \mbox{$\mathbb Q\subset\mathbb Q(\mathbf i)\subset\mathbb Q(\zeta_p,\mathbf i)$} yields that the ramification index of \mbox{$\wp$} over \mbox{$\mathfrak p$} is at least \mbox{$p-1$}. From~(\ref{sum}) and \mbox{$[\mathbb Q(\zeta_p,\mathbf i):\mathbb Q(\mathbf i)]=p-1$}, the ramification index of \mbox{$\wp$} over \mbox{$\mathfrak p$} must be \mbox{$p-1$}; i.e., \mbox{$\mathfrak p$} is totally ramified in \mbox{$\mathbb Q(\zeta_p,\mathbf i)/\mathbb Q(\mathbf i)$}. Moreover, \mbox{$\mathbb Z[\mathbf i]/\mathfrak p$} is a field extension of \mbox{$\mathbb Z/p\mathbb Z$}; thus the characteristic of \mbox{$\mathbb Z[\mathbf i]/\mathfrak p$} must be \mbox{$p$}. \mbox{$\mathfrak p$} is totally and tamely ramified in \mbox{$\mathbb Q(\zeta_p,\mathbf i)/\mathbb Q(\mathbf i)$}.
\end{example}
Theorem~\ref{locnon} presents our new sufficient condition for obtaining non-norm elements.
\begin{thm}
\label{locnon}
Let \mbox{$q$} be a prime number, \mbox{$k$} be a positive integer, \mbox{$\mathbb F/\mathbb E$} be a cyclic extension with degree \mbox{$q^k$}, and \mbox{$\mathfrak p$} be totally and tamely ramified in \mbox{$\mathbb F/\mathbb E$}. If \mbox{$\gamma\in\mathcal O_\mathbb E\setminus\mathfrak p$} and \mbox{$x^q\equiv \gamma\pmod{\mathfrak p}$} is not solvable in \mbox{$\mathcal O_\mathbb E$}, then \mbox{$\gamma$} is a non-norm element of \mbox{$\mathbb F/\mathbb E$}.
\end{thm}
\begin{proof}
Since \mbox{$\mathfrak p$} is totally ramified in \mbox{$\mathbb F/\mathbb E$}, there is a unique nonzero prime ideal \mbox{$\mathfrak P$} of \mbox{$\mathcal O_\mathbb F$} with \mbox{$\mathfrak P\mid\mathfrak p$}. By Lemma~\ref{hasse}, to show that \mbox{$\gamma$} is a non-norm element of \mbox{$\mathbb F/\mathbb E$}, it suffices to prove $$\gamma^{q^{k-1}}\not\in N_{\mathbb F_\mathfrak P/\mathbb E_\mathfrak p}(\mathbb F_\mathfrak P^*).$$ The condition \mbox{$\gamma\in\mathcal O_\mathbb E\setminus\mathfrak p$} gives $$\gamma^{q^{k-1}}\in\mathcal O_{\mathbb E_\mathfrak p}\setminus\mathfrak p\mathcal O_{\mathbb E_\mathfrak p};$$ the condition that \mbox{$\mathfrak p$} is totally and tamely ramified in \mbox{$\mathbb F/\mathbb E$} implies that \mbox{$\mathbb F_\mathfrak P/\mathbb E_\mathfrak p$} is a totally and tamely ramified extension with degree \mbox{$q^k$}~\cite[Ch. II, Thm. 3.8]{J96}. Thereby every \mbox{$\alpha\in\mathcal O_{\mathbb E_\mathfrak p}\setminus\mathfrak p\mathcal O_{\mathbb E_\mathfrak p}$} satisfies \mbox{$\alpha\in N_{\mathbb F_\mathfrak P/\mathbb E_\mathfrak p}(\mathbb F_\mathfrak P^*)$} if and only if $$\alpha\equiv \beta^{q^k}\pmod{\mathfrak p\mathcal O_{\mathbb E_\mathfrak p}}$$ for some \mbox{$\beta\in\mathcal O_{\mathbb E_\mathfrak p}$}~\cite[Ch. IV, Sec. 1.5]{FV02}. Together with \mbox{$\mathcal O_{\mathbb E_\mathfrak p}/\mathfrak p\mathcal O_{\mathbb E_\mathfrak p}\simeq\mathcal O_{\mathbb E}/\mathfrak p$}~\cite[Ch. II, Cor. 2.7]{J96}, it leaves to show that
\begin{equation}
\label{coneq}
x^{q^k}\equiv \gamma^{q^{k-1}}\pmod{\mathfrak p}
\end{equation}
is not solvable in \mbox{$\mathcal O_\mathbb E$}. Suppose, on the contrary, that~(\ref{coneq}) has a solution \mbox{$\theta\in \mathcal O_\mathbb E$}. Since \mbox{$\mathcal O_\mathbb E/\mathfrak p$} is a finite field, \mbox{$(\mathcal O_\mathbb E/\mathfrak p)^*$} is a cyclic group. Let \mbox{$\lambda\in\mathcal O_\mathbb E\setminus \mathfrak p$} be a generator of \mbox{$(\mathcal O_\mathbb E/\mathfrak p)^*$}. Then \mbox{$\theta\equiv\lambda^s\pmod{\mathfrak p}$} and \mbox{$\gamma\equiv\lambda^t\pmod{\mathfrak p}$} for some positive integers \mbox{$s$} and \mbox{$t$} such that $$\lambda^{s q^k}\equiv\lambda^{tq^{k-1}}\pmod{\mathfrak p};$$ i.e., the order of \mbox{$(\mathcal O_\mathbb E/\mathfrak p)^*$} divides \mbox{$q^{k-1}(sq-t)$}. On the other hand, \mbox{$q^k$} must divide the order of \mbox{$(\mathcal O_\mathbb E/\mathfrak p)^*$} because \mbox{$\mathfrak p$} is totally and tamely ramified in \mbox{$\mathbb F/\mathbb E$}~\cite[Ch. IV, Sec. 1.5]{FV02}. Therefore \mbox{$q\mid t$} and \mbox{$\lambda^{t/q}$} becomes a solution of \mbox{$x^q\equiv \gamma\pmod{\mathfrak p}$} in \mbox{$\mathcal O_\mathbb E$}, a contradiction.
\end{proof}
\section{The Generality of \mbox{$1+\mathbf i$} as a Non-Norm Element}
\label{method}
This section first shows that for every odd \mbox{$n$}, a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$n$} and \mbox{$1+\mathbf i$} a non-norm element can always be constructed. This paper aims to establish such an extension for each positive integer \mbox{$n$}. According to Lemma~\ref{comp}, it suffices to further construct a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$2^k$} and \mbox{$1+\mathbf i$} a non-norm element for each positive integer \mbox{$k$}.
Lemma~\ref{spl} has been used extensively in~\cite{EKPKL06,GX09,KR05,LC09}.
\begin{lem}[{\cite[Thm. 1]{KR05}}]
\label{spl}
Let \mbox{$\mathbb F/\mathbb Q$} be a cyclic extension with degree \mbox{$n$} and \mbox{$\mathbb F\cap\mathbb Q(\mathbf i)=\mathbb Q$}, \mbox{$\ell$} be an inert prime number in \mbox{$\mathbb F/\mathbb Q$}, and \mbox{$\mathfrak p$} be a nonzero prime ideal of \mbox{$\mathbb Z[\mathbf i]$} with \mbox{$\mathfrak p\mid \ell$} and residue class degree $1$ over $\ell$. If \mbox{$\gamma\in\mathbb Z[\mathbf i]$} satisfies \mbox{$\gamma\in\mathfrak p\setminus\mathfrak p^2$}, then \mbox{$\gamma$} is a non-norm element of \mbox{$\mathbb F(\mathbf i)/\mathbb Q(\mathbf i)$}.
\end{lem}
Denote \mbox{$\varphi(\cdot)$} as Euler's totient function. If \mbox{$m$} is an odd prime power then \mbox{$\mathbb Q(\zeta_m)/\mathbb Q$} is a cyclic extension with degree \mbox{$\varphi(m)$}. In that case, there is a unique intermediate field \mbox{$\mathbb F$} of \mbox{$\mathbb Q(\zeta_{m})/\mathbb Q$} such that \mbox{$\mathbb F/\mathbb Q$} is a cyclic extension with degree \mbox{$n$} whenever \mbox{$n\mid\varphi(m)$}. Motivated by Lemma~\ref{spl} and \mbox{$\varphi(q^{k+1})=q^k(q-1)$}, we naturally consider \mbox{$m=q^{k+1}$} and \mbox{$n=q^k$} for each odd prime number \mbox{$q$} and positive integer \mbox{$k$}, shown as follows.
\begin{prop}
\label{power}
Let \mbox{$q$} be an odd prime number, \mbox{$k$} be a positive integer, and \mbox{$\mathbb F/\mathbb Q$} be the unique \mbox{degree-$q^k$} sub-extension of \mbox{$\mathbb Q(\zeta_{q^{k+1}})/\mathbb Q$}. If \mbox{$\ell$} is a prime number such that \mbox{$\ell\neq q$} and \mbox{$q^2\nmid \ell^{q-1}-1$} then \mbox{$\ell$} is inert in \mbox{$\mathbb F/\mathbb Q$}.
\end{prop}
\begin{proof}
For a nonzero prime ideal \mbox{$\wp$} of \mbox{$\mathcal O_{\mathbb Q(\zeta_{q^{k+1}})}$} with \mbox{$\wp\mid\ell$}, the residue class degree \mbox{$f$} of \mbox{$\wp$} over \mbox{$\ell$} is the smallest positive integer \mbox{$r$} such that \mbox{$\ell^r\equiv 1\pmod{q^{k+1}}$}, denoted by \mbox{$\mathrm{ord}_{q^{k+1}}(\ell)$}~\cite[Ch. 13, Sec. 2, Thm. 2]{IR90}. Fermat's little theorem and \mbox{$\ell\ne q$} imply \mbox{$q\mid\ell^{q-1}-1$}; thus \mbox{$q^2\nmid \ell^{q-1}-1$} results in \mbox{$\ell^{q-1}=1+sq$} with \mbox{$q\nmid s$}. From \mbox{$\ell^{q(q-1)}=(1+sq)^q=1+sq^2+tq^3$}, \mbox{$q^2\mid\ell^{q(q-1)}-1$} and \mbox{$q^3\nmid \ell^{q(q-1)}-1$}. Continuing this argument leads to $$q^{k+1}\nmid \ell^{q^{k-1}(q-1)}-1,$$ i.e., \mbox{$\mathrm{ord}_{q^{k+1}}(\ell)\nmid q^{k-1}(q-1)$}. Moreover, Euler's theorem states $$\mathrm{ord}_{q^{k+1}}(\ell)\mid\varphi(q^{k+1})=q^k(q-1).$$ Therefore \mbox{$q^k\mid\mathrm{ord}_{q^{k+1}}(\ell)=f$}. Let \mbox{$\mathfrak P=\wp\cap\mathcal O_\mathbb F$}, and \mbox{$f'$} and \mbox{$f''$} be the residue class degrees of \mbox{$\wp$} over \mbox{$\mathfrak P$} and \mbox{$\mathfrak P$} over \mbox{$\ell$}, respectively. By~(\ref{tower}), \mbox{$q^k\mid f'f''$}, and by~(\ref{sum}), $$f'\mid [\mathbb Q(\zeta_{q^{k+1}}):\mathbb F]=q^k(q-1)/q^k=q-1.$$ Hence, \mbox{$q^k\mid f''$}. Again, by~(\ref{sum}) and \mbox{$[\mathbb F:\mathbb Q]=q^k$}, \mbox{$f''=[\mathbb F:\mathbb Q]$}; \mbox{$\ell$} is inert in \mbox{$\mathbb F/\mathbb Q$}.
\end{proof}
Together Lemma~\ref{spl} with Proposition~\ref{power}, for each odd prime number \mbox{$q$} with \mbox{$q^2\nmid 2^{q-1}-1$} and positive integer \mbox{$k$}, a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$q^k$} and \mbox{$1+\mathbf i$} a non-norm element can always be constructed. Nonetheless, there do exist odd prime numbers \mbox{$q$} with \mbox{$q^2\mid 2^{q-1}-1$}, called the Wieferich primes. Although the only known Wieferich primes are \mbox{$q=1093$} and \mbox{$3511$}, Proposition~\ref{residue} below is capable of handling more general scenarios.
\begin{prop}
\label{residue}
Let \mbox{$q$} be a prime number, \mbox{$k$} be a positive integer, \mbox{$p$} be an odd prime number with \mbox{$q^k\mid p-1$}, and \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} be the unique \mbox{degree-$q^k$} sub-extension of \mbox{$\mathbb Q(\zeta_{p},\mathbf i)/\mathbb Q(\mathbf i)$}. If \mbox{$x^q\equiv 2\pmod{p}$} is not solvable in \mbox{$\mathbb Z$}, then \mbox{$1+\mathbf i$} is a non-norm element of \mbox{$\mathbb F/\mathbb Q(\mathbf i)$}.
\end{prop}
\begin{proof}
Let \mbox{$\mathbb K/\mathbb Q$} be the unique \mbox{degree-$q^k$} sub-extension of \mbox{$\mathbb Q(\zeta_{p})/\mathbb Q$}. From Example~\ref{ttr}, \mbox{$p$} is totally and tamely ramified in \mbox{$\mathbb K/\mathbb Q$}. Theorem~\ref{locnon} states that \mbox{$2$} is a non-norm element of \mbox{$\mathbb K/\mathbb Q$}. Suppose, on the contrary, that \mbox{$1+\mathbf i$} is not a non-norm element of \mbox{$\mathbb F/\mathbb Q(\mathbf i)$}, i.e., \mbox{$N_{\mathbb F/\mathbb Q(\mathbf i)}(\alpha)=(1+\mathbf i)^r$} for some \mbox{$\alpha\in\mathbb F^*$} and \mbox{$1\le r<q^k$}. By the transitivity of norm, $$2^r=N_{\mathbb Q(\mathbf i)/\mathbb Q}((1+\mathbf i)^r)=N_{\mathbb Q(\mathbf i)/\mathbb Q}(N_{\mathbb F/\mathbb Q(\mathbf i)}(\alpha))=N_{\mathbb F/\mathbb Q}(\alpha)=N_{\mathbb K/\mathbb Q}(N_{\mathbb F/\mathbb K}(\alpha))\in N_{\mathbb K/\mathbb Q}(\mathbb K^*),$$ a contradiction to \mbox{$2$} being a non-norm element of \mbox{$\mathbb K/\mathbb Q$}.
\end{proof}
By Chebotarev's density theorem or~\cite[Thm. 4]{AR51}, there exist infinitely many odd prime numbers \mbox{$p$} such that \mbox{$q^k\mid p-1$} and \mbox{$x^q\equiv 2\pmod{p}$} is not solvable in \mbox{$\mathbb Z$} for each odd prime number \mbox{$q$} and positive integer \mbox{$k$}; thus a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$q^k$} and \mbox{$1+\mathbf i$} a non-norm element can always be constructed. Note that this statement is not true for \mbox{$q=2$} and \mbox{$k\ge 3$}; otherwise \mbox{$q^k\mid p-1$} implies \mbox{$8\mid p-1$} thereby that \mbox{$x^2\equiv 2\pmod{p}$} is always solvable in \mbox{$\mathbb Z$}. This is also the reason why there is no cyclic extension over \mbox{$\mathbb Q$} with degree \mbox{$n$} such that \mbox{$2$} is inert whenever \mbox{$8\mid n$}. In other words, it is infeasible to find a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with each of such degrees \mbox{$n$} and \mbox{$1+\mathbf i$} a non-norm element by simply beginning at Lemma~\ref{spl}.
Let \mbox{$n=2^k$} with \mbox{$k=1$} or \mbox{$2$}. Both \mbox{$x^2\equiv 2\pmod{3}$} and \mbox{$x^2\equiv 2\pmod{5}$} are not solvable in \mbox{$\mathbb Z$}. By Proposition~\ref{residue}, \mbox{$\mathbb Q(\zeta_3,\mathbf i)/\mathbb Q(\mathbf i)$} and \mbox{$\mathbb Q(\zeta_5,\mathbf i)/\mathbb Q(\mathbf i)$} are cyclic extensions with degrees \mbox{$2$} and \mbox{$4$}, respectively, and \mbox{$1+\mathbf i$} a non-norm element. Integrated by Lemma~\ref{comp}, whenever \mbox{$8\nmid n$}, a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$n$} and \mbox{$1+\mathbf i$} a non-norm element can always be constructed.
\begin{example}[\mbox{$n=6$}]
\mbox{$\mathbb Q(\zeta_{9},\mathbf i)/\mathbb Q(\mathbf i)$} is a cyclic extension with degree \mbox{$6$}. Since \mbox{$3^2\nmid 2^2-1$}, Lemma~\ref{spl} and Proposition~\ref{power} state that \mbox{$1+\mathbf i$} is a non-norm element for the unique \mbox{degree-$3$} sub-extension \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} of \mbox{$\mathbb Q(\zeta_{9},\mathbf i)/\mathbb Q(\mathbf i)$}. Moreover, \mbox{$1+\mathbf i$} is a non-norm element of \mbox{$\mathbb Q(\zeta_3,\mathbf i)/\mathbb Q(\mathbf i)$}. By Lemma~\ref{comp}, \mbox{$1+\mathbf i$} is a non-norm element of the \mbox{degree-$6$} extension \mbox{$\mathbb F(\zeta_{3})/\mathbb Q(\mathbf i)$}. In fact, \mbox{$\mathbb F(\zeta_{3})=\mathbb Q(\zeta_9,\mathbf i)$} because both \mbox{$\mathbb F$} and \mbox{$\mathbb Q(\zeta_3,\mathbf i)$} are subfields of \mbox{$\mathbb Q(\zeta_9,\mathbf i)$}.
\end{example}
For \mbox{$n=2^k$} with \mbox{$k\ge 3$}, let \mbox{$p$} be an odd prime number with \mbox{$2^k\mid p-1$}, \mbox{$\mathfrak p$} be a nonzero prime ideal of \mbox{$\mathbb Z[\mathbf i]$} with \mbox{$\mathfrak p\mid p$}, and \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} be the unique \mbox{degree-$2^k$} sub-extension of \mbox{$\mathbb Q(\zeta_p,\mathbf i)/\mathbb Q(\mathbf i)$}. Example~\ref{ttr} states that \mbox{$\mathfrak p$} is totally and tamely ramified in \mbox{$\mathbb F/\mathbb Q(\mathbf i)$}. From \mbox{$1+\mathbf i\in\mathbb Z[\mathbf i]\setminus\mathfrak p$} and Theorem~\ref{locnon}, \mbox{$1+\mathbf i$} is a non-norm element of \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} as long as \mbox{$x^2\equiv 1+\mathbf i\pmod{\mathfrak p}$} is not solvable in \mbox{$\mathbb Z[\mathbf i]$}.
By definitions, if \mbox{$1+\mathbf i$} is a non-norm element of some cyclic extension \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} with degree \mbox{$2^r$}, then it is also a non-norm element of the unique \mbox{degree-$2^k$} sub-extension of \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} for every \mbox{$k\le r$}. The odd prime number \mbox{$1+2647\cdot 2^{1000}$} permits \mbox{$1+\mathbf i$} to be a non-norm element of \mbox{$\mathbb Q(\zeta_{1+2647\cdot 2^{1000}},\mathbf i)/\mathbb Q(\mathbf i)$}. Indeed, \mbox{$1+\mathbf i$} is always a generator of \mbox{$(\mathbb Z[\mathbf i]/\mathfrak p)^*$} whenever \mbox{$\mathfrak p$} is a nonzero prime ideal of \mbox{$\mathbb Z[\mathbf i]$} with \mbox{$\mathfrak p\mid 1+2647\cdot 2^{1000}$}. Consequently, \mbox{$1+\mathbf i$} is a non-norm element of the unique \mbox{degree-$2^k$} sub-extension of \mbox{$\mathbb Q(\zeta_{1+2647\cdot 2^{1000}},\mathbf i)/\mathbb Q(\mathbf i)$} for each \mbox{$k\le 1000$}. Generally speaking, Chebotarev's density theorem shows that for every \mbox{$k\ge 3$}, there exist infinitely many odd prime numbers \mbox{$p$} such that
\begin{enumerate}
\item \mbox{$p\equiv 1\pmod{2^k}$};
\item \mbox{$x^2\equiv 1+\mathbf i\pmod{\mathfrak p}$} is not solvable in
\mbox{$\mathbb Z[\mathbf i]$} for a nonzero prime ideal \mbox{$\mathfrak p$} of \mbox{$\mathbb Z[\mathbf i]$} with \mbox{$\mathfrak p\mid p$}.
\end{enumerate}
Hence, a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$2^k$} and \mbox{$1+\mathbf i$} a non-norm element can always be constructed for each positive integer \mbox{$k$}.
\begin{thm}
\label{gen}
There exists a cyclic extension over \mbox{$\mathbb Q(\mathbf i)$} with degree \mbox{$n$} and \mbox{$1+\mathbf i$} a non-norm element for each positive integer \mbox{$n$}.
\end{thm}
Recall~\cite[Lem. 3.1]{LC09} that for an odd prime number \mbox{$p$} with \mbox{$n\mid p-1$}, the unique \mbox{degree-$n$} sub-extension \mbox{$\mathbb F/\mathbb Q(\mathbf i)$} of \mbox{$\mathbb Q(\zeta_p,\mathbf i)/\mathbb Q(\mathbf i)$} can be obtained through
\begin{equation}
\label{sub}
\mathbb F= \mathbb Q(\eta,\mathbf i)\mbox{ with }\eta=\sum_{i=0}^{\frac{p-1}{n}-1}\zeta_p^{c^{ni}},
\end{equation}
where \mbox{$c$} is a primitive root modulo \mbox{$p$}. Example~\ref{real} demonstrates the generalized constructional procedure by instancing \mbox{$n=8$} and \mbox{$16$}.
\begin{example}[\mbox{$n=8$} and \mbox{$16$}]
\label{real}
\mbox{$\mathbb Q(\zeta_{17},\mathbf i)/\mathbb Q(\mathbf i)$} is a cyclic extension with degree \mbox{$16$}. According to~(\ref{sub}), \mbox{$\eta=\zeta_{17}+\zeta_{17}^{-1}$} constructs \mbox{$\mathbb Q(\eta,\mathbf i)/\mathbb Q(\mathbf i)$} to be the unique \mbox{degree-$8$} sub-extension of \mbox{$\mathbb Q(\zeta_{17},\mathbf i)/\mathbb Q(\mathbf i)$}. From \mbox{$17\mathbb Z[\mathbf i]=(1+4\mathbf i)\mathbb Z[\mathbf i](1-4\mathbf i)\mathbb Z[\mathbf i]$}, the residue class degree of \mbox{$1+4\mathbf i$} over \mbox{$17$} is \mbox{$1$}; in other words, \mbox{$\mathbb Z[\mathbf i]/(1+4\mathbf i)\simeq\mathbb Z/17\mathbb Z$}. From \mbox{$\mathbf i\equiv 4\pmod{1+4\mathbf i}$}, to solve \mbox{$x^2\equiv 1+\mathbf i\pmod{1+4\mathbf i}$} in \mbox{$\mathbb Z[\mathbf i]$} is equivalent to solve \mbox{$x^2\equiv 1+4\pmod{17}$} in \mbox{$\mathbb Z$}. Whereas \mbox{$x^2\equiv 5\pmod{17}$} is not solvable in \mbox{$\mathbb Z$}, \mbox{$1+\mathbf i$} is a non-norm element of both the \mbox{degree-$8$} extension \mbox{$\mathbb Q(\eta,\mathbf i)/\mathbb Q(\mathbf i)$} and \mbox{degree-$16$} extension \mbox{$\mathbb Q(\zeta_{17},\mathbf i)/\mathbb Q(\mathbf i)$}.
\end{example}
Table~\ref{table1} illustrates cyclic extensions over \mbox{$\mathbb Q(\mathbf i)$} with degrees ranging from \mbox{$2$} to \mbox{$100$} and \mbox{$1+\mathbf i$} a non-norm element, by jointing those designated primitive roots of unity to \mbox{$\mathbb Q(\mathbf i)$}. Note that none of \mbox{$\{\pm 1,\pm\mathbf i\}$} can be a non-norm element when the extension degree over \mbox{$\mathbb Q(\mathbf i)$}, i.e., the number of transmit antennas, \mbox{$n\ge 5$}; thus Theorem~\ref{gen} also proves that \mbox{$1+\mathbf i$} is an integer non-norm element with the smallest absolute value over QAM for every \mbox{$n\geq 5$}.
\begin{rem}
An analogous series of arguments can show how to construct a cyclic extension over \mbox{$\mathbb Q(\zeta_3)$} with degree \mbox{$n$} and \mbox{$\sqrt{-3}$} a non-norm element for every positive integer $n$. Table~\ref{table2} illustrates cyclic extensions over \mbox{$\mathbb Q(\zeta_3)$} with degrees ranging from \mbox{$2$} to \mbox{$100$} and \mbox{$\sqrt{-3}$} a non-norm element, by jointing those designated primitive roots of unity to \mbox{$\mathbb Q(\zeta_3)$}. Note that none of \mbox{$\{\pm 1,\pm\zeta_3,\pm\zeta_3^2\}$} can be a non-norm element when \mbox{$n\ge 7$}. Similarly, \mbox{$\sqrt{-3}$} is an integer non-norm element with the smallest absolute value over hexagonal modulations (HEX) for every \mbox{$n\geq 7$}.
\end{rem}
\begin{table*}
\centering
\begin{tabular}{|c|c||c|c||c|c||c|c|}
\hline
Extension & Roots & Extension & Roots & Extension & Roots & Extension & Roots \\
Degrees & of Unity & Degrees & of Unity & Degrees & of Unity & Degrees & of Unity \\
\hline
& & \mbox{$26$} & \mbox{$\zeta_{53}$} & \mbox{$51$} & \mbox{$\zeta_{103}$} & \mbox{$76$} & \mbox{$\zeta_{229}$} \\
\mbox{$2$} & \mbox{$\zeta_3$} & \mbox{$27$} & \mbox{$\zeta_{81}$} & \mbox{$52$} & \mbox{$\zeta_{53}$} & \mbox{$77$} & \mbox{$\zeta_{463}$} \\
\mbox{$3$} & \mbox{$\zeta_7$} & \mbox{$28$} & \mbox{$\zeta_{29}$} & \mbox{$53$} & \mbox{$\zeta_{107}$} & \mbox{$78$} & \mbox{$\zeta_{169}$} \\
\mbox{$4$} & \mbox{$\zeta_5$} & \mbox{$29$} & \mbox{$\zeta_{59}$} & \mbox{$54$} & \mbox{$\zeta_{81}$} & \mbox{$79$} & \mbox{$\zeta_{317}$} \\
\mbox{$5$} & \mbox{$\zeta_{11}$} & \mbox{$30$} & \mbox{$\zeta_{61}$} & \mbox{$55$} & \mbox{$\zeta_{121}$} & \mbox{$80$} & \mbox{$\zeta_{187}$} \\
\mbox{$6$} & \mbox{$\zeta_{9}$} & \mbox{$31$} & \mbox{$\zeta_{311}$} & \mbox{$56$} & \mbox{$\zeta_{493}$} & \mbox{$81$} & \mbox{$\zeta_{163}$} \\
\mbox{$7$} & \mbox{$\zeta_{29}$} & \mbox{$32$} & \mbox{$\zeta_{97}$} & \mbox{$57$} & \mbox{$\zeta_{361}$} & \mbox{$82$} & \mbox{$\zeta_{83}$} \\
\mbox{$8$} & \mbox{$\zeta_{17}$} & \mbox{$33$} &$\zeta_{67}$ & \mbox{$58$} & \mbox{$\zeta_{59}$} & \mbox{$83$} & \mbox{$\zeta_{167}$} \\
\mbox{$9$} & \mbox{$\zeta_{19}$} & \mbox{$34$} & \mbox{$\zeta_{307}$} & \mbox{$59$} & \mbox{$\zeta_{709}$} & \mbox{$84$} & \mbox{$\zeta_{203}$} \\
\mbox{$10$} & \mbox{$\zeta_{11}$} & \mbox{$35$} & \mbox{$\zeta_{71}$} & \mbox{$60$} & \mbox{$\zeta_{61}$} & \mbox{$85$} & \mbox{$\zeta_{1021}$} \\
\mbox{$11$} & \mbox{$\zeta_{23}$} & \mbox{$36$} & \mbox{$\zeta_{37}$} & \mbox{$61$} & \mbox{$\zeta_{367}$} & \mbox{$86$} & \mbox{$\zeta_{173}$} \\
\mbox{$12$} & \mbox{$\zeta_{13}$} & \mbox{$37$} & \mbox{$\zeta_{149}$} & \mbox{$62$} & \mbox{$\zeta_{373}$} & \mbox{$87$} & \mbox{$\zeta_{349}$} \\
\mbox{$13$} & \mbox{$\zeta_{53}$} & \mbox{$38$} & \mbox{$\zeta_{361}$} & \mbox{$63$} & \mbox{$\zeta_{379}$} & \mbox{$88$} & \mbox{$\zeta_{391}$} \\
\mbox{$14$} & \mbox{$\zeta_{29}$} & \mbox{$39$} & \mbox{$\zeta_{79}$} & \mbox{$64$} & \mbox{$\zeta_{193}$} & \mbox{$89$} & \mbox{$\zeta_{179}$} \\
\mbox{$15$} & \mbox{$\zeta_{61}$} & \mbox{$40$} & \mbox{$\zeta_{187}$} & \mbox{$65$} & \mbox{$\zeta_{131}$} & \mbox{$90$} & \mbox{$\zeta_{181}$} \\
\mbox{$16$} & \mbox{$\zeta_{17}$} & \mbox{$41$} & \mbox{$\zeta_{83}$} & \mbox{$66$} & \mbox{$\zeta_{67}$} & \mbox{$91$} & \mbox{$\zeta_{547}$} \\
\mbox{$17$} & \mbox{$\zeta_{103}$} & \mbox{$42$} & \mbox{$\zeta_{147}$} & \mbox{$67$} & \mbox{$\zeta_{269}$} & \mbox{$92$} & \mbox{$\zeta_{235}$} \\
\mbox{$18$} & \mbox{$\zeta_{19}$} & \mbox{$43$} & \mbox{$\zeta_{173}$} & \mbox{$68$} & \mbox{$\zeta_{515}$} & \mbox{$93$} & \mbox{$\zeta_{373}$} \\
\mbox{$19$} & \mbox{$\zeta_{191}$} & \mbox{$44$} & \mbox{$\zeta_{115}$} & \mbox{$69$} & \mbox{$\zeta_{139}$} & \mbox{$94$} & \mbox{$\zeta_{283}$} \\
\mbox{$20$} & \mbox{$\zeta_{25}$} & \mbox{$45$} & \mbox{$\zeta_{181}$} & \mbox{$70$} & \mbox{$\zeta_{211}$} & \mbox{$95$} & \mbox{$\zeta_{191}$} \\
\mbox{$21$} & \mbox{$\zeta_{49}$} & \mbox{$46$} & \mbox{$\zeta_{139}$} & \mbox{$71$} & \mbox{$\zeta_{569}$} & \mbox{$96$} & \mbox{$\zeta_{679}$} \\
\mbox{$22$} & \mbox{$\zeta_{67}$} & \mbox{$47$} & \mbox{$\zeta_{283}$} & \mbox{$72$} & \mbox{$\zeta_{323}$} & \mbox{$97$} & \mbox{$\zeta_{389}$} \\
\mbox{$23$} & \mbox{$\zeta_{47}$} & \mbox{$48$} & \mbox{$\zeta_{119}$} & \mbox{$73$} & \mbox{$\zeta_{293}$} & \mbox{$98$} & \mbox{$\zeta_{197}$} \\
\mbox{$24$} & \mbox{$\zeta_{119}$} & \mbox{$49$} & \mbox{$\zeta_{197}$} & \mbox{$74$} & \mbox{$\zeta_{149}$} & \mbox{$99$} & \mbox{$\zeta_{199}$} \\
\mbox{$25$} & \mbox{$\zeta_{101}$} & \mbox{$50$} & \mbox{$\zeta_{101}$} & \mbox{$75$} & \mbox{$\zeta_{707}$} & \mbox{$100$} & \mbox{$\zeta_{101}$} \\
\hline
\end{tabular}
\caption{Primitive roots of unity jointed to \mbox{$\mathbb Q(\mathbf i)$}.}\label{table1}
\end{table*}
\begin{table*}
\centering
\begin{tabular}{|c|c||c|c||c|c||c|c|}
\hline
Extension & Roots & Extension & Roots & Extension & Roots & Extension & Roots \\
Degrees & of Unity & Degrees & of Unity & Degrees & of Unity & Degrees & of Unity \\
\hline
& & \mbox{$26$} & \mbox{$\zeta_{53}$} & \mbox{$51$} & \mbox{$\zeta_{409}$} & \mbox{$76$} & \mbox{$\zeta_{761}$} \\
\mbox{$2$} & \mbox{$\zeta_5$} & \mbox{$27$} & \mbox{$\zeta_{109}$} & \mbox{$52$} & \mbox{$\zeta_{53}$} & \mbox{$77$} & \mbox{$\zeta_{463}$} \\
\mbox{$3$} & \mbox{$\zeta_7$} & \mbox{$28$} & \mbox{$\zeta_{29}$} & \mbox{$53$} & \mbox{$\zeta_{107}$} & \mbox{$78$} & \mbox{$\zeta_{79}$} \\
\mbox{$4$} & \mbox{$\zeta_5$} & \mbox{$29$} & \mbox{$\zeta_{59}$} & \mbox{$54$} & \mbox{$\zeta_{163}$} & \mbox{$79$} & \mbox{$\zeta_{317}$} \\
\mbox{$5$} & \mbox{$\zeta_{11}$} & \mbox{$30$} & \mbox{$\zeta_{31}$} & \mbox{$55$} & \mbox{$\zeta_{253}$} & \mbox{$80$} & \mbox{$\zeta_{187}$} \\
\mbox{$6$} & \mbox{$\zeta_{7}$} & \mbox{$31$} & \mbox{$\zeta_{311}$} & \mbox{$56$} & \mbox{$\zeta_{113}$} & \mbox{$81$} & \mbox{$\zeta_{163}$} \\
\mbox{$7$} & \mbox{$\zeta_{29}$} & \mbox{$32$} & \mbox{$\zeta_{128}$} & \mbox{$57$} & \mbox{$\zeta_{361}$} & \mbox{$82$} & \mbox{$\zeta_{415}$} \\
\mbox{$8$} & \mbox{$\zeta_{17}$} & \mbox{$33$} &$\zeta_{161}$ & \mbox{$58$} & \mbox{$\zeta_{233}$} & \mbox{$83$} & \mbox{$\zeta_{167}$} \\
\mbox{$9$} & \mbox{$\zeta_{19}$} & \mbox{$34$} & \mbox{$\zeta_{103}$} & \mbox{$59$} & \mbox{$\zeta_{709}$} & \mbox{$84$} & \mbox{$\zeta_{203}$} \\
\mbox{$10$} & \mbox{$\zeta_{25}$} & \mbox{$35$} & \mbox{$\zeta_{71}$} & \mbox{$60$} & \mbox{$\zeta_{155}$} & \mbox{$85$} & \mbox{$\zeta_{1133}$} \\
\mbox{$11$} & \mbox{$\zeta_{23}$} & \mbox{$36$} & \mbox{$\zeta_{95}$} & \mbox{$61$} & \mbox{$\zeta_{367}$} & \mbox{$86$} & \mbox{$\zeta_{173}$} \\
\mbox{$12$} & \mbox{$\zeta_{35}$} & \mbox{$37$} & \mbox{$\zeta_{149}$} & \mbox{$62$} & \mbox{$\zeta_{961}$} & \mbox{$87$} & \mbox{$\zeta_{349}$} \\
\mbox{$13$} & \mbox{$\zeta_{53}$} & \mbox{$38$} & \mbox{$\zeta_{361}$} & \mbox{$63$} & \mbox{$\zeta_{127}$} & \mbox{$88$} & \mbox{$\zeta_{89}$} \\
\mbox{$14$} & \mbox{$\zeta_{29}$} & \mbox{$39$} & \mbox{$\zeta_{79}$} & \mbox{$64$} & \mbox{$\zeta_{256}$} & \mbox{$89$} & \mbox{$\zeta_{179}$} \\
\mbox{$15$} & \mbox{$\zeta_{31}$} & \mbox{$40$} & \mbox{$\zeta_{187}$} & \mbox{$65$} & \mbox{$\zeta_{131}$} & \mbox{$90$} & \mbox{$\zeta_{209}$} \\
\mbox{$16$} & \mbox{$\zeta_{17}$} & \mbox{$41$} & \mbox{$\zeta_{83}$} & \mbox{$66$} & \mbox{$\zeta_{161}$} & \mbox{$91$} & \mbox{$\zeta_{911}$} \\
\mbox{$17$} & \mbox{$\zeta_{103}$} & \mbox{$42$} & \mbox{$\zeta_{43}$} & \mbox{$67$} & \mbox{$\zeta_{269}$} & \mbox{$92$} & \mbox{$\zeta_{235}$} \\
\mbox{$18$} & \mbox{$\zeta_{19}$} & \mbox{$43$} & \mbox{$\zeta_{173}$} & \mbox{$68$} & \mbox{$\zeta_{137}$} & \mbox{$93$} & \mbox{$\zeta_{373}$} \\
\mbox{$19$} & \mbox{$\zeta_{191}$} & \mbox{$44$} & \mbox{$\zeta_{89}$} & \mbox{$69$} & \mbox{$\zeta_{139}$} & \mbox{$94$} & \mbox{$\zeta_{283}$} \\
\mbox{$20$} & \mbox{$\zeta_{25}$} & \mbox{$45$} & \mbox{$\zeta_{181}$} & \mbox{$70$} & \mbox{$\zeta_{211}$} & \mbox{$95$} & \mbox{$\zeta_{191}$} \\
\mbox{$21$} & \mbox{$\zeta_{43}$} & \mbox{$46$} & \mbox{$\zeta_{139}$} & \mbox{$71$} & \mbox{$\zeta_{569}$} & \mbox{$96$} & \mbox{$\zeta_{896}$} \\
\mbox{$22$} & \mbox{$\zeta_{67}$} & \mbox{$47$} & \mbox{$\zeta_{283}$} & \mbox{$72$} & \mbox{$\zeta_{323}$} & \mbox{$97$} & \mbox{$\zeta_{389}$} \\
\mbox{$23$} & \mbox{$\zeta_{47}$} & \mbox{$48$} & \mbox{$\zeta_{119}$} & \mbox{$73$} & \mbox{$\zeta_{293}$} & \mbox{$98$} & \mbox{$\zeta_{197}$} \\
\mbox{$24$} & \mbox{$\zeta_{119}$} & \mbox{$49$} & \mbox{$\zeta_{197}$} & \mbox{$74$} & \mbox{$\zeta_{149}$} & \mbox{$99$} & \mbox{$\zeta_{199}$} \\
\mbox{$25$} & \mbox{$\zeta_{101}$} & \mbox{$50$} & \mbox{$\zeta_{101}$} & \mbox{$75$} & \mbox{$\zeta_{601}$} & \mbox{$100$} & \mbox{$\zeta_{101}$} \\
\hline
\end{tabular}
\caption{Primitive roots of unity jointed to \mbox{$\mathbb Q(\zeta_3)$}.}\label{table2}
\end{table*}
\begin{table*}
\centering
\begin{tabular}{|c||c|c|c|}\hline
\multicolumn{4}{|c|}{\mbox{$n=8$}}\\
\hline Coding Schemes & New Code & \cite{LC09} & \cite{GX09}\\
\hline &&& \\ \mbox{$\xi(\mathbf{S})$} & \mbox{$\frac{1}{278130^8}$} & \mbox{$\frac{1}{414430^8}$} & \mbox{$\frac{1}{888380^8}$}\\ &&&\\
\hline \mbox{$\gamma$} & \mbox{$1+\mathbf i$} & \multicolumn{2}{c|}{\mbox{$2+\mathbf i$}}\\
\hline \mbox{$\eta$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}+\zeta_{17}^{16}$}} & \mbox{$\zeta_{32}+\zeta_{32}^{15}$}\\
\mbox{$\sigma(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^3+\zeta_{17}^{14}$}} & \mbox{$\zeta_{32}^5+\zeta_{32}^{11}$}\\
\mbox{$\sigma^2(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^8+\zeta_{17}^{9}$}} & \mbox{$\zeta_{32}^{25}+\zeta_{32}^{23}$}\\
\mbox{$\sigma^3(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^7+\zeta_{17}^{10}$}} & \mbox{$\zeta_{32}^{29}+\zeta_{32}^{19}$}\\
\mbox{$\sigma^4(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^4+\zeta_{17}^{13}$}} & \mbox{$\zeta_{32}^{17}+\zeta_{32}^{31}$}\\
\mbox{$\sigma^5(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^5+\zeta_{17}^{12}$}} & \mbox{$\zeta_{32}^{21}+\zeta_{32}^{27}$}\\
\mbox{$\sigma^6(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^2+\zeta_{17}^{15}$}} & \mbox{$\zeta_{32}^{9}+\zeta_{32}^{7}$}\\
\mbox{$\sigma^7(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^6+\zeta_{17}^{11}$}} & \mbox{$\zeta_{32}^{13}+\zeta_{32}^{3}$}\\
\hline \hline \multicolumn{4}{|c|}{\mbox{$n=16$}}\\
\hline Coding Schemes & New Code & \cite{LC09} & \cite{GX09}\\
\hline &&& \\ \mbox{$\xi(\mathbf{S})$} & \mbox{$\frac{1}{6016^{16}}$} & \mbox{$\frac{1}{11776^{16}}$} & \mbox{$\frac{1}{\left(1.7051\cdot 10^{11}\right)^{16}}$}\\ &&&\\
\hline \mbox{$\gamma$} & \mbox{$1+\mathbf i$} & \multicolumn{2}{c|}{\mbox{$2+\mathbf i$}}\\
\hline \mbox{$\eta$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}$}} & \mbox{$\zeta_{64}+\zeta_{64}^{31}$}\\
\mbox{$\sigma(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^3$}} & \mbox{$\zeta_{64}^5+\zeta_{64}^{27}$}\\
\mbox{$\sigma^2(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^9$}} & \mbox{$\zeta_{64}^{25}+\zeta_{64}^{7}$}\\
\mbox{$\sigma^3(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{10}$}} & \mbox{$\zeta_{64}^{61}+\zeta_{64}^{35}$}\\
\mbox{$\sigma^4(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{13}$}} & \mbox{$\zeta_{64}^{49}+\zeta_{64}^{47}$}\\
\mbox{$\sigma^5(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^5$}} & \mbox{$\zeta_{64}^{53}+\zeta_{64}^{43}$}\\
\mbox{$\sigma^6(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{15}$}} & \mbox{$\zeta_{64}^{9}+\zeta_{64}^{23}$}\\
\mbox{$\sigma^7(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{11}$}} & \mbox{$\zeta_{64}^{45}+\zeta_{64}^{51}$}\\
\mbox{$\sigma^8(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{16}$}} & \mbox{$\zeta_{64}^{33}+\zeta_{64}^{63}$}\\
\mbox{$\sigma^9(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{14}$}} & \mbox{$\zeta_{64}^{37}+\zeta_{64}^{59}$}\\
\mbox{$\sigma^{10}(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{8}$}} & \mbox{$\zeta_{64}^{57}+\zeta_{64}^{39}$}\\
\mbox{$\sigma^{11}(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{7}$}} & \mbox{$\zeta_{64}^{29}+\zeta_{64}^{3}$}\\
\mbox{$\sigma^{12}(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{4}$}} & \mbox{$\zeta_{64}^{17}+\zeta_{64}^{15}$}\\
\mbox{$\sigma^{13}(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{12}$}} & \mbox{$\zeta_{64}^{21}+\zeta_{64}^{11}$}\\
\mbox{$\sigma^{14}(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{2}$}} & \mbox{$\zeta_{64}^{41}+\zeta_{64}^{55}$}\\
\mbox{$\sigma^{15}(\eta)$} & \multicolumn{2}{c|}{\mbox{$\zeta_{17}^{6}$}} & \mbox{$\zeta_{64}^{13}+\zeta_{64}^{19}$}\\
\hline
\end{tabular}
\caption{The normalized diversity products and corresponding coding structures.}\label{table3}
\end{table*}
\section{Comparisons With Existing Codes}
\label{numerical}
Given a cyclic extension \mbox{$\mathbb Q(\eta,\mathbf i)/\mathbb Q(\mathbf i)$} with degree \mbox{$n$}, a generator \mbox{$\sigma$} of \mbox{$\mathrm{Gal}(\mathbb Q(\eta,\mathbf i)/\mathbb Q(\mathbf i))$}, and a non-norm element \mbox{$\gamma$} of \mbox{$\mathbb Q(\eta,\mathbf i)/\mathbb Q(\mathbf i)$}, the generating full-rate space-time block code with nonvanishing determinants can be expressed as~\cite{EKPKL06,GX09,KR05,LC09} $$\mathbf{S}=\left(\begin{array}{ccccc}s_1 & \gamma\sigma(s_n) & \gamma\sigma^2(s_{n-1}) & \cdots & \gamma\sigma^{n-1}(s_2)\\ s_2 & \sigma(s_1) & \gamma\sigma^2(s_n) & \cdots & \gamma\sigma^{n-1}(s_3)\\ s_3 & \sigma(s_2) & \sigma^2(s_1) & \cdots & \gamma\sigma^{n-1}(s_4)\\ \vdots & \vdots & \vdots & \ddots & \vdots\\ s_n & \sigma(s_{n-1}) & \sigma^2(s_{n-2}) & \cdots & \sigma^{n-1}(s_1)\end{array}\right),$$ where $$s_i= \sum_{j=1}^n x_{i,j}\eta^{j-1}$$ for each \mbox{$i=1,\ldots,n$} with \mbox{$(x_{i,j})_{1\le i,j\le n}\in \mathbb Z(\mathbf i)^{n\times n}$} representing the \mbox{$n^2$} input symbols. When \mbox{$\gamma$} is an integer non-norm element, the minimum determinant of \mbox{$\mathbf{S}$}, $$\delta(\mathbf{S})= \min_{(x_{i,j})_{1\le i,j\le n}\in \mathbb Z(\mathbf i)^{n\times n}\setminus \mathbf{0}_{n\times n}}|\det(\mathbf{S})|^2,$$ always equals \mbox{$1$}, where \mbox{$\mathbf{0}_{k\times m}$} denotes the \mbox{$k\times m$} matrix all of whose entries are zero. An asymptotic measure of the performance for \mbox{$\mathbf{S}$} over QAM can be determined via the normalized diversity product~\cite{GX09}: $$\xi(\mathbf{S})= E^{-n}=\left(\sum_{i=0}^{n-1} \left\|\left(\begin{array}{cc}\mathbf{I}_{n-i} & \mathbf{0}_{(n-i)\times i}\\ \mathbf{0}_{i\times(n-i)} & \gamma\mathbf{I}_{i} \end{array}\right)\Xi\right\|^2_F\right)^{-n},$$ where \mbox{$\mathbf{I}_{k}$} is the identity matrix of size \mbox{$k\times k$}, $$\Xi=\left(\begin{array}{ccccc}1 & \eta & \eta^2 & \cdots & \eta^{n-1}\\ 1 & \sigma(\eta) & (\sigma(\eta))^2 & \cdots & (\sigma(\eta))^{n-1}\\ 1 & \sigma^2(\eta) & (\sigma^2(\eta))^2 & \cdots & (\sigma^2(\eta))^{n-1}\\ \vdots & \vdots & \vdots & \ddots & \vdots\\ 1 & \sigma^{n-1}(\eta) & (\sigma^{n-1}(\eta))^2 & \cdots & (\sigma^{n-1}(\eta))^{n-1}\end{array}\right),$$ and \mbox{$\|\cdot\|_F$} stands for the Frobenius norm of input matrix. In other words, \mbox{$E$} represents the total energy needed for encoding \mbox{$\mathbf{S}$}. Table~\ref{table3} lists the normalized diversity products and corresponding coding structures for the new codes constructed in Example~\ref{real} and those in~\cite{GX09} and~\cite{LC09} with \mbox{$n=8$} and \mbox{$16$}; the smaller the absolute value of an integer non-norm element \mbox{$\gamma$}, the larger the normalized diversity product \mbox{$\xi(\mathbf{S})$}.
\section{Conclusions}
\label{conc}
This paper presents a newly constructive method that proves \mbox{$1+\mathbf i$} is an integer non-norm element with the smallest absolute value over QAM for every \mbox{$n\geq 5$}. The proposed designs achieve better normalized diversity products and the optimal diversity-multiplexing gains tradeoff.
|
\section{Introduction}
\label{sec_introduction}
The dynamical stability of self-gravitating systems is an important
problem in astrophysics. This problem was first considered by Jeans
(1902,1929) who studied the linear dynamical stability of an infinite
homogeneous collision-dominated gas described by the Euler
equations. He found that this system becomes unstable if the
wavelength of the perturbation exceeds a critical value
$\lambda_c=(\frac{\pi}{G\rho})^{1/2}c_s$ (where $c_s$ is the velocity
of sound) nowadays called the Jeans length. In that case, the
perturbation grows exponentially rapidly without oscillating. By
contrast, for small wavelengths $\lambda<\lambda_c$, the perturbation
oscillates without attenuation and behaves like a gravity-modified
sound wave. As is well-known, the Jeans approach suffers from a
mathematical inconsistency at the start. Indeed, an infinite
homogeneous gravitating system cannot be in static equilibrium since
there are no pressure gradients to balance the gravitational
force. Jeans removed this inconsistency by assuming that the Poisson
equation describes only the relationship between the perturbed
gravitational potential and the perturbed density. However, this
assumption is {\it ad hoc} and is known as the {\it Jeans swindle}
(Binney \& Tremaine 1987). A Jeans-type analysis can however be
justified in certain situations: (i) In cosmology, when we work in the
comoving frame, the expansion of the universe introduces a sort of
``neutralizing background'' in the Poisson equation (like in the
Jellium model of plasma physics). In that case, an infinite
homogeneous self-gravitating medium can be in static equilibrium
(Peebles 1980). Therefore, the Jeans instability mechanism is relevant to
understand the formation of galaxies from the homogeneous early
universe. (ii) If we consider a uniformly rotating system, the
gravitational attraction can be balanced by the centrifugal
force. Therefore, a homogeneous system can be in static equilibrium in
the rotating frame provided that the angular velocity and the density
satisfy a well-defined relation (Chandrasekhar 1961). (iii) The Jeans
procedure is approximately correct if we only consider perturbations
with small wavelengths, much smaller than the Jeans length (Binney \&
Tremaine 1987).
The dynamical stability of an infinite homogeneous encounterless
stellar system described by the Vlasov equation was considered much
later by Lynden-Bell (1962)\footnote{See also Simon (1961) and Liboff
(1963).} who used methods similar to those introduced by Landau (1946)
in plasma physics. There exists indeed many analogies between
self-gravitating systems and plasmas since the Coulombian and
Newtonian interactions both correspond to an inverse square
law. However, there exists also crucial differences. First of all,
gravity is attractive whereas electricity is repulsive. On the other
hand, as already indicated, the self-gravity of a uniform gravitating
system is usually not neutralized, contrary to a plasma where the
presence of electrons and ions provide electrical neutrality. In
order to circumvent the Jeans swindle, Lynden-Bell considered the case
of a rotating system and noted that the angular velocity plays a role
similar to the magnetic field in plasma physics (Bernstein 1958). He
found that the system becomes unstable above a critical
wavelength. For a Maxwellian distribution, this critical wavelength
$\lambda_c=(\frac{\pi}{G\rho})^{1/2}\sigma$ is equivalent to the Jeans
length if we identify the r.m.s. velocity of the stars $\sigma$ in one
direction with the velocity of sound $c_s$ in an isothermal gas. In
that case, the perturbation grows exponentially rapidly without
oscillating. By contrast, for small wavelengths $\lambda<\lambda_c$,
the perturbation is damped exponentially. This is similar to the
Landau damping in plasma physics that is derived in the complete
absence of collisions.
Sweet (1963) considered the gravitational instability of a system of
gas and stars in relative motion and provided a detailed analysis of
the dispersion relation in different cases. He considered in
particular the situation where the gas is at rest ($U=0$) and the star
system is made of two identical interpenetrating streams with
Maxwellian distribution moving in opposite direction with equal
velocity $\pm v_a$. This corresponds to the Kapteyn-Eddington
two-stream representation in the solar neighborhood. He mentioned that
the two humps could be asymmetric but he did not study the consequence
of this asymmetry in detail. One important conclusion of his work is
that the critical Jeans wavelength is reduced by the relative motion
of the gas and stars and that it vanishes when the relative velocity
$v_a$ exceeds a limit of the order of the effective velocity of sound
$c_s$ in the gas. In particular, the star streaming in the solar
neighborhood can cause the gas to be unstable at all
wavelengths. A similar conclusion was reached by Talwar \& Kalra (1966) who
considered the contrastreaming instability of two self-gravitating
gaseous streams with velocity $\pm U$ described by fluid equations. In
the subsonic regime $U<c_s$, they found that the critical Jeans
wavelength is reduced and tends to zero as the streaming velocity $U$
approaches the velocity of sound $c_s$ (by contrast, in the supersonic
regime $U>c_s$, the critical wavelength is larger than the Jeans
length without streaming). The works of Sweet (1963) and Talwar \&
Kalra (1966) were further developed by Ikeuchi {et al.} (1974) who
presented various stability diagrams for two-stream stellar systems,
gaseous-stellar systems, two gaseous systems and plasmas. For a purely
two-stream stellar systems (without gas), they found that the critical
Jeans length becomes larger due to the stabilization effect of
relative velocity, contrary to the case where a gas component is
present. The two-stream instability was also discussed by Araki (1987)
for infinite homogeneous and uniformly rotating stellar systems.
The seminal works of Jeans (for gaseous systems) and Lynden-Bell (for
stellar systems) have been continued in several directions. For
example, the Jeans instability of a self-gravitating gas in the
presence of a magnetic field or an overall rotation has been studied
in detail by Chandrasekhar (1961). On the other hand, the
gravitational stability of a disk of stars is treated in the classical
paper of Toomre (1964). Other interesting contributions are reviewed
by Binney \& Tremaine (1987). However, this problem is largely
academic because real stellar systems (like stars and galaxies) are
not spatially homogeneous and are limited in space. Now, when spatial
inhomogeneity is properly taken into account, the picture becomes very
different. The dynamical stability of stars has been considered by
various authors such as Eddington (1918), Ledoux (1945) and
Chandrasekhar (1964) and different stability criteria for
the Euler-Poisson system have been obtained. For example, it can be
shown that polytropic stars are Euler stable if their index $n\le 3$ while they
are unstable if $3<n\le 5$ (polytropic stars with $n>5$, including
isothermal stars $n\rightarrow +\infty$, have infinite mass). On the
other hand, the Vlasov stability of stellar systems has been studied by
Antonov (1960a), Doremus et al. (1971), Kandrup \& Sygnet (1985)
and others. They have shown that all stellar systems with a
distribution function $f=f(\epsilon)$ which is a monotonically
decreasing function of the energy are linearly dynamically stable with
respect to the Vlasov-Poisson system. In particular, all polytropic
galaxies with index in the range $3/2\le n\le 5$ (such that the total
mass is finite) are stable\footnote{The difference between the
dynamical stability of gaseous stars and stellar systems, which is
related to the Antonov (1960b) first law, can also be related to a notion of
``ensemble inequivalence'' like in thermodynamics (see Chavanis
2006a).}. Very recently, it was proven rigorously by Lemou et
al. (2009) that these systems are also {\it nonlinearly} dynamically
stable with respect to the Vlasov-Poisson system. These results show
that, when the spatial inhomogeneity of the system is properly
accounted for, the Jeans instability is suppressed\footnote{A form of
Jeans instability is recovered for spatially inhomogeneous isothermal
($n=\infty$) and polytropic ($n>5$) systems confined within a box
(Padmanabhan 1990, Semelin et al. 2001, Chavanis 2002a,2002b, Taruya \& Sakagami 2003a).}.
Recently, there was a renewed interest for this classical problem due
to analogies with other physical systems. For example, the community
of statistical mechanics involved in the dynamics and thermodynamics
of systems with long-range interactions (Dauxois et al. 2002) has
studied a toy model called the Hamiltonian Mean Field (HMF) model
which presents many features similar to self-gravitating
systems\footnote{In fact, this model is directly inspired by
astrophysics. It was introduced by Pichon (1994) as a simplified model
to describe the formation of bars in disk galaxies (see a detailed
historic in Chavanis et al. (2005)).}. In particular, this model
presents an instability below a critical temperature that is similar
to the Jeans instability in astrophysics. Interestingly, a spatially
homogeneous phase exists for this model at any temperature (it is
stable for $T>T_c$ and unstable for $T<T_c$). Therefore, there is no
mathematical inconsistency (i.e. no ``Jeans swindle'') when we perform
the stability analysis of the homogeneous phase of this system
(Inagaki \& Konishi 1993, Choi \& Choi 2003, Yamaguchi et al. 2004,
Chavanis \& Delfini 2009). On the other hand, in biology, several
researchers in physics and applied mathematics have studied the
phenomenon of chemotaxis using the Keller-Segel (1970) model in order
to explain the self-organization of bacterial colonies, cells or even
social insects. Recently, it was shown that the chemotactic
instability in biology bears some deep analogies with the Jeans
instability in astrophysics (Chavanis 2006b)\footnote{Curiously, Keller \& Segel (1970) did not point out this analogy and interpreted instead the chemotactic instability in relation to B\'enard convection.}. Interestingly, in the
biological problem, there is no ``Jeans swindle'' because the Poisson
equation is replaced by a more complex reaction-diffusion equation
that allows for the existence of infinite and homogeneous
distributions.
These analogies were a source of motivation to study anew the
classical Jeans problem in astrophysics (for the Euler-Poisson and
Vlasov-Poisson systems). In fact, we discovered that the study of this
classical problem was still incomplete. For example, the case of
(spatially homogeneous) polytropic distribution functions has not been
investigated in detail and the case of a symmetric double-humped
distribution considered by Ikeuchi et al. (1974)
needs further discussion. On the
other hand, the stability of an asymmetric double-humped distribution
has never been considered in the astrophysical literature (although
the interest of such distributions was mentioned early by Sweet
(1963)) and could form an interesting theoretical problem. The
stability criteria for homogeneous stellar systems can be established
very easily with the powerful Nyquist method (Nyquist 1932) used in
plasma physics. This method can be readily applied to homogeneous
self-gravitating systems. However, since the gravitational interaction
is attractive while the electric interaction is repulsive, a crucial
sign changes in the dispersion relation and the problem must be
reconsidered. In particular, this change of sign is the reason for the
Jeans instability in astrophysics and the Nyquist method gives a nice
graphical illustration of this instability. This method does not seem
to be well-known by astrophysicists and is not exposed in classical
monographs\footnote{As mentioned by the referee, an early application
of the Nyquist method was made by Lynden-Bell (1967) in a not very easily accessible paper. Nyquist's method has also been used by Weinberg (1991) to study the stability of spherical stellar systems. We are not aware of other references.}. Although it essentially has an academic interest since
real stellar systems are not spatially
homogeneous\footnote{In fact, the Nyquist method can be generalized to apply to spatially {\it inhomogeneous} self-gravitating
systems (see Weinberg 1991).}, we think that it deserves a detailed pedagogical
exposition. We thus propose an exhaustive study of the linear
dynamical stability of infinite and homogeneous gaseous (Euler) and
stellar (Vlasov) systems that uses the Nyquist method and completes
previous works on that problem.
The paper is organized as follows. In Sec. \ref{sec_gas}, we consider
a self-gravitating barotropic gas described by the Euler-Poisson
system. We recall the classical Jeans instability criterion and, for
future comparison, consider explicitly the case of isothermal
and polytropic equations of state. In Sec. \ref{sec_stellar}, we
consider a collisionless stellar system described by the
Vlasov-Poisson system. Using the Nyquist method, we derive the Jeans
instability criterion for an infinite and homogeneous distribution. We
show that, for any single humped distribution function of the form
$f=f(v^2)$ with $f'(v^2)<0$, the Jeans instability criterion for a
stellar system is equivalent to the Jeans instability criterion for
the corresponding barotropic gas with the same equation of state. Of
course, the dispersion relation and the evolution of the perturbation
is different in the two systems (gas and stellar system) but the
threshold of the instability is the same. This generalizes the result
obtained by Lynden-Bell (1962) for the Maxwellian distribution. In
Secs. \ref{sec_vb} and
\ref{sec_vb2}, we specifically consider the case of isothermal and
polytropic distribution functions and derive analytical expressions
for the growth rate and damping rate of the perturbation by adapting
methods of plasma physics to the gravitational context. In
Sec. \ref{sec_vhq}, we consider a symmetric double-humped distribution
made of two counter-streaming Maxwellians with velocities $\pm v_a$
and establish the stability diagram. In Sec. \ref{sec_vha}, we
generalize our results to the case of an asymmetric double-humped
distribution. We find that: (i) The system is stable for
$\lambda<(\lambda_c)_*$ and unstable for $\lambda>(\lambda_c)_*$ where
the critical wavelength $(\lambda_c)_*$ is a non-trivial function of
the relative velocity $v_a$ and asymmetry parameter $\Delta$. (ii) The
critical wavelength $(\lambda_c)_*$ is always larger than the critical
wavelength in the absence of streaming ($v_a=0$) so that the
instability is delayed. (iii) The phase velocity of the unstable mode
corresponds to the global maximum of the distribution function. In
Sec. \ref{sec_plasmas}, we consider the case of plasmas. We recall the
well-known fact that single-humped distributions are always
stable. Then, we consider the case of symmetric and asymmetric
double-humped distributions. For an asymmetry parameter
$\Delta<\Delta_*=3.3105$: (i) the system is stable for
$v_a^2<y_c(\Delta)T$ where $y_c(\Delta)$ depends on the asymmetry
($y_c=1.708$ for a symmetric distribution); (ii) for
$v_a^2>y_c(\Delta)T$, the system is stable for $\lambda<(\lambda_c)_*$
and unstable for $\lambda>(\lambda_c)_*$ where $(\lambda_c)_*$ depends
on the relative velocity $v_a$ and on the asymmetry parameter
$\Delta$. For an asymmetry parameter $\Delta>\Delta_*=3.3105$: (i) the
system is stable for $v_a^2<y_*(\Delta)T$ where $y_*(\Delta)$ depends
on the asymmetry; (ii) for $y_*(\Delta)T<v_a^2<y_c(\Delta)T$, the system is stable
for $\lambda<(\lambda_c)_1$, unstable for
$(\lambda_c)_1<\lambda<(\lambda_c)_2$ and stable for
$\lambda>(\lambda_c)_2$ where $(\lambda_c)_1$ and $(\lambda_c)_2$
depend on the relative velocity $v_a$ and on the asymmetry parameter
$\Delta$; (iii) for $v_a^2>y_c(\Delta)T$, the system is stable
for $\lambda<(\lambda_c)_1$ and unstable for
$\lambda<(\lambda_c)_1$. Throughout the paper, we also compare our results obtained
for stellar systems (Vlasov) to those obtained for gaseous systems
(Euler).
\section{Self-gravitating barotropic gas}
\label{sec_gas}
\subsection{The Euler-Poisson system}
\label{sec_ep}
We consider a self-gravitating gas described by the
Euler-Poisson system
\begin{equation}
{\partial\rho\over\partial t}+\nabla \cdot (\rho {\bf u})=0,
\label{ep1}
\end{equation}
\begin{equation}
{\partial {\bf u}\over\partial t}+({\bf u}\cdot \nabla) {\bf u}=-{1\over\rho}\nabla p-\nabla\Phi,
\label{ep2}
\end{equation}
\begin{equation}
\Delta\Phi=4\pi G\rho.
\label{ep3}
\end{equation}
These are the usual equations used to study the dynamical stability of
a gaseous medium like a molecular cloud or a star. We assume that the
gas is a perfect gas in local thermodynamic equilibrium (L.T.E).
Therefore, at each point, the distribution function of the particles
(atoms or molecules) is of the form
\begin{equation}
f({\bf r},{\bf v},t)=\biggl \lbrack {1\over 2\pi T({\bf r},t)}\biggr\rbrack^{3/2}\rho({\bf r},t)e^{-{ \lbrack {\bf v}-{\bf u}({\bf r},t)\rbrack^{2}\over 2 T({\bf r},t)}}.
\label{ep4}
\end{equation}
Note that the kinetic temperature $T({\bf r},t)=\frac{1}{3}\langle
w^2\rangle$ where ${\bf w}={\bf v}-{\bf u}({\bf r},t)$ is the relative
velocity\footnote{Throughout this paper, by an abuse of language, we
shall define the kinetic temperature by $T({\bf
r},t)\equiv \frac{1}{d}\langle ({\bf v}-{\bf u}({\bf r},t))^2\rangle$ where
$d$ is the dimension of space. Therefore, the kinetic temperature is a
measure of the local velocity dispersion of the particles in one
dimension. It is related to the real kinetic temperature by $T({\bf
r},t)=k_B T_{real}({\bf r},t)/m$. We do that in order to extend this
definition to the case of collisionless stellar systems described by
the Vlasov equation where the individual mass of the stars does not
appear.}. Introducing the pressure $p={1\over 3}\int f w^{2}d{\bf v}$, we find that the local equation of state is
\begin{equation}
p({\bf r},t)=\rho({\bf r},t)T({\bf
r},t).
\label{ep5}
\end{equation}
The system of equations (\ref{ep1})-(\ref{ep3}) is not closed since it
must be supplemented by an equation for the energy or, equivalently,
for the temperature $T({\bf r},t)$. Here, we shall restrict ourselves
to the case of a {\it barotropic gas} for which the pressure depends
only on the density
\begin{equation}
p({\bf r},t)=p[\rho({\bf r},t)].
\label{ep5b}
\end{equation}
The local velocity of sound is
\begin{equation}
c_{s}^{2}({\bf r},t)=p'(\rho({\bf r},t)).
\label{ep6}
\end{equation}
A steady state of the Euler-Poisson system satisfies ${\bf u}={\bf 0}$
and the condition of hydrostatic equilibrium
\begin{equation}
\nabla p+\rho\nabla\Phi={\bf 0}.
\label{ep7}
\end{equation}
Combining the condition of hydrostatic equilibrium (\ref{ep7}) and
the equation of state $p=p(\rho)$, we find that the density is a function $\rho=\rho(\Phi)$ of the gravitational potential given by
\begin{eqnarray}
\int^{\rho}{p'(\rho')\over\rho'}d\rho'=-\Phi.
\label{ep8}
\end{eqnarray}
We note that $p'(\rho)=-\rho/\rho'(\Phi)$. Since $p'(\rho)=c_s^2\ge 0$, we
find that $\rho'(\Phi)\le 0$. The density is a monotonically
decreasing function of the gravitational potential. The total energy
of the star is
\begin{eqnarray}
{\cal W}[\rho,{\bf
u}]=\int\rho\int_{0}^{\rho}{p(\rho')\over\rho^{'2}}d\rho'd{\bf r}
+{1\over 2}\int \rho\Phi d{\bf r}+\int \rho
{{\bf u}^{2}\over 2}d{\bf r},\nonumber\\
\label{ep9}
\end{eqnarray}
including the internal energy, the potential energy and the kinetic
energy of the mean motion. The mass $M[\rho]$ and the energy ${\cal
W}[\rho,{\bf u}]$ are conserved by the barotropic Euler-Poisson
system: $\dot M=\dot{\cal W}=0$. Therefore, a minimum of the energy
functional at fixed mass determines a steady state of the barotropic
Euler-Poisson system that is nonlinearly dynamically stable. Here, we
have stability in the sense of Lyapunov. This means that the size of
the perturbation is bounded by the size of the initial perturbation
for all times. We are led therefore to considering the minimization
problem
\begin{equation}
\min_{\rho,{\bf u}} \ \left\lbrace {\cal W}[\rho]\quad |\quad M[\rho]=M\right \rbrace.\label{ep10}
\end{equation}
The critical points of energy at fixed mass are determined by the
Euler-Lagrange equation $\delta{\cal W}-\mu\delta M=0$ where $\mu$ is
a Lagrange multiplier. This yields the condition of hydrostatic
equilibrium (\ref{ep7}). Then, a critical point of energy at fixed
mass is a (local) energy minimum iff
\begin{equation}
\delta^2{\cal W}=\int \frac{p'(\rho)}{2\rho}(\delta\rho)^2\, d{\bf r}+\frac{1}{2}\int \delta\rho\delta\Phi\, d{\bf r}\ge 0
\label{ep10b}
\end{equation}
for all perturbations $\delta\rho$ that conserve mass: $\delta M=0$.
\subsection{The Jeans instability criterion}
\label{sec_jeans}
We consider the linear dynamical stability of an infinite homogeneous
gaseous medium described by the hydrodynamic equations (\ref{ep1}),
(\ref{ep2}), (\ref{ep3}) and (\ref{ep5b}). This is the classical Jeans
(1902,1929) problem. Linearizing Eqs. (\ref{ep1})-(\ref{ep3}) around a
solution $\rho=\Phi={\rm const.}$ and ${\bf u}={\bf 0}$, making the
Jeans swindle and decomposing the perturbations in normal modes of the
form $e^{i({\bf k}\cdot {\bf r}-\omega t)}$, we obtain the classical
dispersion relation (Binney \& Tremaine 1987):
\begin{eqnarray}
\omega^{2}=c_{s}^{2}k^{2}-4\pi G\rho,
\label{jeans1}
\end{eqnarray}
where $c_{s}^{2}=dp/d\rho$ is the velocity of sound. Since $\omega^2$ is real, the complex
pulsation $\omega$ is either real or purely imaginary. The condition $\omega=0$
determines a critical wavenumber
\begin{eqnarray}
k_{c}^{2}=\frac{4\pi G\rho}{c_{s}^{2}},
\label{jeans2}
\end{eqnarray}
called the Jeans wavenumber for a barotropic gas.
The dispersion relation can be rewritten
\begin{eqnarray}
\frac{\omega^{2}}{4\pi G\rho}=\frac{k^{2}}{k_{c}^{2}}-1.
\label{jeans3}
\end{eqnarray}
For $k>k_{c}$, we have $\omega=\omega_r=\pm
\sqrt{\omega^2}$ so that the perturbation oscillates like $e^{-i\omega_r t}$ with a pulsation
\begin{eqnarray}
\omega_r=\pm\sqrt{c_{s}^{2}k^{2}-4\pi G\rho},
\label{jeans4}
\end{eqnarray}
without attenuation. This corresponds to gravity-modified sound waves. In that case the system is stable.
For $k<k_{c}$, we have $\omega=\omega_{i}=\pm
i\sqrt{-\omega^2}$ so that the perturbation grows like $e^{\omega_i t}$ with a growth rate
\begin{eqnarray}
\omega_i=\sqrt{4\pi G\rho-c_{s}^{2}k^{2}},
\label{jeans5}
\end{eqnarray}
without oscillation (the second mode is attenuated exponentially
rapidly so that the growing mode dominates). In that case, the system
is unstable. This is the so-called Jeans instability. The growth rate
is maximum for $k=0$ and its value is $\omega_i(k=0)=\sqrt{4\pi
G\rho}$.
In conclusion, a perturbation with wavenumber $k$ is stable if
\begin{eqnarray}
k>k_{c}=\frac{\sqrt{4\pi G\rho}}{c_{s}},
\label{jeans6}
\end{eqnarray}
and unstable otherwise. These results are valid for an
arbitrary barotropic equation of state $p=p(\rho)$. We now specialize
on particular cases, namely isothermal and polytropic equations of
state.
\subsection{Isothermal equation of state}
\label{sec_ip}
We first consider an isothermal equation of state
\begin{equation}
p({\bf r},t)=\rho({\bf r},t)T,
\label{ip1}
\end{equation}
where the temperature $T({\bf r},t)=T$ is a uniform. At hydrostatic
equilibrium, according to Eq. (\ref{ep8}), the relation between the
density and the gravitational potential is given by the Boltzmann law
\begin{equation}
\rho=A'e^{-{\Phi\over T}}, \label{ip2}
\end{equation}
where $A'$ is a constant. The energy functional (\ref{ep9}) reads
\begin{eqnarray}
{\cal W}[\rho,{\bf u}]=k_{B}T\int {\rho}\ln {\rho}
d{\bf r}+{1\over 2}\int \rho\Phi d{\bf r} +\int \rho {{\bf
u}^{2}\over
2}d{\bf r},\nonumber\\
\label{ip3}
\end{eqnarray}
and it can be viewed as a Boltzmann free energy $F=E-TS$ where $E$ is the macroscopic energy and $S$ the Boltzmann entropy. The velocity of sound (\ref{ep6}) is uniform:
\begin{equation}
c_{s}^{2}=T.
\label{ip4}
\end{equation}
Finally, for an infinite homogeneous isothermal gas, the Jeans instability criterion takes the classical form
\begin{eqnarray}
k^{2}<k^2_{c}=\frac{4\pi G\rho}{T}.
\label{ip5}
\end{eqnarray}
\subsection{Polytropic equation of state}
\label{sec_ip2}
The equation of state of a polytropic gas\footnote{The polytropic equation of
state corresponds to an adiabatic transformation for which the
specific entropy $s({\bf r},t)=s$ is uniform. In that case,
$\gamma=c_{p}/c_{v}$ is the ratio of specific heats at constant
pressure and constant volume. For a monoatomic gas $\gamma=5/3$.
This law describes convective regions of stellar interior. An
approximate polytropic equation of state with index $\gamma\simeq
3.25$ also holds in the radiative region of a star. Finally,
classical white dwarf stars are described by a polytropic equation
of state with index $\gamma=5/3$ ($n=3/2$). This equation of state
is valid for a completely degenerate gas of electrons described by
the Fermi-Dirac statistics at $T=0$ (Chandrasekhar 1942).} is
\begin{equation}
p({\bf r},t)=K \rho({\bf r},t)^{\gamma},
\label{ip6}
\end{equation}
where $K$ is a constant. The polytropic index $n$ is
defined by $\gamma=1+1/n$. For $n\rightarrow +\infty$, we recover
the isothermal case with $\gamma=1$.
At hydrostatic equilibrium, according to Eq. (\ref{ep8}), the relation between
the density and the gravitational potential can be written
\begin{eqnarray}
\rho=\biggl\lbrack \lambda-{\gamma-1\over K\gamma}\Phi\biggr
\rbrack^{1\over \gamma-1}, \label{ip7}
\end{eqnarray}
where $\lambda$ is a constant. The energy
functional (\ref{ep9}) can be put in the form
\begin{eqnarray}
{\cal W}[\rho,{\bf u}]={K\over \gamma-1}\int
(\rho^{\gamma}-\rho)d{\bf r}+{1\over 2}\int \rho\Phi d{\bf r} +\int
\rho {{\bf u}^{2}\over
2}d{\bf r}.\nonumber\\
\label{ip8}
\end{eqnarray}
In the limit $\gamma\rightarrow 1$, we recover the
results (\ref{ip2}) and (\ref{ip3}) obtained for isothermal
distributions. For a polytropic gas, the local temperature $T({\bf
r},t)$ given by Eq. (\ref{ep5}) reads
\begin{equation}
T({\bf r},t)=K\rho({\bf r},t)^{\gamma-1}.
\label{ip9}
\end{equation}
We note that the temperature is position dependent (while the
specific entropy $s$ is uniform) and related to the density by a
power law (this is the local version of the usual isentropic law
$TV^{\gamma-1}={\rm const.}$ in thermodynamics). The polytropic index $n$
is related to the gradients of temperature and density according
to
\begin{equation}
{\nabla\ln T}={1\over n}\nabla\ln\rho. \label{ip10}
\end{equation}
Combining Eqs. (\ref{ip7}) and (\ref{ip9}), we note that, for a
polytropic gas at equilibrium, the relation between the kinetic
temperature and the gravitational potential is {\it linear} so that
\begin{equation}
\nabla T=-\frac{\gamma-1}{\gamma}\nabla\Phi. \label{ip11}
\end{equation}
The coefficient of proportionality is related to the adiabatic index $\gamma$.
The velocity of sound (\ref{ep6}) can be expressed as
\begin{equation}
c_{s}({\bf r},t)^{2}=\gamma T({\bf r},t),
\label{ip12}
\end{equation}
and it is usually position dependent. However, when we consider an
infinite and homogeneous distribution (making the Jeans swindle), the
velocity of sound and the temperature are uniform. In that case, we
can speak of ``the temperature $T$ of the polytropic gas'' and write the
Jeans instability criterion as
\begin{eqnarray}
k^{2}<k^2_{c}=\frac{4\pi G\rho}{\gamma T}.
\label{ip13}
\end{eqnarray}
We thus find that the critical Jeans wavenumber $k_{c}^{(poly)}$ of a polytropic gas is related to the critical Jeans wavenumber $k_{c}^{(iso)}$ of an isothermal gas with the same temperature $T$ by
\begin{eqnarray}
k_{c}^{(poly)}=\frac{1}{\sqrt{\gamma}}k_{c}^{(iso)},
\label{ip14}
\end{eqnarray}
where the proportionality factor is the polytropic index $\gamma$.
The positivity of $c_{s}^{2}=dp/d\rho$ implies that $\gamma\ge 0$,
i.e. $n\ge 0$ or $n\le -1$ (for $-1<n<0$, the gas is always unstable,
even in the absence of gravity, since $\omega^{2}<0$). For $\gamma=0$
($n=-1$), the gas is unstable to all wavelengths since
$k_{c}^{(poly)}=+\infty$. For $0<\gamma<1$ ($n<-1$),
$k_{c}^{(poly)}>k_{c}^{(iso)}$. For $\gamma=1$ ($n=\infty$),
$k_{c}^{(poly)}=k_{c}^{(iso)}$ (isothermal case). For $\gamma>1$
($n>0$), $k_{c}^{(poly)}<k_{c}^{(iso)}$. For $\gamma=+\infty$ ($n=0$), the
gas is stable to all wavelengths since $k_{c}^{(poly)}=0$ (solid
medium). For an isentropic gas for which $\gamma=c_{p}/c_{v}$, the
Mayer relation $c_{p}-c_{v}=k_{B}$ implies that $\gamma>1$
($n>0$). Therefore, for an isentropic gas, the gravitational
instability is always delayed (i.e., it occurs for larger wavelengths)
with respect to an isothermal gas with the same temperature.
\section{Collisionless stellar systems}
\label{sec_stellar}
\subsection{The Vlasov-Poisson system}
\label{sec_vh}
We consider a self-gravitating system described by the Vlasov-Poisson
system
\begin{equation}
{\partial f\over\partial t}+{\bf v}\cdot {\partial f\over\partial {\bf r}}+{\bf F}\cdot {\partial f\over\partial {\bf v}}=0,\label{vh0}
\end{equation}
\begin{equation}
\Delta\Phi=4\pi G\int f d{\bf v},\label{vh1}
\end{equation}
where ${\bf F}=-\nabla\Phi$ is the self-consistent gravitational field
produced by the particles. The Vlasov description assumes that the
evolution of the system is encounterless. This is a very good
approximation for most astrophysical systems like galaxies and dark
matter because the relaxation time due to close encounters is in general
larger than the age of the universe by several orders of magnitude
(Binney \& Tremaine 1987). The Vlasov equation admits an infinite number of stationary solutions
given by the Jeans theorem (Jeans 1915). For example,
distribution functions of the form $f=f(\epsilon)$ which depend only on the individual energy
$\epsilon={v^2}/{2}+\Phi({\bf r})$ of the stars are particular steady states of the
Vlasov equation. They describe spherical stellar systems.
The Vlasov
equation conserves the mass $M[f]=\int f\, d{\bf r}d{\bf v}$, the
energy $E[f]=\frac{1}{2}\int fv^2 \, d{\bf r}d{\bf
v}+\frac{1}{2}\int\rho\Phi\, d{\bf r}$ and the Casimirs $I_{h}=\int
h(f)\, d{\bf r}d{\bf v}$ for any continuous function $h$. Let us
introduce the functionals
\begin{equation}
S=-\int C(f)\, d{\bf r}d{\bf v},\label{vh2}
\end{equation}
where $C$ is a convex function ($C''>0$). These particular Casimirs
will be called ``pseudo entropies''\footnote{The functionals $H=-\int
C(\bar{f})\, d{\bf r}d{\bf v}$ defined in terms of the coarse-grained
distribution $\bar{f}$ are called generalized $H$-functions (Tremaine
et al. 1986).}.
{\it The two constraints problem:} since the Vlasov equation
conserves $M$, $E$ and $S$, the {maximization} problem
\begin{equation}
\max_{f} \ \left\lbrace S[f]\quad |\quad E[f]=E, \quad M[f]=M\right \rbrace,\label{vh3}
\end{equation}
determines a steady state of the Vlasov-Poisson system that is
nonlinearly dynamically stable. The critical points of pseudo entropy
at fixed mass and energy are determined by the variational principle
$\delta S-\beta\delta E-\alpha\delta M=0$, where $\beta$ (pseudo
inverse temperature) and $\alpha$ (pseudo chemical potential) are
Lagrange multipliers. This yields $C'(f)=-\beta\epsilon-\alpha$. Since
$C$ is convex, this relation can be reversed to give
$f=F(\beta\epsilon+\alpha)$ where $F=(C')^{-1}(-x)$. We have
$f'(\epsilon)=-\beta/C''(f)$. Therefore, if $\beta>0$, the critical
points of pseudo entropy at fixed mass and energy determine
distributions of the form $f=f(\epsilon)$ with $f'(\epsilon)<0$: the
distribution function is a monotonically decreasing function of the
energy\footnote{More generally, the solutions of (\ref{vh3}) are of
the form $f=f(\epsilon)$ where $f$ is monotonic, decreasing at
positive temperatures and increasing at negative temperatures. For
realistic stellar systems, the DF should decrease close to the escape
energy $\epsilon=0$. Therefore, for the class of distributions
considered, $f$ must decrease everywhere implying $\beta>0$.}. Then, a
critical point of pseudo entropy at fixed mass and energy is a (local)
maximum iff
\begin{equation}
\delta^2{S}=-\frac{1}{2}\int C''(f)(\delta f)^2\, d{\bf r}d{\bf v}-\frac{1}{2}\beta\int \delta\rho\delta\Phi\, d{\bf r}< 0,
\label{vh3b}
\end{equation}
for all perturbations $\delta f$ that conserve mass and energy at
first order: $\delta M=\delta E=0$.
{\it The one constraint problem:} the minimization problem
\begin{equation}
\min_{f} \ \left\lbrace F[f]=E[f]-TS[f]\quad | \quad M[f]=M\right \rbrace,\label{vh4}
\end{equation}
where $T=1/\beta>0$ is prescribed, determines a steady state of the Vlasov-Poisson system that is
nonlinearly dynamically stable. The functional $F[f]$ is a pseudo free
energy. The critical points of pseudo free energy at fixed mass are
determined by the variational principle $\delta F+\alpha T\delta M=0$,
where $\alpha$ (pseudo chemical potential) is a Lagrange
multiplier. This yields $C'(f)=-\beta\epsilon-\alpha$. Then, a
critical point of pseudo free energy at fixed mass is a (local)
minimum iff
\begin{equation}
\delta^2{F}=\frac{1}{2}T\int C''(f)(\delta f)^2\, d{\bf r}d{\bf v}+\frac{1}{2}\int \delta\rho\delta\Phi\, d{\bf r}> 0,
\label{vh4b}
\end{equation}
for all perturbations $\delta f$ that conserve mass: $\delta M=0$.
{\it The no constraint problem:} the maximization problem
\begin{equation}
\max_{f} \ \left\lbrace G[f]=S[f]-\beta E[f]-\alpha M[f]\right \rbrace,\label{erica1}
\end{equation}
where $\beta$ and $\alpha$ are prescribed, determines a steady state of
the Vlasov-Poisson system that is nonlinearly dynamically stable. The functional $G[f]$ is a
pseudo grand potential. The critical points of pseudo
grand potential $G$, satisfying $\delta G=0$, are given by $C'(f)=-\beta\epsilon-\alpha$.
Then, a
critical point of pseudo grand potential is a (local)
maximum iff
\begin{equation}
\delta^2{G}=-\frac{1}{2}\int C''(f)(\delta f)^2\, d{\bf r}d{\bf v}-\frac{1}{2}\beta\int \delta\rho\delta\Phi\, d{\bf r}< 0,
\label{erica2}
\end{equation}
for all perturbations $\delta f$.
The optimization problems (\ref{vh3}), (\ref{vh4}) and (\ref{erica1}) have the {\it
same} critical points (canceling the first order variations). Furthermore, a solution
of (\ref{erica1}) is always a solution of the more constrained dual
problem (\ref{vh4}). Indeed, if inequality (\ref{erica2}) is true for
all perturbations, it is true a fortiori for all
perturbations that conserve mass. Similarly, a solution of (\ref{vh4}) is always a solution of the more constrained dual
problem (\ref{vh3}). Indeed, if inequality (\ref{vh4b}) is true for
all perturbations that conserve mass, it is true a fortiori for all
perturbations that conserve mass {\it and} energy. However, the
reciprocal is wrong. A solution of (\ref{vh4}) is not necessarily a solution
of (\ref{erica1}), and a solution of (\ref{vh3}) is not necessarily a solution
of (\ref{vh4}). This is similar to the notion of ensemble inequivalence in
thermodynamics (Ellis et al. 2000, Bouchet \& Barr\'e 2005, Chavanis 2006a). Indeed, the two
constraints problem (\ref{vh3}) is similar to a condition of
microcanonical stability, the one constraint problem (\ref{vh4}) is
similar to a condition of canonical stability, and the no constraint problem (\ref{erica1})
is similar to a condition of grand canonical stability. The implication
$(\ref{erica1})\Rightarrow (\ref{vh4})\Rightarrow (\ref{vh3})$ is similar to the fact that
grand canonical stability implies canonical stability which itself implies
microcanonical stability (but not the
converse) in thermodynamics. Therefore, (\ref{erica1}) provides just a
{\it sufficient} condition of nonlinear dynamical stability that is
less refined than (\ref{vh4}), and (\ref{vh4}) provides just a
{\it sufficient} condition of nonlinear dynamical stability that is
less refined than (\ref{vh3}).
{\it The most refined problem:} the minimization problem
\begin{equation}
\min_{f} \ \left\lbrace E[f]\quad | {\rm \ all \ the \ Casimirs\ } I_h \right\rbrace,\label{erica3}
\end{equation}
determines a steady state of the Vlasov-Poisson system that is
nonlinearly dynamically stable. A distribution function is a critical point of energy for symplectic
perturbations (i.e. perturbations that conserve all the Casimirs) iff $f({\bf r},{\bf v})$ is a steady state of the Vlasov equation (Bartholomew 1971, Kandrup 1991). Furthermore, if we consider spherical stellar systems for which $f=f(\epsilon)$, it can be shown (Bartholomew 1971, Kandrup 1991) that a critical point of energy for symplectic perturbations is a (local) minimum iff
\begin{equation}
\delta^2{E}=-\frac{1}{2}\int \frac{(\delta f)^2}{f'(\epsilon)}\, d{\bf r}d{\bf v}+\frac{1}{2}\int \delta\rho\delta\Phi\, d{\bf r}> 0,
\label{erica4}
\end{equation}
for all perturbations $\delta f$ that conserve energy and all the Casimirs at first order: $\delta E=\delta I_h=0$. Such symplectic (physically accessible) perturbations are of the form $\delta f=D\delta g$ where $D={\bf v}\cdot\nabla_{\bf r}-\nabla\Phi\cdot \nabla_{\bf v}$ is the advective operator in phase space and $\delta g({\bf r},{\bf v})$ is {\it any} perturbation. Inequality (\ref{erica4}) corresponds to the Antonov (1960) criterion of dynamical stability that was obtained by investigating the linear dynamical stability of a steady state of the Vlasov equation. Since this is the most constrained criterion, this is the most refined one. We note that inequality (\ref{erica4}) is equivalent to inequalities (\ref{vh3b}), (\ref{vh4b}) and (\ref{erica2}) if we use the identity $f'(\epsilon)=-\beta/C''(f)$ derived above. However, the classes of perturbations to consider are different: in (\ref{erica2}), we must consider all perturbations, in (\ref{vh4b}) we must consider perturbations that conserve mass, in (\ref{vh3b}) we must consider perturbations that conserve mass and energy, and in (\ref{erica4}) we must consider perturbations that conserve energy and all the Casimirs. Therefore, we have the implications $(\ref{erica1}) \Rightarrow (\ref{vh4}) \Rightarrow (\ref{vh3}) \Rightarrow (\ref{erica3})$. The connection between (\ref{erica3}) and the preceding optimization problems can be understood as follows. The minimization problem (\ref{erica3}), conserving all the Casimirs, is clearly more refined than the minimization problem
\begin{equation}
\min_{f} \ \left\lbrace E[f]\quad | \quad M[f]=M, \quad S[f]=S\right\rbrace,\label{erica5}
\end{equation}
where only the mass and {\it one} Casimir of the form (\ref{vh2}) is conserved. Now, it is easy to show that the minimization problem (\ref{erica5}) is equivalent to the maximization problem (\ref{vh3}). Hence, we have the chain of relations
\begin{equation}
(\ref{erica1}) \Rightarrow (\ref{vh4}) \Rightarrow (\ref{vh3}) \Leftrightarrow (\ref{erica5}) \Rightarrow (\ref{erica3}).
\label{erica6}
\end{equation}
To summarize, the minimization problem (\ref{erica3}) is the most refined stability criterion because it tells that, in order to settle the dynamical stability of a stellar system, we just need considering symplectic (i.e. dynamically accessible) perturbations. Of course, if inequality (\ref{erica4}) is satisfied by a larger class of perturbations, as implied by problems (\ref{erica1}), (\ref{vh4}), (\ref{vh3}) and (\ref{erica5}), the system will be stable a fortiori. Therefore, we have the implications (\ref{erica6}). Problems (\ref{erica1}), (\ref{vh4}), (\ref{vh3}) and (\ref{erica5}) provide sufficient (but not necessary) conditions of dynamical stability. A steady state can be stable according to (\ref{vh4}) while it does not satisfy (\ref{erica1}), or it can be stable according to (\ref{vh3}) and (\ref{erica5}) while it does not satisfy (\ref{vh4}), or it can be stable according to (\ref{erica3}) while it does not satisfy (\ref{vh3}) and (\ref{erica5}). Therefore, the criterion (\ref{erica3}) is more refined than (\ref{erica5}) or (\ref{vh3}), which is itself more refined than (\ref{vh4}), which is itself more refined than (\ref{erica1}). Let us give an astrophysical illustration of all that (Chavanis 2006a). Using (\ref{vh4}), we can show that stellar polytropes with $3/2\le n\le 3$ are dynamically stable because they are minima of $F$ at fixed mass $M$. By contrast, stellar polytropes with $3<n<5$ are not minima of $F$ at fixed mass. However, using (\ref{vh3}), we can show that stellar polytropes with $3/2<n<5$ are dynamically stable because they are maxima of $S$ at fixed mass and energy. However, not all stellar systems of the form $f=f(\epsilon)$ are maxima of $S$ at fixed mass and energy. But, using (\ref{erica3}), we can show that all spherical galaxies $f=f(\epsilon)$ with $f'(\epsilon)<0$ are dynamically stable. This last statement has been shown for linear stability by Kandrup (1991). A rigorous mathematical proof has been given recently by Lemou et al. (2009) for nonlinear stability.
{\it Remark}: similar results are obtained in 2D fluid mechanics based on the Euler-Poisson system (see Chavanis 2009). Criterion (\ref{erica3}) is equivalent to the so-called Kelvin-Arnol'd energy principle, criterion (\ref{erica1}) is equivalent to the standard Casimir-energy method (see Holm et al. 1985) introduced by Arnol'd (1966) and criterion (\ref{vh3}) is equivalent to the refined stability criterion given by Ellis et al. (2002).
\subsection{The corresponding barotropic star}
\label{sec_vhc}
To any stellar system with $f=f(\epsilon)$ and $f'(\epsilon)<0$, we
can associate a corresponding barotropic star with the same
equilibrium density distribution (Lynden-Bell \& Sanitt 1969). Indeed,
defining the density and the pressure by $\rho=\int f d{\bf v}$ and
$p={1\over 3}\int f v^{2}d{\bf v}$, we get $\rho=\rho(\Phi)$ and
$p=p(\Phi)$. Eliminating the potential $\Phi$ between these two
expressions, we find that $p=p(\rho)$. Furthermore, taking the
gradient of the pressure and using the chain of identities
\begin{eqnarray}
\nabla p={1\over 3}\int {\partial f\over\partial {\bf r}}v^{2}d{\bf v}={1\over 3}\nabla \Phi \int f'(\epsilon) v^{2}d{\bf v}\nonumber\\
={1\over 3}\nabla \Phi \int {\partial f\over\partial {\bf v}}\cdot {\bf v}\, d{\bf v}=-\nabla \Phi \int f \, d{\bf v}=-\rho\nabla \Phi,
\label{vhc1}
\end{eqnarray}
we obtain the condition of hydrostatic equilibrium (\ref{ep7}). Finally, we define the kinetic temperature of an isotropic stellar system by the relation
\begin{eqnarray}
\frac{3}{2}T({\bf r})=\frac{1}{2}\langle v^2\rangle.
\label{vhc2a}
\end{eqnarray}
Therefore, the quantity
\begin{eqnarray}
T({\bf r})=\frac{1}{3}\langle v^2\rangle={{1\over 3}\int f v^{2}d{\bf v}\over \int f d{\bf v}}={p({\bf r})\over \rho({\bf r})}.
\label{vhc2b}
\end{eqnarray}
measures the velocity dispersion (in one direction) of an isotropic
stellar system. In general, the kinetic
temperature is position dependent.
Finally, we can show that the variational principles $(\ref{vh4})$ and
$(\ref{ep10})$ are equivalent, i.e. a stellar system is a minimum of
pseudo free energy at fixed mass iff the corresponding barotropic gas
is a minimum of energy at fixed mass (see Appendix \ref{sec_eqa}). This leads to a
nonlinear version of the Antonov (1960b) first law: ``a stellar system with
$f=f(\epsilon)$ and $f'(\epsilon)<0$ is nonlinearly dynamically stable
with respect to the Vlasov-Poisson system if the corresponding
barotropic gas is nonlinearly dynamically stable with respect to the
Euler-Poisson system''. However, the reciprocal is wrong because, as
we have already indicated, (\ref{vh4}) provides just a {\it
sufficient} condition of nonlinear dynamical stability with respect to
the Vlasov equation. A galaxy can be dynamically stable according to
criterion (\ref{vh3}) [or even more generally (\ref{erica3})]
while it fails to satisfy criterion
(\ref{vh4}). Therefore, the nonlinear Antonov first law is similar to
a notion of ensembles inequivalence between microcanonical and
canonical ensembles in thermodynamics (Chavanis 2006a).
\subsection{The dispersion relation}
\label{sec_lsv}
We shall study the linear dynamical stability of a spatially
homogeneous stationary solution of the Vlasov equation described by a
distribution function $f=f({\bf v})$. Like for an infinite homogeneous
gas, we make the Jeans swindle. Linearizing the Vlasov equation around
this steady state, taking the Laplace-Fourier transform of this
equation and writing the perturbations in the form $e^{i({\bf k}\cdot
{\bf r}-\omega t)}$, we obtain the classical dispersion relation
(Binney \& Tremaine 1987):
\begin{eqnarray}
\epsilon({\bf k},\omega)\equiv 1+{4\pi G\over k^{2}}\int {{\bf k}\cdot {\partial f\over\partial {\bf v}}\over {\bf k}\cdot {\bf v}-\omega}d{\bf v}=0,
\label{lsv1}
\end{eqnarray}
where $\epsilon({\bf k},\omega)$ is similar to the ``dielectric
function'' of plasma physics (with the sign $+$ instead of $-$). For a given distribution $f({\bf v})$, this equation determines the complex pulsation(s)
$\omega=\omega_{r}+i\omega_{i}$ of a perturbation with wavevector
${\bf k}$. Since the time dependence of the perturbation is $\delta f\sim e^{-i\omega_r t}e^{\omega_i t}$, the system is linearly
stable if $\omega_{i}<0$ and linearly unstable if $\omega_{i}>0$. The condition of marginal stability corresponds to $\omega_i=0$.
If we take the wavevector ${\bf k}$ along the $z$-axis\footnote{If the distribution function is isotropic, there is no
restriction in making this choice.}, the dispersion relation becomes
\begin{eqnarray}
\epsilon({k},\omega)\equiv 1+{4\pi G\over k^{2}}\int_{C} {{k}{\partial f\over\partial {v_z}}\over {k} {v_z}-\omega}d{v}_x d{v}_y d{v}_z=0,
\label{lsv2}
\end{eqnarray}
where the integral must be performed along the Landau contour $C$ (Binney \& Tremaine 1987). We define
\begin{eqnarray}
g(v_z)=\int f dv_x dv_y.
\label{lsv3}
\end{eqnarray}
In the following, we shall note $v$ instead of $v_z$ and $f$ instead of $g$. With these conventions, the dispersion relation (\ref{lsv2}) can be rewritten
\begin{eqnarray}
\epsilon({k},\omega)\equiv 1+{4\pi G\over k^{2}}\int_{C} {f'(v)\over {v}-\frac{\omega}{k}}d{v}=0.
\label{lsv4}
\end{eqnarray}
This is the fundamental equation of the problem. For future
reference, let us recall that
\begin{eqnarray}
\int_{C} {{f'(v)}\over v-\frac{\omega}{k}}dv=\int_{-\infty}^{+\infty} {{f'(v)}\over v-\frac{\omega}{k}}dv, \quad (\omega_i>0),
\label{lsv5}
\end{eqnarray}
\begin{eqnarray}
\int_{C} {{f'(v)}\over v-\frac{\omega}{k}}dv=P\int_{-\infty}^{+\infty} {{f'(v)}\over v-\frac{\omega}{k}}dv+i\pi f'\left (\frac{\omega}{k}\right ), \quad (\omega_i=0),\nonumber\\
\label{lsv6}
\end{eqnarray}
\begin{eqnarray}
\int_{C} {{f'(v)}\over v-\frac{\omega}{k}}dv=\int_{-\infty}^{+\infty} {{f'(v)}\over v-\frac{\omega}{k}}dv+ 2\pi i f'\left (\frac{\omega}{k}\right ), \quad (\omega_i<0),\nonumber\\
\label{lsv7}
\end{eqnarray}
where $P$ denotes the principal value.
In general, the dispersion relation (\ref{lsv4}) cannot be solved explicitly to obtain $\omega(k)$ except in some very simple cases \footnote{There exists less simple cases were explicit solutions can be obtained. We should mention in this respect the extensive set of closed-form solutions for generalized Lorentzian distributions due to Summers \& Thorne (1991) and Thorne \& Summers (1991). They correspond to Tsallis (polytropic) distributions of the form (\ref{pdr1}) with negative index $n<-1$.}. For example, for cold systems described by the distribution function $f=\rho \delta({v}-{v}_0)$, we obtain after an integration by parts
\begin{eqnarray}
\omega=v_0 k\pm i\sqrt{4\pi G\rho}.
\label{lsv8}
\end{eqnarray}
In particular, when $v_0=0$, we get
\begin{eqnarray}
\omega=\pm i\sqrt{4\pi G\rho}.
\label{lsv9}
\end{eqnarray}
The system is unstable to all wavenumbers and the perturbation grows
with a growth rate $\omega_i=\sqrt{4\pi G\rho}$. We also note that the
dispersion relation (\ref{lsv9}) coincides with the dispersion
relation (\ref{jeans5}) of a self-gravitating gas with $c_s=0$.
\subsection{The condition of marginal stability}
\label{sec_cms}
For $\omega_{i}=0$, the real and imaginary parts of the dielectric function $\epsilon(k,\omega_{r})=\epsilon_r(k,\omega_{r})+i\epsilon_i(k,\omega_{r})$ are given by
\begin{eqnarray}
\epsilon_r({k},\omega_r)= 1+\frac{4\pi G}{k^2}P\int_{-\infty}^{+\infty} {f'(v)\over {v}-{\omega_r}/{k}}d{v},
\label{cms1}
\end{eqnarray}
\begin{eqnarray}
\epsilon_i({k},\omega_r)= \frac{4\pi^2 G}{k^2}f'\left ({\omega_r}/{k}\right ).
\label{cms2}
\end{eqnarray}
The condition of marginal stability corresponds to $\epsilon(k,\omega)=0$ and $\omega_i=0$, i.e. $\epsilon_r({k},\omega_r)=\epsilon_i({k},\omega_r)=0$. We obtain therefore the equations
\begin{eqnarray}
1+\frac{4\pi G}{k^2}P\int_{-\infty}^{+\infty} {f'(v)\over {v}-{\omega_r}/{k}}d{v}=0,
\label{cms1b}
\end{eqnarray}
\begin{eqnarray}
f'\left ({\omega_r}/{k}\right )=0.
\label{cms2b}
\end{eqnarray}
The second condition (\ref{cms2b}) imposes that the phase velocity
$\omega_r/k=v_{ext}$ is equal to a velocity where $f(v)$ is extremum
($f'(v_{ext})=0$). The first condition (\ref{cms1b}) then determines
the wavenumber(s) $k_c$ corresponding to marginal stability. It can be
written
\begin{eqnarray}
1+\frac{4\pi G}{k_c^2}\int_{-\infty}^{+\infty} {f'(v)\over {v}-v_{ext}}d{v}=0,
\label{cms3}
\end{eqnarray}
where the principal value is not needed anymore. The wavenumber(s) corresponding to marginal stability are therefore given by
\begin{eqnarray}
k_{c}=\left (-4\pi G\int_{-\infty}^{+\infty} \frac{f'(v)}{v-v_{ext}}\, dv\right )^{1/2}.
\label{cms4}
\end{eqnarray}
Finally, the pulsation(s) corresponding to marginal stability are $\omega_r=v_{ext}k_c$ and we have $\delta f\sim e^{-i\omega_r t}$. Note that the
distribution $f(v)$ can be relatively arbitrary. There can be pure oscillations ($\omega=\omega_r\neq 0$) only if $f(v)$ has some extrema at $v\neq 0$. If $f(v)$ has a
single maximum at $v=0$, then $\omega_{r}=0$ (implying $\omega=0$) and
the condition of marginal stability becomes
\begin{eqnarray}
k_{c}=\left (-4\pi G\int_{-\infty}^{+\infty} \frac{f'(v)}{v}\, dv\right )^{1/2}.
\label{cms5}
\end{eqnarray}
\subsection{The Nyquist method}
\label{sec_n}
To determine whether the distribution $f=f(v)$ is stable or unstable
for a perturbation with wavenumber $k$, one possibility is to solve
the dispersion relation (\ref{lsv4}) and determine the sign of the
imaginary part of the complex pulsation. This can be done analytically
in some simple cases (see Secs. \ref{sec_mon}, \ref{sec_vb} and
\ref{sec_vb2}). We can also apply the Nyquist method introduced in plasma
physics. This is a graphical method that does not require to solve the
dispersion relation. The details of the method are explained by
Nicholson (1992) and we just recall how it works in practice. In the
$\epsilon$-plane, taking $\omega_i=0$, we plot the Nyquist
curve\footnote{This curve is also called an hodograph.}
$(\epsilon_{r}(k,\omega_{r}),\epsilon_{i}(k,\omega_{r}))$
parameterized by $\omega_{r}$ going from $-\infty$ to $+\infty$ (for a
given wavenumber $k$). This curve is closed and always rotates in the
counterclockwise sense. If the Nyquist curve does not encircle the
origin, the system is stable (for the corresponding wavenumber
$k$). If the Nyquist curve encircles the origin one or more times, the
system is unstable. The number $N$ of tours around the origin gives
the number of zeros of $\epsilon(k,\omega)$ in the upper half
$\omega$-plane, i.e. the number of unstable modes with $\omega_{i}>0$.
The Nyquist method by itself does not give the growth rate of the
instability.
Let us consider the asymptotic behavior of
$(\epsilon_{r}(k,\omega_{r}),\epsilon_{i}(k,\omega_{r}))$ defined by
Eqs. (\ref{cms1}) and (\ref{cms2}) for $\omega_{r}\rightarrow
\pm\infty$. Since $f(v)$ is positive and tends to zero for
$v\rightarrow \pm\infty$, we conclude that
$\epsilon_{i}(k,\omega_{r})\rightarrow 0$ for $\omega_{r}\rightarrow
\pm\infty$ and that $\epsilon_{i}(k,\omega_{r})>0$ for
$\omega_{r}\rightarrow -\infty$ while $\epsilon_{i}(k,\omega_{r})<0$ for
$\omega_{r}\rightarrow +\infty$. On the other hand, integrating by
parts in Eq. (\ref{cms1}), we obtain
\begin{eqnarray}
\epsilon_{r}(k,\omega_{r})=1+\frac{4\pi G}{k^2}{P}\int_{-\infty}^{+\infty} {{f(v)}\over (v-\omega_{r}/k)^{2}}dv,
\label{gp4}
\end{eqnarray}
provided that $f(v)$ decreases sufficiently rapidly. Therefore, for $\omega_{r}\rightarrow \pm\infty$, we obtain at leading order
\begin{eqnarray}
\epsilon_{r}(k,\omega_{r})\simeq 1+\frac{4\pi G\rho}{\omega_{r}^{2}}, \qquad (\omega_{r}\rightarrow \pm\infty).
\label{gp5}
\end{eqnarray}
In particular, $\epsilon_{r}(k,\omega_{r})\rightarrow 1^+$ for
$\omega_{r}\rightarrow
\pm\infty$.
From these results, we conclude that the behavior of the Nyquist curve
close to the limit point $(1,0)$ is like the one represented in
Fig. \ref{maxwell}. In addition, according to Eq. (\ref{cms2}), the
Nyquist curve crosses the $x$-axis at each value of $\omega_{r}/k$
corresponding to an extremum of $f(v)$. For $\omega_r/k=v_{ext}$, where
$v_{ext}$ is a velocity at which the distribution is extremum
$(f'(v_{ext})=0)$, the imaginary part of the dielectric function
$\epsilon_i(k,k v_{ext})=0$ and the real part of the dielectric function
\begin{equation}
\epsilon_r(k,k v_{ext})= 1+\frac{4\pi G}{k^2} \int_{-\infty}^{\infty}\frac{f'(v)}{v-v_{ext}}\, dv.
\end{equation}
Subtracting the value $f'(v_{ext})=0$ in the numerator of the integrand, and integrating by parts, we obtain
\begin{equation}
\epsilon_r(k, k v_{ext})= 1-\frac{4\pi G}{k^2} \int_{-\infty}^{\infty}\frac{f(v_{ext})-f(v)}{(v-v_{ext})^2}\, dv.
\label{gp6}
\end{equation}
If $v_{Max}$ denotes the velocity corresponding to the global maximum of
the distribution, we clearly have
\begin{eqnarray}
\epsilon_{r}(k, k v_{Max})= 1-\frac{4\pi G}{k^2} \int_{-\infty}^{\infty}\frac{f(v_{Max})-f(v)}{(v-v_{Max})^2}\, dv<1.\nonumber\\
\label{gp7}
\end{eqnarray}
Furthermore, $\epsilon_r(k,kv_{Max})<0$ for sufficiently small
$k$. Therefore, by tuning $k$ appropriately, we can always make the
Nyquist curve encircle the origin. We conclude that a spatially
homogeneous self-gravitating system is always unstable to some
wavelengths.
\subsection{Single-humped distributions}
\label{sec_sh}
Let us assume that the distribution $f(v)$ has a single maximum at
$v=v_{0}$ (so that $f'(v_0)=0$) and tends to zero for $v\rightarrow
\pm \infty$. Then, the Nyquist curve cuts the $x$-axis ($\epsilon_{i}(k,\omega_{r})$ vanishes) at the limit point
$(1,0)$ where $\omega_{r}\rightarrow \pm \infty$ and at the point
$(\epsilon_{r}(k, k v_0),0)$ where $\omega_{r}/k=v_{0}$. Due to its behavior
close to the limit point $(1,0)$, the fact that it rotates in the
counterclockwise sense, and the property that $\epsilon_r(k, k v_0)<1$ according to Eq. (\ref{gp7}),
the Nyquist curve must necessarily behave like in
Fig. \ref{maxwell}. Therefore, the Nyquist curve starts on the real
axis at $\epsilon_r(k, \omega_r) =1$ for $\omega_r \rightarrow -\infty$,
then going in counterclockwise sense it crosses the real axis at the
point $\epsilon_r(k, k v_0)<1$ and returns on the real axis at
$\epsilon_r(k,\omega_r) =1$ for $\omega_r \rightarrow +
\infty$. According to the Nyquist criterion exposed in Sec. \ref{sec_n}, we
conclude that a single-humped distribution is linearly
stable with respect to a perturbation with wavenumber $k$ if
\begin{eqnarray}
\epsilon_{r}(k, k v_0)=1+\frac{4\pi G}{k^2}\int_{-\infty}^{+\infty} {{f'(v)}\over v-v_0}dv>0,
\label{sh1}
\end{eqnarray}
and linearly unstable if $\epsilon_{r}(k, k v_0)<0$. The equality
corresponds to the marginal stability condition
(\ref{cms3}). Therefore, the system is stable iff
\begin{eqnarray}
k>k_{c}=\left (-4\pi G\int_{-\infty}^{+\infty} \frac{f'(v)}{v-v_{0}}\, dv\right )^{1/2},
\label{sh2}
\end{eqnarray}
where $k_{c}$ is the critical Jeans wavenumber for a stellar system. Note that an infinite homogeneous stellar system whose DF has a single hump is always unstable to sufficiently small wavenumbers. For the unstable wavenumbers $k<k_c$, there is only one mode of instability $\omega_i>0$ since the Nyquist curve rotates only one time
around the origin. This stability criterion is valid for any single-humped distribution. In particular, a symmetric distribution $f=f(v)$ with a single maximum at $v_0=0$ is linearly dynamically stable to a perturbation with wavenumber $k$ iff
\begin{eqnarray}
k>k_{c}=\left (-4\pi G\int_{-\infty}^{+\infty} \frac{f'(v)}{v}\, dv\right )^{1/2}.
\label{sh3}
\end{eqnarray}
We shall make the
connection between the stability of an infinite homogeneous stellar system and the stability
of the corresponding barotropic gas in Sec. \ref{sec_fo}. In
particular, using identity (\ref{fo3}), we will show that
Eq. (\ref{sh3}) is equivalent to
Eq. (\ref{jeans6}).
\subsection{Double-humped distributions}
\label{sec_doubh}
Let us consider a double-humped distribution with a global maximum at
$v_{Max}$, a minimum at $v_{min}$ and a local maximum at $v_{max}$. We
assume that $v_{Max}<v_{min}<v_{max}$. The Nyquist curve will cut the
$x$-axis at the limit point $(1,0)$ and at three other points
$(\epsilon_{r}(k,k v_{Max}),0)$, $(\epsilon_{r}(k,k v_{min}),0)$ and
$(\epsilon_{r}(k,k v_{max}),0)$. We also know that the Nyquist curve can only
rotate in the counterclockwise sense and that $\epsilon_{r}(k,k v_{Max}) <1$ according to Eq.
(\ref{gp7}). Then, we can convince ourselves, by making drawings, of the following results. If\footnote{In the following, in order to simplify the notations, we note $\epsilon_{r}(v_{Max})$ for $\epsilon_{r}(k,kv_{Max})$ etc.}
$(+++)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})>0$, $\epsilon_{r}(v_{max})>0$,
$(+--)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})<0$, $\epsilon_{r}(v_{max})<0$,
$(--+)$: $\epsilon_{r}(v_{Max})<0$, $\epsilon_{r}(v_{min})<0$, $\epsilon_{r}(v_{max})>0$,
$(+-+)$ : $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})<0$, $\epsilon_{r}(v_{max})>0$,
\noindent the Nyquist curve does not encircle the origin so the
system is stable. If
$(---)$: $\epsilon_{r}(v_{Max})<0$, $\epsilon_{r}(v_{min})<0$, $\epsilon_{r}(v_{max})<0$,
$(-++)$: $\epsilon_{r}(v_{Max})<0$, $\epsilon_{r}(v_{min})>0$, $\epsilon_{r}(v_{max})>0$,
$(++-)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})>0$, $\epsilon_{r}(v_{max})<0$,
\noindent the Nyquist curve rotates one time
around the origin so that there is one mode of instability. Finally, if
$(-+-)$: $\epsilon_{r}(v_{Max})<0$, $\epsilon_{r}(v_{min})>0$, $\epsilon_{r}(v_{max})<0$,
\noindent the Nyquist curve rotates two times around the origin
so that there are two modes of instability. Cases $(+++)$, $(---)$, $(-++)$ and $(-+-)$ are observed in
Sec. \ref{sec_vha} for an asymmetric double-humped distribution
made of two Maxwellians. The other cases cannot be obtained from this
distribution but they may be obtained from other distributions.
If the double-humped distribution is symmetric with respect to the
origin with two maxima at $\pm v_{*}$ and a minimum at $v=0$, only
three cases can arise. If
$(+++)$: $\epsilon_{r}(v_{*})>0$, $\epsilon_{r}(0)>0$,
$(+-+)$: $\epsilon_{r}(v_{*})>0$, $\epsilon_{r}(0)<0$,
\noindent the Nyquist curve does not encircle the origin so the
system is stable. If
$(---)$: $\epsilon_{r}(v_{*})<0$, $\epsilon_{r}(0)<0$,
\noindent the Nyquist curve rotates one time
around the origin so that there is one mode of instability. Finally, if
$(-+-)$: $\epsilon_{r}(v_{*})<0$, $\epsilon_{r}(0)>0$,
\noindent the Nyquist curve rotates two times around the origin
so that there are two modes of instability. Cases $(+++)$, $(---)$ and $(-+-)$ are
observed in Sec. \ref{sec_vh} for a symmetric double-humped
distribution made of two Maxwellians.
\subsection{Particular solutions of $\epsilon(k,\omega)=0$}
\label{sec_mon}
We can look for a solution of the dispersion relation
$\epsilon(k,\omega)=0$ in the form $\omega=i\omega_{i}$ corresponding to
$\omega_{r}=0$. In that case, the perturbation grows ($\omega_{i}>0$)
or decays ($\omega_{i}<0$) without oscillating. For $\omega_{i}>0$,
the equation $\epsilon(k,\omega)=0$ becomes
\begin{eqnarray}
1+{4\pi G\over k^{2}}\int_{-\infty}^{+\infty}\frac{f'(v)}{v-i\frac{\omega_{i}}{k}}\, dv=0.
\label{mon1}
\end{eqnarray}
Multiplying the numerator by $v+i\omega_{i}/k$ and separating real and imaginary parts, we obtain
\begin{eqnarray}
1+{4\pi G\over k^{2}}\int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+\frac{\omega_{i}^2}{k^2}}\, dv=0,
\label{mon2}
\end{eqnarray}
\begin{eqnarray}
\int_{-\infty}^{+\infty}\frac{f'(v)}{v^2+\frac{\omega_{i}^2}{k^2}}\, dv=0.
\label{mon3}
\end{eqnarray}
If we consider distribution functions $f(v)$ that are symmetric with
respect to $v=0$, Eq. (\ref{mon3}) is always satisfied. Then, the
growth rate $\omega_i>0$ is given by Eq. (\ref{mon2}).
For $\omega_{i}<0$, the equation $\epsilon(k,\omega)=0$ becomes
\begin{eqnarray}
1+{4\pi G\over k^{2}}\int_{-\infty}^{+\infty}\frac{f'(v)}{v-i\frac{\omega_{i}}{k}}\, dv+i\frac{8\pi^2G}{k^2} f'\left (\frac{i\omega_{i}}{k}\right )=0.
\label{mon4}
\end{eqnarray}
Multiplying the numerator by $v+i\omega_{i}/k$ and assuming that
$f(v)$ is even, we obtain
\begin{eqnarray}
1+{4\pi G\over k^{2}}\int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+\frac{\omega_{i}^2}{k^2}}\, dv+i \frac{8\pi^2G}{k^2} f'\left (\frac{i\omega_{i}}{k}\right )=0,
\label{mon5}
\end{eqnarray}
which determines the damping rate $\omega_i<0$.
Let us introduce the function $K(x)$ defined by
\begin{eqnarray}
K(x)=\frac{1}{I}\int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+x^2}\, dv \ (x\ge 0)
\label{kk1}
\end{eqnarray}
\begin{eqnarray}
K(x)=\frac{1}{I}\left\lbrace \int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+x^2}\, dv+2\pi i f'(ix)\right\rbrace \ (x\le 0)
\label{kk2}
\end{eqnarray}
where
\begin{eqnarray}
I=\int_{-\infty}^{+\infty}\frac{f'(v)}{v}\, dv.
\label{kk3}
\end{eqnarray}
This function is normalized such that $K(0)=1$. The dispersion relations (\ref{mon2}) and (\ref{mon5}) can then be written
\begin{eqnarray}
1-\frac{k_c^2}{k^2}K\left (\frac{\omega_i}{k}\right )=0,
\label{kk4}
\end{eqnarray}
where $k_c$ is the maginal wavenumber corresponding to $\omega=0$ given by Eq. (\ref{cms5}). The pulsation $\omega_i(k)$ is given by
\begin{eqnarray}
\frac{\omega_i}{k_c}=\frac{k}{k_c}K^{-1}\left (\frac{k^2}{k_c^2}\right ).
\label{kk5}
\end{eqnarray}
Setting $u=\omega_i/k$, it can also be written in parametric form as
\begin{eqnarray}
\frac{\omega_i}{k_c}=u\sqrt{K(u)}, \qquad \frac{k}{k_c}=\sqrt{K(u)},
\label{kk6}
\end{eqnarray}
where $u$ goes from $-\infty$ to $+\infty$.
Let us obtain some asymptotic expansions of these relations (valid for
symmetric distributions):
(i) Let us first consider the case $\omega_i>0$ and $k\rightarrow 0$ corresponding to instability. Integrating Eq. (\ref{mon2}) by parts, we obtain
\begin{eqnarray}
1-4\pi G\int_{-\infty}^{+\infty}\frac{f(v)(\omega_{i}^{2}-k^2v^2)}{(k^2v^2+\omega_{i}^2)^{2}}\, dv=0.
\label{mon6}
\end{eqnarray}
Expanding the integrand in powers of $kv/\omega_i\ll 1$, we find that
\begin{eqnarray}
\omega_{i}^2=4\pi G\rho-3Tk^2-... \qquad (k\rightarrow 0),
\label{mon7q}
\end{eqnarray}
with $T=\langle v^2\rangle$ (where we recall that $v=v_z$ in the
present case). This
expression can be compared with the corresponding expression
(\ref{fe16}) valid for a gas. This identification yields $c_s^2=3T$
so that large wavelength perturbations in a collisionless stellar system
correspond to one dimensional isentropic perturbations with index $\gamma=3$ in a gas (see
Appendix \ref{sec_fee}).
(ii) The case $\omega_i<0$ and $k\rightarrow +\infty$ corresponding to
stability cannot be treated at a general level because the result
depends on the behavior of the distribution function for large
velocities. The case of a Maxwellian distribution is specifically
considered in Sec. \ref{sec_vb}.
(iii) Let us finally consider the behavior of the dispersion relation
(\ref{mon2}) or (\ref{mon5}) close to the point of marginal
stability $k=k_c$. For $\omega_i=0$, we obtain the critical wavenumber $k_c$
given by Eq. (\ref{cms5}). Let us consider $\omega_i\rightarrow 0^+$
and $k\rightarrow k_c^{-}$ in Eq. (\ref{mon2}). We introduce the
function
\begin{eqnarray}
F(x)=\int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+x^2}\, dv,
\label{mon8}
\end{eqnarray}
for any real $x$. For $x\rightarrow 0$, we have the Taylor expansion $F(x)=F(0)+F'(0)x+...$ with
\begin{eqnarray}
F(0)=\int_{-\infty}^{+\infty}\frac{f'(v)}{v}\, dv,
\label{mon9}
\end{eqnarray}
\begin{eqnarray}
F'(x)=-\int_{-\infty}^{+\infty}\frac{2xv f'(v)}{(v^2+x^2)^2}\, dv=-\int_{-\infty}^{+\infty}\frac{x f''(v)}{v^2+x^2}\, dv,\nonumber\\
\label{mon10}
\end{eqnarray}
where we have used an integration by parts to get the last expression. Under this form, we cannot take the limit $x\rightarrow 0$ in the integral because the integral is not convergent for $x=0$. However, if we write Eq. (\ref{mon10}) in the equivalent form
\begin{eqnarray}
F'(x)=-x\int_{-\infty}^{+\infty}\frac{f''(v)-f''(0)}{v^2+x^2}\, dv-\int_{-\infty}^{+\infty}\frac{ x f''(0)}{v^2+x^2}\, dv,\nonumber\\
\label{mon11}
\end{eqnarray}
we obtain
\begin{eqnarray}
F'(0)=-\lim_{x\rightarrow 0}x f''(0)\int_{-\infty}^{+\infty}\frac{dv}{v^2+x^2}=-\pi f''(0) {\rm sign}(x).\nonumber\\
\label{mon12}
\end{eqnarray}
Regrouping the previous results, we find that the dispersion relation
(\ref{mon2}) becomes for $\omega_i\rightarrow 0^+$:
\begin{eqnarray}
1+\frac{4\pi G}{k^2}\left\lbrace \int_{-\infty}^{+\infty}\frac{f'(v)}{v}\, dv-\pi f''(0)\frac{\omega_i}{k}\right\rbrace=0.
\label{mon13}
\end{eqnarray}
Similarly, the dispersion relation
(\ref{mon5}) becomes for $\omega_i\rightarrow 0^-$:
\begin{eqnarray}
1+\frac{4\pi G}{k^2}\left\lbrace \int_{-\infty}^{+\infty}\frac{f'(v)}{v}\, dv+\pi f''(0)\frac{\omega_i}{k}\right\rbrace\nonumber\\
-\frac{8\pi^2G}{k^2}f''(0)\frac{\omega_i}{k}=0,
\label{mon14}
\end{eqnarray}
which takes the same form as Eq. (\ref{mon13}). Therefore,
Eq. (\ref{mon13}) is valid for $\omega_i\rightarrow 0$ whatever its
sign. To leading order, we obtain\footnote{In the derivation, we have
assumed that $f''(0)\neq 0$. If $f''(0)=0$, we need to develop the
Taylor expansion to the next order.}
\begin{eqnarray}
\omega_i=\frac{-k_c^3}{4\pi^2 G f''(0)}\left (1-\frac{k^2}{k_c^2}\right ), \qquad (k\rightarrow k_c).
\label{mon15}
\end{eqnarray}
This formula leads to the following result. First of all, we recall that the mode of marginal stability that we consider corresponds to $\omega_r/k=v_{ext}=0$, i.e. to the extremum value of the distribution function $f(v)$ at $v_{ext}=0$. If the distribution is maximum at $v=0$, so that $f''(0)<0$, we find that the mode $\omega=i\omega_i$ is stable for $k>k_c$ and unstable for $k<k_c$. Alternatively, if the distribution is minimum at $v=0$, so that $f''(0)>0$, we find that the mode $\omega=i\omega_i$ is stable for $k<k_c$ and unstable for $k>k_c$. This result will be illustrated in connection to Fig. \ref{GRAVdiagphasesym} for the symmetric double humped distribution.
In this section, we have discussed particular solutions of the
dispersion relation of the form $\omega=i\omega_i$. Of course, there
may exist other solutions to the equation $\epsilon(k,\omega)=0$ with
$\omega_r\neq 0$\footnote{The modes $\omega=\omega_r+i\omega_i$ with $\omega_i>0$ and $\omega_r\neq 0$ are sometimes called {\it overstable}.}. However, for single-humped distributions and
unstable wavenumbers $k<k_c$, the Nyquist curve encircles the origin
only once (see Sec. \ref{sec_sh}) so that, when it exists,
the solution $\omega=i\omega_i$ with $\omega_{i}>0$ is the only solution of
$\epsilon(k,\omega)=0$ (for single-humped distributions and stable
wavenumbers $k>k_c$, there may exist other solutions of
$\epsilon(k,\omega)=0$ with $\omega_{r}\neq 0$ and
$\omega_{i}<0$ as discussed in Sec. \ref{sec_mn}). Explicit solutions of the dispersion relation
$\epsilon(k,\omega)=0$ with $\omega=i\omega_{i}$ are given in Secs. \ref{sec_vb} and \ref{sec_vb2}
for isothermal and polytropic distributions.
\subsection{Equivalence between the Jeans instability criterion for a stellar system and the Jeans instability criterion for the corresponding barotropic gas}
\label{sec_fo}
Lynden-Bell (1962) first observed that the critical Jeans length for a
stellar system described by a Maxwellian distribution function is
equal to the critical Jeans length for an isothermal gas if we replace
the velocity of sound by the velocity dispersion in one direction. In
this section, we provide the appropriate generalization of this result
for an {\it arbitrary} distribution of the form $f=f(v^2)$ with $f'<0$.
As indicated in Sec. \ref{sec_vhc}, to any stellar system with a
distribution function $f=f(\epsilon)$ and $f'(\epsilon)<0$, we can
associate a corresponding barotropic gas with an equation of state
$p=p(\rho)$. According to Eq. (\ref{ep8}), the condition of
hydrostatic equilibrium can be written
\begin{eqnarray}
p'(\rho)=-\frac{\rho}{\rho'(\Phi)}.
\label{fo1}
\end{eqnarray}
Now, using Eq. (\ref{ep6}) and the fact that $\rho=\int f(\epsilon)\,
d{\bf v}$, we get
\begin{eqnarray}
c_{s}^2({\bf r})=-\frac{\rho({\bf r})}{\int f'(\epsilon)\, d{\bf v}}=-\frac{\rho({\bf r})}{\int \frac{\frac{\partial f}{\partial v_z}}{v_z} d{\bf v}}=-\frac{\rho({\bf r})}{\int_{-\infty}^{+\infty} \frac{\frac{\partial g}{\partial v_z}}{v_z} d{v}_{z}},\nonumber\\
\label{fo2}
\end{eqnarray}
where $g({\bf r},v_z)=\int f(\epsilon) \, dv_xdv_y$. For a spatially homogeneous system, we obtain the identity
\begin{eqnarray}
c_{s}^2=-\frac{\rho}{\int_{-\infty}^{+\infty} \frac{f'(v)}{v}\, d{v}},
\label{fo3}
\end{eqnarray}
where we have noted $v$ for $v_z$ and $f$ for $g$ like in
Sec. \ref{sec_lsv}. This identity is explicitly checked in Appendix \ref{sec_cali} for the isothermal and polytropic distributions. Since $f(v)$ is symmetric with respect to $v=0$ and
has a single maximum at $v=0$, the Jeans instability criterion can be written (see Sec. \ref{sec_sh}):
\begin{eqnarray}
k^2<k_{c}^2=-4\pi G\int_{-\infty}^{+\infty} \frac{f'(v)}{v}\, dv.
\label{fo4}
\end{eqnarray}
Using identity (\ref{fo3}), it can be rewritten
\begin{eqnarray}
k^2<k_{c}^2=\frac{4\pi G\rho}{c_{s}^{2}}.
\label{fo5}
\end{eqnarray}
Therefore, the criterion of dynamical stability for a spatially homogeneous stellar system
coincides with the criterion of dynamical stability for the
corresponding barotropic gas (see Sec. \ref{sec_jeans}). We insist on the fact
that this equivalence is valid for an arbitrary distribution function
of the form $f=f(v^2)$ with $f'<0$, not only for the
Maxwellian. We conclude that: an infinite homogeneous stellar system
with $f=f(v^2)$ and $f'<0$ is dynamically stable with respect to a
perturbation with wavenumber $k$ if and only if the corresponding
barotropic gas is dynamically stable with respect to this
perturbation. This is the proper formulation of the Antonov first law
for spatially homogeneous systems: in the present case, we have
equivalence\footnote{This ``equivalence'' for
the dynamical stability of a homogeneous stellar system and the
corresponding barotropic gas differs from the case of inhomogeneous
systems where the limits of stability of stars and galaxies are generically different (see,
e.g., Chavanis 2006a).}. Of course, although the thresholds of
stability/instability coincide, the evolution of the perturbation is
different in a stellar system and in a fluid system (see Secs. \ref{sec_vb} and \ref{sec_vb2}). We
should also emphasize that, in general, the velocity of sound $c_s$ in
the corresponding barotropic gas is {\it not} equal to the velocity
dispersion $\langle v^2\rangle^{1/2}$ in the stellar system, except
when $f(v)$ is the Maxwellian distribution leading to Lynden-Bell's result.
\section{Isothermal stellar systems}
\label{sec_vb}
\subsection{The equation of state}
\label{sec_mes}
We consider an isothermal stellar system described by the distribution function
\begin{eqnarray}
f=A e^{-\beta \epsilon}, \label{mes1}
\end{eqnarray}
where $\beta=1/T$ is a pseudo inverse temperature.
We justify here this distribution as a particular steady state of the
Vlasov equation\footnote{The statistical equilibrium state of a
self-gravitating system (resulting from a ``collisional'' relaxation)
is also described by an isothermal distribution of the form
(\ref{mes1}). In that case, $f=Ae^{-\beta m \epsilon}$ is the
Boltzmann distribution, $\beta=1/(k_B T)$ is the inverse thermodynamic
temperature and $S_B[f]=-k_B\int
\frac{f}{m}\ln \frac{f}{m} d{\bf r}d{\bf v}$ is the Boltzmann
entropy (Padmanabhan 1990).}. The associated pseudo entropy is
\begin{eqnarray}
S[f]=-\int f\ln (f/f_0) d{\bf r}d{\bf v},\label{mes2}
\end{eqnarray}
where $f_0$ is a constant introduced to make the term in the logarithm
dimensionless (it will play no role in the following since it appears
in an additional constant term). The distribution (\ref{mes1}) is
obtained by extremizing the pseudo entropy (\ref{mes2}) at fixed mass
and energy, writing $\delta S-\beta\delta E-\alpha\delta M=0$. For an
isothermal distribution, the kinetic temperature (velocity dispersion)
defined by Eq. (\ref{vhc2b}) is spatially uniform, and it coincides
with the pseudo temperature: $T({\bf r})=T$. The barotropic gas
corresponding to the isothermal stellar system defined by
Eq. (\ref{mes1}) is the isothermal gas with an equation of state
$p({\bf r})=\rho({\bf r}) T$. The velocity of sound is also spatially
uniform and it coincides with the velocity dispersion in one
direction: $c_{s}^{2}({\bf r})=T$. The density is related to the
gravitational potential by Eq. (\ref{ip2}). It can be obtained by
integrating Eq. (\ref{mes1}) on the velocities leading to
Eq. (\ref{ip2}) with $A'=A(2\pi/\beta)^{3/2}$. It can also be obtained
by using Eq. (\ref{ep8}) with $p({\bf r})=\rho({\bf r}) T$, or by
extremizing the functional (\ref{ip3}) at fixed mass (Chavanis 2006a).
Finally, combining Eqs (\ref{mes1}) and (\ref{ip2}), we can express
the distribution function in terms of the density profile according to
\begin{eqnarray}
f=\biggl ({\beta\over 2\pi}\biggr )^{3/2}\rho({\bf r}) \ e^{-\beta {v^{2}\over 2}}.
\label{mes4}
\end{eqnarray}
{\it Remark}: the stability of box-confined isothermal stellar systems has been studied by Antonov (1962), Lynden-Bell \& Wood (1968), Padmanabhan (1990) and Chavanis (2006a).
\subsection{The dispersion relation}
\label{sec_md}
Let us now consider an infinite homogeneous isothermal system described by the Maxwellian distribution function (\ref{mes4}) with uniform
density $\rho({\bf r})=\rho$. The reduced distribution (\ref{lsv3}) is
\begin{eqnarray}
f=\biggl ({\beta\over 2\pi}\biggr )^{1/2}\rho\ e^{-\beta {v^{2}\over 2}}.
\label{md1}
\end{eqnarray}
The Maxwellian distribution has a single maximum at $v=0$. Therefore,
the condition of marginal stability (\ref{cms2b}) implies $\omega_{r}=0$. From
Eq. (\ref{cms1b}), we find that the Maxwellian distribution is marginally stable
for $k=k_c$ where we have introduced the critical wavenumber
\begin{equation}
\label{md2}
k_{c}^2=\frac{4\pi G\rho}{T}.
\end{equation}
According to the criterion (\ref{sh2}), the Maxwell distribution is
linearly dynamically stable if $k>k_{c}$ and linearly dynamically
unstable if $k<k_{c}$. The critical Jeans wavenumber (\ref{md2}) for
an isothermal stellar system is the same as the critical Jeans
wavenumber (\ref{ip5}) for an isothermal gas. This is to be expected
on account of the {general} result of Sec. \ref{sec_fo}.
The dielectric function (\ref{lsv4}) associated to the Maxwellian distribution is
\begin{eqnarray}
\epsilon(k,\omega)=1-\frac{4\pi G}{k^2}\left (\frac{\beta}{2\pi}\right )^{1/2}\rho\int_{C}\frac{\beta v}{v-\frac{\omega}{k}}e^{-\beta {v^{2}\over 2}}\ dv.
\label{md3}
\end{eqnarray}
Introducing the critical
Jeans wavenumber (\ref{md2}), it can be rewritten
\begin{eqnarray}
\epsilon({k},\omega)=1-{k_{c}^{2}\over k^{2}} W\biggl ({\sqrt{\beta}\omega\over k}\biggr ),
\label{md4}
\end{eqnarray}
where
\begin{eqnarray}
W(z)={1\over\sqrt{2\pi}}\int_{C}{x\over x-z}e^{-{x^{2}\over 2}}dx,
\label{md5}
\end{eqnarray}
is the $W$-function of plasma physics (Ichimaru 1973). We note that $W(0)=1$. For any complex number $z$, we have the analytical expression
\begin{eqnarray}
W(z)=1-z e^{-{z^{2}\over 2}}\int_{0}^{z} e^{x^{2}\over 2}\, dx+i{\sqrt{\pi\over 2}}z e^{-{z^{2}\over 2}}.
\label{md6}
\end{eqnarray}
\subsection{Growth rate and damping rate}
\label{sec_mg}
We look for particular solutions of the dispersion relation $\epsilon({
k},\omega)=0$ in the form $\omega=i\omega_i$ where $\omega_i$ is
real. Using Eq. (\ref{md6}), we note that
\begin{eqnarray}
\epsilon({k},i\omega_i)=1-{k_{c}^{2}\over k^{2}}H\biggl (\frac{\sqrt{\beta}\omega_i}{k}\biggr ),
\label{mg1}
\end{eqnarray}
where we have introduced the function
\begin{equation}
\label{mg2} H(x)\equiv W(ix)= 1-\sqrt{\frac{\pi}{2}}x e^{\frac{x^{2}}{2}}{\rm erfc}\left (\frac{x}{\sqrt{2}}\right ).
\end{equation}
This function has the following asymptotic behaviors: (i) For
$x\rightarrow 0$, $H(x)=1-\sqrt{\frac{\pi}{2}}x+...$. (ii) For
$x\rightarrow +\infty$, $H(x)=\frac{1}{x^2}(1-{3\over
x^{2}}+...)$. (iii) For $x\rightarrow -\infty$, $H(x)\sim
-\sqrt{2\pi}x e^{\frac{x^{2}}{2}}$ (see Fig. \ref{h}). Using
$\epsilon(k,i\omega_i)=0$, the relation between $\omega_i$ and $k$
(for fixed $T$) can be written
\begin{eqnarray}
1-{k_{c}^{2}\over k^{2}}H\biggl ({\sqrt{\beta}\omega_i\over k}\biggr )=0,
\label{mg3}
\end{eqnarray}
or more explicitly\footnote{This relation is established by Binney \& Tremaine (1987) for $\omega_i\ge 0$. The present analysis shows that it is also valid for $\omega_i<0$.}
\begin{equation}
\label{mg6}1-{k_{c}^{2}\over k^{2}}\biggl \lbrace 1-\sqrt{\pi\beta \over 2}{\omega_i\over k}e^{{\beta\omega_i^{2}\over 2k^{2}}}{\rm erfc}\biggl(\sqrt{\frac{\beta}{2}} {\omega_i\over k}\biggr )\biggr\rbrace=0.
\end{equation}
This equation can also be obtained directly from Eqs. (\ref{lsv5}) and
(\ref{lsv7}) (see Appendix \ref{sec_wi}). If we set
$x=\sqrt{\beta}\omega_i/k$, we can rewrite Eq. (\ref{mg3}) in the
parametric form
\begin{eqnarray}
\frac{\omega_i}{\sqrt{4\pi G\rho}}=x\sqrt{H(x)}, \qquad \frac{k^2}{k_{c}^2}={H(x)}.
\label{mg4}
\end{eqnarray}
By varying $x$ between $-\infty$ and $+\infty$, we obtain the full
curve giving $\omega_i$ as a function of the wavenumber $k$ (see
Fig. \ref{pulsation}). Since the time dependence of the perturbation
is $\delta f\sim e^{\omega_i t}$, the case of neutral stability
$\omega_i=0$ corresponds to $k=k_{c}$, the case of instability
$\omega_i>0$ corresponds to $k<k_{c}$ and the case of stability
$\omega_i<0$ corresponds to $k>k_{c}$. We have the asymptotic
behaviors
\begin{eqnarray}
{\omega_i\over \sqrt{4\pi G\rho}}\sim 1-\frac{3}{2}{k^{2}\over k_{c}^{2}},\qquad (k/k_{c}\rightarrow 0),
\label{mg5}
\end{eqnarray}
\begin{eqnarray}
{\omega_i\over \sqrt{4\pi G\rho}}\sim \sqrt{\frac{2}{\pi}}\biggl (1-{k^{2}\over k_{c}^{2}}\biggr ), \qquad (k/k_c\rightarrow 1),
\label{mg6b}
\end{eqnarray}
\begin{eqnarray}
{\omega_i\over \sqrt{4\pi G\rho}}\sim -\sqrt{{2k^{2}\over k_{c}^{2}}\ln\biggl ({k^{2}\over k_{c}^{2}}\biggr )},\qquad (k/k_c\rightarrow +\infty).
\label{mg7}
\end{eqnarray}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{h.eps}
} \caption[]{The function $H(x)$.}
\label{h}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{pulsation.eps}
} \caption[]{Pulsation $\omega$ as a function of the wavenumber $k$
for an isothermal stellar system. For $k<k_J$, the system is unstable
and $\omega=i\omega_i$ with $\omega_i>0$. For $k>k_J$, the system is
stable. There exists many branches of solutions
$\omega=\omega_r+i\omega_i$ with $\omega_i<0$ (see Sec. \ref{sec_mn}) but we
have only represented the branch corresponding to
$\omega_r=0$. We have also compared these results with the case of an isothermal gas. For $k<k_J$, the system is unstable
and $\omega=\pm i\omega_i$. For $k>k_J$, the system is
stable and $\omega=\pm\omega_r$.}
\label{pulsation}
\end{figure}
Equation (\ref{mg4}) provides a particular solution of the dispersion
relation $\epsilon(k,\omega)=0$ of the form $\omega=i\omega_i$ with
$\omega_r=0$. The dispersion relation may have other solutions with
$\omega_r\neq 0$. However, for single humped distributions, we know
that there is only one unstable mode with $\omega_i>0$ for given
$k<k_c$ (see Sec. \ref{sec_sh}). Since the solutions $\omega=i\omega_i$ given by
Eq. (\ref{mg4}) exist for any $k<k_c$, we conclude that they are the
only solutions in that range of wavenumbers. Therefore, for the
unstable wavenumbers $k<k_c$, the perturbation grows exponentially
rapidly without oscillating. In other words, there are no overstable modes for the Maxwell distribution\footnote{Binney \& Tremaine (1987) show this result by a different method.}.
This is the same behavior as in a fluid
system (see Sec. \ref{sec_gas}) except that the growth rate $\omega_i>0$ in the
stellar system [see Eq. (\ref{mg4})] and in the fluid system [see
Eq. (\ref{jeans5})] are different\footnote{They only coincide for a
cold system $T=c_s^2=0$ (see Sec. \ref{sec_lsv}) or for $k=0$ (see
Sec. \ref{sec_mon}).}. For the stable wavenumbers $k>k_c$, the perturbations in a
stellar system are damped exponentially rapidly ($\omega_i<0$). We
have exhibited particular solutions (\ref{mg4}) that are damped
without oscillating $\omega_r=0$ but these are not the only solutions
of the dispersion relation. There also exists modes that are damped
exponentially ($\omega_i<0$) and oscillate $\omega_r\neq 0$ (see
asymptotic results in Sec. \ref{sec_mn}). This form of damping for collisionless
stellar systems is similar to the Landau damping for a
plasma\footnote{In plasma physics, for the Maxwellian distribution, there is no solution to the dispersion relation of the form $\omega=i\omega_i$. The pulsation
$\omega_r$ is non zero (see Sec. \ref{sec_pvp}).}. The situation is
very different in a fluid system. In that case, the stable modes with
wavenumbers $k>k_c$ correspond to gravity-modified sound waves that
propagate without attenuation ($\omega_r\neq 0$, $\omega_i=0$).
The pulsation of the perturbations in an infinite homogeneous
isothermal stellar system is plotted in Fig. \ref{pulsation} as a
function of the wavenumber $k$. For comparison, we have also indicated
the pulsation of the perturbations in an infinite homogeneous
isothermal gas.
\subsection{Other branches for $k\rightarrow +\infty$}
\label{sec_mn}
Let us solve the dispersion relation $\epsilon(k,\omega)=0$ for an
isothermal distribution in the limit $k\rightarrow
+\infty$\footnote{We here adapt the method of plasma physics developed by
Landau (1946), Jackson (1960) and Balescu (1963).}. We look for a solution of the equation
$\epsilon(k,\omega)=0$ of the form $\omega=\omega_r+i\omega_i$ with
$\omega_i<0$ (damping) and $\omega_i\gg\omega_r$. This corresponds to
heavily damped perturbations. We shall check this approximation a
posteriori. Using Eq. (\ref{lsv7}) for $\omega_{i}<0$ ,
Eq. (\ref{lsv4}) can be written
\begin{eqnarray}
1+\frac{4\pi G}{k^2}\int_{-\infty}^{+\infty} {{f'(v)}\over v-\frac{\omega}{k}}\, dv+\frac{8\pi^2 G}{k^2}i f'\left (\frac{\omega}{k}\right)=0.
\label{gh1}
\end{eqnarray}
If $f(v)$ decreases like $e^{-\beta v^2/2}$ for real $v\rightarrow \pm\infty$, then for complex $\omega=\omega_r+i\omega_i$ with $\omega_i\gg \omega_r$, $f'(\omega/k)$ will increase like $e^{\beta\omega_i^2/2k^2}$. Therefore, to leading order, the foregoing equation reduces to
\begin{eqnarray}
1+\frac{8\pi^2 G}{k^2}i f'\left (\frac{\omega}{k}\right)=0.
\label{gh2}
\end{eqnarray}
Separating real and imaginary parts, we obtain two transcendant equations
\begin{eqnarray}
{\rm Re}\left\lbrack\frac{8\pi^2 G}{k^2}i f'\left (\frac{\omega_r+i\omega_i}{k}\right)\right\rbrack=-1,
\label{gh3}
\end{eqnarray}
\begin{eqnarray}
{\rm Im}\left\lbrack i f'\left (\frac{\omega_r+i\omega_i}{k}\right)\right\rbrack=0,
\label{gh4}
\end{eqnarray}
which crucially depend on the form of the distribution. For the Maxwellian (\ref{md1}), they can be rewritten to leading order in the limit $\omega_i/\omega_r\gg 1$ as
\begin{eqnarray}
\frac{8\pi^2 G}{k^3}\left (\frac{\beta}{2\pi}\right )^{1/2}\rho\beta\omega_i e^{\frac{\beta\omega_{i}^{2}}{2k^2}}\cos\left (\frac{\beta\omega_r\omega_i}{k^2}\right )=-1,
\label{gh5}
\end{eqnarray}
\begin{eqnarray}
\sin\left (\frac{\beta\omega_r\omega_i}{k^2}\right )=0.
\label{gh6}
\end{eqnarray}
Equation (\ref{gh6}) implies $\beta \omega_r\omega_i/k^2=m\pi$. Eq. (\ref{gh5}) will have a solution provided that $m$ is even. Then, Eq. (\ref{gh5}) gives
\begin{eqnarray}
\frac{8\pi^2 G}{k^3}\left (\frac{\beta}{2\pi}\right )^{1/2}\rho\beta\omega_i e^{\frac{\beta\omega_{i}^{2}}{2k^2}}=-1,
\label{gh7}
\end{eqnarray}
which determines $\omega_i$. By a graphical construction, it is easy to see that $|\omega_i|$ is an increasing function of $k$. For $k\rightarrow +\infty$, we find the asymptotic behaviors
\begin{eqnarray}
\omega_i=-\frac{2}{\sqrt{\beta}}k\sqrt{\ln k}, \qquad \omega_r=-m\frac{\pi}{2}\frac{1}{\sqrt{\beta}}\frac{k}{\sqrt{\ln k}}.
\label{gh8}
\end{eqnarray}
Since $\omega_{i}/\omega_{r}\sim \ln k\rightarrow +\infty$, our basic assumption is satisfied. Therefore, for $k>k_c$ we have several branches of solutions parameterized by the even integer $m$. For $m=0$, we recover the results of Sec. \ref{sec_mg}.
{\it Remark}: by analogy with plasma physics, we could also look for
solutions of the dispersion relation (\ref{lsv4}) of the form
$\omega=\omega_r+i\omega_i$ with $\omega_i\ll \omega_r$. This
corresponds to weakly damped perturbations. In plasma physics, these
solutions are valid for $k\rightarrow 0$ and lead to the usual Landau
damping formula. However, a self-gravitating system is unstable for
$k<k_c$. Furthermore, it is easy to show that there is no solution of
that form to Eq. (\ref{lsv4}) whatever the form of the distribution
$f(v)$ and the wavenumber $k$. This implies that for attractive
interactions (like gravity) the perturbations are unstable for $k<k_c$
and heavily damped for $k>k_c$ while for repulsive interaction (like
plasmas) they are weakly damped for $k<k_D$ and heavily damped for
$k>k_D$.
\subsection{Nyquist curve}
\label{sec_mnb}
It will be convenient in the following to work with dimensionless quantities.
We introduce the dimensionless wavenumber and the dimensionless pulsation
\begin{eqnarray}
\eta=\frac{4\pi G\rho}{T k^2}, \qquad \Omega=\frac{\omega}{\sqrt{4\pi G\rho}},
\label{mn1}
\end{eqnarray}
Noting that $\sqrt{\beta}\omega/k=\sqrt{\eta}\Omega$, the dielectric function (\ref{md4}) can be rewritten
\begin{eqnarray}
\epsilon({\eta},\Omega)=1-\eta W(\sqrt{\eta}\Omega).
\label{mn3}
\end{eqnarray}
When $\Omega_{i}=0$, the real and imaginary parts of the dielectric function $\epsilon(\eta,\Omega_{r})=\epsilon_{r}(\eta,\Omega_{r})+i\epsilon_{i}(\eta,\Omega_{r})$ can be written
\begin{equation}
\label{mn4}
\epsilon_{r}(\eta,\Omega_{r})=1-\eta W_{r}\left (\sqrt{\eta}\Omega_r\right ),
\end{equation}
\begin{equation}
\label{mn5}
\epsilon_{i}(\eta,\Omega_{r})=-\eta W_{i}\left (\sqrt{\eta}\Omega_r\right ),
\end{equation}
with
\begin{equation}
\label{mn6}
W_{r}(z)=1-z e^{-\frac{z^2}{2}}\int_{0}^{z}e^{\frac{x^2}{2}}\, dx,
\end{equation}
\begin{equation}
\label{mn7}
W_{i}(z)=\sqrt{\frac{\pi}{2}}z e^{-\frac{z^2}{2}},
\end{equation}
where $z$ is here a real number. The condition of marginal stability corresponds to $\epsilon_{r}(\eta,\Omega_{r})=\epsilon_{i}(\eta,\Omega_{r})=0$. The condition $\epsilon_{i}(\eta,\Omega_{r})=0$, which is equivalent to $f'(\sqrt{\eta}\Omega_{r})=0$, implies $\Omega_{r}=0$. Then, the condition $\epsilon_{r}(\eta,\Omega_{r})=0$ leads to $\eta=\eta_{c}$ with
\begin{equation}
\label{mn8}
\eta_c=1.
\end{equation}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{maxwell.eps}
} \caption[]{Nyquist curve for the Maxwellian distribution
(\ref{md1}). The DF is stable for perturbations with $\eta<\eta_c$
($k>k_c$), marginally stable for perturbations with $\eta=\eta_c$
($k=k_c$) and unstable for perturbations with $\eta>\eta_c$
($k<k_c$). We have taken $\eta=2,1,0.5$ from the outer to the inner
curve.}
\label{maxwell}
\end{figure}
To apply the Nyquist method, we need to plot the curve
$(\epsilon_{r}(\eta,\Omega_{r}),\epsilon_{i}(\eta,\Omega_{r}))$ in the
$\epsilon$-plane. For $\Omega_{r}\rightarrow \pm \infty$, this curve
tends to the point $(1,0)$ in the manner described in
Sec. \ref{sec_sh}. On the other hand, for $\Omega_r=0$, it crosses the
$x$-axis at $(\epsilon_r(\eta,0)=1-\eta,0)$. The Nyquist curve is
represented in Fig. \ref{maxwell} for several values of the wavenumber
$\eta$. For $\eta<1$ (i.e. $k>k_c$), the Nyquist curve does not
encircle the origin so that the Maxwellian distribution is stable. For
$\eta>1$ (i.e. $k<k_c$) the Nyquist curve encircles the origin so that
the Maxwellian distribution is unstable. For $\eta=1$ (i.e. $k=k_c$)
the Nyquist curve passes through the origin so that the Maxwellian
distribution is marginally stable. The Nyquist method provides a nice
graphical illustration of the Jeans instability criterion for an
infinite homogeneous stellar system.
\section{Stellar polytropes}
\label{sec_vb2}
\subsection{The equation of state}
We consider a stellar polytrope (or polytropic galaxy) described by the distribution function
\begin{eqnarray}
f=\biggl \lbrack \mu-{\beta (q-1)\over q}\epsilon\biggr \rbrack_+^{1\over q-1},
\label{vb1}
\end{eqnarray}
where $\beta=1/T$ is a pseudo inverse temperature
and $q$ is a parameter related to the
traditional polytropic index $n$ by
\begin{eqnarray}
n={3\over 2}+{1\over q-1}.
\label{vb2}
\end{eqnarray}
This relation is plotted in Fig. \ref{nq} and specific values considered
in the sequel are highlighted. We justify here the polytropic distribution function (\ref{vb1}) as a particular steady state of the Vlasov equation\footnote{Some authors (Plastino \& Plastino (1993), Lima et al. (2002), Silva \& Alcaniz (2004), Lima \& de Souza (2005), Taruya \& Sakagami (2003a), Leubner (2005), Kronberger et al. (2006), Du Julin (2006)) have interpreted the polytropic distribution (\ref{vb1}) and the functional (\ref{vb3}) in terms of Tsallis (1988) generalized thermodynamics. Here, we use a more conventional approach (Ipser 1974, Ipser \& Horwitz 1979, Binney \& Tremaine 1987, Chavanis 2006a). We interpret the polytropic distribution (\ref{vb1}) as a particular steady state of the Vlasov equation and the functional (\ref{vb3}) as a pseudo entropy. Its maximization at fixed mass and energy provides a condition of nonlinear dynamical stability with respect to the Vlasov equation (see Sec. \ref{sec_vh}), not a condition of ``generalized thermodynamical stability''. In particular, as argued by Chavanis \& Sire (2005) and Campa et al. (2008), the instabilities reported by Taruya \& Sakagami (2003b) correspond to Vlasov dynamical instabilities, not ``generalized thermodynamical instabilities''. In the present context, the analogies with Tsallis thermodynamics are purely coincidental. They are the mark of a {\it thermodynamical analogy} (Chavanis 2006a). Tsallis generalized thermodynamics applies in different contexts (see Chavanis 2008a).}. The associated pseudo-entropy is
\begin{eqnarray}
S=-{1\over q-1}\int (f^{q}-f_0^{q-1}f)d{\bf r}d{\bf v},
\label{vb3}
\end{eqnarray}
where $f_0$ is a constant introduced for reasons of homogeneity (it will play no role in the following since it appears in a term proportional to the mass that is conserved). The DF (\ref{vb1}) is obtained by
extremizing the pseudo entropy (\ref{vb3}) at fixed mass and energy, writing
$\delta S-\beta\delta E-\alpha\delta M=0$. The condition that $C(f)$ must be convex imposes $q>0$. On the other hand, we shall assume that $f(\epsilon)$ is a decreasing function of the energy so that $\beta>0$. It is important to note that $1/\beta$ does {\it not} represent the kinetic temperature (or velocity dispersion) of the polytropic distribution (see Sec. \ref{sec_other}).
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{qn.eps}
} \caption[]{The relation between the polytropic index $n$ and the
parameter $q$. Some particular values $(q,n)$ are indicated for
reference.} \label{nq}
\end{figure}
We need to distinguish two cases depending on the sign of $q-1$. For $q>1$ ($n>3/2$), the
distribution function can be written
\begin{eqnarray}
f=A(\epsilon_{m}-\epsilon)_+^{1\over q-1},
\label{vb4}
\end{eqnarray}
where we have set $A=\lbrack\beta(q-1)/q\rbrack^{1\over q-1}$ and
$\epsilon_{m}=q\mu/\lbrack \beta(q-1)\rbrack$. Such distributions have a compact support since they
vanish at $\epsilon=\epsilon_m$. For $\epsilon>\epsilon_m$, we set $f=0$. Therefore, the notation $[x]_+=x$ for
$x>0$ and $[x]_+=0$ for $x<0$. At a given position, the distribution function vanishes for $v\ge v_{m}({\bf
r})=\sqrt{2(\epsilon_{m}-\Phi({\bf r}))}$. For $q\rightarrow 1$
($n\rightarrow +\infty$), we recover the isothermal distribution (\ref{mes1}) and for
$n=3/2$ the distribution function is a step function (see
Sec. \ref{sec_fermi}). This is the distribution function of a Fermi
gas at zero temperature describing classical white dwarf stars
(Chandrasekhar 1942). For $0<q<1$, the distribution function can be written
\begin{eqnarray}
f=A(\epsilon_{0}+\epsilon)^{1\over q-1},
\label{vb5}
\end{eqnarray}
where we have set $A=\lbrack\beta(1-q)/q\rbrack^{1\over q-1}$ and
$\epsilon_{0}=q\mu/\lbrack\beta(1-q)\rbrack$. Such distributions are defined for all velocities. At a given position, the distribution function behaves, for large velocities, as
$f\sim v^{2/(q-1)}\sim v^{2n-3}$. We shall only consider distribution functions for which the
density $\rho=\int f\, d{\bf v}$ and the pressure $p=\frac{1}{3}\int f v^2\, d{\bf v}$ are
finite. This implies
$3/5<q<1$ ($n<-1$)\footnote{If we allow $\beta$ to be negative, then it is possible to construct stellar polytropes with index $1/2<n<3/2$ which are mathematically well-behaved (see Binney \& Tremaine 1987). However, for those polytropes, the distribution function increases with the energy (and diverges at $\epsilon=\epsilon_m$) so they may not be physical.}.
Let us now determine the equation of state of the barotropic gas corresponding to a stellar polytrope.
For $n>3/2$, the density and
the pressure can be expressed as
\begin{eqnarray}
\rho=4\pi\sqrt{2}A(\epsilon_{m}-\Phi)^{n}{\Gamma(3/2)\Gamma(n-1/2)\over \Gamma(n+1)},
\label{vb6}
\end{eqnarray}
\begin{eqnarray}
p={1\over n+1}4\pi\sqrt{2}A(\epsilon_{m}-\Phi)^{n+1}{\Gamma(3/2)\Gamma(n-1/2)\over \Gamma(n+1)},
\label{vb7}
\end{eqnarray}
where $\Gamma(x)$ denotes the Gamma function.
For $n<-1$, the density and the pressure can be expressed as
\begin{eqnarray}
\rho=4\pi\sqrt{2}A(\epsilon_{0}+\Phi)^{n}{\Gamma(-n)\Gamma(3/2)\over \Gamma(3/2-n)},
\label{vb8}
\end{eqnarray}
\begin{eqnarray}
p=-{1\over n+1}4\pi\sqrt{2}A(\epsilon_{0}+\Phi)^{n+1}{\Gamma(-n)\Gamma(3/2)\over \Gamma(3/2-n)}.
\label{vb9}
\end{eqnarray}
Eliminating the gravitational potential between the expressions (\ref{vb6})-(\ref{vb7}) and (\ref{vb8})-(\ref{vb9}), one finds that
\begin{eqnarray}
p=K\rho^{\gamma}, \qquad \gamma=1+{1\over n},
\label{vb10}
\end{eqnarray}
where
\begin{eqnarray}
K={1\over n+1}\biggl \lbrace 4\pi\sqrt{2}A {\Gamma(3/2)\Gamma(n-1/2)\over \Gamma(n+1)}
\biggr \rbrace^{-{1\over n}} (n>3/2)\qquad
\label{vb11}
\end{eqnarray}
\begin{eqnarray}
K=-{1\over n+1}\biggl \lbrace 4\pi\sqrt{2}A{\Gamma(-n)\Gamma(3/2)\over\Gamma(3/2-n)}\biggr \rbrace^{-{1\over n}} (n<-1).
\label{vb12}
\end{eqnarray}
Therefore, a stellar polytrope has the same equation of state
(\ref{vb10}) as a polytropic star. However, they do not have the same
distribution function (compare Eq. (\ref{vb1}) to Eq.
(\ref{ep4})) except for $n\rightarrow \infty$ corresponding to the isothermal case.
The density is related to the gravitational potential by Eq. (\ref{ip7}).
It can be obtained by integrating Eq. (\ref{vb1}) on the velocity leading to Eqs. (\ref{vb6}) and (\ref{vb8}) from which, using Eqs. (\ref{vb11}) and (\ref{vb12}), we deduce Eq. (\ref{ip7}) with $\lambda=\epsilon_m/(K(n+1))$ for $n\ge 3/2$ and $\lambda=\epsilon_0/(-K(n+1))$ for $n<-1$. It can also be obtained by using Eq. (\ref{ep8}) with Eq. (\ref{vb10}) or by extremizing the functional (\ref{ip8}) at fixed mass (Chavanis 2006a). Note that Eqs. (\ref{ip7}) and (\ref{ip8}) are similar to
Eqs. (\ref{vb1}) and ({\ref{vb3}) with $\gamma$ playing the role of the
parameter $q$ and $K$ playing the role of the pseudo temperature $T=1/\beta$.
Polytropic distributions (including the isothermal one) are apparently the only
distributions for which $f(\epsilon)$ and $\rho(\Phi)$ have a similar mathematical form.
{\it Remark}: isolated stellar polytropes have a finite mass iff $3/2\le n\le 5$ and they are dynamically stable (Binney \& Tremaine 1987). The stability of box-confined polytropes is studied by Taruya \& Sakagami (2003a) in the context of generalized thermodynamics and by Chavanis (2006a) in the context of Vlasov dynamical stability.
\subsection{Other expressions of the distribution function}
\label{sec_other}
We can write the distribution function of stellar polytropes
(\ref{vb1}) in different forms that all have their own
interest. This will show that different notions of ``temperature'' exist for
polytropic distributions\footnote{We recall that, for collisionless stellar systems, the mass of the stars does not appear in the Vlasov equation and the different ``temperatures'' that we introduce have the dimension of a velocity squared.}.
(i) Pseudo temperature $T=1/\beta$: as indicated previously, the form (\ref{vb1}) of the polytropic distribution directly comes from the variational principle (\ref{vh3}) when we write the pseudo entropy in the form (\ref{vb3}). Therefore, $\beta=1/T$ is the Lagrange multiplier associated with the conservation of energy. It is called ``pseudo inverse temperature''. Note, however, that $T=1/\beta$ does not have the dimension of a temperature (squared velocity).
(ii) Dimensional temperature $\theta=1/b$: we can define a quantity that has the dimension of a temperature (squared velocity) by setting $b=\beta/q\mu$. If we define furthermore $f_*=\mu^{1/(q-1)}$, the polytropic distribution (\ref{vb1}) can be rewritten
\begin{eqnarray}
f=f_*\biggl \lbrack 1-{b (q-1)}\epsilon\biggr \rbrack_+^{1\over q-1}.
\label{na1}
\end{eqnarray}
Using Eqs. (\ref{vb6}) and (\ref{vb8}), the relation between the density and the gravitational potential can be written
\begin{eqnarray}
\rho=\rho_{*}\biggl \lbrack 1-{b (q-1)}\Phi\biggr \rbrack^{n},
\label{na2}
\end{eqnarray}
with
\begin{eqnarray}
\rho_*=2\pi f_{*}\left (\frac{2n-3}{b}\right )^{3/2}\frac{\Gamma(3/2)\Gamma(n-1/2)}{\Gamma(n+1)} \ (n>3/2),\nonumber\\
\label{na3}
\end{eqnarray}
\begin{eqnarray}
\rho_*=2\pi f_{*}\left (\frac{3-2n}{b}\right )^{3/2}\frac{\Gamma(3/2)\Gamma(-n)}{\Gamma(3/2-n)} \ (n<-1).
\label{na4}
\end{eqnarray}
(iii) Polytropic temperature $K$: eliminating the gravitational potential between Eqs. (\ref{vb4}) and (\ref{vb6}), and between Eqs. (\ref{vb5}) and (\ref{vb8}), we can express the distribution function in terms of the density according to
\begin{eqnarray}
f={1\over Z}\biggl \lbrack \rho({\bf r})^{1/n}-{v^{2}/2\over (n+1)K}\biggr\rbrack_+^{n-3/2},
\label{vb19}
\end{eqnarray}
\begin{eqnarray}
Z=4\pi\sqrt{2}{\Gamma(3/2)\Gamma(n-1/2)\over\Gamma(n+1)}\lbrack K(n+1)\rbrack^{3/2} \ (n>3/2)
\label{vb20}
\end{eqnarray}
\begin{eqnarray}
Z=4\pi\sqrt{2}{\Gamma(-n)\Gamma(3/2)\over\Gamma(3/2-n)}\lbrack -K(n+1)\rbrack^{3/2}\ (n<-1).
\label{vb21}
\end{eqnarray}
This is the polytropic counterpart of expression (\ref{mes4}) for the
isothermal distribution. The constant $K$ plays a role similar to the temperature $T$ in an isothermal distribution. In particular, it is uniform in a polytropic distribution as is the temperature in an isothermal system. For that reason, it is sometimes called a polytropic temperature.
(iv) Kinetic temperature $T({\bf r})$: for a polytropic distribution, the kinetic temperature (velocity dispersion) defined by Eq. (\ref{vhc2b}) is given by
\begin{eqnarray}
T({\bf r})=K\rho({\bf r})^{\gamma-1}.
\label{vb21bis}
\end{eqnarray}
For an inhomogeneous stellar polytrope, the kinetic temperature $T({\bf r})$ is position dependent and differs from the pseudo temperature $T=1/\beta$. The velocity of sound is given by
\begin{eqnarray}
c_s^2({\bf r})=K\gamma\rho({\bf r})^{\gamma-1}=\gamma T({\bf r}).
\label{vb21tris}
\end{eqnarray}
It is also position dependent and differs from the velocity dispersion (they differ by a factor $\gamma$). Using Eq. (\ref{vb21bis}), the distribution function
(\ref{vb19}) can be written
\begin{eqnarray}
f=B_{n}{\rho({\bf r})\over \lbrack 2\pi T({\bf r})\rbrack^{3/2}}\biggl\lbrack 1-{v^{2}/2\over (n+1)T({\bf r})}\biggr\rbrack_+^{n-3/2},
\label{vb23}
\end{eqnarray}
\begin{eqnarray}
B_{n}={\Gamma(n+1)\over\Gamma(n-1/2)(n+1)^{3/2}}, \quad (n>3/2)
\label{vb24}
\end{eqnarray}
\begin{eqnarray}
B_{n}={\Gamma(3/2-n)\over\Gamma(-n)\lbrack -(n+1)\rbrack^{3/2}} \quad (n<-1).
\label{vb25}
\end{eqnarray}
Note that for $n>3/2$, the maximum velocity can be expressed in terms of the kinetic temperature by
\begin{eqnarray}
v_{m}({\bf r})=\sqrt{2(n+1)T({\bf r})}.
\label{vb26}
\end{eqnarray}
Using $\Gamma(z+a)/\Gamma(z)\sim z^{a}$ for $z\rightarrow +\infty$, we recover the isothermal distribution (\ref{mes1}) for $n\rightarrow +\infty$.
On the other hand, from Eqs. (\ref{vb21bis}) and (\ref{ip7}), we
immediately get $T({\bf r})=K(\lambda-(\gamma-1)\Phi({\bf
r})/K\gamma)$ so that
\begin{eqnarray}
\nabla T=-{\gamma-1\over\gamma}\nabla\Phi.
\label{vb22}
\end{eqnarray}
This shows that, for a stellar
polytrope, the kinetic temperature (velocity dispersion) is a linear
function of the gravitational potential\footnote{For any spherical stellar system with $f=f(\epsilon)$, we have $\rho=\rho(\Phi)$ and $p=p(\Phi)$ so that the kinetic temperature (velocity dispersion) $T=p/\rho$ is a function $T=T(\Phi)$ of the gravitational potential. For a polytropic distribution function, this relation turns out to be linear.}. The coefficient of
proportionality is related
to the polytropic index by $(\gamma-1)/\gamma=1/(n+1)=2(q-1)/(5q-3)$.
This relation can also be obtained directly from Eq. (\ref{vb4}) [or Eq. (\ref{vb5})] noting that
\begin{eqnarray}
f=A(\epsilon_{m}-\Phi)^{n-3/2}\left \lbrack 1-\frac{v^{2}/2}{\epsilon_{m}-\Phi}\right\rbrack_+^{n-3/2},
\label{vb22ndl}
\end{eqnarray}
and comparing with Eq. (\ref{vb23}).
\subsection{The dispersion relation}
\label{sec_pdr}
Let us now consider an infinite homogeneous polytropic stellar system described by the polytropic distribution function (\ref{vb23}) with uniform density $\rho({\bf r})=\rho$ and uniform kinetic temperature $T({\bf r})=T$. From now on, $T=1/\beta$ will denote the kinetic temperature (\ref{vhc2b}), not the Lagrange multiplier appearing in Eq. (\ref{vb1}). The kinetic temperature is uniform because we have assumed that the density is uniform. The reduced distribution function (\ref{lsv3}) is\footnote{If we justify the distribution function (\ref{pdr1}) from the 3D distribution function (\ref{vb23}) integrated on $v_x$ and $v_y$, then it is valid for $n>3/2$ and $n<-1$. However, as far as mathematics is concerned, the distribution (\ref{pdr1}) is normalizable and has a finite variance in the range of indices $n\ge 1/2$ and $n<-1$.}
\begin{equation}
\label{pdr1}
f(v)=B_n \frac{\rho}{\sqrt{2\pi T}}\left [1-\frac{v^2}{2(n+1)T} \right ]_+^{n-1/2},
\end{equation}
where $\rho$ is the density, $T=1/\beta=\langle v^2\rangle$ is the velocity dispersion in one direction and $B_n$ is a normalization constant given by
\begin{eqnarray}
\label{pdr2}
B_n&=&\frac{\Gamma(n+1)}{\Gamma(n+1/2)(n+1)^{1/2}}, \qquad n>{1\over 2},\\
B_n&=&\frac{\Gamma(1/2-n)}{\Gamma(-n)[-(n+1)]^{1/2}}, \qquad n<{-1}.
\end{eqnarray}
The polytropic distribution has a single maximum at $v=0$. Therefore,
the condition of marginal stability (\ref{cms2b}) implies $\omega_{r}=0$. From
Eq. (\ref{cms1b}), we find that the polytropic distribution is marginally stable
for $k=k_c$ where we have introduced the critical wavenumber
\begin{equation}
\label{pdr3}
k_{c}^2=\frac{4\pi G\rho}{\gamma T}.
\end{equation}
According to the criterion (\ref{sh2}), the polytropic distribution is
linearly dynamically stable if $k>k_{c}$ and linearly dynamically
unstable if $k<k_{c}$. The critical Jeans wavenumber (\ref{pdr3}) for
a stellar polytrope is the same as the critical Jeans wavenumber
(\ref{ip13}) for a polytropic gas. This is to be expected on account
of the {general} result of Sec. \ref{sec_fo}. It should be stressed
that the quantity that appears in the critical wavenumber (\ref{pdr3})
is the velocity of sound $c_s^2=\gamma T$ in the corresponding
barotropic gas, not the velocity dispersion $T$. They coincide for the
Maxwellian distribution ($\gamma=1$), but this is not general.
The dielectric function (\ref{lsv4}) associated to the polytropic distribution is
\begin{eqnarray}
\label{pdr4}
\epsilon(k,\omega)=1- \frac{4\pi G}{k^2}\frac{\rho}{\sqrt{2 \pi T}} B_n \left (n-\frac{1}{2}\right )\frac{1}{(n+1)T} \nonumber\\
\times \int_{C} \frac{v\left [1-\frac{v^2}{2(n+1)T} \right ]_+^{n-3/2}}{v-\frac{\omega}{k}}\, dv.
\end{eqnarray}
Introducing the critical wavenumber (\ref{pdr3}), we obtain
\begin{eqnarray}
\epsilon({k},\omega)=1-{k_{c}^{2}\over k^{2}}W_{n}\biggl ({\omega\over k\sqrt{T}}\biggr ),
\label{pdr5}
\end{eqnarray}
where
\begin{eqnarray}
\label{pdr6}
W_{n}(z)= \frac{1}{\sqrt{2\pi}}\frac{B_n}{n}\left (n-\frac{1}{2}\right )\int_{C}\frac{x[1-\frac{x^2}{2(n+1)}]_+^{n-3/2}}{x-z}dx,\nonumber\\
\end{eqnarray}
is a generalization of the $W$-function of plasma physics. We note that $W_{n}(0)=1$. For $n\rightarrow +\infty$, we recover the
$W$-function (\ref{md5}). Equation (\ref{pdr6}) is therefore a
generalization of this function to the case of polytropic
distributions.
\subsection{Growth rate and damping rate}
\label{sec_pgd}
We look for particular solutions of the dispersion relation $\epsilon({
k},\omega)=0$ in the form $\omega=i\omega_i$ where $\omega_i$ is
real. First, we note that
\begin{eqnarray}
\epsilon({k},i\omega_i)=1-{k_{c}^{2}\over k^{2}}H_{n}\biggl (\frac{\omega_i}{k\sqrt{T}}\biggr ),
\label{pgd1}
\end{eqnarray}
where we have introduced the function $H_{n}(x)\equiv W_{n}(ix)$. For $x>0$, we have
\begin{eqnarray}
\label{pgd2}
H_{n}(x)= \frac{1}{\sqrt{2\pi}}\frac{B_n}{n}\left (n-\frac{1}{2}\right )\int_{-\infty}^{+\infty}\frac{t^2[1-\frac{t^2}{2(n+1)}]_+^{n-3/2}}{t^2+x^2}dt,\nonumber\\
\end{eqnarray}
and for $x<0$, we have
\begin{eqnarray}
\label{pgd3}
H_{n}(x)= \frac{1}{\sqrt{2\pi}}\frac{B_n}{n}\left (n-\frac{1}{2}\right )\biggl\lbrace \int_{-\infty}^{+\infty}\frac{t^2[1-\frac{t^2}{2(n+1)}]_+^{n-3/2}}{t^2+x^2}dt\nonumber\\
-2\pi x \left\lbrack1+\frac{x^2}{2(n+1)}\right\rbrack^{n-3/2}\biggr\rbrace.\qquad\qquad\qquad
\end{eqnarray}
Using $\epsilon(k,i\omega_i)=0$, the relation between $\omega_i$ and $k$ (for
fixed $T$) can be written
\begin{eqnarray}
1-{k_{c}^{2}\over k^{2}}H_{n}\biggl ({\omega_i\over k\sqrt{T}}\biggr )=0.
\label{pgd4}
\end{eqnarray}
For $n\rightarrow +\infty$, we recover Eq. (\ref{mg3}). If we set $x=\sqrt{\beta}\omega_i/k$, we
can rewrite Eq. (\ref{pgd4}) in the parametric form
\begin{eqnarray}
\frac{\omega_i}{\sqrt{4\pi G\rho}}=x\sqrt{\frac{H_{n}(x)}{\gamma}}, \qquad \frac{k^2}{k_{c}^2}={H_{n}(x)}.
\label{pgd6}
\end{eqnarray}
By varying $x$ between $-\infty$ and $+\infty$, we obtain the full curve giving $\omega_i$ as a function of the wavenumber $k$. Since the time dependence of the perturbation is $\delta f\sim e^{\omega_i
t}$, the case of neutral stability $\omega_i=0$ corresponds to $k=k_{c}$, the
case of instability $\omega_i>0$ corresponds to $k<k_{c}$ and the case of stability $\omega_i<0$ corresponds to $k>k_{c}$. The discussion of the different regimes is similar to the one given in Sec. \ref{sec_mg}. For $k\rightarrow k_c$, using the general approximate formula (\ref{mon15}), the pulsation $\omega_i$ is given by
\begin{eqnarray}
\frac{\omega_i}{\sqrt{4\pi G\rho}}=\sqrt{\frac{2}{\pi}}\left (\frac{n}{1+n}\right )^{1/2}\frac{n}{B_n}\frac{1}{n-\frac{1}{2}}\left (1-\frac{k^2}{k_c^2}\right ).
\label{kkn}
\end{eqnarray}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{gr.eps}
} \caption[]{Pulsation $\omega$ as a function of the wavenumber $k$
for a stellar polytrope with $n\ge 1/2$ (we have represented $n=0.5,1,1.5,2,3,5,10,100$).
For $k<k_c$, the system is unstable
and $\omega=i\omega_i$ with $\omega_i>0$. For $k>k_c$, the system is
stable. There exists many branches of solutions
$\omega=\omega_r+i\omega_i$ with $\omega_i<0$ but we
have only represented the branch corresponding to
$\omega_r=0$. We have also compared these results to the case of a polytropic gas. For $k<k_c$, the system is unstable
and $\omega=\pm i\omega_i$. For $k>k_c$, the system is
stable and $\omega=\pm\omega_r$. The instability occurs for $k<k_{c}^{(poly)}=k_{c}^{(iso)}/\sqrt{\gamma}$ where $\gamma=1+1/n$. For $n\ge 1/2$, $k_{c}^{(poly)}\le k_{c}^{(iso)}$. }
\label{tsallispos}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{disneg.eps}
} \caption[]{Pulsation $\omega$ as a function of the wavenumber $k$
for a stellar polytrope and a polytropic gas with $n\le -1$ (we have represented $n=-1.5,-2,-3,-5,-10,-100$). The instability occurs for $k<k_{c}^{(poly)}=k_{c}^{(iso)}/\sqrt{\gamma}$ where $\gamma=1+1/n$. For $n\le -1$, $k_{c}^{(poly)}\ge k_{c}^{(iso)}$.}
\label{tsallisneg}
\end{figure}
The pulsation of the perturbation in an infinite homogeneous stellar polytrope is plotted in Figs. \ref{tsallispos} and \ref{tsallisneg} as a function of the wavenumber $k$. For comparison, we have also indicated the pulsation of the perturbation in an infinite homogeneous polytropic gas. For $n=\infty$, we recover the isothermal case shown in Fig. \ref{pulsation}.
\subsection{Particular cases}
\label{sec_fermi}
The case $n=3/2$ deserves a particular attention. In that case, the distribution function $f({\bf r},{\bf v})$ is a step function so that $f=\eta_{0}$ for $v\le v_{m}({\bf r})\equiv \sqrt{2(\epsilon_{m}-\Phi({\bf r}))}$ and $f=0$ otherwise. This corresponds to the distribution function of the self-gravitating Fermi gas at $T=0$ which describes classical white dwarf stars (Chandrasekhar 1942). The density is $\rho={4\pi\over 3}\eta_{0}v_{m}^{3}$ and the pressure $p={4\pi\over 15}\eta_{0}v_{m}^{5}$. Eliminating $v_{m}$ from these two relations, we get a polytropic equation of state $p=K\rho^{5/3}$ with $n=3/2$, $\gamma=5/3$ and $K={1\over 5}({3\over 4\pi\eta_{0}})^{2/3}$. The kinetic temperature is $T=\frac{1}{5}v_m^2$ and the velocity of sound is $c_{s}^{2}={1\over 3}v_{m}^{2}$. For an infinite and homogeneous medium, the reduced distribution function (\ref{pdr1}) corresponding to $n=3/2$ is a parabola
\begin{eqnarray}
f(v)=\frac{3\rho}{4\sqrt{5T}}\left (1-\frac{v^2}{5T}\right ).
\label{par}
\end{eqnarray}
The critical Jeans wavenumber (\ref{sh3}) is $k_{c}^{2}=12\pi G\rho/5T=12\pi G\rho/v_{m}^{2}$ which is fully consistent with the expression (\ref{pdr3}) with $\gamma=1+1/n=5/3$. For $n=3/2$, we can obtain an explicit expression of the function (\ref{pgd2})-(\ref{pgd3}). For $x>0$, we have
\begin{eqnarray}
H_{3/2}(x)=1-\frac{x}{\sqrt{5}}\arctan\left (\frac{\sqrt{5}}{x}\right ),
\label{fermi2}
\end{eqnarray}
and for $x<0$, we have
\begin{eqnarray}
H_{3/2}(x)=1-\frac{x}{\sqrt{5}}\left\lbrack \arctan\left (\frac{\sqrt{5}}{x}\right )+\pi\right\rbrack.
\label{fermi3}
\end{eqnarray}
The index $n=1/2$ is also special and corresponds to the water-bag model. In that case, the reduced distribution $f(v)$ is a step function so that $f=\eta_{0}$ for $|v|\le v_{m}$ and $f=0$ otherwise. The amplitude $\eta_0$ is determined by the density according to the relation $\rho=2\eta_{0}v_{m}$. The kinetic temperature is $T=\langle v^2\rangle=\frac{1}{3}v_m^2$.
The derivative of the distribution function is $f'(v)=\eta_0\lbrack\delta(v+v_m)-\delta(v-v_m)\rbrack$.
The critical Jeans wavenumber (\ref{sh3}) is $k_{c}^{2}=4\pi G\rho/v_{0}^{2}=4\pi G\rho/3T$ which is fully consistent with the expression (\ref{pdr3}) with $\gamma=1+1/n=3$. For $n=1/2$, we can obtain an explicit expression of the dielectric function (\ref{pdr4}). We get
\begin{eqnarray}
\epsilon(k,\omega)=1-\frac{k_c^2}{k^2}W_{1/2}\left (\frac{\omega}{k\sqrt{T}}\right ),
\label{diela}
\end{eqnarray}
with
\begin{eqnarray}
W_{1/2}=\frac{1}{1-\frac{1}{3}z^2}.
\label{dielb}
\end{eqnarray}
The condition $\epsilon(k,\omega)=0$ determines the pulsation. For $k>k_c$, the system is stable and the perturbation presents pure oscillations with pulsation
\begin{eqnarray}
\omega=\pm\sqrt{3T}(k^2-k_c^2)^{1/2}.
\label{dielc}
\end{eqnarray}
For $k<k_c$, the system is unstable and the perturbation has a growth rate (and a decay rate) given by
\begin{eqnarray}
\omega=\pm i\sqrt{3T}(k_c^2-k^2)^{1/2}.
\label{dielcb}
\end{eqnarray}
We note that, for the water-bag distribution, the general asymptotic behavior (\ref{mon7q}) becomes exact for all $k\le k_c$. We also note that for the specific index $n=1/2$ ($\gamma=3$), the dispersion relation in a stellar system takes the same form as in a gas (see Sec. \ref{sec_gas}).
\subsection{The Nyquist curve}
\label{sec_polyn}
Introducing the dimensionless wavenumber and dimensionless pulsation (\ref{mn1}), the
dielectric function (\ref{pdr5}) can be rewritten
\begin{eqnarray}
\epsilon({\eta},\Omega)=1-\frac{1}{\gamma}\eta W_{n}(\sqrt{\eta}\Omega).
\label{pn1}
\end{eqnarray}
When $\Omega_{i}=0$, the real and imaginary parts of the dielectric function $\epsilon(\eta,\Omega_{r})=\epsilon_{r}(\eta,\Omega_{r})+i\epsilon_{i}(\eta,\Omega_{r})$ are given by
\begin{equation}
\label{pn2}
\epsilon_{r}(\eta,\Omega_{r})=1-\frac{1}{\gamma}\eta W_{r}^{(n)}\left (\sqrt{\eta}\Omega_r\right ),
\end{equation}
\begin{equation}
\label{pn3}
\epsilon_{i}(\eta,\Omega_{r})=-\frac{1}{\gamma}\eta W_{i}^{(n)}\left (\sqrt{\eta}\Omega_r\right ),
\end{equation}
with
\begin{eqnarray}
\label{pn4}
W_{r}^{(n)}(z)=\frac{1}{\sqrt{2\pi}}\frac{B_n}{n}\left (n-\frac{1}{2}\right )\nonumber\\
\times P\int_{-\infty}^{+\infty}\frac{x[1-\frac{x^2}{2(n+1)}]_{+}^{n-3/2}}{x-z}dx,
\end{eqnarray}
\begin{eqnarray}
\label{pn5}
W_{i}^{(n)}(z)=\sqrt{\frac{\pi}{2}}\frac{B_n}{n}\left (n-\frac{1}{2}\right )z\left [1-\frac{z^2}{2(n+1)}\right ]_+^{n-3/2},\nonumber\\
\end{eqnarray}
where $z$ is here a real number. The condition of marginal stability
corresponds to
$\epsilon_{r}(\eta,\Omega_{r})=\epsilon_{i}(\eta,\Omega_{r})=0$. The condition
$\epsilon_{i}(\eta,\Omega_{r})=0$, which is equivalent to
$f'(\sqrt{\eta}\Omega_{r})=0$, implies $\Omega_r=0$. Then, the relation $\epsilon_{r}(\eta,\Omega_{r})=0$ leads to $\eta=\eta_{c}$ with
\begin{equation}
\label{pn6}
\eta_c={\gamma}.
\end{equation}
To apply the Nyquist method, we need to plot the curve
$(\epsilon_{r}(\eta,\Omega_{r}),\epsilon_{i}(\eta,\Omega_{r}))$ in the $\epsilon$-plane.
We have to distinguish different cases according to the value of the index $n$. A general discussion
has been given by Chavanis \& Delfini (2009) in the context of the HMF model. This discussion can be immediately transposed to the present context.
\section{The symmetric double-humped distribution}
\label{sec_vhq}
\subsection{Determination of the extrema}
\label{sec_ext}
We consider a reduced distribution function (\ref{lsv3}) of the form
\begin{equation}
\label{e1}
f(v)=\sqrt{\frac{\beta}{2\pi}} \frac{\rho}{2} \left [ e^{-\frac{\beta}{2}(v-v_a)^2} + e^{-\frac{\beta}{2}(v+v_a)^2} \right ].
\end{equation}
This is a symmetric double-humped distribution corresponding to the
superposition of two Maxwellian distributions with temperature
$T=1/\beta$ centered in $v_a$ and $-v_a$ respectively (see
Fig. \ref{dbumpdist}). This distribution models two streams of
particles in opposite direction. The average velocity is $\langle
v\rangle=0$ and the kinetic temperature $T_{kin}\equiv \langle
v^2\rangle=T+v_{a}^{2}$. The velocitie(s) $v_0$ at which the
distribution function $f(v)$ is extremum satisfy $f'(v_0)=0$. They are
determined by the equation
\begin{equation}
\label{e2}
e^{-2\beta v_{a} v_0}=\frac{v_a-v_0}{v_a+v_0}.
\end{equation}
We note that $v_0\in \lbrack -v_a, +v_a\rbrack$. Introducing the dimensionless velocity and the dimensionless separation
\begin{equation}
\label{e3}
V=\sqrt{\beta}v,\qquad V_a=\sqrt{\beta}v_a,
\end{equation}
Eq. (\ref{e2}) can be rewritten
\begin{equation}
\label{e4}
1=\frac{1}{2V_a V_0}\ln\left (\frac{V_a+V_0}{V_a-V_0}\right ).
\end{equation}
It is convenient to introduce the variables
\begin{equation}
\label{e5}
x=\frac{V_0}{V_a}, \qquad y=V_a^2.
\end{equation}
For a fixed temperature $T$, $x$ plays the role of the velocity $v_0$ at which the distribution is extremum and $y$ plays the role of separation $v_a$. Then, we have to study the function
\begin{equation}
\label{e6}
y(x)=\frac{1}{2x}\ln\left (\frac{1+x}{1-x}\right ),
\end{equation}
for $x \in ] -1, +1[$. This function is plotted in Fig. \ref{ysym}. It has the
following properties:
\begin{equation}
\label{e7}
y(-x)=y(x),
\end{equation}
\begin{equation}
\label{e8}
y(x)\sim -\frac{1}{2}\ln (1-x), \qquad (x\rightarrow 1^{-}),
\end{equation}
\begin{equation}
\label{e9}
y(0)=1.
\end{equation}
The extrema of the distribution function (\ref{e1}) can be deduced
from the study of this function. First, considering Eq. (\ref{e4}), we
note that $f(v)$ always has an extremum at $v_0=0$, for any value of
$\beta$ and $v_a$. This is a ``degenerate'' solution of Eq. (\ref{e6})
corresponding to the vertical line $x=0$ in Fig. \ref{ysym}. On the
other hand, if $y>1$, i.e $\beta v_a^2>1$, there exists two other extrema
$v_{0}=\pm v_{*}$ where $v_{*}=v_a x_*$ with $x_*=y^{-1}(\beta v_a^2)$.
In conclusion, for a given temperature $T$:
$\bullet$ if $y>1$ ($v_a^2>T$), the distribution function $f(v)$ has two maxima at $v_{0}=\pm v_{*}$ and one minimum at $v_{0}=0$.
$\bullet$ if $y\le 1$ ($v_a^2\le T$), the distribution function $f(v)$ has only one
maximum at $v_{0}=0$ (the limit case $\beta v_a^2=1$ corresponds to $f''(0)=0$).
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.3]{dbump-dist.eps}
\caption{Symmetric double-humped distribution made of two Maxwellians with separation $y$ (for a given temperature $T$). If $y>1$, the distribution has two maxima at $\pm V_{*}$ and one minimum at $V_{0}=0$ while for $y<1$, it has only one maximum at $V_{0}=0$.}
\label{dbumpdist}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.3]{y_sym.eps}
\caption{The function $y(x)$ for the symmetric double-humped distribution.}
\label{ysym}
\end{center}
\end{figure}
\subsection{The condition of marginal stability}
\label{sec_sms}
Introducing the dimensionless variables (\ref{mn1}) and (\ref{e5}), the dielectric function associated to the symmetric double-humped distribution (\ref{e1}) is
\begin{eqnarray}
\label{sms1}
\epsilon(\eta,\Omega)= 1-\frac{\eta}{2} \left [ W(\sqrt{\eta}\Omega-\sqrt{y}))+W(\sqrt{\eta}\Omega+\sqrt{y})) \right ],\nonumber\\
\end{eqnarray}
where $W(z)$ is defined in Eq. (\ref{md5}). For a fixed temperature, $\eta$ plays the role of the wavenumber $k$, $\Omega$ plays the role of the pulsation $\omega$ and $y$ plays the role of the separation $v_a$. When $\Omega_{i}=0$, the real and imaginary parts of the dielectric function $\epsilon(\eta,\Omega_{r})=\epsilon_{r}(\eta,\Omega_{r})+i\epsilon_{i}(\eta,\Omega_{r})$ can be written
\begin{eqnarray}
\label{sms2}
\epsilon_{r}(\eta,\Omega_{r})=1-\frac{\eta}{2} \left [ W_r(\sqrt{\eta}\Omega_{r} - \sqrt{y}))+W_r(\sqrt{\eta}\Omega_{r} + \sqrt{y})) \right ],\nonumber\\
\end{eqnarray}
\begin{eqnarray}
\epsilon_{i}(\eta,\Omega_{r})=-\frac{\eta}{2} \left [ W_i(\sqrt{\eta}\Omega_{r} - \sqrt{y}))+W_i(\sqrt{\eta}\Omega_{r} + \sqrt{y})) \right ],\nonumber\\
\end{eqnarray}
where $W_r(z)$ and $W_i(z)$ are defined in
Eqs. (\ref{mn6})-(\ref{mn7}) where $z$ is here a real number. The
condition of marginal stability corresponds to
$\epsilon_{r}(\eta,\Omega_{r})=\epsilon_{i}(\eta,\Omega_{r})=0$. The condition
$\epsilon_{i}(\eta,\Omega_{r})=0$ is equivalent to
\begin{equation}
\label{sms3}
f'(\sqrt{\eta}\Omega_{r})=0.
\end{equation}
The condition $\epsilon_{r}(\eta,\Omega_{r})=0$ leads to
\begin{equation}
\label{sms4}
1-\frac{\eta}{2} \left [ W_r(\sqrt{\eta}\Omega_{r} - \sqrt{y}))+W_r(\sqrt{\eta}\Omega_{r} + \sqrt{y})) \right ]=0.\nonumber\\
\end{equation}
Therefore, according to Eq. (\ref{sms3}), the phase velocity
$\sqrt{\eta}\Omega_{r}$ is equal to a velocity $V_{0}$ at which the distribution
(\ref{e1}) is extremum. The second equation (\ref{sms4}) determines
the value(s) $\eta_{c}(y)$ of the wavenumber at which the
distribution is marginally stable.
\subsubsection{The case $\omega_{r}=0$}
Let us first consider the value $\Omega_{r}=0$ that is solution of
Eq. (\ref{sms3}) for any $v_a$ and $\beta$. In that case, Eq. (\ref{sms4}) becomes
\begin{equation}
\label{sms7}
\eta_{c}^{(0)}(y)=\frac{1}{W_{r}(\sqrt{y})},
\end{equation}
where we have used $W_r(-x)=W_r(x)$. For given separation $y$,
this equation determines the wavenumber $\eta_{c}^{(0)}(y)$
corresponding to a mode of marginal stability with
$\Omega_{r}=0$. The function defined by Eq. (\ref{sms7}) is plotted in
Fig. \ref{w3}. It diverges
at $y=y_{max}=z_{c}^{2}$ where
$z_{c}=1.307$ is the zero of $W_{r}(z)$ (see Appendix A of Chavanis \& Delfini 2009). Then, using $W_r'(z_c)=-1/z_c$, we
find from Eq. (\ref{sms7}) that
\begin{equation}
\label{sms8}
\eta_{c}^{(0)}(y)\sim \frac{2y_{max}}{y_{max}-y}, \qquad (y\rightarrow y_{max}).
\end{equation}
For $y>y_{max}=1.708$, $W_{r}(\sqrt{y})$ is negative so
the branch $\eta_c^{(0)}(y)$ exists only for $y\in[0,y_{max}]$. For
$y\rightarrow 0$, we have
\begin{equation}
\label{sms9}
\eta_{c}^{(0)}(0)=1.
\end{equation}
This result is to be expected since, for $v_a=0$, the distribution
(\ref{e1}) reduces to the Maxwellian. We thus recover the critical
wavenumber (\ref{mn8}).
In conclusion:
$\bullet$ if $y<y_{max}$, there exists a critical wavenumber
$\eta_{c}^{(0)}(y)$ determined by
Eq. (\ref{sms7}) corresponding to a marginal mode
$(\Omega_{r}=0,\Omega_{i}=0)$.
$\bullet$ if $y>y_{max}$, there is no marginal mode
$(\Omega_{r}=0,\Omega_{i}=0)$.
The curve $\eta_{c}^{(0)}(y)$ corresponding to the
marginal mode with zero pulsation $\Omega_r=0$ is plotted in Fig. \ref{GRAVdiagphasesym}.
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.3]{W3.eps}
\caption{$\eta=1/W_r(\sqrt{y})$ as a function of $y$.}
\label{w3}
\end{center}
\end{figure}
\subsubsection{The case $\omega_{r}\neq 0$}
We now consider the cases where $\sqrt{\eta}\Omega_{r}=\pm V_{*}$ is
solution of Eq. (\ref{sms3}) for $y>1$. To determine the
wavenumber(s) at which the distribution (\ref{e1}) is marginally stable, we
have to solve
\begin{equation}
\label{sms10}
1-\frac{\eta}{2} \left [ W_r(V_*- \sqrt{y}))+W_r(V_* + \sqrt{y})) \right ]=0,\nonumber\\
\end{equation}
where $V_{*}$ is given by
\begin{equation}
\label{sms11}
1=\frac{1}{2V_* \sqrt{y}}\ln\left (\frac{\sqrt{y}+V_{*}}{\sqrt{y}-V_{*}}\right ).
\end{equation}
Eliminating $V_*$ between these two expressions, we obtain the critical wavenumber(s) $\eta_{c}^{(\pm)}(y)$ as a function of $y$. However, it is easier to proceed differently. Setting $x=V_*/V_a=V_*/\sqrt{y}$, we obtain the equations
\begin{equation}
\label{sms12}
y=\frac{1}{2x}\ln\left (\frac{1+x}{1-x}\right ),
\end{equation}
\begin{equation}
\label{sms13}
\eta=\frac{2}{\left [ W_{r}(\sqrt{y}(x - 1))+W_{r}(\sqrt{y}(x+ 1)) \right ]}.
\end{equation}
For given $x$, we can obtain $y$ from Eq. (\ref{sms12}) [see also
Fig. \ref{ysym}] and $\eta$ from Eq. (\ref{sms13}). Varying $x$ in the interval $]-1,1[$ yields the
full curve $\eta_{c}^{(\pm)}(y)$. By symmetry, we can restrict
ourselves to the interval $x\in [0,1[$.
For $x=0$, we have $y=1$ and
\begin{equation}
\label{sms18}
\eta_{c}^{(\pm)}(1)=\eta_*=\frac{1}{W_r(1)}=3.633.
\end{equation}
The branch $\eta_{c}^{(\pm)}(y)$ starts at the
point $(1,\eta_{*})$, corresponding to $\sqrt{\eta}\Omega_{r}=\pm V_{*}=0$
(i.e. $x=0$). This point is at the intersection between the branch $\eta_{c}^{(0)}(y)$ along which
$\Omega_{r}=0$ and the line $y=1$
separating the regions where the distribution has one or two maxima.
For $x\rightarrow 1$, we have $y\sim -\frac{1}{2}\ln(1-x)\rightarrow +\infty$ and
\begin{equation}
\label{sms19}
\eta_{c}^{(\pm)}(y)=2+\frac{1}{2y}+... \qquad (y\rightarrow +\infty).
\end{equation}
This result can be understood
simply. For $y\rightarrow +\infty$, the two humps are far away from
each other so that the critical wavenumber
$\eta_{c}^{(\pm)}$ coincides with the critical wavenumber of a
single Maxwellian since they do not ``see'' each other. Noting that
the mass of a single hump is $M/2$, the corresponding critical
wavenumber is $k_{c}=4\pi G(\rho/2)/T$ leading to
$\eta_{c}^{(\pm)}=2$.
In conclusion:
$\bullet$ if $y>1$, there exists a single critical wavenumber
$\eta_{c}^{(\pm)}(y)$ determined by Eqs. (\ref{sms12})-(\ref{sms13})
corresponding to a marginal mode $(\Omega_r=\pm V_*/\sqrt{\eta_c},\Omega_i=0)$. Note
that the modes $\Omega_r=+V_*/\sqrt{\eta_c}$ and $\Omega_r=-V_{*}/\sqrt{\eta_c}$ are
degenerate. This degeneracy can be raised by a small asymmetry
(symmetry breaking) in the distribution (see Sec. \ref{sec_vha}).
The curve $\eta_{c}^{(\pm)}(y)$ corresponding to the
marginal mode with pulsation $\Omega_r=\pm V_{*}/\sqrt{\eta_c}$ is plotted in Fig. \ref{GRAVdiagphasesym}.
\subsection{The stability diagram}
\label{sec_sdsy}
The critical wavenumbers $\eta_c(y)$ corresponding to marginal
stability determined previously are represented as a function of the
separation $y$ in Fig. \ref{GRAVdiagphasesym}. We have also plotted the line
$y=1$. On the left of this line, the distribution has a single maximum at $V_0=0$
and on the right of this line, the distribution has two maxima at $V_0=\pm V_*$ and a
minimum at $V_0=0$. In order to investigate the stability of the
solutions in the different regions, we have used the Nyquist method.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{GRAVdiagphasesym.eps}
} \caption[]{Stability diagram of the symmetric double-humped
distribution (\ref{e1}). The solid line $(\eta_c)_*$ corresponds to
the critical line: below this line the DF is stable and above this
line the DF is unstable. On the left panel (delimited by the dotted
line), there is one mode of instability and on the right panel there
are two modes of instability. The dashed line corresponds to $y=1$:
on the left of this line the DF has one maximum and on the right of
this line the DF has two maxima and one minimum. The symbols in
parenthesis like $(x=0)$ give the values of $x$ or $y$ that
parametrize the marginal curves. The notations like $(- + -)$ give the
positions of the $\epsilon_{r}(v_{ext})$'s in the Nyquist curves as
defined in Sec. \ref{sec_doubh}. }
\label{GRAVdiagphasesym}
\end{figure}
For $y<1$, there exists one wavenumber $\eta_{c}^{(0)}(y)$ at which
the DF is marginally stable. For $\eta=\eta_{c}^{(0)}(y)$, the DF has
one maximum at $V_0=0$. The marginal perturbation does not propagate
($\Omega_r=0$). By considering the Nyquist curves in this region (see
Figs. \ref{doublesym_y0.5_n0.5}-\ref{doublesym_y0.5_n6}), we find that
the DF is stable for $\eta<\eta_{c}^{(0)}(y)$ and unstable for
$\eta>\eta_{c}^{(0)}(y)$.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y0.5_n0.5.eps}
} \caption[]{Nyquist curve for $y<1$ and $\eta<\eta_c^{(0)}$
(specifically $y=0.5$ and $\eta=0.5$). The DF has only one maximum at
$V_0=0$. It is stable (with respect to this perturbation) because the
Nyquist curve does not encircle the origin. Case (+).}
\label{doublesym_y0.5_n0.5}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y0.5_n6.eps}
} \caption[]{Nyquist curve for $y<1$ and $\eta>\eta_c^{(0)}$ (specifically $y=0.5$ and $\eta=6$). The DF has only one maximum at
$V_0=0$. It is unstable (with respect to this perturbation) because the
Nyquist curve encircles the origin. Case (-).}
\label{doublesym_y0.5_n6}
\end{figure}
For $y>y_{max}$, there exists one wavenumber $\eta_{c}^{(\pm)}(y)$ at
which the DF is marginally stable. For $\eta=\eta_{c}^{(\pm)}(y)$, the
DF has two maxima at $V_0=\pm V_*$ and one minimum at $V_0=0$. The
marginal perturbation evolves with a pulsation $\Omega_r=\pm
V_*/\sqrt{\eta_c}$. By considering the Nyquist curves in this region
(see Figs. \ref{doublesym_y2n1}-\ref{doublesym_y2n6}), we find that
the DF is stable for $\eta<\eta_{c}^{(\pm)}(y)$ and unstable for
$\eta>\eta_{c}^{(\pm)}(y)$.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y2n1.eps}
} \caption[]{Nyquist curve for $y>y_{max}$ and $\eta<\eta_c^{(\pm)}$ (specifically $y=2$ and $\eta=1$). The DF has two maxima at $V_0=\pm V_*$ and one minimum at $V_0=0$. The DF is stable (with respect to this perturbation) because the
Nyquist curve does not encircle the origin. Case (+ + +).}
\label{doublesym_y2n1}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y2n6.eps}
} \caption[]{Nyquist curve for $y>y_{max}$ and $\eta>\eta_c^{(\pm)}$
(specifically $y=2$ and $\eta=6$). The DF has two maxima at $V_0=\pm
V_*$ and one minimum at $V_0=0$. The DF is unstable (with respect to
this perturbation) because the Nyquist curve encircles the
origin. Since it rotates twice around the origin, this implies that there are $N=2$ unstable modes $(\omega_r,\omega_i)$ with $\omega_i>0$. Case (- + -).}
\label{doublesym_y2n6}
\end{figure}
For $1<y<y_{max}$, there exists two wavenumbers $\eta_{c}^{(0)}(y)$
and $\eta_{c}^{(\pm)}(y)$ at which the DF is
marginally stable. The DF has
two maxima at $V_0=\pm V_*$ and one minimum at $V_0=0$. For $\eta=\eta_{c}^{(0)}(y)$, the
marginal perturbation does not propagate ($\Omega_r=0$). For
$\eta=\eta_{c}^{(\pm)}(y)$, the marginal perturbation has a pulsation
$\Omega_r=\pm V_*/\sqrt{\eta_c}$. By considering the Nyquist curves in this
region (see Fig. \ref{doublesym_y1.2_n6}), we find that the DF is stable
for $\eta<\eta_{c}^{(\pm)}(y)$ and unstable for $\eta>\eta_{c}^{(\pm)}(y)$.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y1.2_n6.eps}
} \caption[]{Nyquist curve for $1<y<y_{max}$ and $\eta>\eta_c^{(0)}$
(specifically $y=1.2$ and $\eta=6$). The DF has two maxima at $V_0=\pm
V_*$ and one minimum at $V_0=0$. The DF is unstable (with respect to
this perturbation) because the Nyquist curve encircles the
origin. Since it rotates only once around the origin, this implies that there is $N=1$ unstable mode $(\omega_r,\omega_i)$ with $\omega_i>0$. Case (- - -).}
\label{doublesym_y1.2_n6}
\end{figure}
A few comments are in order:
1. In Fig. \ref{GRAVdiagphasesym}, we explicitly see that the system is always unstable with respect to perturbations with sufficiently small $k$ (the case of cold systems $T=0$ is treated in Appendix \ref{sec_cold}). This corroborates the general result given at the end of Sec. \ref{sec_n}. More precisely, we see that the system is stable for perturbations with $k>(k_c)_*$ (corresponding to the solid line) and unstable for perturbations with $k<(k_c)_*$. Furthermore, we note that this critical wavenumber $(k_c)_*$ corresponds to a marginal perturbation where the phase velocity $\omega/k$ coincides with the {\it maximum} of the velocity distribution: for $v_a^2<T$, this is $v_0=0$ and for $v_a^2>T$, this is $v_0=\pm v_*$.
2. For $1<y<y_{max}$ and $\eta_c^{(\pm)}<\eta<\eta_c^{(0)}$, there are two unstable modes since the Nyquist curve encircles the origin twice. One of these two modes is $\omega=i\omega_i$ and the other is $\omega=\omega_r+i\omega_i$ with $\omega_r\neq 0$ (overstable). When we increase $\eta$ and cross the marginal line $\eta=\eta_c^{(0)}$, the mode $\omega=i\omega_i$ becomes stable according to the general result (\ref{mon15}). Indeed, for $y>1$, the DF is {\it minimum} at $v=0$. This is why there is only one unstable mode for $\eta>\eta_c^{(0)}$: the mode $\omega=\omega_r+i\omega_i$ with $\omega_r\neq 0$ (overstable). On the other hand, for $y<1$, the mode $\omega=i\omega_i$ is stable for $\eta<\eta_c^{(0)}$. When we increase $\eta$ and cross the marginal line $\eta=\eta_c^{(0)}$, it becomes unstable according to the general result (\ref{mon15}). Indeed, for $y>1$, the DF is {\it maximum} at $v=0$.
3. In Fig. \ref{GRAVdiagphasesym}, we see that the critical Jeans length $(\lambda_c)_*$ associated to a double-humped DF ($v_a\neq 0$) is always larger than the critical Jeans length associated to the single-humped Maxwellian ($v_a=0$). This means that the presence of streaming in a purely stellar system has a stabilizing role. This is different if there exists a gas component in the system since the critical Jeans length is reduced by the relative motion of the gas and stars (Sweet 1963).
4. For a double-humped stellar system with $v_a<\sqrt{T}$, the critical Jeans length increases with $v_a$ due to the stabilization effect of the relative velocity. Alternatively, in a contra-streaming self-gravitating gas with $v_a<c_s$, the critical Jeans length decreases with $v_a$ and tends to zero as $v_a\rightarrow c_s$ (Talwar \& Kalra 1966, Ikeuchi et al. 1974). On the other hand, for a double-humped stellar system with $v_a>\sqrt{T}$ or for a contrastreaming self-gravitating gas with $v_a>c_s$, the critical Jeans length decreases with $v_a$ (but remains larger than the classical Jeans wavelength corresponding to $v_a=0$) and overstable modes appear.
\section{The asymmetric double-humped distribution}
\label{sec_vha}
\subsection{Determination of the extrema}
\label{sec_ae}
We now assume that the reduced distribution (\ref{lsv3}) is an asymmetric
double-humped distribution of the form
\begin{equation}
\label{ae1}
f(v)=\sqrt{\frac{\beta}{2\pi}} \frac{\rho}{1+\Delta} \left [ e^{-\frac{\beta}{2}(v-v_a)^2} + \Delta e^{-\frac{\beta}{2}(v+v_a)^2} \right ],
\end{equation}
where $T=1/\beta$ is the temperature of the Maxwellians and $\Delta$
is the asymmetry parameter (we assume here that $\Delta>1$). This distribution is plotted in Fig. \ref{asymdistr}. The symmetric case is recovered for
$\Delta=1$. The average velocity is $\langle
v\rangle=-[(\Delta-1)/(\Delta+1)]v_a$ and the kinetic temperature
$T_{kin}\equiv \langle (v-\langle v\rangle)^2\rangle=T+[4\Delta/(\Delta+1)^2]v_{a}^{2}$. The velocities $v_0$ at
which the distribution function $f(v)$ is extremum satisfy
$f'(v_0)=0$. They are determined by the equation
\begin{equation}
\label{ae2}
e^{-2\beta v_{a} v_0}=\frac{1}{\Delta}\frac{v_a-v_0}{v_a+v_0}.
\end{equation}
We note that $v_0\in \rbrack -v_a,
+v_a\lbrack$. Introducing the
dimensionless velocity and the dimensionless separation
(\ref{e3}), Eq. (\ref{ae2}) can be rewritten
\begin{equation}
\label{ae3}
1=\frac{1}{2V_a V_0}\ln\left (\frac{V_a+V_0}{V_a-V_0}\right )+\frac{\ln(\Delta)}{2 V_a V_0}.
\end{equation}
It is convenient to introduce the variables
\begin{equation}
\label{ae4}
x=\frac{V_0}{V_a}, \qquad y=V_a^2.
\end{equation}
Then, we have to study the function
\begin{equation}
\label{ae5}
y(x)=\frac{1}{2x}\ln\left (\frac{1+x}{1-x}\right )+\frac{\ln(\Delta)}{2x},
\end{equation}
for $x \in ] -1, +1[$. This function is plotted in
Fig. \ref{yasym}. It has the following properties
\begin{equation}
\label{ae6}
y(x)\sim -\frac{1}{2}\ln (1-x), \qquad (x\rightarrow 1^{-}),
\end{equation}
\begin{equation}
\label{ae7}
y(x)\sim -\frac{1}{2}\ln (1+x), \qquad (x\rightarrow -1^{+}),
\end{equation}
\begin{equation}
\label{ae8}
y(x)\sim \frac{\ln(\Delta)}{2x}, \qquad (x\rightarrow 0).
\end{equation}
Considering the negative velocities $x<0$, we note that $y\ge 0$ iff $x\le x_0$ with
\begin{equation}
\label{ae9}
x_{0}=\frac{1-\Delta}{1+\Delta}.
\end{equation}
Considering the positive velocities $x>0$, we note that the curve $y(x)$ is
minimum at $x=x_*$ where $x_{*}$ is solution of
\begin{equation}
\label{ae10}
\frac{2x_{*}}{1-x_{*}^{2}}-\ln\left (\frac{1+x_{*}}{1-x_{*}}\right )=\ln(\Delta).
\end{equation}
This function is represented in Fig. \ref{xynew}. We note that
Eq. (\ref{ae10}) has a unique solution $x_*$ for each value of
$\Delta>1$. Therefore, the function $y(x)$ has a single minimum at
$x=x_{*}$. The value of this minimum is
\begin{equation}
\label{ae11}
y_{*}=\frac{1}{1-x_{*}^{2}}>1.
\end{equation}
The extrema of the distribution $f(V)$ can be determined from the
study of the function (\ref{ae5}). If $y<y_{*}$, the
distribution $f(V)$ has a single maximum at $V_{0}=V_{-}<0$. If
$y>y_{*}$, the distribution $f(V)$ has two
maxima at $V_{0}=V_{-}<0$ and $V_{0}=V_{+}>0$ and one minimum at
$V_{0}=V_{p}>0$. These different values are given by $V_{0}=V_a y^{-1}(V_a^2)$.
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.3]{asym-distr.eps}
\caption{Asymmetric double-humped distribution made of two Mawellians with separation $y$ and asymmetry $\Delta>1$. If $y>y_{*}$, the DF has one global maximum at $V_{0}=V_{-}<0$, a minimum at $V_{0}=V_{p}>0$ and a local maximum at $V_{0}=V_{+}>0$. If $y<y_{*}$, the DF has only one maximum at $V_{0}=V_{-}<0$.}
\label{asymdistr}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.3]{y_asym.eps}
\caption{The function $y(x)$ for the asymmetric double-humped distribution.}
\label{yasym}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.3]{x-y-new.eps}
\caption{Evolution of $x_*$, $y_*$, $x_c$ and $y_c$ as a function of $\Delta$. The curves intersect each other at $\Delta_{*}=3.3105$.}
\label{xynew}
\end{center}
\end{figure}
In conclusion, for a given asymmetry $\Delta>1$ and temperature $T$:
$\bullet$ if $y>y_{*}$, the distribution function $f(V)$ has a global maximum at $V_{0}=V_{-}<0$, a local maximum at $V_{0}=V_{+}>0$ and one minimum at $V_{0}=V_{p}>0$.
$\bullet$ if $y<y_{*}$, the distribution function $f(V)$ has only one
maximum at $V_{0}=V_{-}<0$.
\subsection{The condition of marginal stability}
\label{sec_ams}
The dielectric function associated to the asymmetric double-humped
distribution is
\begin{eqnarray}
\label{ams1}
\epsilon(\Omega)= 1-\frac{\eta}{1+\Delta} \left [ W(\sqrt{\eta}\Omega - \sqrt{y}))+\Delta W(\sqrt{\eta}\Omega + \sqrt{y})) \right ],\nonumber\\
\end{eqnarray}
where $W(z)$ is defined in Eq. (\ref{md5}). When $\Omega_{i}=0$, the real and imaginary parts of the dielectric function $\epsilon(\eta,\Omega_{r})=\epsilon_{r}(\eta,\Omega_{r})+i\epsilon_{i}(\eta,\Omega_{r})$ can be written
\begin{eqnarray}
\label{ams2}
\epsilon_{r}(\eta,\Omega_{r})=1-\frac{\eta}{1+\Delta} [ W_r(\sqrt{\eta}\Omega_{r} - \sqrt{y}))\nonumber\\
+\Delta W_r(\sqrt{\eta}\Omega_{r} + \sqrt{y})) ],
\end{eqnarray}
\begin{eqnarray}
\label{ams3}
\epsilon_{i}(\eta,\Omega_{r})=-\frac{\eta}{1+\Delta} [ W_i(\sqrt{\eta}\Omega_{r} - \sqrt{y}))\nonumber\\
+\Delta W_i(\sqrt{\eta}\Omega_{r} + \sqrt{y})) ],
\end{eqnarray}
where $W_r(z)$ and $W_i(z)$ are defined in Eqs. (\ref{mn6})-(\ref{mn7})
where $z$ is here a real number. The condition of marginal stability corresponds to $\epsilon_{r}(\eta,\Omega_{r})=\epsilon_{i}(\eta,\Omega_{r})=0$. The condition $\epsilon_{i}(\eta,\Omega_{r})=0$ is equivalent to
\begin{equation}
\label{ams4}
f'(\sqrt{\eta}\Omega_{r})=0.
\end{equation}
The condition $\epsilon_{r}(\eta,\Omega_{r})=0$ leads to
\begin{eqnarray}
\label{ams5}
1-\frac{\eta}{1+\Delta} \left [ W_{r}(\sqrt{\eta}\Omega_{r} - \sqrt{y}))+\Delta W_{r}(\sqrt{\eta}\Omega_{r} + \sqrt{y})) \right ]=0.\nonumber\\
\end{eqnarray}
According to Eq. (\ref{ams4}), the phase velocity $\sqrt{\eta}\Omega_{r}$ is
equal to a velocity $V_{0}$ at which the distribution (\ref{ae1}) is
extremum. The second equation (\ref{ams5}) determines the value(s)
$\eta_{c}(y)$ of the wavenumber at which the distribution is marginally
stable. Therefore, we have to solve
\begin{eqnarray}
\label{ams6}
1-\frac{\eta}{1+\Delta} \left [ W_{r}(V_{0} - \sqrt{y}))+\Delta W_{r}(V_{0}+ \sqrt{y})) \right ]=0,\nonumber\\
\end{eqnarray}
where $V_{0}$ is given by
\begin{equation}
\label{ams7}
1=\frac{1}{2\sqrt{y}V_{0}}\ln\left (\frac{\sqrt{y}+V_{0}}{\sqrt{y}-V_{0}}\right )+\frac{\ln(\Delta)}{2 \sqrt{y} V_{0}}.
\end{equation}
Eliminating $V_0$ between these two expressions yields the
critical wavenumber(s) $\eta_{c}(y)$ as a function of
$y$. However, it is easier to proceed differently. Setting $x=V_0/\sqrt{y}$, we obtain the equations
\begin{equation}
\label{ams8}
y=\frac{1}{2x}\ln\left (\frac{1+x}{1-x}\right
)+\frac{\ln(\Delta)}{2x},
\end{equation}
\begin{equation}
\label{ams9}
\eta=\frac{1+\Delta}{\left [ W_{r}(\sqrt{y}(x - 1))+\Delta W_{r}(\sqrt{y}(x+ 1)) \right ]}.
\end{equation}
For given $x$, we can obtain $y$ from Eq. (\ref{ams8}) [see also
Fig. \ref{yasym}] and $\eta$ from Eq. (\ref{ams9}). Varying $x$ in the interval $]-1,1[$ yields the
full curve $\eta_c(y)$. We have three types of solutions. For $x\in
]-1,x_0]$, we obtain a branch $\eta_c^{(-)}(y)$ where the pulsation of
the marginal mode is negative: $\Omega_{r}=V_{-}/\sqrt{\eta_c}<0$ corresponding to
the global maximum of $f(v)$. For $x\in ]0,x_*]$, we obtain a branch
$\eta_c^{(p)}(y)$ where the pulsation of the marginal mode is
positive: $\Omega_{r}=V_{p}/\sqrt{\eta_c}>0$ corresponding to the minimum of
$f(v)$. For $x\in [x_*,1[$, we obtain a branch $\eta_c^{(+)}(y)$ where
the pulsation of the marginal mode is positive: $\Omega_{r}=V_{+}/\sqrt{\eta_c}>0$
corresponding to the local maximum of $f(v)$. This leads to the curves
reported in Figs. \ref{GRAVdiagphased1.02} and \ref{GRAVdiagphased10}
for two values of the asymmetry factor $\Delta$.
For $x\rightarrow -1$, $y\rightarrow +\infty$ and
\begin{equation}
\label{ams18}
\eta_{c}^{(-)}(y)=\frac{1+\Delta}{\Delta}+\frac{1+\Delta}{4\Delta^2 y}+... \qquad (y\rightarrow +\infty).
\end{equation}
For $x\rightarrow x_0$, $y\rightarrow 0$ and
\begin{equation}
\label{ams19}
\eta_{c}^{(-)}(y)\rightarrow 1 \qquad (y\rightarrow 0).
\end{equation}
This returns the critical wavenumber (\ref{mn8}) associated with the
Maxwellian distribution (\ref{md1}) corresponding to $y=0$.
For $x\rightarrow 1$, $y\rightarrow +\infty$ and
\begin{equation}
\label{ams19b}
\eta_{c}^{(+)}(y)\sim 1+\Delta+\frac{\Delta (1+\Delta)}{4y}+...\qquad (y\rightarrow +\infty).
\end{equation}
Let $x_c$ and $y_c$ be determined by the equations
\begin{equation}
\label{ams21}
y_{c}=\frac{1}{2x_{c}}\ln\left (\frac{1+x_{c}}{1-x_{c}}\right
)+\frac{\ln(\Delta)}{2x_{c}},
\end{equation}
\begin{equation}
\label{ams22}
W_{r}(\sqrt{y_{c}}(x_{c} - 1))+\Delta W_{r}(\sqrt{y_{c}}(x_{c}+ 1))=0.
\end{equation}
For $x\rightarrow x_c$, $y\rightarrow y_c$ and
\begin{equation}
\label{ams19c}
\eta_{c}^{(s)}(y)\propto \frac{1}{y-y_c}\rightarrow +\infty \qquad (y\rightarrow y_c),
\end{equation}
where $s=p$ if $\Delta<\Delta_{*}$ and $s=+$ if
$\Delta>\Delta_{*}$ as will become clear below.
Note that there is no physical solution to
Eqs. (\ref{ams8})-(\ref{ams9}) when $0<x<x_{c}$ ($\eta$ would be
negative) so that the branch $\eta_{c}^{(s)}(y)$ starts at
$(y_c,+\infty)$ corresponding to $x=x_{c}$. The evolution of $x_c$ and
$y_c$ with $\Delta$ is studied in Fig. \ref{xynew} and is compared
to the evolution of $x_*$ and $y_*$. The curves intersect each other
at $\Delta_*=3.3105$. For $\Delta<\Delta_*$, $x_c<x_*$ so that the
stability diagram displays three marginal branches
$\eta_{c}^{(-)}(y)$, $\eta_{c}^{(p)}(y)$ and $\eta_{c}^{(+)}(y)$ (see
Fig. \ref{GRAVdiagphased1.02}). The branches $\eta_{c}^{(p)}(y)$ and
$\eta_{c}^{(+)}(y)$ connect each other at $(y_{*},\eta_{*})$
corresponding to $x=x_{*}$. At that point they touch the line
$y=y_{*}$ separating distributions with one or two
maxima. For
$\Delta=\Delta_*$, $x_c=x_*$ so that the branch $\eta_{c}^{(p)}(y)$ is rejected to infinity and
only the branches $\eta_{c}^{(-)}(y)$ and
$\eta_{c}^{(+)}(y)$ remain. For $\Delta>\Delta_*$, $x_c>x_*$ so that the phase
diagram displays only two marginal branches $\eta_{c}^{(-)}(y)$ and
$\eta_{c}^{(+)}(y)$ (see Fig. \ref{GRAVdiagphased10}).
\subsection{The stability diagram}
The critical wavenumbers $\eta_c(y)$ corresponding to marginal
stability determined previously are represented as a function of the
separation $y$ in Figs. \ref{GRAVdiagphased1.02} and \ref{GRAVdiagphased10}
for two values of the asymmetry factor $\Delta$. We
recall that $\eta_{c}^{(-)}(y)$ corresponds to the wavenumber
associated with a marginal mode with pulsation $\Omega_{r}=V_{-}/\sqrt{\eta_c}<0$ (global maximum of $f$),
$\eta_{c}^{(p)}(y)$ corresponds to the wavenumber associated with a
marginal mode with pulsation $\Omega_{r}=V_{p}/\sqrt{\eta_c}>0$ (minimum of $f$) and
$\eta_{c}^{(+)}(y)$ corresponds to the wavenumber associated with a
marginal mode with pulsation $\Omega_{r}=V_{+}/\sqrt{\eta_c}>0$ (local maximum of $f$). We have also
plotted the line $y=y_*$. On the left of this line, the DF has a
single maximum at $V_{-}<0$ and on the right of this line, the DF has two
maxima at $V_{-}<0$ and $V_{+}>0$ and a minimum at $V_{p}>0$. In order
to investigate the stability of the solutions in the different regions
of the parameter space, we have used the Nyquist criterion. The description
of the stability diagram is similar to the one given in
Sec. \ref{sec_sdsy} and the different possible cases can be understood
directly from the reading of Figs. \ref{GRAVdiagphased1.02} and \ref{GRAVdiagphased10}. The best way is to fix $y$ and progressively
increase the value of $\eta$. For $y<y_*$, the distribution
has only one maximum at $V_{0}=V_{-}$ so the Nyquist curve has one
intersection with the $x$-axis (in addition to the limit point
$(1,0)$). For $\eta\rightarrow 0$, we find that $\epsilon_{r}(V_{-})>0$
so the system is stable. As we increase $\eta$ and pass above the solid
line, we find that $\epsilon_{r}(V_{-})<0$
so the system is unstable. For $y>y_*$, the distribution
has two maxima at $V_{0}=V_{-}$ and
$V_{0}=V_{+}$ and one minimum at $V_{0}=V_{p}$. At each intersection
with a marginal line, one of the values $\epsilon_{r}(V_{-})$,
$\epsilon_{r}(V_{p})$ or $\epsilon_{r}(V_{+})$ changes sign. We have indicated by symbols like $(-++)$ the respective signs of $\epsilon_{r}(V_{-})$,
$\epsilon_{r}(V_{p})$ and $\epsilon_{r}(V_{+})$. We
can then easily draw by hands the corresponding Nyquist curve.
Therefore, it is not necessary to show all the possibilities and we
have only indicated a few representative cases in
Figs. \ref{data_delta1.02_y1.1_n1}-\ref{data_delta1.02_y1.1_n6} for illustration.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{GRAVdiagphased1.02.eps}
} \caption[]{Stability diagram of the asymmetric double-humped distribution for $\Delta<\Delta_*$ (specifically $\Delta=1.02$). There exists three marginal branches $\eta_c^{(-)}(y)$, $\eta_c^{(p)}(y)$ and $\eta_c^{(+)}(y)$. The symbols have been defined in the text.}
\label{GRAVdiagphased1.02}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{GRAVdiagphased10.eps}
} \caption[]{Stability diagram of the asymmetric double-humped distribution for $\Delta>\Delta_*$ (specifically $\Delta=10$). There exists two marginal branches $\eta_c^{(-)}(y)$ and $\eta_c^{(+)}(y)$. }
\label{GRAVdiagphased10}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta1.02_y1.1_n1.eps}
} \caption[]{$\Delta<\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta<\eta_c^{(-)}$
(specifically $\Delta=1.02$, $y=1.1$ and $\eta=1$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is stable (with respect to
this perturbation) because the Nyquist curve does not encircle the
origin. Case (+ + +).}
\label{data_delta1.02_y1.1_n1}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta1.02_y1.1_n3.436.eps}
} \caption[]{$\Delta<\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta_c^{(-)}<\eta<\eta_c^{(+)}$
(specifically $\Delta=1.02$, $y=1.1$ and $\eta=3.436$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is unstable (with respect to
this perturbation) because the Nyquist curve encircles the
origin once. There is $N=1$ unstable mode. Case (- + +).}
\label{data_delta1.02_y1.1_n3.436}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta1.02_y1.1_n4.eps}
} \caption[]{$\Delta<\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta_c^{(+)}<\eta<\eta_c^{(p)}$
(specifically $\Delta=1.02$, $y=1.1$ and $\eta=4$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is unstable (with respect to
this perturbation) because the Nyquist curve encircles the
origin twice. There are $N=2$ unstable modes. Case (- + -).}
\label{data_delta1.02_y1.1_n4}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta1.02_y1.1_n6.eps}
} \caption[]{$\Delta<\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta>\eta_c^{(p)}$
(specifically $\Delta=1.02$, $y=1.1$ and $\eta=6$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is unstable (with respect to
this perturbation) because the Nyquist curve encircles the
origin once. There is $N=1$ unstable mode. Case (- - -).}
\label{data_delta1.02_y1.1_n6}
\end{figure}
{\it Remark}: In Figs. \ref{GRAVdiagphased1.02} and \ref{GRAVdiagphased10}, we see that the system is stable for perturbations with $k>(k_c)_*$ (corresponding to the solid line) and unstable for perturbations with $k<(k_c)_*$. This critical wavenumber $(k_c)_*$ corresponds to a marginal perturbation where the phase velocity $\omega/k$ coincides with the {\it global maximum} $v_0=v_{-}$ of the velocity distribution.
\section{The case of plasmas}
\label{sec_plasmas}
We shall now compare the previous results obtained for self-gravitating systems to the case of plasmas.
\subsection{A brief historic}
\label{sec_paz}
The name plasma was introduced by Tonks \& Langmuir (1929) to describe
an ionised gas made of electrons and ions. They found that cold
plasmas oscillate with a natural pulsation $\omega_p=(4\pi\rho
e^2/m^2)^{1/2}$. In order to take into account thermal effects,
plasmas were initially described in terms of hydrodynamic equations. A
dispersion relation was derived by Thomson \& Thomson (1933) for
isothermal perturbations and by Gross (1951) for three dimensional
adiabatic perturbations. However, fluid equations do not provide a
good description of plasmas and rely on arbitrary assumptions
concerning the perturbations (see van Kampen 1957). Indeed, plasmas
are essentially collisionless so that a hydrodynamic description is
not clearly justified. Vlasov (1938,1945) was the first to attempt to
derive the dispersion relation directly from the collisionless
Boltzmann equation (nowdays called the Vlasov equation). He obtained
the expression of the pulsation $\omega(k)$ in the high wavelength
limit\footnote{The same result was derived in a different manner by
Bohm \& Gross (1949) and Jackson (1960).}. However, Landau (1946)
criticized his mathematical treatment showing that there are serious
divergences in the integrals considered by Vlasov. Landau performed a
rigorous mathematical study using an appropriate contour of
integration in the complex plane and showed that the plasma undergoes
damped oscillations. The pulsation is in agreement with the expression
given by Vlasov but, in addition, the plasma undergoes collisionless
damping (nowdays called Landau damping). Following Landau's seminal
work, several authors studied the stability of a plasma. They showed
that single-humped distributions are always stable (Berz 1956,
Appendix by W. Newcomb in Bernstein 1958, Auer 1958, Penrose 1960,
Noerdlinger 1960, Jackson 1960). Then, they considered double humped
distributions modeling two contrastreaming beams (Haeff 1949, Nergaard
1948, Pierce
\& Hebenstreit 1949) and determined particular conditions under which
such distributions are unstable (Twiss 1952, Buneman 1958, Auer 1958,
Kahn 1959, Buneman 1959, Penrose 1960, Noerdlinger 1960, Jackson
1960). They found qualitatively that the plasma becomes unstable when
the stream velocity becomes sensibly larger than the thermal velocity.
After briefly recalling some classical results, we will provide a detailed study of the stability/instability of a double-humped distribution made of two Maxwellians. We shall complete the study of a symmetric double-humped distribution (by explicitly computing the range of unstable wavelengths) and consider for the first time the case of asymmetric double-humped distributions. We will show that the nature of the problem changes above a critical asymmetry $\Delta_*=3.3105$.
\subsection{The Euler-Poisson system}
\label{sec_pep}
We consider a plasma made of electrons with mass $m$ and charge $-e$ and ions with mass $m_i$ and charge $+e$. We assume the ions to be infinitely massive so that they do not contribute to the motion and only provide a neutralizing background. The fluid equations describing the motion of the electrons can be written
\begin{equation}
{\partial\rho\over\partial t}+\nabla \cdot (\rho {\bf u})=0,
\label{pep1}
\end{equation}
\begin{equation}
{\partial {\bf u}\over\partial t}+({\bf u}\cdot \nabla) {\bf u}=-{1\over\rho}\nabla p-\nabla\Phi,
\label{pep2}
\end{equation}
\begin{equation}
\Delta\Phi=-\frac{4\pi e^2}{m^2}(\rho-\rho_i).
\label{pep3}
\end{equation}
These equations differ from the Euler-Poisson system (\ref{ep1})-(\ref{ep3}) describing a self-gravitating barotropic gas only in the sign of the interaction and in the presence of a neutralizing background. This neutralizing background avoids the Jeans swindle since an infinite homogeneous distribution of electrons ($\rho={\rm cst}$, ${\bf u}={\bf 0}$, $\Phi=0$) is a steady state of the fluid equations (\ref{pep1})-(\ref{pep3}) provided that the condition of electroneutrality $\rho=\rho_i$ is satisfied.
Linearizing the Euler-Poisson system around an infinite homogeneous distribution of electrons and decomposing the perturbations in normal modes of the form $e^{i({\bf k}\cdot {\bf r}-\omega t)}$, we obtain the dispersion relation (Nicholson 1992):
\begin{eqnarray}
\label{pep7}
\omega^2=c_s^2k^2+\omega_p^2,
\end{eqnarray}
where $c_s^2=p'(\rho)$ is the velocity of sound and
\begin{eqnarray}
\label{pep5}
\omega_{p}^{2}\equiv \frac{4\pi \rho e^{2}}{m^{2}},
\end{eqnarray}
is the plasma pulsation. For an isothermal equation of state $p=\rho T$, we get
\begin{eqnarray}
\label{pep8b}
\omega^2=T k^2+\omega_p^2,
\end{eqnarray}
and for a polytropic equation of state $p=K\rho^{\gamma}$, we obtain
\begin{eqnarray}
\label{pep8c}
\omega^2=\gamma T k^2+\omega_p^2.
\end{eqnarray}
The dispersion relation (\ref{pep7}) shows that a plasma described by fluid equations is always stable. For $k=0$, the pulsation tends to a constant
\begin{eqnarray}
\label{pep9}
\omega=\pm\omega_p.
\end{eqnarray}
Therefore, in this limit of small wavenumbers, plasma oscillations do not propagate in space: $\rho({\bf r},t)=\rho({\bf r},0)e^{-i\omega_p t}$. This behavior is in marked contrast with all usual types of oscillating phenomena such as acoustic waves for which $\omega=ck$. For large wavenumbers, the dispersion relation approaches that of sound waves: $\omega=\pm c_s k$.
{\it Remark:} Since $\int\delta\rho\delta\Phi\, d{\bf r}=\frac{m^2}{4\pi e^2}\int (\Delta\delta\Phi)^2\, d{\bf r}>0$ for a plasma, the second variations (\ref{ep10b}) of the energy functional (\ref{ep9}) are always positive: $\delta^2{\cal W}>0$. Therefore, the energy functional has a unique (global) minimum and the plasma is nonlinearly dynamically stable with respect to the barotropic Euler-Poisson system.
\subsection{The Vlasov-Poisson system}
\label{sec_pvp}
The fluid equations are not very appropriate to describe a plasma because most plasmas are in a regime where the collisions between charges are negligible. Therefore, from now on, we shall describe Coulombian plasmas by the Vlasov-Poisson system
\begin{equation}
\label{pvp1}
{\partial f\over\partial t}+{\bf v}\cdot {\partial f\over\partial {\bf r}}-\nabla\Phi\cdot {\partial f\over\partial {\bf v}}=0,
\end{equation}
\begin{equation}
\Delta\Phi=-\frac{4\pi e^2}{m^2}(\rho-\rho_i),
\label{pvp2}
\end{equation}
which ignores collisions between charges. These equations are similar to the Vlasov-Poisson system (\ref{vh0})-(\ref{vh1}) describing collisionless stellar systems except that the sign of the interaction is reversed and that there is a neutralizing background. We shall investigate the linear dynamical stability of an infinite homogeneous medium. Therefore, we consider stationary solutions of the Vlasov equation of
the form $f=f({\bf v})$. The dispersion relation can be written (Nicholson 1992):
\begin{eqnarray}
\epsilon(k,\omega)\equiv 1-{4\pi e^2\over m^2k^2}\int_{C} {{f'(v)}\over v-\frac{\omega}{k}}dv=0.
\label{pvp3}
\end{eqnarray}
As in Sec. \ref{sec_lsv}, we have taken ${\bf k}$ along the $z$-axis and noted $v$ for $v_z$ and $f(v)$ for $\int f({\bf v})\, dv_xdv_y$.
In general, the dispersion relation (\ref{pvp3}) cannot be solved explicitly to obtain $\omega(k)$ except in some very simple cases. For example, for cold systems described by the distribution function $f=\rho \delta({v}-{v}_0)$, we obtain after an integration by parts
\begin{eqnarray}
\omega=v_0 k\pm \omega_p.
\label{lsv8s}
\end{eqnarray}
In particular, when $v_0=0$, we get
\begin{eqnarray}
\omega=\pm \omega_p.
\label{lsv9s}
\end{eqnarray}
The system is stable to all wavenumbers. We also note that the dispersion relation (\ref{lsv9s}) coincides with the dispersion relation (\ref{pep7}) obtained with fluid equations with $c_s=0$.
Let us briefly recall important properties of the dispersion relation (\ref{pvp3}) obtained in plasma physics (Balescu 1963, Nicholson 1992). We consider a symmetric distribution and write the complex pulsation in the form $\omega=\omega_r+i\omega_i$. When $\omega_i\ll\omega_r$ (weakly damped perturbations), the complex pulsation is given by\footnote{A more accurate expression (Jackson 1960, Nicholson 1992) is obtained by replacing $f'(\omega_p/k)$ by $f'(\omega_r/k)$. This introduces a factor $e^{3/2}$ in the Landau formula (\ref{pvp5}).}
\begin{eqnarray}
\label{pvp4}
\omega_r^2=\omega_p^2+3Tk^2,\qquad \omega_{i}=\frac{\pi\omega_{p}^{3}}{2\rho k^2}f'\left (\frac{\omega_{p}}{k}\right ),
\end{eqnarray}
with $T=\langle v^2\rangle$ (where we recall that $v=v_z$ in the present case). These relations are valid for $k/k_D\ll 1$. The real part of the pulsation satisfies a dispersion relation of the form (\ref{fe17}) with $c_s^2=3T$. Therefore, large wavelengths perturbations in a collisionless plasma correspond to one dimensional isentropic perturbations with index $\gamma=3$ in a gas (see Appendix \ref{sec_fee}). This relation dispersion is called the Langmuir wave dispersion relation (Nicholson 1992). On the other hand, when $f'(\omega_p/k)<0$, the expression of the imaginary part of the pulsation corresponds to Landau damping. If $f'(v)\le 0$ for all $v\ge 0$, as for the Maxwellian, the plasma is stable. For the Maxwell distribution, we get the Landau formula
\begin{eqnarray}
\omega_{i}=-\sqrt{\frac{\pi}{8}}\omega_p \left (\frac{k_{D}}{k}\right )^{3}e^{-\frac{k_{D}^{2}}{2k^2}},
\label{pvp5}
\end{eqnarray}
for the damping rate, where we have introduced the Debye wavenumber
\begin{eqnarray}
\label{pep6}
k_{D}^{2}\equiv \frac{4\pi e^2\beta\rho}{m^2}=\beta \omega_{p}^{2},
\end{eqnarray}
where $\beta=1/T$ is the inverse temperature. Alternatively, if there exists wavenumbers $k$ such that $f'(\omega_p/k)>0$, the perturbation grows and the plasma is unstable. Since the expression (\ref{pvp4}) is valid for small $k$, this implies that $f'(v)$ must be positive for large velocities. This corresponds for example to the ``bump-on-tail'' situation analyzed in plasma physics (Nicholson 1992). The dispersion relation (\ref{pvp4})-a was first obtained by Vlasov (1938,1945). However, the derivation given by Vlasov was not rigorous since he evaluated the integral (\ref{pvp3}) on the real axis on which there is a singularity. Landau (1946) performed a rigorous mathematical study and showed that, in addition to the oscillations, the waves must be damped exponentially leading to equation (\ref{pvp4})-b for the damping rate. It is interesting to note that the Langmuir wave dispersion relation (\ref{pvp4})-a is rather independent on the precise form of the distribution function (it depends only on the variance $T=\langle v^2\rangle$) while the Landau damping (\ref{pvp4})-b is very sensitive to the distribution function. For example, the expression of Landau damping for polytropic distributions is given by (see Appendix E of Chavanis \& Delfini 2009):
\begin{eqnarray}
\omega_i=-\sqrt{\frac{\pi}{8}}B_n\omega_p\left (\frac{k_D}{k}\right )^3\left (n-\frac{1}{2}\right )\frac{1}{n+1}\nonumber\\
\times\left\lbrack 1-\frac{k_D^2}{2(n+1)k^2}\right\rbrack_+^{n-3/2}.
\label{lapol}
\end{eqnarray}
When $\omega_r\ll\omega_i$ (heavily damped perturbations), the pulsation is given for the Maxwellian distribution by
\begin{eqnarray}
\omega_r=\pm m\frac{\pi\omega_p}{2k_{D}}\frac{k}{\sqrt{\ln k}}, \qquad \omega_i=-\frac{2\omega_{p}}{k_{D}}k\sqrt{\ln k}.
\label{pvp6}
\end{eqnarray}
These expressions are valid for $k/k_D\rightarrow +\infty$. There
exists several branches of solutions parameterized by the odd integer
$m$ (but the branch $m=1$ is the most relevant). Expression
(\ref{pvp6}) of the complex pulsation was first obtained by Landau
(1946). The evolution of $\omega_r(k)$ and $\omega_i(k)$ with $k$ for
a Maxwellian distribution are represented in Fig. 3 of Jackson (1960)
and it can be compared with Fig. \ref{pulsation} for stellar
systems. For small $k$, the damping is weak but it increases rapidly
for increasing wavenumbers so that the waves are never observed at
sufficiently large wavenumbers.
{\it Remark:} Let us consider a plasma with a DF of the form $f=f(v^2)$ with $f'<0$. Since $\int\delta\rho\delta\Phi\, d{\bf r}>0$ for a plasma, the second variations (\ref{vh3b}) of the constrained pseudo entropy (\ref{vh2}) are always negative: $\delta^2{S}\le 0$. Therefore, the pseudo entropy has a unique (global) maximum at fixed mass and energy and the plasma is nonlinearly dynamically stable. These variational technics were introduced in the plasma literature by W. Newcomb in Bernstein (1958) and Gardner (1963); see also the review of Holm et al. (1985).
\subsection{The condition of marginal stability}
When $\omega_i=0$, the real and imaginary parts of the dielectric
function
$\epsilon(k,\omega_{r})=\epsilon_{r}(k,\omega_{r})+i\epsilon_{i}(k,\omega_{r})$
are
\begin{eqnarray}
\epsilon_{r}(k,\omega_{r})=1-{4\pi e^2\over m^2k^2}{P}\int_{-\infty}^{+\infty} {{f'(v)}\over v-\frac{\omega_{r}}{k}}dv,
\label{pny1}
\end{eqnarray}
\begin{eqnarray}
\epsilon_{i}(k,\omega_{r})=-{4\pi^2 e^2\over m^2k^2}f'(\omega_{r}/k).
\label{pny2}
\end{eqnarray}
The condition of marginal stability corresponds to $\epsilon(k,\omega)=0$ and $\omega_i=0$, i.e. $\epsilon_r({k},\omega_r)=\epsilon_i({k},\omega_r)=0$. We obtain therefore the equations
\begin{eqnarray}
1-\frac{4\pi e^2}{m^2k^2}P\int_{-\infty}^{+\infty} {f'(v)\over {v}-{\omega_r}/{k}}d{v}=0,
\label{cms1bp}
\end{eqnarray}
\begin{eqnarray}
f'\left ({\omega_r}/{k}\right )=0.
\label{cms2bp}
\end{eqnarray}
The second condition (\ref{cms2bp}) imposes that the phase velocity $\omega_r/k=v_{ext}$ is equal to a velocity where $f(v)$ is extremum ($f'(v_{ext})=0$). The first condition (\ref{cms1bp}) then determines the wavenumber(s) corresponding to marginal stability. It can be written
\begin{eqnarray}
k_{c}=\left (\frac{4\pi e^2}{m^2}\int_{-\infty}^{+\infty} \frac{f'(v)}{v-v_{ext}}\, dv\right )^{1/2}.
\label{cms4p}
\end{eqnarray}
Finally, the pulsation(s) corresponding to marginal stability are $\omega_r=v_{ext}k_c$ and the perturbation behaves like $\delta f\sim e^{-i\omega_r t}$.
\subsection{Particular solutions of $\epsilon(k,\omega)=0$}
\label{sec_monbis}
We can look for a solution of the dispersion relation
$\epsilon(k,\omega)=0$ in the form $\omega=i\omega_{i}$ corresponding to
$\omega_{r}=0$. In that case, the perturbation grows ($\omega_{i}>0$)
or decays ($\omega_{i}<0$) without oscillating. Like in Sec. \ref{sec_mon}, we shall assume that
$f(v)$ is an even function. For $\omega_i>0$, the growth rate is given by
\begin{eqnarray}
1-{4\pi e^2\over m^2 k^{2}}\int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+\frac{\omega_{i}^2}{k^2}}\, dv=0.
\label{mon2plas}
\end{eqnarray}
For $\omega_{i}<0$, the decay rate is given by
\begin{eqnarray}
1-{4\pi e^2\over m^2 k^{2}}\int_{-\infty}^{+\infty}\frac{v f'(v)}{v^2+\frac{\omega_{i}^2}{k^2}}\, dv-i \frac{8\pi^2 e^2}{m^2k^2} f'\left (\frac{i\omega_{i}}{k}\right )=0.
\label{mon5bb}
\end{eqnarray}
Let us assume that the distribution satisfies
\begin{eqnarray}
\int_{-\infty}^{+\infty}\frac{f'(v)}{v}\, dv>0.
\label{mon2plasb}
\end{eqnarray}
This is not the generic case. For example, the Maxwell distribution does not satisfy this inequality. If inequality (\ref{mon2plasb}) is satisfied, then the marginal mode $\omega=0$ corresponds to the critical wavenumber
\begin{eqnarray}
k_c=\left ({4\pi e^2\over m^2}\int_{-\infty}^{+\infty}\frac{f'(v)}{v}\, dv\right )^{1/2}.
\label{mon2plasc}
\end{eqnarray}
Furthermore, repeating the steps of Sec. \ref{sec_mon}, we obtain for $k\rightarrow k_c$ that\footnote{In the derivation, we have assumed that $f''(0)\neq 0$. If $f''(0)=0$, we need to develop the Taylor expansion to the next order.}
\begin{eqnarray}
\omega_i=\frac{k_c^3m^2}{4\pi^2 e^2f''(0)}\left (1-\frac{k^2}{k_c^2}\right ), \qquad (k\rightarrow k_c).
\label{mon15b}
\end{eqnarray}
This formula leads to the following result. If the distribution is maximum at $v=0$, so that $f''(0)<0$, we find that the mode $\omega=i\omega_i$ is stable for $k<k_c$ and unstable for $k>k_c$. Alternatively, if the distribution is minimum at $v=0$, so that $f''(0)>0$, we find that the mode $\omega=i\omega_i$ is stable for $k>k_c$ and unstable for $k<k_c$. This result will be illustrated in connection to Fig. \ref{PLASMAphasediag} for the symmetric double humped distribution. Note that this result implies that a symmetric distribution satisfying inequality (\ref{mon2plasb}) is always unstable (to some wavenumbers).
\subsection{The Nyquist method}
\label{sec_pny}
To apply the Nyquist method\footnote{Nyquist (1932) introduced his graphical method in relation to servomechanism theory. His method was first applied to plasma physics by Harris (1959), Penrose (1960) and Jackson (1960).}, we have to plot the curve
$(\epsilon_{r}(k,\omega_{r}),\epsilon_{i}(k,\omega_{r}))$ parameterized by
$\omega_{r}$ going from $-\infty$ to $+\infty$ (for a fixed wavenumber $k$). Let us consider the
asymptotic behavior for $\omega_{r}\rightarrow \pm\infty$. Since $f(v)$ is positive
and tends to zero for $v\rightarrow \pm\infty$, we conclude that
$\epsilon_{i}(k,\omega_{r})\rightarrow 0$ for $\omega_{r}\rightarrow
\pm\infty$ and that $\epsilon_{i}(k,\omega_{r})<0$ for
$\omega_{r}\rightarrow -\infty$ while $\epsilon_{i}(k,\omega_{r})>0$ for
$\omega_{r}\rightarrow +\infty$. On the other hand, for
$\omega_{r}\rightarrow \pm\infty$, we obtain at leading order
\begin{eqnarray}
\epsilon_{r}(k,\omega_{r})\simeq 1-{4\pi^2 e^2\rho\over m^2\omega_{r}^{2}}, \qquad (\omega_{r}\rightarrow \pm\infty).
\label{pny3}
\end{eqnarray}
From these results, we conclude that the behavior of the curve close to the point $(1,0)$ is like the one represented in Fig. \ref{maxwellP}. For $\omega_r/k=v_{ext}$, where
$v_{ext}$ is a velocity at which the distribution is extremum
$(f'(v_{ext})=0)$, the imaginary part of the dielectric function
$\epsilon_i(k,k v_{ext})=0$ and the real part of the dielectric function
\begin{equation}
\epsilon_r(k,k v_{ext})= 1-{4\pi^2 e^2\over m^2k^2}\int_{-\infty}^{\infty}\frac{f'(v)}{v-v_{ext}}\, dv.
\end{equation}
Subtracting the value $f'(v_{ext})=0$ in the numerator of the integrand, and integrating by parts, we obtain
\begin{eqnarray}
\epsilon_{r}(k,k v_{ext})=1+{4\pi^2 e^2\over m^2k^2}\int_{-\infty}^{+\infty} \frac{f(v_{ext})-f(v)}{(v-v_{ext})^{2}}\, dv.
\label{pny4}
\end{eqnarray}
If $v_{Max}$ denotes the velocity corresponding to the global maximum of the distribution, we clearly have
\begin{eqnarray}
\epsilon_{r}(k,k v_{Max})> 1.
\label{pny5}
\end{eqnarray}
\subsubsection{Single-humped distributions}
\label{sec_pshr}
Let us assume that the distribution $f(v)$ has a single maximum at
$v=v_{0}$ (so that $f'(v_0)=0$) and tends to zero for $v\rightarrow
\pm \infty$. Then, the Nyquist curve cuts the $x$-axis ($\epsilon_{i}(k,\omega_{r})$ vanishes) at the limit point
$(1,0)$ when $\omega_{r}\rightarrow \pm \infty$ and at the point
$(\epsilon_{r}(k,kv_0),0)$ when $\omega_{r}/k=v_{0}$. Due to its behavior
close to the limit point $(1,0)$, the fact that it rotates in the
counterclockwise sense, and the property that $\epsilon_r(k,k v_0)>1$ according to Eq. (\ref{pny5}), the Nyquist curve must necessarily behave like in
Fig. \ref{maxwellP}. Therefore, the Nyquist curve starts on the
real axis at $\epsilon_r(k,\omega_r) =1$ for $\omega_r \rightarrow
-\infty$, then going in counterclockwise sense it crosses the real
axis at the point $\epsilon_r(k,k v_0)>1$ and returns on the real axis at
$\epsilon_r(k, \omega_r) =1$ for $\omega_r \rightarrow +
\infty$. Therefore, it cannot encircle the origin. According to the Nyquist criterion exposed in
Sec. \ref{sec_n}, we conclude that a single-humped distribution
is always linearly dynamically stable\footnote{The fact that a single-humped distribution is always stable is known as Gardner (1963)'s theorem; see also Jackson (1960).}.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{maxwellP.eps}
} \caption[]{Nyquist curve for the Maxwell distribution. The DF is always stable.}
\label{maxwellP}
\end{figure}
Let us specifically consider the Maxwell distribution (\ref{md1}) for illustration. The dielectric function (\ref{pvp3}) associated to the Maxwell distribution is
\begin{eqnarray}
\epsilon(k,\omega)=1+\frac{4\pi e^2}{m^2k^2}\left (\frac{\beta}{2\pi}\right )^{1/2}\rho\int_{C}\frac{\beta v}{v-\frac{\omega}{k}}e^{-\beta {v^{2}\over 2}}\ dv.
\label{Pmd3}
\end{eqnarray}
Introducing the Debye wavenumber (\ref{pep6}), it can be rewritten
\begin{eqnarray}
\epsilon({k},\omega)=1+{k_{D}^{2}\over k^{2}} W\biggl ({\sqrt{\beta}\omega\over k}\biggr ).
\label{Pmd4}
\end{eqnarray}
It will be convenient in the following to work with dimensionless quantities.
We introduce the dimensionless wavenumber and the dimensionless pulsation
\begin{eqnarray}
\eta=\frac{k_D^2}{k^2}, \qquad \Omega=\frac{\omega}{\omega_p}.
\label{Pmn1}
\end{eqnarray}
Noting that $\sqrt{\beta}\omega/k=\sqrt{\eta}\Omega$, the dielectric function (\ref{Pmd4}) can be rewritten
\begin{eqnarray}
\epsilon({\eta},\Omega)=1+\eta W(\sqrt{\eta}\Omega).
\label{Pmn3}
\end{eqnarray}
When $\Omega_{i}=0$, the real and imaginary parts of the dielectric function $\epsilon(\eta,\Omega_{r})=\epsilon_{r}(\eta,\Omega_{r})+i\epsilon_{i}(\eta,\Omega_{r})$ are given by
\begin{equation}
\label{Pmn4}
\epsilon_{r}(\eta,\Omega_{r})=1+\eta W_{r}\left (\sqrt{\eta}\Omega_r\right ),
\end{equation}
\begin{equation}
\label{Pmn5}
\epsilon_{i}(\eta,\Omega_{r})=\eta W_{i}\left (\sqrt{\eta}\Omega_r\right ).
\end{equation}
The Nyquist curve for the Maxwell distribution is represented in Fig. \ref{maxwellP}.
On the other hand, if we consider a polytrope of index $n=1/2$ (waterbag distribution) as in Sec. \ref{sec_fermi}, we find that the dielectric function is given by
\begin{eqnarray}
\epsilon(k,\omega)=1+\frac{k_D^2}{3k^2}W_{1/2}\left (\frac{\omega}{k\sqrt{T}}\right ),
\label{dielanew}
\end{eqnarray}
where $W_{1/2}(x)$ is given by Eq. (\ref{dielb}).
The condition $\epsilon(k,\omega)=0$ determines the dispersion relation. We get
\begin{eqnarray}
\omega^2=\omega_p^2+3Tk^2.
\label{dielcnew}
\end{eqnarray}
We note that, for the water-bag distribution, the general asymptotic
behavior (\ref{pvp4}) becomes exact. We also note that for the
specific index $n=1/2$ ($\gamma=3$), the dispersion relation in a
collisionless stellar system takes the same form as in a gas (see
Sec. \ref{sec_pep}).
\subsubsection{Double-humped distributions}
\label{sec_dhhr}
Let us consider a double-humped distribution with a global maximum at
$v_{Max}$, a minimum at $v_{min}$ and a local maximum at $v_{max}$. We
assume $v_{Max}<v_{min}<v_{max}$. The Nyquist curves starts at
$(1,0)$, progresses in the counterclockwise sense and crosses the
$x$-axis at $\epsilon_{r}(v_{Max})>1$, then at $\epsilon_{r}(v_{min})$
and $\epsilon_{r}(v_{max})$. We can convince ourselves by making drawings of the following results. If
$(+++)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})>0$, $\epsilon_{r}(v_{max})>0$,
$(+--)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})<0$, $\epsilon_{r}(v_{max})<0$,
$(++-)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})>0$, $\epsilon_{r}(v_{max})<0$,
\noindent the Nyquist curve does not encircle the origin so the
system is stable. If
$(+-+)$: $\epsilon_{r}(v_{Max})>0$, $\epsilon_{r}(v_{min})<0$, $\epsilon_{r}(v_{max})>0$,
\noindent the Nyquist curve rotates one time
around the origin so that there is one mode of instability. Since
$\epsilon_{r}(v_{Max})>0$ there is no mode of marginal stability with
$\omega_{r}/k=v_{Max}$. Cases $(+++)$, $(+-+)$ and $(+--)$ are
observed in Sec. \ref{sec_rass} for an asymmetric double-humped
distribution made of two Maxwellians.
If the double-humped distribution is symmetric with respect to the
origin with two maxima at $\pm v_{*}$ and a minimum at $v=0$, we get
the same results as above with the additional properties
$\epsilon_{r}(v_{Max})=\epsilon_{r}(v_{max})=\epsilon_{r}(v_{*})>1$
and $\epsilon_{r}(v_{min})=\epsilon_{r}(0)$. We have only two cases
$(+++)$ and $(+-+)$. They are observed in Sec. \ref{sec_rsh} for a
symmetric double-humped distribution made of two Maxwellians. Since
$\epsilon_{r}(v_{*})>0$, there is no mode of marginal stability with
$\omega_{r}/k=\pm v_{*}$.
It can be shown that a plasma is unstable (to some wavelengths) iff $f(v)$ has a minimum at a value $v=v_{min}$ such that
\begin{eqnarray}
\int_{-\infty}^{+\infty} {{f'(v)}\over v-v_{min}}dv=\int_{-\infty}^{+\infty} \frac{f(v)-f(v_{min})}{(v-v_{min})^{2}}\, dv>0.
\label{penrose}
\end{eqnarray}
This result was proven by Penrose (1960) and it is sometimes called the Penrose criterion\footnote{In fact, we found that an equivalent criterion was obtained at the same period by Noerdlinger (1960).}.
This criterion can be deduced from the Nyquist method as follows. A double-humped distribution $f(v)$ is unstable if there exists a range of $k$ such that we are in the situation $(+-+)$, i.e. $\epsilon_r(v_{min})<0$ and $\epsilon_r(v_{max})>0$. The first condition can be satisfied (for sufficiently small $k$) iff condition (\ref{penrose}) is realized, which is the Penrose criterion. Then, the range of unstable wavenumbers is
\begin{eqnarray}
\frac{4\pi e^2}{m^2}\int_{-\infty}^{+\infty} {{f'(v)}\over v-v_{max}}dv<k^2<\frac{4\pi e^2}{m^2}\int_{-\infty}^{+\infty} {{f'(v)}\over v-v_{min}}dv.\nonumber\\
\label{penroserange}
\end{eqnarray}
If the first integral is negative, the range of unstable wavenumbers is $k<k_c^{(min)}$. This corresponds to Fig. \ref{pdg1} and to the region $y>y_c$ in Fig. \ref{pdg2}. If the first integral is positive, the range of unstable wavenumbers is $k_c^{(max)}<k<k_c^{(min)}$. This corresponds to the region $y_*<y<y_c$ in Fig. \ref{pdg2} (indeed, $y_c$ corresponds precisely to the case where the first integral becomes equal to zero).
\subsection{The symmetric double-humped distribution}
\label{sec_rsh}
Let us now consider the symmetric double-humped distribution (\ref{e1}). In the plasma case, the dielectric function can be written
\begin{eqnarray}
\label{rsh1}
\epsilon(\eta,\Omega)= 1+\frac{\eta}{2} \left [ W(\sqrt{\eta}\Omega-\sqrt{y}))+W(\sqrt{\eta}\Omega+\sqrt{y})) \right ],\nonumber\\
\end{eqnarray}
where $\eta$ (wavelength) and $\Omega$ (pulsation) are defined by Eq. (\ref{Pmn1}) and $y$ (separation) is defined by Eq. (\ref{e5}).
The condition of marginal stability corresponds to $\epsilon(k,\omega)=0$ and $\omega_i=0$. The condition $\epsilon_i(k,\omega_r)=0$ is equivalent to $f'(\omega_r/k)=0$ so that the phase velocity $\omega_r/k=v_0$ corresponds to the velocities where the distribution is extremum. Then, the wavenumbers at which the distribution is marginally stable are obtained by solving $\epsilon_r(k,\omega_r=k v_0)=0$. Proceeding as in Sec. \ref{sec_vhq} and introducing the parameter $x=V_0/V_a$, we find that the equations determining the critical wavenumbers
$\eta_c(y)$ are given by
\begin{equation}
\label{rsh2}
y=\frac{1}{2x}\ln\left (\frac{1+x}{1-x}\right ),
\end{equation}
\begin{equation}
\label{rsh3}
\eta=\frac{-2}{\left [ W_{r}(\sqrt{y}(x - 1))+W_{r}(\sqrt{y}(x+ 1)) \right ]}.
\end{equation}
We note that only the sign in Eq. (\ref{rsh3}) changes with respect to the study of the gravitational case, so we can readily adapt the results of Sec. \ref{sec_vhq} to the present situation by simply reverting the sign. For $\sqrt{\eta}\Omega_{r}=\pm V_*$, corresponding to $x\neq 0$, there is no physical solution to Eqs. (\ref{rsh2})-(\ref{rsh3}) with positive $\eta>0$. Therefore, in the plasma case, there is no marginal mode with non zero pulsation for the symmetric double-humped distribution (in agreement with the general discussion of Sec. \ref{sec_dhhr}). We now consider the marginal mode with $\Omega_{r}=0$. This corresponds to the ``degenerate" solution $x=0$ (for any $y$) for which Eqs. (\ref{rsh2})-(\ref{rsh3}) reduce to
\begin{equation}
\label{rsh4}
\eta_c^{(0)}(y)=\frac{-1}{W_{r}(\sqrt{y})}.
\end{equation}
According to Fig. \ref{w3}, physical solutions exist only for $y\ge
y_{max}=z_{c}^{2}=1.708$. We note that the range of parameters that was
forbidden in the gravitational case is now allowed in the plasma case
and vice versa. For $y<y_{max}$, the system is stable. This result is to be expected
since, for $y=0$, the distribution (\ref{e1}) reduces to the Maxwellian that is stable
for a repulsive interaction. We now consider the range $y\in \lbrack
y_{max},+\infty\lbrack$.
For $y\rightarrow y_{max}$, we have
\begin{equation}
\label{rsh5}
\eta_{c}^{(0)}(y)\sim \frac{2y_{max}}{y-y_{max}}, \qquad (y\rightarrow y_{max}).
\end{equation}
On the other hand, the curve $\eta_{c}^{(0)}(y)$ has a minimum at $(4.511,3.512)$ (see Appendix A of Chavanis \& Delfini 2009). Finally,
for $y\rightarrow +\infty$, we have
\begin{equation}
\label{rsh6}
\eta_{c}^{(0)}(y)\sim y\rightarrow +\infty, \qquad (y\rightarrow +\infty).
\end{equation}
Again this result is expected because, for $y\rightarrow
+\infty$, the two humps do not ``see" each other and behave as two
independent single-humps distributions that are stable in the repulsive
case.
The critical wavenumber $\eta_c^{(0)}(y)$ corresponding to marginal stability determined previously is represented as a function of the separation $y$ in Fig. \ref{PLASMAphasediag}. We have also plotted the line $y=1$. On the left of this line, the distribution has a single maximum at $V_0=0$ and on the right, the distribution has two maxima at $V_0=\pm V_*$ and a minimum at $V_0=0$. In order to investigate the stability of the solutions in the different regions, we have used the Nyquist criterion.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{PLASMAphasediag.eps}
} \caption[]{Stability diagram of the symmetric double-humped
distribution (\ref{e1}) in the case of plasmas.}
\label{PLASMAphasediag}
\end{figure}
For $y<1$, the DF has only one maximum at $V_0=0$. There is no marginal mode and the distribution is always
stable (see Fig. \ref{doublesym_y0.5_n6P}).
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y0.5_n6P.eps}
} \caption[]{Nyquist curve for $y<1$ (specifically $y=0.5$ and $\eta=6$). The DF is stable for any perturbation. Case (+).}
\label{doublesym_y0.5_n6P}
\end{figure}
For $1<y<y_{max}$, the DF has a minimum at $V_0=0$ and two maxima at $\pm V_*$. There is no marginal mode and the distribution is always
stable (see Figs. \ref{doublesym_y1.2_n6P}).
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y1.2_n6P.eps}
} \caption[]{Nyquist curve for $1<y<y_{max}$ (specifically $y=1.2$ and $\eta=6$). The DF is stable for any perturbation. Case (+ + +).}
\label{doublesym_y1.2_n6P}
\end{figure}
For $y>y_{max}$, the distribution has
a minimum at $V_0=0$ and two maxima at $\pm V_*$. There exists one wavenumber $\eta_{c}^{(0)}$ at which
the distribution is marginally stable. For $\eta=\eta_{c}^{(0)}$, the marginal
perturbation does not propagate ($\Omega_r=0$). By considering the
Nyquist curves in this region (see Figs. \ref{doublesym_y5_n2P}-\ref{doublesym_y5_n6P}), we
find that the DF is stable for $\eta<\eta_{c}^{(0)}$ and unstable
for $\eta>\eta_{c}^{(0)}$.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y5_n2P.eps}
} \caption[]{Nyquist curve for $y>y_{max}$ and $\eta<\eta_{c}^{(0)}$ (specifically $y=5$ and $\eta=2$). The DF is stable for this range of wavenumbers. Case (+ + +).}
\label{doublesym_y5_n2P}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{doublesym_y5_n6P.eps}
} \caption[]{Nyquist curve for $y>y_{max}$ and $\eta>\eta_{c}^{(0)}$ (specifically $y=5$ and $\eta=6$). There is one mode of instability in this range of wavenumbers. Case (- + -).}
\label{doublesym_y5_n6P}
\end{figure}
Let us make some remarks:
1. If we consider all possible perturbations, Fig. \ref{PLASMAphasediag} shows that a symmetric double-humped distribution is stable if $y<y_{max}$ (i.e. $v_a<1.307 \, T^{1/2}$) and unstable if $y>y_{max}$ (i.e. $v_a>1.307 \, T^{1/2}$). Therefore, instability occurs when the drift velocity $v_a$ is sensibly larger than the thermal speed $\sqrt{T}$ so that the two humps are well separated. In particular, at $T=0$, a double-humped distribution is always unstable to some wavelengths (see Appendix \ref{sec_cold}). More precisely, a symmetric double-humped distribution with $y>y_{max}$ is stable for perturbations with wavenumbers $k>(k_c)_*$ (corresponding to the solid line) and unstable for perturbations with wavenumbers $k<(k_c)_*$. The critical wavenumber $(k_c)_*$ corresponds to the marginal perturbation for which the phase velocity $\omega/k$ coincides with the {\it minimum} of the velocity distribution: $v_0=0$.
2. For $y>y_{max}$, the mode $\omega=i\omega_i$ is stable for $\eta<\eta_c^{(0)}$ and it becomes unstable when we increase $\eta$ above the marginal line $\eta=\eta_c^{(0)}$. This is consistent with the general result (\ref{mon15b}) since, for $y>1$, the DF is {\it minimum} at $v_0=0$.
3. For a double-humped distribution, the critical wavelength is infinite for $v_a\le \sqrt{y_{max}T}$ (stable), then decreases, reaches a minimum and increases again. By contrast, if the plasma is modeled as a contrastreaming gas, the critical wavelength is infinite for $v_a< c_s$ (stable), then jumps discontinuously to zero at $v_a=c_s$ and increases for $v_a>c_s$ (Ikeuchi et al. 1974).
\subsection{The asymmetric double-humped distribution}
\label{sec_rass}
The dielectric function associated to the asymmetric double-humped
distribution (\ref{ae1}) in the plasma case is
\begin{eqnarray}
\label{rass1}
\epsilon(\eta,\Omega)= 1+\frac{\eta}{1+\Delta} \biggl [ W(\sqrt{\eta}\Omega - \sqrt{y}))\nonumber\\
+\Delta W(\sqrt{\eta}\Omega + \sqrt{y})) \biggr ].
\end{eqnarray}
Proceeding as in Sec. \ref{sec_vha} and introducing the parameters
$x=V_0/V_a$ and $y=V_a^2$, we find that the equations determining the
critical wavenumbers $\eta_c(y)$ are given by
\begin{equation}
\label{rass2}
y=\frac{1}{2x}\ln\left (\frac{1+x}{1-x}\right
)+\frac{\ln(\Delta)}{2x},
\end{equation}
\begin{equation}
\label{rass3}
\eta=-\frac{1+\Delta}{\left [ W_{r}(\sqrt{y}(x - 1))+\Delta W_{r}(\sqrt{y}(x+ 1)) \right ]}.
\end{equation}
Equation (\ref{rass2}) determines the extrema of
the distribution $f(v)$ and Eq. (\ref{rass3}) determines the wavenumbers
corresponding to the modes of marginal stability. As in
Sec. \ref{sec_vha}, the curve $\eta_c(y)$ can be obtained by varying
$x$ between $-1$ and $+1$. In the plasma case, there exists
physical solutions with positive $\eta$ only for $0<x\le x_c$.
For $x\rightarrow 0^+$, using Eq. (\ref{rass2}), we find that
$y\rightarrow +\infty$. Then, we get
\begin{equation}
\label{rass4}
\eta_c^{(p)}(y)\sim y\rightarrow +\infty, \qquad (y\rightarrow +\infty).
\end{equation}
For $x\rightarrow x_c$, we find that $y\rightarrow y_c$ and
\begin{equation}
\label{rass5}
\eta_c^{(s)}(y)\propto \frac{1}{y-y_c}\rightarrow +\infty, \qquad (y\rightarrow y_c),
\end{equation}
where $s=p$ if $\Delta<\Delta_{*}$ and $s=+$ if
$\Delta>\Delta_{*}$. For $\Delta<\Delta_{*}$, we get only one marginal branch
$\eta_{c}^{(p)}(y)$ corresponding to the mode $\Omega_r=V_{p}/\sqrt{\eta_c}>0$ (see
Fig. \ref{pdg1}). For $\Delta>\Delta_{*}$, we get two marginal
branches $\eta_{c}^{(p)}(y)$ and $\eta_{c}^{(+)}(y)$ corresponding to
the modes $\Omega_r=V_{p}/\sqrt{\eta_c}>0$ and $\Omega_r=V_{+}/\sqrt{\eta_c}>0$ (see
Fig. \ref{pdg2}). They connect each other at $(y_{*},\eta_{*})$
corresponding to $x=x_{*}$. At that point they touch the line
$y=y_{*}$ separating distributions with one or two maxima.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{PLASMAphasediagd1.02.eps}
} \caption[]{Stability diagram of the asymmetric double-humped distribution for $\Delta<\Delta_*$ (specifically $\Delta=1.02$). There exists one marginal branch $\eta_c^{(p)}(y)$.}
\label{pdg1}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{PLASMAphasediagd50.eps}
} \caption[]{Stability diagram of the asymmetric double-humped distribution for $\Delta>\Delta_*$ (specifically $\Delta=50$). There exists two marginal branches $\eta_c^{(p)}(y)$ and $\eta_c^{(+)}(y)$.}
\label{pdg2}
\end{figure}
The stability diagrams corresponding to the asymmetric double-humped
distribution with $\Delta<\Delta_*$ and $\Delta>\Delta_*$ are represented in
Figs. \ref{pdg1} and \ref{pdg2}. To investigate the stability of the solutions in the different regions of the parameter space, we have used the Nyquist criterion. If $\Delta<\Delta_*$, the description of the stability diagram can be made in continuity with the case of a symmetric double-humped distribution: for $y<y_{c}$ (i.e. $v_a^2<y_c(\Delta)T$), the plasma is stable to all perturbations; for $y>y_{c}$ (i.e. $v_a^2>y_c(\Delta)T$), the plasma is stable for $k>k_c^{(p)}(v_a)$ and unstable for $k<k_c^{(p)}(v_a)$. If $\Delta>\Delta_*$, the stability diagram changes: for $y<y_{*}$ (i.e. $v_a^2<y_*(\Delta)T$), the plasma is stable to all perturbations; for $y_{*}<y<y_c$ (i.e. $y_*(\Delta)T<v_a^2<y_c(\Delta)T$), the plasma is stable for $k>k_c^{(p)}(v_a)$, unstable for $k_c^{(+)}(v_a)<k<k_c^{(p)}(v_a)$ and stable for $k<k_c^{(+)}(v_a)$: this is similar to a re-entrant phase; for $y>y_c$ (i.e. $v_a^2>y_c(\Delta)T$), the plasma is stable for $k>k_c^{(p)}(v_a)$ and unstable for $k<k_c^{(p)}(v_a)$. In conclusion, instability occurs when the drift velocity is sensibly larger than the thermal speed so that the two humps are well separated. However, the precise threshold changes whether the asymmetry is smaller or larger than $\Delta_*$. For $\Delta<\Delta_*$, the plasma is stable iff $v_a^2<y_c(\Delta)T$ and for $\Delta>\Delta_*$, the plasma is stable iff $v_a^2<y_*(\Delta)T$.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta50_y5_n5.eps}
} \caption[]{$\Delta>\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta<\eta_c^{(p)}$
(specifically $\Delta=50$, $y=5$ and $\eta=5$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is stable (with respect to
this perturbation) because the Nyquist curve does not encircle the
origin. Case (+ + +).}
\label{data_delta50_y5_n5}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta50_y5_n15.eps}
} \caption[]{$\Delta>\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta_c^{(p)}<\eta<\eta_c^{(+)}$
(specifically $\Delta=50$, $y=5$ and $\eta=15$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is unstable (with respect to
this perturbation) because the Nyquist curve encircles the
origin once. There is $N=1$ unstable mode. Case (+ - +).}
\label{data_delta50_y5_n15}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,angle=0]{data_delta50_y5_n30.eps}
} \caption[]{$\Delta>\Delta_*$: Nyquist curve for $y_*<y<y_c$ and $\eta>\eta_c^{(+)}$
(specifically $\Delta=50$, $y=5$ and $\eta=30$). The DF has a global maximum at $V_{-}$, a minimum at $V_p$ and a local maximum at $V_{+}$. The DF is stable (with respect to
this perturbation) because the Nyquist curve does not encircle the
origin. Case (+ - -).}
\label{data_delta50_y5_n30}
\end{figure}
\section{Conclusion}
We have carried out an exhaustive study of the linear dynamical
stability/instability of an infinite and homogeneous self-gravitating
medium with respect to perturbations with wavenumber $k$. This is the
classical Jeans problem that describes the growth of perturbations and
the emergence of structures (like stars and galaxies) in astrophysics
and cosmology. We have considered the case of a collision-dominated
gas described by the Euler equation or the case of a collisionless
stellar system described by the Vlasov equation. In the latter case,
we have studied single-humped distributions as well as symmetric and
asymmetric double-humped distributions. We have used the Nyquist
theorem to settle the stability of the system. Detailed stability
diagrams in the $(v_a,k)$ plane for different values of the asymmetry
parameter $\Delta$ have been obtained. We have also derived general
analytical results concerning the dispersion relation. Finally, we
have studied the case of Coulombian plasmas in parallel. The general
methods developed here to study the stability of a homogeneous
self-gravitating medium can have applications in other domains, not
necessarily astrophysics, where the system is described by the Euler
or the Vlasov equations coupled to a mean field equation for the
potential. One example is the HMF model (Chavanis \& Delfini 2009).
Although the structure of the problem is the same, the results and the
stability diagrams differ because the potential of interaction is
different. Many generalizations of our work are possible, changing
the potential of interaction or the class of distribution functions,
but the present paper provides many methods and analytical results
that can be useful to investigate more general situations.
|
\section{Introduction}\label{intro}
It has recently been proved by Lott in \cite{Ltt07} that if $g(t)$ is a Ricci flow
solution on a $3$-dimensional compact manifold $M$, with sectional curvatures that
are $\Order(t^{-1})$ and diameter that is $\Order(\sqrt{t})$, then the pullback
Ricci flow solution on the simply connected cover $\tilde{M}$ approaches a
homogeneous expanding Ricci soliton. Among many others, this is certainly a good
motivation to study Ricci solitons in the homogeneous case. A natural question we
are particularly interested in is how much stronger is, for homogeneous metrics, the
Einstein condition compared with the condition of being a Ricci soliton.
From results due to Ivey, Naber, Perelman and Petersen-Wylie, it follows that any
nontrivial homogeneous Ricci soliton must be noncompact, expanding and non-gradient
(see Section \ref{hrs}). Up to now, all known examples are isometric to a
left-invariant metric $g$ on a simply connected Lie group $G$, which when identified
with an inner product on the Lie algebra $\ggo$ of $G$ satisfies
\begin{equation}\label{intro1}
\Ricci(g)=cI+D, \qquad\mbox{for some}\quad c\in\RR,\quad D\in\Der(\ggo),
\end{equation}
where $\Ricci(g)$ is the Ricci operator of $g$. On the other hand the converse is
true: any left invariant metric which satisfies (\ref{intro1}) is automatically a
Ricci soliton. For $G$ nilpotent, these metrics are called {\it nilsolitons} and
have been extensively studied in the last decade, mainly because of the strong
connection with Einstein solvmanifolds (see the survey \cite{cruzchica}). Examples
with $G$ solvable but non-nilpotent have explicitly appeared in \cite{BrdDnl}
($\dim{G}=3$) and \cite{IsnJckLu} ($\dim{G}=4$).
The aim of this paper is to study the structure of solvable Lie groups admitting a
left invariant metric for which (\ref{intro1}) holds; these metrics will be called
{\it solsolitons} from now on. The tools from geometric invariant theory used in
\cite{standard} to prove that any Einstein solvmanifold is standard (see Section
\ref{Tb}), turned out to be useful in the study of solsolitons as well.
It is really easy to get examples of solsolitons from only a nilsoliton and an
abelian Lie algebra of symmetric derivations of its metric Lie algebra. Our main
result asserts that any solsoliton can actually be obtained (up to isometry) by such
a simple construction (see Section \ref{structure}). In particular, any solsoliton
is standard, and if not Einstein, it admits a one-dimensional extension which is an
Einstein solvmanifold, just as for nilsolitons. We are therefore showing that most
of the structural results proved for Einstein solvmanifolds in \cite{Hbr,standard}
are still valid for solsolitons. We also obtain a uniqueness result that
generalizes the known results for Einstein solvmanifolds and nilsolitons: among all
left invariant metrics on a given solvable Lie group, there is at most one
solsoliton up to isometry and scaling (see Section \ref{uniq}). All this is used in
Section \ref{exa} to classify solvable Lie groups admitting solsolitons in dimension
$\leq 4$.
\vs \noindent {\it Acknowledgements.} Part of this research was performed while
attending the Borel trimester `Ricci curvature and Ricci flow'. I am very grateful
to Institut Poincar\'e for their hospitality and to the organizers for inviting me.
I would also like to thank Luca Di Cerbo and Peng Lu for fruitful discussions on the
topic of this paper.
\section{Homogeneous Ricci solitons}\label{hrs}
A complete Riemannian metric $g$ on a differentiable manifold $M$ is said to be a
{\it Ricci soliton} if its Ricci tensor satisfies
\begin{equation}\label{defrs}
\ricci(g)=cg+\Lie_Xg, \qquad\mbox{for some}\quad c\in\RR, \quad
X\in\chi(M)\;\mbox{complete},
\end{equation}
where $\chi(M)$ denotes the space of all differentiable vector fields on $M$ and
$\Lie_X$ the Lie derivative (see \cite{libro} for further information on Ricci
solitons). Recall that if $\theta_t$ is the one-parameter group associated to $X$
then $\Lie_Xg=\ddt|_0\theta_t^*g$, and hence the Ricci soliton condition may be
rephrased as follows: $\ricci(g)$ is tangent at $g$ to the space of all metrics
which are {\it homothetic} (i.e. isometric up to scaling) to $g$. If in addition
$X$ is the gradient field of a smooth function $f:M\longrightarrow\RR$, then
(\ref{defrs}) becomes $\ricci(g)=cg+2\Hess(f)$ and $g$ is called a {\it gradient}
Ricci soliton. In any case, we see that Ricci solitons are very natural
generalizations of {\it Einstein} metrics (i.e. $\ricci(g)=cg$).
The main significance, though, of the concept is that $g$ is a Ricci soliton if and
only if the curve of metrics
\begin{equation}\label{rssol}
g(t)=(-2ct+1)\vp_t^*g,
\end{equation}
is a solution to the Ricci flow
\begin{equation}\label{rf}
\dpar g(t)=-2\ricci(g(t)),
\end{equation}
for some one-parameter group $\vp_t$ of diffeomorphisms of $M$. In other words, the
Ricci flow starting at $g$ stays forever in the space of metrics which are
homothetic to $g$; it is unable to `improve' $g$. According to (\ref{rssol}), Ricci
solitons are called {\it expanding}, {\it steady}, or {\it shrinking} depending on
whether $c<0$, $c=0$, or $c>0$.
We are interested in the following question:
\begin{quote}
Which homogeneous manifolds $G/K$ admit a $G$-invariant Ricci soliton metric?
\end{quote}
Unfortunately, even Einstein homogeneous manifolds are still not well understood
(see \cite{BhmWngZll} and \cite{cruzchica} for the compact and noncompact cases,
respectively). But let us first review to what extent the Ricci soliton condition
is weaker than the Einstein condition for homogeneous manifolds.
Let $(M,g)$ be a Ricci soliton and let us assume it is homogeneous, i.e. its
isometry group acts transitively on $M$. In particular, $g$ has bounded curvature.
If $g$ is steady, it is easy to see from the ODE that the scalar curvature
$\scalar(g(t))$ satisfies that $\ricci(g)=0$, and consequently $g$ must be flat (see
\cite{AlkKml}). In the shrinking case, it follows from \cite[Theorem 1.2]{Nbr} that
$g$ is of gradient type, and it is proved in \cite{PtrWyl} that any homogeneous
gradient Ricci soliton is isometric to a quotient of $N\times\RR^k$, where $N$ is
some homogeneous Einstein manifold with positive scalar curvature and so compact and
with $\Pi_1(N)$ finite (see also \cite{Wyl}). Finally, if $g$ is expanding then $M$
must be noncompact (see \cite{Ivy}). Recall also that it follows from \cite{Prl}
that on a compact manifold all Ricci solitons are of gradient type.
We conclude that,
\begin{quote}
the noncompact expanding case is the only one allowing nontrivial homogeneous Ricci
solitons, and furthermore, they can not be of gradient type.
\end{quote}
All known examples so far of nontrivial homogeneous Ricci solitons are isometric to
a left-invariant metric $g$ on a simply connected Lie group $G$ (see Remark \ref{sc}
concerning non-simply connected Lie groups), and can be obtained in the following
way. Assume that $g$, which is identified with an inner product on the Lie algebra
$\ggo$ of $G$, satisfies
\begin{equation}\label{rsD}
\Ricci(g)=cI+D, \qquad\mbox{for some}\quad c\in\RR,\quad D\in\Der(\ggo),
\end{equation}
where $\Ricci(g)$ also denotes the {\it Ricci operator} of $g$ (i.e.
$\ricci(g)=g(\Ricci(g)\cdot,\cdot)$). If $X_D\in\chi(G)$ is defined by
\begin{equation}\label{defxd}
X_D(p)=\ddt|_0\vp_t(p), \qquad p\in G,
\end{equation}
where $\vp_t\in\Aut(G)$ is the unique automorphism such that
$\dif\vp_t|_e=e^{tD}\in\Aut(\ggo)$ (the existence of $\vp_t$ follows from $G$ being
simply connected), then
$$
\Lie_{X_D}g=\ddt|_0\vp_t^*g=\ddt|_0g(e^{-tD}\cdot,e^{-tD}\cdot) = -2g(D\cdot,\cdot).
$$
This implies that the Ricci tensor equals $\ricci(g)=cg-\unm\Lie_{X_{D}}g$, and
henceforth $g$ is a Ricci soliton. These vector fields $X_D$'s can be viewed as a
generalization to any Lie group of the so called {\it linear vector fields} on
$\RR^n$ (i.e. $X(p)=Ap$, $A\in\glg_n(\RR)$), and they play a nice and important role
in control theory (see \cite{AylTir}). We notice that for the Gaussian soliton on
$\RR^n$ one uses the linear vector field $X(p)=cp$.
Condition (\ref{rsD}) nicely combines the geometric and algebraic features of a
left-invariant metric on a Lie group, providing a neat way to find examples of
homogeneous Ricci solitons. These examples first appeared in \cite{soliton} ($G$
nilpotent), \cite{BrdDnl} ($G$ solvable, $\dim{G}=3$) and \cite{IsnJckLu} ($G$
solvable, $\dim{G}=4$).
\begin{remark}
It is an open question whether any Ricci soliton left invariant metric will satisfy
(\ref{rsD}), and concerning existence, we do not know of any non-solvable Lie group
admitting a nontrivial Ricci soliton.
\end{remark}
In the case when $G$ is nilpotent, metrics for which (\ref{rsD}) holds are called
{\it nilsolitons} and are known to satisfy the following properties (see the recent
survey \cite{cruzchica} for further information on nilsolitons):
\begin{itemize}
\item[(a)] Any left invariant Ricci soliton on $G$ is a nilsoliton.
\item[(b)] A given $G$ can admit at most one nilsoliton up to isometry and scaling among all its left-invariant metrics.
\item[(c)] Nilsolitons are also characterized by the following extremal property:
$$
||\ricci(g)||=\min\left\{ ||\ricci(g')||: g'\;\mbox{left-invariant on $G$},
\;\scalar(g')=\scalar(g)\right\}.
$$
Furthermore, they are the critical points of the functional square norm of the Ricci
tensor on the space of all nilmanifolds of a given dimension and scalar curvature.
\item[(d)] Nilsolitons are precisely the nilpotent parts of Einstein solvmanifolds.
\end{itemize}
Nevertheless, the existence, structural and classification problems on nilsolitons
seem to be far from being satisfactory solved, if at all possible.
\begin{definition}\label{defsols}
A left-invariant metric $g$ on a simply connected solvable Lie group is called a
{\it solsoliton} if the corresponding Ricci operator satisfies (\ref{rsD}).
\end{definition}
The name is inspired by the $3$-dimensional homogeneous geometry Sol from the
Geometrization Conjecture. It is natural to ask, in the case of solsolitons, for
properties analogous to (a)-(d) above. We will consider properties (b) and (d) here
and leave (a) and (c) for a forthcoming paper. Concerning property (a), it is worth
mentioning that any left invariant Ricci soliton on a {\it completely solvable} Lie
group (i.e. the eigenvalues of any $\ad{X}$ are all real) is necessarily a
solsoliton. This follows analogously to the proof of \cite[Proposition
1.1]{soliton} by using that two left-invariant metrics on one of these groups are
isometric if and only if there is an isomorphism which is an isometry between them
(see \cite{Alk}).
\section{Variety of nilpotent Lie algebras}\label{Tb}
Let $G$ be a real reductive group acting linearly on a finite dimensional real
vector space $V$ via $(g,v)\mapsto g.v$, $g\in G,v\in V$. The precise setting is the
one in \cite{RchSld}. We also refer to \cite{EbrJbl, HnzSchStt}, where many results from
geometric invariant theory are adapted and proved over $\RR$. The Lie algebra $\ggo$
of $G$ also acts linearly on $V$ by the derivative of the above action, which will
be denoted by $(\alpha,v)\mapsto\pi(\alpha)v$, $\alpha\in\ggo$, $v\in V$. We
consider a Cartan decomposition $\ggo=\kg\oplus\pg$, where $\kg$ is the Lie algebra
of a maximal compact subgroup $K$ of $G$. Endow $V$ with a from now on fixed
$K$-invariant inner product $\ip$ such that $\pg$ acts by symmetric operators, and
endow $\pg$ with an $\Ad(K)$-invariant inner product, which will be also denoted by
$\ip$.
The function $m:V\smallsetminus\{ 0\}\longrightarrow\pg$ implicitly defined by
\begin{equation}\label{defmm}
\la m(v),\alpha\ra=\isn\la\pi(\alpha)v,v\ra, \qquad \forall\alpha\in\pg,\; v\in V,
\end{equation}
\noindent is called the {\it moment map} for the representation $V$ of $G$. It is
easy to see that $m(cv)=m(v)$ for any nonzero $c\in\RR$ and $m$ is $K$-equivariant:
$m(k.v)=\Ad(k)m(v)$ for all $k\in K$. In the complex case (i.e. for a complex
representation of a complex reductive algebraic group), under the natural
identifications $\pg=\pg^*=(\im\kg)^*=\kg^*$, the function $m$ is precisely the
moment map from symplectic geometry, corresponding to the Hamiltonian action of $K$
on the symplectic manifold $\PP V$ (see \cite[Chapter 8]{Mmf} for further
information).
The functional square norm of the moment map,
\begin{equation}\label{norm}
F:V\smallsetminus\{ 0\}\mapsto\RR, \qquad F(v)=||m(v)||^2,
\end{equation}
is scaling invariant, so it can actually be viewed as a function on any sphere of
$V$ or on the projective space $\PP V$. If $\cca$ denotes the set of critical
points of $F:V\smallsetminus\{ 0\}\longrightarrow\RR$, then it is proved in \cite{Mrn}
(see \cite{Krw1,Nss} for the complex case) that $v\in\cca$ (or equivalently, $v$ is
a minimum for $F|_{G.v}$) if and only if $v$ is an eigenvector of $\pi(m(v))$, and
in that case, the following uniqueness result holds: $G.v\cap\cca=K.v$ (up to
scaling). This was previously proved in \cite{RchSld} (see \cite{KmpNss} for the
complex case) for the zeroes of $m$ (or equivalently, minimal vectors), which can
only appear inside closed orbits (see \cite[Section 11]{cruzchica} for a more
complete overview).
Let us consider the space of all skew-symmetric algebras of dimension $n$, which is
parameterized by the vector space
$$
V:=\lam=\{\mu:\RR^n\times\RR^n\longrightarrow\RR^n : \mu\; \mbox{bilinear and
skew-symmetric}\}.
$$
Then
$$
\nca=\{\mu\in V:\mu\;\mbox{satisfies Jacobi and is nilpotent}\}
$$
is an algebraic subset of $V$ as the Jacobi identity and the nilpotency condition
can both be written as zeroes of polynomial functions. $\nca$ is often called the
{\it variety of nilpotent Lie algebras} (of dimension $n$). There is a natural
linear action of $G=\G$ on $V$ given by
\begin{equation}\label{action}
g.\mu(X,Y)=g\mu(g^{-1}X,g^{-1}Y), \qquad X,Y\in\RR^n, \quad g\in\G,\quad \mu\in V.
\end{equation}
Recall that $\nca$ is $\G$-invariant and the Lie algebra isomorphism classes are
precisely the $\G$-orbits. The action of $\g$ on $V$ obtained by differentiation of
(\ref{action}) is given by
\begin{equation}\label{actiong}
\pi(\alpha)\mu=\alpha\mu(\cdot,\cdot)-\mu(\alpha\cdot,\cdot)-\mu(\cdot,\alpha\cdot),
\qquad \alpha\in\g,\quad\mu\in V.
\end{equation}
We note that $\pi(\alpha)\mu=0$ if and only if $\alpha\in\Der(\mu)$, the Lie algebra
of derivations of the algebra $\mu$. The canonical inner product $\ip$ on $\RR^n$
determines an $\Or(n)$-invariant inner product on $V$, also denoted by $\ip$, as
follows:
\begin{equation}\label{innV}
\la\mu,\lambda\ra= \sum\la\mu(e_i,e_j),\lambda(e_i,e_j)\ra
=\sum\la\mu(e_i,e_j),e_k\ra\la\lambda(e_i,e_j),e_k\ra,
\end{equation}
\noindent and also the standard $\Ad(\Or(n))$-invariant inner product on $\g$ given
by
\begin{equation}\label{inng}
\la \alpha,\beta\ra=\tr{\alpha \beta^{\mathrm t}}=\sum\la\alpha e_i,\beta e_i\ra
=\sum\la\alpha e_i,e_j\ra\la\beta e_i,e_j\ra, \qquad \alpha,\beta\in\g,
\end{equation}
\noindent where $\{ e_1,...,e_n\}$ denotes the canonical basis of $\RR^n$. We have
made several abuses of notation concerning inner products. Recall that $\ip$ has
been used to denote inner products on $\RR^n$, $V$ and $\g$ indistinctly. We note
that $\pi(\alpha)^t=\pi(\alpha^t)$ and $(\ad{\alpha})^t=\ad{\alpha^t}$ for any
$\alpha\in\g$, due to the choice of these canonical inner products everywhere.
We use $\g=\sog(n)\oplus\sym(n)$ as a Cartan decomposition for the Lie algebra $\g$
of $\G$, where $\sog(n)$ and $\sym(n)$ denote the subspaces of skew-symmetric and
symmetric matrices, respectively. It is proved in \cite[Proposition 3.5]{minimal}
that if $\ad_{\mu}$ denotes left multiplication for the algebra $\mu$ then the
moment map $m:V\smallsetminus\{ 0\}\longrightarrow\sym(n)$ for the action (\ref{action})
is given by
$$
m(\mu)=\tfrac{1}{||\mu||^2} \left(-2\sum(\ad_{\mu}{e_i})^t\ad_{\mu}{e_i} +
\sum\ad_{\mu}{e_i}(\ad_{\mu}{e_i})^t\right),
$$
or equivalently, for all $X\in\RR^n$,
\begin{equation}\label{mmv}
\la m(\mu)X,X\ra=\tfrac{1}{||\mu||^2}\left(-2\sum\la\mu(X,e_i),e_j\ra^2
+\sum\la\mu(e_i,e_j),X\ra^2\right).
\end{equation}
Let $\tg$ denote the set of all diagonal $n\times n$ matrices. If $\{
e_1',...,e_n'\}$ is the basis of $(\RR^n)^*$ dual to the canonical basis $\{
e_1,...,e_n\}$ then
$$
\{ v_{ijk}=(e_i'\wedge e_j')\otimes e_k : 1\leq i<j\leq n, \; 1\leq k\leq n\}
$$
is a basis of weight vectors of $V$ for the action (\ref{action}), where $v_{ijk}$
is actually the bilinear form on $\RR^n$ defined by
$v_{ijk}(e_i,e_j)=-v_{ijk}(e_j,e_i)=e_k$ and zero otherwise. The corresponding
weights $\alpha_{ij}^k\in\tg$, $i<j$, are given by
\begin{equation}\label{alfas}
\pi(\alpha)v_{ijk}=(a_k-a_i-a_j)v_{ijk}=\la\alpha,\alpha_{ij}^k\ra v_{ijk},
\quad\forall\alpha=\left[\begin{smallmatrix} a_1&&\\ &\ddots&\\ &&a_n
\end{smallmatrix}\right]\in\tg,
\end{equation}
\noindent where $\alpha_{ij}^k=E_{kk}-E_{ii}-E_{jj}$ and $\ip$ is the inner product
defined in (\ref{inng}). As usual $E_{rs}$ denotes the matrix with $1$ at entry
$rs$ and $0$ elsewhere.
From now on, we will always denote by $\mu_{ij}^k$ the structure constants of a
vector $\mu\in V$ with respect to the basis $\{ v_{ijk}\}$:
$$
\mu=\sum\mu_{ij}^kv_{ijk}, \qquad \mu_{ij}^k\in\RR, \qquad {\rm i.e.}\quad
\mu(e_i,e_j)=\sum_{k=1}^n\mu_{ij}^ke_k, \quad i<j.
$$
Each nonzero $\mu\in V$ uniquely determines an element $\beta_{\mu}\in\tg$ given by
$$
\beta_{\mu}:=\mcc\left\{\alpha_{ij}^k:\mu_{ij}^k\ne 0\right\},
$$
where $\mcc(X)$ denotes the unique element of minimal norm in the convex hull
$\CH(X)$ of a subset $X\subset\tg$. We note that $\beta_{\mu}$ is always nonzero
since $\tr{\alpha_{ij}^k}=-1$ for all $i<j$ and consequently $\tr{\beta_{\mu}}=-1$.
Let $\tg^+$ denote the Weyl chamber of $\g$ given by
\begin{equation}\label{weyl}
\tg^+=\left\{\left[\begin{smallmatrix} a_1&&\\ &\ddots&\\ &&a_n
\end{smallmatrix}\right]\in\tg:a_1\leq...\leq a_n\right\}.
\end{equation}
In \cite{standard}, a $\G$-invariant stratification for $V=\lam$ has been defined by
adapting to this context the construction given in \cite[Section 12]{Krw1} for
reductive group representations over an algebraically closed field. We summarize in
the following theorem the main properties of the stratification, which will become
our main tool in the study of the structure of solsolitons in next section (see
\cite[Section 7]{cruzchica} for a detailed overview).
\begin{theorem}\label{strata}\cite{standard, einsteinsolv}
There exists a finite subset $\bca\subset\tg^+$, and for each $\beta\in\bca$ a
$\G$-invariant subset $\sca_{\beta}\subset V$ (a {\it stratum}) such that
$$
V\smallsetminus\{ 0\}=\bigcup_{\beta\in\bca}\sca_{\beta} \qquad \mbox{(disjoint
union)},
$$
and $\tr{\beta}=-1$ for any $\beta\in\bca$. For $\mu\in\sca_{\beta}$ we have that
\begin{equation}\label{betapos}
\beta+||\beta||^2I \quad\mbox{is positive definite for all}\; \beta\in\bca
\;\mbox{such that}\; \sca_{\beta}\cap\nca\ne\emptyset,\;\mbox{and}
\end{equation}
\begin{equation}\label{bmu}
||\beta||\leq ||m(\mu)||\qquad(\mbox{here, equality holds}\;\Leftrightarrow
m(\mu)\;\mbox{is conjugate to}\; \beta).
\end{equation}
If in addition, $\mu\in\sca_{\beta}$ satisfies $\beta_{\mu}=\beta$ (or equivalently,
$\min\left\{\la\beta,\alpha_{ij}^k\ra:\mu_{ij}^k\ne 0\right\}=||\beta||^2$), which
always holds for some $g.\mu$, $g\in\Or(n)$, then
\begin{equation}\label{adbeta}
\left\la[\beta,D],D\right\ra\geq 0 \qquad\forall\; D\in\Der(\mu) \qquad(\mbox{here,
equality holds}\;\Leftrightarrow [\beta,D]=0),
\end{equation}
\begin{equation}\label{betaort}
\tr{\beta D}=0 \quad\forall\; D\in\Der(\mu), \;\mbox{and}
\end{equation}
\begin{equation}\label{delta}
\left\la\pi\left(\beta+||\beta||^2I\right)\mu,\mu\right\ra\geq 0 \qquad (\mbox{here,
equality holds}\;\Leftrightarrow \beta+||\beta||^2I\in\Der(\mu)).
\end{equation}
\end{theorem}
This stratification is based on instability results and is strongly related to the
moment map in many ways other than (\ref{bmu}) (see \cite{cruzchica}).
\section{Structure of solsolitons}\label{structure}
Let $S$ be a {\it solvmanifold}, that is, a simply connected solvable Lie group
endowed with a left invariant Riemannian metric. $S$ will be often identified with
its metric Lie algebra $(\sg,\ip)$, where $\sg$ is the Lie algebra of $S$ and $\ip$
denotes the inner product on $\sg$ which determines the metric. We consider the
orthogonal decomposition
$$
\sg=\ag\oplus\ngo,
$$
where $\ngo$ is the nilradical of $\sg$ (i.e. maximal nilpotent ideal). The {\it
mean curvature vector} of $S$ is the only element $H\in\ag$ such that $\la
H,A\ra=\tr{\ad{A}}$ for any $A\in\ag$. If $B$ denotes the symmetric map defined by
the Killing form of $\sg$ relative to $\ip$ (i.e. $\la BX,X\ra=\tr{(\ad{X})^2}$ for
all $X\in\sg$) then $B(\ag)\subset\ag$ and $B|_{\ngo}=0$. The Ricci operator
$\Ricci$ of $S$ is given by (see for instance \cite[7.38]{Bss}):
\begin{equation}\label{ricci}
\Ricci=R-\unm B-S(\ad{H}),
\end{equation}
\noindent where $S(\ad{H})=\unm(\ad{H}+(\ad{H})^t)$ is the symmetric part of
$\ad{H}$ and $R$ is the symmetric operator defined by
\begin{equation}\label{R}
\la RX,X\ra=-\unm\sum\la [X,X_i],X_j\ra^2 +\unc\sum\la
[X_i,X_j],X\ra^2,\qquad\forall X\in\sg,
\end{equation}
\noindent where $\{ X_i\}$ is any orthonormal basis of $(\sg,\ip)$.
It follows from (\ref{mmv}) that this anonymous tensor $R$ in the formula of the
Ricci operator satisfies
\begin{equation}\label{mmR}
m([\cdot,\cdot])=\tfrac{4}{||[\cdot,\cdot]||^2}R,
\end{equation}
\noindent where $m:\Lambda^2\sg^*\otimes\sg\longrightarrow\sym(\sg)$ is the moment
map for the natural action of $\Gl(\sg)$ on $\Lambda^2\sg^*\otimes\sg$. In other
words, $R$ may be alternatively defined as follows:
\begin{equation}\label{Rmm}
\tr{RE}=\unc\la \pi(E)[\cdot,\cdot],[\cdot,\cdot]\ra, \qquad\forall E\in\End(\sg),
\end{equation}
\noindent where we are considering the Lie bracket $[\cdot,\cdot]$ of $\sg$ as a
vector in $\Lambda^2\sg^*\otimes\sg$, $\ip$ is the inner product defined in
(\ref{innV}) and $\pi$ is the representation given in (\ref{actiong}) (see the
notation in Section \ref{Tb} and replace $\RR^n$ by $\sg$).
\begin{remark}\label{st2}
In particular, $R$ is orthogonal to any derivation of $\sg$. It is easy to see that
the same holds for $B$.
\end{remark}
\begin{remark}\label{st1}
If $\sg$ is nilpotent then $\Ricci= R$ and hence the scalar curvature is simply
given by $\scalar=\tr{R}=-\unc ||\lb||^2$.
\end{remark}
The following more explicit formula for the Ricci operator of $S$ follows from a
straightforward computation by using (\ref{ricci}) and (\ref{R}):
\begin{equation}\label{ricgen}
\begin{array}{rcl}
\la\Ricci A,A\ra &=& -\unm\sum ||[A,A_i]||^2 -\tr{S(\ad{A}|_{\ngo})^2}, \\ \\
\la\Ricci A,X\ra &=& -\unm\sum\la[A,A_i],[X,A_i]\ra-\unm\tr{(\ad{A}|_{\ngo})^t\ad{X}|_{\ngo}} \\ \\ && -\unm\la[H,A],X\ra, \\ \\
\la\Ricci X,X\ra &=& \unc\sum\la[A_i,A_j],X\ra^2
+\unm\sum\la[\ad{A_i}|_{\ngo},(\ad{A_i}|_{\ngo})^t](X),X\ra \\ \\ && -\unm\sum\la
[X,X_i],X_j\ra^2 +\unc\sum\la [X_i,X_j],X\ra^2 -\la[H,X],X\ra,
\end{array}
\end{equation}
\noindent for all $A\in\ag$ and $X\in\ngo$, where $\{ A_i\}$, $\{ X_i\}$, are any
orthonormal basis of $\ag$ and $\ngo$, respectively, and
$S(\ad{A}|_{\ngo})=\unm(\ad{A}|_{\ngo}+(\ad{A}|_{\ngo})^t)$. It is now clear from
(\ref{ricgen}) that there is a substantial simplification of the expression of
$\Ricci$ under the assumptions $[\ag,\ag]=0$ and $\ad{A}$ symmetric for all
$A\in\ag$. This gives rise to the following natural construction of solsolitons
starting from a nilsoliton.
\begin{proposition}\label{const}
Let $(\ngo,\ip_1)$ be a nilsoliton, say with Ricci operator $\Ricci_1=cI+D_1$,
$c<0$, $D_1\in\Der(\ngo)$, and consider $\ag$ any abelian Lie algebra of symmetric
derivations of $(\ngo,\ip_1)$. Then the solvmanifold $S$ with Lie algebra
$\sg=\ag\oplus\ngo$ (semidirect product) and inner product given by
$$
\ip|_{\ngo\times\ngo}=\ip_1, \qquad \la\ag,\ngo\ra=0, \qquad \la
A,A\ra=-\tfrac{1}{c}\tr{A^2} \qquad\forall A\in\ag,
$$
is a solsoliton with $\Ricci=cI+D$, where $D\in\Der(\sg)$ is defined by
$D|_{\ag}=0$, $D|_{\ngo}=D_1-\ad{H}|_{\ngo}$ and $H$ is the mean curvature vector of
$S$. Furthermore, $S$ is Einstein if and only if $D_1\in\ag$.
\end{proposition}
\begin{remark} The aim of this section is to show that this very simple procedure
actually yields all solsolitons up to isometry.
\end{remark}
\begin{remark} If $\ngo$ is abelian then $\Ricci_1=0$ and so we can take $D_1=-cI$ for any $c<0$.
\end{remark}
\begin{proof}
It follows directly from the hypotheses and (\ref{ricgen}) that $\Ricci|_{\ag}=cI$
and
$$
\Ricci|_{\ngo}=\Ricci_1-H=cI+D_1-\ad{H}|_{\ngo}.
$$
If $\ngo=\ngo_1\oplus...\oplus\ngo_r$ is the orthogonal decomposition with
$[\ngo,\ngo]=\ngo_2\oplus...\oplus\ngo_r$,
$[\ngo,[\ngo,\ngo]]=\ngo_3\oplus...\oplus\ngo_r$, and so on, then for any $A\in\ag$,
$\ad{A}|_{\ngo}=A$ leaves the subspaces $\ngo_i$ invariant as it is a symmetric
derivation (it is actually enough for this that $(\ad{A}|_{\ngo})^t$ be a derivation
as well) and $\ad{X}(\ngo_i)\subset\ngo_{i+1}\oplus...\oplus\ngo_r$ for all $i$.
This implies that $\tr{(\ad{A}|_{\ngo})^t\ad{X}|_{\ngo}}=0$, and thus
$\la\Ricci\ag,\ngo\ra=0$. It only remains to prove that $D\in\Der(\sg)$, for which
it is enough to show that $[\ag,D_1]=0$, but this follows by using that any
symmetric derivation of $(\ngo,\ip_1)$ (and more generally any derivation whose
transpose is also a derivation) commutes with $\Ricci_1$ (see \cite[Lemma
2.2]{Hbr}).
Let us now prove the Einstein assertion. If $S$ is Einstein then
$D_1=\ad{H}|_{\ngo}=H\in\ag$. Conversely, if $D_1\in\ag$, then since
$\tr{\Ricci_1A}=0$ for any $A\in\ag$ (see Remarks \ref{st1} and \ref{st2}) we have
that
$$
\tr{D_1A}=-c\tr{A}=-c\tr{\ad{A}}=-c\la H,A\ra =\tr{HA}=\tr{\ad{H}|_{\ngo}A},
\qquad\forall A\in\ag.
$$
This implies that $D_1=\ad{H}|_{\ngo}$ (i.e. $\Ricci=cI$), completing the proof of
the proposition.
\end{proof}
We now show that $c\geq 0$ is actually not allowed for nontrivial solsolitons. This
gives an alternative proof of the fact that any nonflat solsoliton must be expanding
(see Section \ref{hrs}).
\begin{proposition}\label{cneg}
Let $S$ be a solsoliton, say with $\Ricci=cI+D$, $c\in\RR$, $D\in\Der(\sg)$. If
$c\geq 0$ then $\Ricci=0$.
\end{proposition}
\begin{proof}
Since $D$ is a symmetric derivation we have that $D|_{\ag}=0$, $D(\ngo)\subset\ngo$.
It then follows from just the first line in (\ref{ricgen}) that $c=0$, $[\ag,\ag]=0$
and $\ad{A}^t=-\ad{A}$ for any $A\in\ag$. We also obtain from (\ref{ricgen}) that
$D|_{\ngo}=\Ricci_1$, the Ricci operator of $(\ngo,\ip_{\ngo\times\ngo})$, and so
$\Ricci_1=0$ by Remarks \ref{st1} and \ref{st2}, concluding the proof.
\end{proof}
We will need in what follows the following technical lemma valid for general
solvmanifolds.
\begin{lemma}\label{adA}
Let $S$ be a solvmanifold. Then, for any $A\in\ag$, the following conditions are
equivalent:
\begin{itemize}
\item[(i)] $(\ad{A})^t$ is a derivation of $\sg$.
\item[(ii)] $\ad{A}$ is a normal operator (i.e. $[\ad{A},(\ad{A})^t]=0$).
\end{itemize}
\end{lemma}
\begin{proof}
It follows from (\ref{Rmm}) that
$$
\begin{array}{rcl}
\tr{R[\ad{A},(\ad{A})^t]} &=& \unc\la\pi([\ad{A},(\ad{A})^t])[\cdot,\cdot],[\cdot,\cdot]\ra \\ \\
&=& \unc\la\pi(\ad{A})\pi((\ad{A})^t)[\cdot,\cdot],[\cdot,\cdot]\ra \\ \\
&=& \unc\la\pi((\ad{A})^t)[\cdot,\cdot],\pi(\ad{A})^t[\cdot,\cdot]\ra \\ \\
&=& \unc||\pi((\ad{A})^t)[\cdot,\cdot]||^2,
\end{array}
$$
and so if $\ad{A}$ is normal then $\pi((\ad{A})^t)\lb=0$, that is,
$(\ad{A})^t\in\Der(\sg)$.
Conversely, if $(\ad{A})^t$ is a derivation of $\sg$ then both $\ad{A}$ and
$(\ad{A})^t$ must vanish on $\ag$, since they leave $\ngo$ invariant and their
images are contained in $\ngo$ (this last statement is well-known to be true for any
derivation, see for instance \cite[Lemma 2.6]{GrdWls}). This implies that
$$
(\ad{A})^t([A,X])=[(\ad{A})^t(A),X]+[A,(\ad{A})^t(X)]= [A,(\ad{A})^t(X)],
\qquad\forall X\in\ngo,
$$
which is equivalent to saying that $[\ad{A},(\ad{A})^t]=0$.
\end{proof}
The following structural theorem for solsolitons is the main result of this paper.
\begin{theorem}\label{main}
Let $S$ be a solvmanifold with metric Lie algebra $(\sg,\ip)$ and consider the
orthogonal decomposition $\sg=\ag\oplus\ngo$, where $\ngo$ is the nilradical of
$\sg$. Then $\Ricci=cI+D$ for some $c<0$, $D\in\Der(\sg)$, i.e. $S$ is a
solsoliton, if and only if the following conditions hold:
\begin{itemize}
\item[(i)] $(\ngo,\ip|_{\ngo\times\ngo})$ is a nilsoliton with Ricci
operator $\Ricci_1=cI+D_1$, for some $D_1\in\Der(\ngo)$.
\item[(ii)] $[\ag,\ag]=0$.
\item[(iii)] $(\ad{A})^t\in\Der(\sg)$ (or equivalently,
$[\ad{A},(\ad{A})^t]=0$) for all $A\in\ag$.
\item[(iv)] $\la A,A\ra=-\tfrac{1}{c}\tr{S(\ad{A})^2}$ for all $A\in\ag$,
where $S(\ad{A})=\unm(\ad{A}+(\ad{A})^t)$.
\end{itemize}
\end{theorem}
\begin{proof}
If conditions (i)-(iv) are satisfied by $S$ then we can argue exactly as in the
proof of Proposition \ref{const} to obtain that $\Ricci=cI+D$, where $D$ is defined
by $D|_{\ag}=0$, $D|_{\ngo}=D_1-S(\ad{H}|_{\ngo})$, $H$ the mean curvature vector of
$S$.
Let us then prove the converse assertion. Let $S$ be a solvmanifold with
$\Ricci=cI+D$, $c<0$, $D\in\Der(\sg)$. If $F:=S(\ad{H})+D$ then we obtain from
(\ref{ricci}) and (\ref{Rmm}) that
\begin{equation}\label{einstein}
\tr{\left(cI+\unm B+F\right)E}= \unc\la\pi(E)\lb,\lb\ra, \qquad\forall
E\in\End(\sg).
\end{equation}
By letting $E=\ad{H}+D$ in (\ref{einstein}) and using Remark \ref{st2} we get
\begin{equation}\label{c}
c\tr{F}+\tr{F^2}=0.
\end{equation}
In particular, $\tr{F}\geq 0$ and equality holds if and only if $F=0$. Also, by
applying (\ref{einstein}) to $E$ defined by $E|_{\ag}=0$ and $E|_{\ngo}=I$, it is
easy to see that
\begin{equation}\label{c2}
cn+\tr{F}=\unc\sum||[A_i,A_j]||^2+\unc||\lb_{\ngo\times\ngo}||^2, \qquad
n=\dim{\ngo}.
\end{equation}
We first consider the case when $\ngo$ is abelian. It follows from (\ref{c2}) that
$cn+\tr{F}\geq 0$, and so by (\ref{c}) we get that
$\tr{F^2}\leq\tfrac{1}{n}(\tr{F})^2$. This implies that $[\ag,\ag]=0$, since
equality must hold in (\ref{c2}), and also that $F|_{\ngo}=tI$ for some $t\geq 0$,
but therefore $F|_{\ag}=0$ and $F|_{\ngo}=-cI$. We now obtain from (\ref{ricgen})
that the restrictions to $\ngo$ satisfy
$$
cI+D = \unm\sum[\ad{A_i},(\ad{A_i})^t]-S(\ad{H}),
$$
from which it follows that $\sum[\ad{A_i},(\ad{A_i})^t]=0$. By arguing as in the
proof of Lemma \ref{adA} we get that
\begin{equation}\label{trRadA}
0=\tr\left(R\sum[\ad{A_i},(\ad{A_i})^t]\right)=\unc\sum||\pi((\ad{A_i})^t)\lb||^2,
\end{equation}
\noindent and therefore (iii) follows.
Let us assume from now on that $\ngo$ is not abelian. In order to apply the results
in Section \ref{Tb}, we identify $\ngo$ with $\RR^n$ via an orthonormal basis $\{
e_1,...,e_n\}$ of $\ngo$. In this way, $\mu:=\lb|_{\ngo\times\ngo}$ can be viewed
as a nonzero element of $\nca\subset V$. Thus $\mu\in\sca_{\beta}$ for some
$\beta\in\bca\subset\tg^+$ and there exists $g\in\Or(n)$ such that
$\tilde{\mu}:=g.\mu$ satisfies $\beta_{\tilde{\mu}}=\beta$, so that in addition
(\ref{betaort}) and (\ref{delta}) hold for $\tilde{\mu}$ (see Theorem \ref{strata}).
Let $\tilde{g}$ denote the orthogonal map of $(\sg,\ip)$ defined by
$\tilde{g}|_{\ag}=I$, $\tilde{g}|_{\ngo}=g$, and let $\tilde{\sg}$ be the Lie
algebra which is $\sg$ as a vector space and has Lie bracket
$$
\tilde{g}.[\cdot,\cdot]=\tilde{g}[\tilde{g}^{-1}\cdot,\tilde{g}^{-1}\cdot].
$$
We therefore have that $(\tilde{\sg},\ip)$ is isometric to $(\sg,\ip)$, as
$\tilde{g}:\ggo\longrightarrow\tilde{\ggo}$ is an isometric isomorphism between the
two metric Lie algebras. Since conditions (i)-(iv) hold for $(\tilde{\sg},\ip)$ if
and only if they hold for $(\sg,\ip)$, we can assume from now on that all properties
(\ref{adbeta})-(\ref{delta}) in Theorem \ref{strata} hold for $\mu$.
Consider $E_{\beta}\in\End(\sg)$ defined by
$$
E_{\beta}:=\left[\begin{array}{cc} 0&0\\
0&\beta+||\beta||^2I\end{array}\right], \qquad \mbox{that is}, \quad
E_{\beta}|_{\ag}=0, \quad E_{\beta}|_{\ngo}=\beta+||\beta||^2I.
$$
\begin{lemma}\label{est}
$\la\pi(E_{\beta})\lb,\lb\ra\geq 0$ and it equals the sum of the following three
nonnegative terms:
\begin{equation}\label{ests}
\begin{array}{l}
\unc\sum\la(\beta+||\beta||^2I)[A_i,A_j],[A_i,A_j]\ra \\ \\
+ \unm\sum\la [\beta,\ad{A_i}],\ad{A_i}\ra \\ \\
+ \unc\la\pi(\beta+||\beta||^2I)\mu,\mu\ra.
\end{array}
\end{equation}
\end{lemma}
\begin{proof}
If $\{ A_r\}$ is an orthonormal basis of $\ag$ then
$$
\begin{array}{rcl}
\la\pi(E_{\beta})\lb,\lb\ra &=& \unc\sum\la E_{\beta}[A_r,A_s],[A_r,A_s]\ra \\ \\
&&+ \unm\sum\la E_{\beta}[A_r,e_i],[A_r,e_i]\ra - \unm\sum\la
[A_r,E_{\beta}e_i],[A_r,e_i]\ra, \\ \\
&& + \unc\sum\la E_{\beta}[e_i,e_j]-[E_{\beta}e_i,e_j]
-[e_i,E_{\beta}e_j],[e_i,e_j]\ra,
\end{array}
$$
which in turn equals
\begin{equation}\label{prueba2}
\begin{array}{l}
\unc\sum\la(\beta+||\beta||^2I)[A_r,A_s],[A_r,A_s]\ra \\ \\
+ \unm\sum\la (\beta\ad{A_r}-\ad{A_r}\beta)(e_i),\ad{A_r}(e_i)\ra
+\unc\la\pi(\beta+||\beta||^2I)\mu,\mu\ra,
\end{array}
\end{equation}
\noindent and so (\ref{ests}) follows. The three terms in (\ref{ests}) are $\geq 0$
by (\ref{betapos}), (\ref{adbeta}) and (\ref{delta}), respectively.
\end{proof}
We therefore obtain, by applying (\ref{einstein}) to $E=E_{\beta}$ and using Lemma
\ref{est}, that
\begin{equation}\label{prueba1}
c\tr{E_{\beta}}+\tr{FE_{\beta}}\geq 0.
\end{equation}
Recall that $\tr{\beta}=-1$ (see Theorem \ref{strata}) and so
\begin{equation}\label{prueba3}
\begin{array}{ll}
\tr{E_{\beta}^2} & =\tr(\beta^2+||\beta||^4I+2||\beta||^2\beta)=||\beta||^2(1+n||\beta||^2-2) \\ \\
& = ||\beta||^2(-1+n||\beta||^2)=||\beta||^2\tr{E_{\beta}}.
\end{array}
\end{equation}
On the other hand, we have that
\begin{equation}\label{prueba4}
\tr{FE_{\beta}}=\tr{F|_{\ngo}(\beta+||\beta||^2)}=||\beta||^2\tr{F}
\end{equation}
\noindent by (\ref{betaort}). We now use (\ref{c}), (\ref{prueba1}),
(\ref{prueba3}) and (\ref{prueba4}) and obtain by a straightforward manipulation
that
$$
\tr{F^2}\tr{E_{\beta}^2}\leq (\tr{FE_{\beta}})^2,
$$
that is, we get a `backwards' Cauchy-Schwartz inequality. This has many strong
consequences. A first one is that $F=tE_{\beta}$ for some $t\geq 0$, and thus by
(\ref{c}),
\begin{equation}\label{tfor}
F|_{\ngo}=-\tfrac{c}{||\beta||^2}E_{\beta}.
\end{equation}
Recall that $\tr{F}>0$ since $F\ne 0$ by (\ref{prueba1}) and (\ref{betapos}).
Secondly, the three nonnegative terms in (\ref{ests}) must vanish, which
respectively implies that $[\ag,\ag]=0$ by (\ref{betapos}),
$[\beta,\ad{\ag}|_{\ngo}]=0$ by (\ref{adbeta}) and $\beta+||\beta||^2I\in\Der(\ngo)$
by (\ref{delta}). Thus (ii) holds and so (iv) follows from (\ref{ricgen}).
We also obtain from (\ref{ricgen}) that the restrictions to $\ngo$ satisfy
\begin{equation}\label{cID}
cI+D = \unm\sum[\ad{A_i},(\ad{A_i})^t] + \Ricci_1 - S(\ad{H}),
\end{equation}
\noindent and hence it follows from (\ref{tfor}) that
$$
\unm\sum[\ad{A_i},(\ad{A_i})^t] +\Ricci_1=-\tfrac{c}{||\beta||^2}\beta.
$$
By taking traces we get $-\unc||\mu||^2=\tfrac{c}{||\beta||^2}$ (see Remark
\ref{st1}). This implies that
$$
\begin{array}{l}
\tfrac{1}{8}\sum||\pi((\ad{A_i})^t)\lb||^2 +\tr{\Ricci_1^2} =
\tr{\Ricci_1\sum[\ad{A_i},(\ad{A_i})^t]} +\tr{\Ricci_1^2} \\ \\
= -\tfrac{c}{||\beta||^2}\tr{\Ricci_1\beta} = \unc||\mu||^2\tr{\Ricci_1\beta} =
\tfrac{1}{16}||\mu||^4\la m(\mu),\beta\ra \leq \tfrac{1}{16}||\mu||^4 ||m(\mu)||
||\beta|| \\ \\
\leq \tfrac{1}{16}||\mu||^4 ||m(\mu)||^2 = \tr{\Ricci_1^2}.
\end{array}
$$
The first equality above holds by (\ref{trRadA}) and the last inequality follows
from (\ref{bmu}) and the fact that $m(\mu)=\tfrac{4}{||\mu||^2}\Ricci_1$ (see
(\ref{mmR}) and Remark \ref{st1}). We therefore obtain that
$$
\sum||\pi((\ad{A_i})^t)\lb||^2=0,
$$
and so (iii) follows. We now use (\ref{cID}) and Lemma \ref{adA} to conclude that
$$
\Ricci_1=cI+D+S(\ad{H})\in\RR I +\Der(\ngo),
$$
and therefore part (i) holds.
\end{proof}
It is well-known that a solvmanifold which satisfies conditions (ii) and (iii) in
Theorem \ref{main} is isometric to the one obtained by just changing the Lie bracket
into
\begin{equation}\label{sympart}
[A,X]'=S(\ad{A})X, \qquad [X,Y]'=[X,Y], \qquad\forall A\in\ag,\; X,Y\in\ngo,
\end{equation}
and keeping the same $\ip$ (see for instance \cite[Proposition 2.5]{Hbr}).
Thus the next result follows directly from Theorem \ref{main}.
\begin{corollary}\label{cormain}
Up to isometry, any solsoliton can be constructed as in {\rm Proposition
\ref{const}}.
\end{corollary}
As a byproduct of the proof of Theorem \ref{main}, the following extra structural
properties for solsolitons have been obtained.
\begin{proposition}\label{extras}
Let $S$ be a solsoliton, say with $\Ricci=cI+D$, $c<0$, $D\in\Der(\sg)$, and let us
assume that $\ngo$ is not abelian, $\mu:=\lb|_{\ngo\times\ngo}\in\sca_{\beta}$ and
$\beta_{\mu}=\beta$. If $E_{\beta}\in\End(\sg)$ is defined by
$$
E_{\beta}:=\left[\begin{smallmatrix} 0&0\\
0&\beta+||\beta||^2I\end{smallmatrix}\right], \qquad\mbox{i.e.}\quad
E_{\beta}|_{\ag}=0, \quad E_{\beta}|_{\ngo}=\beta+||\beta||^2I,
$$
then the following conditions hold:
\begin{itemize}
\item[(i)] $E_{\beta}\in\Der(\sg)$ (or equivalently, $[\beta,\ad{\ag}]=0$
and $\beta+||\beta||^2I\in\Der(\ngo)$).
\item[(ii)] $S(\ad{H})+D=-\tfrac{c}{||\mu||^2}E_{\beta}$. In particular,
$S$ is Einstein if and only if $S(\ad{H})=-\tfrac{c}{||\mu||^2}E_{\beta}$.
\item[(iii)] $c=-\unc ||\mu||^2||\beta||^2$ and $m(\mu)=\beta$.
\end{itemize}
\end{proposition}
Recall that condition $\beta_\mu=\beta$ can be assumed to hold up to isometry since
always $\beta_{g.\mu}=\beta$ for some $g\in\Or(\ngo)$ (see Theorem \ref{strata} and
the paragraph before Lemma \ref{est}).
\begin{remark}\label{sc}
Can we use condition (\ref{rsD}) to find examples of Ricci solitons on solvable Lie
groups which are not simply connected? The answer is no, such as for Einstein
solvmanifolds. Indeed, for the field $X_D$ (see (\ref{defxd})) to be defined on a
Lie group $S/\Gamma$ covered by a simply connected solvable Lie group $S$, where
$\Gamma$ is a discrete subgroup of the center of $S$, it is necessary that the
one-parameter group of $\vp_t\in\Aut(S)$ with $\dif\vp_t|_e=e^{tD}$ satisfies
$\vp_t(\Gamma)=\Gamma$ for all $t\in\RR$. But since $\Gamma$ is discrete this
implies that $\vp_t(\gamma)=\gamma$ for all $\gamma\in\Gamma$ and $t$, and
consequently $DX=0$ for some nonzero $X$ in the center $\zg(\sg)$ of $\sg$. It
follows from Theorem \ref{main} (ii) and (iii) that $X\in\ngo$, and since
$D|_{\ngo}=D_1-S(\ad{H})|_{\ngo}$ we obtain that
$$
0=\la DX,X\ra=\la D_1X,X\ra-\la[H,X],X\ra = \la D_1X,X\ra,
$$
which contradicts the fact that $D_1$ is positive definite.
\end{remark}
\section{Uniqueness of solsolitons}\label{uniq}
The structural results obtained in Theorem \ref{main} pave the way for the following
uniqueness result for solsolitons, which is the analogous of the one already known
for nilsolitons.
\begin{theorem}\label{uni}
Let $S$ and $S'$ be two solsolitons which are isomorphic as Lie groups. Then $S$ is
isometric to $S'$ up to scaling.
\end{theorem}
\begin{remark}
In particular, a given solvable Lie group can admit at most one Ricci soliton left
invariant metric up to isometry and scaling.
\end{remark}
\begin{proof}
Without any loss of generality, we can assume that $\sg=\sg'$, $\ag=\ag'$ and
$\ngo=\ngo'$ as vector spaces and that $\ip=\ip'$. Thus the solvmanifolds $S$ and
$S'$ are respectively determined only by their Lie brackets $\lb$ and $\lb'$, which
must satisfy all structural properties in Theorem \ref{main}. From now on, such properties will
be referred to just by (i),...,(iv). Since $S$ and $S'$ are isomorphic,
there exists a Lie algebra isomorphism between $\lb$ and $\lb'$ of the form
$$
f=\left[\begin{array}{cc} g&0\\
j&h\end{array}\right], \qquad g\in\Gl(\ag), \quad h\in\Gl(\ngo), \quad
j:\ag\longrightarrow\ngo,
$$
such that
\begin{equation}\label{ison}
h.\mu=\mu', \qquad \mu:=\lb|_{\ngo\times\ngo}, \quad \mu':=\lb'|_{\ngo\times\ngo},
\end{equation}
\noindent and
\begin{equation}\label{isoa}
h\ad(g^{-1}A)|_{\ngo}h^{-1} = \ad'{A}|_{\ngo} + \ad_{\mu'}(jg^{-1}A), \qquad\forall
A\in\ag.
\end{equation}
These two conditions are also sufficient for $f$ being an isomorphism by (ii). We
can assume up to scaling that the scalars $c$ and $c'$ in the definition of a
solsoliton satisfy $c=c'$, and therefore $||\mu||=||\mu'||$ by Proposition
\ref{extras}, (iii) and the fact that $\mu$ and $\mu'$ belong to the same
$\Gl(\ngo)$-orbit and consequently they must lie in the same stratum $\sca_{\beta}$
(see Theorem \ref{strata}). It then follows from (i) and the uniqueness for
nilsolitons (see either \cite[Theorem 3.5]{soliton} or \cite[Theorem
4.2]{cruzchica}) that $h\in\Or(\ngo)$.
If $\ngo=\ngo_1\oplus...\oplus\ngo_r$ is the orthogonal decomposition with
$[\ngo,\ngo]=\ngo_2\oplus...\oplus\ngo_r$,
$[\ngo,[\ngo,\ngo]]=\ngo_3\oplus...\oplus\ngo_r$, and so on, then by (iii) we have
that for all $i$, $\ngo_i$ is an invariant subspace for any $\ad{A}|_{\ngo},
\ad'{A}|_{\ngo}$, $A\in\ag$, and also for $h$. But since
$\ad_{\mu'}(jg^{-1}A)\ngo_i\subset\ngo_{i+1}\oplus...\oplus\ngo_r$ for all $i$,
condition (\ref{isoa}) implies that $\ad_{\mu'}(jg^{-1}A)=0$ for all $A\in\ag$ and
consequently
$$
\tilde{f}=\left[\begin{array}{cc} g&0\\
0&h\end{array}\right]
$$
is also an isomorphism between $\lb$ and $\lb'$. We finally use (\ref{isoa}) and
(iv) to obtain that
$$
-c\la g^{-1}A,g^{-1}A\ra = \tr{S(\ad(g^{-1}A)|_{\ngo})^2} =
\tr{S(\ad'{A}|_{\ngo})^2} = -c\la A,A\ra,
$$
that is, $g\in\Or(\ag)$. Thus $\tilde{f}$ is an orthogonal isomorphism and so it
determines an isometry between $S$ and $S'$, which concludes the proof of the
theorem.
\end{proof}
It follows from Corollary \ref{cormain} that to classify solsolitons up to isometry
one firstly has to classify nilsolitons and for each of these to consider all
possible abelian Lie algebras of symmetric derivations. The following result gives
us the precise equivalence relation that must be considered between such algebras.
\begin{proposition}\label{toclas}
Let $(\ngo,\ip_1)$ be a nilsoliton and consider two solsolitons $S$ and $S'$
constructed as in {\rm Proposition \ref{const}} for abelian subalgebras
$$
\ag,\ag'\subset\Der(\ngo)\cap\sym(\ngo,\ip_1),
$$
respectively. Then $S$ is isometric to $S'$ if and only if there exists
$h\in\Aut(\ngo)\cap\Or(\ngo,\ip_1)$ such that $\ag'=h\ag h^{-1}$.
\end{proposition}
\begin{proof}
If $S$ and $S'$ are isometric then we can argue as in the proof of Theorem \ref{uni}
to obtain that $h\in\Aut(\ngo)\cap\Or(\ngo,\ip_1)$ (see (\ref{ison})) and that
$\ag'=h\ag h^{-1}$ by (\ref{isoa}).
Conversely, if we define $g:\ag\longrightarrow\ag'$ by $gA=hAh^{-1}$, then
$$
\la gA,gA\ra' = -\tfrac{1}{c}\tr{(gA)^2} = -\tfrac{1}{c}\tr{(hAh^{-1})^2} =
-\tfrac{1}{c}\tr{A^2} = \la A,A\ra.
$$
This implies that $f:=\left[\begin{array}{cc} g&0\\0&h\end{array}\right]$ is an
isometric isomorphism between $(\sg,\ip)$ and $(\sg',\ip')$ and thus $f$ defines an
isometry between $S$ and $S'$, as was to be shown.
\end{proof}
\section{Examples of solsolitons}\label{exa}
Once we have chosen our favorite nilsoliton, it is quite easy to get examples of
solsolitons by using Proposition \ref{const}. The results in Section
\ref{structure} even tell us that any solsoliton can be constructed in this simple
way. We consider in this section the problem of which simply connected solvable Lie
groups of dimension $\leq 4$ admit a solsoliton.
\subsection{Dimension $3$}
It is well known that for any $3$-dimensional real solvable Lie algebra which is not
nilpotent there exists a basis $\{ A,X_1,X_2\}$ such that
\begin{equation}\label{par1}
\begin{array}{l}
\rg_{\alpha}: \quad [A,X_1]=X_1, \quad [A,X_2]=\alpha X_2, \qquad -1\leq\alpha\leq 1, \\ \\
\sg_{\beta}: \quad [A,X_1]=X_2, \quad [A,X_2]= -X_1+\beta X_2, \qquad 0\leq\beta\leq
2.
\end{array}
\end{equation}
The constraints on the parameters $\alpha$ and $\beta$ guarantee that these Lie
algebras are in addition pairwise non-isomorphic (see \cite{Jcb}). It follows from
Proposition \ref{const} that the inner product $\ip$ for which $\{ A,X_1,X_2\}$ is
orthonormal is a solsoliton on $\rg_{\alpha}$ for all $\alpha$. Recall that in all
cases we have $\sg=\ag\oplus\ngo$ for $\ag=\RR A$ and $\ngo=\RR X_1+\RR X_2$, $\ngo$
abelian. For $\sg_{\beta}$, the matrix of $\ad{A}|_{\ngo}$ relative to $\{
X_1,X_2\}$ equals $\left[\begin{smallmatrix} 0&-1\\
1&\beta\end{smallmatrix}\right]$, which has eigenvalues $\tfrac{\beta}{2}\pm
i\sqrt{1-\tfrac{\beta^2}{4}}$, and thus for any $0\leq\beta<2$, there is a basis $\{
Y_1,Y_2\}$ of $\ngo$ with respect to which $\ad{A}|_{\ngo}$ has the normal matrix
$$
\left[\begin{smallmatrix} \tfrac{\beta}{2}&-\sqrt{1-\tfrac{\beta^2}{4}}\\
\sqrt{1-\tfrac{\beta^2}{4}}&\tfrac{\beta}{2}\end{smallmatrix}\right].
$$
This implies that the inner product $\ip_1$ given by $\la
A,A\ra_1=\tr{S(\ad{A})^2}=\tfrac{\beta^2}{2}$ and with $\{ Y_1,Y_2\}$ as an
orthonormal basis for $\ngo$ is a solsoliton for $0<\beta<2$ (see Theorem
\ref{main}). But hence no new example appears, since by (\ref{sympart}) the
solsolitons $(\sg_{\beta},\ip_1)$ are either isometric to the hyperbolic space
$H^3=(\rg_1,\ip)$ for $0<\beta<2$ or to the euclidean space $\RR^3$ for $\beta=0$.
The only remaining case is $\sg_2$, for which we have that $\ad{A}|_{\ngo}$ is not
diagonalizable over $\CC$ and so it can not be normal with respect to any inner
product on $\ngo$. We conclude from Theorem \ref{main} that $\sg_2$ is the only
$3$-dimensional solvable Lie group which does not admit a solsoliton.
{\small
\begin{table}
$$
\begin{array}{ccccccc}
\hline\hline
&&&&&& \\
& \ngo & \ad{A}|_{\ngo} & constraints & unimodular & solsoliton & Einstein \\ \\
\hline
&&&&&&\\
\rg_3 & \RR^2 & \left[\begin{smallmatrix} 1&1\\ &1\end{smallmatrix}\right] & - & - & - & - \\ \\
\rg_{3,\lambda} & \RR^2 & \left[\begin{smallmatrix} 1&\\ &\lambda\end{smallmatrix}\right] & -1\leq\lambda\leq 1 & \lambda=-1 & \checkmark & \lambda=1 \\ \\
\rg_{3,\lambda}' & \RR^2 & \left[\begin{smallmatrix} \lambda&1\\ -1&\lambda\end{smallmatrix}\right] & 0\leq\lambda & \lambda=0 & \checkmark & \checkmark\\ \\
\hline \hline \\
\end{array}
$$
\caption{Classification of $3$-dimensional solvable Lie algebras admitting
solsolitons}\label{s3}
\end{table}}
There are many ways to parametrize $3$-dimensional solvable Lie algebras over $\RR$
other than (\ref{par1}). It is in fact more convenient for our purpose to use the
description given in \cite[Theorem 1.1]{AndBrbDttOvn}, which we show in Table
\ref{s3} together with the information about the existence of solsolitons. In this
case, if $\{ X_1,X_2\}$ is the basis of $\ngo=\RR^2$ that we use to write
$\ad{A}|_{\ngo}$, then the inner product $\ip$ making of $\{ A,X_1,X_2\}$ an
orthonormal basis is always a solsoliton.
We note that for any $0\leq\lambda$, $\rg_{3,\lambda}'$ is isomorphic to
$\sg_{\beta}$ for $\beta=\tfrac{2\lambda}{\sqrt{\lambda^2+1}}$ and $\rg_3$ is
isomorphic to $\sg_2$. A third description can be found in \cite{Mln}.
\subsection{Dimension $4$}
It is enough to consider $4$-dimensional real solvable Lie algebras which are not
either nilpotent or a direct sum of two Lie algebras. If $\sg=\ag\oplus\ngo$ with
$\ngo$ the nilradical of $\sg$, then the only one with $\dim{\ag}=2$ is
$\affg(\CC)$, which is defined for a basis $\{ A_1,A_2\}$ of $\ag$ by
$$
\ad{A_1}|_{\ngo}=\left[\begin{smallmatrix} 1&0\\ 0&1\end{smallmatrix}\right], \quad
\ad{A_2}|_{\ngo}=\left[\begin{smallmatrix} 0&-1\\ 1&0\end{smallmatrix}\right].
$$
It follows from Theorem \ref{main} that $\affg(\CC)$ admits a solsoliton, which is
isometric to $H^3\times\RR$ by (\ref{sympart}).
Any other has $\dim{\ag}=1$, say $\ag=\RR A$, and let $\{ X_1,X_2,X_3\}$ be a basis
of $\ngo$, which will be assumed to satisfy $[X_1,X_2]=X_3$ when $\ngo$ is not
abelian (i.e. $\ngo$ isomorphic to the $3$-dimensional Heisenberg Lie algebra
$\hg_3$). We have used the classification given in \cite[Theorem 1.5]{AndBrbDttOvn}
to summarize all the relevant information in Table \ref{n3}, which has been obtained
as in the $3$-dimensional case above by a direct application of Proposition
\ref{const} and Theorem \ref{main}. The map $\ad{A}|_{\ngo}$ is always written in
terms of the basis $\{ X_1,X_2,X_3\}$ and the inner product $\ip$ for which $\{
A,X_1,X_2,X_3\}$ is an orthonormal basis is always a solsoliton, when there is one.
{\small
\begin{table}
$$
\begin{array}{ccccccc}
\hline\hline
&&&&&& \\
& \ngo & \ad{A}|_{\ngo} & constraints & unimodular & solsoliton & Einstein \\ \\
\hline
&&&&&&\\
\rg_4 & \RR^3 & \left[\begin{smallmatrix} 1&1&\\ &1&1\\ &&1\end{smallmatrix}\right] & - & - & - & - \\ \\
\rg_{4,\lambda} & \RR^3 & \left[\begin{smallmatrix} 1&&\\ &\lambda&1\\ &&\lambda\end{smallmatrix}\right] & -\infty<\lambda<\infty & \lambda=-\unm & - & - \\ \\
\rg_{4,\mu,\lambda} & \RR^3 & \left[\begin{smallmatrix} 1&&\\ &\mu&\\ &&\lambda\end{smallmatrix}\right] & \begin{array}{l} -1<\mu\leq\lambda\leq 1; \\ -1=\mu\leq\lambda<0\end{array} & \mu=-1-\lambda & \checkmark & \mu=\lambda=1\\ \\
\rg_{4,\mu,\lambda}' & \RR^3 & \left[\begin{smallmatrix} \mu&&\\ &\lambda&1\\ &-1&\lambda\end{smallmatrix}\right] & 0<\mu & \mu=-2\lambda & \checkmark & \mu=\lambda \\ \\
\sg_4 & \hg_3 & \left[\begin{smallmatrix} 1&&\\ &-1&\\ &&0\end{smallmatrix}\right] & - & \checkmark & \checkmark & - \\ \\
\sg_{4,\lambda} & \hg_3 & \left[\begin{smallmatrix} \lambda&&\\ &1-\lambda&\\ &&1\end{smallmatrix}\right] & \unm\leq\lambda & - & \checkmark & \lambda=\unm \\ \\
\sg_{4,\lambda}' & \hg_3 & \left[\begin{smallmatrix} \lambda&1&\\ -1&\lambda&\\ &&2\lambda\end{smallmatrix}\right] & 0\leq\lambda & \lambda=0 & \lambda\ne 0 & \lambda\ne 0 \\ \\
\hg_4 & \hg_3 & \left[\begin{smallmatrix} 1&1&\\ &1&\\ &&2\end{smallmatrix}\right] & - & - & - & - \\ \\
\hline \hline \\
\end{array}
$$
\caption{Classification of $4$-dimensional solvable Lie algebras with a
$3$-dimensional nilradical admitting a solsoliton}\label{n3}
\end{table}}
By using (\ref{sympart}), we get that $\rg_{4,\mu,\lambda}'$, $\lambda\ne 0$, is
isometric to $\rg_{4,\lambda/\mu,\lambda/\mu}$ and that $\sg_{4,\lambda}'$ is
isometric to $\sg_{4,1/2}$ for any $\lambda>0$.
One can see in Table \ref{n3} that $\rg_{4,-1/2}$ is the only unimodular solvable
Lie algebra of dimension $4$ which does not admit a solsoliton. It follows however
from the results in \cite{Hng} that $\rg_{4,-1/2}$ does not either admit a {\it
lattice} (i.e. cocompact discrete subgroup). Thus the universal cover of any
$4$-dimensional compact solvmanifold $S/\Gamma$ does admit a solsoliton, a result
which has already been proved in \cite{IsnJckLu}. This is essentially due to the
fact from algebraic number theory that any $A\in\Sl_3(\ZZ)$ with positive real
eigenvalues and at least one of them different from $1$ is necessarily
diagonalizable over $\CC$. The next example shows that this is no longer true for
compact solvmanifolds of dimension $\geq 5$.
\begin{example}
Let $\sg=\RR A\oplus\ngo$ be the solvable Lie algebra defined by: $\ngo$ is abelian,
$\dim{\ngo}=4$ and
$$
\ad{A}|_{\ngo}= \left[\begin{smallmatrix} 0&1&&\\ 0&0&&\\ &&\ln{\lambda}&\\
&&&-\ln{\lambda}\end{smallmatrix}\right], \qquad\lambda=\tfrac{3+\sqrt{5}}{2}.
$$
Thus there exists $\sigma\in\Gl_4(\RR)$ such that
$$
\sigma e^{\ad{A}|_{\ngo}}\sigma^{-1} = \left[\begin{smallmatrix} 1&1&&\\ 0&1&&\\
&&2&1\\ &&1&1\end{smallmatrix}\right]\in\Sl_4(\ZZ),
$$
and it is therefore easy to check that $\Gamma:=\ZZ\ltimes\sigma^{-1}\ZZ^4$ is a
lattice of the solvable Lie group $S$ with Lie algebra $\sg$. Recall that
$S=\RR\ltimes\RR^4$ with multiplication given by
$$
(t,X).(s,Y) = (t+s, X+e^{t\ad{A}|_{\ngo}}Y), \qquad t,s\in\RR, \quad X,Y\in\RR^4.
$$
However, since $\ad{A}|_{\ngo}$ is not diagonalizable over $\CC$, we conclude from
Theorem \ref{main} that $S$ can never admit a solsoliton.
\end{example}
|
\section{Introduction}
First {\it IRAS}\ and {\it ISO}, then SCUBA and {\it COBE}, and most recently
{\it Spitzer}\ have convincingly shown that most of
the star formation in the universe is enshrouded in dust.
Large populations of infrared-luminous galaxies at $z\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$1.5
contribute to about 70-80\% of the far-infrared
and 30\% of the sub-millimeter backgrounds,
and are probably responsible for most of the star-formation activity at
high redshift
\citep[e.g.,][]{chary01,lefloch05,dole06}.
Indeed, half the energy and most of the photons in
the universe come from the infrared spectral region
\citep[e.g.,][]{hauser01,franceschini08}.
That dust is so prominent in the high-redshift universe
may appear surprising, since it has been assumed that
dust would be absent in low-metallicity environments.
However, recent theoretical developments challenge the assumption
that dust is absent in metal-poor gas.
Large amounts of dust can be created on short timescales
by supernovae (SNe)
and Asymptotic Giant Branch stars which evolve in a metal-free
ISM \citep[e.g.,][]{todini01,schneider04,nozawa07,bianchi07,valiante09}.
Observationally,
the copious dust emission observed at very high redshifts
($z\simgt6$) in quasars implies that indeed
dust can form rapidly at early epochs \citep{bertoldi03}.
Although luminous quasars can hardly be considered as
typical examples of star-forming environments, observations
of sub-millimeter galaxies \citep[SMGs,][]{chapman05} and
dust-obscured galaxy populations \citep[DOGs,][]{dey08}
also suggest that early star formation episodes can also be very intense
and relatively brief.
However, exactly how these massive starbursts occur and evolve
is not yet clear.
The short interval in which star formation and the ensuing chemical enrichment
and dust formation convert a dust-free metal-free environment to a dusty
metal-rich one by redshift $\sim$6 is as yet unobserved, and studies of
such transitions remain a major observational challenge.
The Local Universe is home to star-forming dwarf galaxies that are of much
lower metallicity than those observed thus far at high redshift.
Over the last thirty years or so, only a handful
of blue compact dwarf (BCD) galaxies have been discovered
with 12$+$log(O/H)$\sim$7.2. Despite intensive searches, only one
BCD, SBS\,0335$-$052\,W, is known to host \rm H{\sc ii}\ regions with 12$+$log(O/H)$\leq$7.1 \citep{izotov09}.
Even at these extremely low metallicities,
BCD spectral energy distributions (SEDs) can be dominated by
infrared dust reprocessing of ultraviolet radiation
from young massive stars
\citep[e.g., SBS\,0335$-$052\,E\ and I\,Zw\,18:][]{thuan99a,houck04,wu07}.
Because such galaxies are chemically unevolved, they
can provide a window on primordial galaxy formation and evolution.
This is, in some sense, a ``local'' approach to a cosmological problem.
If we can study the properties of a metal-poor ISM and its constituents
locally, we may be able to better understand the high-redshift transition from
metal-free Population\,III (Pop III) stars to the chemically evolved
massive galaxies typical of the current epoch.
We adopt here this ``local'' approach.
With the aim of characterizing the properties of star and dust formation in
a metal-poor ISM,
we have used the {\it Spitzer}\ Space Telescope to study
a sample of nearby low-metallicity star-forming dwarf galaxies.
This is the third in a series of papers;
the first two papers have focused on two particular BCDs, Haro 3, the highest
metallicity object in our sample
\citep{hunt06}, and Mrk 996, a BCD with an unusually dense and
compact nuclear \rm H{\sc ii}\ region \citep{thuan08}.
Here, we present for the first time the entire sample of 23 BCDs, and
report on its spectral properties, as derived from IRS observations\footnote{We
present IRS results for only 22 BCDs, since one, SBS\,0940$+$544, was observed
in a Guaranteed Time Program.}.
The sample has been carefully chosen to span a wide range in
oxygen abundance: 12$+$log(O/H)\ varies between 7.4 and 8.3, with a median of
7.9. About a quarter
of the objects in the sample are in the eXtremely Metal-Deficient (XMD)
or Extremely Metal-Poor Galaxy (EMPG) range
\citep[12$+$log(O/H)$\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$7.65][]{pustilnik07,brown08}.
Thus our sample comprises the largest number of XMDs/EMPGs to date with
good signal-to-noise IRS spectra.
In $\S$\ref{sec:observations}, we describe the sample selection criteria,
together with the {\it Spitzer}\ observations and the data reduction methods.
In $\S$\ref{sec:analysis},
we derive fluxes of lines and features in the spectra by fitting them with
PAHFIT \citep{smith07}.
We present and discuss our results for
the infrared (IR) fine-structure (FS) lines in $\S$\ref{sec:ionized},
the aromatic features in $\S$\ref{sec:aromatics},
and in $\S$\ref{sec:molecules}, the molecular lines.
Our conclusions are summarized in $\S$\ref{sec:summary}.
\section{Sample Definition, Observations, and Data Reduction
\label{sec:observations}}
One of the most important factors in shaping the SED at low metallicity
is the hardness and intensity of the interstellar radiation field (ISRF).
The ISRF hardness and intensity govern
dust properties such as the grain size distribution and temperature.
Hard and intense ISRFs may destroy small grains all together,
thus suppressing the aromatic or polycyclic aromatic hydrocarbon features
(PAHs) which are typical of metal-rich starburst galaxies
\citep[e.g.,][]{voit92,madden06,smith07}.
In the post-{\it ISO}\ and post-{\it Spitzer}\ eras, ratios of IR lines of neon
have often been used
to quantify the hardness of the ISRF in star-forming
galaxies \citep[e.g.,][]{thornley00,dale06}.
However, since we have selected the objects in our sample on the basis
of their optical spectra,
we have quantified ISRF hardness optically, with the ratio
of the nebular \rm He{\sc ii}\ $\lambda$\,4686 emission line relative to H{$\beta$}.
The \rm He{\sc ii}\ line, with an ionization potential of 54.4\,eV,
is a good indicator of a hard ultraviolet (UV) radiation field.
Nebular \rm He{\sc ii}\ lines are predicted to be very strong in metal-free
(or nearly metal- free) primordial galaxies \citep{schaerer02,schaerer03},
although
no \rm He{\sc ii}-emitting high-redshift galaxy (z$\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$5) has yet been discovered
\citep[e.g.,][]{nagao08}.
However, some low-redshift low-metallicity BCDs are known to show
strong nebular \rm He{\sc ii}\ $\lambda$4686 emission with \rm He{\sc ii}/H{$\beta$}$\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$3\%
\citep []{guseva00, izotovthuan04, thuanizotov05}.
We have selected 23 objects so that they span a wide range in \rm He{\sc ii}/H{$\beta$}.
Roughly 1/3 of the sample consists of objects having strong \rm He{\sc ii}\ emission,
with \rm He{\sc ii}/H{$\beta$}$\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$2\%;
another third have intermediate \rm He{\sc ii}\ emission with
\rm He{\sc ii}/H{$\beta$}$\sim$1-2\%,
and the remaining third has weak or no \rm He{\sc ii}\ emission, with
\rm He{\sc ii}/H{$\beta$}$\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$1\%.
This wide range of \rm He{\sc ii}\ strength in our sample
also provides a large range in metallicity as
there is a loose correlation between the two quantities \citep{thuanizotov05}.
The general properties of the sample are given in Table \ref{tab:sample}.
The observations presented in this paper are part of our GO
{\it Spitzer}\ program (PID 3139) for 23 BCDs. In the course of this program,
we have obtained IRS spectra
in the low- and high-resolution modules \citep[SL, SH, LH][]{houckirs},
IRAC images at 4.5 and 8\,\micron\ \citep{fazioirac},
and MIPS images at 24, 71, and 160\,\micron\ \citep{riekemips}.
For all instruments, the data reduction starts with
the Basic Calibrated Data ({\it bcd}).
The corresponding masks (the DCE masks),
furnished by the {\it Spitzer}\ Science Center pipeline,
were then used to flag
potential spurious features in the images, such as
strong radiation hits, saturated pixels, or nonexistent/corrupted data.
\subsection{IRS spectroscopy}
Spectroscopy was performed in the staring mode,
in both orders of the Short Low-resolution module (SL1, SL2)
and with the Short and Long High-resolution modules (SH, LH) \citep{houckirs}.
Thus, we have obtained low-resolution spectra from 5.2 to 14.5\,\micron\ (R$\simeq$64-128), and
high-resolution spectra from 9.6 to 37.2\,\micron\ (R$\simeq$600).
We have used different acquisition schemes, depending on the
brightness of the source as measured by its
scaled H{$\beta$}\ flux.
Sources were centered in the slits by peaking up on 2MASS stars.
The individual {\it bcd} frames were processed by the S13.2.0 version of the
SSC pipeline, which performs ramp fitting, dark current
subtraction, droop and linearity corrections,
flat-fielding, and wavelength and flux calibrations\footnote{See the IRS
Data Handbook, \url{http://ssc.spitzer.caltech.edu/irs/dh}.}.
Because the pipeline does not include background subtraction, we
have constructed,
for the low-resolution long-slit spectra, a coadded
background frame from the {\it bcd} observations
with the source in the opposing nod and off-order positions
\citep[see also][]{weedman06}.
Coadding was performed with the sigma-clipping option
of the {\it imcombine} task
in IRAF\footnote{IRAF is the Image Analysis and
Reduction Facility made available to the astronomical community by the National Optical
Astronomy Observatory, which is operated by AURA, Inc., under
contract with the U.S. National Science Foundation.}.
The inclusion of off-order and off-nod frames in the background image
means that
the total integration time on the background is three times that on the source,
thus improving the signal-to-noise of the two-dimensional (2D) subtraction.
We did not acquire a separate background spectrum
for the high-resolution SH and LH spectra, and the slit size precludes
using the same procedure as for the SL module.
Therefore we subtracted the background from the SH and LH observations
using the one-dimensional (1D) spectra as described below.
The SH and LH 2D {\it bcd} images were coadded in the same way
as for the SL modules,
and corrections for sporadic bad pixels and cosmic ray
hits were carried out manually by inspection of the images at
the separate nod positions.
We extracted the source spectra with {\it SPICE}, the post-pipeline IRS package
furnished by the SSC.
Before extraction, all spectra at separated nod positions were cleaned with the {\it IRSCLEAN}\
algorithm (http://ssc.spitzer.caltech.edu/postbcd/irsclean.html).
The automatic point-source extraction window was used for all modules.
Automatic extraction uses a variable-width extraction window which
scales with wavelength in order to recover a constant fraction of
the diffraction-limited instrument response.
For the SL spectra, this gives a 4-pixel (7\farcs2) length at 6\,\micron, and
an 8-pixel length (14\farcs4) at 12\,\micron;
the slit width is 3\farcs6 for both SL modules.
At high resolution, the {\it SPICE}\ extraction was performed over the entire slit
(4\farcs7$\times$11\farcs3 SH; 11\farcs1$\times$22\farcs3 LH).
Orders were spliced together by averaging, ignoring the noisy
regions at the red end of each order \citep[e.g.,][]{armus04}.
Then the individual spectra were box-car smoothed to a resolution element,
and clipped in order to
eliminate any remaining spikes in the high-resolution data.
Finally, the two spectra for each module (one for each nod position)
were averaged.
Background was subtracted from the 1D SH spectra by
defining a level from the overlap region with the SL spectra,
from which we had subtracted a 2D background.
Specifically, the
difference between the 2D background-subtracted SL and SH spectra
was minimized over
the substantial region where they overlap in wavelength ($\sim$5\,\micron).
The spectral shape of the background was given by
the model of Reach and coworkers\footnote{See
\url{http://ssc.spitzer.caltech.edu/documents/background}.}, and
the multiplicative constant by the minimization process.
Subsequently,
the LH background was subtracted in an analogous
way, by minimizing the difference
between the SH and LH spectra over their overlap region ($\la$1\,\micron).
This ``bootstrapping'' procedure for background subtraction
in the high-resolution modules should in principle
eliminate the need for scaling adjustments.
However, its accuracy depends on the unresolved nature of the sources
in the spectral apertures.
We measured the source sizes by fitting Gaussians to the
IRAC 7.9\,\micron\ and MIPS 24\,\micron\ images.
The mean (and median) image FWHM is 3\farcs7 at 7.9\,\micron, and 6\farcs2 at 24\,\micron.
The SL1 slit width is 3\farcs6, so most of the sources are sufficiently
compact to not significantly exceed the SL1 slit size.
The resolution of MIPS24 is $\sim$6\arcsec, making the BCDs essentially unresolved
at these longer wavelengths;
in fact, the characteristic diffraction pattern is seen in all the MIPS24 images.
The FWHM of the largest source at 24\,\micron\ (Tol\,1924$-$416) is 10\arcsec,
so even in the most extreme case,
the LH slit width of 11\farcs1 completely encompasses the BCD light
at 24\,\micron.
Our assumption of point-like morphologies at {\it Spitzer}\ wavelengths thus
appeared to be reasonably justified.
Nevertheless, for a few (of the largest) sources,
the level of the background needed to be adjusted
slightly to provide a smooth transition among the modules.
In any case, the 24\,\micron\ flux as measured from MIPS is generally
consistent with the continuum level derived for the IRS spectra (see below).
We have obtained IRAC and MIPS photometry for all 23 BCDs in our
sample.
However, we have IRS observations for only 22 BCDs.
One BCD, SBS\,0940$+$544, has an IRS observation in the Guaranteed Time
program (PID\,85). Thus, we present here IRS results for only 22 objects, and
will incorporate SBS\,0940$+$544 into our sample in a future paper.
\subsection{IRAC and MIPS Imaging}
Although the focus of this paper is on the IRS spectra of our sample
objects, we briefly
describe their {\it Spitzer}\ photometric data because we need photometric
quantities such as the
total IR luminosity ($L_{\rm TIR}$) to properly interpret their spectral properties.
A more detailed description of the photometry and reduction
procedures can be found in \citet[][Haro\,3]{hunt06} and
\citet[][Mrk\,996]{thuan08}. The discussion of the photometric data of
the whole sample is deferred to a future paper
on SEDs.
Briefly, we acquired IRAC images in two filters, 4.5 and 7.9\,\micron,
each with two sets of four frames in the high dynamic range mode.
The individual {\it bcd} frames were processed with
the S14.0.0 version of the SSC pipeline
(see the IRAC Data Handbook\footnote{Available from the SSC website
\url{http://ssc.spitzer.caltech.edu/irac/dh/}}).
The {\it bcd} frames (corrected for ``banding'' effects for bright sources)
were coadded using {\it MOPEX}, the image mosaicing and source-extraction package
provided by the SSC \citep{mopex}.
Pixels flagged by the masks were subsequently ignored.
Additional inconsistent pixel values were removed by means of the {\it MOPEX}\ outlier
rejection algorithms, in particular
the dual-outlier technique, together with the multiframe algorithm.
The frames were corrected for geometrical distortion and
projected onto a ``fiducial'' (refined) coordinate system
with pixel sizes of 1\farcs20, roughly equivalent to the original pixels.
Standard linear interpolation was used for the mosaics.
The noise levels in our post-pipeline {\it MOPEX}\ IRAC mosaics are comparable to
or lower than those in the SSC products.
Our MIPS images were acquired in the Fixed Cluster-Offset mode in all
three channels, with offsets of 12\arcsec\ in two additional
pointings.
The individual {\it bcd} frames were processed by the
the S14.4.0 version of the SSC pipeline,
which converts the integration ramps inherent to the MIPS detectors into
slopes, and corrects for temporal variations of the slope images
\citep{gordon05}.
As for the IRAC images, we processed the dithered {\it bcd} frames in
the spatial domain with {\it MOPEX}.
The DCE masks and the static masks were used to flag pixels which were
subsequently ignored.
The {\it MOPEX}\ outlier rejection was used to remove any additional spurious pixel
values, with the dual-outlier and multiframe algorithm as for the IRAC frames.
Geometrical distortion was corrected before projecting the frames onto
a fiducial coordinate system with pixel sizes of 1\farcs20 for MIPS24,
roughly half of the original pixel size of 2\farcs5.
Pixel sizes of the final mosaics at 71 and 160\,\micron\ are also
approximately half
the originals, i.e. 4\farcs95 at 71\micron\ and 8\farcs0 at 160\,\micron.
Unlike the IRAC coadds,
we incorporated the sigma-weighting algorithm because
it gave less noisy MIPS mosaics than without.
Standard linear interpolation was used in all cases.
In all three channels, our {\it MOPEX}\ mosaics are superior to
those provided by the automated post-pipeline reduction.
\subsection{\label{sec:phot} IRAC and MIPS Photometry}
We have performed aperture photometry on the IRAC and MIPS images
with the IRAF photometry package {\it apphot}, taking care to convert
the MJy/sr surface brightness units of the images to integrated flux units.
The background level was determined by averaging several
adjacent empty sky regions.
Fluxes were determined at radii where the growth curve has
become asymptotically flat.
In many cases, there are multiple sources within the IRS apertures,
and we used subjective judgment to match photometry with the spectroscopic
slit.
Following \citet{draine07}, these fluxes were used to derive
the ``color temperature'' between 71 and 160\,\micron,
$\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$,
and the total infrared flux (TIR):
$$L_{\rm TIR}\,=\,0.95\,\langle\nu L_\nu\rangle_{7.9}
+ 1.15\,\langle\nu L_\nu\rangle_{24}
+ \langle\nu L_\nu\rangle_{71}
+ \langle\nu L_\nu\rangle_{160} $$
\noindent
Both quantities will be pertinent for the discussion of the spectra.
In order to check our
reduction and calibration procedures for both the IRS and MIPS,
we have compared the MIPS 24\,\micron\ total fluxes with the
continuum levels in the IRS spectra.
Such a comparison (see below)
shows good general agreement, implying that both the
IRS flux calibration (in the case of a point-source) and background
subtraction and the derivation of the total MIPS flux have been done correctly.
The only exception is UM\,311, which is in a crowded field (there are several
point sources in close proximity to a spiral galaxy), and thus
its photometry is probably contaminated by other sources.
All targets were detected at 24 and 71\,\micron.
However, at 160\,\micron, we obtained only upper limits (ULs) for 6 of them.
As a consequence of a {\it Spitzer}\ queue problem,
SBS\,1030$+$583 has no MIPS24 observations.
\section{Spectral Analysis \label{sec:analysis}}
The flux emitted by aromatic (or PAH) features constitutes a large fraction
of the IR energy budget in a typical star-forming galaxy spectrum.
However, these features
are known to decrease in intensity or even disappear at low
metallicities \citep{engelbracht05,madden06,wu06,engelbracht08}.
Our sample contains many metal-deficient objects and
is thus uniquely able to address the behavior of
PAH emission at low metallicities. In order to directly compare the results for
our low-metallicity sample with those for
an existing well-defined set of more metal-rich galaxies,
we wanted to analyze our IRS spectra in the same manner as this latter set.
This consideration dictated the use of PAHFIT \citep{smith07},
an IDL procedure which was developed for and
applied to the Spitzer Nearby Galaxy Survey \citep[SINGS,][]{kennicutt03}.
PAHFIT is particularly suited for separating emission lines
and PAH features (e.g., the PAH blend at 12.6--12.7\,\micron\ and the \rm [Ne\,{\sc ii}]\
line at 12.8\,\micron), as well as measuring faint PAH features superimposed
on a strong continuum.
Broad ``plateaus'' underlying the 8\,\micron\
and 16-17\,\micron\ emission \citep{peeters02,peeters04b}
and strong silicate absorption at 9.7\,\micron\
can make the measurement of broad PAH bands exceedingly difficult in
those wavelength regions.
PAHFIT overcomes this problem by fitting simultaneously the spectral features,
the underlying continuum, and the extinction, and gives
quantitative estimates and uncertainties for all parameters.
Drude profiles are fitted to the PAH features, and Gaussian profiles
to the molecular hydrogen and fine-structure lines.
In addition to PAHFITting the spectra,
we also measured emission line fluxes over the entire IRS range
by fitting Gaussian profiles.
Our strategy for deriving quantitative measurements from the IRS spectra is described in the following.
\subsection{PAHFIT \label{sec:pahfit}}
Our spectra have a higher resolution than those normally fitted with PAHFIT,
specifically for wavelengths $\geq$14.5\,\micron\ where we have no
low-resolution spectra.
Hence, after binning the spectra to a 0.03\,\micron\ resolution,
and fitting the entire spectrum, we use only
the spectral lines measured with PAHFIT having $\lambda$ $\leq$
14.5\,\micron.
In contrast, we adopt all the dust features fitted by PAHFIT
over the entire IRS range.
We modified the PAHFIT IDL code to allow the central wavelengths
and widths of the Drude profiles to vary, and
experimented with different weighting schemes.
We also tried several different combinations of dust continuum
temperatures, but these did not influence the fit results.
PAHFIT was run with three different options, including
mixed extinction, screen extinction, and fixing extinction to zero.
The nominally best fits (those with lowest $\chi_\nu^2$) for 11 BCDs
had non-zero extinction.
However, in every case but Haro\,3 and Mrk\,996,
the statistical significance of the fit was comparable to similar fits,
but with zero extinction.
Hence, when possible, we opted for the fit with no extinction
[$\tau(9.7)$\,=\,0].
The results are shown in Figure \ref{fig:pahfit1}.
The top panels show the best-fit PAHFIT model as a red curve,
and the bottom panels give the residuals.
The reduced $\chi_\nu^2$ given in each panel in the figures is
calculated from
the residuals of the fit up to 14.5\,\micron, normalized by the
mean spectral uncertainty over the same spectral range.
The figures show that the residuals are generally
quite well-behaved, and that $\chi_\nu^2$ is generally sufficiently
small to imply a good fit.
Uncertainties on the integrated line fluxes were estimated in two ways.
The standard deviation $\sigma$ per pixel
of the continuum was measured individually in a relatively clean region of each SL,
SH, and LH spectrum.
A first estimate of the uncertainty in the line flux,
assumed to be spread over 3 pixels,
is taken to be $3\times3\sigma$. This is compared with a
second uncertainty
estimate, the one given by PAHFIT for each spectral feature. The largest
of the two estimates was then chosen to be the final uncertainty of
the measurement.
Table \ref{tab:fspahfit}
reports all FS lines with $\lambda\leq$14.5\,\micron\
detected at the $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ level with PAHFIT, and
Table \ref{tab:dfpahfit} lists the dust features.
\subsection{Gaussian Fitting of Emission Lines \label{sec:gauss}}
We fitted all possible molecular and FS emission lines with Gaussian profiles,
taking the integrated fitted flux as the measure of the total flux in the line.
When necessary, we used multiple Gaussian profiles to deblend
adjacent lines (e.g., \rm [O\,{\sc iv}]\ and \rm [Fe\,{\sc ii}]\ at $\sim$26\,\micron).
Linear continua were fitted simultaneously, taking care to define
the continuum regions well beyond the Gaussian wings.
In order for a spectral feature to be identified with a known emission
line, its rest-frame wavelength shift from the laboratory wavelength had to
to be $\leq$0.065\,\micron\
(for each object, we correct the observed spectrum to its rest frame by
using the optical redshift listed in Table \ref{tab:sample}).
Each emission line was measured several times, varying the
continuum positions and blending parameters, to obtain an additional
estimate of the uncertainty.
The uncertainty on the integrated flux was then taken to be the larger
of the two following estimates:
the standard deviation of the ensemble of different measurements, or
9 times the standard deviation $\sigma$ of the continuum fluctuations,
as described above (assuming the line is spread over 3 pixels).
To guard against false detections from bad pixels or spectral glitches,
we also required that the average of the two nod positions have a signal-to-noise
$\geq$3.
Table \ref{tab:fsgauss}
reports the long-wavelength FS fluxes we obtain from these measurements.
For the objects where \rm [O\,{\sc iv}]\ or \rm [Fe\,{\sc ii}]\ was detected,
blown-up parts of the spectra showing the 26\micron\ region
are given in Figure \ref{fig:oiv}.
Although we detect \rm [O\,{\sc iv}]\ at 3$\sigma$ (2$\sigma$) in 7 (9) galaxies,
\rm [Fe\,{\sc ii}]\ is detected only in 2 (5 at 2$\sigma$).
This result will be discussed in more detail in $\S$\ref{sec:oxygen_iron}.
The extremely high-excitation lines [Ne{\sc v}] (ionization potential
of 97.1\,eV) at 14 and 24\,\micron\ are not detected in our data for any BCD.
The MIPS 24\,\micron\ continuum level
is also shown in Fig. \ref{fig:oiv} by an open circle.
As stated in $\S$\ref{sec:phot}, the IRS spectra and MIPS 24\,\micron\
total fluxes are consistent, except for UM\,311 which is in a crowded
field.
In addition to the definite presence of \rm H$_2$,
a few spectra shown in Fig. \ref{fig:h2o_oh}
also tentatively suggest the presence of two other
molecules, {\rm OH}\ ($\sim$28.9, 30.3, 30.7\,\micron)
and {\rm H}$_2${\rm O}\ ($\sim$29.8 and 29.9\,\micron).
Table \ref{tab:h2} gives the \rm H$_2$\ line fluxes (including those measured
by PAHFIT for $\lambda\la$ 14.5\,\micron, see above),
and Table \ref{tab:molecules} reports the other molecules.
These detections will be discussed further in $\S$\ref{sec:molecules}.
\subsection{Comparison of Fitting Techniques \label{ref:comparison}}
To judge the reliability of our line flux measurements, we have compared the
fluxes of those emission lines that
fall in the short-wavelength region of the spectra
($\lambda\leq$\,14.3\,\micron), and that have been
measured in two different ways:
first by PAHFIT and second by individual Gaussian fits.
For strong lines with fluxes $\sim 2\times10^{-17}$\,W\,m$^{-2}$,
the difference between these two methods is $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$4\%;
for faint lines with fluxes $\sim 5\times10^{-18}$\,W\,m$^{-2}$,
the difference can be as large as 25\%.
The larger discrepancy for the fainter lines
can be attributed to the choice of different continuum levels in the
two methods.
This is the main reason why we preferred to use PAHFIT
for the short-wavelength
region. However, the overall agreement between the two methods is
sufficiently good to conclude that we can reliably
compare short- and long-wavelength
line fluxes, and that we do not introduce systematic differences
in measuring them with slightly different methods.
Although PAHFIT corrects for dust extinction, we have not
applied an analagous correction to the line
fluxes for $\lambda \geq$14.5\,\micron.
Such corrections would be applicable only to two sources,
Haro\,3 and Mrk\,996.
Haro\,3, with $\tau(9.7)$\,=\,0.39, would have a correction
$\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$25\% for wavelengths longer than 15\,\micron;
Mrk\,996 would have a 2\% correction.
Hence, because the correction for both objects are comparable to
or smaller than uncertainties in the line fluxes, we did not
apply them to the long-wavelength measurements.
\section{The Ionized Component of the ISM \label{sec:ionized}}
The spectral region covered by {\it Spitzer}/IRS provides a
wealth of diagnostics to probe the physics of the ISM at low metallicity.
The FS lines allow us to assess the physical conditions of the ISM,
including the spectral shape or hardness of the ISRF and the density of the ionized gas.
These quantities affect the dust properties by causing changes
in the temperature of the ``classical'' grains, and by altering
the stochastic emission from the aromatic features.
They constrain the molecular content through dissociation
and excitation processes.
\subsection{The Hardness of the Interstellar Radiation Field \label{sec:hardness}}
The IR FS lines span a wide range of ionization potentials
and are thus sensitive diagnostics of the ionized ISM in BCD \rm H{\sc ii}\ regions.
We focus first on diagnostics of the
ISRF hardness. In their discussion, we will attempt
to distinguish among the interdependent effects of
ISRF {\it hardness}, ISRF {\it intensity}, and nebular {\it metallicity}.
Traditionally, flux ratios of high-ionization species have been
used to measure the ISRF hardness, or equivalently the UV slope of its
spectrum, which is sensitive
to the effective temperatures of the ionizing stars.
In the IR range, the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ flux ratios are
usually adopted \citep[e.g.,][]{thornley00,giveon02}.
Such flux ratios also depend on the ionization parameter $U$, defined
as $U\,=\,Q/4\pi R_* n_{\rm e} c$, where $Q$ is the number of ionizing
photons, $n_{\rm e}$\ is the electron density, and $R_*$ is the radius of the
star-forming region \citep[e.g.,][]{thornley00}.
We first examine whether these ratios depend on metallicity as shown
in Fig. \ref{fig:nes_oh}.
The top panel shows that within our sample there may be a very slight dependence
of the \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ ratio on metallicity, although with a very large
scatter (there are no SINGS data for the variation of this flux ratio
with abundance).
The bottom panel shows
that there is very little correlation of the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ flux
ratio with oxygen abundance within our sample alone. However, when the
SINGS data are included, it is clear that
the low-metallicity BCDs show considerably higher neon ratios
than the more metal-rich SINGS galaxies \citep{dale09}.
The SINGS galaxies span an abundance range of
roughly a factor of 10, from 12$+$log(O/H)\,$\sim$8.4 to 9.4,
and our sample fills in the decade below, down to 12$+$log(O/H)\,$\sim$7.5.
Considering the two samples together,
the bottom panel shows an increase of the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ flux ratio by a
factor of $\sim$10 from
12$+$log(O/H)\,$\sim$9.4 down to $\sim$8.2. Then the curve
flattens out
into a ``plateau'' for lower abundances, with \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ $\sim$5.
Again, the scatter of the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ flux
ratio at a given oxygen abundance is large, indicating that
metallicity cannot be the main parameter controlling the hardness of the
ISRF.
Similar behavior is shown also in previous metal-poor samples
\citep{wu06,hao09}.
Our sample was defined on the basis of the optical \rm He{\sc ii}/H{$\beta$}\
ratio which gives an independent estimate of the ISRF hardness.
However, within our sample, there is no correlation
between the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ ratios and
the \rm He{\sc ii}/H{$\beta$}\ ratio as shown in Fig. \ref{fig:nes_heii}.
A possible reason for this lack of correlation
is that the IR and optical ratios do not probe the same excitation energy.
The ionization potentials for \rm [Ne\,{\sc iii}]\ and \rm [Ne\,{\sc ii}]\ are respectively
41.0\,eV and 21.6\,eV, while they are respectively 34.8\,eV and 23.3\,eV
for \rm [S\,{\sc iv}]\ and \rm [S\,{\sc iii}].
These ionization potentials are well below that of He$^+$ which is 54.4\,eV.
Thus, the ISRF UV energies probed
by the neon and sulfur lines are too soft compared to those
probed by the \rm He{\sc ii}\ ratio.
We have therefore examined the relation between
the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ and \rm He{\sc ii}/H{$\beta$}\ ratios;
the ionization potential of \rm [O\,{\sc iv}]\ is
54.9\,eV, similar to that of He$^+$ (54.4\,eV), while that of
\rm [Si\,{\sc ii}]\ is only 8.2\,eV.
Since both silicon and oxygen are products of
core-collapse supernovae (SNe),
the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio should not be subject to
abundance anomalies.
The left panel of Fig. \ref{fig:oiv_heii_oh} shows a weak
correlation between the two ratios for the BCDs in our sample:
a high \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio is associated with a higher \rm He{\sc ii}/H{$\beta$},
at a confidence level of $\sim$93\% (one-tailed).
Higher \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ is associated with lower nebular O/H
abundance (right panel of Fig. \ref{fig:oiv_heii_oh}), at a
$\sim$98\% confidence level (one-tailed).
This last correlation is similar to the one
between \rm He{\sc ii}/H{$\beta$}\ and O/H \citep{thuanizotov05}.
Evidently, lower metallicity galaxies tend
to have generally harder ionizing radiation.
Although the BCDs by themselves show little correlation between
the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio and the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ ratios,
when SINGS galaxies are included, there is a definite correlation.
Fig. \ref{fig:nes_oiv} illustrates this, and suggests
that in star-forming galaxies the presence of hard radiation
as probed by the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio implies also the presence of
less hard radiation as probed by the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ ratios.
\subsection{The Intensity of the Interstellar Radiation Field \label{sec:intensity}}
We now explore diagnostics for the ISRF intensity, as
distinct from the ISRF hardness, since it does not depend on the
spectral shape of the ionizing radiation.
The ISRF intensity is related to the ionization parameter
which measures the number of ionizing photons per hydrogen atom.
MIPS observations are relevant here, because the ratio of
71 to 160\,\micron\ fluxes is sensitive to the temperature of
the large (``classical'') grains, and can thus act as an indicator of the
ISRF intensity \citep{draine07}.
Essentially, we will use the MIPS luminosity ratio (``color temperature'')
$\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$\ as a proxy for the ionization parameter.
The upper panel of Fig. \ref{fig:p24oh_r71} shows the relation between this
``color temperature'' $\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$\
and the quantity $\langle\nu{\rm L}_\nu\rangle$ $_{24}$/($\langle\nu{\rm L}_\nu\rangle$ $_{71}+$ $\langle\nu{\rm L}_\nu\rangle$ $_{160}$);
the latter represents the relative contribution
of the 24\,\micron\ emission to the long-wavelength component of $L_{\rm TIR}$,
the total IR luminosity as defined by \citet[][see $\S$\ref{sec:phot}]{draine07}.
The lower panel of Fig. \ref{fig:p24oh_r71} shows the variation
of the same 71 to 160\,\micron\ luminosity ratio
with oxygen abundance.
SINGS galaxies are also shown in Fig. \ref{fig:p24oh_r71}
using data taken from \citet{dale07}.
The correlations in both panels are highly significant even within the BCD sample
alone:
the correlation in the top panel is at a $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$99\% confidence level
(one-tailed), while the bottom one is at a $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$99.9\% confidence
level (one-tailed).
This indicates that the 24\micron\ fraction of $\langle\nu{\rm L}_\nu (71)\rangle + \langle\nu{\rm L}_\nu (160)\rangle$\ increases
with increasing large-grain
temperature (and decreasing oxygen abundance).
Hence, the relative strength of the 24\,\micron\ luminosity is a
good tracer of the intensity of the ISRF.
As dust gets warmer, more large-grain
emission is shifted toward shorter wavelengths, resulting in
increased 24\micron\ flux \citep[e.g.,][]{dale05}.
Apparently, at low metallicities, 12$+$log(O/H) $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$7.8,
large grains can be heated to extreme
$\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$\ ratios, reaching values of 10--30.
The effect could be even more extreme since the lowest
metallicity BCDs in our sample are not plotted; the MIPS 160\,\micron\
emission for the four galaxies with 12$+$log(O/H)$\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$7.6 is not detected.
Compared to the SINGS sample \citep{dale05,dale07}, the ISRF intensity
as measured by $\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$\ is $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$ 5 times greater on average in the BCDs than
in even the most intense SINGS galaxy\footnote{To do the comparison,
we have converted the flux ratios used by
\citet{dale05} to the $\langle\nu{\rm L}_\nu\rangle$\ ratios used here.}.
Using the 71/160\micron\ luminosity ratio ($\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$) as a tracer of the ISRF
intensity, we next investigate whether any of the FS line ratios,
indicators of the ISRF hardness as shown above, are correlated
with it.
Figure \ref{fig:neoiv_r71} shows the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ line
ratios as a function of the 71/160\micron\ ratio;
SINGS galaxies are also plotted combining data from \citet{dale07} and \citet{dale09}.
The \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ flux ratio shows no correlation, although there is a
weak one in the BCD sample alone (at the $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$97\% confidence level) for \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}];
the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio is larger for warmer large grains.
In the SINGS sample, most of the highest \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\
flux ratios are associated with AGN, so any correlation could be masked by
AGN high-excitation emission-line ratios in galaxies with a relatively low-intensity ISRF.
In any case,
we conclude that the hardness of the ISRF, as measured by \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\
is weakly related to its intensity, as traced by the
71/160\micron\ luminosity ratio. In lower metallicity BCDs, the ISRF
tends to be both harder and more intense.
To summarize, in our low-metallicity BCD sample,
the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio is correlated with \rm He{\sc ii}/H{$\beta$}, 12$+$log(O/H),
and the large-grain dust temperature (71/160\micron\ ratio). This implies
that star-forming galaxies with a lower nebular metallicity
as compared to more metal-rich galaxies,
generally possess an ISRF which is both harder, as indicated
by higher \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ or \rm He{\sc ii}/H{$\beta$}\ flux ratios, and
more intense as indicated by a larger 71/160\micron\ luminosity ratio.
The relatively large scatter in some of these relations,
and the ``plateau-like'' behavior of others, implies
that these quantities do not depend on only one parameter, but that
many parameters are involved.
The ISRFs in the SINGS galaxies are of considerably lower intensity
than those generally found in BCDs.
This is probably because the star-forming regions are more extended,
less extreme, and have a lower ionization parameter $Q$.
Such factors are probably as important as ISRF hardness and metallicity.
\subsection{Electron densities\label{sec:densities}}
We have derived electron densities $n_{\rm e}$\ in the ionized gas from the ratio
of the temperature-insensitive \rm [S\,{\sc iii}]\ IR lines \citep{draine09}.
Fig. \ref{fig:densities} shows these densities as compared with those
derived using the optical
\rm [S\,{\sc ii}]\ $\lambda$6717/$\lambda$6731 ratio.
If the IR emission were to originate in regions suffering from higher
extinction than the optical regions,
we might expect the IR $n_{\rm e}$\ to be higher than those derived from
the optical lines.
Fig. \ref{fig:densities} shows that this is true for some BCDs,
but not all.
On the contrary, a few galaxies
show higher optical than IR densities.
The reason for this behavior is unclear;
beam dilution could play a role
since the IRS aperture is several times larger than the optical one and
thus includes lower density outer regions.
In the optical, high-ionization radiation responsible for ions such as \rm [Ne\,{\sc v}]\ is found
to be associated with extremely massive compact super star clusters
and high optical densities \citep{thuanizotov05}.
The idea that compact star-forming regions were associated with
higher electron densities $n_{\rm e}$\ was first proposed on theoretical
grounds by \citet{hirashita04}.
The size-density correlation in \rm H{\sc ii}\ regions
in which smaller regions are denser
is observed over six orders of magnitude
from Galactic objects to distant BCDs \citep{hunt09}.
If fast radiative shocks were responsible for the ionizing radiation
as proposed by \citet{thuanizotov05}, then we would expect such
processes to be more efficient at high densities.
However, in the current BCD sample,
only two (three) of 7 (9) galaxies with a 3$\sigma$ (2$\sigma$) \rm [O\,{\sc iv}]\ detection
has a $n_{\rm e}$\ above the low-density limit of $\sim$30\,cm$^{-3}$.
The electron densities as derived from the \rm [S\,{\sc iii}]\ IR lines
are apparently not correlated with the hardness
of the ISRF, as measured by \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}].
One reason for this non-correlation
may be that \rm [O\,{\sc iv}]\, with an ionization potential
of 54.9 eV is not probing sufficiently high ionization
energies (the \rm [Ne\,{\sc v}]\ line has an ionization potential of 97.1\,eV).
Another reason may be that the \rm [S\,{\sc iii}]\ IR lines arise from a region that
is less dense than the \rm [O\,{\sc iv}]\ region.
More work on a larger sample is needed to explore the connection between
density, geometry, and the ISRF.
\subsection{\rm [O\,{\sc iv}]\ and \rm [Fe\,{\sc ii}]\ in BCDs \label{sec:oxygen_iron}}
The two FS lines \rm [O\,{\sc iv}]\ and \rm [Fe\,{\sc ii}]\ lie at adjacent wavelengths
($\Delta\lambda\,=\,$0.098\micron) near 26.0\,\micron. They are, however,
sufficiently separated to be resolved by LH IRS spectra.
While \rm [O\,{\sc iv}]\ traces high-excitation gas, \rm [Fe\,{\sc ii}]\ has an ionization
potential of 7.9\,eV, more similar to \rm [Si\,{\sc ii}]\ (see $\S$\ref{sec:SiII}).
Both lines may be excited by photoionization and shocks
\citep{lutz98,ohalloran06,allen08}.
By fitting the two lines with Gaussian profiles, we are able to deblend
them in our spectra.
We detect \rm [O\,{\sc iv}]\ at 3$\sigma$ in 7/22 observed galaxies, but
\rm [Fe\,{\sc ii}]\ at 3$\sigma$ in only Mrk\,5 and Mrk\,996 (see Fig. \ref{fig:oiv}).
At low metallicity, \rm [O\,{\sc iv}]\ emission is almost 4 times as common as
\rm [Fe\,{\sc ii}]\ (32\% vs. 9\%).
This conclusion contrasts with the findings of \citet{ohalloran06,ohalloran08}.
Using {\it Spitzer}\ archival IRS data,
they have published 26.0\,\micron\ \rm [Fe\,{\sc ii}]\ fluxes for a sample of starbursts,
including the 22 BCDs reported here.
While the uncertainties in the spectra estimated by us (see $\S$\ref{sec:analysis})
are generally more conservative
than those reported by \citet{ohalloran08}, these
authors find \rm [Fe\,{\sc ii}]\ emission in all objects, but report no \rm [O\,{\sc iv}]\ emission.
The \rm [Fe\,{\sc ii}]\ emission fluxes they find are also larger than ours, suggesting
contamination from the \rm [O\,{\sc iv}]\ emission.
Figure \ref{fig:oiv} shows the identifications of the
two lines; it is evident that the \rm [Fe\,{\sc ii}]\ flux can be securely measured
only when \rm [O\,{\sc iv}]\ is taken into account.
No spectra are shown by \citet{ohalloran08}, so we cannot directly
compare their line identifications with ours to resolve the source of
discrepancy.
\subsection{\rm [Si\,{\sc ii}]\ Emission \label{sec:SiII}}
The \rm [Si\,{\sc ii}]\ line is detected in more than half the BCDs in our sample.
At an ionization potential of only 8.2\,eV, \rm [Si\,{\sc ii}]\ traces
the neutral gas, rather than the more highly excited ionized gas traced
by the neon, sulfur, and oxygen lines.
Figure \ref{fig:tir} shows the \rm [Si\,{\sc ii}]\ emission normalized to the TIR luminosity.
The mean \rm [Si\,{\sc ii}]/TIR ratio
is more than a factor of two lower than that of the SINGS galaxies,
$8\times10^{-4}$ versus $2\times10^{-3}$ (compare the dotted and dashed
horizontal lines in Fig. \ref{fig:tir}).
This slight difference is probably due to the lower metal abundance of the BCDs,
and the consequently less efficient cooling of the gas through FS lines.
It could also be due to a more intense ISRF, and thus a lower Si$^+$ abundance.
In any case,
the ratio of \rm [Si\,{\sc ii}]\ emission to $L_{\rm TIR}$\ is relatively constant with luminosity
(excluding UM\,311 as discussed previously).
These considerations suggest that the \rm [Si\,{\sc ii}]\ line is an efficient coolant,
comparable to the [C{\sc ii}] line \citep[e.g.,][]{malhotra97},
and could be used as a proxy for $L_{\rm TIR}$\ to within a factor of two, even at low
metallicities.
\section{Aromatic Features \label{sec:aromatics}}
The wealth of spectral data obtained by {\it ISO}\
combined with much theoretical work and laboratory measurements
has shown convincingly that aromatic features
are associated with stochastic or
single-photon emission from large molecules called
polycyclic aromatic hydrocarbons or PAHs
\citep{joblin95,verstraete01,hony01,peeters02,vandiedenhoven04,peeters04a,bauschlicher08,bauschlicher09}.
PAH emission dominates the mid-infrared spectra of star-forming
galaxies \citep{brandl06,smith07}, and
contributes importantly to their IR energy budget.
Through the photoelectric effect, PAHs heat
the gas in the ISM \citep[e.g.,][]{hollenbach97}, thereby
reinforcing the coupling of the gas and dust components.
Because they are so dominant in the MIR spectral range,
PAHs have been used to identify galaxies with intense star
formation at redshifts of $\sim2-3$, and to estimate their
bolometric luminosities \citep[e.g.,][]{houck07,weedman08,dey08}.
In most star-forming galaxies,
PAH emission shows a surprisingly narrow range of properties,
well characterized by a ``standard set'' of
features with fixed wavelengths and FWHMs \citep{smith07}.
However, first {\it ISO}\ and later {\it Spitzer}\ have shown that metal-poor
star-forming galaxies are deficient in PAH emission
\citep{thuan99a,engelbracht05,madden06,wu06,engelbracht08}.
Although it has generally been concluded that the reason for
this deficiency is low metallicity, it is not yet clear whether
it is the metallicity directly, or rather the
indirect effects of a low metallicity environment.
In the first case, there would be a consequent
lack of raw material (carbon and nitrogen) with which to form PAHs;
in the second,
the increased ISRF hardness or intensity would lead to their destruction.
Here, we examine in detail the PAH emission in our sample to
examine this question, and
show for the first time that some PAHs can survive even in
extremely metal-poor environments.
\subsection{PAH Properties at Low Metallicity\label{sec:pahtrends}}
The 7.7\,\micron\ blend is the most common PAH feature in our sample,
being detected in 15 of 22 objects.
The 7.7\,\micron\ feature is also the strongest one, comprising
on average 49\% of the total PAH power.
The remaining PAHs are significantly weaker, and less frequently detected:
13 BCDs show the 11.3\,\micron\ blend, 9 the 8.6\,\micron\ feature,
and 7 have a detection of the 6.2\,\micron\ PAH.
No 17\,\micron\ PAH features were detected, and only
Mrk\,1329 showed a significant detection at 16.4\,\micron.
The 6.2, 7.7, 8.6, 11.3, and 12.6\,\micron\ features
by themselves constitute $\sim$72\% of the total PAH power in BCDs
(see Table \ref{tab:dfpahfit});
the SINGS sample has about $\sim$85\% of the total PAH power
in these features \citep{smith07}.
Similarly to the SINGS galaxies,
other weaker features (not listed in Table \ref{tab:dfpahfit})
which contribute to the PAH luminosity
include the
5.7, 6.6, 6.7, 8.3, 10.6, and 10.8\,\micron\ features.
In our sample, these features comprise roughly 28\% of the PAH luminosity,
but in SINGS galaxies, only 15\%.
Part of the reason for this difference
may be the difficulty in detecting long-wavelength
PAHs in BCDs; only 1 BCD has the 16.4\,\micron\ feature detected,
while in SINGS galaxies the 16-17\,\micron\ blend makes up
$\sim$6\% of the total PAH power.
The fractional power of the four strongest features
relative to the total PAH luminosity is illustrated in
Figure \ref{fig:pahfrac}.
The horizontal lines in each panel correspond to the SINGS medians
(dashed) and the BCD means (dotted).
The BCD means are calculated taking into account all galaxies
with 7.7\,\micron\ detections; thus they are a sort of weighted
average which considers frequency of detection together with intensity.
This is the reason that the mean for the 6.2\,\micron\ PAH lies
below most of the BCD data points: that feature was detected only in
7 galaxies, while the other features in Fig. \ref{fig:pahfrac} were
detected with a frequency more similar to the 7.7\,\micron\ PAH.
The relative PAH fractions for the BCD and SINGS 6.2 and 7.7\,\micron\ features
are virtually indistinguishable: 0.10 vs. 0.11 for 6.2\,\micron,
and 0.49 vs. 0.42 for 7.7\,\micron\ \citep{smith07}.
However, for the longer wavelength PAHs at 8.6 and 11.3\,\micron,
the BCD fractions are $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$ 40\% larger:
0.10 vs. 0.07 (8.6) and 0.17 vs. 0.12 (11.3).
Although there is considerable scatter, the BCD mean PAH relative fractions for these
features tend to exceed even the 10 to 90th percentile spreads of the
SINGS galaxies.
Figure \ref{fig:pahfwhm} shows the mean Drude profile widths
and central wavelengths of six aromatic features detected
at $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ in our sample; the numbers of BCDs with each
feature are given in parentheses.
The analogous quantities for the SINGS galaxies are also plotted.
In no case, were the PAHFIT profiles fixed to the SINGS galaxy
parameters able to fit the BCD data; every BCD was better fitted by allowing
the Drude profile widths (FWHMs) and central wavelengths to vary.
In some cases, the differences between the $\chi_\nu^2$
with the SINGS ``standard'' profile parameters
and the best-fit parameters were small, $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$15\%;
but in others, the improvement of $\chi_\nu^2$ was almost a factor of 2.
The only {\it systematic} difference between the two samples
is the narrower profile width for
the 8.6\,\micron\ feature; only the very broadest BCD profiles are
as broad as those in more metal-rich systems.
The larger relative intensities of the 8.6 and 11.3\,\micron\
feature in BCDs and the narrow profile width
of the 8.6\,\micron\ PAH could be due to a different size distribution
of PAH populations at low metal abundances.
\citet{bauschlicher08} find that the 8.6\,\micron\ band arises from
large PAHs, with $N_{\rm min} \lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$ 100 carbon atoms.
Hence, they suggest that the relative intensity of
the 8.6\,\micron\ band can be taken as an indicator of
the relative amounts of large and small PAHs in a given population.
The large 8.6\,\micron\ intensity relative to total PAH power in BCDs
could be a signature of fewer small PAHs (or more larger ones) at low metallicity.
The relative lack of small PAHs could also explain the
low detection rate of the 6.2\,\micron\ feature,
since most of its intensity comes from PAHs with less than 100 C atoms
\citep{schutte93,hudgins05}.
The size of a PAH molecule grows with the number of carbon
atoms in it \citep[e.g.,][]{draine07}.
For a given maximum number of carbon atoms $N_{\rm max}$, a {\it larger
minimum number} $N_{\rm min}$ theoretically gives narrower profiles \citep{joblin95,verstraete01}.
Although fitting intrinsically asymmetric PAH profiles by symmetric Drude
profiles in PAHFIT is a simplification,
the narrower width of the 8.6\,\micron\ feature
could be a signature of relatively larger PAHs at low metallicity.
The difference disappears at 7.6\,\micron\ and is less pronounced
at longer wavelengths than 8.6\,\micron\ because
both small and large PAH sizes contribute to
these bands\footnote{The 7.8\,\micron\ band may be dominated by larger PAHs
only \citep{bauschlicher09}; but
this band is present in only two of the BCDs in our sample, Haro\,3 and
II\,Zw\,70.}
\citep{schutte93,peeters02,vandiedenhoven04,bauschlicher09}.
Another indication that low-metallicity BCDs may be lacking the
smallest size PAHs comes from a correlation analysis.
In a detailed study of Galactic \rm H{\sc ii}\ regions, young stellar objects,
reflection and planetary nebulae (RNe, PNe), and evolved stars, \citet{hony01}
found only a few correlations among PAH features.
One of these is between 6.2 and the 12.7\,\micron\
emission, which arises from the CC stretch mode in ionized PAHs,
in contrast to the 11.3\,\micron\ band which is attributed to
the neutral CH out-of-plane bending mode \citep{hony01,peeters02}.
The comparison between the 6.2/11.3 and 12.7/11.3 flux ratios
is shown for our sample of BCDs in Fig. \ref{fig:pah62127};
we show galaxies with $\simgt3\sigma$ detections in at least two features.
With the exception of Mrk\,1315, galaxies with 6.2\,\micron\ PAH detections
and 12$+$log(O/H)$\geq$8.1 (filled circles)
are similar to the \rm H{\sc ii}\ regions studied by \citet{hony01}.
On the other hand, BCDs with lower O/H
occupy a region of the plot with small 6.2/11.3 ratios.
Although the statistics are poor,
small 6.2/11.3 are found for most of the lowest-metallicity
BCDs. They lie in the same region as the RNe and PNe studied by \citet{hony01}.
Because the 11.3\,\micron\ feature arises from non-adjacent or ``solo''
CH groups, dominant 11.3\,\micron\ emission implies large
PAH complexes, with $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$100-200 carbon atoms and long straight
edges \citep{hony01,bauschlicher08,bauschlicher09}.
Moreover, fewer small PAHs reduce the intensity
of the 6.2\,\micron\ emission,
as the bulk of this feature comes from
PAHs containing less than 100 C atoms \citep{schutte93,hudgins05}.
Therefore, small 6.2/11.3\,\micron\ band ratios at low metallicites
could imply an overall deficit of small PAHs.
A third indication that there is a deficit of small
PAHs in a metal-poor ISM is provided by the intensity of the
8.6\,\micron\ feature relative to the 7.7\,\micron\ blend.
As mentioned above,
\citet{bauschlicher08} find that 8.6\,\micron\ band arises from
large PAHs, with $N_{\rm min} \lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$ 100 carbon atoms.
On the other hand, the 7.7\,\micron\ feature
contains emission from both small and large PAH sizes
\citep{schutte93,peeters02,vandiedenhoven04}.
For the 9 BCDs with the 7.7 and 8.6\,\micron\
PAH features detected at $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$,
the mean 8.6/7.7 flux ratio is 0.48, and the median is 0.33.
This ratio is at least double that in the more metal-rich SINGS galaxies,
with 8.6/7.7 $\sim$0.18 in the mean, and a range of 0.11$-$0.21
\citep{smith07}; 7 of 9 BCDs are outside the upper 90th percentile
limit of 0.21.
Again, although the number statistics are small,
the large 8.6/7.7\,\micron\ flux ratios of
the PAHs in these low-metallicity BCDs
appear to be dominated by the largest complexes.
In summary, there are three different lines of evidence
which tentatively suggest that the PAH populations in low-metallicity BCDs
are deficient in the smallest sizes ($N_{\rm min} \lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$50 C atoms).
The IRS spectra show flux ratios typical of
the largest PAHs modeled so far, with $N_{\rm min} \lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$100 C atoms.
Apparently, the ISM environment in the BCD \rm H{\sc ii}\ regions
is not propitious to the existence of the smallest PAHs and
has led to their destruction.
Perhaps only the PAH population with the largest complexes can
survive the extreme physical conditions in the low-metallicity \rm H{\sc ii}\
regions in the BCDs.
\subsection{PAHs and the Energy Budget\label{sec:budget}}
PAHs emit about 10\% of the total infrared power in typical
solar metallicity star-forming
galaxies (12$+$log(O/H)\ $\sim$ 8.7-9.0), with a maximum of $\sim$20\% \citep{smith07}.
In our low-metallicity BCD sample, with oxygen abundances 12$+$log(O/H)\ ranging from
7.4 to 8.3, the fraction of PAH emission ranges from $\sim$0.1\% to
$\sim$1.6\%, with a median (and mean) of $\sim$0.5\%\footnote{Because of
crowding and source structure, the TIR luminosity in UM\,311 may be overestimated;
nevertheless even without UM\,311, the mean is only slightly larger, 0.57\% vs. 0.54\%.}.
Figure \ref{fig:pahtot_tir} shows how the sum of all $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$
PAH emission features normalized to the total IR luminosity $L_{\rm TIR}$,
$\Sigma$(PAH)/TIR, varies as a function of
$L_{\rm TIR}$\ for our BCD (filled circles) and SINGS galaxies
(non-AGN open squares, AGN open triangles) \citep{smith07}.
For the SINGS galaxies, $L_{\rm TIR}$\ has been calculated {\it within the spectral
extraction aperture}, while for the BCDs, $L_{\rm TIR}$\ is taken to be the
total luminosity as described in $\S$\ref{sec:phot}.
Because most of the
BCDs are point-like at 24\,\micron\ and are generally unresolved with MIPS,
this is usually comparable to $L_{\rm TIR}$\ within the spectral aperture (see Fig.
\ref{fig:oiv}).
Were the source sizes larger at 160\,\micron\ \citep[e.g., Haro3: ][]{hunt06},
$L_{\rm TIR}$\ could be overestimated, and the PAH fractions somewhat higher.
The infrared luminosities $L_{\rm TIR}$\ of the SINGS sample\footnote{The normalizing
factor, $L_{\rm TIR}$, is calculated by different groups with slightly different
algorithms; however, the general
consistency of the data suggests that our conclusions do not depend
on the exact method used to derive $L_{\rm TIR}$.}
vary by a range of more than three
orders of magnitude, while the BCD luminosities occupy the middle portion of
this range.
Figure \ref{fig:pahtot_tir} shows clearly the
segregation of the SINGS points, with high $\Sigma$(PAH)/TIR, and the BCD points
with low $\Sigma$(PAH)/TIR. The AGN, with intermediate $\Sigma$(PAH)/TIR, lie between
the metal-rich star-forming SINGS galaxies and BCDs.
Figure \ref{fig:pahtot_tir} also shows that there is no trend of the PAH
fraction $\Sigma$(PAH)/TIR with $L_{\rm TIR}$\ in either sample.
We might expect the fraction of IR luminosity in PAHs to vary with metallicity.
We explore this possibility in the left panel of Fig. \ref{fig:pahtot_ohne},
where $\Sigma$(PAH)/TIR is plotted against oxygen abundance.
Although
there is no correlation of PAH fraction with metallicity within the BCD sample,
the inclusion of SINGS galaxies does show a trend of
increasing PAH fraction with increasing metallicity.
We next investigate other physical factors which may control the PAH
fraction, such as the ISRF hardness.
The right panel of Fig. \ref{fig:pahtot_ohne} shows $\Sigma$(PAH)/TIR as a function
of the IR neon line ratio, \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ (only those BCDs with at least
a 3$\sigma$
detection in the neon lines are shown).
As discussed in $\S$\ref{sec:ionized}, the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ ratio is a
good tracer of ISRF hardness, up to an intermediate energy range.
There is a weak correlation between $\Sigma$(PAH)/TIR and \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}],
significant at the 95\% (one-tailed) confidence level.
Thus,
over the range in UV energies $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$41\,eV, the fraction
of $L_{\rm TIR}$\ emerging as PAH emission may depend weakly on the hardness of the ISRF.
This suggests that the PAH emission fraction in BCDs is not directly
controlled by metallicity, but
by the hardness of the radiation field. The latter could also be responsible
for the destruction of the smallest PAH particles, as discussed
in the previous section.
We have checked for a correlation of $\Sigma$(PAH)/TIR with ISRF intensity
as measured by $\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$, but found none.
To conclude our discussion of PAHs and the IR energy budget, we examine
the diagnostic proposed by \citet{houck07}: the quantity
$\nu L_\nu$(7.7\micron),
where $L_\nu$(7.7\,\micron) is determined from the flux density at the
peak of the 7.7$+$7.8\,\micron\ blend, with no correction for the underlying
continuum.
Because it arises from a range of sizes,
the 7.7\,\micron\ band is arguably the most reliable PAH diagnostic,
as it avoids preferential selection of either small or large PAH populations
\citep{schutte93,peeters02}.
Figure \ref{fig:pah77} compares $\nu L_\nu$(7.7\micron)/$L_{\rm TIR}$\
for our BCDs with that for a sample of local starbursts \citep{brandl06,houck07}.
Considering only those BCDs with a 160\,\micron\ detection,
the BCDs have a mean $\nu L_\nu$(7.7\micron)/$L_{\rm TIR}$\
roughly four times lower than that of the starbursts.
The means of the two samples are shown by horizontal dotted lines.
Part of this difference may be
due to a luminosity effect since the BCDs are more than 100 times less
luminous in the mean than the starburst galaxies
(see for example the SINGS galaxies in Fig. \ref{fig:pahtot_tir}).
It may also result from variations of physical conditions
in the ISM, since the BCDs with the highest $\nu
L_\nu$(7.7\micron)/$L_{\rm TIR}$\ clearly overlap with
the range observed in local starbursts that are however much
more luminous in $L_{\rm TIR}$.
\subsection{PAH Diagnostics for AGN/Starbursts Revisited\label{sec:pahdiagnostics}}
Several diagnostics based on IR spectra have been proposed to separate AGN from
star-forming regions in galactic nuclei.
Previous work has also focused on the comparison of the flux ratios of PAH
features and emission line ratios from FS transitions.
In this section, we examine our BCD sample in the context of such diagnostics
and assess their usefulness in the low metallicity regime.
\subsubsection{7.7/11.3\,\micron\ Band Ratios}
We saw in the previous section that the total PAH fraction $\Sigma$(PAH)/$L_{\rm TIR}$\
is correlated with the IR neon ratio \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ (Fig. \ref{fig:pahtot_ohne}).
Here, following \citet{smith07},
we examine the variations of the 7.7/11.3\,\micron\ band ratios with
\rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ for both the BCD and SINGS samples as shown in
Fig. \ref{fig:pah_ne}.
Among the SINGS galaxies, there is a weak correlation, mostly for AGN,
in the sense that a smaller 7.7/11.3 ratio is associated with
a larger \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}] ratio \citep[e.g.,][]{smith07}.
The BCDs extend the correlation to considerably
higher neon ratios (by one order of magnitude), that is to harder
ISRFs.
The plot shows that the radiation fields in the low metallicity
BCDs can be as extreme as those in the SINGS AGN.
Such hard, intense ISRFs could suppress the ionized 7.7\,\micron\ feature, relative to the
neutral 11.3\,\micron\ feature.
This may be due to the dominant contribution of
large PAHs to the 11.3\,\micron\ band,
as compared to the mixed size distribution
thought to be responsible for the 7.7\,\micron\ band;
the small PAH molecules
contributing to the 7.7\,\micron\ emission could be destroyed while
the large PAH molecules responsible for the 11.3\,\micron\ emission are not
(see $\S$\ref{sec:pahtrends}).
\subsubsection{AGN/Starburst Diagnostics}
Because a hard, intense radiation
field changes the PAH properties, a deficit of PAH emission,
combined with FS line indicators of excitation, has often been
used as a AGN diagnostic in solar metallicity star-forming galaxies.
However, we have seen above that the radiation fields in low metallicity BCDs
can be just as intense as
those in AGN (see Fig. \ref{fig:pah_ne}), possibly destroying the
smallest PAHs. Thus a PAH emission deficit
may not necessarily be a reliable diagnostic for AGN.
We compare our BCD sample with AGN taken from \citet{genzel98},
using the diagnostics proposed by the latter authors and \citet{laurent00}.
Figure \ref{fig:genzel_diag} shows the
\rm [O\,{\sc iv}]/\rm [Ne\,{\sc ii}]\ ratio \citep[or 1.7$\times$33.5\micron\ \rm [S\,{\sc iii}], see][]{genzel98}
as a function of the ``strength'' of the 7.7\,\micron\ PAH band.
\rm [O\,{\sc iv}]/\rm [Ne\,{\sc ii}]\ is another measure of ISRF hardness (see $\S$\ref{sec:ionized}
where we used \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]).
The ``strength'' of the 7.7\,\micron\ PAH feature was
determined by estimating the continuum level between 5 and 14\,\micron\,
and dividing the PAH flux by this level;
this method is the same as that used by \citet{genzel98}, so we preferred it
to the simpler equivalent width measurement given by PAHFIT.
For metal-rich systems, plotting these two ratios against one another
results in a clear distinction between star-forming galaxies
(stars and squares)
and AGN (triangles), the latter having a higher \rm [O\,{\sc iv}]/\rm [Ne\,{\sc ii}]\ ratio and
a weaker 7.7\,\micron\ PAH feature than the former.
However, the situation for
the BCDs in our sample is different. They do not fall
in the star-forming but rather closer to the AGN region.
The starburst with the highest \rm [O\,{\sc iv}]\ fraction in the \citet{genzel98} sample
is NGC\,5253 (open star symbol), a BCD-like galaxy
with an oxygen abundance of 12$+$log(O/H)$\sim$8.2 \citep{kobulnicky99}.
Two BCDs in our sample would
have been classified as pure AGN on the basis of this diagnostic diagram.
Thus, the strength of \rm [O\,{\sc iv}]\ combined with the 7.7\,\micron\ PAH deficit
in metal-poor starbursts makes them difficult to distinguish from AGN.
The second AGN diagnostic diagram often used is illustrated in Fig. \ref{fig:laurent_diag}.
This diagnostic, proposed by \citet{laurent00},
exploits the steeply rising MIR continuum in starbursts,
compared to the flatter continua in AGN and quiescent star-forming regions
(or PDRs).
Most star-forming galaxies have both PDR and \rm H{\sc ii}\ region components in their
spectra; the PDR component tends to have a flat MIR continuum and be
dominated by PAH bands, while the \rm H{\sc ii}\ regions have
a deficiency of PAHs and a steeply rising MIR continuum from warm dust.
To analyze our BCD sample, we have not used the original
diagnostic diagram as proposed by \citet{laurent00}, but a
slightly modified version of it as formulated by \citet{peeters04a}.
It differs from the original diagram in the continuum wavelength range that
enters in the denominator of the labels on both axes.
Fig. \ref{fig:laurent_diag} plots the 14$-$15\,\micron\ continuum against
the 6.2\,\micron\ PAH feature, both normalized by the 5.3$-$5.8\,\micron\
continuum.
In this diagram, because of their steep MIR continua,
the BCDs (filled circles) are clearly distinguished
from the AGN (triangles), ULIRGs (squares), and metal-rich
star-forming galaxies (stars), lying above them.
A few BCDs are consistent with metal-rich Galactic \rm H{\sc ii}\ regions
($\times$),
but others are more extreme, both in their steeper continuum slope and in their
enhanced PAH strength (or faint 5\,\micron\ continuum).
The four BCDs with the steepest continua and most enhanced PAH strength
are also the most metal-rich galaxies in our sample
(Haro\,3, Mrk\,450, Mrk\,1315, and UM\,311).
We conclude that the AGN/starburst diagnostic diagram proposed
by \citet{laurent00} based on the continuum MIR slopes of these
types of objects effectively separates AGN from starbursts, even
at low metallicity.
\section{Molecules \label{sec:molecules}}
A long-standing puzzle of metal-poor star formation has been
the apparent lack of cool molecular gas from which to form stars.
Numerous attempts
to detect carbon monoxide in metal-poor BCDs,
with nebular oxygen abundances
12$+$log(O/H) $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$8.2, have failed.
Despite vigorous on-going star formation,
there seems to be very little CO in BCDs
\citep{sage92,taylor98,gondhalekar98,barone00,leroy05}.
Intriguingly, 12$+$log(O/H) $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$8.2 appears to be
the same metallicity ``threshold'' below
which PAH emission is thought to be suppressed
\citep{engelbracht05,wu06,madden06,engelbracht08}.
Here we show that warm molecular gas in the form of \rm H$_2$\
does exist at very low metallicity.
Our new IRS spectra reveal a variety of \rm H$_2$\ rotational lines,
and more than a third of the objects in our sample (8 BCDs) have $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$
detections in one or more of the four lowest-order transitions of \rm H$_2$.
\subsection{\rm H$_2$\ Emission \label{sec:h2}}
Rotational transitions of \rm H$_2$\ are important diagnostics of the
warm neutral phase of the ISM, at temperatures between $\sim$100 and 1000\,K.
These transitions, observable from space between 5 and 28\,\micron,
are a main coolant of the warm molecular component.
Massive stars are the main source of excitation \citep{hollenbach97},
as \rm H$_2$\ molecules can be pumped by non-ionizing far-ultraviolet (FUV) photons
into excited states.
They then return to the ground state through fluorescence or collisional
deexcitation.
The pure fluorescence mechanism is likely to be responsible for
the near-infrared roto-vibrational transitions
\citep[e.g.,][]{puxley88},
because of the higher critical densities needed for collisional deexcitation.
Critical densities of the lower-order rotational transitions are sufficiently low
($\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}\,10^3$\,cm$^{-3}$\ for S(0) to S(3))
that the lower rotational levels should be thermalized.
For the pure rotational \rm H$_2$\ transitions studied here, $\Delta v$ is 0;
we will refer to the $\Delta J$ series of rotational transitions as S(0), S(1), ..., S(7)
\citep[for a complete tabulation, see][]{roussel07}.
We first examine the \rm H$_2$\ emission in terms of its contribution to the
energy budget in the neutral ISM.
Following \citet{roussel07}, we use the sum of the S(0) to S(2) transitions to
assess \rm H$_2$\ cooling.
The bulk of \rm H$_2$\ cooling occurs through these lines.
Figure \ref{fig:h2tir} shows the sum of the S(0) to S(2)
\rm H$_2$\ emission, normalized to the total IR luminosity, $L_{\rm TIR}$.
As before, both the SINGS and BCD samples are plotted;
the SINGS galaxies are confined to the hatched region.
The $\Sigma[{\rm H}_2(0-2)]$/TIR ratios range from $1\times10^{-4}$ to $2\times10^{-3}$,
with a mean of $\sim 5\times10^{-4}$ for the BCD sample
\footnote{We have excluded UM\,311 because of the probable
overestimate of its IR flux.}.
The $\Sigma[{\rm H}_2(0-2)]$/TIR for the SINGS galaxies range from
$2.5\times10^{-4}$ to $7.5\times10^{-4}$ \citep{roussel07}.
Relative to $L_{\rm TIR}$,
the warm molecular content of the BCDs is comparable to that of
the SINGS galaxies.
Because warm \rm H$_2$\ emission originates in PDRs, the UV-illuminated
parts of molecular clouds, \rm H$_2$\ and PAH emission should
be related, since both species are excited by UV photons.
Although our sample contains too few points to explore such
a correlation,
the mean ratio of $\Sigma[{\rm H}_2(0-2)]$\ to $\Sigma$(PAH)\ of $\sim$6\% is roughly ten times
larger than that for the SINGS galaxies \citep{roussel07}.
Because the \rm H$_2$\ emission of BCDs, relative to TIR, is similar to the SINGS sample,
this difference arises mainly from the deficit in PAH emission at low metallicites.
Models show that the intensity of rotational transitions of \rm H$_2$\ peaks inside
molecular clouds at
extinctions A$_{\rm V}$ $\sim$2 \citep{hollenbach97}, while PAHs are expected to
be excited predominantly on the cloud surface layers.
Deeper into the cloud, densities are higher and temperatures lower,
and PAHs would be coagulated onto grain mantles \citep{boulanger90}.
The similarity of \rm H$_2$/$L_{\rm TIR}$\ over a wide range of metallicity
implies that \rm H$_2$\ might be somewhat self-shielded in BCDs.
Although PAHs can survive the impact
of FUV photons with energies between $\sim$11 and 13.6\,eV,
\rm H$_2$\ would be dissociated without self-shielding.
It is also difficult to understand how, without self-shielding, the \rm H$_2$\
could survive the intense UV radiation field in low metallicity BCDs
that is destroying the smallest PAHs (see $\S$\ref{sec:pahtrends}).
Although more than a third of our sample
show definite low-order \rm H$_2$\ emission in their IRS spectra, all
attempts to detect \rm H$_2$\ in BCDs in UV {\it FUSE} absorption
spectra\footnote{Half a dozen have been observed, none of which are in the present sample.}
have failed \citep{thuan05}.
The absence of \rm H$_2$\ absorption lines in {\it FUSE} spectra
implies that the warm \rm H$_2$\ detected through IR emission must be quite
clumpy. The {\it FUSE} observations are not sensitive to such a clumpy
\rm H$_2$\ distribution because they can only probe
the transparent UV sight lines, not being able to penetrate dense clouds
with $N(H_2)\simgt10^{20}$\,cm$^{-2}$ \citep{hoopes04}.
\subsubsection{\rm H$_2$\ Excitation Temperatures \label{sec:excitation}}
Excitation diagrams for the \rm H$_2$\ lines help infer the
excitation temperatures for molecular hydrogen.
For four galaxies in our sample,
there were sufficient data to construct excitation diagrams.
These diagrams show, as a function of the upper level energy $kT_i$,
the natural logarithm of the column density of the species $N_i$ in
the upper level of each transition, divided by the statistical
weight, $g_i$.
Assuming local thermodynamic equilibrium (LTE),
the total column density N$_{\rm tot}$\ then follows from the Boltzmann equation:
\begin{equation}
\frac{N_i}{N_{\rm tot}}\,=\,\frac{g_i}{Z(T_{ex})} \exp \left( - \frac{T_i}{T_{ex}} \right)
\end{equation}
where
$T_{ex}$ is the excitation temperature for the $i^{th}$ level,
and $Z(T_{ex})$ is the partition function.
The statistical weight is given by $g_i\,=\,(2I+1)/(2J+1)$;
the spin number takes the values $I=0$ for
even $J$ (para) transitions, and $I=1$ for odd $J$ (ortho) transitions.
\citet{herbst96} gives a convenient expression for the partition function:
$Z(T_{ex})\,=\,0.0247\,T/[1-\exp(-6000\,{\rm K}/T)]$, valid for temperatures $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$40\,K.
The column density for a given transition $N_i$ is calculated from the
observed \rm H$_2$\ flux $F_i$ in that transition:
\begin{equation}
N_i\,=\, \frac{F_i 4\pi}{ h\nu_i A_i \Omega}
\end{equation}
where $A_i$ is the transition's spontaneous emission coefficient,
$\nu_i$ is the frequency of the transition,
and $\Omega$ is the beam solid angle.
Figure \ref{fig:h2excitation} shows the resulting excitation diagrams for the
four BCDs which have at least one detection at the 3$\sigma$ level or
greater in a low-order transition
(except for SBS\,1152$+$579 for which we have used 2.5$\sigma$
detections), and at least two other significant ($\simgt3\sigma$)
detections in higher-order lines.
Since the inferred column density for a given transition depends
on the beam size, we have experimented
using different beam sizes of the IRS modules for the various transitions.
The main result of this exercise is that the most appropriate beam size
is the largest LH beam, equal to 11\farcs1$\times$22\farcs3.
Because the scale of our spectra was ultimately set
by the LH background subtraction
(see $\S$\ref{sec:observations}), and it was never necessary to multiply
by a scaling factor larger than the ratios of the apertures,
this is a reasonable conclusion.
However, the sources observed at 24\,\micron\ with MIPS
are either
unresolved or only slightly extended, which would imply a slightly
different beam dilution for each galaxy at longer wavelengths;
applying such a correction would have been highly problematic.
Nevertheless, since we have used
the IRS point-source flux calibration, this may not be a serious problem,
and, because of the way our spectra have been scaled,
the excitation diagrams should not be affected
significantly by this approximation.
With a beam size that is too small,
column densities are overestimated and total masses are underestimated.
Hence, in Fig. \ref{fig:h2excitation}, we have taken the most conservative
approach for column densities, using the LH beam as discussed above.
At the median distance of 21\,Mpc for our sources,
this corresponds to a region of $\sim$1.6\,kpc in diameter.
The excitation temperature of the line-emitting gas is the reciprocal of the
slope of the excitation curve;
in the absence of fluorescence,
this temperature would be the gas kinetic temperature in LTE.
Assuming thermal emission\footnote{This assumption is almost
certainly invalid for the higher-order transitions, because that
would require densities $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 10^5$\,cm$^{-3}$.
In any case, the excitation diagrams remain a useful tool for
inferring physical conditions of the line-emitting gas.},
Fig. \ref{fig:h2excitation} implies
a range of temperatures for the molecular gas.
As a crude approximation to the probable continuous temperature
distribution of the gas (see the parabolic fit for CGCG\,$005-027$),
we have fitted the data points in each diagram by two linear segments
\citep[see also][]{rigopoulou02,higdon06,roussel07}.
The temperature inferred is a strong function of the set of transitions
from which it is derived:
lower-order transitions probe lower-temperature gas than
the higher-order ones.
The implied temperature for the warm gas (from the lower-order transitions) is
$\sim$245\,K (not considering Mrk\,996 at 98\,K,
the only galaxy with a significant S(0) detection).
The temperatures of the hotter molecular component (from the high-order transitions)
range from $\sim$820\,K to $\sim$1600\,K.
Considering the transitions from which they are derived,
these temperatures are consistent with those found for starbursts
\citep{rigopoulou02}, SINGS galaxies \citep{roussel07}, and
ULIRGs \citep{higdon06}.
We have not considered any departure from the standard
ortho-para ratio of 3 for LTE \cite[e.g.,][]{burton92,roussel07}.
Departure from thermalization of ortho and para levels would be
suggested by a non-monotonic increase of inferred excitation
temperatures with increasing rotational transition.
The temperatures derived from mixed-parity \rm H$_2$\ transitions for the
BCDs in our sample are
generally consistent, indicating no departure from
thermalization for most objects,
although Fig. \ref{fig:h2excitation} does show
some sign of a discrepancy for SBS\,1152$+$579.
A detailed discussion of this topic at low metallicity is
precluded by the paucity of the BCD data presented here.
\subsubsection{Column Densities and Gas Masses \label{sec:h2mass}}
Once we have an estimate of the gas temperature for each galaxy,
we can derive the
\rm H$_2$\ column density and total \rm H$_2$\ mass within the IRS beam.
The derived masses and column densities
depend on the size of the IRS beam used to relate
the observed flux to the column densities.
As stated above,
we adopted a conservative approach for the column densities, and used the
large LH beam, which may give masses that are slightly overestimated.
In our discussion of \rm H$_2$\ column densities and masses,
we will consider only those objects
in which ``secure'' temperatures have been derived, i.e. those that have
two relatively reliable detections (at the $2.5\sigma$ level or greater)
of lower-order transitions\footnote{Now including S(3) to consider
correctly parity in the derivation of temperature.}.
Four objects meet these criteria: CGCG\,$005-027$, HS\,0837$+$4717, Mrk\,996,
and SBS\,1152$+$579.
Mrk\,996 and CGCG\,$005-027$\ have
a total warm (T$\sim$100-120\,K) \rm H$_2$\ column density N$_{\rm H2}$\ $\sim
3\times10^{21}$cm$^{-2}$\ or 53\,\ensuremath{M_\odot}\,pc$^{-2}$.
At a higher temperature of $\sim$340\,K,
SBS\,1152$+$579 has a much lower molecular column density
with N$_{\rm H2}$\ $\sim 3\times 10^{18}$ (0.04\,\ensuremath{M_\odot}\,pc$^{-2}$).
With a warm/total \rm H$_2$\ fraction of $\sim$5\% \citep{rigopoulou02,roussel07},
the total molecular column density
in SBS\,1152$+$579 would be $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$1\,\ensuremath{M_\odot}\,pc$^{-2}$.
The column density for HS\,0837$+$4717 is more uncertain,
because it has only 2.5$\sigma$ detections; the inferred
N$_{\rm H2}$\ is high, $\sim 5\times 10^{21}$ (78\,\ensuremath{M_\odot}\,pc$^{-2}$),
at a temperature of $\sim$95\,K.
With the exception of SBS\,1152$+$579,
the inferred \rm H$_2$\ column densities derived for the BCDs
are rather high, compared to the values normally thought to hold
for metal-poor galaxies.
Converting to masses, the two most extreme cases
are Mrk\,996 with a total warm (98\,K) \rm H$_2$\ mass
in the IRS beam of 1.5$\times10^8$\,\ensuremath{M_\odot}\
and CGCG\,$005-027$\ (at 120\,K) of 1.3$\times10^8$\,\ensuremath{M_\odot}.
Compared to the SINGS galaxies, these are rather high warm \rm H$_2$\ masses.
There are only three SINGS galaxies with warm ($\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$100\,K)
\rm H$_2$\ masses which exceed $10^8$\,\ensuremath{M_\odot}, and only one (NGC\,7552)
with a warm \rm H$_2$\ mass greater than that of Mrk\,996.
The other BCDs in our sample, with at least one
low-order detection have warm (250-470\,K) \rm H$_2$\ masses ranging from
$4\times10^3$\,\ensuremath{M_\odot}\ (Mrk\,209) to $5\times10^7$\,\ensuremath{M_\odot}\ (HS\,0837+4717).
The median SINGS warm \rm H$_2$\ mass is $\sim3.4\times10^6$\,\ensuremath{M_\odot},
and ranges from $\sim10^3$ to $3\times10^8$\,\ensuremath{M_\odot}\ \citep{roussel07}.
Most of the BCDs have $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 10^3$ times more warm \rm H$_2$\ mass than the SINGS
low-metallicity galaxies with CO detections, NGC\,6822 and NGC\,2915
\citep{roussel07}.
As stated above, these mass estimates may be overestimates.
The true beam size is
probably smaller than that of LH. Moreover, molecular hydrogen is not
distributed uniformly within that beam but in dense clumps.
However, even if we reduced the mass estimates by a factor of five
(roughly the difference in beam area between LH and SH),
the BCDs' molecular content would still be quite high.
This is contrary to the conventional view that \rm H$_2$\ molecules
do not exist in a low-metallicity environment, because of the lack of
grains on which to form them.
Clearly, metal abundance is not the only factor guiding \rm H$_2$\
formation in low-metallicity BCDs.
Finally, we can compare the warm \rm H$_2$\ mass to that of the neutral atomic gas.
According to the star formation models by \citet{krumholz09},
the \rm H$_2$\ column density of Mrk\,966
would correspond to an ISM molecular fraction $\sim$50\%, given its metal abundance.
Mrk\,996 has an \rm H{\sc i}\ mass of 1.5$\times10^8$\,\ensuremath{M_\odot} \citep{thuan99b};
thus its \rm H$_2$\ warm mass is comparable to that of
its neutral atomic gas mass, in agreement with the theoretical
predictions of \citet{krumholz09}.
Surprisingly, our data suggest that
metallicity and \rm H$_2$\ content are not directly linked.
Of the four BCDs discussed above with significant low-order \rm H$_2$\ detections,
three are below the sample median 12$+$log(O/H)\ of 7.9.
Other factors must play a role in determining the \rm H$_2$\ fraction.
In star-forming regions where substantial amounts of dust are concentrated
in high-density regions, self-shielding can promote \rm H$_2$\ formation;
thus the region can cool more effectively and star-formation rate
is enhanced \citep{hirashita04}.
Thus \rm H$_2$\ content may be more correlated with compactness (size and density)
than with metal abundance.
\subsection{Water and the Hydroxyl Radical \label{sec:water}}
Our IRS spectra tentatively suggest the presence of {\rm H}$_2${\rm O}\ and OH
in six objects (see Table \ref{tab:molecules}).
Figure \ref{fig:h2o_oh} shows the $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ detections of
{\rm H}$_2${\rm O}\ in three BCDs and OH in four.
The {\rm H}$_2${\rm O}\ lines at 29.836\,\micron\ and 29.885\,\micron\ result
from rotational transitions of ${\rm H_2\,^{16}O}\ 7_{25}-6_{16}$ and
$5_{42}-4_{13}$, respectively.
The OH emission lines correspond to rotationally excited levels
$^{2}\Pi_{1/2}\rightarrow ^{2}\Pi_{3/2} 7/2-5/2$ (28.940\,\micron),
$^{2}\Pi_{3/2} 19/2-17/2$ (30.346\,\micron), and
$^{2}\Pi_{1/2} 17/2-15/2$ (30.657\,\micron).
Because some of the spectra also show spurious spectral features
near the tentative detections, not associated with
known emission lines, these molecule detections are uncertain.
Nevertheless, if real, the transitions of ``hot'' {\rm H}$_2${\rm O}\ and OH
would be, to the best of our knowledge,
the first such detections in extragalactic objects.
Molecular emission from {\rm H}$_2${\rm O}\ and OH is perhaps not unexpected,
given the physical conditions in some BCDs.
{\rm H}$_2${\rm O}\ emission from vapor is thought to arise from
slow nondissociative, or C-type shocks \citep[e.g.,][]{draine80}.
In such a situation, at a shock velocity of $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$10\,km\,s$^{-1}$,
temperatures would exceed 300\,K, and subsequent reactions could rapidly
convert all gas-phase oxygen not bound in CO into {\rm H}$_2${\rm O}\ \citep{elitzur78}.
If there is a connection between
the dissociative shocks responsible for {\rm H}$_2${\rm O}\ vapor emission
and the fast radiative shocks possibly responsible for \rm [O\,{\sc iv}]\ emission
(see $\S$ \ref{sec:oxygen_iron}), we would expect both types of emission.
Indeed, all three galaxies with {\rm H}$_2${\rm O}\ emission in our sample
also show $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ \rm [O\,{\sc iv}]\ detections.
Nevertheless, to be consistent with the shock scenario for {\rm H}$_2${\rm O}\
production, the \rm H$_2$\ density would need to be rather high,
$\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 10^5$cm$^{-3}$\ \citep{elitzur78}.
The electron densities in the ionized gas
inferred from the \rm [S\,{\sc iii}]\ line ratio ($\S$ \ref{sec:densities})
are considerably smaller than this, but the \rm [O\,{\sc iv}]\ and {\rm H}$_2${\rm O}\ emission
could arise from different regions that are much denser.
We have no way of verifying this condition with the present data.
With {\it Herschel}, it may be possible to further pursue observations of
{\rm H}$_2${\rm O}\ vapor at low metallicity, and better understand the necessary
physical conditions and formation scenarios for this molecule.
OH emission could be associated
with the photo-dissociation of {\rm H}$_2${\rm O}\ at photon energies $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$9\,eV,
similar to the situation in Galactic outflows \citep{tappe08}.
Interestingly, the four galaxies with OH detections
also have high-order \rm H$_2$\ detections,
implying the presence of hot dense gas, thought to be necessary
for the neutral reactions leading to OH and {\rm H}$_2${\rm O}\ formation
\citep[e.g.,][]{hollenbach79}.
\section{Summary and Conclusions\label{sec:summary}}
We have presented low- and high-resolution IRS spectra,
supplemented by IRAC and MIPS measurements,
of 22 BCD galaxies, obtained during
our Cycle 1 GO {\it Spitzer}\ program (PID 3139). The BCD sample was chosen
to span a wide range in oxygen abundance [12$+$log(O/H)\ between 7.4 and 8.3],
and in ISRF hardness as measured
by the intensity of the nebular \rm He{\sc ii}\ $\lambda$\,4686 emission line
relative to H{$\beta$}.
The IRS spectra provide a variety of fine-structure
lines, aromatic features, and molecular lines which
probe the physical conditions in a metal-poor ISM,
and enable a study of
the dust properties as function of metallicity, hardness and
intensity of the ISRF.
We have used the PAHFIT routine to fit simultaneously the
spectral features, the underlying continuum and
the extinction in the IRS spectra. Fluxes have also been derived by
fitting Gaussian profiles to all FS and molecular lines. To place the
results for our low-metallicity BCD sample in
perspective, we have compared its properties to those of the SINGS
galaxy sample \citep{kennicutt03}.
We have obtained the following results:
\begin{enumerate}[(1)]
\item
The \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ and \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ flux ratios are
good diagnostics for the softer UV radiation ($\leq$ 40 eV).
The \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ ratio depends on metallicity,
being 1--2 orders of magnitude greater in our metal-poor
BCD sample than in the more metal-rich SINGS galaxies, but
the ratio flattens out at 12$+$log(O/H)$\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$8.3.
The \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio is an effective measure of harder radiation,
and
there is a strong correlation of the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ ratio with
oxygen abundance, implying that lower metallicity galaxies have harder
ionizing radiation than more metal-rich ones.
\item
We detect at 3$\sigma$ \rm [O\,{\sc iv}]\ (ionization potential of 54.9\,eV)
in 7/22 observed galaxies, but
\rm [Fe\,{\sc ii}]\ (ionization potential of 7.9\,eV) in only 2.
At low metallicity, \rm [O\,{\sc iv}]\ emission is almost 4 times as common as
\rm [Fe\,{\sc ii}]\ (32\% vs. 9\%). This is another indication that the ISRF of
low-metallicity galaxies is very hard.
\item
Electron densities derived from the IR \rm [S\,{\sc iii}]\ line ratio do not
correlate with those inferred from the optical \rm [S\,{\sc ii}]\ ratio,
or with the presence of \rm [O\,{\sc iv}]\ emission.
We would expect a correlation of \rm [O\,{\sc iv}]\ intensity with electron density,
if fast radiative shocks were responsible for the ionizing radiation
that produces \rm [O\,{\sc iv}].
Perhaps the \rm [S\,{\sc iii}]\ IR lines arise in and probe a region that
is less dense than the \rm [O\,{\sc iv}]\ region.
\item
The ratio of
71 to 160\micron\ fluxes $\langle\nu{\rm L}_\nu (71)\rangle/\langle\nu{\rm L}_\nu (160)\rangle$\ is sensitive to the temperature of
the large (``classical'') grains, and is thus a good indicator of the
ISRF intensity. The latter is $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$5 times greater in the BCDs than
in the SINGS galaxy with the most intense ISRF, implying
that metal-poor star-forming galaxies have not only a harder ISRF,
but also a more intense one than metal-rich galaxies.
\item
Two-thirds of the BCDs show PAH features.
The flux ratios of the different bands
are typical of the largest PAHs modeled so far,
with $N_{\rm min} \lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$100 C atoms.
We interpret these trends as an
indication that low-metallicity BCDs contain
relatively larger PAHs than more metal-rich environments.
Apparently, only these large PAHs are able to survive the hard and
intense ISRF in a low-metallicity ISM.
\item
The fraction $\Sigma$(PAH)/TIR of PAH emission $\Sigma$(PAH)\ normalized to the total
IR luminosity TIR is considerably smaller in
low-metallicity BCDs ($\sim$0.5\%) than in the more metal-rich SINGS
galaxies ($\sim$10\%).
At low metallicity, just as in metal-rich galaxies,
the PAH fraction of $L_{\rm TIR}$\ is constant.
\item
There is a good correlation between $\Sigma$(PAH)/TIR and the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]
flux ratio, but not with metallicity or with ISRF intensity.
Evidently, over the range in UV energies $\lower.5ex\hbox{$\;\buildrel<\over\sim\;$}$41\,eV, the fraction
of TIR emerging as PAH emission depends on the hardness of the ISRF.
This suggests that the PAH fraction in BCDs is not directly
controlled by metallicity, but rather
by the hardness of the radiation field, also responsible
for the destruction of the smallest PAH particles.
\item
The hardness of the ISRF in BCDs can be comparable to
that in AGN. Because of this, the \rm [O\,{\sc iv}]/\rm [Ne\,{\sc ii}]\
flux ratio, often used as a diagnostic to distinguish
star-forming galaxies from AGN at solar abundances, cannot play that
role in the low-metallicity regime: in the \rm [O\,{\sc iv}]/\rm [Ne\,{\sc ii}]\ vs. 7.7\,\micron\
PAH band diagnostic diagram, low-metallicity BCDs occupy the same region
as metal-rich AGN.
On the contrary, the AGN/starburst diagnostic diagram proposed
by \citep{laurent00}, based on the continuum MIR slopes of these
types of objects, does a good job at separating AGN from starbursts, even
at low metallicity.
\item
Our IRS spectra reveal a variety of \rm H$_2$\ rotational lines,
and more than one third of the objects in our sample (8 BCDs) have $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$
detections in one or more of the four lowest-order transitions of \rm H$_2$.
The BCDs contain more \rm H$_2$\ than most SINGS galaxies, relative to $L_{\rm TIR}$.
This is quite contrary to the usual assertion that \rm H$_2$\ molecules
do not exist in a low-metallicity environment, because of the lack of
grains on which to form them.
Clearly, metal abundance cannot be the only factor guiding \rm H$_2$\
formation in low-metallicity BCDs.
\item
The mean ratio of \rm H$_2$\ to PAH emission is $\sim$6\%, or 10 times
larger than that for the SINGS galaxies (but comparable to SINGS AGN).
This difference arises mainly from the BCD deficit in PAH emission.
While PAHs can survive the impact
of FUV photons with energies between $\sim$11 and 13.6\,eV,
\rm H$_2$\ would be dissociated without self-shielding,
implying that \rm H$_2$\ could be somewhat self-shielded in BCDs.
\item
The mean excitation temperature derived for the warm gas
from the lower-order \rm H$_2$\ transitions is
$\sim$245\,K.
The temperatures of the hotter molecular component from the high-order
\rm H$_2$\ transitions
range from $\sim$820\,K (in Mrk\,996 and SBS\,1152$+$579) to $\sim$1600\,K.
These temperatures are consistent with those found for starburst
SINGS galaxies and ULIRGs.
The warm molecular gas masses in our BCDs range from $10^3$ to $10^8$\,\ensuremath{M_\odot},
and can be comparable to the neutral hydrogen gas mass.
\item
Some of our IRS spectra suggest tentatively the presence of molecules
other than \rm H$_2$, such as {\rm H}$_2${\rm O}\ and OH.
All three galaxies with {\rm H}$_2${\rm O}\ emission
also show a $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ \rm [O\,{\sc iv}]\ line,
consistent with a shock scenario for {\rm H}$_2${\rm O}\ production.
All four galaxies with OH emission also have high-order \rm H$_2$\ detections,
implying
the presence of hot, dense gas which would favor OH and {\rm H}$_2${\rm O}\ formation.
\end{enumerate}
Our IRS data have shown that the ISRF in metal-poor BCDs can
be as extreme in hardness as the radiation in AGN.
In fact, the deficit of PAH emission at low metallicity
and their hard radiation as indicated by \rm [O\,{\sc iv}]\ luminosity
would have placed some BCDs
as AGN in diagnostic diagrams such as that proposed by \citet{genzel98}.
Our analysis suggests that in the absence of an AGN,
low metal abundance is a necessary condition for such extreme fields.
However, we have also shown that
low metallicity alone does not guarantee an extreme ISRF
in terms of hardness and intensity.
Other factors such as compactness of the star-forming region
must play a role.
We have inferred a deficit of small PAHs, and find that
PAH destruction mechanisms are more closely related
to the hardness of the radiation field, rather than to metallicity.
This is a rather difficult exercise, since most of the \rm [O\,{\sc iv}]\
detections are in galaxies where we have little information on the
PAH population, and the converse is also true.
Nevertheless, there appears to be a close connection between
hardness of the ISRF and the properties of the PAHs;
this connection is tighter than that with metallicity.
In fact, we find correlations of PAH properties with ISRF
hardness, but almost none with metallicity, in spite of the
good correlations between hardness tracers and metal abundance.
It is not clear whether the origin of such hard radiation
can be stellar.
In their detailed modeling of the \rm [O\,{\sc iv}]\ line in Mrk 996,
\citet{thuan08} ran Costar models by \citet{schaerer97}
using stars with the highest effective temperature
(T$_{eff}$\,=53,000 K) and the hardest radiation possible
(corresponding to the hottest O3 stars).
They failed to reproduce the observed intensity of the \rm [O\,{\sc iv}]\ line
by a factor of several. They then considered Wolf-Rayet stars of type
WNE-w. According to calculations by \citet{crowther99}, models for
WNE-w stars (early nitrogen Wolf-Rayet stars with weak lines) show a
strong ionizing flux above 54.4 eV, in contrast to the WCE and WNL
stars, which show negligible fluxes above that energy. The line
intensity of \rm [O\,{\sc iv}]\ in Mrk 996 can be reproduced with such a population
of WNE-w stars. However,
\citet{thuan08} do not favor such an hypothesis because WNE-w stars are
very rare. Rather, those authors favor fast radiative shocks
\citep{thuanizotov05}
as the origin of the high-ionization radiation.
Shocks are a more realistic expectation in dense compact regions
with intense star
formation because of SN explosions and massive stellar
winds.
At low metallicity, the less efficient cooling may intensify
the effect of the shock.
We postpone an in-depth discussion of the origin of
the hard ionizing radiation for the whole of our sample to a future paper.
There, we will present detailed photoionization and shock models to account
for both optical$+$ IR emission line intensities.
\acknowledgments
This work is based on observations made with the {\it Spitzer Space Telescope},
which is operated by JPL/Caltech under NASA contract 1407.
Support for this work was provided by NASA {\it Spitzer}\ GO grant
JPL-1263707.
We acknowledge financial contribution from contract ASI-INAF
I/016/07/0.
\begin{deluxetable}{lrrcrr}
\tablecaption{Sample Galaxies \label{tab:sample}}
\tablewidth{0pt}
\tablehead{
\colhead{Name} &
\colhead{Distance\tablenotemark{a}} &
\colhead{Redshift} &
\colhead{12$+$log(O/H)} &
\colhead{$c_{H\beta}$\tablenotemark{b}} &
\colhead{He{\sc ii}/H$\beta$} \\
}
\startdata
CGCG007-025 & 20.8 & 0.00483 & 7.76 & 0.29 & 0.0136\\
HS0837+4717 & 174.3 & 0.04195 & 7.60 & 0.29 & 0.0228\\
HS1442+4250 & 12.6 & 0.00211 & 7.63 & 0.11 & 0.0297\\
HS2236+1344 & 86.4 & 0.02062 & 7.47 & 0.16 & 0.0106\\
Haro3 & 17.4 & 0.00323 & 8.32 & 0.24 & 0.0000\\
IIZw70 & 21.8 & 0.00394 & 8.04 & 0.21 & 0.0056\\
J0519+0007 & 181.8 & 0.04476 & 7.43 & 0.30 & 0.0249\\
Mrk5 & 15.4 & 0.00264 & 8.04 & 0.42 & 0.0387\\
Mrk36 & 7.6 & 0.00215 & 7.81 & 0.02 & 0.0244\\
Mrk209 & 4.8 & 0.00094 & 7.81 & 0.00 & 0.0117\\
Mrk450 & 15.1 & 0.00288 & 8.15 & 0.14 & 0.0031\\
Mrk724 & 20.0 & 0.00402 & 8.03 & 0.12 & 0.0000\\
Mrk996 & 21.8 & 0.00541 & 8.10 & 0.53 & 0.0000\\
Mrk1315 & 13.1 & 0.00282 & 8.25 & 0.16 & 0.0000\\
Mrk1329 & 13.1 & 0.00544 & 8.25 & 0.16 & 0.0000\\
SBS0917+527 & 35.4 & 0.00776 & 7.90 & 0.09 & 0.0221\\
SBS0946+558 & 25.5 & 0.00517 & 8.04 & 0.18 & 0.0149\\
SBS1030+583 & 35.3 & 0.00757 & 7.83 & 0.00 & 0.0234\\
SBS1152+579 & 74.5 & 0.01720 & 7.85 & 0.26 & 0.0152\\
SBS1415+437 & 11.7 & 0.00203 & 7.61 & 0.00 & 0.0227\\
Tol1924-416 & 38.4 & 0.00945 & 7.94 & 0.26 & 0.0190\\
UM311 & 23.0 & 0.00559 & 8.31 & 0.15 & 0.0000\\
\enddata
\tablenotetext{a}{Distance in Mpc, taken from NED.}
\tablenotetext{b}{Extinction coefficient: A$_{\rm H\beta}$=2.5,$c_{H\beta}$.}
\end{deluxetable}
\begin{deluxetable}{lrrrr}
\tablecaption{Short-Wavelength Fine-Structure Lines\tablenotemark{a} \label{tab:fspahfit}}
\tabletypesize{\footnotesize}
\tablewidth{0pt}
\tablecolumns{ 5}
\tablehead{
\colhead{Name}
& \colhead{\rm [Ar\,{\sc ii}]}
& \colhead{\rm [Ar\,{\sc iii}]}
& \colhead{\rm [S\,{\sc iv}]}
& \colhead{\rm [Ne\,{\sc ii}]}
\\
& \multicolumn{1}{c}{ 6.985\micron}
& \multicolumn{1}{c}{ 8.991\micron}
& \multicolumn{1}{c}{ 10.511\micron}
& \multicolumn{1}{c}{ 12.814\micron}
\\
& \multicolumn{1}{c}{ 15.759\,eV}
& \multicolumn{1}{c}{ 27.629\,eV}
& \multicolumn{1}{c}{ 34.830\,eV}
& \multicolumn{1}{c}{ 21.564\,eV}
}
\startdata
CGCG\,007$-$025 & \nodata (0.33) & 0.66 (0.18) & 5.98 (0.43) & \nodata (1.29) \\
Haro\,3 & 3.44 (0.97) & 10.60 (1.67) & 26.80 (3.70) & 27.50 (3.59) \\
HS\,0837$+$4717 & \nodata (0.13) & 0.28 (0.09) & 3.07 (0.41) & \nodata (1.23) \\
HS\,1442$+$4250 & \nodata (0.21) & \nodata (0.21) & \nodata (1.60) & \nodata (1.60) \\
HS\,2236$+$1344 & \nodata (0.12) & \nodata (0.12) & 2.50 (0.56) & \nodata (1.68) \\
II\,Zw\,70 & \nodata (0.72) & \nodata (0.72) & 4.03 (0.26) & 1.34 (0.30) \\
J0519$+$0007 & \nodata (0.22) & \nodata (0.22) & \nodata (1.17) & \nodata (1.17) \\
Mrk\,5 & 0.54 (0.08) & 0.52 (0.08) & 1.64 (0.13) & 1.08 (0.13) \\
Mrk\,36 & \nodata (0.27) & \nodata (0.27) & 2.36 (0.14) & \nodata (0.43) \\
Mrk\,209 & 1.26 (0.33) & 1.58 (0.26) & 8.63 (0.18) & \nodata (0.45) \\
Mrk\,450 & \nodata (0.21) & 1.46 (0.13) & 4.45 (0.35) & 0.91 (0.29) \\
Mrk\,724 & \nodata (0.28) & \nodata (0.28) & 3.92 (0.44) & \nodata (1.31) \\
Mrk\,996 & \nodata (1.23) & 1.87 (0.40) & 3.90 (0.32) & 4.55 (0.32) \\
Mrk\,1315 & 0.73 (0.05) & 1.36 (0.12) & 10.20 (0.48) & \nodata (1.43) \\
Mrk\,1329 & \nodata (0.46) & 1.80 (0.15) & 12.30 (0.27) & 1.27 (0.27) \\
SBS\,0917$+$527 & \nodata (0.12) & \nodata (0.12) & 1.46 (0.12) & \nodata (0.36) \\
SBS\,0946$+$558 & \nodata (0.15) & \nodata (0.15) & \nodata (1.63) & \nodata (1.63) \\
SBS\,1030$+$583 & \nodata (0.10) & \nodata (0.10) & 1.93 (0.59) & \nodata (1.77) \\
SBS\,1152$+$579 & \nodata (0.40) & 0.92 (0.13) & 6.32 (0.12) & \nodata (0.36) \\
SBS\,1415$+$437 & \nodata (0.18) & \nodata (0.18) & 1.02 (0.22) & \nodata (0.49) \\
Tol\,1924$-$416 & \nodata (0.80) & 2.56 (0.27) & 11.40 (0.34) & 1.88 (0.34) \\
UM\,311 & \nodata (0.29) & 1.71 (0.23) & 5.22 (0.69) & \nodata (2.08) \\
\enddata
\tablenotetext{a}{Fluxes are measured by integrating Gaussian profiles
as fit by PAHFIT, and
are in units of $10^{-17}$\,W\,m$^{-2}$.
$1\sigma$ uncertainties are given in parentheses; when there is no
flux available, the value reported in parentheses is the $3\sigma$ upper limit.}
\end{deluxetable}
\begin{deluxetable}{lrrrrrrrc}
\rotate
\tablecaption{Dust Features\tablenotemark{a} \label{tab:dfpahfit}}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.04in}
\tablewidth{0pt}
\tablecolumns{9}
\tablehead{
\colhead{Name}
& \colhead{DF\_5}
& \colhead{DF\_6}
& \colhead{DF\_7\tablenotemark{b}}
& \colhead{DF\_8}
& \colhead{DF\_11}
& \colhead{DF\_12}
& \colhead{DF\_17}
& \colhead{$\Sigma$PAH} \\
& \multicolumn{1}{c}{ 5.7\micron}
& \multicolumn{1}{c}{ 6.2\micron}
& \multicolumn{1}{c}{ 7.7\micron}
& \multicolumn{1}{c}{ 8.6\micron}
& \multicolumn{1}{c}{ 11.3\micron}
& \multicolumn{1}{c}{ 12.6\micron}
& \multicolumn{1}{c}{ 17\micron}
& Total\tablenotemark{c}\\
& & & & & & & Complex
}
\startdata
CGCG\,007$-$025 & \nodata (0.33) & \nodata (0.33) & 6.34 (0.82) & \nodata (0.33) & 2.81 (0.51) & \nodata (1.28) & \nodata (0.33) & 12.79 (1.12) \\
& \nodata & \nodata & 7.459 & \nodata & 11.199 & \nodata \\
& \nodata & \nodata & 0.470 & \nodata & 0.067 & \nodata \\
Haro\,3 & 20.30 (3.53) & 101.00 ( 10.60) & 188.50 ( 31.85) & 56.30 ( 14.40) & 148.90 ( 41.95) & 111.00 ( 19.30) & \nodata (1.39) & 680.60 ( 60.62) \\
& 5.744 & 6.245 & 7.808 & 8.682 & 11.335 & 12.673 \\
& 0.188 & 0.194 & 0.362 & 0.279 & 0.544 & 0.798 \\
HS\,0837$+$4717 & \nodata (2.91) & \nodata (0.13) & \nodata (0.13) & \nodata (0.13) & 1.37 (0.41) & \nodata (1.23) & \nodata (0.13) & 2.95 (0.57) \\
& \nodata & \nodata & \nodata & \nodata & 11.214 & \nodata \\
& \nodata & \nodata & \nodata & \nodata & 0.090 & \nodata \\
HS\,1442$+$4250 & \nodata (0.21) & 0.94 (0.24) & \nodata (0.21) & \nodata (0.57) & 3.46 (0.53) & \nodata (1.60) & \nodata (0.21) & 6.15 (0.79) \\
& \nodata & 6.294 & \nodata & \nodata & 11.347 & \nodata \\
& \nodata & 0.094 & \nodata & \nodata & 0.545 & \nodata \\
HS\,2236$+$1344 & \nodata (0.12) & \nodata (0.12) & \nodata (0.12) & \nodata (0.12) & \nodata (1.68) & \nodata (1.68) & \nodata (0.12) & \nodata (0.12) \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
II\,Zw\,70 & \nodata (0.72) & \nodata (3.78) & 10.81 (1.98) & \nodata (0.72) & \nodata (0.66) & \nodata (0.66) & \nodata (0.72) & 12.39 (2.03) \\
& \nodata & \nodata & 7.808 & \nodata & \nodata & \nodata \\
& \nodata & \nodata & 0.207 & \nodata & \nodata & \nodata \\
J0519$+$0007 & \nodata (0.22) & \nodata (0.22) & 2.45 (0.55) & \nodata (0.22) & \nodata (1.17) & \nodata (1.17) & \nodata (0.22) & 4.24 (0.67) \\
& \nodata & \nodata & 7.405 & \nodata & \nodata & \nodata \\
& \nodata & \nodata & 0.467 & \nodata & \nodata & \nodata \\
Mrk\,1315 & \nodata (0.15) & 3.97 (0.12) & 1.91 (0.10) & 3.47 (0.48) & 1.45 (0.48) & \nodata (1.43) & \nodata (0.15) & 15.24 (0.75) \\
& \nodata & 6.201 & 7.500 & 8.648 & 11.256 & \nodata \\
& \nodata & 0.279 & 0.165 & 0.431 & 0.088 & \nodata \\
Mrk\,1329 & \nodata (5.70) & \nodata (3.03) & 9.31 (0.99) & 4.70 (0.87) & 6.94 (0.54) & 1.89 (0.27) & 1.85 (0.49) & 24.69 (1.53) \\
& \nodata & \nodata & 7.520 & 8.537 & 11.272 & 12.627 \\
& \nodata & \nodata & 0.474 & 0.166 & 0.541 & 0.265 \\
Mrk\,209 & \nodata (2.82) & \nodata (0.32) & \nodata (2.37) & \nodata (0.32) & \nodata (0.45) & \nodata (0.45) & \nodata (0.32) & 7.41 (1.28) \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
Mrk\,36 & \nodata (3.00) & \nodata (0.27) & \nodata (0.27) & \nodata (0.27) & \nodata (0.43) & \nodata (0.43) & \nodata (0.27) & 4.73 (1.32) \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
Mrk\,450 & 2.06 (0.25) & 6.16 (0.23) & 16.40 (1.42) & 5.95 (0.85) & 5.46 (1.61) & \nodata (0.76) & \nodata (0.21) & 36.03 (2.33) \\
& 5.800 & 6.291 & 7.520 & 8.520 & 11.331 & \nodata \\
& 0.169 & 0.283 & 0.659 & 0.255 & 0.388 & \nodata \\
Mrk\,5 & 2.17 (0.20) & 3.27 (0.14) & 7.47 (0.71) & 3.45 (0.39) & 5.42 (0.64) & \nodata (0.39) & \nodata (0.24) & 24.33 (1.08) \\
& 5.791 & 6.287 & 7.801 & 8.708 & 11.326 & \nodata \\
& 0.214 & 0.283 & 0.281 & 0.440 & 0.544 & \nodata \\
Mrk\,724 & \nodata (0.28) & \nodata (0.28) & \nodata (0.28) & \nodata (2.16) & \nodata (1.31) & \nodata (1.31) & \nodata (0.28) & \nodata (0.28) \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
& \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
Mrk\,996 & 7.03 (1.45) & 15.50 (4.32) & 17.90 (4.35) & \nodata (9.63) & 11.80 (2.02) & 3.53 (0.71) & \nodata (0.66) & 60.66 (6.71) \\
& 5.757 & 6.236 & 7.493 & \nodata & 11.276 & 12.684 \\
& 0.101 & 0.179 & 0.472 & \nodata & 0.250 & 0.082 \\
SBS\,0917$+$527 & \nodata (0.12) & \nodata (0.12) & \nodata (0.12) & 1.42 (0.23) & 1.00 (0.29) & 0.37 (0.12) & \nodata (0.12) & 2.79 (0.39) \\
& \nodata & \nodata & \nodata & 8.559 & 11.230 & 12.601 \\
& \nodata & \nodata & \nodata & 0.167 & 0.539 & 0.082 \\
SBS\,0946$+$558 & \nodata (0.15) & \nodata (0.15) & 0.89 (0.28) & \nodata (0.15) & 2.74 (0.54) & \nodata (1.63) & \nodata (0.15) & 12.11 (1.03) \\
& \nodata & \nodata & 7.520 & \nodata & 11.230 & \nodata \\
& \nodata & \nodata & 0.474 & \nodata & 0.539 & \nodata \\
SBS\,1030$+$583 & \nodata (0.10) & \nodata (0.10) & 4.75 (0.40) & 0.67 (0.14) & \nodata (1.77) & \nodata (1.77) & \nodata (0.10) & 5.42 (0.42) \\
& \nodata & \nodata & 7.520 & 8.534 & \nodata & \nodata \\
& \nodata & \nodata & 0.474 & 0.166 & \nodata & \nodata \\
SBS\,1152$+$579 & 3.30 (0.70) & \nodata (0.40) & 12.20 (0.67) & 3.41 (0.46) & \nodata (0.36) & \nodata (0.36) & \nodata (0.40) & 24.15 (1.11) \\
& 5.800 & \nodata & 7.520 & 8.710 & \nodata & \nodata \\
& 0.305 & \nodata & 0.474 & 0.510 & \nodata & \nodata \\
SBS\,1415$+$437 & \nodata (0.18) & \nodata (0.18) & 12.84 (0.25) & \nodata (2.04) & \nodata (0.49) & \nodata (0.49) & \nodata (0.18) & 12.84 (0.25) \\
& \nodata & \nodata & 7.782 & \nodata & \nodata & \nodata \\
& \nodata & \nodata & 0.258 & \nodata & \nodata & \nodata \\
Tol\,1924$-$416 & \nodata (0.80) & \nodata (0.80) & 6.39 (1.40) & 2.08 (0.62) & 2.32 (0.58) & 3.09 (0.41) & \nodata (0.80) & 13.88 (1.69) \\
& \nodata & \nodata & 7.507 & 8.510 & 11.230 & 12.643 \\
& \nodata & \nodata & 0.473 & 0.166 & 0.180 & 0.266 \\
UM\,311 & \nodata (2.37) & 12.70 (2.14) & 38.76 (5.37) & 4.27 (1.29) & 5.53 (0.69) & \nodata (2.08) & \nodata (0.29) & 64.98 (6.00) \\
& \nodata & 6.216 & 7.822 & 8.634 & 11.237 & \nodata \\
& \nodata & 0.163 & 0.207 & 0.196 & 0.067 & \nodata \\
\enddata
\tablenotetext{a}{Dust features are integrated over the entire
best-fit Drude profile as performed by PAHFIT.
In the first line, fluxes are in units of $10^{-17}$\,W\,m$^{-2}$.
$1\sigma$ uncertainties are given in parentheses; when there is no
flux available, the value reported in parentheses is the $3\sigma$ upper limit.
The PAHFIT best-fit central wavelengths in \micron\ are given in the second line,
and the best-fit FWHM in \micron\ in the third. }
\tablenotetext{b}{This complex includes the sum of features
from 7.4 to 7.8\,\micron; the FWHM corresponds to the longest wavelength feature.}
\tablenotetext{c}{This total includes the sum of features
in the table, together with features identified as PAHs by PAHFIT
with wavelengths from
6.6 to 6.8\,\micron,
8.3\,\micron, and
10.6 to 10.8\,\micron.
}
\end{deluxetable}
\begin{deluxetable}{lrrrrrrrrrr}
\rotate
\tablecaption{Long-Wavelength Fine-Structure Lines\tablenotemark{a} \label{tab:fsgauss}}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.04in}
\tablewidth{0pt}
\tablecolumns{ 9}
\tablehead{
\colhead{Name}
& \colhead{\rm [Ne\,{\sc iii}]}
& \colhead{\rm [S\,{\sc iii}]}
& \colhead{\rm [O\,{\sc iv}]}
& \colhead{\rm [Fe\,{\sc ii}]}
& \colhead{\rm [Fe\,{\sc iii}]}
& \colhead{\rm [S\,{\sc iii}]}
& \colhead{\rm [Si\,{\sc ii}]}
& \colhead{\rm [Ne\,{\sc iii}]}
\\
& \multicolumn{1}{c}{ 15.555\micron}
& \multicolumn{1}{c}{ 18.713\micron}
& \multicolumn{1}{c}{ 25.890\micron}
& \multicolumn{1}{c}{ 25.988\micron}
& \multicolumn{1}{c}{ 33.038\micron}
& \multicolumn{1}{c}{ 33.481\micron}
& \multicolumn{1}{c}{ 34.815\micron}
& \multicolumn{1}{c}{ 36.014\micron}
\\
& \multicolumn{1}{c}{ 40.962\,eV}
& \multicolumn{1}{c}{ 23.330\,eV}
& \multicolumn{1}{c}{ 54.934\,eV}
& \multicolumn{1}{c}{ 7.870\,eV}
& \multicolumn{1}{c}{ 16.180\,eV}
& \multicolumn{1}{c}{ 23.330\,eV}
& \multicolumn{1}{c}{ 8.151\,eV}
& \multicolumn{1}{c}{ 40.962\,eV}
}
\startdata
CGCG\,007$-$025 & 5.55 (0.49) & \nodata (1.28) & \nodata (1.49) & \nodata (1.49) & \nodata (1.77) & 3.48 (0.67) & 2.89 (0.50) & \nodata (1.49) \\
Haro\,3 & 79.58 (0.27) & 40.99 (0.27) & 5.38 (0.93) & \nodata (2.80) & 27.13 (0.93) & 116.41 (0.93) & 53.47 (0.93) & 12.11 (0.93) \\
HS\,0837$+$4717 & 1.54 (0.41) & \nodata (1.23) & \nodata (0.92) & \nodata (0.92) & \nodata (0.92) & \nodata (0.92) & \nodata (0.92) & \nodata (0.92) \\
HS\,1442$+$4250 & \nodata (1.60) & \nodata (1.60) & \nodata (1.00) & \nodata (1.00) & \nodata (1.00) & \nodata (1.00) & \nodata (1.00) & \nodata (1.00) \\
HS\,2236$+$1344 & \nodata (1.86) & \nodata (1.68) & \nodata (0.54) & \nodata (0.54) & \nodata (0.54) & \nodata (0.54) & \nodata (0.54) & \nodata (0.54) \\
II\,Zw\,70 & 6.78 (0.35) & 3.95 (0.22) & 0.77 (0.17) & \nodata (0.60) & \nodata (0.51) & 10.09 (0.17) & 6.65 (0.53) & \nodata (0.51) \\
J0519$+$0007 & \nodata (1.17) & \nodata (1.17) & \nodata (0.67) & \nodata (0.67) & \nodata (0.67) & \nodata (0.67) & \nodata (0.99) & \nodata (0.67) \\
Mrk\,5 & 4.31 (0.13) & 1.90 (0.13) & \nodata (0.21) & 0.24 (0.07) & \nodata (0.21) & 5.02 (0.07) & 2.45 (0.07) & 1.01 (0.07) \\
Mrk\,36 & 2.70 (0.14) & 1.53 (0.14) & 1.11 (0.30) & \nodata (0.91) & \nodata (0.91) & 2.63 (0.30) & 2.97 (0.30) & \nodata (0.91) \\
Mrk\,209 & 5.57 (0.15) & 2.69 (0.15) & 1.46 (0.19) & \nodata (0.57) & \nodata (0.57) & 5.66 (0.19) & \nodata (0.57) & \nodata (0.57) \\
Mrk\,450 & 7.00 (1.33) & 4.27 (0.25) & \nodata (0.64) & \nodata (0.64) & \nodata (0.64) & 5.68 (0.21) & 3.03 (0.21) & 1.96 (0.21) \\
Mrk\,724 & 3.98 (0.44) & 2.89 (0.44) & \nodata (1.73) & \nodata (1.73) & \nodata (1.73) & \nodata (1.73) & \nodata (1.73) & \nodata (1.73) \\
Mrk\,996 & 8.95 (0.32) & 7.31 (0.32) & 1.46 (0.21) & 0.96 (0.32) & 0.90 (0.21) & 16.43 (0.21) & 5.75 (0.21) & \nodata (0.63) \\
Mrk\,1315 & 9.74 (0.48) & 5.09 (0.48) & \nodata (0.80) & \nodata (0.80) & \nodata (0.80) & 7.26 (0.27) & 1.81 (0.27) & \nodata (0.80) \\
Mrk\,1329 & 15.51 (0.27) & 8.07 (0.32) & \nodata (2.10) & \nodata (2.10) & \nodata (2.10) & 8.59 (0.70) & 2.83 (0.70) & \nodata (2.10) \\
SBS\,0917$+$527 & 1.00 (0.12) & 0.67 (0.12) & \nodata (0.76) & \nodata (0.76) & \nodata (0.76) & \nodata (0.76) & \nodata (0.76) & \nodata (0.76) \\
SBS\,0946$+$558 & \nodata (1.63) & \nodata (1.63) & \nodata (0.51) & \nodata (0.51) & \nodata (0.51) & 1.56 (0.17) & 0.57 (0.17) & \nodata (0.51) \\
SBS\,1030$+$583 & \nodata (1.77) & \nodata (1.77) & \nodata (0.83) & \nodata (0.83) & \nodata (0.83) & \nodata (0.83) & \nodata (0.83) & \nodata (0.83) \\
SBS\,1152$+$579 & 2.45 (0.12) & 1.38 (0.14) & \nodata (0.76) & \nodata (0.76) & \nodata (0.76) & 2.68 (0.25) & \nodata (0.76) & \nodata (0.76) \\
SBS\,1415$+$437 & 1.22 (0.16) & 1.09 (0.16) & 0.61 (0.13) & \nodata (0.40) & \nodata (0.40) & 1.41 (0.13) & 2.15 (0.13) & \nodata (0.40) \\
Tol\,1924$-$416 & 23.99 (0.34) & 7.40 (0.34) & 3.33 (0.61) & \nodata (3.51) & \nodata (1.84) & 18.57 (0.61) & 15.23 (0.74) & \nodata (6.54) \\
UM\,311 & 13.75 (0.69) & 2.53 (0.69) & \nodata (1.99) & \nodata (1.99) & \nodata (1.99) & \nodata (1.99) & 4.01 (0.66) & \nodata (1.99) \\
\enddata
\tablenotetext{a}{Fluxes are measured by integrating Gaussian profiles, and
are in units of $10^{-17}$\,W\,m$^{-2}$.
$1\sigma$ uncertainties are given in parentheses; when there is no
flux available, the value reported in parentheses is the $3\sigma$ upper limit.}
\end{deluxetable}
\begin{deluxetable}{lrrrrrrrr}
\rotate
\tablecaption{\rm H$_2$\ Lines\tablenotemark{a} \label{tab:h2}}
\tabletypesize{\footnotesize}
\setlength{\tabcolsep}{0.04in}
\tablewidth{0pt}
\tablecolumns{9}
\tablehead{
\colhead{Name} &
\colhead{S(7)} &
\colhead{S(6)} &
\colhead{S(5)} &
\colhead{S(4)} &
\colhead{S(3)} &
\colhead{S(2)} &
\colhead{S(1)\tablenotemark{b}} &
\colhead{S(0)\tablenotemark{b}} \\
& \multicolumn{1}{c}{ 5.511\micron}
& \multicolumn{1}{c}{ 6.109\micron}
& \multicolumn{1}{c}{ 6.910\micron}
& \multicolumn{1}{c}{ 8.025\micron}
& \multicolumn{1}{c}{ 9.665\micron}
& \multicolumn{1}{c}{ 12.279\micron}
& \multicolumn{1}{c}{ 17.035\micron}
& \multicolumn{1}{c}{ 28.219\micron}
}
\startdata
CGCG\,007$-$025 & 0.46 (0.13) & \nodata (0.33) & \nodata (0.33) & 0.72 (0.21) & \nodata (0.33) & 1.67 (0.43) & 5.45 (0.43) & \nodata (1.49) \\
Haro\,3 & 2.03 (0.46) & \nodata (1.39) & \nodata (1.39) & \nodata (1.39) & \nodata (1.39) & \nodata (0.82) & \nodata (0.82) & \nodata (2.80) \\
HS\,0837$+$4717 & 0.44 (0.19) & \nodata (0.13) & \nodata (0.13) & \nodata (0.13) & 0.18 (0.07) & \nodata (1.23) & 1.07 (0.41) & \nodata (0.92) \\
HS\,1442$+$4250 & 0.47 (0.17) & 0.41 (0.14) & \nodata (0.21) & \nodata (0.21) & 0.17 (0.08) & \nodata (1.60) & \nodata (1.60) & \nodata (1.00) \\
II\,Zw\,70 & \nodata (0.72) & 1.53 (0.64) & \nodata (0.72) & 1.34 (0.46) & 0.69 (0.31) & \nodata (0.66) & \nodata (0.66) & \nodata (0.51) \\
J0519$+$0007 & 0.90 (0.17) & \nodata (0.22) & \nodata (0.22) & 0.34 (0.11) & 0.22 (0.07) & \nodata (1.17) & \nodata (1.17) & \nodata (0.67) \\
Mrk\,5 & 0.38 (0.08) & 0.60 (0.08) & 0.35 (0.08) & 0.56 (0.08) & \nodata (0.24) & \nodata (0.39) & \nodata (0.39) & \nodata (0.21) \\
Mrk\,36 & \nodata (0.27) & \nodata (0.27) & \nodata (0.27) & \nodata (0.27) & 0.94 (0.14) & \nodata (0.43) & \nodata (0.43) & \nodata (0.91) \\
Mrk\,209 & \nodata (0.32) & \nodata (0.32) & \nodata (0.32) & \nodata (0.32) & 0.52 (0.23) & 0.49 (0.16) & \nodata (0.45) & \nodata (0.57) \\
Mrk\,450 & \nodata (0.21) & 0.78 (0.08) & 0.58 (0.07) & 0.39 (0.18) & 0.20 (0.08) & \nodata (0.76) & \nodata (0.76) & \nodata (0.64) \\
Mrk\,996 & \nodata (0.66) & \nodata (0.66) & 1.19 (0.48) & \nodata (0.66) & 1.15 (0.35) & \nodata (0.97) & 0.74 (0.32) & 0.71 (0.21) \\
Mrk\,1315 & \nodata (0.15) & \nodata (0.15) & 0.18 (0.05) & \nodata (0.15) & \nodata (0.15) & \nodata (1.43) & \nodata (1.43) & \nodata (0.80) \\
Mrk\,1329 & \nodata (0.46) & 1.05 (0.21) & \nodata (0.46) & 0.78 (0.18) & \nodata (0.46) & \nodata (0.80) & \nodata (0.80) & \nodata (2.10) \\
SBS\,0917$+$527 & \nodata (0.12) & \nodata (0.12) & \nodata (0.12) & 0.38 (0.16) & \nodata (0.12) & \nodata (0.36) & \nodata (0.36) & \nodata (0.76) \\
SBS\,0946$+$558 & \nodata (0.15) & \nodata (0.15) & 0.21 (0.05) & \nodata (0.15) & \nodata (0.15) & \nodata (1.63) & \nodata (1.63) & \nodata (0.51) \\
SBS\,1030$+$583 & 0.53 (0.13) & \nodata (0.10) & \nodata (0.10) & \nodata (0.10) & \nodata (0.10) & \nodata (1.77) & \nodata (1.77) & \nodata (0.83) \\
SBS\,1152$+$579 & \nodata (0.40) & \nodata (0.40) & \nodata (0.40) & 0.41 (0.15) & 0.39 (0.13) & 0.30 (0.12) & \nodata (0.36) & \nodata (0.76) \\
SBS\,1415$+$437 & \nodata (0.18) & \nodata (0.18) & 2.04 (0.42) & \nodata (0.18) & \nodata (0.18) & 0.43 (0.20) & \nodata (0.49) & \nodata (0.40) \\
Tol\,1924$-$416 & \nodata (0.80) & \nodata (0.80) & \nodata (0.80) & \nodata (0.80) & 2.35 (0.27) & \nodata (1.03) & \nodata (1.03) & \nodata (1.84) \\
UM\,311 & \nodata (0.29) & \nodata (0.29) & 0.66 (0.27) & 0.73 (0.35) & \nodata (0.29) & \nodata (2.08) & 1.49 (0.69) & \nodata (1.99) \\
\enddata
\tablenotetext{a}{Fluxes are measured by integrating Gaussian profiles, and
are in units of $10^{-17}$\,W\,m$^{-2}$.
Unlike other tables, lines shown here as detections are 2$\sigma$.
$1\sigma$ uncertainties are given in parentheses; when there is no
flux available, the value reported in parentheses is the $3\sigma$ upper limit.}
\tablenotetext{b}{Unlike the shorter-wavelength lines which are fitted with
Gaussian profiles by PAHFIT, these are
measured by independent Gaussian profiles as described in the text.
}
\end{deluxetable}
\begin{deluxetable}{lrrrrrrrrrr}
\tablecaption{Water and Hydroxyl Lines\tablenotemark{a} \label{tab:molecules}}
\tabletypesize{\footnotesize}
\tablewidth{0pt}
\tablecolumns{6}
\tablehead{
\colhead{Name}
& \colhead{OH1/2-3/2}
& \colhead{o-H2O}
& \colhead{p-H2O}
& \colhead{OH19/2-17/2}
& \colhead{OH19/2-17/2}
\\
& \multicolumn{1}{c}{ 28.939\micron}
& \multicolumn{1}{c}{ 29.836\micron}
& \multicolumn{1}{c}{ 29.885\micron}
& \multicolumn{1}{c}{ 30.346\micron}
& \multicolumn{1}{c}{ 30.657\micron}
\\
}
\startdata
Haro\,3 & \nodata (2.80) & \nodata (2.80) & 4.57 (0.93) & \nodata (2.80) & \nodata (2.80) \\
II\,Zw\,70 & \nodata (0.51) & \nodata (0.51) & \nodata (0.51) & 0.59 (0.17) & \nodata (0.51) \\
Mrk\,5 & \nodata (0.21) & \nodata (0.21) & \nodata (0.21) & \nodata (0.21) & 0.32 (0.07) \\
Mrk\,996 & \nodata (0.63) & 2.48 (0.40) & \nodata (0.63) & \nodata (0.63) & \nodata (0.63) \\
Mrk\,1315 & 1.65 (0.27) & \nodata (0.80) & \nodata (0.81) & \nodata (0.80) & \nodata (0.81) \\
Tol\,1924$-$416 & 2.67 (0.61) & \nodata (1.84) & 2.30 (0.61) & \nodata (1.84) & \nodata (1.84) \\
\enddata
\tablenotetext{a}{Fluxes are measured by integrating Gaussian profiles, and
are in units of $10^{-17}$\,W\,m$^{-2}$.
$1\sigma$ uncertainties are given in parentheses; when there is no
flux available, the value reported in parentheses is the $3\sigma$ upper limit.}
\end{deluxetable}
\begin{figure}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1a.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1b.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1c.ps}
}
}
\vspace{0.3cm}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1d.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1e.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1f.ps}
}
}
\vspace{0.3cm}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1g.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1h.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1i.ps}
}}
\caption{IRS SL$+$SH spectra from 1 to 15\micron, with the best-fitting PAHFIT
model superimposed shown as a red curve.
The vertical axis is in units of mJy.
Significant emission-line identifications are shown as labeled vertical lines.
The top panels report the best-fit reduced $\chi^2_{\nu}$, obtained over the wavelength
region shown.
The bottom panels show the residuals from the PAHFIT models.
\label{fig:pahfit1}}
\end{figure}
\clearpage
\setcounter{figure}{0}
\begin{figure}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1j.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1k.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1l.ps}
}}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1m.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1n.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1o.ps}
}}
\vspace{0.3cm}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1p.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1q.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1r.ps}
}}
\caption{\ continued. }
\end{figure}
\clearpage
\setcounter{figure}{0}
\begin{figure}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1s.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1t.ps}
\hspace{-0.1cm}
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1u.ps}
}}
\vspace{0.3cm}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=40 251 587 713]{f1v.ps}
}}
\caption{\ continued. }
\end{figure}
\begin{figure}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2a.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2b.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2c.ps}
}}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2d.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2e.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2f.ps}
}}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2g.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f2h.ps}
}}
\caption{IRS LH spectra of the 26\micron\ region around \rm [O\,{\sc iv}]\ and \rm [Fe\,{\sc ii}]\
for those objects with either \rm [O\,{\sc iv}]\ or \rm [Fe\,{\sc ii}]\ detections.
As in Fig. \ref{fig:pahfit1}, all marked features are considered significant
(see Table \ref{tab:fsgauss}).
The MIPS24 total flux is marked with a red open circle.
Only those spectral points with S/N$>$3 are shown.
\label{fig:oiv}}
\end{figure}
\clearpage
\begin{figure}
\centerline{
\hbox{
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f3a.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f3b.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f3c.ps}
}}
\centerline{
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f3d.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f3e.ps}
\includegraphics[angle=0,width=0.33\linewidth,bb=17 162 590 570]{f3f.ps}
\hbox{
}}
\caption{Close-up of tentative {\rm H}$_2${\rm O}\ (29.8\micron)
and OH (28.9\,\micron) detections.
All marked lines are considered to be significant detections (see Table \ref{tab:molecules}).
Only those spectral points with S/N$>$3 are shown.
\label{fig:h2o_oh}}
\end{figure}
\clearpage
\begin{figure}
\bigskip
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f4.ps} }
\caption{\rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ (18.7\,\micron) flux ratios (top panel) and \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ ratios
(bottom) plotted against the nebular oxygen abundance, 12$+$log(O/H).
The BCDs in our sample are shown as filled circles,
while the SINGS galaxies \citep{dale09}
as open squares (\rm H{\sc ii}\ region-nuclei) and open triangles (AGN).
The horizontal dotted line in the lower panel shows the mean
neon ratio (5.42) for 12$+$log(O/H)$\leq$8.3 (not considering lower limits).
\label{fig:nes_oh}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f5.ps} }
\caption{\rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ (18.7\,\micron) flux ratios (top panel) and \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ ratios
(bottom) plotted against the optical \rm He{\sc ii}/H{$\beta$}\ ratio.
\label{fig:nes_heii}}
\end{figure}
\bigskip
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.5\linewidth,bb=19 159 588 547]{f6a.ps}
\hspace{-0.094\linewidth}
\includegraphics[angle=0,width=0.5\linewidth,bb=19 159 588 547]{f6b.ps} }
\caption{\rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ flux ratios plotted against the
optical \rm He{\sc ii}/H{$\beta$}\ ratio (left panel) and 12$+$log(O/H)\ (right).
The figure shows only those objects with 3$\sigma$ detections or better
in the \rm [O\,{\sc iv}]\ line.
\label{fig:oiv_heii_oh}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f7.ps} }
\caption{\rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]/ and \rm [S\,{\sc iv}]/\rm [S\,{\sc iii}]\ (18.7\,\micron) plotted against the \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ flux ratios.
As in Fig. \ref{fig:nes_oh}, BCDs are shown as filled circles,
and the SINGS galaxies \citep{dale09}
as open squares (\rm H{\sc ii}\ region-nuclei) and open triangles (AGN).
The figure shows only those objects with 3$\sigma$ detections or better
in the \rm [O\,{\sc iv}]\ line.
\label{fig:nes_oiv}}
\end{figure}
\clearpage
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f8.ps} }
\caption{$\langle\nu{\rm L}_\nu\rangle$ $_{24}$/($\langle\nu{\rm L}_\nu\rangle$ $_{71}+$ $\langle\nu{\rm L}_\nu\rangle$ $_{160}$ (upper panel)
and 12$+$log(O/H)\ (lower) plotted against the MIPS ratio $\langle\nu{\rm L}_\nu\rangle$ $_{71}$/$\langle\nu{\rm L}_\nu\rangle$ $_{160}$.
As in Fig. \ref{fig:nes_oh}, BCDs are shown as filled circles,
and the SINGS galaxies \citep{dale09}
as open squares (\rm H{\sc ii}\ region-nuclei) and open triangles (AGN).
Both panels show correlations significant within the BCD sample
at $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$}$99\% confidence level, as described in the text.
\label{fig:p24oh_r71}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f9.ps} }
\caption{\rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]/ (lower panel) and \rm [O\,{\sc iv}]/\rm [Si\,{\sc ii}]\ (upper) plotted against
MIPS ratio $\langle\nu{\rm L}_\nu\rangle$ $_{71}$/$\langle\nu{\rm L}_\nu\rangle$ $_{160}$.
As in Fig. \ref{fig:nes_oh}, BCDs are shown as filled circles,
and the SINGS galaxies \citep{dale09}
as open squares (\rm H{\sc ii}\ region-nuclei) and open triangles (AGN).
In the upper panel,
the $3\sigma$ \rm [Si\,{\sc ii}]\ upper limit (with \rm [O\,{\sc iv}]\ detection) is shown
as an open circle.
\label{fig:neoiv_r71}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f10.ps} }
\caption{Electron densities $n_{\rm e}$\ derived from
IR \rm [S\,{\sc iii}]\ line ratios vs. those inferred from optical \rm [S\,{\sc iii}]\ line
ratios.
The dotted line denotes equality.
\label{fig:densities}}
\end{figure}
\clearpage
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth,bb=19 171 590 545]{f11.ps} }
\caption{\rm [Si\,{\sc ii}]\ emission normalized by and plotted against TIR.
The blue dotted horizontal line indicates the BCD mean,
without any MIPS ULs and
without UM\,311 ($L_{\rm TIR}$$\sim 3\times10^{9}$\ensuremath{L_\odot}), since its TIR is overestimated
because of crowding (see text).
The red dashed line corresponds to the
mean for the SINGS sample \citep{roussel07}.
\label{fig:tir} }
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.8\linewidth,bb=18 156 591 593]{f12.ps} }
\caption{PAH strengths relative to total PAH luminosity vs. the total
PAH power normalized to TIR.
The four panels show the the following dust features (DF):
6.2 \micron\ (bottom left), 7.7 \micron\ (bottom right),
8.6 \micron\ (top left), 11.3 (top right) \micron\ features.
BCDs are plotted as filled (blue) circles, with TIR lower limits shown as open circles;
the TIR lower limits do not affect the vertical axis, only the placement along the horizontal one.
Open (green) squares correspond to SINGS HII nuclei, and open (red) triangles
to SINGS AGN \citep{smith07}.
The horizontal dashed lines give the SINGS sample medians \citep{smith07},
and the dotted ones the means for the BCD sample,
taking into account all objects with 7.7\,\micron\ detections.
The vertical dotted line corresponds to the BCD mean PAH power normalized to TIR.
\label{fig:pahfrac}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth,bb=19 171 590 545]{f13.ps} }
\caption{Median Drude profile widths (FWHM) and central wavelengths
for seven aromatic features detected ($\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$) in our sample.
The error bars are the standard deviations of the measurements,
and the numbers in parentheses are the numbers of BCDs with
significant detections.
Open squares correspond to SINGS sample means \citep{smith07}.
\label{fig:pahfwhm}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f14.ps} }
\caption{The integrated flux in the PAH features at 6.2\,\micron,
11.2\,\micron, and 12.7\,\micron, plotted in two sets of ratios:
6.2/11.2 vs. 12.7/11.2.
Only those BCDs with at least one ratio with 11.2\,\micron\ are shown:
filled (blue) circles correspond to those BCDs with 12$+$log(O/H)$\geq$8.1;
open ones to those objects with lower metallicities.
The other data are taken from \citet{hony01} and correspond
to planetary nebulae (PNe, open red triangles),
reflection nebulae (RNe, open red squares),
intermediate-mass star-forming regions (six-sided open green stars),
and \rm H{\sc ii}\ regions (open green hexagons).
The SINGS average is shown as a (black) $\times$, with error
bars reporting standard deviations over the sample \citep{smith07}.
\label{fig:pah62127}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f15.ps} }
\caption{The integrated flux in all the PAH features detected at $\geq 3\sigma$,
relative to the total IR flux, plotted against the
the total IR flux as described in the text.
BCDs are plotted as filled blue circles.
The non-AGN SINGS galaxies are plotted as open green squares, and the
AGN SINGS galaxies (Seyferts and LINERs) as open red triangles.
The blue dotted line shows the median $\Sigma$(PAH)/TIR
for our sample of 0.54\%; the red dashed line shows the
mean $\Sigma$(PAH)/TIR for the SINGS sample of $\sim$10\% \citep{smith07}.
\label{fig:pahtot_tir}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.5\linewidth,bb=19 159 588 547]{f16a.ps}
\hspace{-0.094\linewidth}
\includegraphics[angle=0,width=0.5\linewidth,bb=19 159 588 547]{f16b.ps} }
\caption{The integrated flux in all the PAH features detected at $\geq 3\sigma$,
relative to the total IR flux, plotted against the
the oxygen abundance (left panel) and the \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}]\ flux ratio (right).
BCDs are shown as filled blue circles,
the non-AGN SINGS galaxies as open green squares, and the
AGN SINGS galaxies (Seyferts and LINERs) as open red triangles.
The horizontal dotted line shows the BCD sample mean $\Sigma$(PAH)/TIR of 0.54\%,
and the dashed line the SINGS mean of 10\%.
The right panel shows only those objects with 3$\sigma$ detections or
better in each line; the best-fit linear regression is shown as a dotted line.
\label{fig:pahtot_ohne}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth,bb=19 159 588 547]{f17.ps} }
\caption{$\nu L_\nu$(7.7\micron)/TIR plotted again $L_{\rm TIR}$\ for the local
starbursts in \citet{houck07} (shown as open red squares)
and the BCDs (filled blue circles).
The dotted horizontal line segments correspond to the sample means:
0.044 for the BCDs, and 0.167 for the starbursts.
\label{fig:pah77}}
\end{figure}
\clearpage
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f18.ps} }
\caption{The ratios of the 7.7/11.3\,\micron\ bands plotted against \rm [Ne\,{\sc iii}]/\rm [Ne\,{\sc ii}].
BCDs with $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ detections are shown as filled blue circles,
SINGS AGN as open red triangles, and SINGS HII nuclei as open green squares.
The mean 7.7/11.3\,\micron\ ratio for the SINGS sample is shown as a horizontal
dotted line \citep{smith07}.
\label{fig:pah_ne}
}
\end{figure}
\clearpage
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth]{f19.ps} }
\caption{\rm [O\,{\sc iv}]\/\rm [Ne\,{\sc ii}]\ (or 1.7$\times$33.5\micron\ \rm [S\,{\sc iii}]) vs. the strength of
the 7.7\,\micron\ PAH feature.
The data for starburst galaxies, ULIRGs, and AGN are taken from \citet{genzel98},
and are marked by green open stars, green open squares, and red open triangles, respectively.
Our BCDs as filled blue circles, with upper limits for the PAH features shown
by left-pointing arrows; only the (8) BCDs with $\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 3\sigma$ \rm [O\,{\sc iv}]\ detections
are shown.
The ``mixing curve'' shown by a dashed line is made following \citet{genzel98};
various fractions of total luminosity in an AGN and a starburst are combined,
assuming a pure AGN has \rm [O\,{\sc iv}]/\rm [Ne\,{\sc ii}]/ $\sim$1, PAH strength $\sim$0.04,
and a pure starburst \rm [O\,{\sc iv}]\rm [Ne\,{\sc ii}]/ $\sim$0.002, PAH strength $\sim$3.6.
The 50\% mark is shown as a line segment along the mixing curve.
\label{fig:genzel_diag} }
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth,bb=19 159 588 547]{f20.ps} }
\caption{Diagnostic diagram as in \citet{peeters04a}, adapted from \citet{laurent00}.
The templates for the ``mixing'' curves are empirical,
defined to be the positions of M\,17 (\rm H{\sc ii}\ region), NGC\,7023
(an exposed PDR), and a pure AGN (assumed to have negligible 6.2\,\micron\ flux
in the PAH band).
Data are taken from \citet{peeters04a}, and include
Seyfert 1s (filled red triangles), Seyfert 2s (open red triangles),
ULIRGs (open green squares),
starburst galaxies (open green stars), and
Galactic star-forming regions (green $\times$).
Again, BCDs are shown as filled blue circles, with upper limits
for the 6.2\,\micron\ PAH feature shown as left-pointing arrows.
\label{fig:laurent_diag} }
\end{figure}
\clearpage
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.6\linewidth,bb=19 159 588 547]{f21.ps} }
\caption{Sum of S(0) to S(2) \rm H$_2$\ luminosity
normalized by and plotted against TIR.
All detections with $\simgt2\sigma$ are plotted; a $\times$
denotes the 3$\sigma$ detections.
Different symbols distinguish which \rm H$_2$\ lines are
considered in the sum:
Mrk\,996 with S(0) and S(1), and CGCG\,$005-027$\ with S(1) and S(2),
are shown by open circles;
the remaining BCDs with only S(1) or S(2) by open squares.
The red hatched area shows the range for SINGS galaxies
\citep{roussel07}, and
the blue dotted horizontal line indicates the BCD mean,
without UM\,311 ($L_{\rm TIR}$$\sim 3\times10^{9}$\ensuremath{L_\odot}) because
of the likely overestimate of its TIR because of source crowding (see text).
\label{fig:h2tir} }
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=0,width=0.28\linewidth,bb=19 145 588 712]{f22a.ps}
\hspace{-0.9cm}
\includegraphics[angle=0,width=0.28\linewidth,bb=19 145 588 712]{f22b.ps}
\hspace{-0.9cm}
\includegraphics[angle=0,width=0.28\linewidth,bb=19 145 588 712]{f22c.ps}
\hspace{-0.9cm}
\includegraphics[angle=0,width=0.28\linewidth,bb=19 145 588 712]{f22d.ps}
}
\caption{Excitation diagrams for the four BCDs with multiple significant \rm H$_2$\
detections and at least one significant ($\lower.5ex\hbox{$\;\buildrel>\over\sim\;$} 2.5\sigma$ for SBS\,1152$+$579)
detection in a low-order transition (see text).
3$\sigma$ upper limits are shown with vertical arrows.
The dotted lines correspond to best-fit estimates of the temperature of the
line-emitting gas (see text).
The parabolic fit for CGCG\,$005-027$\ is an experiment to show the
continuity of the data, and thus the continuous range of gas temperatures.
\label{fig:h2excitation} }
\end{figure}
\clearpage
|
\section{Introduction}
This paper is concerned with the development of harmonic analysis and the related quantum mechanics in superspace (see e.g. \cite{DBS5,DBE1,DBS9,CDBS1}) and more specifically the Hermite polynomials introduced in that setting (see \cite{DBS3}).
To fix ideas, let us first consider the quantum harmonic oscillator in $\mathbb{R}^{m}$ with rotational $O(m)$-symmetry. This system is described by the Schr\"odinger equation
\begin{equation}
-\frac{\nabla^{2}}{2} \psi + \frac{r^{2}}{2} \psi = E \psi
\label{HOclass}
\end{equation}
and is typically solved in two different ways: either one uses cartesian co-ordinates or one uses spherical co-ordinates, yielding two types of Hermite polynomials. The use of cartesian co-ordinates reduces the problem to $m$ one-dimensional oscillators and yields solutions of the type
\begin{equation}
\psi_{k_{1}, \ldots, k_{m}}^{b} \propto H_{k_{1}}(x_{1}) \ldots H_{k_{m}}(x_{m}) e^{-r^{2}/2} , \qquad E= \frac{m}{2} + \sum_{i=1}^{m} k_{i}
\label{cartHerm}
\end{equation}
with $H_{k}(x) = (-1)^{k} \exp{(-x^2)} \frac{d^k}{dx^k} \exp{(-x^2)} $ the one-dimensional Hermite polynomials. If, on the other hand, one uses spherical co-ordinates, the eigenfunctions are expressed in terms of spherical harmonics, yielding
\begin{equation}
\phi_{j,k,l}^{b} \propto L_{j}^{\frac{m}{2} + k-1}(r^2) H_{k}^{(l)} e^{-r^{2}/2} , \qquad E= \frac{m}{2} + (2j+ k)
\label{sphHerm}
\end{equation}
with $L_{\alpha}^{\beta}$ the generalized Laguerre polynomials and $H_{k}^{(l)}$ a basis for the space of spherical harmonics of degree $k$. We will adopt the name spherical Hermite polynomials in this case. Both techniques completely solve the quantum harmonic oscillator and yield bases $\{\psi_{k_{1}, \ldots, k_{m}}^{b}\}$ and $\{ \phi_{j,k,l}^{b} \}$ of $L_{2}(\mathbb{R}^{m}, dV(\underline{x}))$ which are moreover orthonormal.
Second, it is possible to restrict the $O(m)$-symmetry in equation (\ref{HOclass}) to a finite reflection group $\mathcal{G} < O(m)$. The related quantum system is then of Calogero-Moser-Sutherland type (see \cite{vD}) and given by
\[
-\frac{\Delta_{\kappa}}{2} \psi + \frac{r^{2}}{2} \psi = E \psi
\]
where $\Delta_{\kappa}$ is the so-called Dunkl Laplacian related to $\mathcal{G}$ (see e.g. \cite{MR951883}, \cite{MR1827871}). Again, this equation can be solved using two types of Hermite polynomials. The first type is a generalization of the cartesian type (formula (\ref{cartHerm})) and was introduced by R\"osler (see \cite{MR1620515}). The second type (see \cite{MR1199124}) generalizes the spherical Hermite polynomials (formula (\ref{sphHerm})) to the Dunkl setting. (Explicit formulae will be presented in section \ref{DunklHermite}).
Again both types of functions form orthonormal bases for the weighted $L_{2}$-space $L_{2}(\mathbb{R}^{m}, w_{\kappa}(\underline{x}) dV(\underline{x}))$. In this notation, $w_{\kappa}(\underline{x}) dV(\underline{x})$ is the $\mathcal{G}$-invariant measure in $\mathbb{R}^{m}$.
Let us now turn our attention to the problem at hand. Again we make a change of symmetry by considering the symplectic group $Sp(2n)$ instead of $O(m)$ or $\mathcal{G}$. The corresponding quantum problem is now formulated in a Grassmann algebra (which can be seen as a purely fermionic superspace). Again, as we will show in section \ref{FermionicMehler} and the beginning of section \ref{supersuper}, two bases of this Grassmann algebra exist (mimicking the previously discussed bases), which are now orthogonal with respect to a canonically defined inner product using the Berezin integral (see \cite{MR732126}).
However, the picture changes dramatically when one considers a full superspace with symmetry $O(m)\times Sp(2n)$. Although there still exist two types of Hermite polynomials, we will prove that the spherical Hermite polynomials are in this case not orthogonal with respect to the canonically defined inner product. As a consequence, Schr\"odinger operators for e.g. anharmonic oscillators in superspace are not self-adjoint with respect to this inner product, and we also do not immediately have an $O(m)\times Sp(2n)$ invariant Mehler formula for the spherical Hermite polynomials. Recall that the classical Mehler formula (see e.g. \cite{Sz}) is given by
\begin{equation*}
\sum_{k=0}^{\infty} \frac{e^{i k \alpha}}{2^k k! \sqrt{\pi}} H_k(x) H_k(y) = \left(\pi (1- e^{2i \alpha})\right)^{-1/2} \exp{\frac{ 2 e^{i \alpha} x y - e^{2i \alpha}(-x^2 + y^2)}{1- e^{2i \alpha}}}
\end{equation*}
and connects the one-dimensional Hermite polynomials with the kernel of the (fractional) Fourier transform.
Let us thus summarize the main aims of this paper in three questions:
\begin{itemize}
\item \textbf{Q1:} Can we construct a new inner product in full superspace such that the spherical Hermite polynomials are orthogonal?
\item \textbf{Q2:} Can we restore the self-adjointness of a class of Schr\"odinger operators of anharmonic type?
\item \textbf{Q3:} Can we obtain an $O(m)\times Sp(2n)$ invariant Mehler formula?
\end{itemize}
We will provide a positive answer to these 3 questions. Because it is possible to split integration in superspace in radial and spherical parts, the main technical difficulty lies in finding a positive definite inner product on the space of (super) spherical harmonics. This is the subject of the technical lemmas \ref{SSin1}, \ref{SSin2}, \ref{SSin3} and \ref{SSin4}, where use is made of the decomposition of harmonics in superspace under the action of $O(m)\times Sp(2n)$ obtained in \cite{DBE1} and of recent results on integration over the supersphere (see \cite{CDBS1}). These lemmas culminate in theorem \ref{defsuper}, where the new inner product is given.
The paper is organized as follows. In section \ref{classHarmAnalysis} we give a brief review of what is known in both the case of $O(m)$ and $\mathcal{G} < O(m)$ symmetry. We focus on the $\mathfrak{sl}_{2}$ algebra generated by the Laplace operator and the squared length of a vector, introduce the two types of Hermite polynomials and discuss their orthogonality. We show how they give rise to $O(m)$ and $\mathcal{G} < O(m)$ invariant Mehler formulae. In section \ref{Hamaninsup} we give the basic notions on superspaces needed for the sequel. We start with introducing Grassmann algebras, then proceed to full superspaces and discuss the notion of Schr\"odinger equations in superspaces. Next, in section \ref{FermionicMehler} we construct an inner product on the Grassmann algebra and obtain its basic properties. We show that the spherical Hermite functions, related to the symplectic symmetry, form an orthogonal basis of the Grassmann algebra with respect to this inner product. Finally we obtain
a
Mehler formula for the kernel of the purely fermionic Fourier transform. In doing so, we also determine the reproducing kernel for spaces of fermionic harmonics and express them in terms of regularized Gegenbauer polynomials. In section \ref{supersuper} we first discuss in detail where the orthogonality of the spherical Hermite polynomials in a full superspace fails. We then proceed to construct a new inner product and show that this inner product satisfies the desired properties, thus solving questions \textbf{Q1} and \textbf{Q2}. We then discuss how the spherical Hermite polynomials give rise to an $O(m)\times Sp(2n)$ invariant Mehler formula, answering question \textbf{Q3}. In section \ref{conclSummary} we summarize our results. We present in two extensive tables the differences and analogies that exist between the different types of symmetries considered in this paper. This also serves as a list of notations. Finally, we give several directions for further research. In
the appendix we recall some well-known facts about Hermite, Laguerre and Gegenbauer polynomials on the real line that will be used implicitly throughout the paper.
\section{Orthogonality of Hermite polynomials for $O(m)$ and finite reflection groups}
\setcounter{equation}{0}
\label{classHarmAnalysis}
\subsection{Classical harmonic analysis}
\label{ClassHO}
Harmonic analysis in $\mathbb{R}^{m}$ is governed by the following three operators
\begin{eqnarray*}
\nabla^{2}_b &=& \sum_{i=1}^{m}\partial_{x_{i}}^{2}\\
r^{2} &=& \sum_{i=1}^{m}x_{i}^{2}\\
\mathbb{E}_{b} &=& \sum_{i=1}^{m} x_{i} \partial_{x_{i}}
\end{eqnarray*}
with $\Delta_{b}$ the Laplace operator and $\mathbb{E}_{b}$ the Euler operator. The subindex $b$ denotes that we are working with bosonic or commuting co-ordinates. The operators $E = r^{2}/2$, $F =-\nabla^2_{b}/2$ and $H =\mathbb{E}_{b} + m/2$ are invariant under $O(m)$ and generate the Lie algebra $\mathfrak{sl}_{2}$ (see e.g. \cite{MR1151617}):
\begin{equation}
\label{sl2relclass}
\big[H,E\big] = 2E,\>\> \big[H,F\big] = -2F,\>\>\big[E,F\big] = H.
\end{equation}
The space of polynomials in $\mathbb{R}^{m}$ is given by $Pol = \mathbb{R}[x_{1}, \ldots, x_{m}]$ and the space of homogeneous polynomials of degree $k$ by $Pol_{k}$. We then define the space $\mathcal{H}_{k}^{b}$ of spherical harmonics of degree $k$ by $\mathcal{H}_{k}^{b} = \ker{\nabla^2_{b}} \cap Pol_{k}$.
The Schr\"odinger equation of the harmonic oscillator is given by the following partial differential equation
\begin{equation}
-\frac{\nabla^2_{b}}{2} \psi + \frac{r^{2}}{2} \psi = E \psi
\end{equation}
and has two complete sets of solutions. Using cartesian co-ordinates one obtains
\begin{equation}
\psi_{k_{1}, \ldots, k_{m}}^{b} = \frac{1}{\sqrt{2^{\sum_{i=1}^{m} k_{i}} k_{1}! \ldots k_{m}! \pi^{m/2}}} H_{k_{1}}(x_{1}) \ldots H_{k_{m}}(x_{m}) e^{-r^{2}/2}
\label{cartbasisorthogonal}
\end{equation}
with $k_{i} \in \mathbb{N}$ and with $H_{k}(x)$ the one-dimensional Hermite polynomials. The energy associated to $\psi_{k_{1}, \ldots, k_{m}}^{b}$ is given by $E= \frac{m}{2} + \sum_{i=1}^{m} k_{i}$.
If one uses spherical co-ordinates, the eigenfunctions are expressed in terms of spherical harmonics, yielding the so-called spherical Hermite functions (see e.g. \cite{MR926831})
\begin{equation}
\phi_{j,k,l}^{b} = \frac{1}{\sqrt{\frac{1}{2}4^{2j}j!\Gamma (j+\frac{m}{2}+k)}} \left[ (-\nabla^2_{b} - 4 r^2 + 4\mathbb{E}_{b} + 2m)^{j} H_{k}^{(l)} \right] e^{-r^{2}/2},
\label{CHbosbasis}
\end{equation}
where the associated energy is $E= \frac{m}{2} + (2j+ k)$ and with $j, k \in \mathbb{N}$. In (\ref{CHbosbasis}), $\{ H_k^{(l)} \}$ ($l \in 1, \ldots, \dim \mathcal{H}_{k}^{b}$) denotes a (real) orthonormal basis of $\mathcal{H}_{k}^{b}$, i.e.
\begin{equation}
\label{bosharmbasis}
\int_{\mathbb{S}^{m-1}} H_k^{(l_{1})}(\xi) \overline{H_k^{(l_{2})}}(\xi) d \sigma(\xi) = \delta_{l_{1} l_{2}},
\end{equation}
with $d\sigma$ the unique $O(m)$-invariant measure on $\mathbb{S}^{m-1}$.
These functions can be written more explicitly as
\begin{eqnarray*}
\phi_{j,k,l}^{b}&=& \sqrt{\frac{2 j!}{\Gamma (j+\frac{m}{2}+k)}} L_{j}^{\frac{m}{2} + k-1}(r^2) H_k^{(l)}e^{-r^{2}/2}
\end{eqnarray*}
with $L_{j}^{\frac{m}{2} + k-1}$ the generalized Laguerre polynomials.
Both the sets $\{\psi_{k_{1}, \ldots, k_{m}}^{b}\}$ and $\{ \phi_{j,k,l}^{b} \}$ are orthonormal bases of $L_{2}(\mathbb{R}^{m}, dV(\underline{x}))$:
\begin{eqnarray*}
\langle \psi_{k_{1}, \ldots, k_{m}}^{b} , \psi_{l_{1}, \ldots, l_{m}}^{b} \rangle_{L_{2}} = \int_{\mathbb{R}^{m}} \psi_{k_{1}, \ldots, k_{m}}^{b} \overline{\psi_{l_{1}, \ldots, l_{m}}^{b}} dV(\underline{x}) &=& \delta_{k_{1} l_{1}} \ldots \delta_{k_{m} l_{m}}\\
\langle \phi_{j_{1},k_{1},l_{1}}^{b} , \phi_{j_{2},k_{2},l_{2}}^{b} \rangle_{L_{2}}=\int_{\mathbb{R}^{m}} \phi_{j_{1},k_{1},l_{1}}^{b} \overline{\phi_{j_{2},k_{2},l_{2}}^{b}} dV(\underline{x}) &=& \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}},
\end{eqnarray*}
with $dV(\underline{x})$ the Lebesgue measure in $\mathbb{R}^{m}$.
Recall that the classical Fourier transform is given by
\begin{equation}
\mathcal{F}^{-}_{m|0}(f) = (2 \pi)^{-\frac{m}{2}} \int_{\mathbb{R}^{m}} e^{-i\langle \underline{x},\underline{y}\rangle} f(\underline{x}) dV(\underline{x}), \quad \langle \underline{x},\underline{y}\rangle = \sum_{i=1}^{m} x_{i}y_{i}
\label{classFT}
\end{equation}
or in exponential operator notation by
\[
\mathcal{F}^{-}_{m|0} = e^{ \frac{i \pi m}{4}} e^{\frac{i \pi}{4}(\nabla^2_{b} - r^{2})}.
\]
Again, both sets $\{\psi_{k_{1}, \ldots, k_{m}}^{b}\}$ and $\{ \phi_{j,k,l}^{b} \}$ act as eigenfunction bases for the Fourier transform
\begin{eqnarray*}
\mathcal{F}^{-}_{m|0}(\psi_{k_{1}, \ldots, k_{m}}^{b}) &=& (-i)^{\sum_{i=1}^{m}k_{i}}\psi_{k_{1}, \ldots, k_{m}}^{b}\\
\mathcal{F}^{-}_{m|0}(\phi_{j,k,l}^{b}) &=&(-i)^{2j+k}\phi_{j,k,l}^{b}.
\end{eqnarray*}
The Mehler formula for the one-dimensional Hermite polynomials is given by
\begin{equation*}
\sum_{k=0}^{\infty} \frac{e^{i k \alpha}}{2^k k! \sqrt{\pi}} H_k(x) H_k(y) = \left(\pi (1- e^{2i \alpha})\right)^{-1/2} \exp{\frac{ 2 e^{i \alpha} x y - e^{2i \alpha}(x^2 + y^2)}{1- e^{2i \alpha}}}.
\end{equation*}
If $\alpha = - \pi/2$ this formula gives a decomposition of the kernel of the one-dimensional Fourier transform. For a discussion of this formula, we refer the reader to \cite{Sz} or \cite{W}. A nice combinatorial proof can be found in \cite{MR0498167}.
There exist several generalizations of this formula. We will discuss the $O(m)$- and $\mathcal{G}$-invariant cases in this and the next section. Note that there also exists a Mehler formula for the $q$-Hermite polynomials, see e.g. \cite{MR578207}.
The one-dimensional Mehler formula has an important property that can be deduced from the proof of the Mehler formula in \cite{W}.
\begin{lemma}
\label{Mehler1}
For any polynomial $D$ in four variables, one has
\begin{eqnarray*}
&&\sum_{k=0}^{\infty}D(x,y,\partial_x,\partial_y) \frac{e^{i k \alpha}}{2^k k! \sqrt{\pi}} H_k(x) H_k(y)\\ &=& \left(\pi (1- e^{2i \alpha})\right)^{-1/2} D(x,y,\partial_x,\partial_y)\exp{\frac{ 2 e^{i \alpha} x y - e^{2i \alpha}(x^2 + y^2)}{1- e^{2i \alpha}}}.
\end{eqnarray*}
The Hermite polynomials are defined as above and the series is absolutely convergent for $x,y,\alpha\in\mathbb{R}$.
\end{lemma}
One can also construct $m$-dimensional Mehler formulae. Multiplying $m$ copies of the one-dimensional Mehler formula yields
\begin{eqnarray*}
&&\sum_{k_{1}, \ldots, k_{m}} e^{i \alpha \sum_{i=1}^{m} k_{i}}\psi_{k_{1}, \ldots, k_{m}}^{b}(\underline{x}) \psi_{k_{1}, \ldots, k_{m}}^{b}(\underline{y})\\ &=& \left(\pi (1- e^{2i \alpha})\right)^{-\frac{m}{2}} e^{\frac{ 4 e^{i \alpha}\langle \underline{x},\underline{y} \rangle - (1 +e^{2i \alpha})(r^2 + r_{\underline{y}}^2)}{2- 2e^{2i \alpha}}}, \qquad r_{\underline{y}}^2 = \sum_{i=1}^{m} y_{i}^{2}.
\end{eqnarray*}
As this series is absolutely convergent, we can rearrange terms in the left-hand side. Using the fact that the change of basis from $\{\psi_{k_{1}, \ldots, k_{m}}^{b}\}$ to $\{ \phi_{j,k,l}^{b} \}$ is orthogonal in each eigenspace of the harmonic oscillator we obtain
\[
\sum_{j,k,l} e^{i \alpha (2j+k)}\phi_{j,k,l}^{b}(\underline{x}) \phi_{j,k,l}^{b}(\underline{y}) = \left(\pi (1- e^{2i \alpha})\right)^{-\frac{m}{2}} e^{\frac{ 4 e^{i \alpha}\langle \underline{x},\underline{y} \rangle - (1 +e^{2i \alpha})(r^2 + r_{\underline{y}}^2)}{2- 2e^{2i \alpha}}}.
\]
This formula can be simplified even further by making use of the reproducing kernel of the spaces of spherical harmonics, given by (see e.g. \cite{MR0499342})
\begin{align}
\label{bosreprkern}
\begin{split}
F_k(\underline{x},\underline{y}) &=\sum_{l=1}^{\dim \mathcal{H}_{k}^{b}}H_{k}^{(l)}(\underline{x}) H_{k}^{(l)}(\underline{y})\\ &= \frac{2k+m-2}{m-2} \frac{\Gamma(m/2)}{2 \pi^{m/2}} (|\underline{x}||\underline{y}|)^k C^{(m-2)/2}_k (\langle \frac{\underline{x}}{|\underline{x}|}, \frac{\underline{y}}{|\underline{y}|}\rangle)
\end{split}
\end{align}
with $C^{(m-2)/2}_k$ the Gegenbauer polynomial of degree $k$ (see Appendix) and where we have homogenized the right-hand side. This yields the $O(m)$-invariant version of the Mehler formula for every $\underline{x},\underline{y}\in\mathbb{R}^m$ and $\alpha\in\mathbb{R}$:
\begin{align}
\label{OmMehler2}
\begin{split}
&\sum_{j,k}\frac{2j! e^{i \alpha (2j+k)}}{\Gamma(j+\frac{m}{2}+k)} L_j^{\frac{m}{2}+k-1}(r^2)L_j^{\frac{m}{2}+k-1}(r_{\underline{y}}^2)F_k(\underline{x},\underline{y}) \\
=& \left(\pi (1- e^{2i \alpha})\right)^{-\frac{m}{2}} e^{\frac{ 2 e^{i \alpha}\langle \underline{x},\underline{y} \rangle - e^{2i \alpha}(r^2 + r_{\underline{y}}^2)}{1- e^{2i \alpha}}}.
\end{split}
\end{align}
In this formula only $r^2$, $r_{\underline{y}}^2$ and $\langle \underline{x},\underline{y} \rangle$ appear, which are three arbitrary real numbers satisfying $r^2\ge 0$, $r_{\underline{y}}^2\ge 0$ and $\langle \underline{x},\underline{y} \rangle^2\le r^2r_{\underline{y}}^2$. This leads to the following corollary.
\begin{corollary}
For every $a,b\in\mathbb{R}^+$ and $c\in\mathbb{R}$ with $c^2\le a^2b^2$, one has
\begin{eqnarray*}
&&\sum_{j,k}\frac{2j! e^{i \alpha (2j+k)}}{\Gamma(j+\frac{m}{2}+k)} L_j^{\frac{m}{2}+k-1}(b^2)L_j^{\frac{m}{2}+k-1}(a^2)(ab)^k G_k\left(\frac{c}{ab}\right)\\
&=& \left(\pi (1- e^{2i \alpha})\right)^{-m/2} \exp{\frac{ 2 e^{i \alpha} c - e^{2i \alpha}(a^2 +b^2)}{1- e^{2i \alpha}}}
\end{eqnarray*}
where
\[
G_k\left(\frac{c}{ab}\right) = \frac{2k+m-2}{m-2} \frac{\Gamma(m/2)}{2 \pi^{m/2}} \, C^{(m-2)/2}_k \left(\frac{c}{ab}\right).
\]
\label{OmMehler3}
\end{corollary}
In this way, the Mehler formula has been stripped of its geometrical meaning in $\mathbb{R}^{m}$ and is reduced to a statement about orthogonal polynomials on the real line. We will need this in the proof of the $O(m)\times Sp(2n)$-invariant Mehler formula.
\begin{remark}
\label{Mehlerafleiden}
Lemma \ref{Mehler1} is immediately adapted to the $O(m)$-invariant case in equation (\ref{OmMehler2}) or corollary \ref{OmMehler3}.
\end{remark}
\subsection{Hermite polynomials related to the Dunkl Laplacian}
\label{DunklHermite}
Denote by $\langle .,. \rangle$ the standard Euclidean scalar product in $\mathbb{R}^{m}$ and by $r = \langle \underline{x}, \underline{x}\rangle^{1/2}$ the associated norm. For $\underline{\alpha} \in \mathbb{R}^{m} - \{ 0\}$, the reflection $r_{\underline{\alpha}}$ in the hyperplane orthogonal to $\underline{\alpha}$ is given by
\[
r_{\underline{\alpha}}(\underline{x}) = \underline{x} - 2 \frac{\langle \underline{\alpha}, \underline{x}\rangle}{|\underline{\alpha}|^{2}}\underline{\alpha}, \quad \underline{x} \in \mathbb{R}^{m}.
\]
A root system is a finite subset $R \subset \mathbb{R}^{m}$ of non-zero vectors such that, for every $\underline{\alpha} \in R$, the associated reflection $r_{\underline{\alpha}}$ preserves $R$. We will assume that $R$ is reduced, i.e. $R \cap \mathbb{R} \underline{\alpha} = \{ \pm \underline{\alpha}\}$ for all $\underline{\alpha} \in R$. Each root system can be written as a disjoint union $R = R_{+} \cup (-R_{+})$, where $R_{+}$ and $-R_{+}$ are separated by a hyperplane through the origin. The subgroup $\mathcal{G} \subset O(m)$ generated by the reflections $\{r_{\underline{\alpha}} | \underline{\alpha} \in R\}$ is called the finite reflection group associated with $R$. We will also assume that $R$ is normalized such that $\langle \underline{\alpha}, \underline{\alpha}\rangle = 2$ for all $\underline{\alpha} \in R$. For more information on finite reflection groups we refer the reader to \cite{Humph}.
A multiplicity function $\kappa$ on the root system $R$ is a $\mathcal{G}$-invariant function $\kappa: R \rightarrow \mathbb{C}$, i.e. $\kappa(\underline{\alpha}) = \kappa(h \underline{\alpha})$ for all $h \in \mathcal{G}$. We will denote $\kappa(\underline{\alpha})$ by $\kappa_{\underline{\alpha}}$.
Fixing a positive subsystem $R_{+}$ of the root system $R$ and a multiplicity function $\kappa$, we introduce the Dunkl operators $T_{i}$ associated to $R_{+}$ and $\kappa$ by (see \cite{MR951883}, \cite{MR1827871})
\[
T_{i} f(\underline{x})= \partial_{x_{i}} f(\underline{x}) + \sum_{\underline{\alpha} \in R_{+}} \kappa_{\underline{\alpha}} \alpha_{i} \frac{f(\underline{x}) - f(r_{\underline{\alpha}}(\underline{x}))}{\langle \underline{\alpha}, \underline{x}\rangle}, \qquad f \in C^{1}(\mathbb{R}^{m}).
\]
An important property of the Dunkl operators is that they commute, i.e. $T_{i} T_{j} = T_{j} T_{i}$.
The Dunkl Laplacian is given by $\Delta_{\kappa} = \sum_{i=1}^{m} T_i^2$, or more explicitly by
\[
\Delta_{\kappa} f(\underline{x}) = \nabla^{2} f(\underline{x}) + 2 \sum_{\underline{\alpha} \in R_{+}} \kappa_{\underline{\alpha}} \left( \frac{\langle \nabla f(\underline{x}), \underline{\alpha} \rangle}{\langle \underline{\alpha}, \underline{x} \rangle} - \frac{f(\underline{x}) - f(r_{\underline{\alpha}}(\underline{x}))}{\langle \underline{\alpha}, \underline{x} \rangle^{2}} \right)
\]
with $\nabla$ the gradient operator.
If we let $\Delta_{\kappa}$ act on $r^2$ we find $\Delta_{\kappa} r^2 = 2m + 4 \gamma = 2 \mu$, where $\gamma = \sum_{\underline{\alpha} \in R_+} \kappa_{\underline{\alpha}}$. We refer to $\mu$ as the Dunkl dimension, because most special functions related to $\Delta_{\kappa}$ behave as if one would be working with the classical Laplace operator in a space with dimension $\mu$.
We also denote by $\mathcal{H}_k^{\mathcal{D}}$ the space of Dunkl-harmonics of degree $k$, i.e. $\mathcal{H}_k^{\mathcal{D}} = \ker{\Delta_{\kappa}} \cap Pol_k$. The space of Dunkl-harmonics of degree $k$ has the same dimension as the classical space of spherical harmonics of degree $k$ and a basis can e.g. be constructed using Maxwell's representation (see \cite{Xu}).
It is possible to construct an intertwining operator $V_{\kappa}$ connecting the classical derivatives $\partial_{x_{j}}$ with the Dunkl operators $T_{j}$ such that $T_{j} V_{\kappa} = V_{\kappa} \partial_{x_{j}}$ (see e.g. \cite{MR1273532}). Note that explicit formulae for $V_{\kappa}$ are only known in a few special cases.
The operators
\[ E:= \frac{1}{2}r^2,\>\> F:= -\frac{1}{2}\Delta_{\kappa} \quad\text{and}\>\>
H:= \mathbb{E} +\mu/2
\]
again satisfy the defining relations \eqref{sl2relclass} of the Lie algebra $\mathfrak{sl}_2$, see \cite{He}.
We further define the Fischer inner product on $Pol$ by (see \cite{MR1145585})
\[
[p,q]_{\kappa} = \left(p(T/2) q\right) (0), \qquad p,q \in Pol
\]
where $p(T)$ means substituting the variables $x_{i}$ for Dunkl operators $T_{i}$. Using an analog of a result of MacDonald (see \cite{MR0558859}), this inner product can be rewritten as
\begin{equation}
[p,q]_{\kappa} = d_{\kappa}^{-1} \int_{\mathbb{R}^{m}} \left(e^{-\Delta_{\kappa}/4} p(\underline{x})\right)\left(e^{-\Delta_{\kappa}/4} q(\underline{x})\right) e^{-r^{2}} w_{\kappa}(\underline{x})dV(\underline{x}), \label{FischerMcDo}
\end{equation}
with $w_{\kappa}(\underline{x}) = \prod_{\alpha \in R_{+}} |\langle \underline{\alpha}, \underline{x}\rangle |^{2 \kappa_{\alpha}}$ the weight function corresponding to $\mathcal{G}$ and with $d_{\kappa} =\int_{\mathbb{R}^{m}} e^{-r^{2}} w_{\kappa}(\underline{x})dV(\underline{x})$ (see e.g. \cite{MR2022853}).
Now let $\{p_\nu\,, \nu\in \mathbb{Z}_+^m\}$ be a basis of $Pol$ such that $p_\nu\in \mathcal Pol_{|\nu|}$ (with $|\nu|=\sum_{i=1}^m\nu_i$) and moreover $[p_\nu, p_\mu]_{\kappa} = \delta_{\nu \mu}$. The `cartesian' Hermite polynomials related to $\mathcal{G}$ are defined as follows by R\"osler (see \cite{MR1620515,MR2022853}).
\begin{definition}
\label{DefHermRos}
The generalized Hermite polynomials $\{H_\nu\,, \>\nu\in \mathbb{Z}_+^m\}$
associated with the basis $\{p_\nu\}$ on $\mathbb{R}^m$ are given by
\[
H_\nu(\underline{x}):= 2^{|\nu|}e^{-\Delta_{\kappa}/4}p_\nu(\underline{x}) =
2^{|\nu|}\sum_{n=0}^{\lfloor|\nu|/2\rfloor} \frac{(-1)^n}{4^n n!}\,
\Delta_{\kappa}^n p_\nu(\underline{x}).
\]
Moreover, the generalized Hermite functions on $\mathbb{R}^m$ are then defined by
\[
\psi_\nu^{\mathcal{D}}(\underline{x}):= \frac{1}{2^{|\nu|} \sqrt{d_{\kappa} }}H_\nu(\underline{x}) e^{-r^2/2}, \quad \nu\in \mathbb{Z}_+^m.
\]
\end{definition}
\begin{remark}
Similar exponential formulas for Hermite polynomials have been given earlier in \cite{MR1471336} by Baker and Forrester, following (unpublished) work of Lassalle. The Hermite polynomials given in Definition \ref{DefHermRos} yield the polynomials of \cite{MR1471336} after symmetrization.
When $\mathcal{G}$ is the group $S_{m}$ or the group $B_{m}$, the polynomials in Definition \ref{DefHermRos} include, under a suitable choice of basis $\{p_\nu\}$, the polynomials studied in \cite{MR1456121, MR1646546} and previously introduced in symmetric version in \cite{MR1105634, MR1133488}.
\end{remark}
The Rodrigues formula for $H_{\nu}(\underline{x})$ is given by
\[
H_{\nu}(\underline{x}) = (-1)^{|\nu|} e^{r^2}p_{\nu}(\underline{T})e^{-r^2},
\]
so if we consider $\kappa=0$ and take the standard monomial basis of $Pol$ we reobtain the cartesian Hermite polynomials for the orthogonal case.
Using formula (\ref{FischerMcDo}) we immediately obtain that the set $\{\psi_\nu^{\mathcal{D}} \}$ forms an orthonormal basis of the weighted $L_{2}$-space $L_{2}(\mathbb{R}^{m}, w_{\kappa}(\underline{x}) dV(\underline{x}))$ :
\[
\int_{\mathbb{R}^{m}} \psi_{\nu}^{\mathcal{D}} \overline{\psi_{\mu}^{\mathcal{D}}} w_{\kappa}(\underline{x}) dV(\underline{x}) = \delta_{\nu \mu}.
\]
The related quantum system in this case is given by the following PDE with difference terms
\begin{equation}
-\frac{\Delta_{\kappa}}{2} \psi + \frac{r^{2}}{2} \psi = E \psi,
\label{HOdunkl}
\end{equation}
which is a so-called Calogero-Moser-Sutherland system (with harmonic confinement). For a review of this type of quantum systems we refer the reader to \cite{vD,MR2022853}.
If $p_\nu$ is a homogeneous polynomial of degree $|\nu|$, then the associated Hermite function $\psi_{\nu}^{\mathcal{D}}$ is a solution of (\ref{HOdunkl}) with corresponding energy $E = \mu/2 + |\nu|$ (see \cite{MR1620515}).
It is also possible to introduce a generalization of the spherical Hermite functions related to the Dunkl Laplacian (see a.o. \cite{MR1199124,Said} or \cite{DBSIGMA} for an approach using Clifford analysis). They are defined as follows
\begin{equation}
\phi_{j,k,l}^{\mathcal{D}} = \frac{1}{\sqrt{\frac{1}{2}4^{2j}j!\Gamma (j+\frac{\mu}{2}+k)}} \left[ (-\Delta_{\kappa} - 4 r^2 + 4\mathbb{E}_{b} + 2\mu)^{j} H_{k}^{(l)} \right] e^{-r^{2}/2},
\end{equation}
with associated energy $E= \frac{\mu}{2} + (2j+ k)$ and where $j, k \in \mathbb{N}$. In this notation, $\{ H_k^{(l)} \}$, $l \in 1, \ldots, \dim \mathcal{H}_{k}^{\mathcal{D}}$, is a (real) orthonormal basis of $\mathcal{H}_{k}^{\mathcal{D}}$ satisfying
\[
\int_{\mathbb{S}^{m-1}} H_k^{(l_{1})}(\xi) \overline{H_k^{(l_{2})}}(\xi) w_{\kappa}(\xi) d \sigma(\xi) = \delta_{l_{1} l_{2}}.
\]
More explicitly, they are given by
\begin{eqnarray*}
\phi_{j,k,l}^{\mathcal{D}}&=& \sqrt{\frac{2 j!}{\Gamma (j+\frac{\mu}{2}+k)}} L_{j}^{\frac{\mu}{2} + k-1}(r^2) H_k^{(l)}e^{-r^{2}/2}.
\end{eqnarray*}
The set of functions $\{ \phi_{j,k,l}^{\mathcal{D}} \}$ again forms an orthonormal basis of the space $L_{2}(\mathbb{R}^{m}, w_{\kappa}(\underline{x}) dV(\underline{x}))$, i.e.
\[
\int_{\mathbb{R}^{m}} \phi_{j_{1},k_{1},l_{1}}^{\mathcal{D}} \overline{\phi_{j_{2},k_{2},l_{2}}^{\mathcal{D}}} w_{\kappa}(\underline{x}) dV(\underline{x}) = \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}.
\]
There exists a Fourier transform related to the Dunkl Laplacian (see \cite{MR1199124,deJ}). This so-called Dunkl transform $\mathcal \mathcal{F}_{\kappa}: L^1(\mathbb{R}^m, w_{\kappa}(\underline{x})dx)\to C(\mathbb{R}^m)$ is defined as follows
\[
\mathcal \mathcal{F}_{\kappa} f(\underline{y}):= c_{\kappa}^{-1} \int_{\mathbb{R}^m} f(\underline{x})\,D(\underline{x},-i\underline{y})\,w_{\kappa}(\underline{x}) dV(\underline{x}) \quad
(\underline{y} \in \mathbb{R}^m)
\]
with $c_{\kappa} = \int_{\mathbb{R}^m} e^{-r^{2}/2}w_{\kappa}(\underline{x}) dV(\underline{x})$ the Mehta constant related to $\mathcal{G}$ and where $D(\underline{x},\underline{y})$ is the Dunkl kernel. This kernel is the unique solution of the system
\[
T_{i, x} D(\underline{x},\underline{y}) = y_{i} D(\underline{x},\underline{y}), \quad i=1, \ldots, m
\]
which is real-analytic in $\mathbb{R}^{m}$ and satisfies $K(0,\underline{y})=1$. For general reflection groups this kernel is not explicitly known. There also exists an operator exponential expression for the Dunkl transform
\begin{equation*}
\mathcal{F}_{\kappa} = e^{ \frac{i \pi \mu}{4}} e^{\frac{i \pi}{4}(\Delta_{\kappa} - r^{2})}
\end{equation*}
as has been studied in-depth in \cite{Said}. Note that both bases $\{\psi_\nu^{\mathcal{D}} \}$ and $\{ \phi_{j,k,l}^{\mathcal{D}} \}$ are eigenfunctions of the Dunkl transform, i.e.
\begin{eqnarray*}
\mathcal{F}_{\kappa} (\psi_\nu^{\mathcal{D}}) &=& (-i)^{|\nu|} \psi_\nu^{\mathcal{D}}\\
\mathcal{F}_{\kappa} (\phi_{j,k,l}^{\mathcal{D}}) &=& (-i)^{2j + k} \phi_{j,k,l}^{\mathcal{D}}.
\end{eqnarray*}
For the first basis this was obtained in \cite{MR1620515}, for the second basis in \cite{MR1199124,Said}.
A Mehler formula for the `cartesian' Hermite polynomials related to $\mathcal{G}$ was obtained in \cite{MR1620515}, Theorem 3.12. It states that, for $u\in \mathbb{C}$ with $|u|<1$ and all $\underline{x}, \underline{y} \in \mathbb{R}^{m}$, the following holds
\[
\sum_{\nu\in \mathbb{Z}_+^m} d_{\kappa} u^{|\nu|} \psi_\nu^{\mathcal{D}}(\underline{x}) \psi_\nu^{\mathcal{D}}(\underline{y}) =(1-u^{2})^{-\mu/2} D\left( \frac{2u\underline{x}}{1-u^{2}},\underline{y}\right)e^{-\frac{1+u^{2}}{2-2u^{2}}(r^{2}+r_{\underline{y}}^{2})}.
\]
Going over to the basis $\{\phi_{j,k,l}^{\mathcal{D}}\}$ we then immediately obtain
\[
\sum_{j,k,l} d_{\kappa} u^{2j+k} \phi_{j,k,l}^{\mathcal{D}}(\underline{x}) \phi_{j,k,l}^{\mathcal{D}}(\underline{y}) =(1-u^{2})^{-\mu/2} D\left( \frac{2u\underline{x}}{1-u^{2}},\underline{y}\right)e^{-\frac{1+u^{2}}{2-2u^{2}}(r^{2}+r_{\underline{y}}^{2})}.
\]
Note that when $\kappa=0$, then $d_{\kappa}= \pi^{m/2}$ and the formulae reduce to the formulae in section \ref{ClassHO}. The last formula can be simplified by using the explicit expression of $\{\phi_{j,k,l}^{\mathcal{D}}\}$ in terms of generalized Laguerre polynomials and Dunkl harmonics. Then one can apply the reproducing kernel for $\mathcal{H}_{k}^{\mathcal{D}}$, which is again given by a suitable Gegenbauer polynomial, see \cite{MR1402890}, as
\begin{eqnarray*}
\sum_{l=1}^{\dim \mathcal{H}_{k}^{\mathcal{D}}}H_{k}^{(l)}(\underline{x}/|\underline{x}|) H_{k}^{(l)}(\underline{y}/|\underline{y}|) &=& \frac{2k+\mu-2}{\mu-2} \frac{\Gamma(\mu/2)}{2^{(1-\mu)/2} c_{\kappa}}\\
&&\times [V_{\kappa} C^{(\mu-2)/2}_k \left(\langle \cdot , \underline{y}/|\underline{y}|\rangle \right)](\underline{x}/ |\underline{x}|),
\end{eqnarray*}
with $V_{\kappa}$ the intertwining operator. Homogenizing this reproducing kernel then leads to a similar formula as (\ref{OmMehler2}), which is now $\mathcal{G}$-invariant.
\begin{remark}
In \cite{MR2134314,MR2401813} and the recent preprint \cite{Orsted2}, the role played by Mehler type formulae for orthogonal and finite reflection group symmetries is further elucidated, for radially deformed operators still satisfying the defining relations of $\mathfrak{sl}_{2}$.
\end{remark}
\section{Harmonic analysis in superspace}
\setcounter{equation}{0}
\label{Hamaninsup}
\subsection{Grassmann algebras}
Consider the complex Grassmann algebra $\Lambda_{2n}$ generated by $2n$ anti-commuting generators ${x \grave{}}_i$ satisfying ${x \grave{}}_j {x \grave{}}_k = - {x \grave{}}_k {x \grave{}}_j$. An arbitrary element $f \in \Lambda_{2n}$ can always be written as $f = \sum_A f_A {x \grave{}}_A$ with ${x \grave{}}_A = {x \grave{}}_1^{\, \alpha_1} \ldots {x \grave{}}_{2n}^{\, \alpha_{2n}}$, $A = (\alpha_{1}, \ldots, \alpha_{2n}) \in \{0,1\}^{2n}$ and $f_A \in \mathbb{C}$. The dimension of $\Lambda_{2n}$ as a $\mathbb{C}$-vectorspace is hence $2^{2n}$. $\Lambda_{2n}$ decomposes as
\[
\Lambda_{2n} = \oplus_{k=0}^{2n} \Lambda_{2n}^{k}
\]
with $\Lambda_{2n}^{k}$ the space of homogeneous elements of degree $k$.
Introducing $\theta^2 = -\sum_{j=1}^n {x\grave{}}_{2j-1} {x\grave{}}_{2j}$ and the so-called fermionic Laplace operator $\nabla^2_f =- 4 \sum_{j=1}^n \partial_{{x \grave{}}_{2j-1}} \partial_{{x \grave{}}_{2j}}$, we have the following decomposition of $\Lambda_{2n}$ (see \cite{DBE1}):
\begin{equation}
\Lambda_{2n} = \bigoplus_{k=0}^{n} \left(\bigoplus_{j=0}^{n-k} \theta^{2j} \mathcal{H}^f_k \right),
\label{Fischer}
\end{equation}
where $\mathcal{H}^f_k$ is the space of fermionic spherical harmonics of degree $k$, i.e. $\mathcal{H}_{k}^{f} = \ker \nabla^2_f \cap \Lambda_{2n}^{k}$. The dimension of $\mathcal{H}^f_{k}$ is given by $\binom{2n}{k} - \binom{2n}{k-2}$. Formula (\ref{Fischer}) is the decomposition of $\Lambda_{{2n}}$ into irreducible pieces under the action of $Sp(2n)$ (see \cite{DBE1}).
The fermionic Euler operator is given by $\mathbb{E}_f = \sum_{j=1}^{2n} {x \grave{}}_{j} \partial_{{x \grave{}}_{j}}$ and satisfies the equation
\begin{eqnarray}
\label{fermEuler}
[\nabla^2_f,\theta^2]&=&4\mathbb{E}_f-4n.
\end{eqnarray}
We also have $[\nabla^2_{f},\mathbb{E}_f-n ] = 2 \nabla^2_{f}$ and $[\theta^{2},\mathbb{E}_f-n] = -2 \theta^{2}$ so the $\mathfrak{sl}_{2}$ relations from both the orthogonal and the Dunkl case extend to the Grassmann case. It is moreover easily seen that $\theta^2$, $-\nabla^2_f$ and $\mathbb{E}_{f}-n$ are invariant under the action of $Sp(2n)$. So we have obtained a change of symmetry from either $O(m)$ or $\mathcal{G} < O(m)$ to $Sp(2n)$.
\begin{remark}
In previous papers the notations $\Delta_f=-\nabla^2_f$ and $\underline{x} \grave{}^2=-\theta^2$ were used.
\end{remark}
We have the following important lemma concerning the action of the Laplace operator (see \cite{DBS5}, Corollary 2).
\begin{lemma}
\label{Laplfischer}
Let $H_k \in \mathcal{H}^f_k$. Then for $j \leq n-k $ one has
\[
\nabla^{2i}_f(\theta^{2j} H_k) = \left\{
\begin{array}{l}
c_{i,j,k} \theta^{2j-2i} H_k, \quad i \leq j\\
\vspace{-2mm}\\
0, \quad i > j
\end{array}
\right.
\]
with
\[
c_{i,j,k} = (-1)^i 4^{i} \frac{j!}{(j-i)!} \frac{(n+i-j-k)!}{(n-j-k)!}.
\]
\end{lemma}
The fermionic Gaussian function is given by the finite Taylor expansion $\exp(-\theta^2/2) =\sum_{j=0}^n(-1)^j \theta^{2j}/(2^{j}j!)$.
We consider a basis $H_k^{(l)}$ of $\mathcal{H}^f_k$. The functions
\begin{align}
\label{CliffordHermiteFunctions}
\begin{split}
\varphi_{j,k,l}^{f} =& (-\nabla^2_{f} -\theta^2 + 2\mathbb{E} -2n)^j H_k^{(l)} \exp(-\theta^2/2)\\
=&2^{2j}j! L_j^{-n+k-1}(\theta^2) H_k^{(l)}\exp(-\theta^2/2)
\end{split}
\end{align}
for $j = 0, \ldots, n-k$; $k = 0, \ldots, n$ and $l = 1, \ldots, \dim \mathcal{H}_k^{f}$ and with $L_\beta^{\alpha}$ the generalized Laguerre polynomials, are the so-called fermionic spherical Hermite functions (see \cite{DBS3}). They constitute a basis of $\Lambda_{2n}$, which follows immediately from the decomposition (\ref{Fischer}).
Note that it is possible to explicitly construct a basis of $\mathcal{H}_k^f$ by decomposing this space under the action of the subgroup $Sp(2) \times Sp(2n-2)$ of $Sp(2n)$. This leads to the following theorem.
\begin{theorem}
\label{basisrecursief}
If $1<k\leq n$, then the space $\mathcal{H}_k^f({x \grave{}}_1 , \ldots, {x \grave{}}_{2n})$ decomposes as
\begin{eqnarray*}
&& \mathcal{H}_k^f({x \grave{}}_3 , \ldots, {x \grave{}}_{2n}) \; \oplus \; \mathcal{H}_1^f({x \grave{}}_1 ,{x \grave{}}_2)\otimes \mathcal{H}_{k-1}^f({x \grave{}}_3 , \ldots, {x \grave{}}_{2n})\\
&& \oplus \; \left[{x \grave{}}_1 {x \grave{}}_2
+ \frac{1}{k-n-1} ({x \grave{}}_3 {x \grave{}}_4 + \ldots +{x \grave{}}_{2n-1} {x \grave{}}_{2n}) \right] \mathcal{H}_{k-2}^f({x \grave{}}_3 , \ldots, {x \grave{}}_{2n}).
\end{eqnarray*}
\end{theorem}
The integration we use on $\Lambda_{2n}$ is the so-called Berezin integral (see \cite{MR732126}, \cite{DBS5}), defined by
\[
\int_{B,x} = \pi^{-n} \partial_{{x \grave{}}_{2n}} \ldots \partial_{{x \grave{}}_{1}} = \frac{ \pi^{-n}}{4^n n!} \nabla_f^{2n}.
\]
\subsection{Full superspace}
Next we consider a general superspace, with $m$ commuting and $2n$ anti-commuting variables. The space of polynomials is then denoted by $\mathcal{P}=\mathbb{R}[x_1,\ldots,x_m]\otimes \Lambda_{2n}$. The full Laplace operator is given by
\[
\nabla^2=\nabla^2_b+\nabla^2_f = \sum_{i=1}^m\partial_{x_i}^2 -4 \sum_{j=1}^n \partial_{{x \grave{}}_{2j-1}} \partial_{{x \grave{}}_{2j}}
\]
and the generalization of $r^{2}$ is
\[
R^2=r^2+\theta^2 = \sum_{i=1}^m x_i^2 - \sum_{j=1}^n {x\grave{}}_{2j-1} {x\grave{}}_{2j}.
\]
\begin{remark}
In previous papers the notations $\bold{x}^2=-R^2$ and $\Delta=-\nabla^2$ were used.
\end{remark}
The super Euler operator $\mathbb{E}$ is given by $\mathbb{E} = \mathbb{E}_{b}+ \mathbb{E}_{f}$. We can introduce a dimension parameter, the so-called super-dimension, by
\[
M=m-2n=\frac{1}{2}\nabla^2(R^2).
\]
The operators $E = R^{2}/2$, $F = -\nabla^2/2$ and $H =\mathbb{E} + M/2$ again satisfy the defining commutation relations of $\mathfrak{sl}_{2}$ (see formula (\ref{sl2relclass})), although the symmetry of the operators is now given by $O(m)\times Sp(2n)$, see \cite{DBE1}. In particular
\begin{equation}
\label{commsuper}
\left[\nabla^2/2,R^2/2\right]=\mathbb{E}+M/2
\end{equation}
holds. The symmetry group is realized as the matrices $A\in\mathbb{R}^{(m+2n)\times(m+2n)}$ which satisfy $A^TGA=G$ with
\begin{eqnarray}
\label{metricG}
G&=&\left( \begin{array}{c|c} \mathbb{I}_m&0\\ \hline \vspace{-3.5mm} \\0&J
\end{array}\right)
\end{eqnarray}
and
\[
J = \left( \begin{array}{cccccc} 0&1&&&\\-1&0&&&\\&&\ddots&&\\&&&0&1\\&&&-1&0
\end{array}
\right).
\]
The action of $O(m)\times Sp(2n)$ on superfunctions is then given by
\begin{eqnarray}
\label{actieOSp}
Af(\bold{x})&=&f(A^T\bold{x}).
\end{eqnarray}
The space $\mathcal{H}_{k}$ of (super) spherical harmonics of degree $k$ is defined by $\mathcal{H}_{k} = \ker \nabla^2 \cap \mathcal{P}_{k}$, with $\mathcal{P}_{k}$ the space of homogeneous polynomials of degree $k$. If the super-dimension $M=m-2n$ is not even and negative, we have the following decomposition of $\mathcal{P}$
\begin{equation}
\label{superFischer}
\mathcal{P} = \bigoplus_{k=0}^{\infty} \mathcal{P}_k= \bigoplus_{j=0}^{\infty} \bigoplus_{k=0}^{\infty} R^{2j}\mathcal{H}_k.
\end{equation}
This decomposition is closely related to the Howe dual pair $(\mathfrak{osp}(m|2n),\mathfrak{sl}_2)$. Each subspace $\bigoplus_{j=0}^{\infty} R^{2j}\mathcal{H}_k$ is an irreducible $\mathfrak{sl}_2$-representation, the weight vectors are $R^{2j}\mathcal{H}_k$. The blocks $R^{2j}\mathcal{H}_k$ are exactly the irreducible pieces of the representation of $\mathfrak{osp}(m|2n)$ on $\mathcal{P}$ when $M>0$, see \cite{ZHANG}.
There exist explicit projection operators for the decompositions (\ref{Fischer}) and (\ref{superFischer}), satisfying
\[
\mathbb{P}_i^k (R^{2j} \mathcal{H}_{k-2j}) = \delta_{ij} \mathcal{H}_{k-2j}.
\]
They were constructed in \cite{DBE1} (see also \cite{DBThesis}). More specifically, $\mathbb{P}_0^k$ is given by
\begin{eqnarray}
\label{projectie}
\mathbb{P}_0^k &=& \sum_{j=0}^{\lfloor k/2 \rfloor} \frac{1}{4^{j} j!} \frac{1}{(k+M/2-j-1)_j} R^{2j}\nabla^{2j},
\end{eqnarray}
with $(a)_j=a(a+1)\cdots(a+j-1)$ the Pochhammer symbol.
We can again introduce the spherical Hermite functions, and they will now form a basis of $\mathcal{P} \exp (-r^2/2)$. Because $\exp(-\theta^2/2)$ is an invertible element of $\Lambda_{2n}$, we have
\[\mathcal{P} \exp (-r^2/2)=\mathcal{P}\exp(-r^2/2)\exp(-\theta^2/2)=\mathcal{P} \exp (-R^2/2).\]
If we now consider a basis $H_k^{(l)}$ of $\mathcal{H}_k$, then the functions
\begin{eqnarray*}
\varphi_{j,k,l}& =& (-\nabla^2 -R^2 + 2\mathbb{E} +M)^j H_k^{(l)} \exp{(-R^2/2)}
\label{CliffordHermiteFunctionssuper}
\end{eqnarray*}
with $j, k \in \mathbb{N}$ and $l = 1, \ldots, \dim \mathcal{H}_k$, are the so-called spherical Hermite functions in (full) superspace (see \cite{DBS3}). They constitute a basis of $\mathcal{P} \exp (-R^2/2)$ when $M\not\in-2\mathbb{N}$, which follows immediately from the decomposition (\ref{superFischer}). In terms of classical orthogonal polynomials the functions $\varphi_{j,k,l}$ are given by
\begin{equation}
\varphi_{j,k,l} = 2^{2j} j! L^{\frac{M}{2}+k-1}_{j}(R^2) H_k^{(l)} \exp{(-R^2/2)}
\label{CHLaguerre}
\end{equation}
with $L^{\alpha}_{\beta}$ the generalized Laguerre polynomials (see \cite{DBS3}, theorem 8) when $M\not\in-2\mathbb{N}$ or when $m=0$. In the limit $n\to 0$, the spherical Hermite functions (\ref{CliffordHermiteFunctionssuper}) reduce to the purely bosonic spherical Hermite functions (see \cite{MR926831} or formula (\ref{CHbosbasis})). The spherical Hermite functions satisfy
\begin{equation}
\label{differentiaalvgl}
(\nabla^2+R^2+2\mathbb{E}+M)\varphi_{j,k,l}=-8j(2j+M+2k-2)\varphi_{j-1,k,l}
\end{equation}
for all $(m,n)$.
If we gauge $\nabla^2$ and $(\nabla^2+R^2 + 2\mathbb{E} +M)$ with the generalized Gaussian, we obtain
\begin{equation}
\label{berekexpdelta}
\exp(R^2/2) \nabla^2\exp{(-R^2/2)}= \nabla^2+R^2 -2\mathbb{E} -M
\end{equation}
and
\begin{equation}
\label{CHexp}
\exp(R^2/2)(\nabla^2+R^2 + 2\mathbb{E} +M) \exp{(-R^2/2)}=\nabla^2.
\end{equation}
In the full superspace, lemma \ref{Laplfischer} takes the following form.
\begin{lemma}
Let $H_k \in \mathcal{H}_k$ and $M \not \in -2 \mathbb{N}$. Then for all $i,j,k \in \mathbb{N}$ one has
\[
\nabla^{2i}(R^{2j} H_k) = \left\{
\begin{array}{l}
c_{i,j,k} R^{2j-2i} H_k, \quad i \leq j\\
\vspace{-2mm}\\
0, \quad i > j
\end{array}
\right.
\]
with
\[
c_{i,j,k} = 4^{i} \frac{j!}{(j-i)!} \frac{\Gamma(k+M/2+j)}{\Gamma(k+M/2+j-i)}.
\]
\label{laplonpieces}
\end{lemma}
The integration we use on a full superspace is defined by
\begin{equation}
\label{superint}
\int_{\mathbb{R}^{m | 2n}} = \int_{\mathbb{R}^m} d V(\underline{x}) \int_B=\int_B \int_{\mathbb{R}^m} d V(\underline{x}).
\end{equation}
The Schwartz space in superspace is given by $\mathcal{S}(\mathbb{R}^m)_{m|2n}=\mathcal{S}(\mathbb{R}^m)\otimes\Lambda_{2n}$. In \cite{DBS9}, a Fourier transform in superspace was introduced for $f\in\mathcal{S}(\mathbb{R}^m)_{m|2n}$ by
\[
\mathcal{F}_{m | 2n}^{\pm}(f) = (2 \pi)^{-\frac{M}{2}} \int_{\mathbb{R}^{m | 2n},x} \exp{(\pm i \langle \bold{x},\bold{y} \rangle)}f(\bold{x})
\]
with
\begin{equation}
\label{innerprodvectorssuper}
\langle \bold{x} , \bold{y} \rangle = \sum_{i=1}^{m} x_i y_i -\frac{1}{2} \sum_{j=1}^{n}({x \grave{}}_{2j-1}{y \grave{}}_{2j} - {x \grave{}}_{2j} {y \grave{}}_{2j-1})= \langle\underline{x},\underline{y}\rangle+ \langle \underline{x} \grave{},\underline{y} \grave{}\rangle.
\end{equation}
In this notation, ${x \grave{}}_{i}$ and ${y \grave{}}_{j}$ generate a Grassmann algebra $\Lambda_{4n}$. In particular, ${x \grave{}}_{i}$ and ${y \grave{}}_{j}$ are mutually anti-commutative and $\langle \bold{x},\bold{y} \rangle$ is symmetric.
Note that this Fourier transform can also be defined as the operator exponential
\begin{equation}
\label{FTSuperExp}
\mathcal{F}_{m | 2n}^{\pm} =e^{\mp \frac{ i \pi M}{4} } e^{\pm \frac{i \pi}{4}(-\nabla^2 +R^2)}.
\end{equation}
The super Fourier transform satisfies similar properties as the classical bosonic Fourier transform defined in (\ref{classFT}). The inverse of $\mathcal{F}^+$ is given by $\mathcal{F}^-$, i.e.
\[
\mathcal{F}_{m | 2n}^{\pm} \circ \mathcal{F}_{m | 2n}^{\mp} = \mbox{id}_{\, \mathcal{S}(\mathbb{R}^m)_{m|2n}}.
\]
The Fourier transform of the spherical Hermite functions is
\[
\mathcal{F}_{m | 2n}^{\pm} (\varphi_{j,k,l}(\bold{x}) )=(\pm i )^{2j+ k} \varphi_{j,k,l}(\bold{y}).
\]
The extension of the Fourier transform from $\mathcal{S}(\mathbb{R}^m)_{m|2n}$ to $L_{2}(\mathbb{R}^m)_{m|2n} =L_2(\mathbb{R}^m)\otimes\Lambda_{2n}$ is trivial because clearly $\mathcal{F}_{m|2n}^\pm=\mathcal{F}_{m|0}^\pm\circ\mathcal{F}_{0|2n}^\pm$.
When $M\not\in-2\mathbb{N}$ or when $m=0$, we define the fractional Fourier transform $\mathcal{F}^{\alpha}_{m|2n}$ on $\mathcal{S}(\mathbb{R}^m)\otimes\Lambda_{2n}$, by its action on the basis functions (see \cite{DBS9}):
\begin{equation}
\label{fracFourbasis}
\mathcal{F}^{\alpha}_{m|2n} (\varphi_{j,k,l}(\bold{x})) = e^{ i \alpha (2j+k)} \varphi_{j,k,l}(\bold{y}),
\end{equation}
where $\alpha \in [-\pi/2,\pi/2]$. The fractional Fourier transform thus rotates the basis functions over a multiple of the angle $\alpha$. In the limit case $\alpha = \pm \pi/2$, the fractional Fourier transform reduces to the ordinary Fourier transform. We have the following integral representation (see \cite{DBS9}, theorem 12).
\begin{theorem}
On $\mathcal{S}(\mathbb{R}^m)_{m|2n}$, the fractional Fourier transform is given by
\[
\mathcal{F}^{\alpha}_{m|2n}(f(\bold{x}))= c \int_{\mathbb{R}^{m|2n},x} \exp{\frac{ 4 e^{i \alpha} \langle \bold{x},\bold{y} \rangle - (1+ e^{2i \alpha})(R^2 + R_{\bold{y}}^2)}{2- 2e^{2i \alpha}}} f(\bold{x}),
\]
with $c= \left(\pi (1- e^{2i \alpha})\right)^{-M/2}$.
\label{fracfourthm}
\end{theorem}
\begin{remark}
Note that the sign convention we use in the definition of the Fourier transform (see formula (\ref{innerprodvectorssuper})) is slightly different than the one used in \cite{DBS9}.
\end{remark}
Finally, we repeat some important facts about spherical harmonics in superspace. For proofs we refer the reader to \cite{DBE1}.
\begin{lemma}
\label{polythm}
If $0\le q \le n$ and $0\le k\le n-q$, there exists a homogeneous polynomial $f_{k,p,q}=f_{k,p,q}(r^2,\theta^2)$ (unique up to a multiplicative constant) of total degree $k$ such that $f_{k,p,q} \mathcal{H}_p^b \otimes \mathcal{H}_q^f \neq 0$ and $\nabla^2 (f_{k,p,q} \mathcal{H}_p^b \otimes \mathcal{H}_q^f) = 0$.
This polynomial is given explicitly by
\[
f_{k,p,q}=\sum_{s=0}^ka_sr^{2k-2s}\theta^{2s},\quad a_s=\binom{k}{s}\frac{(n-q-s)!}{\Gamma (\frac{m}{2}+p+k-s)}\frac{\Gamma(\frac{m}{2}+p+k)}{(n-q-k)!}.
\]
\end{lemma}
In particular $f_{0,p,q}=1$ holds. To simplify the subsequent formulas, we have used a different normalization of the $f_{k,p,q}$ as in \cite{DBE1}. Using these polynomials we can obtain a full decomposition of the space of spherical harmonics of degree $k$.
\begin{theorem}[Decomposition of $\mathcal{H}_k$]
\label{decompintoirreps}
Under the action of $SO(m) \times Sp(2n)$ the space $\mathcal{H}_k$ decomposes as
\[
\mathcal{H}_{k} = \bigoplus_{j=0}^{\min(n, k)} \bigoplus_{l=0}^{\min(n-j,\lfloor \frac{k-j}{2} \rfloor)} f_{l,k-2l-j,j} \mathcal{H}^b_{k-2l-j} \otimes \mathcal{H}^f_{j},
\]with $f_{l,k-2l-j,j}$ the polynomials determined in lemma \ref{polythm}.
\end{theorem}
The integration on superspace (\ref{superint}) for $m\not=0$ is linked with an integration over the supersphere defined by the following Pizzetti formula (see \cite{DBS5,DBE1}):
\begin{equation}
\int_{SS} f = \sum_{k=0}^{\infty} \frac{2 \pi^{M/2}}{2^{2k} k!\Gamma(k+M/2)} (\nabla^{2k} f)(0), \quad f \in \mathcal{P}.
\label{defintss}
\end{equation}
The supersphere is formally defined as the algebraic object $R^2-1=0$. It can be proven that (\ref{defintss}) is, up to a constant, the unique linear functional on $\mathcal{P}$ satisfying $\int_{SS}R^2 f=\int_{SS}f$, which is invariant under the group $SO(m)\times Sp(2n)$ and which makes spherical harmonics of different degree `orthogonal' (see \cite{DBE1}, \cite{CDBS1}). In fact, this orthogonality condition can be made even stronger (see \cite{DBE1}, theorem 8):
\begin{theorem}
\label{integorth}
One has that $f_{i,p,q} \mathcal{H}^b_{p} \otimes \mathcal{H}^f_{q} \quad \bot \quad f_{j,r,s} \mathcal{H}^b_{r} \otimes \mathcal{H}^f_{s}$, meaning
\[
\int_{SS} \left(f_{i,p,q} \mathcal{H}^b_{p} \otimes \mathcal{H}^f_{q}\right) \left(f_{j,r,s} \mathcal{H}^b_{r} \otimes \mathcal{H}^f_{s}\right) =0
\]
with respect to the Pizzetti integral, if and only if $(i,p,q) \neq (j,r,s)$.
\end{theorem}
In case $M\not\in-2\mathbb{N}$, for $P_k$ a homogeneous polynomial of degree $k$ we also have
\begin{equation}
\label{superint2}
\int_{\mathbb{R}^{m|2n}}P_k\exp (-R^2) =\frac{1}{2}\Gamma (\frac{k+M}{2})\int_{SS} P_k.
\end{equation}
Finally, the supersphere integration given by (\ref{defintss}) can be extended to more general (non-polynomial) functions as follows (see \cite{CDBS1}, theorem 8)
\begin{eqnarray}
\label{SSintnieuw}
\int_{SS}f&=& \sum_{j=0}^n\int_{\mathbb{S}^{m-1}} d \sigma \int_B\frac{(-1)^j\theta^{2j}}{j!}\left[(\frac{\partial}{\partial r^2})^jr^{m-2}f\right]_{r=1}.
\end{eqnarray}
\subsection{Schr\"odinger equations in superspace}
\label{SEsuper}
Schr\"odinger equations in superspace are equations of the type
\begin{equation}
\label{SchrodEq}
-\frac{\nabla^2}{2} \psi + V \psi = E \psi
\end{equation}
with wave function $\psi \in L_{2}(\mathbb{R}^{m})_{m|2n}$ and the energy $E$ a complex number. For more background on Schr\"odinger equations in superspace and supersymmetric quantum mechanics, see e.g. \cite{MR0780110, MR0723957, MR0952024, MR2277080, MR2025382, Witten, MR0683171}. The potential $V$ is expressed in terms of commuting and anti-commuting variables. In the case $m=0$ this is a purely algebraic eigenvalue problem. When $m \neq 0$, equation (\ref{SchrodEq}) is equivalent with a system of PDEs, as can be observed by expanding $\psi$ in the anti-commuting variables.
Several authors have studied explicit examples of such Schr\"odinger equations. The (purely fermionic) harmonic oscillator was studied in \cite{MR830398}. Anharmonic extensions were studied in \cite{MR967935,MR1019514,MR1032208}. In \cite{ZHANG} the hydrogen atom in superspace (or quantum Kepler problem) was studied using Lie superalgebra techniques. Also the delta potential has been studied, see \cite{DBS8}.
In this paper we consider general potentials of the form $V(R^{2})$, with $V$ a polynomial. The simplest case is then the harmonic oscillator described by the hamiltonian
\begin{equation}
\label{hamiltoniaan}
H= \frac{1}{2}(-\nabla^2+R^2) = \sum_{i=1}^m a_i^+ a_i^- +\sum_{i=1}^{2n} b_i^+ b_i^- + \frac{M}{2}
\end{equation}
with
\[
\begin{array}{llll}
a^+_i = \frac{\sqrt{2}}{2}(x_i - \partial_{x_i} ) &\quad&a^-_i = \frac{\sqrt{2}}{2}(x_i+\partial_{x_i})\\
\vspace{-1mm}\\
b^+_{2i} = \frac{1}{2}({x \grave{}}_{2i} + 2 \partial_{{x \grave{}}_{2i-1}})&\quad&b^-_{2i} =\frac{1}{2} ({x \grave{}}_{2i-1} + 2 \partial_{{x \grave{}}_{2i}})\\
\vspace{-1mm}\\
b^+_{2i-1} = \frac{1}{2}({x \grave{}}_{2i-1} - 2 \partial_{{x \grave{}}_{2i}})&\quad&b^-_{2i-1} = \frac{1}{2}(-{x \grave{}}_{2i} + 2 \partial_{{x \grave{}}_{2i-1}})\\
\vspace{-1mm}\\
\end{array}
\]
the bosonic and fermionic creation and annihilation operators. The spherical Hermite functions defined in (\ref{CliffordHermiteFunctionssuper}) form a basis of eigenvectors of $H$ for $\mathcal{P}\exp(-R^2/2)$, satisfying (see \cite{DBS3})
\begin{equation}
H \varphi_{j,k,l} = \left( \frac{M}{2} + (2j+k)\right) \varphi_{j,k,l}.
\label{HOCH}
\end{equation}
\section{Hermite polynomials in Grassmann algebras}
\setcounter{equation}{0}
\label{FermionicMehler}
\subsection{Inner product on a Grassmann algebra}
We first define two vector space isomorphisms of the Grassmann algebra $\Lambda_{2n}$.
\begin{definition}
\label{defmaptilde}
The transformation $\widetilde{.}:\Lambda_{2n}\to \Lambda_{2n} $ is a linear transformation defined by
\begin{eqnarray*}
\widetilde{{x \grave{}}_{2i-1}}&=&{x \grave{}}_{2i}\\
\widetilde{{x \grave{}}_{2i}}&=&-{x \grave{}}_{2i-1}\\
\widetilde{ab}&=&\widetilde{b}\widetilde{a}, \qquad a,b \in \Lambda_{2n}.
\end{eqnarray*}
\end{definition}
This transformation satisfies the following property.
\begin{lemma}
\label{Sptilde}
For each $B\in Sp(2n)$, define $D\in Sp(2n)$ as $D=J^TBJ$ (as matrix multiplications). The transformation $\widetilde{\cdot}$ intertwines with the action of $Sp(2n)$ in the following way:
\[
\widetilde{\cdot}\circ B= D\circ\widetilde{\cdot}.
\]
\end{lemma}
\begin{proof}
For a monomial ${x\grave{}}_A={x\grave{}}_{i_1}\cdots {x\grave{}}_{i_k}$, definition \ref{defmaptilde} yields
\begin{eqnarray*}
\widetilde{{x\grave{}}_A}=\sum_{j_1,\cdots,j_k}J_{i_1j_1}\cdots J_{i_kj_k}{x\grave{}}_{j_k}\cdots {x\grave{}}_{j_1}.
\end{eqnarray*}
The definition of the action of $Sp(2n)$ in formula \eqref{actieOSp} implies
\begin{eqnarray*}
B{x\grave{}}_A&=&\sum_{j_1,\cdots,j_k}B_{j_1i_1}\cdots B_{j_ki_k}{x\grave{}}_{j_1}\cdots {x\grave{}}_{j_k}.
\end{eqnarray*}
Combining these transformations yields
\begin{eqnarray*}
\widetilde{B{x\grave{}}_A}&=&\sum_{j_1,\cdots,j_k}B_{j_1i_1}\cdots B_{j_ki_k}\sum_{l_1,\cdots, l_k}J_{j_1l_1}\cdots J_{j_k l_k}{x\grave{}}_{l_k}\cdots {x\grave{}}_{l_1}
\end{eqnarray*}
and
\begin{eqnarray*}
D\widetilde{{x\grave{}}_A}&=&\sum_{j_1,\cdots,j_k}J_{i_1j_1}\cdots J_{i_kj_k}\sum_{l_1,\cdots,l_k}D_{l_1j_1}\cdots D_{l_kj_k}{x\grave{}}_{l_k}\cdots {x\grave{}}_{l_1}.
\end{eqnarray*}
For the last two expressions to be identical, $DJ^T=J^TB$ needs to hold, which is equivalent with $D=J^TBJ$.
\end{proof}
\begin{definition}
\label{defstar}
The star map $\ast$ maps monomials ${x \grave{}}_A = {x \grave{}}_1^{\, \alpha_1} \ldots {x \grave{}}_{2n}^{\, \alpha_{2n}}$ of degree $k$ to monomials $\ast {x \grave{}}_A = \pm 2^{k-n} {x \grave{}}_1^{1-\alpha_1} \ldots {x \grave{}}_{2n}^{1-\alpha_{2n}}$ of degree $(2n-k)$ where the sign is chosen such that
\[
{x \grave{}}_A (\ast {x \grave{}}_A )= 2^{k-n} {x \grave{}}_{1} \ldots {x \grave{}}_{2n}.
\]
By linearity, $\ast$ is extended to the whole of $\Lambda_{2n}$.
\end{definition}
It is easy to check that $\ast \ast {x \grave{}}_A = (-1)^{k (k-2n)} {x \grave{}}_A$, so $\ast$ behaves very similarly as the Hodge star map acting on the space of differential forms on a Riemannian manifold. Because the dimension $(-2n)$ in our case is always even, we obtain
\[\ast\ast {x\grave{}}_A=(-1)^k{x\grave{}}_A=\widetilde{\widetilde{{x\grave{}}_A}}.
\]
Note that $\ast$ and $\widetilde{.}$ are vector space isomorphisms on $\Lambda_{2n}$. It is also easy to check that $\widetilde{.}$ leaves $\mathcal{H}_{k}^{f}$ invariant.
Using the star map we define the following inner product on $\Lambda_{2n}$ (where we use the bra-ket notation for convenience):
\begin{definition}
\label{inprodGrass}
The inner product $\langle.|.\rangle:\Lambda_{2n}\times \Lambda_{2n}\to \mathbb{C}$ is given by
\begin{eqnarray*}
\langle f|g \rangle &=& \int_{B,x} f (\ast \overline{g})\\
&=& \pi^{-n} \partial_{{x \grave{}}_{2n}} \ldots \partial_{{x \grave{}}_{1}} f (\ast \overline{g})\\
&=& \frac{1}{(2\pi)^n} \sum_A 2^{|A|} f_A \overline{g_{A}},
\end{eqnarray*}
with $\overline{\, \cdot\, }$ the standard complex conjugation, $f=\sum_{A}f_A{x\grave{}}_A$ and $g=\sum_{A}g_A{x\grave{}}_A$.
\end{definition}
This inner product, in a different formulation, was also used in \cite{MR2025382}. When we will introduce new inner products in superspace (see section \ref{supersuper}), we will use the notation $\langle.|.\rangle_{\Lambda_{2n}}$ for the inner product in definition \ref{inprodGrass} to avoid confusion. The fact that $\langle.|.\rangle$ is an inner product follows immediately from the last expression in the definition. This immediately implies the following lemma.
\begin{lemma}
\label{triviaalstelemma}
If for $p\in \Lambda_{2n}$ it holds that $\int_B pq=0$ for every $q$ $\in \Lambda_{2n}$, then $p=0$.
\end{lemma}
\begin{proof}
Using the fact that the star map is an isomorphism, $\int_B pq=0$ for every $q$ $\in \Lambda_{2n}$ implies
\[
\int_B p (\ast \overline{p})=0= \langle p | p\rangle.
\]
This is only possible when $p=0$.
\end{proof}
It is easy to see that any inner product on the Grassman algebra $\Lambda_{2n}$ can be written similarly to definition \ref{inprodGrass} as $\langle f|g \rangle=\sum_{A,B} f_A P_{AB}\overline{ g_B}$ with $P$ a hermitian positive definite $2^{2n}\times2^{2n}$ matrix. This is an inner product of the form $\int_{B,x} f (\Phi\overline{g})$, with $\Phi$ a different isomorphism on the Grassman algebra.
The inner product in definition \ref{inprodGrass} is not covariant as the classical bosonic inner product $\int_{\mathbb{R}^m} f\overline{g} dV(\underline{x})$, because the star map does not behave covariantly for transformations with determinant equal to one. However, it can be shown that there are no covariant positive definite inner products on $\Lambda_{2n}$. Before doing so, we reobtain the well-known fact that the Berezin integral itself does behave covariantly.
\begin{lemma}
\label{Berezincov}
Let $A$ be a transformation on the generators of the Grassmann algebra given by ${x\grave{}}_k=\sum_{j=1}^{2n}A_{kj}{y\grave{}}_j$. Then one has
\begin{equation*}
\int_{B,y} f(\underline{x} \grave{} (\underline{y} \grave{})) \equiv \int_{B,y} f(A.\underline{y} \grave{}) =\det (A) \int_{B,x}f(\underline{x} \grave{}).
\end{equation*}
\end{lemma}
\begin{proof}
The transformation preserves the degree of $f$, so we only have to consider the case $f(\underline{x} \grave{})={x \grave{}}_{1} \ldots {x \grave{}}_{2n}$. Because of the anticommutation rules we immediately have $f(\underline{x} \grave{}(\underline{y} \grave{}))=\det (A) {y \grave{}}_{1} \ldots {y \grave{}}_{2n}$. As $\int_{B,y}$ is defined by $ \pi^{-n} \partial_{{y \grave{}}_{2n}} \ldots \partial_{{y \grave{}}_{1}}$, the lemma follows.
\end{proof}
So for transformations with determinant one the integration is independent of the choice of coordinates.
\begin{remark}
From the calculation above it is clear that the Jacobian determinant appears on the other side in comparison to the bosonic case. Therefore, the above formula for substitution in the Berezin integral is usually written after multiplying both sides with $1/\det (A)$. The factor $1/\det (A)$ is then called the Berezinian (see \cite{MR732126}), the fermionic equivalent of the Jacobian. In general the Berezinian is a superdeterminant, but in the purely fermionic case this is equivalent to the inverse.
\end{remark}
\begin{proposition}
\label{geencov}
There are no inner products on the Grassmann algebra $\Lambda_{2n}$ with the property that
\begin{eqnarray*}
\langle f(\underline{x} \grave{}) |g(\underline{x} \grave{})\rangle_x &=& \langle f(\underline{x} \grave{}(\underline{y} \grave{}) |g(\underline{x} \grave{}(\underline{y} \grave{}))\rangle_y
\end{eqnarray*}
for transformations $A$ as in lemma \ref{Berezincov} with $\det (A)=1$.
\end{proposition}
\begin{proof}
We already mentioned that every inner product on the Grassmann algebra is of the form
\[
\langle f|g\rangle=\sum_{A,B}f_AP_{AB} \overline{g_B},
\]
for some hermitian positive definite $2^{2n}\times 2^{2n}$ matrix $P$. Now we only consider the case with $f$ and $g$ homogeneous of degree one. Then we get a submatrix of $P$, the hermitian positive definite $2n\times 2n$ matrix $Q$. For the inner product to be covariant, $Q$ has to satisfy $R^H QR=Q$ for every matrix $R$ with determinant one and with $\cdot^H$ the hermitian conjugate. Basic linear algebra shows that such a non-zero $Q$ does not exist.
\end{proof}
Proposition \ref{geencov} can be made even stronger. By similar arguments one can show that there are no inner products on the Grassmann algebra which are invariant under symplectic transformations. In the subsequent proposition \ref{Spinprod} the behavior of the inner product with respect to the symplectic group will be studied.
Now we derive the properties of the inner product given in definition \ref{inprodGrass}. We start with the following lemma, which shows that the adjoint of ${x\grave{}}_j$ is given by the usual adjoint in supersymmetric quantum mechanics.
\begin{lemma}
\label{adjoints}
The adjoint of $\partial_{{x \grave{}}_{j}}$ with respect to the inner product $\langle . | . \rangle $ on $\Lambda_{2n}$ is given by ${x \grave{}}_{j}/2$. This property determines the inner product uniquely, up to a multiplicative constant.
\end{lemma}
\begin{proof}
We need to prove that $\langle \partial_{{x \grave{}}_{j}}f|g \rangle = \langle f| \frac{{x \grave{}}_{j}}{2} g \rangle$.
Due to linearity, it suffices to take $f = a {x \grave{}}_j {x \grave{}}_A$, $g = b {x \grave{}}_A$ with $a,b \in \mathbb{C}$. Then
\[
\langle \partial_{{x \grave{}}_{j}}f|g \rangle = \int_{B,x} a {x \grave{}}_A (\ast \overline{b} {x \grave{}}_A)= \frac{\pi^{-n}}{2^n} 2^{|A|} a \overline{b}
\]
and on the other hand
\[
\langle f| {x \grave{}}_{j}g \rangle = \int_{B,x} a {x \grave{}}_{j}{x \grave{}}_A (\ast \overline{b} {x \grave{}}_{j}{x \grave{}}_A)
= \frac{\pi^{-n}}{2^n} 2^{|A| +1} a \overline{b} ,
\]
proving the first part of the lemma. Now suppose we have a different inner product on $\Lambda_{2n}$, $(.|.)$, for which the same property holds. For two monomials ${x \grave{}}_A = {x \grave{}}_{1}^{\, \alpha_1} \ldots {x \grave{}}_{2n}^{\, \alpha_{2n}}$ of degree $k$ and ${x \grave{}}_B = {x \grave{}}_1^{\, \beta_1} \ldots {x \grave{}}_{2n}^{\, \beta_{2n}}$ of degree $l$ (with $k\ge l$) we find
\begin{eqnarray*}
({x \grave{}}_A|{x \grave{}}_B)&=&2^k(1|\partial_{{x \grave{}}_{2n}}^{\, \alpha_{2n}} \ldots \partial_{{x \grave{}}_{1}}^{\, \alpha_{1}}{x \grave{}}_B).
\end{eqnarray*}
Now, $\partial_{{x \grave{}}_{2n}}^{\, \alpha_{2n}} \ldots \partial_{{x \grave{}}_{1}}^{\, \alpha_{1}}{x \grave{}}_B$ is zero when $k>l$, or when $k=l$ and ${x \grave{}}_A\not={x \grave{}}_B$. This means that we find $({x \grave{}}_A,{x \grave{}}_B)=0$ when $A\not= B$ and $({x \grave{}}_A,{x \grave{}}_A)=2^k(1|1)$, which corresponds to definition \ref{inprodGrass}.
\end{proof}
Contrary to the bosonic $L_2$ inner product, $\partial_{{x\grave{}}_j}$ is defined on the entire space $\Lambda_{2n}$, which simplifies the notion of an adjoint operator. For the sequel, we need the adjoints of the generators of the $\mathfrak{sl}_{2}$ algebra.
\begin{corollary}
\label{hermtoeg}
With respect to the inner product $\langle . | . \rangle $ on $\Lambda_{2n}$ the adjoints of $\theta^2$, $\nabla^2_f$ and $\mathbb{E}_f-n$ are given by
\begin{equation*}
(\theta^2)^\dagger=-\nabla^2_f, \qquad (\nabla^2_f)^\dagger=-\theta^2, \qquad (\mathbb{E}_f-n)^\dagger=(\mathbb{E}_f-n).
\end{equation*}
\end{corollary}
\begin{proof}
The first two adjoints follow immediately from the previous lemma, e.g. $({x \grave{}}_{2j-1}{x \grave{}}_{2j})^\dagger={x \grave{}}_{2j}^\dagger {x \grave{}}_{2j-1}^\dagger=-4\partial_{{x \grave{}}_{2j-1}}\partial_{{x \grave{}}_{2j}}$.
The last one follows from the first two and formula (\ref{fermEuler}):
\[
(\mathbb{E}_f-n)^\dagger=\frac{1}{4}[\nabla^2_f,\theta^2]^\dagger=\frac{1}{4}[(\theta^2)^\dagger,(\nabla^2_f)^\dagger]=\frac{1}{4}[-\nabla^2_f,-\theta^2].
\]
\end{proof}
Recall that the hamiltonian of the fermionic harmonic oscillator is given by (see (\ref{hamiltoniaan}))
\[
H = (-\nabla^2_f +\theta^2)/2.
\]
Using corollary \ref{hermtoeg} we immediately obtain that $H$ is self-adjoint, hence a notation such as $\langle f|H |g \rangle$ makes sense.
It would of course be preferable to have an inner product for which $(\theta^2)^\dagger=\theta^2$ and $(\nabla^2_f)^\dagger=\nabla^2_f$, similar to the bosonic $L_2$ inner product. With such an inner product other hamiltonians would be hermitian too. An interesting class of relevant hamiltonians, as already indicated in section \ref{SEsuper}, is of the form $\Delta_f/2 +V(\theta^2)$, with $V$ a polynomial. Several examples have already been studied in \cite{MR967935,MR1019514,MR1032208}. It is however easy to see that such an inner product does not exist.
\begin{proposition}
\label{onmogelijkinprod}
There are no inner products on the Grassman algebra $\Lambda_{2n}$ for which $(\theta^2)^\dagger=\theta^2$.
\end{proposition}
\begin{proof}
If multiplication with $\theta^2$ is a hermitian operation then
\begin{eqnarray*}
\langle \theta^{2n}|\theta^{2n}\rangle &=& \langle \theta^{2n-2}|\theta^{2n+2} \rangle=0
\end{eqnarray*}
as $\theta^{2n+2}=0$. Hence $\langle.|.\rangle $ is not positive definite and thus not an inner product.
\end{proof}
This proposition does not really form a limitation for negative dimensional quantum mechanics per se, as can be seen from the isotropic anharmonic oscillator. In \cite{MR967935} it was calculated that the eigenvalues for the hamiltonian $H=-\nabla^2_f+\theta^2-\lambda \theta^4$ (with $\lambda$ real) can be complex. Such a hamiltonian can therefore never be hermitian with respect to an inner product.
\vspace{3mm}
We now investigate the action of the star map in more detail. Therefore we first state some useful properties of the Berezin integral.
\begin{lemma}
\label{Berezin}
If $h\in \Lambda_{2n}^{k}$ and $f,g \in \Lambda_{2n}$, then the following relations hold
\begin{align*}
&(i) \int_B (\partial_{{x \grave{}}_{j}} f )h=(-1)^k\int_B f\partial_{{x \grave{}}_{j}} h &\quad&(iii) \int_B (\theta^2 f) g=\int_B f\theta^2 g\\
&(ii) \int_B (\nabla^2_f f) g=\int_B f\nabla^2_f g &\quad&(iv) \int_B ((\mathbb{E}_f-n)f)g=-\int_B f(\mathbb{E}_f-n)g.
\end{align*}
\end{lemma}
\begin{proof}
The first property follows from $\int_B \partial_{{x \grave{}}_{j}}=0$ and the fact that we only have to consider $f$ of degree $2n-k+1$. $(ii)$ immediately follows from $(i)$. $(iii)$ is trivial and $(iv)$ follows from $(ii)$ and $(iii)$ by equation (\ref{fermEuler}).
\end{proof}
Using the previous lemma we obtain the following calculation rules for the star map.
\begin{lemma}
\label{eigstar}
If $f\in \Lambda_{2n}^{k}$ and $g\in \Lambda_{2n}$, then the following relations hold
\begin{align*}
& (i)*{x \grave{}}_{j}f=(-1)^k2\partial_{{x \grave{}}_{j}}*f&\quad &(iv)*\theta^2 g = -\nabla^2_f *g\\
& (ii)*\partial_{{x \grave{}}_{j}}f=-(-1)^k\frac{1}{2}{x \grave{}}_{j}*f&\quad &(v)*(\mathbb{E}_f-n)g= -(\mathbb{E}_f-n)* g\\
& (iii)*\nabla^2_f g= -\theta^2 * g&\quad &(vi)*1=\frac{1}{2^n}{x \grave{}}_{1}\ldots{x \grave{}}_{2n}=\frac{(-1)^n\theta^{2n}}{2^n n!}.
\end{align*}
\end{lemma}
\begin{proof}
The proof follows easily from the lemmas \ref{triviaalstelemma}, \ref{adjoints} and \ref{Berezin}. As an example we prove $(i)$. For every $g\in\Lambda_{2n}$, we have
\begin{eqnarray*}
\int_B g(\ast {x\grave{}}_jf)&=&2\int_B(\partial_{{x\grave{}}_j}g)(\ast f)\\
&=&2(-1)^{2n-k}\int_Bg(\partial_{{x\grave{}}_j}\ast f).
\end{eqnarray*}
This implies $(i)$ by lemma \ref{triviaalstelemma}.
\end{proof}
Now we will calculate the action of the star map on the spherical Hermite functions defined in (\ref{CliffordHermiteFunctions}). We start with two auxiliary results.
\begin{lemma}
\label{astHk}
For $H_k \in \mathcal{H}_k^f$, the following holds:\\
\begin{equation*}
*H_k=\widetilde{H}_k \frac{(-1)^{n-k}\theta^{2n-2k}}{2^{n-k}(n-k)!}.
\end{equation*}
\end{lemma}
\begin{proof}
We prove this lemma by induction on $n$. For $n=1$ the result is trivial. If it holds for $n-1$, then theorem \ref{basisrecursief} provides us with a useful basis for $\mathcal{H}_k^f$ in $\Lambda_{2n}$. We take $\Lambda_{2n-2}$ the Grassmann algebra without ${x\grave{}}_1$ and ${x\grave{}}_2$ and put $\theta^{2}_{2n-2} =- \sum_{j=2}^n {x\grave{}}_{2j-1} {x\grave{}}_{2j}$. Using definition \ref{defstar} we find that for a monomial ${x\grave{}}_A\in\Lambda_{2n-2}$ of degree $k$, $\ast_{n}{x\grave{}}_A=(\ast_{n-1}{x\grave{}}_A)\frac{{x\grave{}}_1{x\grave{}}_2}{2}$. By linearity this holds for every element of $\Lambda_{2n-2}$. In theorem \ref{basisrecursief} there are 3 different types of spherical harmonics. For the first type (namely $H_{k} \in \mathcal{H}_k^f({x\grave{}}_3, \ldots, {x\grave{}}_{2n})$) we find, using the induction hypothesis,
\begin{eqnarray*}
\ast_nH_k&=&(\ast_{n-1}H_k)\frac{{x\grave{}}_1{x\grave{}}_2}{2}\\
&=&\widetilde{H}_k \frac{(-1)^{n-k-1}\theta^{2n-2k-2}_{2n-2}}{2^{n-k-1}(n-k-1)!}\frac{{x\grave{}}_1{x\grave{}}_2}{2}\\
&=&\widetilde{H}_k \frac{(-1)^{n-k}\theta^{2n-2k}}{2^{n-k}(n-k)!}.
\end{eqnarray*}
Indeed, we have
\[
(-{x\grave{}}_1{x\grave{}}_2 + \theta_{2n-2}^{2})^{n-k} = \theta_{2n-2}^{2n-2k} - (n-k) {x\grave{}}_1{x\grave{}}_2 \, \theta_{2n-2}^{2n-2k-2}
\]
and $\widetilde{H}_k\theta^{2n-2k}_{2n-2} = 0$ because of equation (\ref{Fischer}).
For the two other types the lemma follows from the result for the first type and the calculation rules in lemma \ref{eigstar}.
\end{proof}
\begin{lemma}
\label{expstar}
For all $H_k \in \mathcal{H}_k^f$ and for $i+k\le n$ one has
\begin{eqnarray*}
& &(i)*\theta^{2i}H_k= (-1)^{n-k}2^i i! \frac{\theta^{2n-2k-2i}}{2^{n-k-i}(n-k-i)!} \widetilde{H}_k \\
& &(ii)*H_k\exp(-\theta^2/2)=\widetilde{H}_k\exp(-\theta^2/2).
\end{eqnarray*}
\end{lemma}
\begin{proof}
Using lemma \ref{Laplfischer}, lemma \ref{eigstar} and lemma \ref{astHk} we obtain
\begin{eqnarray*}
*\theta^{2i}H_k&=&(-1)^{i+n-k}\nabla_f^{2i} \frac{\theta^{2n-2k}}{2^{n-k}(n-k)!}\widetilde{H}_k\\
&=&(-1)^{n-k}4^i\frac{(n-k)!}{(n-k-i)!}\frac{(n+i-k-n+k)!}{(n-k-n+k)!}\frac{\theta^{2n-2k-2i}}{2^{n-k}(n-k)!}\widetilde{H}_k\\
&=&(-1)^{n-k}4^i\frac{i!}{(n-k-i)!}\frac{\theta^{2n-2k-2i}}{2^{n-k}}\widetilde{H}_k.
\end{eqnarray*}
This proves the first formula. Using this result we then have
\begin{eqnarray*}
*H_k \exp(-\theta^2/2)&=&\sum_{i=0}^{n-k} \frac{1}{2^ii!}*(-1)^i\theta^{2i} H_k\\
&=&\sum_{i=0}^{n-k}\frac{(-1)^{n-k-i}\theta^{2n-2k-2i}}{2^{n-k-i}(n-k-i)!} \widetilde{H}_k
=\sum_{j=0}^{n-k}\frac{(-1)^j\theta^{2j}}{2^jj!}\widetilde{H}_k.
\end{eqnarray*}
\end{proof}
We can now explicitly state the action of the star map on the spherical Hermite functions defined in (\ref{CliffordHermiteFunctions}).
\begin{theorem}
\label{CHstar}
If $H_k^{(l)}\in \mathcal{H}_k^f$ and $j+k\le n$, then the following holds for the $\varphi^{f}_{j,k,l}$ defined in equation (\ref{CliffordHermiteFunctions}):
\begin{eqnarray*}
*L_j^{-n+k-1}(\theta^2) H_k^{(l)}\exp(-\theta^2/2)=(-1)^j L_j^{-n+k-1}(\theta^2)\widetilde{H}_k^{(l)}\exp(-\theta^2/2).
\end{eqnarray*}
\end{theorem}
\begin{proof}
This follows from the lemmas \ref{eigstar} and \ref{expstar} $(ii)$ as follows
\begin{eqnarray*}
*\varphi^{f}_{j,k,l} &=& *(-\nabla^2_{f} -\theta^2 + 2\mathbb{E} -2n)^j H_k^{(l)} \exp{(-\theta^2/2)} \\
&=& (\theta^2+ \nabla^2_{f} - 2\mathbb{E} +2n)^j *H_k^{(l)} \exp{(-\theta^2/2)}\\
&=&(-1)^j(-\nabla^2_f-\theta^2+2\mathbb{E}-2n)^j\widetilde{H}_k^{(l)}\exp(-\theta^2/2)\\
&=&(-1)^j 2^{2j}j! L_j^{-n+k-1}(\theta^2)\widetilde{H}_k^{(l)}\exp(-\theta^2/2).
\end{eqnarray*}
\end{proof}
Although the inner product is not invariant under the symplectic group, it does behave canonically with respect to it.
\begin{proposition}
\label{Spinprod}
For $f,g\in\Lambda_{2n}$ and $A\in Sp(2n)$, with action as defined in formula (\ref{actieOSp}), the following relation holds
\begin{eqnarray*}
\langle Af|g\rangle &=&\langle f|A^Tg\rangle,
\end{eqnarray*}
which implies $A^\dagger=A^T$, with $A^T$ the matrix transpose of $A$. This is equivalent with
\begin{eqnarray*}
\langle AJf|JAg\rangle&=&\langle f|g\rangle,
\end{eqnarray*}
for all $A\in Sp(2n)$.
\end{proposition}
\begin{proof}
Lemma \ref{Sptilde} and \ref{expstar} imply that $\ast A g= (J^TAJ)\ast g$ for $A\in Sp(2n)$. This implies that
\begin{eqnarray*}
\langle (AJ)f|(JA)g\rangle&=&\int_B (AJf)(\ast (JA\overline{g}))\\
&=&\int_B (AJf)(AJ\ast \overline{g})=\int_B AJf(\ast \overline{g}).
\end{eqnarray*}
Lemma \ref{Berezincov} then yields
\begin{eqnarray*}
\langle (AJ)f|(JA)g\rangle={\det(AJ)^T}\int_B f(\ast \overline{g})=\int_B f(\ast \overline{g}),
\end{eqnarray*}
since $\det A=1$ for all $A\in Sp(2n)$. It is easily checked that the map $A\to AJ$ is a bijection of $Sp(2n)$. Since for $A\in Sp(2n)$, the relation $(AJ)^{-1}=(JA)^{T}$ holds, the claim $\langle AJf|JAg\rangle=\langle f|g\rangle$ for all $A\in Sp(2n)$, is equivalent with stating $\langle Af|g\rangle=\langle f|A^Tg\rangle$ for all $A\in Sp(2n)$.
\end{proof}
\subsection{Orthogonality of the spherical Hermite polynomials}
Now choose a (real) orthogonal basis $\{ H_k^{(l)}\}$, $l =1, \ldots, \dim \mathcal{H}_{k}^{f}$ of $\mathcal{H}_k^f$ such that
\begin{equation}
\label{OrthSphHarm}
\langle H_k^{(p)} \exp(-\theta^2/2)|H_k^{(q)} \exp(-\theta^2/2) \rangle = \frac{\delta_{pq}}{(n-k)!}.
\end{equation}
A straightforward calculation shows that this is equivalent with
\begin{equation}
\label{OrthSphHarm2}
\langle H_k^{(p)} |H_k^{(q)} \rangle = \frac{\delta_{pq}}{2^{n-k}(n-k)!}.
\end{equation}
Using this basis, we consider the spherical Hermite functions $\varphi_{j,k,l}^{f}$ defined in (\ref{CliffordHermiteFunctions}). These functions are eigenfunctions of the harmonic oscillator (see (\ref{HOCH})), which immediately implies
\begin{equation*}
\langle \varphi_{j,k,l}^{f} |\varphi_{p,q,r}^{f} \rangle = 0
\end{equation*}
whenever $2j + k \neq 2p+q$. This can be generalized as follows.
\begin{theorem}
\label{orthocliffherm}
The spherical Hermite functions defined in equation (\ref{CliffordHermiteFunctions}), using a basis of fermionic harmonics satisfying (\ref{OrthSphHarm}), are orthogonal with respect to the inner product $\langle .|.\rangle $ on $\Lambda_{2n}$:
\begin{eqnarray*}
\langle \varphi_{j,k,l}^{f} |\varphi_{p,q,r}^{f} \rangle& =& \delta_{jp}\delta_{kq}\delta_{lr}\frac{4^{2j}j!}{(n-k-j)!}.
\end{eqnarray*}
\end{theorem}
\begin{proof}
Without loss of generality we assume $j\ge p$. Using equation \eqref{CliffordHermiteFunctions}, corollary \ref{hermtoeg} and equation (\ref{differentiaalvgl}) and we obtain
\begin{eqnarray*}
&&\langle \varphi_{j,k,l}^{f}|\varphi_{p,q,r}^{f} \rangle \\
&=& \langle H_k^{(l)}\exp (-\theta^2/2) | (\nabla^2_f+\theta^2 + 2\mathbb{E}-2n)^j \varphi_{p,q,r}^{f}\rangle\\
&=&\delta_{jp}\frac{4^{2j}j!(n-q)!}{(n-q-j)!}\langle H_k^{(l)}\exp (-\theta^2/2) | H_q^{(r)}\exp (-\theta^2/2) \rangle.
\end{eqnarray*}
Because $H_k^{(l)}\exp (-\theta^2/2)$ and $H_q^{(r)}\exp (-\theta^2/2)$ belong to a different eigenspace of $H$ if $k\not=q$ and because of (\ref{OrthSphHarm}), we get the desired result.
\end{proof}
Denoting the normalization constants $\gamma^f_{j,k} = \langle \varphi_{j,k,l}^{f} |\varphi_{j,k,l}^{f} \rangle=\frac{4^{2j} j!}{(n-k-j)!}$ we have that the set of functions $\phi_{j,k,l}^{f} = \varphi_{j,k,l}^{f} / \sqrt{\gamma^f_{j,k}}$ satisfies
\[
\langle \phi_{j,k,l}^{f} |\phi_{p,q,r}^{f} \rangle = \delta_{jp} \delta_{kq} \delta_{lr}.
\]
Finally, we will need the following corollary.
\begin{corollary}
\label{orthsphharmferm}
For $H_k\in\mathcal{H}_k^f$ and $H_l\in\mathcal{H}_l^f$ with $k\not=l$ and $p(\theta^2)$ and $q(\theta^2)$ polynomials in $\theta^2$, the following holds
\[
\langle p(\theta^2) H_k|q(\theta^2)H_l\rangle =0.
\]
\end{corollary}
\begin{proof}
Since $p(\theta^2)H_k\in\mbox{span} \{\varphi^f_{j,k,l}| j\le n-k,l\le\dim\mathcal{H}_k^f\}$, theorem \ref{orthocliffherm} implies the corollary.
\end{proof}
\subsection{Mehler formula}
Now we have all tools necessary to obtain a Mehler formula in the Grassmann algebra $\Lambda_{2n}$. Let $f(\underline{x} \grave{})$ be an element of $\Lambda_{2n}$. Then it can be decomposed as $f(\underline{x} \grave{}) = \sum_{j,k,l} a_{j,k,l} \phi_{j,k,l}^{f}(\underline{x} \grave{})$ with $a_{j,k,l} = \langle f |\phi_{j,k,l}^{f}\rangle$. We calculate the fractional Fourier transform of $f$, defined in formula (\ref{fracFourbasis}), as
\begin{eqnarray*}
\mathcal{F}^{\alpha}_{0|2n} (f) &=& \sum_{j,k,l} a_{j,k,l} e^{i \alpha (2j+k)} \phi_{j,k,l}^{f}(\underline{y} \grave{})\\
&=& \sum_{j,k,l} \langle f |\phi_{j,k,l}\rangle e^{i \alpha (2j+k)} \phi_{j,k,l}^{f}(\underline{y} \grave{})\\
&=&\int_{B,x} f(\underline{x} \grave{}) \sum_{j,k,l} \ast (\phi_{j,k,l}^{f})(\underline{x} \grave{}) e^{i \alpha (2j+k)} \phi_{j,k,l}^{f}(\underline{y} \grave{})\\
&=&\int_{B,x} \sum_{j,k,l} \ast (\phi_{j,k,l}^{f})(\underline{x} \grave{}) e^{i \alpha (2j+k)} \phi_{j,k,l}^{f}(\underline{y} \grave{}) f(\underline{x} \grave{}).
\end{eqnarray*}
Comparing this expression with the integral representation of the fractional Fourier transform in theorem \ref{fracfourthm} (and using lemma \ref{triviaalstelemma}) yields the Mehler formula for Grassmann algebras:
\begin{eqnarray*}
&&\sum_{j,k,l} \ast (\phi_{j,k,l}^{f})(\underline{x} \grave{}) e^{i \alpha (2j+k)} \phi_{j,k,l}^{f}(\underline{y} \grave{})\\
& =& \left(\pi (1- e^{2i \alpha})\right)^{n} \exp{\frac{ 4 e^{i \alpha} \langle \underline{x} \grave{} , \underline{y} \grave{} \rangle -(1+ e^{2i \alpha})(\theta^2 + \theta_{\underline{y} \grave{}}^2)}{2- 2e^{2i \alpha}}}
\end{eqnarray*}
with $k = 0, \ldots, n$; $j = 0, \ldots, n-k$ and $l = 1, \ldots, \dim \mathcal{H}_{k}^{f}$. As this is a finite summation, there is no question of convergence. Using theorem \ref{CHstar} the left-hand side can be rewritten as:
\[
\sum_{j,k,l}\frac{(-1)^j}{\gamma^f_{j,k}} 4^{2j}(j!)^2 L_j^{(-n+k-1)}(\theta^2)\widetilde{H}_k^{(l)}(\underline{x} \grave{}) e^{i \alpha (2j+k)}L_j^{(-n+k-1)}(\theta_{\underline{y} \grave{}}^2)H_k^{(l)}(\underline{y} \grave{})\exp(-\frac{\theta^2+\theta_{\underline{y} \grave{}}^2}{2}).
\]
In this expression there is a summation $\sum_l \widetilde{H}_k^{(l)}(\underline{x} \grave{}) H_k^{(l)}(\underline{y} \grave{})$. This summation can be interpreted as a reproducing kernel for the space $\mathcal{H}_{k}^{f}$. In the following theorem we obtain an explicit expression for this sum.
\begin{theorem}
\label{reprkern}
The function $F_k(\underline{x} \grave{},\underline{y} \grave{})= \sum_l \widetilde{H}_k^{(l)}(\underline{x} \grave{}) H_k^{(l)}(\underline{y} \grave{})$ is given by
\begin{equation}
F_k(\underline{x} \grave{},\underline{y} \grave{})=\sum_{j=0}^{\lfloor \frac{k}{2}\rfloor}c^k_j (\langle \underline{x} \grave{} , \underline{y} \grave{}\rangle)^{k-2j}\theta^{2j}\theta_{\underline{y} \grave{}}^{2j}
\label{reprkernfermformula}
\end{equation}
with constants $c_j^k=2^{k-2j}\pi^n\frac{(n+1-k)}{(k-2j)!j!(n+1+j-k)!}$.
\end{theorem}
\begin{proof}
We consider an element $\alpha(\underline{x} \grave{},\underline{y} \grave{})$ of the Grassmann algebra $\Lambda_{4n}$ generated by $\{{x\grave{}}_{1},{x\grave{}}_{2},\cdots,{x\grave{}}_{2n},{y\grave{}}_{1},\cdots,{y\grave{}}_{2n}\}$ which is harmonic, homogeneous of degree $k$ in both $\underline{x} \grave{}$ and $\underline{y} \grave{}$ and which satisfies
\begin{equation}
\label{fermreprkernelprop}
\langle\alpha(\underline{x} \grave{},\underline{y} \grave{})|H_k^{(r)}(\underline{y} \grave{})\rangle_{\underline{y} \grave{}}=\frac{\widetilde{H}_k^{(r)}(\underline{x} \grave{})}{2^{n-k}(n-k)!}.
\end{equation}
The harmonicity implies that $\alpha$ has to be of the from $\sum_{l,t} c_{lt} \widetilde{H}_k^{(l)}(\underline{x} \grave{}) H_k^{(t)}(\underline{y} \grave{})$ for some constants $c_{lt}$. Formula (\ref{fermreprkernelprop}) then implies $c_{lt}=\delta_{lt}$, so $\alpha$ is unique. This means $F_k$ is uniquely determined by these properties.
On the other hand, for any $R_k\in\Lambda_{2n}(\underline{y} \grave{})$ of degree $k$ we calculate, using lemma \ref{adjoints},
\begin{eqnarray*}
\langle \langle \underline{x} \grave{},\underline{y} \grave{}\rangle^k|R_k(\underline{y} \grave{})\rangle_{\underline{y} \grave{}}&=&\frac{1}{2}\sum_{j=1}^{2n}\widetilde{x\grave{}}_j\langle \langle \underline{x} \grave{},\underline{y} \grave{}\rangle^{k-1}{{y\grave{}}_j}|R_k(\underline{y} \grave{})\rangle_{\underline{y} \grave{}}\\
&=&\left(\frac{1}{2}\right)^k\sum_{j_1,\cdots,j_k=1}^{2n}\widetilde{x\grave{}}_{j_k}\cdots\widetilde{x\grave{}}_{j_1}\langle {{y\grave{}}_{j_1}}\cdots {{y\grave{}}_{j_k}}|R_k(\underline{y} \grave{})\rangle_{\underline{y} \grave{}}\\
&=&\sum_{j_1,\cdots,j_k=1}^{2n}\widetilde{{x\grave{}}_{j_1}\cdots {x\grave{}}_{j_k}}\langle1|\partial_{{y\grave{}}_{j_k}}\cdots\partial_{{y\grave{}}_{j_1}}R_k(\underline{y} \grave{})\rangle_{\underline{y} \grave{}}\\
&=&\frac{k!}{(2\pi)^n}\widetilde{R_k(\underline{x} \grave{})}.
\end{eqnarray*}
This means that the normalized harmonic part of the Fischer decomposition of $\langle \underline{x} \grave{},\underline{y} \grave{}\rangle^k$ given by equation (\ref{projectie}) will satisfy the conditions which uniquely define $F_k$. We hence conclude that
\[
F_k(\underline{x} \grave{},\underline{y} \grave{})=\mathbb{P}_0^k\left(\frac{(2\pi)^n}{k!2^{n-k}(n-k)!}\langle \underline{x} \grave{},\underline{y} \grave{}\rangle^k\right).
\]
This can be calculated using the explicit form of the projection operators in equation (\ref{projectie}) and the fact that $\nabla_f^{2j} \langle \underline{x} \grave{},\underline{y} \grave{}\rangle^k = \frac{k!}{(k-2j)!}\langle \underline{x} \grave{},\underline{y} \grave{}\rangle^{k-2j}\theta_{\underline{y} \grave{}}^{2j}$, yielding
\begin{eqnarray*}
&&\mathbb{P}_0^k\left(\frac{(2\pi)^n}{k!2^{n-k}(n-k)!}\langle \underline{x} \grave{},\underline{y} \grave{}\rangle^k\right)\\
&=&\sum_{j=0}^{\lfloor \frac{k}{2}\rfloor}\frac{(2\pi)^n}{k!2^{n-k}(n-k)!}\frac{(n-k+1)!}{4^jj!(n-k+1+j)!}\frac{k!}{(k-2j)!}\theta^{2j}\theta_{\underline{y} \grave{}}^{2j}\langle \underline{x} \grave{},\underline{y} \grave{}\rangle^{k-2j}\\
&=&\sum_{j=0}^{\lfloor \frac{k}{2}\rfloor}2^{k-2j}\pi^n\frac{(n+1-k)}{(k-2j)!j!(n+1+j-k)!} \langle \underline{x} \grave{} , \underline{y} \grave{}\rangle^{k-2j}\theta^{2j}\theta_{\underline{y} \grave{}}^{2j},
\end{eqnarray*}
which is the proposed formula.
\end{proof}
The formula for $F_{k}$ given in theorem \ref{reprkern} can be seen as a dimensional continuation of the purely bosonic case in formula (\ref{bosreprkern}). The quotient of Gamma functions in the explicit expression of the Gegenbauer polynomials in (\ref{GegenbauerCoeffs}) (see Appendix) can be replaced by a Pochhammer symbol. This allows to define the Gegenbauer polynomials for $\alpha<-1/2$. Inspired by equation \eqref{bosreprkern} we start from $C^{(-n-1)}(t)$ and calculate
\begin{eqnarray*}
C^{(-n-1)}_{k}(t) &=& \sum_{j=0}^{\lfloor k/2\rfloor}\frac{(-1)^j(-n-1)_{k-j}}{j!(k-2j)!}(2t)^{k-2j}\\
&=&\sum_{j=0}^{\lfloor k/2\rfloor}\frac{(-1)^j(-n-1)(-n)\cdots(-n-2+k-j)}{j!(k-2j)!}(2t)^{k-2j}\\
&=&\sum_{j=0}^{\lfloor k/2\rfloor}\frac{(-1)^k(n+1)(n)\cdots(n+2-k+j)}{j!(k-2j)!}(2t)^{k-2j}\\
&=&(-1)^k(n+1)!\sum_{j=0}^{\lfloor k/2\rfloor}\frac{1}{(k-2j)!j!(n+1+j-k)!}(2t)^{k-2j}.
\end{eqnarray*}
Comparison with (\ref{reprkernfermformula}) then gives
\[
F_k(\underline{x} \grave{},\underline{y} \grave{})= \pi^{n}(-1)^k\frac{n+1-k}{(n+1)!} \left(\sqrt{\theta^{2}\theta_{\underline{y} \grave{}}^{2}}\right)^{k} C^{(-n-1)}_k\left(\frac{\langle \underline{x} \grave{} , \underline{y} \grave{}\rangle}{\sqrt{\theta^{2}\theta_{\underline{y} \grave{}}^{2}} }\right).
\]
\begin{remark}
It is interesting to note that the reproducing kernel for spaces of harmonics is always expressed by Gegenbauer polynomials, in the case of orthogonal symmetry, finite reflection group symmetry as well as symplectic symmetry.
\end{remark}
Using the expression for $F_k$, we obtain the following simplification of the Mehler formula
\begin{eqnarray*}
&&\sum_{k=1}^n\sum_{j=1}^{n-k}(-1)^jj!(n-k-j)! L_j^{k-n-1}(\theta^2) e^{i \alpha (2j+k)}L_j^{k-n-1}(\theta_{\underline{y} \grave{}}^2 ) F_k(\underline{x} \grave{},\underline{y} \grave{})\\
&=& \left(\pi (1- e^{2i \alpha})\right)^{n} \exp{\frac{ 2 e^{i \alpha} \langle \underline{x} \grave{} , \underline{y} \grave{} \rangle - e^{2i \alpha}(\theta^2 +\theta_{\underline{y} \grave{}}^2)}{1- e^{2i \alpha}}}.
\end{eqnarray*}
In the limit case $\alpha = \pm \pi/2$ (corresponding with the classical Fourier transform) this formula reduces to
\begin{eqnarray*}
&&\sum_{j,k}j!(n-k-j)! L_j^{k-n-1}(\theta^2) (\pm i)^kL_j^{k-n-1}(\theta_{\underline{y} \grave{}}^2 ) F_k(\underline{x} \grave{},\underline{y} \grave{}) \exp(-\frac{\theta^2+\theta_{\underline{y} \grave{}}^2}{2})\\
&&= \left(2\pi \right)^{n} \exp{ \pm i \langle \underline{x} \grave{} , \underline{y} \grave{} \rangle}.
\end{eqnarray*}
\begin{remark}
By making use of the expression for the reproducing kernel, we immediately see that both the left-hand and right-hand side in the Mehler formula are invariant under the symplectic group $Sp(2n)$, acting simultaneously on $\underline{x} \grave{}$ and $\underline{y} \grave{}$. So we have indeed constructed a symplectic analog of the $O(m)$-invariant Mehler formula.
\end{remark}
\section{Hermite polynomials for $O(m)\times Sp(2n)$ symmetry}
\setcounter{equation}{0}
\label{supersuper}
\subsection{Orthogonality of spherical Hermite polynomials and inner products}
\label{innerprodsSuper}
We start by introducing the canonical inner product on $L_2(\mathbb{R}^m)\otimes \Lambda_{2n}$ by combining the standard bosonic $L_2$ inner product and the fermionic inner product from definition \ref{inprodGrass}.
\begin{definition}
\label{Canonischin}
The inner product $\langle .|. \rangle_1: L_2(\mathbb{R}^{m})_{m|2n}\times L_2(\mathbb{R}^{m})_{m|2n}\to \mathbb{C}$ is given by
\[
\langle f | g\rangle_1 = \int_{\mathbb{R}^{m|2n}} f (* \overline{g})
\]
where the star map acts on $\Lambda_{2n}$ as in definition \ref{defstar} and leaves the bosonic variables invariant.
\end{definition}
This inner product is uniquely determined by demanding that the creation and annihilation operators of the harmonic oscillator are mutually adjoint. This is the subject of the following theorem.
\begin{theorem}
The inner product $\langle . | . \rangle_1$ is, up to a multiplicative constant, the unique inner product on $\mathcal{P} \exp(-R^2/2)$ such that $(a_i^\pm)^{\dagger}=a_i^\mp$ and $(b_j^\pm)^{\dagger}=b_j^\mp$, with $a_j$ and $b_j$ the creation and annihilation operators appearing in the hamiltonian of the harmonic oscillator (\ref{hamiltoniaan}).
\end{theorem}
\begin{proof}
The condition $(a_i^\pm)^{\dagger}=a_i^\mp$ is equivalent to $x_j^\dagger=x_j$ and $\partial_{x_j}^\dagger=-\partial_{x_j}$. Demanding that $(b_j^{\pm})^{\dagger}=b_j^\mp$ for every $j$ is equivalent to demanding $\partial_{{x \grave{}}_{j}}^\dagger={x \grave{}}_{j}/2$. These conditions are clearly fulfilled for $\langle . | . \rangle_1$. Because these conditions determine the fermionic inner product (lemma \ref{adjoints}) and the bosonic inner product completely this also holds for the full inner product $\langle . | . \rangle_1$.
\end{proof}
As a consequence of this theorem, it is easy to compute that the set of functions
\begin{eqnarray*}
\psi_{k_{1}, \ldots, k_{m}; l_{1}, \ldots, l_{2n}}&=& \frac{1}{\sqrt{ k_{1}! \ldots k_{m}! \pi^{M/2}}}\\
&&\times (a_1^+)^{k_1}(a_2^+)^{k_2}\ldots (a_m^+)^{k_m} (b_1^+)^{l_1}\ldots (b_{2n}^+)^{l_{2n}}\exp (-R^{2}/2),
\end{eqnarray*}
with $k_{i} \in \mathbb{N}$ and $l_{j} \in \{0,1\}$ is an orthonormal basis of $\mathcal{P} \exp(-R^2/2)$, i.e.
\[
\langle \psi_{k_{1}, \ldots, k_{m}; l_{1}, \ldots, l_{2n}}, \psi_{p_{1}, \ldots, p_{m}; q_{1}, \ldots, q_{2n}}\rangle_{1} = \delta_{k_{1} p_{1}} \ldots \delta_{k_{m} p_{m}} \delta_{l_{1} q_{1}} \ldots \delta_{l_{2n} q_{2n}}.
\]
This basis should be considered as the super-analog of the cartesian basis $\{\psi_{k_{1}, \ldots, k_{m}}^{b}\}$ introduced in the case of $O(m)$ symmetry.
The inner product $\langle . | . \rangle_1$ also has several undesirable properties, namely
\begin{itemize}
\item $\nabla^2$ and $R^{2}$ are neither self-adjoint nor mutually adjoint
\item the spherical Hermite functions are in general not orthogonal
\item contrary to the purely fermionic case, the star map does not preserve harmonicity.
\end{itemize}
Indeed, using corollary \ref{hermtoeg} we find $(R^2)^\dagger=r^2-\nabla^2_f$ and $(\nabla^2)^\dagger=\nabla^2_b-\theta^2$. This means for instance that
\begin{equation}
\label{toegDelta}
(-\nabla^2 -R^2 +2\mathbb{E} +M)^\dagger =(-\nabla^2_b -r^2+\theta^2+\nabla^2_f -2\mathbb{E}_b -m+2\mathbb{E}_f -2n).
\end{equation}
This implies that the procedure, used in theorem \ref{orthocliffherm} to prove the orthogonality of the spherical Hermite functions, is no longer possible. Moreover, an easy example shows that the spherical Hermite functions are indeed not orthogonal.
\begin{example}
We take the spherical harmonics $H_2=2x_1^2-{x\grave{}}_1{x\grave{}}_2$, $H_0=1$ and form $L_1^{\frac{M}{2}-1}(R^2)H_0=-R^2+M/2$. By using equation (\ref{CHexp}) for the purely bosonic and fermionic case we find
\begin{equation*}
(-\nabla^2_b -r^2+\theta^2+\nabla^2_f -2\mathbb{E}_b -m+2\mathbb{E}_f -2n)\exp(-R^2/2)=\exp(-R^2/2)(-\nabla^2_b+\nabla^2_f ).
\end{equation*}
Using this and equations \eqref{CliffordHermiteFunctionssuper}, (\ref{toegDelta}) and equation \eqref{CHexp} we obtain
\begin{eqnarray*}
&&\langle H_2\exp (-R^2/2) | 4L_1^{\frac{M}{2}-1}(R^2)H_0\exp (-R^2/2)\rangle_1 \\
&=& \langle H_2\exp (-R^2/2) | (-\nabla^2 -R^2+2\mathbb{E}+M)\exp (-R^2/2)\rangle_1\\
&=&\langle [(-\nabla^2_b+\nabla^2_f) H_2]\exp (-R^2/2) |\exp (-R^2/2)\rangle_1\\
&=&\langle -8\exp(-R^2/2)|\exp (-R^2/2)\rangle_1.
\end{eqnarray*}
Since $\langle \exp(-R^2/2)|\exp (-R^2/2)\rangle_1 \neq 0$, these functions are not orthogonal.
\end{example}
However, it is still true that the hamiltonian $H$ of the harmonic oscillator is hermitian. The spherical Hermite functions are eigenvectors of this operator, see (\ref{HOCH}). Hence, the spherical Hermite functions belonging to different eigenvalues are still orthogonal. In the example, we see that both $H_2\exp(-R^2/2)$ and $L_1^{\frac{M}{2}-1}(R^2)H_0\exp(-R^2/2)$ have eigenvalue $M/2+2$.
We can make this partial orthogonality even stronger.
\begin{lemma}
\label{bijnaorth}
With $\{ H_{k}^{b(l)} \}$ the basis in formula \eqref{bosharmbasis} for $\mathcal{H}_k^b$ and $\{ H_p^{f(q)} \}$ the basis in formula \eqref{OrthSphHarm} for $\mathcal{H}_{p}^{f}$, one has
\[
\langle L_j^{\frac{M}{2}+2i+k+p-1}(R^2)f_{i,k,p}H_{k}^{b(l)}H_p^{f(q)}\exp(-R^2/2)| L_s^{\frac{M}{2}+2u+t+v-1}(R^2)f_{u,t,v}H_{t}^{b(w)}H_v^{f(z)}\exp(-R^2/2) \rangle_{1} = 0
\]
unless $k=t$, $p=v$, $l=w$, $q=z$ and $j+i=s+u$.
\end{lemma}
\begin{proof}
Recall that the inner product $\langle . |. \rangle_{1}$ is a combination of the known bosonic and fermionic inner product. Hence, using
corollary \ref{orthsphharmferm} and the orthogonality of (bosonic) spherical harmonics over the unit sphere, it follows that the integral is zero unless $k=t$, $p=v$, $l=w$ and $q=z$. The self-adjointness of the hamiltonian of the harmonic oscillator then implies $2j+2i+k+p=2s+2u+t+v$, from which we obtain $j+i=s+u$.
\end{proof}
As a corollary we obtain the following decomposition of $\mathcal{P} \exp(-R^2/2)$.
\begin{corollary}
With respect to $\langle\cdot|\cdot\rangle_1$, the space $\mathcal{P} \exp(-R^2/2)$ decomposes in mutually orthogonal subspaces of dimension at most $n+1$, spanned by the spherical Hermite functions.
\end{corollary}
\begin{proof}
This follows from lemma \ref{bijnaorth} together with the fact that for given $H_k^b\in\mathcal{H}_k^b$ and $H_p^{f} \in \mathcal{H}_{p}^{f}$ there are at most $n+1$ polynomials $f_{i,k,p}$ (see theorem \ref{decompintoirreps}).
\end{proof}
Finally, another problem with the inner product $\langle . |. \rangle_{1}$ is that, contrary to the purely fermionic case (see lemma \ref{expstar} $(ii)$), $* H_k \exp (-R^2/2)$ with $H_{k} \in \mathcal{H}_{k}$ will not always be an element of $\mathcal{H}_k \exp (-R^2/2)$. This problem occurs for spherical harmonics of the form $f_{k,p,q}H_p^bH_q^f$ with $k>0$ (see theorem \ref{decompintoirreps}). It is however possible to introduce a new inner product on $\mathcal{P}\exp(-R^2/2)$, which makes the spherical Hermite functions orthogonal and also solves the other problems.
From now on, we will always assume $M>0$. At the end of the section we will explicitly show that the following constructions are not possible for $M\le 0$ (see theorem \ref{Nogonegdim}). The assumption $M>0$ means that we have a Fischer decomposition $(\ref{superFischer})$ and that the spherical Hermite polynomials constitute a basis for $\mathcal{P}$.
The orthogonality of the bosonic spherical Hermite polynomials depends on two facts. The spherical harmonics are orthogonal with respect to integration over the unit sphere and the radial part (given by Laguerre polynomials) is orthogonal with respect to the radial integration. Therefore, our first aim is to construct an inner product for the spherical harmonics using integration over the supersphere. We start with a few technical lemmas.
\begin{lemma}
\label{SSin1}
For the polynomials $f_{k,p,q}$ introduced in lemma \ref{polythm}, $H_p^b \in \mathcal{H}_p^b$ and $H_q^f \in \mathcal{H}_q^f$, the following algebraic relation holds
\begin{eqnarray*}
f_{k,p,q}(r^2,\theta^2)H_p^bH_q^f&\equiv&(-1)^ka_{k,p,q}r^{2k}H_p^bH_q^f\mod R^2
\end{eqnarray*}
with $a_{k,p,q}=\frac{\Gamma(M/2+p+q+2k-1)}{\Gamma(M/2+p+q+k-1)}$.
\end{lemma}
\begin{proof}
To calculate $a_{k,p,q}$ explicitly we start from
\[
f_{k,p,q}H_p^bH_q^f \equiv \sum_{s=0}^ka_sr^{2k-2s}(-r^2)^{s}H_p^bH_q^f\mod R^2
\]
which leads to
\[
a_{k,p,q}=\frac{\Gamma(\frac{m}{2}+p+k)}{(n-q-k)!}\sum_{s=0}^k(-1)^{k-s}\binom{k}{s}\frac{(n-q-s)!}{\Gamma(\frac{m}{2}+p+k-s)}
\]
and the result follows from lemma \ref{SSin2}.
\end{proof}
\begin{lemma}
\label{SSin2}
For $\nu\in\mathbb{N}$, $\mu\in\mathbb{R}$ and $\mu>\nu$ the following relation holds
\begin{eqnarray*}
\sum_{s=0}^k(-1)^{k-s}\binom{k}{s}\frac{(\nu-s)!}{\Gamma(\mu+k-s)}&=&\frac{\Gamma(\mu-\nu+2k-1)}{\Gamma(\mu-\nu+k-1)}\frac{(\nu-k)!}{\Gamma(\mu+k)}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
We denote $c^k_{\mu,\nu}=\sum_{s=0}^k(-1)^{k-s}\binom{k}{s}\frac{(\nu-s)!}{\Gamma(\mu+k-s)}$. Then $c_{\mu,\nu}^0=\nu!/\Gamma(\mu)$ holds and for $k=1$ we find
\begin{eqnarray*}
c^1_{\mu,\nu}&=&-\frac{\nu!}{\Gamma(\mu+1)}+\frac{(\nu-1)!}{\Gamma(\mu)}=\frac{(\nu-1)!}{\Gamma(\mu+1)}\left(\mu-\nu \right)\\
&=&\frac{\Gamma(\mu-\nu+1)}{\Gamma(\mu-\nu)}\frac{(\nu-1)!}{\Gamma(\mu+1)}.
\end{eqnarray*}
From the definition of $c_{\mu,\nu}^k$ we calculate
\begin{eqnarray*}
c_{\mu,\nu}^k&=&\sum_{s=0}^{k-1}(-1)^{k-s}\binom{k-1}{s}\frac{(\nu-s)!}{\Gamma(\mu+k-s)}+\sum_{s=1}^k(-1)^{k-s}\binom{k-1}{s-1}\frac{(\nu-s)!}{\Gamma(\mu+k-s)}\\
&=&\sum_{s=0}^{k-1}(-1)^{k-s}\binom{k-1}{s}\frac{(\nu-s)!}{\Gamma((\mu+1)+k-1-s)}\\
&& -\sum_{s=0}^{k-1}(-1)^{k-s}\binom{k-1}{s}\frac{(\nu-s-1)!}{\Gamma(\mu+k-s-1)}\\
&=&-c_{\mu+1,\nu}^{k-1}+c^{k-1}_{\mu,\nu-1}.
\end{eqnarray*}
Since $\mu-\nu+1>0$, this implies that, if the lemma holds for $k-1$, then
\begin{eqnarray*}
c_{\mu,\nu}^k&=&-\frac{\Gamma(\mu-\nu+2k-2)}{\Gamma(\mu-\nu+k-1)}\frac{(\nu-k+1)!}{\Gamma(\mu+k)}\\
&& +\frac{\Gamma(\mu-\nu+2k-2)}{\Gamma(\mu-\nu+k-1)}\frac{(\nu-k)!}{\Gamma(\mu+k-1)}\\
&=&\frac{\Gamma(\mu-\nu+2k-2)}{\Gamma(\mu-\nu+k-1)}\frac{(\nu-k)!}{\Gamma(\mu+k)}\left(\mu+k-1-(\nu-k+1)\right)\\
&=&\frac{\Gamma(\mu-\nu+2k-1)}{\Gamma(\mu-\nu+k-1)}\frac{(\nu-k)!}{\Gamma(\mu+k)},
\end{eqnarray*}
so the lemma is proven by induction.
\end{proof}
We will also need the following result.
\begin{lemma}
\label{SSin3}
For every $k\in\mathbb{N}$ and $\alpha\in\mathbb{R}$
\[
\sum_{s=0}^k(-1)^s\binom{k}{s}(\alpha-s)_{k}=k!
\]
with $(a)_k=a(a+1)\cdots(a+k-1)$ the Pochhammer symbol.
\end{lemma}
\begin{proof}
We prove this lemma again by induction. For $k=1$ the lemma is trivial. We will use lemma 5 in \cite{DBThesis} which states
\begin{eqnarray}
\sum_{s=0}^k(-1)^s\binom{k}{s}(\alpha-s)_{k-1}=0.
\label{lemThesis}
\end{eqnarray}
From the definition we calculate, using (\ref{lemThesis}),
\begin{eqnarray*}
\sum_{s=0}^k(-1)^s\binom{k}{s}(\alpha-s)_{k}&=& \sum_{s=0}^k(-1)^s\binom{k}{s}(\alpha-s+k-1)(\alpha-s)_{k-1}\\
&=&- \sum_{s=1}^k(-1)^ss\binom{k}{s}(\alpha-s)_{k-1}\\
&=&\sum_{s=0}^{k-1}(-1)^s\frac{(s+1) k!}{(k-s-1)!(s+1)!}(\alpha-s-1)_{k-1}\\
&=&k\sum_{s=0}^{k-1}(-1)^s\binom{k-1}{s}((\alpha-1)-s)_{k-1},
\end{eqnarray*}
so the lemma follows by induction.
\end{proof}
Now we have all the necessary tools to prove the following lemma.
\begin{lemma}
\label{SSin4}
For $\{H_p^{b(l)}\}$ the orthonormal basis of $\mathcal{H}_{p}^{b}$ in formula \eqref{bosharmbasis}, $\{H_q^{f(t)}\}$ the orthonormal basis of $\mathcal{H}_{q}^{f}$ in equation \eqref{OrthSphHarm} and $f_{k,p,q}$ as defined in lemma \ref{polythm}, the following relation holds
\begin{eqnarray*}
&&\int_{SS}f_{k,p,q}H_p^{b(l_1)}H_q^{f(t_1)}\, f_{k,p,q}H_p^{b(l_2)}\widetilde{H}_q^{f(t_2)}\\
&=&(-1)^k\frac{ k!\,\Gamma(\frac{m}{2}+p+k)}{(\frac{M}{2}+p+q+2k-1)\Gamma(\frac{M}{2}+p+q+k-1)(n-q-k)!}\delta_{l_1l_2}\delta_{t_1t_2}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
Lemma \ref{SSin1} implies there exists a polynomial $P$ of degree $2k+p+q-2$, such that $f_{k,p,q}H_p^bH_q^f=(-1)^ka_{k,p,q} r^{2k}H_p^bH_q^f+R^2 P$. Since $M>0$, $P$ has a Fischer decomposition containing only spherical harmonics of degree smaller than or equal to $2k+p+q-2$. The orthogonality of spherical harmonics of different degree and the property $\int_{SS}R^2f=\int_{SS}f$ then imply
\begin{eqnarray*}
&&\int_{SS}f_{k,p,q}H_p^{b(l_1)}H_q^{f(t_1)}\, f_{k,p,q}H_p^{b(l_2)}\widetilde{H}_q^{f(t_2)}\\
&=&(-1)^ka_{k,p,q}\int_{SS}r^{2k} f_{k,p,q}H_p^{b(l_1)}H_p^{b(l_2)} H_q^{f(t_1)}\widetilde{H}_q^{f(t_2)}\\
&=&(-1)^ka_{k,p,q}\sum_{s=0}^ka_s\int_{SS}r^{4k-2s} \theta^{2s}H_p^{b(l_1)}H_p^{b(l_2)} H_q^{f(t_1)}\widetilde{H}_q^{f(t_2)}.
\end{eqnarray*}
Using expression \eqref{SSintnieuw} for the supersphere integration, we then obtain
\begin{eqnarray*}
&=&(-1)^ka_{k,p,q}\sum_{s=0}^ka_s\int_{\mathbb{S}^{m-1}} d \sigma \int_B\frac{(-1)^{n-q-s}\theta^{2n-2q-2s}}{(n-q-s)!} \\
&&\times \left[\left(\frac{\partial}{\partial r^2}\right)^{n-q-s}r^{4k-2s+m-2} \theta^{2s}H_p^{b(l_1)}H_p^{b(l_2)} H_q^{f(t_1)}\widetilde{H}_q^{f(t_2)}\right]_{r=1}\\
&=&(-1)^ka_{k,p,q}\sum_{s=0}^k(-1)^sa_s\delta_{l_1l_2}\frac{\delta_{t_1t_2}}{(n-q-s)!}\left[\left(\frac{\partial}{\partial r^2}\right)^{n-q-s}r^{2(2k-s+m/2+p-1)}\right]_{r=1}\\
&=&(-1)^k\delta_{l_1l_2}\delta_{t_1t_2}a_{k,p,q}\sum_{s=0}^k(-1)^s\frac{a_s}{{(n-q-s)!}}\frac{\Gamma(2k-s+\frac{m}{2}+p)}{\Gamma(2k-s+\frac{m}{2}+p-n+q+s)}\\
&=&(-1)^k\delta_{l_1l_2}\delta_{t_1t_2}\frac{a_{k,p,q}}{{\Gamma(2k+\frac{M}{2}+p+q)}}\frac{\Gamma(\frac{m}{2}+p+k)}{(n-q-k)!}\\
&&\times \sum_{s=0}^k(-1)^s\binom{k}{s}\frac{\Gamma(2k-s+\frac{m}{2}+p)}{\Gamma(\frac{m}{2}+p+k-s)}\\
&=&(-1)^k\delta_{l_1l_2}\delta_{t_1t_2}\,a_{k,p,q}\,b_{k,p,q}.
\end{eqnarray*}
Lemma \ref{SSin3} implies
\begin{equation}
\label{bkpq}
b_{k,p,q}=\frac{k!}{{\Gamma(2k+\frac{M}{2}+p+q)}}\frac{\Gamma(\frac{m}{2}+p+k)}{(n-q-k)!}.
\end{equation}
Substituting $a_{k,p,q}$ from lemma \ref{SSin1} yields the desired result.
\end{proof}
Now we introduce the following isomorphism $T: \mathcal{H}_{k} \rightarrow \mathcal{H}_{k}$ given by
\[
T(H_k)=(-1)^{i}f_{i,k-2i-j,j}(r^2,\theta^2) H^b_{k-2i-j}\widetilde{H}^f_j,
\]
for $H_{k} = f_{i,k-2i-j,j}(r^2,\theta^2) H^b_{k-2i-j}H^f_j$ with $H^b_{k-2i-j} \in \mathcal{H}^b_{k-2i-j}$, $H^f_j \in \mathcal{H}^f_j$ and extended by linearity to the whole of $\mathcal{H}_{k}$.
\begin{remark}
\label{TOSp}
Using lemma \ref{Sptilde} it is straightforward to prove that for $A\in O(m)\times Sp(2n)$
\[
T\circ A = (G^TAG)\circ T
\]
holds, with $G$ given in equation \eqref{metricG}.
\end{remark}
Using the isomorphism $T$ we can construct a new inner product on $\bigoplus_{k=0}^{\infty} \mathcal{H}_k \exp (-R^2/2)$. This is the subject of the following theorem.
\begin{theorem}
\label{defsuperH}
Put $\mathcal{H} = \bigoplus_{k=0}^{\infty} \mathcal{H}_k = \ker{\nabla^2} \, \cap \, \mathcal{P}$.
The product $\langle .|.\rangle: \mathcal{H} \exp (-R^2/2) \times \mathcal{H} \exp (-R^2/2)\to \mathbb{C}$ given by
\[
\langle H_1\exp( -R^2/2)| H_2\exp( -R^2/2)\rangle =\int_{\mathbb{R}^{m|2n}}H_1\overline{T({H}_2)}\exp (-R^2)
\]
with $H_{1}, H_{2} \in \mathcal{H}$ is an inner product on $\mathcal{H} \exp (-R^2/2)$ satisfying
\[
\langle \mathcal{H}_k \exp (-R^2/2) , \mathcal{H}_l \exp (-R^2/2)\rangle = 0
\]
if $k\not= l$.
\end{theorem}
\begin{proof}
The product is clearly linear. It is positive definite since
\[
\langle\left(f_{i,p,q} \mathcal{H}^b_{p} \otimes \mathcal{H}^f_{q}\right) \exp (-R^2/2) ,\left(f_{j,r,s} \mathcal{H}^b_{r} \otimes \mathcal{H}^f_{s}\right) \exp (-R^2/2)\rangle = 0
\]
when $(i,p,q) \neq (j,r,s)$ (see theorem \ref{integorth}) and $\langle f_{i,p,q}H_p^bH_q^f|f_{i,p,q}H_p^bH_q^f\rangle>0$ (see lemma \ref{SSin4}). Moreover, lemma \ref{SSin4} also implies the inner product is symmetric.
\end{proof}
This inner product can now be extended to the whole space using decomposition (\ref{superFischer}).
\begin{theorem}
\label{defsuper}
The product $\langle .|. \rangle_{2}: \mathcal{P} \exp (-R^2/2)\times \mathcal{P} \exp (-R^2/2)\to \mathbb{C}$ given by
\[
\langle R^{2i}H_k\exp (-R^2/2) | R^{2j}H_l\exp (-R^2/2)\rangle_2 = \int_{\mathbb{R}^{m|2n}} R^{2i+2j} H_k \, \overline{T({H}_l)} \, \exp (-R^2)
\]
with $H_{k} \in \mathcal{H}_{k}$, $H_{l} \in \mathcal{H}_{l}$ and extended by linearity is an inner product.
\end{theorem}
\begin{proof}
The product is clearly linear and symmetric. Using (\ref{superint2}) and theorem \ref{defsuperH} we subsequently obtain
\begin{eqnarray*}
&&\langle R^{2i}H_k\exp (-R^2/2) | R^{2j}H_k\exp (-R^2/2)\rangle_{2}\\
&=&\frac{1}{2}\Gamma \left(\frac{i+j+2k+M}{2}\right)\int_{SS}H_k \overline{T({H}_l)}\\
&=&\frac{\Gamma \left(\frac{i+j+2k+M}{2}\right)}{\Gamma \left(\frac{2k+M}{2}\right)} \int_{\mathbb{R}^{m|2n}} H_k \, \overline{T({H}_l)} \, \exp (-R^2)\\
&\geq&0.
\end{eqnarray*}
Hence $\langle.|. \rangle_{2}$ is positive definite and defines an inner product.
\end{proof}
The behavior of the inner product with respect to $O(m)\times Sp(2n)$ is given by
\begin{lemma}
For $A\in O(m)\times Sp(2n)$ and for $f,g\in\mathcal{P}\exp(-R^2/2)$, with action on superfunctions as defined in formula \ref{actieOSp}, the relation
\[
\langle Af|g\rangle=\langle f|A^Tg\rangle
\]
holds. This implies $A^\dagger=A^T$ for all $A\in O(m)\times Sp(2n)$ and is equivalent with
\[
\langle AG f|GAg\rangle_2=\langle f|g\rangle_2,
\]
with $G$ given in equation \eqref{metricG}.
\end{lemma}
\begin{proof}
The proof is similar to the proof of proposition \ref{Spinprod}. Remark \ref{TOSp} and the $O(m)\times Sp(2n)$-invariance of $\int_{\mathbb{R}^{m|2n}}$ (see lemma \ref{Berezincov}) lead to the proposed formula.
\end{proof}
The inner product of theorem \ref{defsuper} can be written more concisely as
\[
\langle f | g \rangle_{2} = \int_{\mathbb{R}^{m|2n}} f \overline{T(g)}, \qquad f, g \in \mathcal{P} \exp(-R^{2}/2)
\]
by extending $T$ to $\mathcal{P}\exp(-R^2/2)$ such that
\begin{equation*}
T( R^{2j}H_l\exp (-R^2/2))=R^{2j}T({H}_l)\exp (-R^2/2), \quad H_{l} \in \mathcal{H}_{l}.
\end{equation*}
So $T$ is now a map $\mathcal{P}\exp(-R^2/2)\to\mathcal{P}\exp(-R^2/2)$ satisfying
\begin{equation*}
T(L_j^{\frac{M}{2}+k-1}(R^2)H_k^{(l)}\exp(-R^2/2))= L_j^{\frac{M}{2}+k-1}(R^2)T({H}_k^{(l)})\exp(-R^2/2).
\end{equation*}
It is clear by construction that $T$ preserves harmonicity. Now we can show that $\langle .|.\rangle_2 $ also satisfies the other desirable properties that $\langle.|.\rangle_1$ did not satisfy. We start by proving that $R^2$ and $\nabla^2$ are self-adjoint.
\begin{lemma}
\label{adjointsnew}
One has
\begin{eqnarray*}
(R^2)^{\dagger}&=&R^2\\
(\nabla^2)^{\dagger}&=&\nabla^2\\
(2\mathbb{E}+M)^{\dagger}&=&-(2\mathbb{E}+M)
\end{eqnarray*}
with respect to the inner product $\langle. |.\rangle_2$ on $\mathcal{P}\exp (-R^2/2)$.
\end{lemma}
\begin{proof}
The proof of the first property is trivial. The second property is obtained as follows. Because of lemma \ref{Berezin} we know that
\[
\int_{\mathbb{R}^{m|2n}} f \nabla^2 g=\int_{\mathbb{R}^{m|2n}} \nabla^2(f) g.
\]
Using lemma \ref{laplonpieces} and equation (\ref{berekexpdelta}) we obtain that for every piece of the Fischer decomposition
\begin{eqnarray*}
&&T(\nabla^2 R^{2j}H_k\exp(-R^2/2))\\
&=&T([c_{1,j,k} R^{2j-2}H_k+R^{2j+2}H_k+(4j+2k+M)R^{2j}H_k]\exp(-R^2/2))\\
&=&[c_{1,j,k} R^{2j-2}+R^{2j+2}+(4j+2k+M)R^{2j}]T(H_k)\exp(-R^2/2)\\
&=&\nabla^2 R^{2j}T(H_k)\exp(-R^2/2)\\
&=&\nabla^2 T( R^{2j} H_k\exp(-R^2/2)),
\end{eqnarray*}
with $H_{k} \in \mathcal{H}_{k}$.
Combining these two results yields the second property.
Finally, the result for $2\mathbb{E} + M$ follows immediately from equation \eqref{commsuper}.
\end{proof}
To show that the spherical Hermite functions are orthogonal, we first need an orthogonal basis of spherical harmonics. The knowledge of orthonormal bases for the bosonic and fermionic harmonic polynomials suffices to find an orthonormal basis $\{ H_{k}^{(l)}\}$ for the space $\mathcal{H}_{k}$, satisfying
\begin{equation}
\label{orthbasissuper}
\int_{SS}H_k^{(l)}T({H}_q^{(r)})=\delta_{kq}\delta_{lr}.
\end{equation}
Indeed, using the bases in equations \eqref{bosharmbasis} and \eqref{OrthSphHarm}, it is easy to check that
\begin{eqnarray*}
&&\left\{\frac{f_{i,k-2i-j,j}H_{k-2i-j}^{b(l)}H_{j}^{f(t)}}{\sqrt{a_{i,k-2i-j,j} b_{i,k-2i-j,j}}}|0\le j\le \min(n,k-1)-1,\, \right. \\
&&\left. 0\le i\le \min(n-j,\lfloor\frac{k-j}{2}\rfloor)\right\}
\end{eqnarray*}
with $l\le\dim\mathcal{H}_{k-2i-j}^b$ and $t\le \dim\mathcal{H}_j^f$ is an orthonormal basis for $\mathcal{H}_k$.
For this basis of spherical harmonics $\{H_k^{(l)}\}$ the following theorem holds.
\begin{theorem}[Orthogonality spherical Hermite functions]
\label{orthsuperCH}
The set of functions $\{ \varphi_{j,k,l} \}$ in formula \eqref{CliffordHermiteFunctionssuper}, constructed using the basis of spherical harmonics in formula \eqref{orthbasissuper}, forms an orthogonal basis for $\mathcal{P}\, exp(-R^{2}/2)$ with respect to the inner product $\langle .|.\rangle_2$. The normalization is given by
\begin{equation}
\langle \varphi_{j,k,l}| \varphi_{p,q,r} \rangle_2 =\frac{1}{2}4^{2j}j!\Gamma (j+k + M/2)\delta_{jp}\delta_{kq}\delta_{lr}=\gamma^M_{j,k}\delta_{jp}\delta_{kq}\delta_{lr}.
\label{normCH}
\end{equation}
\end{theorem}
\begin{proof}
This result is proven with the same technique as in theorem \ref{orthocliffherm}, using the results we obtained in lemma \ref{adjointsnew}.
\end{proof}
Note that the normalization constants only depend op the super-dimension $M$ and not on the bosonic and fermionic dimensions separately. In particular this implies that the normalization constants are equal to the case with $M$ bosonic variables (and no anti-commuting variables).
Both the inner products $\langle . | . \rangle_1$ and $\langle . | . \rangle_2$ have their own advantages. The hamiltonian of the harmonic oscillator $\frac{1}{2}(-\nabla^2+R^2)$ is hermitian for both. For the inner product $\langle . | . \rangle_1$ we even have $(a_i^{\pm})^\dagger=a_i^\mp$ and $(b_j^\pm)^\dagger=b_j^{\mp}$ or equivalently $x_i^\dagger=x_i$ and $\partial_{{x \grave{}}_{j}}^\dagger={x \grave{}}_{j}/2$. However, for the inner product $\langle . | . \rangle_2$ we have $(R^2)^\dagger= R^2$ and $(\nabla^2)^\dagger= \nabla^2$. This is of major importance to study other potentials in superspace, such as anharmonic oscillators. Using $\langle . | . \rangle_2$, they still have symmetric hamiltonians. That is why we will study the extension of $\langle.|.\rangle_2$ from $\mathcal{P} \exp (-R^{2}/2)$ to the super Schwartz and $L_2$-space in a forthcoming article.
We end this section with two no-go results. First, it is not possible to construct an inner product which has the advantages of both inner products $\langle . | . \rangle_1$ and $\langle . | . \rangle_2$. This is the subject of the following theorem.
\begin{theorem}
\label{Nogo1}
There is no inner product on $\mathcal{P}\exp(-R^2/2)$ for which $(R^2)^\dagger=R^2$ and $(r^2)^\dagger=r^2$.
\end{theorem}
\begin{proof}
$(R^2)^\dagger=R^2$ and $(r^2)^\dagger=r^2$ imply that $(\theta^2)^\dagger=\theta^2$. This is impossible because of the same reason as in theorem \ref{onmogelijkinprod}.
\end{proof}
We have only found an inner product with the property that $(R^2)^\dagger=R^2$ and $(\nabla^2)^\dagger=\nabla^2$ for the case $M>0$. Now we will show that such an inner product does not exist in case $M\le 0$. We also prove that (for all $M$) there does not exist an inner product with the properties of the purely fermionic inner product (see corollary \ref{hermtoeg}) when $m\not=0$.
\begin{theorem}
\label{Nogonegdim}
There is no inner product on $\mathcal{P}\exp (-R^2/2)$ for which $(R^2)^\dagger=R^2$ and $(\nabla^2)^\dagger=\nabla^2 $ in case $M\le 0$.
There also does not exist an inner product with the properties $(R^2)^\dagger=-\nabla^2$ and $(\nabla^2)^\dagger=-R^2$ for arbitrary $M$ with $m\not=0$.
\end{theorem}
\begin{proof}
If $(R^2)^\dagger=R^2$ and $(\nabla^2)^\dagger=\nabla^2 $, using equation \eqref{commsuper} we obtain that $(2\mathbb{E}+M)^\dagger=\frac{1}{2}([\nabla^2, R^2])^\dagger=-(2\mathbb{E}+M)$. So assume there exists an inner product satisfying these properties for $M\le 0$. Then we can calculate $\langle \varphi_{j,k,l}|\varphi_{j,k,l}\rangle $ using (\ref{CHexp}) and (\ref{differentiaalvgl}), yielding
\begin{eqnarray*}
\langle \varphi_{j,k,l}|\varphi_{j,k,l}\rangle&=&\langle (-R^2-\nabla^2+2\mathbb{E}+M)\varphi_{j-1,k,l} |\varphi_{j,k,l} \rangle\\
&=&\langle \varphi_{j-1,k,l} |(-R^2-\nabla^2-2\mathbb{E}-M)\varphi_{j,k,l} \rangle\\
&=&8j(2j+M+2k-2)\langle \varphi_{j-1,k,l}|\varphi_{j-1,k,l}\rangle.
\end{eqnarray*}
Now in the case $M<0$ we can choose $k=0$ and $j$ small enough, but larger than zero, to make the factor $2j+M-2$ negative. This means that either $\langle \varphi_{j,0,1}|\varphi_{j,0,1}\rangle$ or $\langle \varphi_{j-1,0,1}|\varphi_{j-1,0,1}\rangle$ is negative, proving that the inner product is not positive definite. If $M=0$ we find that $\langle \varphi_{1,0,1}|\varphi_{1,0,1}\rangle=0$, proving again that the inner product is not positive definite.
If, on the other hand, $(R^2)^\dagger=-\nabla^2$ and $(\nabla^2)^\dagger=-R^2$, we know that $(2\mathbb{E}+M)^\dagger=\frac{1}{2}([\nabla^2, R^2])^\dagger=(2\mathbb{E}+M)$ and we obtain in a similar fashion
\[
\langle \varphi_{j,k,l}|\varphi_{j,k,l}\rangle=-8j(2j+M+2k-2)\langle \varphi_{j-1,k,l}|\varphi_{j-1,k,l}\rangle.
\]
Now for the case where $j$ and $k$ are big enough we find that either $\langle \varphi_{j,k,l}|\varphi_{j,k,l}\rangle$ or $\langle \varphi_{j-1,k,l}|\varphi_{j-1,k,l}\rangle$ has to be negative.
\end{proof}
\subsection{Mehler formula}
We are now able to establish a Mehler formula for the super spherical Hermite polynomials when $M>0$.
We start from the basis $\{ \varphi_{j,k,l} \}$ considered in theorem \ref{orthsuperCH}. Normalizing this basis, according to formula (\ref{normCH}), yields
\[
\phi_{j,k,l} = \varphi_{j,k,l} / \sqrt{\gamma^M_{j,k}},
\]
which is an orthonormal basis with respect to $\langle\cdot|\cdot\rangle_2$. Using the integral expression for the inner product $\langle .|.\rangle_2$, the basis $\{\phi_{j,k,l}\}$ and the general fractional Fourier transform (see theorem \ref{fracfourthm}) we obtain that, formally,
\begin{eqnarray*}
&&\sum_{j,k,l} T(\phi_{j,k,l})(\bold{x}) \,e^{i \alpha (2j+k)} \,\phi_{j,k,l}(\bold{y})\\ &=& \left(\pi (1- e^{2i \alpha})\right)^{-M/2} \exp{\frac{ 4 e^{i \alpha} \langle \bold{x},\bold{y} \rangle - (1+ e^{2i \alpha})(R^2 + R_{\bold{y}}^2)}{2- 2e^{2i \alpha}}}.
\end{eqnarray*}
Using the explicit expression for the spherical Hermite polynomials in terms of Laguerre polynomials (see (\ref{CHLaguerre})) and the normalization (\ref{normCH}) we then find
\begin{eqnarray*}
&&\sum_{j,k,l} \frac{2j! e^{i \alpha (2j+k)}}{\Gamma(j+\frac{M}{2}+k)} L_j^{\frac{M}{2}+k-1}(R^2)L_j^{\frac{M}{2}+k-1}(R_{\bold{y}}^2) T(H_{k}^{(l)})(\bold{x}) H_{k}^{(l)}(\bold{y}) \\
&=& \left(\pi (1- e^{2i \alpha})\right)^{-M/2} \exp{\frac{ 2 e^{i \alpha} \langle \bold{x},\bold{y} \rangle - e^{2i \alpha}(R^2 + R_{\bold{y}}^2)}{1- e^{2i \alpha}}}.
\end{eqnarray*}
This equation can be simplified again for $M>1$ with the explicit calculation of the summation $F_k(\bold{x},\bold{y})=\sum_lT({H}_k^l(\bold{x}))H_k^l(\bold{y})$. These functions $F_k$ were already constructed in \cite{DBS5} as a consequence of a Funk-Hecke theorem on the supersphere.
\begin{lemma}[Reproducing kernel]
\label{superrepr}
Let $M>1$. Then
\[
G_{k}(\bold{x},\bold{y}) = \frac{2k+M-2}{M-2} \,\frac{\Gamma(M/2)}{2 \pi^{M/2}} \, C^{(M-2)/2}_k (\langle \bold{x},\bold{y} \rangle)
\]
with $C^{(M-2)/2}_k$ a Gegenbauer polynomial, is a reproducing kernel for the space $\mathcal{H}_k$, i.e.
\[
\int_{SS,x} H_l(\bold{x}) G_{k}(\bold{x},\bold{y}) = \delta_{kl} H_l(\bold{y}) \mbox{ mod }(R_{\bold{y}}^2-1), \quad \mbox{for all } H_l \in \mathcal{H}_l.
\]
\end{lemma}
\begin{proof}
See \cite{DBS5}, corollary 5. We have rewritten the Legendre polynomials used there in terms of Gegenbauer polynomials.
\end{proof}
Making this function homogeneous, using equation (\ref{superint2}) and the fact that $\int_{SS}R^2P=\int_{SS}P$ yields
\begin{eqnarray*}
& &\int_{\mathbb{R}^{m|2n}} H_l(\bold{x})\frac{2k+M-2}{M-2} \frac{\Gamma(M/2)}{2 \pi^{M/2}} (R^2R_{\bold{y}}^2)^{(k/2)} C^{(M-2)/2}_k \left(\frac{\langle \bold{x},\bold{y} \rangle}{(R^2R_{\bold{y}}^2)^{(1/2)}} \right) \exp (-R^2)\\
&=&\frac{\Gamma((k+l+M)/2)} {2}\int_{SS,x} H_l(\bold{x}) F_{k}(\bold{x},\frac{\bold{y}}{\sqrt{R_{\bold{y}}^2}})(R_{\bold{y}}^2)^{k/2} \\
&=&\delta_{kl}\frac{\Gamma(k+M/2)} {2}H_l(\bold{y}).
\end{eqnarray*}
Hence we conclude
\begin{eqnarray*}
F_k(\bold{x},\bold{y}) &=& \sum_{l=1}^{\dim \mathcal{H}_{k}} T({H}_k^{(l)}(\bold{x}))H_k^{(l)}(\bold{y})\\
&=&\frac{2k+M-2}{M-2} \frac{\Gamma(M/2)}{2 \pi^{M/2}} (R^2R_{\bold{y}}^2)^{(k/2)} C^{(M-2)/2}_k \left(\frac{\langle \bold{x},\bold{y} \rangle}{(R^2R_{\bold{y}}^2)^{(1/2)}} \right)
\end{eqnarray*}
for a basis of $\mathcal{H}_k$ as in equation (\ref{orthbasissuper}). Putting everything together then yields the following Mehler formula in superspace
\begin{eqnarray}
\nonumber
&&\sum_{j,k}\frac{2j! e^{i \alpha (2j+k)}}{\Gamma(j+\frac{M}{2}+k)} L_j^{\frac{M}{2}+k-1}(R^2)L_j^{\frac{M}{2}+k-1}(R_{\bold{y}}^2) F_k (\bold{x},\bold{y})
\\
&=& \left(\pi (1- e^{2i \alpha})\right)^{-\frac{M}{2}} \exp{\frac{ 2 e^{i \alpha} \langle \bold{x},\bold{y} \rangle - e^{2i \alpha}(R^2 + R_{\bold{y}}^2)}{1- e^{2i \alpha}}}.
\label{superMehler2}
\end{eqnarray}
So far, we have only established this formula formally. We now show that we have actual convergence.
\begin{theorem}
For $\underline{x},\underline{y}\in\mathbb{R}^m$, $\alpha\in\mathbb{R}$ and $M>1$, the series in equation (\ref{superMehler2}) converges pointwise.
\end{theorem}
\begin{proof}
Both sides in equation \eqref{superMehler2} are functions of $\langle \bold{x},\bold{y}\rangle$, $R^2$ and $R_{\bold{y}}^2$. This equation can therefore be written as
\[
\sum_{j,k}g_{j,k}(\langle \bold{x},\bold{y}\rangle,R^2,R_{\bold{y}}^2)=g(\langle \bold{x},\bold{y}\rangle,R^2,R_{\bold{y}}^2).
\]
As the functions $g_{j,k}$ are polynomials in $\langle \bold{x},\bold{y}\rangle$, $R^2$ and $R_{\bold{y}}^2$, they can be written as a Taylor expansion in the anti-commuting variables
\begin{eqnarray*}
g_{j,k}(\langle \bold{x},\bold{y}\rangle,R^2,R_{\bold{y}}^2)&=&\sum_{p,q=0}^n\sum_{r=0}^{2n}\frac{\theta^{2p}\theta_{\underline{y} \grave{}}^{2q}\langle \underline{x} \grave{},\underline{y} \grave{}\rangle^r}{p!q!r!}\\
&&\times (\frac{\partial}{\partial a^2})^p(\frac{\partial}{\partial b^2})^q(\frac{\partial}{\partial c})^rg_{j,k}(c,a^2,b^2)
\end{eqnarray*}
with $a^2=r^2$, $b^2=r_{\underline{y}}^2$ and $c=\langle\underline{x},\underline{y}\rangle$. This also holds for
\begin{eqnarray*}
g(\langle \bold{x},\bold{y}\rangle,R^2,R_{\bold{y}}^2)&=&\left(\pi (1- e^{2i \alpha})\right)^{-\frac{M}{2}} \exp{\frac{ 2 e^{i \alpha} \langle \bold{x},\bold{y} \rangle - e^{2i \alpha}(R^2 + R_{\bold{y}}^2)}{1- e^{2i \alpha}}}\\
&=&\left(\pi (1- e^{2i \alpha})\right)^{-\frac{M}{2}} \exp{\frac{ 2 e^{i \alpha}c -e^{2i \alpha}(a^2 + b^2)}{1- e^{2i \alpha}}} \\
&&\times \exp{\frac{ 2 e^{i \alpha} \langle \underline{x} \grave{} , \underline{y} \grave{} \rangle - e^{2i \alpha}(\theta^2 + \theta_{\underline{y} \grave{}}^2)}{1- e^{2i \alpha}}}.
\end{eqnarray*}
From corollary \ref{OmMehler3} we find
\[
\sum_{j,k}g_{j,k}(c,a^2,b^2)=g(c,a^2,b^2).
\]
Remark \ref{Mehlerafleiden} implies that arbitrary derivatives with respect to $c$, $a^2$ and $b^2$ can be brought inside the summation with the convergence still holding. So we can take the Taylor series of both sides and the theorem is proven.
\end{proof}
\section{Conclusions and summary}
\setcounter{equation}{0}
\label{conclSummary}
In this paper, we have given a detailed treatment of Hermite type polynomials related to three different symmetries: $O(m)$ (orthogonal symmetry), $\mathcal{G} < O(m)$ (finite reflection group symmetry) and $O(m) \times Sp(2n)$ (superspace symmetry). In each case, it was possible to define two types of Hermite polynomials, namely cartesian Hermite polynomials and spherical Hermite polynomials. In the cases of $O(m)$, $\mathcal{G}$ and $Sp(2n)$, both types of polynomials turned out to be orthogonal with respect to the canonical inner product. In the full superspace case of $O(m) \times Sp(2n)$, we obtained that the spherical Hermite polynomials are not orthogonal with respect to the canonical inner product. We presented a detailed analysis of this lack of orthogonality. Then we gave a construction of a new inner product, which restores the orthogonality of the spherical Hermite polynomials but destroys the orthogonality of the cartesian Hermite polynomials.
We have summarized all these results in two tables. In Table \ref{default}, we give an extensive overview of the different types of symmetry and compare their analogies and differences. For the superspace case, the purely fermionic ($Sp(2n)$) case is given in a separate column. In Table \ref{default2} we restrict ourselves to the full superspace case and give a comparison between the two inner products that we have considered. We give the adjoints of the relevant operators as well as the differences in orthogonality of the two types of Hermite polynomials.
As already mentioned, the inner product we have constructed in section \ref{innerprodsSuper} has only been defined for a weighted space of polynomials. It is possible to extend this inner product to broader function spaces, such as $\mathcal{S}(\mathbb{R}^{m})_{m|2n}$. As this is a technical matter requiring subtle estimates, we postpone this to a subsequent paper (see \cite{CDBHilbert}).
The results obtained in this paper allow to study several other interesting problems in the future.
First of all, as we now have a new inner product on superspace that makes hamiltonians of the type
\[
H = -\frac{1}{2}\nabla^2 + V(R^{2})
\]
self-adjoint, we can make a thorough study of such systems and the related Schr\"odinger equations. This should allow to shed new light on the results obtained in e.g. \cite{MR967935,ZHANG} and to put them together in one theory.
In \cite{DBS3} we also introduced a second class of polynomials in superspace, namely the spherical Gegenbauer polynomials. It is expected that the results obtained here in combination with the new view on supersphere integration given in \cite{CDBS1} will allow to obtain orthogonality properties of these polynomials on the superball.
Next, the obtained Mehler formulas in superspace should allow to study the radial behavior of the super Fourier transform and its relation with the classical Hankel transform. More generally, we have now the tools available to make a complete study in superspace of the holomorphic semigroup
\[
e^{z (-\nabla^2 +R^{2})}, \quad z \in \mathbb{C}, \Re z \geq 0
\]
in the sense of \cite{Folland,Howe}. Note that the choice $z= i \pi/4$ leads to the super Fourier transform (see formula (\ref{FTSuperExp})).
Another important direction for further research is in the context of radial deformations. The study of such deformations has recently arisen as a new and exciting topic in harmonic analysis. In the orthogonal situation, one special radial deformation has been studied in \cite{MR2134314,MR2401813}. The Dunkl case has been considered in \cite{Orsted2}. There, the authors introduce a radial deformation parameter in the $\mathfrak{sl}_{2}$ relations satisfied by the Dunkl Laplacian and make a detailed study of the related analysis. Very recently, an even more general radial deformation in the context of Dirac operators has been realized in \cite{DBOrsted}.
It is expected that also the superspace representation of $\mathfrak{sl}_{2}$ (as well as its Dirac counterpart given by $\mathfrak{osp}(1|2)$) can be radially deformed, i.e. that it would be possible to replace the super Laplace operator $\nabla^2$ and $R^{2}$ by radially deformed operators in such a way that the $\mathfrak{sl}_{2}$ relations are preserved. It would be very interesting to see to what extent the theory of radial deformations can be established in the setting of superspaces and whether the Hermite polynomials related to these new deformations have similar orthogonality properties as established in this paper.
Finally, Table \ref{default} suggests that there is a type of symmetry missing in the current scheme. This is the case of a realization of
$\mathfrak{sl}_{2}$ which is only invariant under a (finite) subgroup of $Sp(2n)$, thus establishing the symplectic counterpart of the theory of Dunkl operators. It is at this point not entirely clear whether such a deformation is feasible in all generality in the framework of superspaces, but it would give a very satisfying unifying picture.
Note that in the special case of a superspace with $2n$ commuting and $2n$ anti-commuting variables, so with invariance $O(2n) \times Sp(2n)$, one has already established analogs of various Calogero-Sutherland systems (see e.g. \cite{MR1070940, MR1608453, MR2025382}). Although the hamiltonians considered in those papers don't contain the fermionic Laplace operator $\nabla^{2}_{f}$, contrary to formula (\ref{SchrodEq}), this still hints at possible generalizations to the non-supersymmetric case.
\begin{sidewaystable}[htdp]
\caption{Summary - 4 types of symmetry}
\begin{center}
\scalebox{0.75}{\begin{tabular}{c||c|c|c|c}
Symmetry&$O(m)$& $\mathcal{G} < O(m)$& $Sp(2n)$&$O(m) \times Sp(2n)$\\ \vspace{-3mm} &&&& \\
&orthogonal&finite reflection group& symplectic& full superspace\\ \vspace{-3mm} &&&& \\ \hline \hline \vspace{-3mm} &&&& \\
Basic function space& $L_{2}(\mathbb{R}^{m})$&$L_{2}(\mathbb{R}^{m}, w_{\kappa}(\underline{x}) dV(\underline{x}))$& $\Lambda_{2n}$ & $L_{2}(\mathbb{R}^{m})\otimes \Lambda_{2n}$\\ \vspace{-3mm} &&&& \\
Generators of $\mathfrak{sl}_{2}$& $\nabla^2_{b}$&$\Delta_{\kappa}$&$\nabla^2_{f}$& $\nabla^2 = \nabla^2_{b} + \nabla^2_{f}$\\ \vspace{-3mm} &&&& \\
&$r^{2}$&$r^{2}$&$\theta^{2}$&$R^{2} = \theta^{2}+r^{2}$\\ \vspace{-3mm} &&&& \\
& $\mathbb{E}_{b} + \frac{m}{2}$& $\mathbb{E}_{b} + \frac{\mu}{2}$ & $\mathbb{E}_{f} -n $&$\mathbb{E} + \frac{m-2n}{2}$\\ \vspace{-3mm} &&&& \\
Dimension& $m$&$\mu = m + 2\sum_{\alpha \in R_+} \kappa_{\alpha}$&$-2n$&$M = m -2n$\\ \vspace{-3mm} &&&& \\
Spaces of harmonics & $\mathcal{H}_{k}^{b} =\ker{\nabla^2_{b}} \cap Pol_{k}$ &$\mathcal{H}_{k}^{\mathcal{D}} =\ker{\Delta_{\kappa}} \cap Pol_{k}$ & $\mathcal{H}_{k}^{f} =\ker{\nabla^2_{f}} \cap \Lambda_{2n}^{k}$ & $\mathcal{H}_{k} =\ker{\nabla^2} \cap \mathcal{P}_{k}$\\ \vspace{-3mm} &&&& \\
Reproducing kernel & $C^{(m-2)/2}_{k}$ & $V_{\kappa}(C^{(\mu-2)/2}_{k})$& $C^{(-n-1)}_{k}$& $C^{(M-2)/2}_{k}$\\ \vspace{-3mm} &&&& \\ \hline \vspace{-3mm} &&&& \\
Related quantum system& $\frac{1}{2}(-\nabla^2_{b} + r^{2}) \psi = E \psi
$ & $\frac{1}{2}(-\Delta_{\kappa} + r^{2}) \psi = E \psi
$& $\frac{1}{2}(-\nabla^2_{f}+\theta^2) \psi = E \psi $ & $\frac{1}{2}(-\nabla^2+R^2) \psi = E \psi $\\ \vspace{-3mm} &&&& \\
& PDE&PDE + difference terms& system of algebraic equations& system of PDEs\\ \vspace{-3mm} &&&& \\
Fourier transform & $e^{ \frac{i \pi m}{4}} e^{\frac{i \pi}{4}(\nabla^2_{b} - r^{2})}$
&$e^{ \frac{i \pi \mu}{4}} e^{\frac{i \pi}{4}(\Delta_{\kappa} - r^{2})}$&$e^{-\frac{i \pi n}{2} } e^{ \frac{i \pi }{4}(\nabla^2_{f} -\theta^2)}$ &$e^{\frac{ i \pi M}{4} } e^{\frac{i \pi}{4}(\nabla^2 -R^2)}$ \\ \vspace{-3mm} &&&& \\
&$(2 \pi)^{-\frac{m}{2}} \int_{\mathbb{R}^{m}} e^{-i\langle \underline{x},\underline{y}\rangle} f(\underline{x}) dV(\underline{x})$&$ c_{\kappa}^{-1} \int_{\mathbb{R}^m} D(\underline{x},-i\underline{y}) f(\underline{x}) w_{\kappa}(\underline{x}) dV(\underline{x})$& $(2 \pi)^{n} \int_{B,x} \exp{(- i \langle \underline{x} \grave{} , \underline{y} \grave{} \rangle)}f(\bold{x})$& $(2 \pi)^{-\frac{M}{2}} \int_{\mathbb{R}^{m|2n}} \exp{(- i \langle \bold{x},\bold{y} \rangle)}f(\bold{x})$\\ \vspace{-3mm} &&&& \\
&$\langle \underline{x},\underline{y}\rangle = \sum_{i=1}^{m}x_{i}y_{i}$& $D(\underline{x},-i\underline{y})$ in general unknown&$\langle \underline{x} \grave{},\underline{y} \grave{}\rangle = -\frac{1}{2} \sum_{j=1}^{n}({x \grave{}}_{2j-1}{y \grave{}}_{2j} - {x \grave{}}_{2j} {y \grave{}}_{2j-1})$ &$\langle \bold{x},\bold{y} \rangle = \langle \underline{x},\underline{y}\rangle + \langle \underline{x} \grave{},\underline{y} \grave{}\rangle $\\ \vspace{-3mm} &&&& \\
Integration & Lebesgue integral&weighted Lebesgue integral& Berezin integral $\int_{B}$& $\int_{\mathbb{R}^{m|2n}} = \int_{B} \int_{\mathbb{R}^{m}}dV(\underline{x})$ \\ \vspace{-3mm} &&&& \\ \hline \vspace{-3mm} &&&& \\
Cartesian Hermite functions & $\psi_{k_{1}, \ldots, k_{m}}^{b}$ & $\psi_\nu^{\mathcal{D}}$ & $\psi_{ l_{1}, \ldots, l_{2n}}^{f}$& $\psi_{k_{1}, \ldots, k_{m}; l_{1}, \ldots, l_{2n}}$\\ \vspace{-3mm} &&&& \\
Energy & $\frac{m}{2} + \sum_{i=1}^{m}k_{i}$& $\frac{\mu}{2} + |\nu|$& $-n + \sum_{i=1}^{2n}l_{i}$& $\frac{M}{2} + \sum_{i=1}^{m}k_{i} + \sum_{i=1}^{2n}l_{i}$ \\ \vspace{-3mm} &&&& \\
Spherical Hermite functions & $\phi_{j,k,l}^{b}$ & $\phi_{j,k,l}^{\mathcal{D}}$& $\phi_{j,k,l}^{f}$ & $\phi_{j,k,l}$ \\ \vspace{-3mm} &&&& \\
Energy & $\frac{m}{2} + (2j + k)$&$\frac{\mu}{2} + (2j + k)$&$-n + (2j + k)$&$\frac{M}{2} + (2j + k)$ \\ \vspace{-3mm} &&&& \\ \hline \vspace{-3mm} &&&& \\
Canonical inner product & $\langle f, g \rangle_{L_{2}} = \int_{\mathbb{R}^{m}} f \overline{g} dV(\underline{x})$ &$\langle f, g \rangle_{L_{2}} = \int_{\mathbb{R}^{m}} f \overline{g} w_{\kappa}(\underline{x})dV(\underline{x})$& $\langle f|g \rangle_{\Lambda_{2n}} = \int_{B} f (\ast \overline{g})$& $\langle f|g \rangle_{1} =\int_{\mathbb{R}^{m|2n}} f (* \overline{g}) $ \\ \vspace{-3mm} &&&& \\
Orth. cartesian Hermites &$\langle \psi_{k_{1}, \ldots, k_{m}}^{b} ,\psi_{l_{1}, \ldots, l_{m}}^{b} \rangle_{L_{2}}$ & $\langle \psi_\nu^{\mathcal{D}} ,\psi_\mu^{\mathcal{D}} \rangle_{L_{2}}$&$\langle \psi_{l_{1}, \ldots, l_{2n}}^{f}, \psi_{q_{1}, \ldots, q_{2n}}^{f}\rangle_{\Lambda_{2n}}$&$\langle \psi_{k_{1}, \ldots, k_{m}; l_{1}, \ldots, l_{2n}}, \psi_{p_{1}, \ldots, p_{m}; q_{1}, \ldots, q_{2n}}\rangle_{1}$\\ \vspace{-3mm} &&&& \\
& $ = \delta_{k_{1} l_{1}} \ldots \delta_{k_{m} l_{m}}$& $= \delta_{\mu \nu}$&$ = \delta_{l_{1} q_{1}} \ldots \delta_{l_{2n} q_{2n}}$&$ = \delta_{k_{1} p_{1}} \ldots \delta_{k_{m} p_{m}} \delta_{l_{1} q_{1}} \ldots \delta_{l_{2n} q_{2n}}$\\ \vspace{-3mm} &&&& \\ \hline \vspace{-3mm} &&&& \\
Orth. spherical Hermites &$\langle \phi_{j_{1},k_{1},l_{1}}^{b} , \phi_{j_{2},k_{2},l_{2}}^{b}\rangle_{L_{2}}$& $\langle \phi_{j_{1},k_{1},l_{1}}^{\mathcal{D}} , \phi_{j_{2},k_{2},l_{2}}^{\mathcal{D}}\rangle_{L_{2}} $& $\langle \phi_{j_{1},k_{1},l_{1}}^{f} , \phi_{j_{2},k_{2},l_{2}}^{f}\rangle_{\Lambda_{2n}}$& $\langle \phi_{j_{1},k_{1},l_{1}} , \phi_{j_{2},k_{2},l_{2}}\rangle_{1} $\\ \vspace{-3mm} &&&& \\
&$ = \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}$& $= \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}$& $= \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}$&${\bf\neq} \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}$\\ \vspace{-3mm} &&&& \\
\end{tabular}}
\end{center}
\label{default}
\end{sidewaystable}%
\begin{sidewaystable}[htdp]
\caption{Inner products in the full superspace case ($M >0$)}
\begin{center}
\begin{tabular}{c||c|c}
Inner product & $\langle f|g \rangle_{1} =\int_{\mathbb{R}^{m|2n}} f (* \overline{g}) $& $\langle f|g \rangle_{2} =\int_{\mathbb{R}^{m|2n}} f \overline{T(g)} $\\
\vspace{-3mm}&&\\ \hline \hline
\vspace{-3mm}&&\\
Function space & $L_{2}(\mathbb{R}^{m})\otimes \Lambda_{2n}$ & $\mathcal{P} \exp(-R^{2}/2)$ \\
\vspace{-3mm}&&\\ \hline
\vspace{-3mm}&&\\
Adjoints: &&\\ \vspace{-3mm}&&\\
$(a_{i}^{\pm})^{\dagger}$ &$a_{i}^{\mp}$ & ? \\\vspace{-3mm}&&\\
$(b_{i}^{\pm})^{\dagger}$ &$b_{i}^{\mp}$ & ? \\\vspace{-3mm}&&\\
$(\nabla^2)^{\dagger} $& $\nabla^2_{b}-\theta^{2}$ & $\nabla^2$ \\\vspace{-3mm}&&\\
$(R^{2})^{\dagger} $& $r^{2} -\nabla^2_{f}$ & $R^{2}$ \\\vspace{-3mm}&&\\
$(\mathbb{E} + \frac{M}{2})^{\dagger} $& $\mathbb{E}_{f} -\mathbb{E}_{b} -\frac{m}{2} -n$ &$-(\mathbb{E} + \frac{M}{2})$\\\vspace{-3mm}&&\\
$H^{\dagger} = \frac{1}{2}(-\nabla^2+R^2)^{\dagger} $ &$\frac{1}{2}(-\nabla^2+R^2)$ & $\frac{1}{2}(-\nabla^2+R^2)$ \\ \vspace{-3mm}&&\\ \vspace{-3mm}&&\\ \hline\vspace{-3mm}&&\\
Orth. cartesian Hermites & $\langle \psi_{k_{1}, \ldots, k_{m}; l_{1}, \ldots, l_{2n}}, \psi_{p_{1}, \ldots, p_{m}; q_{1}, \ldots, q_{2n}}\rangle_{1}$ & $\langle \psi_{k_{1}, \ldots, k_{m}; l_{1}, \ldots, l_{2n}}, \psi_{p_{1}, \ldots, p_{m}; q_{1}, \ldots, q_{2n}}\rangle_{2}$\\
&$ = \delta_{k_{1} p_{1}} \ldots \delta_{k_{m} p_{m}} \delta_{l_{1} q_{1}} \ldots \delta_{l_{2n} q_{2n}}$&$ \neq \delta_{k_{1} p_{1}} \ldots \delta_{k_{m} p_{m}} \delta_{l_{1} q_{1}} \ldots \delta_{l_{2n} q_{2n}}$\\ \vspace{-3mm}&&\\ \hline \vspace{-3mm}&&\\
Orth. spherical Hermites &$\langle \phi_{j_{1},k_{1},l_{1}} , \phi_{j_{2},k_{2},l_{2}}\rangle_{1} $ &$\langle \phi_{j_{1},k_{1},l_{1}} , \phi_{j_{2},k_{2},l_{2}}\rangle_{2} $ \\
&${\bf\neq} \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}$& $ = \delta_{j_{1} j_{2}} \delta_{k_{1} k_{2}} \delta_{l_{1} l_{2}}$\\
\end{tabular}
\end{center}
\label{default2}
\end{sidewaystable}
\section{Appendix}
\setcounter{equation}{0}
The Hermite polynomials $H_k$ for $k\in\mathbb{N}$ are defined by their Rodrigues formula as $H_{k}(t) = (-1)^{k} \exp{(t^2)} \frac{d^k}{dt^k} \exp{(-t^2)}$ and are given explicitly by
\begin{eqnarray*}
H_k(t)&=&\sum_{j=0}^{\lfloor k/2\rfloor}(-1)^j\frac{2^{k-2j}k!}{(k-2j)!j!}t^{k-2j}.
\end{eqnarray*}
They satisfy the orthogonality relation
\begin{eqnarray*}
\int_{-\infty}^\infty H_k(t)H_l(t)\exp(-t^2)dt&=&\delta_{kl}k!2^k\sqrt{\pi}.
\end{eqnarray*}
The generalized Laguerre polynomials $L_k^{(\alpha)}$ for $k\in\mathbb{N}$ are defined as
\begin{eqnarray}
L_k^{(\alpha)}(t)&=&\sum_{j=0}^{k}\frac{\Gamma(k+\alpha+1)}{j!(k-j)!\Gamma(j+\alpha+1)}(-t)^j
\label{genLagdef}
\end{eqnarray}
and satisfy the orthogonality relation (when $\alpha>-1$)
\begin{eqnarray*}
\int_{0}^\infty t^{\alpha} L^{(\alpha)}_k(t)L^{(\alpha)}_l(t)\exp(-t)dt&=&\delta_{kl}\frac{\Gamma(k+\alpha+1)}{k!}.
\end{eqnarray*}
The Hermite polynomials can be expressed in terms of the generalized Laguerre polynomials by
\begin{eqnarray*}
H_{2k}(t)=(-1)^k2^{2k}k!L_k^{(-\frac{1}{2})}(t^2) &\mbox{and}&H_{2k+1}(t)=(-1)^k2^{2k+1}k!tL_k^{(\frac{1}{2})}(t^2).
\end{eqnarray*}
The Gegenbauer polynomials $C^{(\alpha)}_k(t)$ are a special case of the Jacobi polynomials. For $k\in\mathbb{N}$ and $\alpha>-1/2$ they are defined as
\begin{eqnarray}
C_k^{(\alpha)}(t)&=&\sum_{j=0}^{\lfloor k/2\rfloor}(-1)^j\frac{\Gamma(k-j+\alpha)}{\Gamma(\alpha)j!(k-2j)!}(2t)^{k-2j}
\label{GegenbauerCoeffs}
\end{eqnarray}
and satisfy the orthogonality relation
\begin{eqnarray*}
\int_{-1}^1C_k^{(\alpha)}(t)C_l^{(\alpha)}(t)(1-t^2)^{\alpha-\frac{1}{2}}dt&=&\delta_{kl}\frac{\pi2^{1-2\alpha}\Gamma(k+2\alpha)}{k!(k+\alpha)(\Gamma(\alpha))^2}.
\end{eqnarray*}
|
\section{Introduction}
This is a continuation of \cite{wy3} in which the spatial limit of the new quasilocal energy defined in \cite{wy1} and \cite{wy2} is analyzed. In the present article, we address the question of the null limit in Bondi--Sachs coordinates for an asymptotically flat spacetime.
Let $N$ be a spacetime with metric $g_{\alpha \beta}$ in Bondi--Sachs coordinates given by
$$
-UVdw^2-2Udwdr+\sigma_{ab}(dx^a+W^adw)(dx^b+W^bdw)\,\,\, a, b=2,3
$$
where
$$ W^a =O(r^{-2}),$$
$$U=1-\frac{X^2+Y^2}{2r^2}+o(r^{-2}),$$
$$V=1-\frac{2m}{r}+o(r^{-1})$$
and the metric $\sigma_{ab}$ is given by
\[ \left( \begin{array}{ccc}
r^2+2Xr+2(X^2+Y^2) & -2Yr \sin \theta \\
-2Yr \sin \theta & \sin^2 \theta[r^2-2Xr+2(X^2+Y^2)]
\end{array} \right)\] with
\[\det \sigma_{ab}=r^4\sin^2\theta.\]
The inverse of the metric $g_{\alpha\beta}$ is
\[g^{ww}=g^{wa}=0, g^{wr}=-U^{-1}, g^{rr}=U^{-1} V, g^{ra}=U^{-1} W^a, \text{\,and\,} g^{ab}=\sigma^{ab}.\]
Throughout the paper, coordinates are labeled by $x^0=w, x^1=r, x^2=\theta, x^3=\phi$ and the indexes are for $\alpha, \beta, \gamma \cdots =0, 1,2 ,3$,
$i, j, k \cdots=1, 2, 3$, and $a, b \cdots=2,3$.
At a retarded time $w=c$, the Bondi--Sachs energy-momentum vector (\cite{bvm} \cite{s}) is defined as
\begin{equation}
(E, P_1, P_2, P_3)=\frac{1}{8 \pi} (\int_{S^2} 2m dS^2, \int_{S^2} 2m\tilde{X}_1dS^2, \int_{S^2} 2m\tilde{X}_2dS^2, \int_{S^2} 2m\tilde{X}_3dS^2)
\end{equation} where $m=m(c, \theta, \phi)$ is the mass aspect function in the expansion of $V$, and
$\tilde{X}_i$, $i=1,2,3$ are the three eigenfunctions $\sin\theta \sin\phi$, $\sin\theta \cos\phi$ and $\cos\theta $ of the Laplace operator $\widetilde{\Delta}$ on $S^2$ with eigenvalue $-2$.
We recall that given a spacelike 2-surface $\Sigma$ in a spacetime, a quasilocal energy $E(\Sigma, X, T_0$) is defined in \cite{wy1}, \cite{wy2} with respect to an isometric embedding $X:\Sigma\rightarrow \mathbb R^{3,1}$ and a constant future timelike vector $T_0\in \mathbb R^{3,1}$. For a family of surfaces $\Sigma_r$ and a family of isometric embeddings $X_r$ of $\Sigma_r$ into $\mathbb R^{3,1}$, the limit of $E(\Sigma_r, X_r, T_0)$ is evaluated in \cite[Theorem 2.1]{wy3} under the assumption that
\begin{equation} \label{ratio}
\lim_{r \to \infty} \frac{|H_0|}{|H|}=1
\end{equation}
where $H$ and $ H_0$ are spacelike mean curvature vectors of $\Sigma_r$ in $N$ and the image of $X_r$ in $\mathbb R^{3,1}$, respectively.
In fact, the limit of $E(\Sigma_r, X_r, T_0)$ with respect to a constant future timelike vector $T_0 \in \mathbb R^{3,1}$ is given by
\begin{equation}\label{linear}
\lim_{r\to \infty}\frac{1}{8\pi} \int_{\Sigma_{r}} \left[-\langle T_0, \frac{{J}_0}{|H_0|}\rangle (|H_0|-|H|) - \langle\nabla^{\mathbb R^{3,1}}_{\nabla\tau} \frac{{J}_0}{|H_0|}, \frac{H_0}{|H_0|}\rangle+ \langle\nabla^{N}_{\nabla\tau} \frac{{J}}{|H|}, \frac{H}{|H|}\rangle \right] d \Sigma_r
\end{equation} where $\tau=-\langle T_0, X_r\rangle$ is the time function with respect to $T_0$, and $J_0$ and $J$ are the future timelike normal vectors dual to $H_0$ and $H$. This expression is linear in $T_0$ and defines an energy-momentum 4-vector at infinity.
In this article, we consider a family of 2-surfaces $\Sigma_r$ on a null cone $w=c$ as $r$ goes to infinity in Bondi--Sachs coordinates.
The limit of the quasilocal energy is first computed with respect to isometric embeddings $X_r$ into $\mathbb R^3$ which are essentially unique and satisfy (\ref{ratio}). We show in particular, \begin{equation}\lim_{r\to \infty}\frac{1}{8\pi} \int_{\Sigma_{r}} (|H_0|-|H|) d\Sigma_r=E,\text{ and } \lim_{r\to \infty}\frac{1}{8\pi} \int_{\Sigma_{r}}\langle\nabla^{N}_{-\nabla X_i} \frac{{J}}{|H|}, \frac{H}{|H|}\rangle d \Sigma_r=P_i\end{equation} where $(X_1, X_2, X_3)$ are the coordinate functions of the isometric embedding $X_r$ into $\mathbb R^3$. We remark that exactly the same limit expression on coordinate spheres of asymptotically flat hypersurface gives the ADM energy-momentum in \cite{wy3}. The computation is stable with respect to any $O(1)$ perturbation of $X_r$ in $\mathbb R^{3,1}$ and is equivariant with respect to Lorentzian transformation acting on $X_r$.
In \cite{wy1} and \cite{wy2}, the quasilocal mass of a 2-surface $\Sigma$ is defined to be the minimum of $E(\Sigma, X, T_0)$ among all admissible pairs $(X, T_0)$ and the Euler-Lagrange equation is derived for an optimal isometric embedding. In the last section, we show that an analytic solution of the optimal isometric embedding equation can be obtained as an $O(1)$ perturbation of embeddings into a boosted totally geodesic slice in $\mathbb R^{3,1}$ whose timelike normal is in the direction of the total energy-momentum 4-vector. This solution locally minimizes the quasilocal energy.
Brown--Lau--York \cite{blk} and Lau \cite{la} compute the null limit of the Brown--York energy and we compare our calculation with theirs in the following:
1) Brown--York mass is gauge dependent. After fixing a reference isometric embedding (either to flat $ \mathbb R^3$ \cite{blk} or to the null cone in $\mathbb R^{3,1}$\cite{la} ), a gauge is chosen arbitrarily so that the limit of the mass coincide with the Bondi mass.
In contrast, in our case, once a reference isometric embedding is picked, the quasilocal energy is determined by the canonical gauge condition (Eq (1.1) in \cite{wy3}). Our calculation is robust with respect to the choice of reference isometric embedding. In particular, the reference family can be arranged to be asymptotically flat or asymptotically null in $\mathbb R^{3,1}$.
2) In \cite{blk}, the momentum part came from the smear energy while in our case, the momentum part came from the connection one-form associated with the mean curvature gauge. This one form gives the right momentum contribution in the asymptotically flat case as well (see \cite{wy3}).
3) In \cite{blk}, the energy and momentum are defined separately. In our case, the Lorentzian symmetric is recovered at infinity and the energy-momentum form a covariant 4-(co)vector. We show that this (co)vector is equivariant with respect to the reference isometric embeddings into $\mathbb R^{3,1}$.
Acknowledgement: Part of the work is done while the authors are visiting the Taida Institute for Mathematical Sciences in Taipei, Taiwan.
\section{The geometry of 2-surface $\Sigma_r$ in Bondi coordinates}
Let $N$ be an asymptotically flat spacetime with Bondi--Sachs coordinates.
Let $\Sigma_r$ be the 2-surface defined by $w=c$ and a fixed $r$. In this section, we compute the mean curvature vector $ H$ of $\Sigma_r$ in $N$ and the connection one-form of the normal bundle of $\Sigma_r$ in the mean curvature gauge. Denote $W_a=\sigma_{ab} W^b$ and let $\delta^a W_a$ be the divergence of the 1-form $W_a$ on $\Sigma_r$ with respect to the induced metric $\sigma_{ab}$.
\begin{lemma}
Let $\Sigma_r$ be the 2-surface defined by $w=c$ and a fixed $r$. The mean curvature vector $H$ of $\Sigma_r$ in $N$ is given by
\begin{equation} \label{eq:H1}
H= \frac{1}{U} \left[ \frac{2}{r} ({\frac{\partial}{\partial w}} -W^{a}{\frac{\partial}{\partial x^a}} ) - (\frac{2V}{r}+ \delta^a W_a ){\frac{\partial}{\partial r}}\right].
\end{equation}
In particular, $H$ is spacelike when $r$ is large enough with
\begin{equation}\label{eq:H2}
|H|^2 = \frac{4}{Ur} (\frac{V}{r} + \delta^a W_a).
\end{equation} Suppose $J$ is the future timelike normal vector dual to $H$, then
\begin{equation}\label{eq:H3} \langle \nabla _{\frac{\partial}{\partial x^b}} {J}, {H} \rangle =
\frac{2}{rU} \partial_{b} (\frac{V}{r}+\delta ^a W_a)-\frac{2}{r U^2}(\frac{V}{r}+{\delta^a W_a})\sigma_{cb}\partial_r W^c.\end{equation}
\end{lemma}
\begin{proof}
By definition, we have
\begin{align*}
H= {} & \sigma^{ab} (\nabla _{\frac{\partial}{\partial x^a}}{\frac{\partial}{\partial x^b}} - (\nabla _{\frac{\partial}{\partial x^a}}{\frac{\partial}{\partial x^b}}) ^{T})\\
= {} & \sigma^{ab} (\Gamma_{ab}^r {\frac{\partial}{\partial r}} + \Gamma_{ab}^w {\frac{\partial}{\partial w}} + (\Gamma_{ab}^c -\langle \nabla _{\frac{\partial}{\partial x^a}}{\frac{\partial}{\partial x^b}},{\frac{\partial}{\partial x^d}}\rangle \sigma^{dc} ){\frac{\partial}{\partial x^c}} ).
\end{align*}
The last coefficient can be computed explicitly as
\begin{align*}
& \Gamma_{ab}^c -\langle \nabla _{\frac{\partial}{\partial x^a}}{\frac{\partial}{\partial x^b}},{\frac{\partial}{\partial x^d}}\rangle \sigma^{dc} \\
={} & \Gamma_{ab}^c -(\Gamma_{ab}^r g_{rd}+\Gamma^{w}_{ab} g_{wd}+\Gamma_{ab}^e\sigma_{ed}) \sigma^{dc}\\
={} & - \Gamma_{ab}^r g_{rd}\sigma^{dc} - \Gamma_{ab} ^w g_{wd} g^{dc} \\
={} & \Gamma_{ab} ^w g_{wr} g^{rc}.
\end{align*}
Thus $$
H= \sigma^{ab} (\Gamma_{ab}^r {\frac{\partial}{\partial r}} + \Gamma_{ab}^w {\frac{\partial}{\partial w}} + \Gamma_{ab} ^w g_{wr} g^{rc} {\frac{\partial}{\partial x^c}} ).
$$
The relevant Christoffel symbols of $g_{\alpha \beta}$ are given by
$$
\Gamma^w_{ab} = \frac{1}{2} U^{-1} \partial_r \sigma _{ab},
$$ and
\[
\Gamma^r_{ab}=- \frac{1}{2} U^{-1} [ \partial_b W_a + \partial_a W_b - \partial_w \sigma _{ab} +V \partial_r \sigma _{ab}- 2 \gamma^d_{ab}W_d] \] where $\gamma_{ab}^d$
is the Christoffel symbol of the metric $\sigma_{ab}$.
When tracing with $\sigma^{ab}$, we notice that
\begin{equation}\label{eq:diff}
\sigma^{ab} \partial_{{\alpha}} \sigma _{ab} = \partial_{{\alpha}} (\ln \det \sigma_{ab}) = \partial_{{\alpha}} \ln (r^4 \sin^2 \theta) .
\end{equation}
Thus, we obtain equations (\ref{eq:H1}) and (\ref{eq:H2}).
To compute the connection one-form, we rewrite equation (\ref{eq:H1}) as
\begin{align*}
U {H} ={} & -(\frac{2V}{r}+{\delta ^a W_a}) \frac{\partial}{\partial r} + \frac{2}{r}( \frac{\partial}{\partial w} -W^c \frac{\partial}{\partial x^c}) \\
= {} & -(\frac{V}{r}+{\delta ^a W_a}) \frac{\partial}{\partial r} + \frac{2}{r}( \frac{\partial}{\partial w} -W^c \frac{\partial}{\partial x^c} -\frac{V}{2} \frac{\partial}{\partial r} )
\end{align*}
where $ \frac{\partial}{\partial r}$ and $ \frac{\partial}{\partial w} -W^c \frac{\partial}{\partial x^c} -\frac{V}{2} \frac{\partial}{\partial r}$ are null vectors.
Thus we have
$$
U {J} = -(\frac{V}{r}+{\delta ^a W_a}) \frac{\partial}{\partial r} - \frac{2}{r}( \frac{\partial}{\partial w} -W^c \frac{\partial}{\partial x^c} -\frac{V}{2} \frac{\partial}{\partial r} ).
$$
For simplicity, let's denote $ \frac{\partial}{\partial r}$ and $ \frac{\partial}{\partial w} -W^c \frac{\partial}{\partial x^c} -\frac{V}{2} \frac{\partial}{\partial r}$ by $ \vec n_1$ and $\vec n_2$ and the coefficients $\frac{V}{r}+\delta ^a W_a$ and $\frac{2}{r}$ by $x$ and $y$ in the following computation. Then,
\begin{equation}\label{JH}\begin{split}
\langle \nabla _{\frac{\partial}{\partial x^b}} J, H \rangle ={} & U^{-2} \langle\nabla _{\frac{\partial}{\partial x^b}} x\vec{n}_1+y\vec{n}_2, x
\vec{n}_1-y\vec{n}_2\rangle\\
={} &U^{-2} \{ [ (\partial_{b}x)(-y) + (\partial_{b}y)(x)] \langle \vec{n}_1, \vec{n}_2 \rangle -xy( \langle \nabla _{\frac{\partial}{\partial x^b}}
\vec{n}_1, \vec{n}_2
\rangle- \langle \nabla _{\frac{\partial}{\partial x^b}} \vec{n}_2, \vec{n}_1 \rangle) \}\\
={} &U^{-2} [ (\partial_{b}x)(-y) \langle \vec{n}_1, \vec{n}_2\rangle -2xy\langle \nabla _{\frac{\partial}{\partial x^b}} \vec{n}_1, \vec{n}_2
\rangle +xy \partial_b \langle \vec{n}_2, \vec{n}_1 \rangle ]
\end{split}\end{equation}
On the other hand,
\[ \langle \nabla _{\frac{\partial}{\partial x^b}} \vec{n}_1, \vec{n}_2 \rangle=\langle \Gamma^r_{br}
\frac{\partial}{\partial r} , \frac{\partial}{\partial w} -W^d \frac{\partial}{\partial x^d}
-\frac{V}{2} \frac{\partial}{\partial r} \rangle= -U\Gamma_{br}^r\] because $\Gamma_{br}^w=0$ by direct computation, and $\frac{\partial}{\partial x^c}$ is perpendicular to the null normal $ \frac{\partial}{\partial w} -W^d \frac{\partial}{\partial x^d}-\frac{V}{2} \frac{\partial}{\partial r} $.
Substitute in $\Gamma_{br}^r=\frac{1}{2} U^{-1}\partial_b U-\frac{1}{2}U^{-1} \sigma_{bc} \partial_r W^c$, and we obtain
\begin{equation} \label{n1n2}\langle \nabla _{\frac{\partial}{\partial x^b}} \vec{n}_1, \vec{n}_2 \rangle=-\frac{1}{2}\partial_b U+\frac{1}{2}\sigma_{bc}\partial_r W^c.\end{equation}
Plug \eqref{n1n2} and $\langle \vec{n}_1,\vec{n}_2\rangle=-U$ into \eqref{JH}, we derive
\begin{align*} \langle \nabla _{\frac{\partial}{\partial x^b}} {J}, {H} \rangle =&
U^{-2}y[(\partial_b x) U-x\sigma_{bc}\partial_r W_c],\end{align*}
and \eqref{eq:H3} follows in view of the definitions of $x$ and $y$.
\end{proof}
\section{Limit of quasilocal energy}
In this section, we compute the limit of quasilocal energy with respect to a family of isometric embeddings $X_r$ of $\Sigma_r$ as an $O(1)$ perturbations of a boosted totally geodesic slice in $\mathbb R^{3,1}$. First we quote the following lemma whose proof can be found in \cite{fst}:
\begin{lemma}\label{isom_R^3}
Let $\sigma^r_{ab}$ be a family of metrics on $\Sigma_r\simeq S^2$ with $\sigma^r_{ab}=r^2\tilde{\sigma}_{ab}+O(r)$ in which $\tilde{\sigma}_{ab}$ is the standard round metric on $S^2$. Let $X_r=(X_1, X_2, X_3)$ be the isometric embedding into $\mathbb R^3$ for $r$ large and $H_0$ be the mean curvature of $X_r$. Then
\[|{H}_0| = \frac{2}{r} +O(r^{-2}) \text{ and }\int_{\Sigma_r} |{H}_0| d\Sigma_r = 4 \pi r + \frac{Area(\Sigma_r)}{r}+O(r^{-1}).
\]
\end{lemma}
We note that up to an isometry of $\mathbb R^3$, $X_r$ can be arranged so that the coordinate functions satisfy $X_i=r\tilde{X}_i+ O(1)$.
\begin{theorem}\label{maintheorem}
Let $\Sigma_r$ be the 2-surface defined by $w=c$ and a fixed $r$ in an asymptotically flat spacetime with Bondi--Sachs coordinates. Suppose $X_r$ is the (unique) family of isometric embeddings of $\Sigma_r$ into $\mathbb R^{3}$ for $r$ large, the limit of quasilocal energy with respect to $T_0=(\sqrt{1+|a|^2}, a^1,a^2,a^3 )$ is
\begin{equation}
\lim_{r\rightarrow\infty} E(\Sigma_r, X_r, T_0)=\frac{1}{8 \pi} \int_{S^2} 2m (\sqrt{1+|a|^2}+ a^i\tilde{X}_i)dS^2.
\end{equation}
\end{theorem}
\begin{proof}
Let $(0,X_1,X_2,X_3)$ be the isometric embedding $X_r$ of $\Sigma_r$ into $\mathbb R^3\subset \mathbb R^{3,1}$. In this case, $\frac{J_0}{|H_0|}$ is simply the vector $(1,0,0,0)$. By the assumption on $\sigma_{ab}$ we can apply Lemma \ref{isom_R^3} and
$$
\int_{\Sigma_r} |H_0| d\Sigma_r = 8 \pi r +O(r^{-1}).
$$ On the other hand, from equation (\ref{eq:H2}) and the expansion for $V$, we obtain
\begin{equation} \label{normH}
|H| = \frac{2}{r} -\frac{2m}{r^2} + \delta^a W_a +O(r^{-3})
\end{equation} and thus
$$
\int_{\Sigma_r} |H| d\Sigma_r = 8 \pi r - \int_{S^2} 2m \, dS^2 + O(r^{-1}).
$$
Next we compute the physical hamiltonian
\[ \frac{1}{8 \pi} \int_{\Sigma_r} \langle \nabla^N_{\nabla \tau} \frac{J}{|H|} , \frac{ H}{|H|} \rangle d\Sigma_r=-a^i \frac{1}{8 \pi} \int_{\Sigma_r} \langle \nabla^N_{\nabla X_i} \frac{J}{|H|} , \frac{ H}{|H|} \rangle d\Sigma_r.\]
From equation (\ref{eq:H3}) and the asymptotic expansions of $V$ and $W^a$ , we derive
\[
\langle \nabla _{\frac{\partial}{\partial x^b}} {J}, {H} \rangle =\frac{2}{r}[\partial_{ b} (\delta ^a W_a-\frac{2m}{r^2})] + \frac{4}{r^3} W_b+O(r^{-4}).\]
Let $V$ denote the connection one-from $\langle \nabla^N \frac{J}{|H|} , \frac{ H}{|H|} \rangle$. From the above computation,
\[ div_{\Sigma_r} V = \frac{1}{r}[\frac{1}{2}(\widetilde{\Delta}+2)({\delta^a}{W_a} -\frac{2m}{r^2}) + \frac{2m}{r^2}] +O(r^{-4}). \]
The limit of $\int_{\Sigma_r} \langle \nabla^N_{\nabla X_i} \frac{J}{|H|} , \frac{ H}{|H|} \rangle d\Sigma_r$ as $r\rightarrow \infty$ is thus the same as
\[ \lim_{r\rightarrow \infty}\int_{\Sigma_r} X_i div_{\Sigma_r} V d\Sigma_r = \int_{S^2} \tilde X_i [\frac{1}{2}(\widetilde{\Delta}+2)(\widetilde{\delta^a}\widetilde{W_a} -2m)+ 2m ]dS^2 = \int_{S^2} \tilde X_i 2m \, dS^2\]
In this case, the reference Hamiltonian term is zero as $\frac{J_0}{|H_0|}$ is a constant vector. In view of expression (\ref{linear}), the theorem is proved.
\end{proof}
Next we show that the limit of the quasilocal energy is invariant under any $O(1)$ perturbations of embeddings into totally geodesic $\mathbb R^3$ and that it is Lorentzian equivariant.
\begin{cor}
Suppose $X_r=(\tau_0, X_1, X_2, X_3)$ is a family of isometric embeddings of $\Sigma_r$ into $\mathbb R^{3,1}$ with $ \tau_0 = \tilde \tau_0+O(r^{-1})$ for some function $\tilde \tau_0$ on $S^2$.
Then we still have
\[\lim_{r\rightarrow\infty} E(\Sigma_r, X_r, T_0)=\frac{1}{8 \pi} \int_{S^2} 2m (\sqrt{1+|a|^2}+ a^i\tilde{X}_i)dS^2.\]
\end{cor}
\begin{proof}
Let $\hat{X}_r$ be the embedding of $\Sigma_r$ by projecting $X_r$ onto $\mathbb R^{3}$ which is given by, $(0, X_1, X_2, X_3)$. It is not hard to check that the induced metric by the embedding $\hat{X}_r$ agrees with the standard round metric of radius $r$ up to the top order term and its area agrees with that of the standard round metric of radius $r$ up to the second order term.
The mean curvature of the embedding $\hat{X}_r$ is then $ (0, \widehat\Delta X_1, \widehat\Delta X_2 , \widehat\Delta X_3)$. By Lemma \ref{isom_R^3}, the mean curvature $\hat{H}_0$ satisfies
\begin{equation}\label{hat_H}|\hat{H}_0| = \frac{2}{r} +O(r^{-2}), \text{ and }\int_{\Sigma_r} |\hat{H}_0| d\Sigma_r = 8 \pi r + O(r^{-1}).
\end{equation}
The mean curvature $ H_0$ of $X_r$ is given by
\[ (\Delta \tau_0, \Delta X_1, \Delta X_2 , \Delta X_3). \]
The difference between $\sigma_{ab}$ and $\hat{\sigma}_{ab}$ is of order
\[ \sigma_{ab} -\hat{\sigma}_{ab} =O(1) \text{ and } \sigma^{ab} -\hat{\sigma}^{ab} =O(r^{-4}). \]
As a result, the difference between the two laplace operators is of order
\[ \widehat\Delta X_i - \Delta X_i =O(r^{-3}). \]
Hence $|\hat{H}_0|^2 -|H_0|^2 =O(r^{-4})$, and thus $|\hat{H}_0| -|H_0| =O(r^{-3})$.
By equation (\ref{linear}), the limit of quasilocal energy with respect to the embedding $(\tau_0, X_1, X_2, X_3)$ is thus
\begin{align*}
\frac{1}{8 \pi}\lim_{ r\to \infty} \int_{\Sigma_r} (|H_0| -|H|) d\Sigma_r=\frac{1}{8 \pi} \lim_{ r\to \infty} \int_{\Sigma_r} (|\hat{H}_0| -|H|) d\Sigma_r.
\end{align*}
Unlike the previous case, $\frac{{J}_0}{|H_0|} $ is no longer a constant vector for such an isometric embedding $X_r$. However, the asymptotic expansion
$$
-\langle T_0, \frac{{J}_0}{|H_0|} \rangle = \sqrt{1+|a|^2} +O(r^{-1})
$$ is valid and the energy component is the same as the limit of quasilocal energy of the isometric embedding into $\mathbb R^3$ in view of (\ref{hat_H}).
Next we compute the physical hamiltonian. Since the induced metric on the projection still agrees with the standard one up to lower order term, up to an isometry of $\mathbb R^3$, $X_i=r\tilde{X}_i+O(1)$. The corresponding time function $\tau$ is
$$ \tau = - (\sum_i a_i X_i) + \tau_0\sqrt{1+\sum_i a_i^2} = -(\sum_i a_i \tilde X_i) r + O(1).$$
Thus the physical hamiltonian remains the same.
Lastly, we claim that the reference hamiltonian goes to $0$ as $r$ goes to infinity.
\begin{lemma}\label{meancurvature}
Let $\sigma_r = r^2 \tilde \sigma +O(r)$ be a family of metrics on $S^2$.
Given an $O(1)$ time function $\tau_0$, let $ \Sigma^0_r $ be the images of the isometric embedding $X_r$ into $\mathbb R^{3,1}$ determined by $\tau_0$.
Let $V_0$ be the vector dual to the one form $\langle \nabla_{(\cdot) } \frac{ J_0 }{|H_0|}, \frac{H_0}{|H_0|} \rangle$ on $\Sigma^0_r$
then
\[ div_{\Sigma_r}V_0 = \frac{1}{2r^3} \widetilde{\Delta}(\widetilde{\Delta}+2) \tau_0 +O(r^{-4}). \]
\end{lemma}
\begin{proof}
We need to compute $\frac{ J_0}{|H_0|}$ up to $O(r^{-2})$. For this purpose, it is enough to assume that
the embedding is $(\tau_0, r \tilde{X}_1, r \tilde{X}_2, r\tilde{X}_3 ) $.
Using $\sum_{i} \partial_c\tilde{X}_i\partial_b \tilde{X}_i=\tilde{\sigma}_{bc}$, we derive that a normal vector is
$$
(1, \frac{1}{r} \partial_a \tau_0 \partial_b \tilde X_i\tilde \sigma^{ab})
$$
where $\tilde \sigma$ is the standard metric on $S^2$. As $\sum_{i} \tilde{X}_i\partial_a \tilde{X}_i=0$, we check that $$
(1, \frac{1}{r} \partial_a \tau_0 \partial_b \tilde X_i\tilde \sigma^{ab}) +\frac{\Delta \tau_0}{|H_0|^2} H_0
$$ is a normal vector perpendicular to $H_0$. Thus, $\frac{ J_0}{|H_0|}$, which is the unit normal perpendicular to $ H_0$ is, up to lower order, given by the same expression. As a result, we compute
$$
\langle \nabla_{\frac{\partial}{\partial x^a} } \frac{ J_0 }{|H_0|}, H_0 \rangle=
\sum_i \partial_a \left[\frac{1}{r} (\partial_b \tau_0 \partial_c \tilde X_i)\tilde \sigma^{bc}+\frac{\Delta \tau_0}{|H_0|^2}\Delta X_i \right] \Delta X_i.
$$
The right hand side equals to
$$\left[\sum_i \frac{1}{r} (\partial_b \tau_0 )(\partial_a \partial_c \tilde X_i)\tilde \sigma^{bc}\Delta X_i\right] + \partial_a \Delta \tau_0 +O(r^{-3}).$$
Here one uses again that $\sum_{i} \tilde{X}_i\partial_a \tilde{X}_i=0$ and thus from the first term, one has non-zero contribution only when the derivative $ \partial_a $ falls on $ \partial_c \tilde X_i$. For the second term, the leading term of $|H_0|$ is independent of $\theta$ and $\phi$ and $\sum _i (\Delta X_i) ^2 =|H_0|^2$ up to lower order terms. Thus one only has contribution when the derivative hits $\Delta \tau_0$. Direct computation using $\sum_i \tilde{X}_i \partial_a\partial_c\tilde{X}_i=-\tilde{\sigma}_{ac}$ shows that
$$
\sum_i \frac{1}{r} (\partial_b \tau_0) (\partial_a\partial_c \tilde X_i)\tilde \sigma^{bc}\Delta X_i = \frac{2}{r^2}\partial_a \tau_0.
$$
As a result,
\[ div_{\Sigma_r}V_0 = \frac{1}{2r^3} \widetilde{\Delta}(\widetilde{\Delta}+2) \tau_0 +O(r^{-4}) \]
\end{proof}
Using the above lemma, the reference hamiltonian at infinity is
\[
\lim_{r\to\infty}\int_{\Sigma_r} X_i div_{\Sigma_r} V_0 d\Sigma= \int_{S^2} \frac{1}{2} \tilde X_i\widetilde{\Delta}(\widetilde{\Delta}+2) \tau_0 dS^2=0
\]
\end{proof}
\begin{cor}
Suppose $X_r'$ is another family of isometric embeddings of $\sigma_r$ into $\mathbb R^{3,1}$ such that $X'_r=\tilde{L}_r X_r$ for some $X_r$ in the previous Corollary, and a family of Lorentzian transformation $\tilde{L}_r$ such that the limit of the $SO(3,1)$ part of $\tilde{L}_r$ converges to an $L_\infty$, then the energy-momentum 4-vector also transform by $L_\infty$.
\end{cor}
\begin{proof}
Both $|H_0|$ and the connection one form $\langle\nabla_{(\cdot)}^{\mathbb R^{3,1}} \frac{{J}_0}{|H_0|}, \frac{H_0}{|H_0|}\rangle$ are invariant under Lorentzian transformation, while $\langle T_0, \frac{{J}_0}{|H_0|}\rangle$ and $\nabla\tau= - \nabla \langle T_0, X_r\rangle$ are Lorentzian equivariant.
\end{proof}
For example, if we take a family of isometric embedding $X_r$ into $\mathbb R^3$ and define $X'_r=X_r+r$, it is not hard to see that the hypersurface spanned by $X'_r$ is asymptotically null.
\section{Optimal embedding equation}\label{opt}
The optimal embedding equation for minimizing the quasilocal energy is derived in \cite[Proposition 6.2]{wy2}. The equation reads
\begin{equation}\label{optimal}
-(\widehat{H}\hat{\sigma}^{ab} -\hat{\sigma}^{ac} \hat{\sigma}^{bd} \hat{h}_{cd})\frac{\nabla_b\nabla_a \tau}{\sqrt{1+|\nabla\tau|^2}}+ div_\Sigma (\frac{\nabla\tau}{\sqrt{1+|\nabla\tau|^2}} \cosh\theta|{H}|-\nabla\theta-V)=0\end{equation}
where $\sinh \theta =\frac{-\Delta \tau}{|{H}|\sqrt{1+|\nabla \tau|^2}} $ and $V$ is the connection one-form $\langle \nabla^N_{(\cdot)} \frac{J}{|H|}, \frac{H}{|H|}\rangle$. To solve for this equation, we start with data on the 2-surface $\Sigma$ given by $(\sigma_{ab}, |H|, V)$. Take a function $\tau$ on $\Sigma$ and consider the isometric embedding $\hat{X}:(\Sigma, \hat{\sigma}) \rightarrow \mathbb R^3$ with the metric $\hat{\sigma}_{ab}=\sigma_{ab}+\tau_a \tau_b$. $\widehat{H}$ and $\hat{h}_{ab}$ are the mean curvature and the second fundamental form of $\hat{\Sigma}$, the image of $\hat{X}$ in $\mathbb R^3$, respectively.
Let $X:\Sigma\rightarrow \mathbb R^{3,1}$ be the embedding of the graph of $\tau$ over $\hat{\Sigma}$ in $\mathbb R^{3,1}$, and $\Sigma_0$ be the image of $X$ with the induced metric isometric to $\sigma$. The optimal isometric embedding equation can be written in terms of the geometry of $\Sigma_0$. In fact, the quasilocal energy of $\Sigma_0$ with respect to itself as a reference is zero and thus minimizing.
Equation \eqref{optimal} is automatically true on $\Sigma_0$ and we deduce
\begin{equation}\label{optimal2}
-(\widehat{H}\hat{\sigma}^{ab} -\hat{\sigma}^{ac} \hat{\sigma}^{bd} \hat{h}_{cd})\frac{\nabla_b\nabla_a \tau}{\sqrt{1+|\nabla\tau|^2}}+ div_\Sigma (\frac{\nabla\tau}{\sqrt{1+|\nabla\tau|^2}} \cosh\theta_0|{H_0}|-\nabla\theta_0-V_0)=0
\end{equation} where $V_0$, $H_0$, $\theta_0$ are the corresponding data on $\Sigma_0$. Equation \eqref{optimal2} can be checked directly for spacelike surfaces in $\mathbb R^{3,1}$.
Subtracting equation (\ref{optimal}) from equation (\ref{optimal2}), equation (\ref{optimal}) is then equivalent to \begin{equation}\label{optimal2.5} div_{\Sigma} [ \frac{\nabla \tau }{\sqrt{1+ |\nabla \tau|^2}}(\cosh \theta |H| - \cosh \theta_0 |H_0|)- \nabla(\theta - \theta_0) -V + V_0] =0.\end{equation}
By the definition of $\theta$ and $\theta_0$, we derive
\[ \cosh \theta |H| - \cosh \theta_0 |H_0| =\sqrt{|H|^2 +\frac{(\Delta \tau)^2}{1+ |\nabla \tau|^2}} - \sqrt{|H_0|^2 +\frac{(\Delta \tau)^2}{1+ |\nabla \tau|^2}} \] and
\[ \sinh (\theta -\theta _0 ) = \frac{\Delta \tau}{|H|| H_0| \sqrt{1+ |\nabla \tau|^2}}(\sqrt{|H|^2 +\frac{(\Delta \tau)^2}{1+ |\nabla \tau|^2}} - \sqrt{|H_0|^2 +\frac{(\Delta \tau)^2}{1+ |\nabla \tau|^2}}). \]
Set
\begin{equation}\label{form_f} \begin{split}f &= \frac{\sqrt{|H_r|^2 +\frac{(\Delta \tau)^2}{1+ |\nabla \tau|^2}} - \sqrt{|H_0|^2 +\frac{(\Delta \tau)^2}{1+ |\nabla \tau|^2}} }{ \sqrt{1+ |\nabla \tau|^2}}\\
&=\frac{|H_r|^2-|H_0|^2}{\sqrt{|H_r|^2(1+|\nabla\tau|^2)+(\Delta\tau)^2}+\sqrt{|H_0|^2(1+|\nabla\tau|^2)+(\Delta\tau)^2}}. \end{split}\end{equation}and equation \eqref{optimal} is equivalent to
\begin{equation} \label{optimal3}
div_\Sigma(f \nabla \tau) - \Delta [ \sinh^{-1} (\frac{\Delta \tau f}{|H||H_0|})]-( div_\Sigma V -div_\Sigma V_0)=0.
\end{equation}
In this equation, $V$ are $|H|$ come from the physical data, and $V_0$ and $|H_0|$ only depend on the embedding $X:\Sigma\rightarrow \mathbb R^{3,1}$, while $\tau=-\langle X, T_0\rangle$ depends on both $X$ and $T_0$. Equation \eqref{optimal3} together with the isometric embedding equation $\langle dX, dX\rangle=\sigma$ form the optimal isometric embedding system.
We shall solve the system for a family of spacelike 2-surfaces at null or spatial infinity such that the family of isometric embeddings $X_r$ into $\mathbb R^{3,1}$ is of the form $X_r=B_r\hat{X}_r$ where $B_r$ is a family in $SO(3,1)$ and $\hat{X}_r$ is an $O(1)$ perturbation of isometric embeddings into $\mathbb R^3$.
We observe that momentum become an obstruction to solving the optimal embedding equation for $\hat{X}_r$ and then discuss how this can be resolved by boosting the embedding by $B_r$. The discussion covers the spatial infinity case discussed in \cite{wy3} as well. In the last subsection, we show the solution obtained is locally energy-minimizing up to lower order terms in $r$.
\subsection{Embedding near $\mathbb R^3$}
In this subsection, we study the geometry of a family of isometric embeddings $\hat{X}_r$ that is near a totally geodesic $\mathbb R^3$ in $\mathbb R^{3,1}$.
\begin{lemma}\label{mean_curv} Suppose $\hat{X}_r$ is a family of isometric embeddings into $\mathbb R^{3,1}$ for a given family of metrics $\sigma_r=r^2\tilde{\sigma}+r\sigma^{(1)}+\sum_{k=0}^\infty r^{-k} \sigma^{(-k)}$ and
\begin{equation}\label{hat_X}\hat{X}_r= r\hat{X}^{(1)}+ \sum_{k=0}^\infty r^{-k} \hat{X}^{(-k)}\end{equation}
with $\hat{X}^{(1)}=\tilde{X}=(0, \tilde{X}_1, \tilde{X}_2, \tilde{X}_3)$, the standard embedding of $S^2$ into $\mathbb R^3$.
Denote by $\hat{\tau}^{(k)}$ the time component of $\hat{X}^{(k)}$ and by $G(\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l)},\sigma) $
a term that depends on $\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l)}$ and $\sigma$. Then
\[|H_0|={2}r^{-1}+r^{-2} h_0^{(-2)}+\sum_{k=3}^\infty r^{-k} h_0^{(-k)}\]
where $h_0^{(-k)}=G(\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-k+3)},\sigma)$ for $k\geq 3$ and $h_0^{(-2)}$ depends only on $\sigma$.
\end{lemma}
\begin{proof} In the proof, we suppress the subscript $r$ and write $\hat{X}$ for $\hat{X}_r$ and $\sigma$ for $\sigma_r$. From the expansion of $\sigma$, $\sigma=r^2\tilde{\sigma}+r\sigma^{(1)}+\cdots$, we deduce
\[2\sum_{i=1}^3 d \tilde{X}_i d\hat{X}_i^{(0)}=\sigma^{(1)}\]
This can be transformed into a linear elliptic equation for $\hat{X}_i^{(0)}$ which can be solved (see section 6 of \cite{n}).
In general,
\[2 \sum_{i=1}^3 d \tilde{X}_i d\hat{X}_i^{(-l-1)}+\sum_{m=0}^l \langle d\hat{X}^{(-m)}, d \hat{X}^{(-l+m)}\rangle=\sigma^{(-l)}\]
and thus $\hat{X}_i^{(-l-1)}$ is determined by $\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l)}$ and $\sigma$.
Recall the mean curvature is given by $H_0={\Delta}\hat{X}$ where ${\Delta}$ is the Laplace operator with respect to $\sigma$.
For a function $g$ on $S^2$, we compute
\[\begin{split}\Delta g&=\sigma^{ab}(\partial_a\partial_b g-\gamma_{ab}^c\partial_c g)\\
&=(r^{-2}\tilde{\sigma}^{ab}+r^{-3} \sigma^{(-3) ab}+O(r^{-4}))(\widetilde{\nabla}^2_{ab}g-r^{-1} \gamma^{(-1)c}_{ab} \partial_c g+O(r^{-2}))\\
\end{split}\]
where $\gamma_{ab}^c=\tilde{\gamma}^{c}_{ab}+r^{-1}\gamma^{(-1)c}_{ab}+\cdots$ is the expansion of the Christoffel symbol $\gamma_{ab}^c$ of
$\sigma_{ab}$ and $\widetilde{\nabla}^2_{ab}g=\partial_a\partial_b g-\tilde{\gamma}_{ab}^c\partial_c g$ is the Hessian of $g$ with respect to $\tilde{\sigma}_{ab}$.
Therefore, we obtain the following formula:
\begin{equation}\label{Delta_g}\Delta g=r^{-2}\widetilde{\Delta} g+r^{-3} (\sigma^{(-3) ab} \widetilde{\nabla}^2_{ab} g-\tilde{\sigma}^{ab}\gamma_{ab}^{(-1)c}\partial_c g)+O(r^{-4})\end{equation}
By \eqref{hat_X}, $H_0=\Delta \hat{X}$ has the following expansion:
\[ H_0=r\Delta \hat{X}^{(1)}+\sum_{k=0}^\infty r^{-k} \Delta\hat{X}^{(-k)}.\]
Since $\hat{X}^{(1)}=\tilde{X}$, the standard embedding of $S^2$, we compute
\[\Delta \tilde{X}=-2r^{-2}\tilde{X}+r^{-3}(-\tilde{\sigma}_{ab}\sigma^{(-3)ab}\tilde{X}-\tilde{\sigma}^{ab}{\gamma^{(-1)}}_{ab}^c\partial_c \tilde{X})+O(r^{-4})\]
where we use $\widetilde{\nabla}^2_{ab}\tilde{X}=-\tilde{\sigma}_{ab}\tilde{X}$.
Therefore, we obtain
\[H_0=\Delta \hat{X}= -2r^{-1}\tilde{X}+r^{-2}H_0^{(-2)}+\sum_{k=3}^\infty r^{-k}H_0^{(-k)}\]
where
\[H_0^{(-2)}=\widetilde{\Delta}\hat{X}^{(0)}-(\tilde{\sigma}_{ab}{\sigma^{(-3)}}^{ab}\tilde{X}+\tilde{\sigma}^{ab}{\gamma^{(-1)}}_{ab}^c\partial_c\tilde{X})\]
and
\[H_0^{(-k)}=\widetilde{\Delta}\hat{X}^{(-k+2)}+G(\hat{\tau}^{(0)},\cdots \hat{\tau}^{(-k+3)},\sigma)\] for $k\geq 3$.
We compute
\[|H_0|^2=4r^{-2}+r^{-3}(2\widetilde{\Delta}\tilde{X}_i\widetilde{\Delta}\hat{X}_i^{(0)}+G(\sigma)) +\sum_{k=3}^\infty r^{-k-1}(2\sum_{i=1}^3\widetilde{\Delta}\tilde{X}_i\widetilde{\Delta} \hat{X}_i^{(-k+2)}+G(\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-k+3)},\sigma) ).\] Since $\hat{X}_i^{(-k+2)}$ depends only on $\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-k+3)}$ and $\hat{X}_i^{(0)}$ depends only on $\sigma$, we obtain
\[|H_0|^2={4}r^{-2}+r^{-3}G(\sigma)+\sum_{k=0}^\infty r^{-k-4} G(\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-k)},\sigma)\] and the expansion for $|H_0|$ follows by taking square root.
\end{proof}
\begin{lemma}\label{conn}
Under the same assumption as Lemma \ref{mean_curv}.
Let $V_0$ be the one form $\langle \nabla_{(\cdot) } \frac{ J_0 }{|H_0|}, \frac{H_0}{|H_0|} \rangle$ on the image of $\hat{X}_r$
then
\[div_{\Sigma_r}V_0=r^{-3}[\frac{1}{2} \widetilde\Delta(\widetilde \Delta +2) \hat{\tau}^{(0)}]+ \sum_{k=4}^{\infty} r^{-k} v_0^{(-k)} \]
where
\[v^{(-k)}=\frac{1}{2} \widetilde\Delta(\widetilde \Delta +2) \hat{\tau}^{(-k+3)} + G(\hat{\tau}^{(0)} , \cdots , \hat{\tau}^{(-k+4)} , \sigma). \]
\end{lemma}
\begin{proof}
To compute the connection one-form in mean curvature gauge, we need a normal vector $I$ that is perpendicular to $H_0$. Suppose $I$ is of the form
\[I=I^{(0)}+\sum_{l=1}^\infty r^{-k} I^{(-k)}\] with $I^{(0)}=(1,0,0,0)$. The condition $\langle I, \partial_a \hat{X}\rangle=0$ implies
\[\langle I^{(0)}, \partial_a \hat{X}^{(0)}\rangle+\langle I^{(-1)}, \partial_a\tilde{X}\rangle=0,\] and $\langle I, H_0\rangle=0$ implies
\[\langle I^{(0)}, H_0^{(-2)}\rangle+\langle I^{(-1)}, -2\tilde{X}\rangle=0.\] Since $\tilde{X}, \partial_1\tilde{X}, \partial_2\tilde{X}$ form an orthonormal basis in $\mathbb R^3$ and we can assume $I^{(-1)}$ is perpendicular to $I^{(0)}$, it is not hard to check that
\[I^{(-1)}=\partial_c \hat{\tau}^{(0)}\tilde{\sigma}^{cb}\partial_b\tilde{X}-\frac{1}{2}\tilde{\Delta} \hat{\tau}^{(0)}\tilde{X}\] and in general
\[I^{(-k)}=\partial_c \hat{\tau}^{(-k+1)}\tilde{\sigma}^{cb}\partial_b\tilde{X}-\frac{1}{2}\tilde{\Delta} \hat{\tau}^{(-k+1)}\tilde{X}+G( \hat{\tau}^{(0)}, \cdots, \hat{\tau}^{(-k+2)})\] for $k\geq 1$.
The connection in mean curvature gauge is thus
\[\langle \nabla_a \frac{I}{\sqrt{-\langle I, I\rangle}}, \frac{H_0}{|H_0|}\rangle=\frac{1}{\sqrt{-\langle I, I\rangle}{|H_0|}}\langle \nabla_a I, H_0\rangle.\]
We compute
\[\begin{split}& \langle \nabla_a I, H_0 \rangle =-2r^{-2}\langle \partial_a I^{(-1)}, \tilde{X}\rangle+\sum_{k=3}^\infty r^{-k}(-2\langle \partial_a I^{(-k+1)}, \tilde{X}\rangle+\cdots+\langle \partial_a I^{(-1)}, H_0^{(-k+1)}\rangle)\\
&=r^{-2}\partial_a(\widetilde{\Delta} \hat{\tau}^{(0)}+2\hat{\tau}^{(0)})+\sum_{k=3}^\infty r^{-k}[\partial_a(\widetilde{\Delta} \hat{\tau}^{(-k+2)}+2\hat{\tau}^{(-k+2)})+G( \hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-k+3)})].\end{split}\]
Since the leading term of $|H_0|$ is $\frac{2}{r}$, the leading term of $\langle I, I\rangle$ is $-1$, and the leading term of $div_{\Sigma_r}\tilde{\alpha}$ is $r^{-2}\widetilde{div}\tilde{\alpha}$ for a one-form $\tilde{\alpha}$ on $S^2$, we obtain the desired expansion for $div_{\Sigma_r} V_0$.
\end{proof}
\subsection{Boost $\hat{X}_r$ in $\mathbb R^{3,1}$}
Suppose $\Sigma_r$ is a family of spacelike 2-surfaces in spacetime such that
\noindent (1) The induced metric satisfies $\sigma_r = r^2 \tilde \sigma+O(r)$.
\noindent(2) The norm of the mean curvature vector satisfies $|H| =\frac{2}{r}+\frac{h^{(-2)}}{r^2} + O(r^{-3})$.
\noindent (3) The connection one-form in mean curvature gauge $V$ satisfies $ div_{\Sigma_r} V = \frac{v^{(-3)}}{r^3} + O(r^{-4})$.
These assumptions hold on coordinate spheres of an asymptotically flat hypersurface as well as the $r$ level surfaces at a retarded time in Bondi-Sachs coordinates. Altogether they guarantee the limit of the quasilocal energy-momentum $(e, p_1, p_2, p_3)$ with respect to isometric embeddings of $\sigma^r_{ab}$ into $\mathbb R^3$ is well-defined.
We assume the family of isometric embeddings $X_r:\Sigma\times [r_0, \infty)\rightarrow \mathbb R^{3,1}$ is of the form
\begin{equation}\label{X_r} X_r=B_r \hat{X}_r\end{equation} where $\hat{X}_r$ is a family of isometric embeddings of $\sigma_r$ into $\mathbb R^{3,1}$ that is analytic in $r$ and $B$ is a family of elements in $SO(3,1)$ that is analytic in $r$. We assume that $B=\lim_{k\rightarrow \infty} B_{k}$ where \[B_{k}=e^{\frac{1}{r^k} b^{(-k)}}\cdots e^{\frac{1}{r} b^{(-1)}}e^{b^{(0)}}, k=0,1,2,\cdots\] for $b^{(-k)}\in \mathfrak{so}(3,1)$, the Lie algebra of $SO(3,1)$. We assume $\hat{X}_r=r\hat{X}^{(1)}+\sum_{k=0}^{\infty} r^{-k} \hat{X}^{(-k)}$
and the time function of the embedding $\hat{X}_r$ is given by
\[\sum_{k=0}^{\infty} r^{-k} \hat{\tau}^{(-k)}.\]
Therefore we may assume $\hat{X}^{(1)}=(0, \tilde{X}_1, \tilde{X}_2, \tilde{X}_3$) is a standard embedding of $S^2$ into $\mathbb R^3$. We shall show that all $\hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-k)}$ and $b^{(0)},\cdots, b^{(-k)}$ can be solved inductively. To prepare for the induction, we compute the corresponding terms in the optimal isometric embedding for $X_r$ of the given form \eqref{X_r}.
Fix an $l$, we denote
\begin{equation}\label{X_l+1} X_{l+1}=e^{r^{-l} b^{(-l)}}\cdots e^{{r}^{-1} b^{(-1)}}e^{b^{(0)}}\hat{X}.\end{equation}
Suppose $\hat{\tau}^{(0)}\cdots \hat{\tau}^{(-l)}$ and all $b^{(0)}, \cdots b^{(-l)}$ are known, we see that other components of $\hat{X}^{(0)}\cdots \hat{X}^{(-l)}$ are known and by \eqref{X_l+1} all $(X_{l+1})^{(0)}, \cdots (X_{l+1})^{(-l)}$ are known.
It is not hard to see that $X_{l+2}^{(-l)}=X_{l+3}^{(-l)}=\cdots$ is stabilized at the $r^{-l}$ term and thus $X^{(-l)}=X_{l+2}^{(-l)}$. By definition, $X_{l+2}$ is
\[\begin{split}&\{I+\cdots+r^{-l-1}[b^{(-l-1)}+G(b^{(0)},\cdots, b^{(-l)})]+\cdots\}(rB\hat{X}^{(1)}+\cdots+r^{-l} B\hat{X}^{(-l)}+\cdots)\\&=rB\hat{X}^{(1)}+\cdots +r^{-l}(b^{(-l-1)}B\hat{X}^{(1)}+B\hat{X}^{(-l)}+G(b^{(0)},\cdots, b^{(-l)}, \hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l+1)}))+O(r^{-l-1}),\end{split}\] and thus
\[X_{l+2}^{(1)}=B\hat{X}^{(1)},\]
\[X_{l+2}^{(-l)}=b^{(-l-1)}B\hat{X}^{(1)}+B\hat{X}^{(-l)}+G(b^{(0)},\cdots, b^{(-l)}, \hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l+1)}),\]
and $X_{l+1}^{(m)}$ for $0\geq m\geq -l+1$ is of the form $G( b^{(0)},\cdots, b^{(-l)}, \hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l+1)})$.
Denote by $\tau$ the time function of $X_{l+2}$. We shall plug $\tau$ into the optimal equation, find $b^{(-l-1)}$ so that $\hat{\tau}^{(-l-1)}$ is solvable, and then solve for $\hat{\tau}^{(-l-1)}$. Denote by $\mathfrak{G}^{(k)}$ a term of order $r^{k}$ whose coefficients depend only on $\hat{\tau}^{(0)}, \cdots, \hat{\tau}^{(-l+1)}$ and $b^{(0)}, \cdots, b^{(-l)}$ and the physical data $\sigma, |H|$, and $V$.
\begin{lemma} Write $e^{b^{(0)}}=(B_{\alpha\beta})$, then $\tau=(X_{l+2})_0$ has the following expansion:
\begin{equation}\label{tau_exp}\tau=r\tau^{(1)}+r^{-l} \tau^{(-l)}+\mathfrak{G}^{(0)}+\cdots+\mathfrak{G}^{(-l+1)}+O(r^{-l-1})\end{equation} where \[\tau^{(1)}=\sum_{i=1}^3 B_{0i}\tilde{X}_i\] and \begin{equation}\tau^{(-l)}=\sum_{i=1}^3\sum_{\alpha=0}^3 b_{0\alpha}^{(-l-1)} B_{\alpha i}\tilde{X}_i+B_{00} \hat{\tau}^{(-l)}+G( b^{(0)},\cdots, b^{(-l)}, \hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l+1)})\end{equation}
\end{lemma}
\begin{proof}
We compute \[(B\hat{X}^{(1)})_0=B_{0\beta}(\hat{X})^{(1)}_\beta\] and recall $(\hat{X})_0^{(1)}=0$ and $(\hat{X})_i^{(1)}=\tilde{X}_i$ for $i=1, 2, 3$. Likewise, \[(b^{(-l-1)}B\hat{X}^{(1)})_0=b^{(-l-1)}_{0\alpha}B_{\alpha\beta}(\hat{X})_\beta^{(1)}.\]
On the other hand $(B\hat{X}^{(-l)})_0=B_{00} \hat{\tau}^{(-l)}+B_{0i}\hat{X}_i^{(-l)}$, and we already know that $\hat{X}_i^{(-l)}$ depends only on $\hat{\tau}^{(-l+1)}$.
\end{proof}
Now we proceed to calculate the terms in the optimal isometric embedding equation. For simplicity, we denote $\sum_{i=1}^3\sum_{\alpha=0}^3 b_{0\alpha}^{(-l-1)} B_{\alpha i}\tilde{X}_i$ by $g$.
\begin{lemma}
With $\tau$ given by \eqref{tau_exp}, we have
\[\Delta \tau=-2r^{-1}\tau^{(1)}-2r^{-l-2}g+\mathfrak{G}^{(-2)}+\cdots +\mathfrak{G}^{(-l-2)}+O(r^{-l-3})\] and
\[\begin{split}|\nabla\tau|^2=\sum_i c_i^2-(\tau^{(1)})^2+2r^{-l-1}(\sum_i c_i d_i-\tau^{(1)}g)+\mathfrak{G}^{(-1)}+\cdots +\mathfrak{G}^{(-l-1)}+O(r^{-l-2})\end{split}\]
where
\begin{equation}\label{c_i}
c_i=B_{0i}
\end{equation}
\begin{equation}\label{d_i}
d_i=\sum_{\alpha=0}^3B_{\alpha i}b_{0 \alpha} ^{(-l-1)}
\end{equation}
\end{lemma}
\begin{proof}
We use the formula that if two functions $A$ and $B$ on $S^2$ are given by $A=A_i\tilde{X}_i$ and $B=B_j\tilde{X}_j$, then $\widetilde{\nabla}A\cdot\widetilde{\nabla}B=\sum A_i B_i-AB$.
\end{proof}
Recalling the function $f$ defined in \eqref{form_f}, we compute the expansion of $f$ and the expansions of terms that appear on the optimal isometric embedding equation \eqref{optimal3} in the following:
\begin{lemma}\label{opt_exp}
Suppose $|H_r|={2}{r}^{-1}+r^{-2} h^{(-2)}_r+\mathfrak{G}^{(-3)}+\cdots+\mathfrak{G}^{(-l-3)}+O(r^{-l-4})$ and $|H_0|={2}r^{-1}+r^{-2} h^{(-2)}_0+\mathfrak{G}^{(-3)}+\cdots+\mathfrak{G}^{(-l-3)}+O(r^{-l-4})$ then
\[f=r^{-2} f^{(-2)}+r^{-l-3}f^{(-l-3)}+\mathfrak{G}^{(-3)}+\cdots +\mathfrak{G}^{(-l-3)}+O(r^{-l-4})\] where
\begin{equation}\label{f-2} f^{(-2)}=\frac{h^{(-2)}_r-h^{(-2)}_0}{(1+|c|^2)^{1/2}}\end{equation} and
\begin{equation}\label{f-3}f^{(-l-3)}=-\frac{(h^{(-2)}_r-h^{(-2)}_0)\sum_i c_i d_i}{(1+|c|^2)^{3/2}}=f^{(-2)}(-\frac{\sum_i c_i d_i}{1+|c|^2}).\end{equation}
\[\begin{split}\Delta \sinh^{-1}\frac{(\Delta\tau)f}{|H_r||H_0|}&=\frac{1}{4} r^{-3}\widetilde{\Delta}[(\widetilde{\Delta}\tau^{(1)}) f^{(-2)}]+\frac{1}{4} r^{-l-4}\widetilde{\Delta} (f^{(-l-3)}\widetilde{\Delta} \tau^{(1)}+f^{(-2)}\widetilde{\Delta}g)\\
&+\mathfrak{G}^{(-4)}+\cdots +\mathfrak{G}^{(-l-4)}+O(r^{-l-5})\end{split}\]
\[\begin{split} div(f\nabla\tau)&=r^{-3}\widetilde{div}(f^{(-2)}\widetilde{\nabla}\tau^{(1)})+r^{-l-4}[\widetilde{div}(f^{(-2)}\widetilde{\nabla}g)
+\widetilde{div}(f^{(-l-3)}\widetilde{\nabla}\tau^{(1)})]\\
&+\mathfrak{G}^{(-4)}+\cdots +\mathfrak{G}^{(-l-4)}+O(r^{-l-5})\end{split}\]
\end{lemma}
\subsection{Solving the optimal isometric embedding equation for all orders}
\begin{theorem} \label{leading_order}
Suppose $\Sigma_r$ satisfies (1), (2), and (3) and the limit of the quasilocal energy-momentum $(e,p_1,p_2,p_3)$ is timelike.
There is a function $\tau^{(0)}$ on $S^2$ such that isometric embeddings $X_r:\Sigma\rightarrow \mathbb{R}^{3,1}$ with the time function $\tau$ given below solves equation (\ref{optimal}) up to $O(r^{-3})$.
\[ \tau = (\sum_{i=1}^{3} c_{i} \tilde X_i )r + \tau^{(0)}+ O(r^{-1}) \]
where $(c_{1}, c_{2}, c_{3})$ satisfy $\frac{c_{i}}{ \sqrt{1+|c|^2}} = \frac{p_i}{e}$.
\end{theorem}
\begin{proof}
Under the assumption, the energy momentum vector $(e,p_1,p_2,p_3)$ and given by \cite[Equation (2.6)]{wy3}
\begin{equation}\label{energy_momentum} \int_{S^2} (h_0^{(-2)}-h^{(-2)}) dS^2 = 8 \pi e \text{ and } -\int_{S^2}(v_0^{(-3)}- v^{(-3)}) \tilde X_i dS^2 = 8 \pi p_i. \end{equation}
From Lemma \ref{opt_exp}, the coefficient of the leading $r^{(-3)}$ term of $div(f\nabla\tau)-\Delta \sinh^{-1}\frac{(\Delta\tau)f}{|H_r||H_0|}$ is \[\widetilde{div}(f^{(-2)}\widetilde{\nabla}\tau^{(1)})-\frac{1}{4} \widetilde{\Delta} (f^{(-2)}\widetilde{\Delta} \tau^{(1)}),\] while the $r^{(-3)}$ term of $div V_r-div V_0$ is
\[v^{(-3)}-\frac{1}{2}\widetilde{\Delta}(\widetilde{\Delta}+2)\hat{\tau}^{(0)}.\]
Thus $\hat{\tau}^{(0)}$ is solvable if
\[\int_{S^2}[\widetilde{div}(f^{(-2)}\widetilde{\nabla}\tau^{(1)})-\frac{1}{4} \widetilde{\Delta} (f^{(-2)}\widetilde{\Delta} \tau^{(1)})-v^{(-3)}]\tilde{X}_i dS^2\]
is zero. Computing using $\widetilde{\nabla}\tau^{(1)}\widetilde{\nabla}\tilde{X}_i=c_i-\tau^{(1)}\tilde{X}_i$, we find this is equivalent to
\[ \int_{S^2} (c_i f^{(-2)}+v^{(-3)}\tilde{X}_i) dS^2=0.\] From \eqref{energy_momentum} we see that $\int_{S^2} f^{(-2)}dS^2=\frac{-e}{\sqrt{1+|c|^2}} 8\pi$ and $\int_{S^2} v^{(-3)} \tilde{X}_i dS^2=8\pi p_i$. Thus we can take
\[ \frac{c_i}{ \sqrt{1+|c|^2}} = \frac{p_i}{e}\] and $\hat{\tau}^{(0)}$ is solvable. By equation (\ref{c_i}), $b^{(0)}$ is determined.
\end{proof}
Now we solve the optimal embedding equation to all order of $r$ by induction.
\begin{theorem}\label{all_order} Under the same assumption as in Theorem \ref{leading_order}.
Suppose the family of optimal isometric embeddings $X_r$ into $\mathbb R^{3,1}$ is of the form \eqref{X_r} described in the previous subsection. There exists $b^{(-k)}\in \mathfrak{so}(3,1)$ such that all $\hat{\tau}^{(-k)}$ can be solved for $k\geq 0$.
\end{theorem}
\begin{proof}
From Lemma \ref{opt_exp}, the coefficient of the $r^{(-l-4)}$ term of $div(f\nabla\tau)-\Delta \sinh^{-1}\frac{(\Delta\tau)f}{|H_r||H_0|}-div(V- V_0)$ is
\[\widetilde{div}(f^{(-2)}\widetilde{\nabla}g+f^{(-l-3)}\widetilde{\nabla}\tau^{(1)})-\frac{1}{4} \widetilde{\Delta} (f^{(-l-3)}\widetilde{\Delta} \tau^{(1)}+f^{(-2)}\widetilde{\Delta}g)+\frac{1}{2}\widetilde{\Delta}(\widetilde{\Delta}+2)\hat{\tau}^{(-l-1)}+\mathfrak{G}\]
where $\mathfrak{G}$ is a term that depend on $\hat{\tau}^{(0)}, \cdots \hat{\tau}^{(-l)}$, $b^{(0)}, \cdots b^{(-l)}$, $\sigma$, $V$ and $H$.
Thus the solvability depends on whether, {for each $i=1, 2, 3$}, we can make the following expression equal to $-\int_{S^2}\mathfrak{G} \tilde X_i \, dS^2$ by choosing suitable $d_j$ in $g$:
\[\int_{S^2}\left[ \widetilde{div}(f^{(-2)}\widetilde{\nabla}g+f^{(-l-3)}\widetilde{\nabla}\tau^{(1)})-\frac{1}{4} \widetilde{\Delta} (f^{(-l-3)}\widetilde{\Delta} \tau^{(1)}+f^{(-2)}\widetilde{\Delta}g) \right]\tilde{X}_i \, dS^2.\]
We integrate by parts and compute that this expression is equal to
\[\begin{split}&-\int_{S^2} f^{(-2)}(\tilde{X}_ig+\widetilde{\nabla}\tilde{X}_i\cdot\widetilde{\nabla}g)
+f^{(-l-3)}(\tilde{X}_i\tau^{(1)}+\widetilde{\nabla}\tilde{X}_i\cdot\widetilde{\nabla}\tau^{(1)}) \, dS^2 \\
&=-\int_{S^2} f^{(-2)}d_i+f^{(-l-3)}c_i \, dS^2\end{split}\] where we use $\widetilde{\nabla}g\widetilde{\nabla}\tilde{X}_i=d_i-g\tilde{X}_i$.
This term is \[(\int_{S^2} f^{(-2)} \,dS^2 )\sum_{j}\left(\delta_{ij}-\frac{c_i c_j}{1+\sum_k c_k^2}\right)d_j \]
Since the energy component is positive and the matrix $\delta_{ij}-\frac{c_i c_j}{1+\sum_k c_k^2}$ is positive definite, we can choose $d_j$ so that $\forall i$.
\[(\int_{S^2} f^{(-2)} \,dS^2 )\sum_{j}\left(\delta_{ij}-\frac{c_i c_j}{1+\sum_k c_k^2}\right)d_j = -\int_{S^2}\mathfrak{G} \tilde X_i \, dS^2 \]
\end{proof}
\subsection{Locally energy minimizing at all orders}
In this subsection, we show the solution $X_r=B_r\hat{X}_r$ obtained in Theorem \ref{all_order} is locally energy-minimizing at all orders. By this we mean that the second variation of the quasilocal energy is positive if we vary any $\hat{\tau}^{(-k)}$ or $b^{(-k)}$. Let $\tau$ be the time function of $X_r$ and let $\delta\tau$ be a variation. From Proposition 6.2 in \cite{wy2}, the first variation of the energy is
\[\int_{{\Sigma_r}}\left[-(\widehat{H} \hat{\sigma}^{ab} -\hat{\sigma}^{ac} \hat{\sigma}^{bd} \hat{h}_{cd})\frac{\nabla_b\nabla_a \tau}{\sqrt{1+|\nabla\tau|^2}}+ div_{\Sigma_r} (\frac{\nabla\tau}{\sqrt{1+|\nabla\tau|^2}} \cosh\theta|{H}|-\nabla\theta-V_r) \right] \delta\tau d\Sigma_r. \]
By the derivation in \S\ref{opt}, we can rewrite this as
\[ \int_{\Sigma_r} \left[ div(f\nabla\tau)-\Delta \sinh^{-1}\frac{(\Delta\tau)f}{|H_r||H_0|}-div(V- V_0) \right] \delta \tau d\Sigma_r.\]
From Lemma \ref{opt_exp}, the coefficient of the $r^{-l-4}$ term of $div(f\nabla\tau)-\Delta \sinh^{-1}\frac{(\Delta\tau)f}{|H_r||H_0|}-div(V- V_0)$ is
\[\begin{split}&\widetilde{div}(f^{(-2)}\widetilde{\nabla}g+f^{(-l-3)}\widetilde{\nabla}\tau^{(1)})-\frac{1}{4} \widetilde{\Delta} (f^{(-l-3)}\widetilde{\Delta} \tau^{(1)}+f^{(-2)}\widetilde{\Delta}g)\\
&+\frac{1}{2}\widetilde{\Delta}(\widetilde{\Delta}+2)\hat{\tau}^{(-l-1)}+{G}( b^{(0)}, \cdots b^{(-l)}, \hat{\tau}^{(0)}, \cdots \hat{\tau}^{(-l)})\end{split}\] where $g=\sum_{i=1}^3\sum_{\alpha=0}^3 b_{0\alpha}^{(-l-1)} B_{\alpha i}\tilde{X}_i$.
On the other hand, the $r^{-l}$ term of $\tau$ is
\begin{equation}\tau^{(-l)}=g+B_{00} \hat{\tau}^{(-l)}+G( b^{(0)},\cdots, b^{(-l)}, \hat{\tau}^{(0)},\cdots, \hat{\tau}^{(-l+1)}).\end{equation}
When we consider the variation of $\hat{\tau}^{(-l)}$, $\delta\hat{\tau}^{(-l)}$, the leading term term of the second variation is of the order $r^{-2l-3}$ with coefficient
\[\frac{B_{00}}{2}\int_{S^2} [\widetilde{\Delta}(\tilde{\Delta}+2)\delta \hat{\tau}^{(-l)}]\delta \hat{\tau}^{(-l)} dS^2.\]
We may assume $\int_{S^2} \delta \hat{\tau}^{(-l)} dS^2=0$ by normalization.
By decomposing $\delta \hat{\tau}^{(-l)}$ into sum of eigenfunctions of $S^2$ and noting that the first non-zero eigenvalue of $S^2$ is $-2$, this is always positive.
Varying $b^{(-l-1)}$ is equivalent to varying $g$ and $\delta g=\sum_{i=1}^3\sum_{\alpha=0}^3 \delta b_{0\alpha}^{(-l-1)} B_{\alpha i}\tilde{X}_i$.
The coefficient of the leading $r^{-2l-4}$ term of the second variation with respect to $b^{(-l-1)}$ is then
\[\int_{S^2} [\widetilde{div}(f^{(-2)} \widetilde\nabla \delta g) - \frac{\widetilde \Delta(f_2\widetilde \Delta \delta g)}{4} ]\delta g dS^2.\]
Using integration by part, we derive that this is equal to
\[-\int_{S^2} f^{(-2)} [ |\widetilde\nabla \delta g|^2 + (\delta g )^2] dS^2.\]
This simply gives a positive multiple of quasilocal energy since
\[ |\widetilde\nabla \delta g|^2 + (\delta g )^2 = \sum_{i}(\sum_\alpha \delta b_{0\alpha}^{(-l-1)} B_{\alpha i})^2 \] is a positive constant and $-\int_{S^2} f^{(-2)} dS^2=\frac{8\pi e}{\sqrt{1+|c|^2}}$.
|
\section{\label{sec:Introduction}Introduction}
LuCuGaO$_4$ encompasses several topics of interest in condensed matter physics. It has the same crystal structure as the multiferroic LuFe$_2$O$_4$~\cite{ikeda_2005}: both contain triangular bilayers (Fig.~\ref{fig:LuCuGaO4topology}), a two-dimensional geometrically frustrated lattice which has so far not been widely studied. Theories of resonating valence bond (RVB) states and other spin liquid ground states~\cite{normand_2009} typically address low-dimensional frustrated magnets with $s=1/2$ and experimental realizations are sought-after for comparison. In LuCuGaO$_4$ the magnetic Cu$^{2+}$ ions have $s=1/2$, suggesting the possibility of quantum magnetic phenomena, but are strongly diluted by diamagnetic Ga$^{3+}$ giving a material with quenched disorder and charge frustration~\cite{anderson_1956,cava_1998}.
At a classical level, spins with antiferromagnetic couplings residing on lattices with triangular geometry are frustrated: all pairwise interactions cannot be minimized simultaneously. Many other systems map onto such models, for example the configurations of two cations on such a lattice are identical to those of antiferromagnetically coupled Ising spins on the same lattice~\cite{anderson_1956}, a situation which is known as charge frustration. In LuCuGaO$_4$ the Cu$^{2+}$ and Ga$^{3+}$ cations are distributed over a single site forming well separated triangular bilayers. A single bilayer contains two opposed triangular lattices (see Fig.~\ref{fig:LuCuGaO4topology}) and the shortest distance between cations is between opposite faces of the bilayer, forming a puckered honeycomb lattice. The combination of honeycomb and triangular geometries may be conveniently viewed as a $J_1-J_2$ model on the honeycomb lattice~\cite{fouet_2001,mattsson_1994}, which is a frustrating geometry for magnetic order (see Fig.~\ref{fig:LuCuGaO4topology}). LuCuGaO$_4$ is therefore expected to exhibit both charge and magnetic frustration.
The combination of a frustrating lattice geometry and $s=1/2$ is expected to lead to unconventional groundstates such as a spin liquid, as the energy can be further reduced below the classical groundstate by quantum fluctuations (many types and definitions of spin liquids are reviewed by Normand in Ref.~\onlinecite{normand_2009}). Whilst geometrically frustrated magnets have typically been viewed as clean systems in which the effect of frustration can be isolated (in contrast to spin glasses, where structural disorder is also inevitable), it has become apparent that weak structural disorder or non-magnetic impurities can determine the low temperature behaviour of frustrated magnets, where any perturbation of a fragile groundstate is magnified in importance. Examples include the pinning of dimers around non-magnetic impurities in the $s=1/2$ kagome Heisenberg antiferromagnet~\cite{rousochatzakis_2009,lauchli_2007,dommange_2003} or the strain-induced spin glass transition in the pyrochlore Heisenberg antiferromagnet with weak bond disorder~\cite{saunders_2007}.
\begin{figure}
\centering
\includegraphics[trim=40 520 110 80,clip=true,scale=0.5]{fennell_lucu_fig1a.pdf}\\
\includegraphics[trim=40 480 110 120,clip=true,scale=0.5]{fennell_lucu_fig1b.pdf}\\
\caption{The topology of the triangular bilayers formed by Cu$^{2+}$ and Ga$^{3+}$ cations in LuCuGaO$_4$. The two faces of the bilayer are shown in black and white. The shortest bonds (called $J_1$ in the text) are those that link the two faces of the bilayer forming a puckered honeycomb lattice (grey). Although the honeycomb lattice is bipartite and cation order is therefore possible, the side views (middle and bottom) show that the puckering of this lattice means that such order would cause charge separation across the layer. The bond distance within the faces of the bilayer are slightly longer and form triangular lattices (black and dashes) on which cation (and magnetic) order will be highly frustrated.}
\label{fig:LuCuGaO4topology}
\end{figure}
LuCuGaO$_4$ is a member of the series $RMM'$O$_4$, where $R$ is a small rare earth such as Yb$^{3+}$ or Lu$^{3+}$, $M$ is a late $3d$ transition metal cation with $+2$ charge and $M'$ is a $+3$ cation from the late transition metals or main group. Much attention has been focussed on LuFe$_2$O$_4$, where charge ordering of Fe$^{2+}$ and Fe$^{3+}$ is possible and is thought to underlie the multiferroicity~\cite{ikeda_2005}. Numerous examples with two different cations have also been characterized. As well as LuCuGaO$_4$, YMnFeO$_4$, Yb$M$FeO$_4$ ($M$ = Mg, Fe, Co, Cu), YbCuGaO$_4$, Lu$M$FeO$_4$ ($M$ = Zn, Fe, Co, Cu) and LuCoGaO$_4$ have all been studied~\cite{wiedenmann_1983,iida_1990,iida_1990_2,isobe_1990,tanaka_1990,iida_1991,matsumoto_1992,ikeda_1995,tanaka_1995,todate_1997,cava_1998,todate_1998_jpcm,todate_1998_prb,sunaga_2001,dabkowska_2002,yoshii_2007}. Structurally, all are characterized as described above, with randomly distributed $M$ and $M'$ ions on the bilayer (the crystal structure is illustrated in Fig.~\ref{fig:LuCuGaO4xtalstructure} and described in greater detail below). With the exception of $R$CuGaO$_4$ ($R$ = Yb, Lu)~\cite{cava_1998} all these materials exhibit a splitting of field cooled and zero field cooled magnetic susceptibilities at temperatures of order 20 K (and in some cases considerably higher), indicative of freezing or spin glass transitions. Previously the most extensively investigated stoichiometry was LuMgFeO$_4$~\cite{wiedenmann_1983,ikeda_1995,tanaka_1995,todate_1997,todate_1998_prb}. Studies using magnetization measurements, neutron scattering and M\"ossbauer spectroscopy established that there are short range magnetic correlations associated with frustrated two-dimensional ordering and suggested that random dilution leads to the formation of clusters and two types of site: those lying in the body of clusters and having a full complement of magnetic neighbors, and those lying on extended branches of clusters and having just one or two neighbors.
Previous studies of LuCuGaO$_4$ also suggest that the copper and gallium ions are randomly distributed on the sites of the bilayer and consequently that LuCuGaO$_4$ is both spin and charge frustrated~\cite{cava_1998}. Indeed, despite a Curie-Weiss temperature of $\theta_{CW} = -69$ K, no magnetic ordering or freezing transition is observed down to 0.4 K where a broad peak in the ac-susceptibility is currently attributed to a spin glass transition \cite{cava_1998}. The primary difference between those $RMM'$O$_4$ with high temperature freezing transitions and LuCuGaO$_4$ is the presence of (only) $s = 1/2$ moments. In this paper we use polarized neutron scattering, $\mu$SR and neutron spectroscopy to produce a detailed microscopic characterization of LuCuGaO$_4$ which suggests that such a state exists, even in the presence of strong structural disorder.
\begin{figure}
\centering
\includegraphics[trim=300 290 450 230,clip=true,scale=0.3]{fennell_lucu_fig2.pdf}
\caption{Crystal structure of LuCuGaO$_4$ (the $c$-axis is vertical on the page, the projection axis is the $a-b$ diagonal) as obtained from the Rietveld refinement shown in Fig.~\ref{fig:LuCuGaO4_structure}. The bilayers (black bonds) are formed of edge-sharing $M^{(')}$O$_5$ trigonal bipyramids ($M^{(')}$ are shown in green, O$^{2-}$ in red). The bilayers are separated by layers of edge-sharing LuO$_6$ octahedra. The Lu$^{3+}$ cations randomly occupy one of two sites displaced from the centre of each octahedron (blue).
We suggest (see discussion) that the minimization of the Coulomb interaction amongst the cations is frustrated, and that short range cation correlations are formed amongst the Cu$^{2+}$ and Ga$^{3+}$ on the bilayers. The siting of a particular Lu$^{3+}$ ion would also be controlled by the cation distributions in a triangle on the face of the bilayer immediately above and below the LuO$_6$ octahedra.}
\label{fig:LuCuGaO4xtalstructure}
\end{figure}
\section{\label{sec:Experimental_methods}Experimental methods}
A 20 gram powder sample of LuCuGaO$_4$ was prepared by reaction of the appropriate stoichiometric quantities of Lu$_2$O$_3$ (Alfa Aesar 99.99\%), CuO (Sigma Aldrich, 99.995\%) and Ga$_2$O$_3$ (Alfa Aesar 99.999\%). CuO and Lu$_2$O$_3$ were pre-treated by heating in air for 24 hours at 900 and 1200 $^\mathrm{o}$C respectively. The starting materials were ground together, pelletized and heated in air at 1050 $^\mathrm{o}$C for a total of 120 hours with 1 intermediate grinding. The field cooled and zero field cooled (FC/ZFC) dc-susceptibility was measured between 1.8 and 300 K in a field of 1000 G using a Quantum Design MPMS-7 SQUID magnetometer.
Neutron powder diffraction patterns were obtained at 70 K and 1.5 K using the D1A diffractometer at the ILL with an incident wavelength of 1.91 \AA~and standard Orange cryostat. Polarized neutron scattering using the $xyz$ technique was performed on D7\cite{stewart_2009} at the ILL using a wavelength of 4.8 \AA~at temperatures of 0.08, 0.5, 5 and 50 K. Low temperatures were achieved by using a dilution refrigerator insert in a standard Orange cryostat. The sample was contained in a copper can with $^3$He exchange gas to provide thermalization below 1 K. An inelastic neutron scattering experiment was carried out on the direct geometry chopper spectrometer IN4 at the ILL, at temperatures of 1.6, 5 and 50 K using an incident energy of 17 meV.
A zero field and longitudinal field muon spin rotation study was conducted between 0.06 and 50 K with fields up to 2500 G using the MuSR instrument at the ISIS pulsed muon source. The sample was pressed tightly in a silver disk-shaped holder approximately 2 mm deep and cooled with a dilution refrigerator.
\section{\label{sec:Results}Results}
\subsection{\label{sec:Characterization}Characterization}
We have used powder neutron diffraction and magnetic susceptibility measurements to establish the basic quality of the sample and compare to previous studies. The powder neutron diffraction data show that the sample is phase pure and highly crystalline. The Rietveld method was used to refine three structural models as described in Ref.~\onlinecite{cava_1998} (see also Ref.~\onlinecite{matsumoto_1992}). The crystal structure has the space group $R\overline{3}m$ and consists of double layers of edge-sharing $M^{(')}$O$_5$ triangular bipyramids connected by triangular layers of RO$_6$ octahedra. All three models have Cu$^{2+}$ and Ga$^{3+}$ randomly distributed on the $6c$ site, and differ with respect to the position of Lu$^{3+}$. The $R^{3+}$ ions lie at the centre of distorted octahedra of O$^{2-}$ ions, directly above and below the centre of triangles of transition metal ions. The simplest model has Lu$^{3+}$ positioned at $(0,0,0)$, midway between adjacent bilayers. The more complex models allow anisotropic motion of the Lu$^{3+}$ ion, or a shift to a lower symmetry site with random half occupation at $(0,0,\pm z)$ ($z = 0.009\pm0.0002$, corresponding to a displacement of 0.22 \AA). In this case, the position of each Lu$^{3+}$ is thought to be controlled by the distribution of cations on the coordinating triangles of adjacent bilayers~\cite{cava_1998}. As in Ref. \onlinecite{cava_1998} the third model provides the best combination of fit and realistic parameters. There is no indication of any structural change between 70 K and 1.5 K. The crystal structure is illustrated in Fig.~\ref{fig:LuCuGaO4xtalstructure}.
We have fitted the inverse magnetic susceptibility to the Curie-Weiss law between 50 K and 300 K, obtaining $\theta_{CW} = -62.1\pm1.8$ K . This is somewhat smaller than the previously reported value (-69 K), but in Ref.~\onlinecite{cava_1998} the fitting range appears to be restricted to 50 K to 150 K. The magnetic moment obtained from the Curie constant is $2.2 \pm 0.03 ~\mu_\mathrm{B}$ per Cu$^{2+}$ (the expected value for $S = 1/2$ is 1.7 $\mu_\mathrm{B}$ atom$^{-1}$), and the effective moment at 295 K is $\mu_{\mathrm{eff}}\equiv\sqrt{8\chi T}=1.73 \pm 0.06 ~\mu_\mathrm{B}$ per Cu$^{2+}$. The effective moment falls at low temperature where the inverse susceptibility becomes non-linear, indicating increasing antiferromagnetic spin correlations. At these temperatures there is no discernible splitting between field cooled and zero field cooled measurements.
We have also fitted the entire temperature dependence of the inverse susceptibility with a model of ``orphan'' spins~\cite{schiffer_1997}. This is essentially a Curie-Weiss law with an additional Curie contribution due to isolated spins, which is expected to become dominant at low temperature. Since these spins are, in principle, isolated, they should have $\theta_{CW} = 0$. In this case $\theta_{CW1} = -67 \pm 2.9$ K, $\mu_{1} = 2.1\pm0.008~ \mu_\mathrm{B}$, $\theta_{CW2} = 1.5 \pm 0.8$ K and $\mu_{2} = 0.33 \pm 0.06~\mu_\mathrm{B}$. The downward curving form of $\chi^{-1}$ is a rather general feature of many frustrated antiferromagnets (see Ref.~\onlinecite{isoda_2008} and references therein) and the reality of the orphan spins remains debatable\cite{schiffer_1997,isoda_2008}.
\begin{figure}
\centering
\includegraphics[clip = true, viewport = 0.0in 4.5in 7.1in 9.15in, width=1.0\columnwidth]{fennell_lucu_fig3.pdf}
\caption{Neutron powder diffraction at 1.5 K. The experimental data are the red crosses. The fitted line (green) was obtained by using the Rietveld method to refine the model M3 of Ref.~\onlinecite{cava_1998}. The corresponding structure is shown in Fig.~\ref{fig:LuCuGaO4xtalstructure}. The calculated peak positions are indicated by the black ticks and the difference between experimental and fitted profiles by the magenta line.}
\label{fig:LuCuGaO4_structure}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{fennell_lucu_fig4.pdf}
\caption{Susceptibility and (inset) inverse susceptibility for FC dc SQUID measurements at 1000 G. The solid line is a fit to the Curie-Wiess law between 200 and 60 K. ZFC results are identical to FC within experimental error. Also shown is the fit to a model with ``orphan'' spins~\cite{schiffer_1997}.}
\label{fig:LuCuGaOsus}
\end{figure}
\subsection{\label{sec:XYZneutrons}Polarized Neutron Scattering}
The $xyz$-polarization analysis technique was used to separate nuclear, spin incoherent and magnetic contributions to the total scattering~\cite{stewart_2009}. The temperatures were selected to compare the paramagnetic regime, where the susceptibility is still described by the Curie-Weiss law (50 K), with the regime in which the inverse susceptibility becomes non-linear (5 K), and the low temperature regime around the proposed spin glass transition. Furthermore, the separation of the contributions allows us to distinguish the origin of any short range order (i.e nuclear or magnetic). The nuclear channel (not shown) contains only Bragg scattering. There is a clear change in magnetic scattering between 50 K and all lower temperatures studied. However, there is no significant change in the low temperature scattering from 5 K down to 0.08 K, in particular between 0.5 and 0.08 K across the proposed spin glass transition, so we have combined the data sets from 0.08, 0.5 and 5 K. As shown in Fig. \ref{fig:LuCuGaO4D7deVries}, at 50 K the magnetic scattering can be approximately fitted by the form factor of Cu$^{2+}$, as expected for a paramagnet. At low temperature the form factor response at small $|Q|$ is suppressed and a broad feature appears, centered around 1.25 \AA$^{-1}$ . There is some additional scattering at small $|Q|$, above the fitted form factor, which we attribute to air scattering.
The magnetic scattering is very weak and has therefore not been observed using conventional neutron powder diffractometers (there is no sign of this feature in Fig.~\ref{fig:LuCuGaO4_structure} or the earlier study of LuCuGaO$_4$ \cite{cava_1998}).
Using an expression for the structure factor of disordered near neighbor antiferromagnetic dimers,
\begin{equation}
I(Q) \propto F^{2}(Q)\left(1 - \frac{\sin(Qd)}{Qd}\right)
\label{AFMdimers}
\end{equation}
a fit was obtained, as shown in Fig. \ref{fig:LuCuGaO4D7deVries}, with $F^2(Q)$ being the form factor for Cu$^{2+}$ and $d$ the distance between the AFM correlated Cu$^{2+}$. No appreciable difference in the fit is observed if the value of the Cu-Cu distance is 3.44~\AA ~(the distance between Cu-Cu (or Cu-Ga etc) in the triangular layers), 3.04~\AA ~(the Cu-Cu distance across the bilayer), or a weighted average of the two (i.e. due to neighbors on both faces of the bilayer).
This form suggests that there are magnetic correlations between nearest neighbor Cu$^{2+}$ ions, which develop in the non-linear region of $\chi^{-1}$. The position and form agrees with other members of the series where the larger moments of Fe$^{3+}$ or Co$^{2+}$ render the diffuse magnetic scattering more readily observable~\cite{wiedenmann_1983,cava_1998}. In LuMgFeO$_4$, the dominant feature is also a peak at $Q\approx 1.25$ \AA$^{-1} $, but the diffuse scattering profile is much more detailed and correlations could be detected up to the twelfth neighbor shell. This fitting strategy involving further neighbor correlations~\cite{wiedenmann_1983} produced unphysical values for further neighbor correlation functions in LuCuGaO$_4$, so we conclude that the correlations are very short ranged.
We gain no particular understanding of the anomaly previously observed in the susceptibility and specific heat at 0.4 K~\cite{cava_1998} as there is no difference between the data measured at 0.08 K and 5 K. This is also a feature of LuMgFeO$_4$ where no change in the neutron scattering intensity is seen across the susceptibility cusp~\cite{ikeda_1995,todate_1998_jpcm}.
\begin{figure}
\includegraphics[width=1.0\columnwidth]{fennell_lucu_fig5.pdf}
\caption{Separated magnetic component from $xyz$ polarization analysis experiment. The combined low temperature results (0.08 to 5 K) fit well to a model of nearest neighbor antiferromagnetic correlations. At 50 K LuCuGaO$_4$ is paramagnetic and the scattering approximately follows the magnetic form factor for Cu$^{2+}$~\cite{brown_ff}.}
\label{fig:LuCuGaO4D7deVries}
\end{figure}
\subsection{\label{sec:muSR_res}$\mu$SR}
The development of the spin correlations at low temperature was further investigated using $\mu$SR. The decay of the implanted muon polarization was measured in zero field as a function of temperature and as a function of magnetic field at selected temperatures. $\mu$SR is sensitive to magnetic fluctuations with nuclear and electronic origins on a timescale of 10$^{-5}$ - 10$^{-7}$ seconds. In a paramagnet the electronic fluctuations are mostly very rapid and lie outside the window of the $\mu$SR experiment. Normally in such a situation, any observed relaxation is a consequence of the nuclear component. The nuclear component of the muon relaxation can usually be decoupled with a field of the order of $10-100$ G. We have applied longitudinal fields at several temperatures. At 50 K, where the polarized neutron scattering also shows the sample to be paramagnetic, the $\mu$SR signal is decoupled in fields of less than 100 G, suggesting the signal is entirely due to relaxation of the nuclear spins. However, at low temperatures, where the polarized neutron scattering shows that spin correlations are developing, a field of 2500 G is insufficient to fully decouple the relaxation. This indicates that at low temperatures there is an increasing electronic contribution, i.e. that as the temperature is decreased, magnetic fluctuations are slowing and falling into the muon time window.
We have therefore fitted the zero-field muon polarization decay for different temperatures (examples are shown in the inset of Fig.~\ref{fig:LuCuGaO4lambdavstemp}) with two components: a temperature independent Kubo-Toyabe function to represent the nuclear component, combined with a stretched exponential relaxation of the type $\exp ((-\lambda t)^{\beta})$, where $\beta$ is a constant and $\lambda$ is the muon depolarization rate.
The temperature dependence of $\lambda$ is shown in Fig.~\ref{fig:LuCuGaO4lambdavstemp}. It can clearly be seen that the depolarization rate begins to increase at $T \approx 10$ K, as the inverse susceptibility becomes non-linear and the peak appears in the diffuse neutron scattering. There is a continual and relatively rapid increase until it reaches a plateau at $T\approx 0.4$ K, close to the proposed spin glass transition.
\begin{figure}
\centering
\includegraphics[clip = true, width=1.0\columnwidth]{fennell_lucu_fig6.pdf}
\caption{The muon depolarization rate in zero field as a function of temperature obtained from fits to a temperature independent nuclear response and stretched exponential electronic contribution. Typical fitted muon responses at different temperatures are shown in the inset.
}
\label{fig:LuCuGaO4lambdavstemp}
\end{figure}
\subsection{\label{sec:IN4_ILL}Inelastic neutron scattering}
A limited inelastic neutron scattering investigation was also made. The results are shown in Fig. \ref{fig:LuCuGaO4IN4}. The peak in the magnetic scattering previously described and attributed to nearest neighbor AFM correlations is at $|Q| \approx 1.25 $ \AA$^{-1}$. In Fig.~\ref{fig:LuCuGaO4IN4} it can be seen that at 1.7 K inelastic scattering extends from this position. There is no sign of any sharp feature, often visible in powder samples of ordered magnets, even when dispersion surfaces are averaged. The intensity is confined to this single feature, which is rather broad and weakly dependent on the energy transfer. Constant energy cuts further highlight the fact that the magnetic part of the inelastic signal can be fitted by the antiferromagnetic dimer structure factor at all energy transfers (Fig.~\ref{fig:LuCuGaO4IN4}b). A constant $|Q|$-cut shows no indication of a gap (Fig.~\ref{fig:LuCuGaO4IN4}c).
\begin{figure}
\centering
\includegraphics[trim=1 1 1 1,clip=true,scale=0.35]{fennell_lucu_fig7a.pdf}
\includegraphics[trim=1 1 1 1,clip=true,scale=0.35]{fennell_lucu_fig7b.pdf}
\includegraphics[trim=1 1 1 1,clip=true,scale=0.35]{fennell_lucu_fig7c.pdf}
\caption{a: $S(Q,\omega)$ at 1.7 K with $E_i = 17$ meV.
b: Constant energy cuts through the feature at $|Q|\approx 1.2$ \AA$^{-1}$ visible in a, with fits to the nearest neighbor AFM structure factor. The peak around $Q = 1.2$ \AA$^{-1}$ fits well to this model at all energies. There is an increasing phonon contribution at higher $|Q|$. c: Cut at $|Q|= 1.2$ \AA$^{-1}$ and 1.9 \AA$^{-1}$ (background).}
\label{fig:LuCuGaO4IN4}
\end{figure}
\section{\label{sec:Discussion}Discussion}
A spin liquid may be defined by analogy with an actual liquid, as a spin system with short ranged, gapless, dynamic correlations~\cite{fak_2008}. We seek to make the case that despite strong quenched disorder and the previous description as a spin glass, LuCuGaO$_4$ is actually a spin liquid. We first discuss the structural disorder and secondly the liquid-like spin dynamics.
In all the models fitted to the powder diffraction data the Cu$^{2+}$ and Ga$^{3+}$ ions were treated as randomly distributed on the triangular bilayers, as in all previous works on related materials. However, although the ordering of two different cation species on the triangular bilayer is frustrated (see Fig.~\ref{fig:LuCuGaO4topology}), a completely random distribution seems unlikely: triangles containing entirely Cu$^{2+}$ or Ga$^{3+}$ will incur a Coulombic cost. We therefore advance the suggestion that in these materials the cations must be correlated. The nearest neighbor part of the Coulomb energy will be minimized by configurations of the two cations which map to the groundstates of an Ising antiferromagnet on the same lattice, as in other systems with two cation species on a frustrated lattice~\cite{anderson_1956}. Structural diffuse scattering would be expected as a result of these short range correlations. Unfortunately, the scattering lengths of copper and gallium are extremely similar ($b_{\mathrm{Cu}} = 7.718$ barn and $b_{\mathrm{Ga}} = 7.288$ barn) so direct detection of short range cation order is effectively impossible.
The importance of this point lies in the resulting quenched disorder configuration. In the case of completely random cations, the topology of the spin system can be compared to percolating clusters, as in Ref.~\onlinecite{ikeda_1995}. In the case of correlated cations, it appears that the topology of the spin system would take the form of branching loops~\cite{banks_unpub}. Less direct evidence exists for the formation of short-range correlations amongst the cations. For example, one can argue that the existence of two defined positions for the Lu$^{3+}$ ions is due to the minimization of electrostatic interactions amongst both the layer and interlayer cations. Without direct evidence for cation correlations, we conclude that the copper and gallium are certainly disordered, and that there is a possibility that they realize a non-trivial network structure.
We turn now to the magnetic correlations. Although we have fitted the peak in the polarized neutron scattering data with the so-called dimer structure factor, we note that this is actually a general description of purely near-neighbor antiferromagnetic correlations. The inelastic scattering data shows that at 1.5 K dynamical correlations exist. The diffuse scattering observed in the polarized neutron scattering experiments may be regarded as integrating over the quasi-elastic part of these fluctuations. The temperature independence of the diffuse peak implies that the correlations remain dynamic at all temperatures measured: if the spins became significantly more static, one would expect to see transfer of spectral weight into the elastic line, certainly resulting in stronger diffuse scattering and possibly a sharper peak shape. The literal interpretation of the dimer structure factor is further precluded as there is no sign of significant singlet formation: the susceptibility does not decrease, there is no clear sign of a singlet-triplet gap in the inelastic neutron scattering data, and the $\mu$SR data implies that the local field in the sample is finite and fluctuating at all measured temperatures.
Canonical spin glasses such as AgMn(0.5 at. \%) have been extensively investigated by $\mu$SR~\cite{keren_1996}. The depolarization rate increases as the system approaches $T_g$ but then falls rapidly. In contrast, in LuCuGaO$_4$ the depolarization rate increases steadily as the spin correlations build up, but there is no evidence of freezing. We therefore also rule out the existence of a spin glass transition.
The literature provides various examples of similar neutron scattering responses, for example energy independent continua at a particular $|Q|$ can be found in deuteronium jarosite ((D$_3$O)Fe$_3$(SO$_4$)$_2$(OD)$_6$, kagome lattice)~\cite{fak_2008}; SCGO (SrCr$_{8-x}$Ga$_{4+x}$O$_{19}$, $x=0.87$, pyrochlore bilayer)~\cite{mutka_2004}; and Y$_{0.5}$Ca$_{0.5}$BaCo$_4$O$_7$ (kagome)~\cite{schweika_2007}. However, the behaviour of LuCuGaO$_4$ is most strikingly similar to that of ZnCu$_3$(OD)$_6$Cl$_2$ (Herbertsmithite), which is thought to be a structurally perfect $s=1/2$ kagome antiferromagnet. In Herbertsmithite, $\chi^{-1}$ curves below the Curie-Weiss straight line, which may be attributable to Zn-Cu disorder~\cite{helton_2007,devries_2009}. Detailed $\mu$SR investigations~\cite{mendels} found a stretched exponential relaxation and increasing depolarization rate at low temperatures which was attributed to slowly relaxing defects, with the remainder of the spins fluctuating too rapidly to be observed by $\mu$SR. The neutron scattering response \cite{devries_2009} consists of a non-dispersive, gapless feature extending across all observed energies at the wavevector transfer characteristic of the near neighbor distance. It can be fitted by the dimer structure factor at all available energies and measured temperatures between 2 and 120 K.
Detailed investigation of the behavior of $\chi''(Q,T)$ extracted from the neutron scattering data shows that it does not have the expected $E/T$ scaling expected in theories of spin liquids (this is also the case in deuteronium jarosite~\cite{fak_2008}). While we do not have sufficient data to compare with these observations, we suggest that the overall similarity of the two compounds means that the same physics is realized and may be a general property of two-dimensional, $s=1/2$ spin liquids, which is robust to the presence of disorder.
\section{\label{sec:Conclusion}Conclusion}
We have extended the characterization of LuCuGaO$_4$ and shown that it is not a spin glass, but a spin liquid. Although a model of randomly disordered cations is sufficient to explain diffraction data, this may be due to the almost identical scattering lengths of copper and gallium and further investigation of the topological properties of these materials promises to be interesting. At moderate temperatures the spin correlations are shown to be dynamic and gapless. There is a complete absence of spin freezing down to the lowest available temperature in LuCuGaO$_4$, suggesting that the dynamic correlations persist despite the structural disorder. LuCuGaO$_4$ therefore provides another opportunity to investigate the scaling behaviour of $\chi(Q,T)$, demonstrated to be important in other two-dimensional spin liquids.
\begin{acknowledgments} We thank Sebastien Turc for assistance with cryogenics on D7 and the EPSRC and STFC (UK) for funding.
\end{acknowledgments}
|
\section{Introduction}
Thermal control in spintronic devices and MRAM applications has in recent years been of great interest due to the associated increase of stability and decrease in power consumption \cite{ref:Prejbeanu, ref:Beech, ref:Daughton}. Recently, thermally excited oscillations in nanocontacts, reaching frequencies of the order of GHz, have been predicted \cite{ref:Kadigrobov}. In this model a ferromagnetic (FM) film is separated from a small FM grain by a point contact having a diameter of a few nanometers. Due to the high current densities reached in such a small area, the FM region within the point contact reaches very high temperatures. When the local temperature is higher than the FM Curie point the exchange coupling through the point contact becomes vanishingly small.
The model is based on two premises. First, a thermally controlled exchange coupling between two FM regions. Second, an increase in resistance when the FM regions decouple. The first criterion can be met in a metallic system with two strong ferromagnets separated by a weakly FM spacer. If this can be combined with a large change in resistance both criteria would be met. To date, the largest changes of resistance obtained in an all metallic structure have been from giant magnetoresistance \cite{ref:Binasch, ref:Baibich} (GMR).
In this work we investigate the possibility of decoupling two strong ferromagnets separated by a weakly ferromagnetic Ni-Cu alloy. To verify if the Ni-Cu alloy, in its paramagnetic phase, can be used as a GMR spacer we study the effects of adding nickel to a copper spacer in a spin-valve structure on the interlayer exchange coupling and GMR. Finally, we design and implement an improved thermionic spin-valve structure, in which the switching and the read out layers are separated.
\section{Experimental details}
All films were deposited on thermally oxidized Si substrates using magnetron sputtering at a base pressure better than $5\cdot10^{-8}$ Torr. The argon pressure during sputtering was kept at 3 mTorr. To demonstrate thermally controlled exchange coupling samples with structure Cu(90)/Ni$_{80}$Fe$_{20}$(8)/Co$_{90}$Fe$_{10}$(2)/ Ni$_{x}$Cu$_{1-x}$(t)/Co$_{90}$Fe$_{10}$(5)/Ta(10) (thickness in nanometers) were deposited. Three different thicknesses, $t = 10, 20, 30$ nm, of the weakly ferromagnetic Ni-Cu alloy were used for studying the interlayer exchange in the system. Variation in $x$ was obtained by cosputtering Ni and Cu onto Si substrates that had been cut into 90 x 10 mm strips. In this way a compositional gradient was created along the Si strips ranging from $x=0.2$ to $x=0.9$. By cutting the strips into smaller pieces a series of samples with different Curie temperatures were obtained.
In order to perform magnetic characterization the samples were placed in a looptracer equipped with a thin-film heater. The temperature was controlled through a feed back loop using a type-T thermocouple in close contact with the samples. Further investigations of the switching behavior at room temperature were performed using a vibrating sample magnetometer (VSM).
To measure the effect of Ni-Cu alloying on GMR, samples with structure Ni$_{80}$Fe$_{20}$(4) / Co$_{90}$Fe$_{10}$(1) / Ni$_{x}$Cu$_{1-x}$(3.5) / Co$_{90}$Fe$_{10}$(2) were deposited. The Ni-Cu alloys were cosputtered from Ni and Cu targets such that the stoichiometry, $x$, could be varied by controlling the relative difference in sputtering rates.
To measure current in plane (CIP) GMR, thin Al wires were bonded to the top of the samples. Before electrical measurements
the separate switching of the NiFe/CoFe bi-layer and CoFe top layer was confirmed using a magnetometer.
\section{Results and Discussion}
A Ni-Cu alloy was chosen for the weakly FM spacer because of the well known dependence of Curie point on the Ni concentration \cite{ref:Hicks, ref:Dutta, ref:Sousa}. The Ni-Cu spacer is used to separate a magnetically softer NiFe/CoFe bi-layer from a magnetically harder CoFe layer. Here NiFe and CoFe stand for Ni$_{80}$Fe$_{20}$ and Co$_{90}$Fe$_{10}$, respectively. By cosputtering Ni and Cu, a series of samples with different Curie temperatures were obtained.
\begin{figure}[htbp]
\centering
\includegraphics[width=3in]{FIG1.eps}
\caption{Normalized magnetization versus applied magnetic field for a sample structure NiFe/CoFe/Ni$_{x}$Cu$_{1-x}$(30)/ CoFe, x $\approx$ 0.7 at (a) room temperature and (b) 110$^{\circ}$C. Upon heating the Ni-Cu alloy goes through a ferromagnetic to paramagnetic phase transition and the NiFe/CoFe bi-layer decouples from the CoFe layer.}
\label{fig:Fig1}
\end{figure}
\subsection{Thermally controlled interlayer exchange coupling}
Fig. 1 (a) shows the magnetization loop for a sample with a 30 nm thick Ni-Cu layer having a Curie point just above room temperature, $x \approx 0.7$. The shape of the curve indicates that the strongly FM layers are weakly exchange coupled through the Ni-Cu alloy. Two distinct regions can be seen --- similar to the magnetic state of a spring-magnet \cite{ref:Davies}. At point $A$ in Fig. 1 (a) the magnetic moment of the soft NiFe/CoFe layer starts to rotate in the external magnetic field. This is a reversible rotation due to the exchange coupling through the Ni-Cu alloy to the harder CoFe layer. The reversible switching continues until point $B$ is reached. By comparing the magnitude of the magnetization at 25 Oe (point B) in Fig. 1 (a) with that in Fig. 1 (b) in which the NiFe/CoFe layer is heated to 110$^{\circ}$C and thereby decoupled from the CoFe layer, it can be seen that at room temperature the NiFe/CoFe layer has not yet finished rotating 180$^{\circ}$ when the hard layer switches. At point $B$ the torque on the CoFe layer is too strong for it to remain in position and an irreversible rotation of all layers follows. This behavior was confirmed by VSM measurements of the same sample at room temperature. Starting at a positive field, high enough so that all the moments were aligned, the magnetization was measured while the external field was swept to -20 Oe and then back again. After the field reversal at -20 Oe the magnetization backtracks the values measured before the reversal. The same behavior was seen for field reversals up to -25 Oe indicating that the rotation is reversible up to this point. For reversal fields any higher than this, the magnetization does not backtrack the values measured before the field reversal. This confirms that the switching behavior is irreversible for fields higher than $\pm$25 Oe.
Fig. 1 (b) shows the same sample at $110^{\circ}$C. The Curie point of the spacer has been reached and the soft and hard FM layers are essentially exchange decoupled as evidenced by the two distinct magnetization transitions at approximately 15 and 45 Oe. As the temperature is reduced to room temperature the two magnetization transitions shift towards each other and the sharp magnetization loop becomes significantly skewed. At room temperature the curve shape is back to the one shown in Fig. 1 (a). This thermally controlled interlayer exchange coupling is perfectly reversible on thermal cycling within the given temperature range.
For samples with the Curie point around room temperature the spacer had to be 20 or 30 nm thick in order to completely diminsh the exchange coupling through the spacer. An explanation for this could be that the alloy is not homogenous after cosputtering at room temperature but contains regions with different Curie points. If the spacer is too thin, these regions could extend to the alloy interfaces and couple the two CoFe films. Another possible explanation is that the alloying is homogenous but the two strong ferromagnets are coupled by exchange interactions through the spacer even at temperatures above the Curie point. It has been indicated that exchange can propagate through paramagnetic regions on length scales of several nanometers \cite{ref:Hernando}, which is believed to be due to enhancement of magnetic order in thin layers caused by the proximity effect of a strong ferromagnet.
\begin{figure}[htbp]
\centering
\includegraphics[width=3in]{FIG2.eps}
\caption{Electron mean free path in bulk Ni-Cu for different Ni concentrations. The data points have been calculated using the Drude model from published data on resistivity measurements at 300, 273 and 250K \cite{ref:Houghton, ref:Ahmad, ref:Schroeder}.}
\label{fig:Fig2}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=3in]{FIG3.eps}
\caption{(a) Current in plane (CIP) giant magnetoresistance (GMR) versus Ni concentration in a NiFe(4)/CoFe(1)/Ni-Cu(3.5)/CoFe(2) spin-valve. The inset shows the normalized magnetization versus applied field for a spin-valve with (b) 29 at.\% Ni in the spacer and (c) 35 at. \% Ni in the spacer.}
\label{fig:Fig3}
\end{figure}
\subsection{CoFe/Ni-Cu/CoFe spin-valve}
To understand the CIP GMR in the above NiFe/CoFe/Ni-Cu/CoFe system, we have to consider the mean free path in the Ni-Cu spacer. When the spacer thickness is much larger than the mean free path the CIP GMR signal vanishes as exp$(-t_{NiCu}/\lambda)$ \cite{ref:Barthelemy}. Here $\lambda$ is the electron mean free path in the spacer material and $t_{NiCu}$ is the spacer thickness. In order to obtain a high GMR signal the thickness of the spacer should be comparable or smaller than $\lambda$. Fig. 2 shows $\lambda$ for different Ni concentrations from our measurements as well as calculated from published experimental data using the Drude model. The mean free path decreases quickly with increasing Ni concentration. In the interesting range of Ni concentrations between 40\% and 70\% where the alloy is weakly ferromagnetic \cite{ref:Hicks}, $\lambda$ is down to $\sim$1 nm. As was detailed in the previous section, our spacers with the Curie point close to room temperature have a minimum thickness of 20 nm in order to completely diminish the interlayer exchange coupling.
Assuming the thickness of the weakly ferromagnetic spacer can be reduced by further material optimization, we next examine how thin a spacer would still provide a measurable GMR signal. The thinnest possible spacer to avoid any significant RKKY coupling is $\sim$3 nm \cite{ref:Parkin}. We therefore choose this thickness and evaluate the CIP GMR versus Ni concentration.
We have fabricated samples with structure NiFe(4)/CoFe(1)/Ni-Cu(3.5)/CoFe(2). CIP GMR measurements for samples with different Ni concentrations are shown in Fig. 3 (a). With pure Cu in the spacer a signal of 3.7 \% is measured at room temperature (5\% at 77 K). When Ni is added to the spacer the signal drops caused by a decrease in $\lambda$. At 29 at. \% Ni the GMR has decreased by more than a factor of three, which we attribute to a sharp decrease of the mean free path on alloying. Fig. 3 (b) shows the magnetization versus applied field for this sample. It can be seen that the switching is still separate at this concentration. At 35 at. \% Ni the magnetic layers couple and the switching is no longer separate, which is shown in Fig. 3 (c). It can thus be concluded that even for a very thin Cu-Ni spacer the CIP GMR signal essentially vanishes for 29 at. \% Ni concentration, where the layers are still decoupled magnetically. The final decrease to zero GMR at 35 at. \% Ni is due to coupling through the spacer and not to a decrease in $\lambda$.
It will be informative to point out that by using current perpendicular to the plane GMR instead of CIP the limiting length scale would be the spin diffusion length \cite{ref:Bass}, $l_{sf}$, and not $\lambda$. However, published experimental data indicate that $l_{sf}$ decrease with the same rate as $\lambda$ and has been measured to be 7.5 nm at 5$^{\circ}$K in an alloy with 22.7\% Ni \cite{ref:Hsu}. This is still below the minimum thickness in our case of 20 nm needed to achieve reliable interlayer decoupling.
\begin{figure}[htbp]
\centering
\includegraphics[width=3in]{FIG4.eps}
\caption{CIP GMR for three different current densities versus applied field measured on a sample structure NiFe(3)/MnIr(15)/CoFe(4)/Cu(3.5)/CoFe(4)/ NiFe(6)/CoFe(2)/NiCu(20)/CoFe(2)/NiFe(3)/ MnIr(12)/ Ta(5). The inset shows a schematic of a device in which thermally controlled exchange coupling is separated from a spin-valve read out. Either giant magnetoresistance or tunnelling magnetoresistance can be used for read out.}
\label{fig:Fig4}
\end{figure}
To overcome the above limitations and use thermally controlled interlayer exchange coupling together with large magnetoresistance we propose a new design in which the GMR read out layer and the weakly FM spacer are separated. A schematic is shown in the inset to Fig. 4. An antiferromagnet (AFM) is used to exchange bias a FM film which works as a reference layer in the spin-valve. The spin-valve uses a metallic spacer for GMR or an insulator for tunneling magnetoresistance. The spin-valve spacer is only used for read out and can be optimized to give the highest possible magnetoresistance signal. The top layer of the spin-valve is exchange coupled through a weakly FM alloy to a top pinned FM. At high temperatures the coupling through the weakly FM alloy is negligible and the top layer of the spin-valve is free to rotate. However, at low temperatures it will be exchange coupled to the top pinned FM. This results in full flexibility when choosing the composition and Curie point of the weakly ferromagnetic alloy while at the same time making it possible to achieve high magnetoresistance.
To demonstrate this new structure we have deposited samples of NiFe(3)/MnIr(15)/CoFe(4)/ Cu(3.5)/CoFe(4)/NiFe(6)/ CoFe(2)/NiCu(20)/CoFe(2)/NiFe(3)/MnIr(12)/Ta(5) with 70 at. \% Ni in the spacer ($T_{C}\approx 100^{\circ}$C). The samples were patterned using photolithography into strips 50 $\mu$m wide and 1 mm long and then bonded at the edges with aluminum wire. The resulting CIP GMR signal for different current densities is shown in FIG. 4. When the current density is increased and the temperature of the device is correspondingly raised due to joule heating, the top layer of the spin-valve decouples from the top pinned CoFe/NiFe bi-layer producing a strong GMR signal. For a current density of $1.3\cdot 10^{6}$ A/cm$^{2}$, $T>T_{C}\approx 100^{\circ}$C, the exchange decoupling is complete and the full CIP GMR signal of 2.5\% for this structure is obtained.
\section{Conclusion}
Thermally controlled exchange coupling between two strongly FM films separated by a weakly FM Ni-Cu spacer is demonstrated. At temperatures higher than the Curie point of the spacer the FM films are decoupled. At lower temperatures the switching behavior can be separated into two regions --- reversible and irreversible.
In a CoFe/Ni-Cu/CoFe spin-valve the CIP GMR signal vanishes due to a sharp reduction of the mean free path on alloying for Ni concentrations above $\sim$30 at. \%. A new design is proposed and demonstrated, combining thermally controlled exchange coupling and large magnetoresistance, which may prove useful for applications in current controlled magneto-resistive oscillators.
\begin{acknowledgments}
This work was supported by EU-FP7-FET-STELE.
\end{acknowledgments}
|
\section{Introduction and notation}
\noindent Statistical models are often defined through estimating
equations
\begin{equation*}
\mathbb{E} \left[g(X,\theta)\right]=0,
\end{equation*}
where $\mathbb{E} [\cdot]$ denotes the mathematical expectation,
$g:=(g_1,\ldots,g_l)^{\top}\in\mathbb{R}^l$ is some specified
vector valued function of a random vector $X\in \mathbb{R}^m$ and
a parameter vector $\theta\in\Theta\subset \mathbb{R}^d$. Examples
of such models are numerous, see e.g.
\noindent\cite{Qin-Lawless1994}, \cite{Haberman1984},
\cite{Sheehy1987}, \cite{McCullagh_Nelder1983}, \cite{Owen2001}
and the references therein. Denoting $M^{1}$ the collection of all
probability measures (p.m.) on the measurable space
$(\mathbb{R}^m,\mathcal{B}(\mathbb{R}^m))$, the submodel
$\mathcal{M}_{\theta}^{1}$, associated to a given value $\theta$
of the parameter, consists of all distributions $Q$ satisfying $l$
linear constraints induced by the vector valued function
$g(.,\theta)$, namely
\begin{equation*}
\mathcal{M}_{\theta}^{1}:=\left\{ Q\in M^{1}~\text{ such that
}~\int g(x,\theta)~dQ(x)=0\right\},
\end{equation*}
with $l\geq d$. The statistical model which we consider can be
written as
\begin{equation}\label{modele}
\mathcal{M}^{1}:=\bigcup_{\theta\in\Theta}\mathcal{M}_{\theta}^{1}.
\end{equation}
\noindent Let $X_{1},...,X_{n}$ denote an i.i.d sample of $X$
with unknown distribution $P_{0}$. We denote $\theta_{0}$, if it
exists, the value of the parameter such that $P_{0}$ belongs to
$\mathcal{M}_{\theta_{0}}^1$, namely the value satisfying
$$\mathbb{E}\left[g(X,\theta_0)\right]=0,$$ and we assume obviously that $\theta_0$
is unique. This paper addresses the two following natural
questions:
\textit{Problem 1}: Does $P_{0}$ belong to the model
$\mathcal{M}^{1}$?
\textit{Problem 2}: When $P_{0}$ is in the model, which is the
value $\theta_{0}$ of the parameter for which
$\mathbb{E}\left[g(X,\theta_{0})\right]=0$? Also can we perform
tests about $\theta_{0}$? Can we construct confidence areas for
$\theta_{0}$?\\
\noindent We note that these problems have been investigated by
many authors. \cite{Hansen1982} considered generalized method of
moments (GMM). \cite{Hansen_Healton_Yaron1996} introduced the
continuous updating (CU) estimate. The empirical likelihood (EL)
approach, developed by \cite{Owen1988} and \cite{Owen1990}, has
been investigated in the context of model (\ref{modele}) by
\cite{Qin-Lawless1994} and \cite{Imbens1997} introducing the EL
estimate. The recent literature in econometrics focusses on such
models; \cite{Smith1997}, \cite{NeweySmith2004} provided a class
of estimates called generalized empirical likelihood (GEL)
estimates which contains the EL and the CU ones.
\cite{Schennach2007} discussed the asymptotic properties of the
empirical likelihood estimate under misspecification; the author
showed the important fact that the EL estimate may cease to be
root $n$ consistent when the functions $g_j$ defining the moments
conditions and the support of $P_0$ are unbounded. Among other
results pertaining to EL, \cite{NeweySmith2004} stated that EL
estimate enjoys optimality properties in term of efficiency when
bias corrected among all GEL estimates including the GMM one.
Moreover, \cite{Corcoran1998} and \cite{Baggerly1998} proved that
in a class of minimum discrepancy statistics (called power
divergence statistics), EL ratio is the only one that is Bartlett
correctable. Confidence areas for the parameter $\theta_{0}$ have
been considered in the seminal paper by \cite{Owen1990}. Problems
1 and 2 have been handled via EL and GEL approaches in
\cite{Qin-Lawless1994}, \cite{Smith1997} and \cite{NeweySmith2004}
under the null hypothesis $\mathcal{H}_0 : P_0\in\mathcal{M}^1$;
the limiting distributions of the GEL estimates and the GEL test
statistics have been obtained under the model and under the null
hypotheses. \cite{Imbens1997} discusses the asymptotic properties
of the EL and exponential tilting estimates under misspecification
and give the formula of the asymptotic variance, using dual
characterizations, without presenting the hypotheses under which
their results hold. \cite{Chen_Hong_Shum2007} give the limiting
distribution of the EL estimate under misspecification as well as
the EL ratio statistic between a parametric model and a moment
condition model. The paper by \cite{Kitamura2007} gives a
discussion of duality for GEL estimates under moment condition
models. \cite{Bertail2006} uses duality to study, under the model,
the asymptotic properties of the EL ratio statistic and its
Bartlett correctability; the author extends his results to
semiparametric problems with
infinite-dimensional parameters.\\
\noindent The main contribution of the present paper is the
precise characterization of duality for a large class of estimates
and test statistics (including GEL and EL ones) and its use in
deriving the limiting properties of both the estimates and the
test statistics under misspecification and under alternatives
hypotheses. Moreover,
\begin{enumerate}
\item [1)] The approach which we develop is based on
\textit{minimum discrepancy estimates}, which extends the EL
method and has common features with minimum distance and GEL
techniques, using merely divergences. We present a wide class of
estimates, test statistics and confidence regions for the
parameter $\theta_{0}$ as well as various test statistics for
\textit{Problems} 1 and 2,
all depending on the choice of the divergence.\\
\item [2)] The limiting distribution of the EL test statistic
under the alternative and under misspecification remains up to
date an open problem. The present paper fills this gap; indeed, we
give the limiting distributions of the proposed estimates and test
statistics (including the EL ones) both under the null hypotheses,
under alternatives and under
misspecification.\\
\item [3)] The limiting distributions of the test statistics under
the alternatives and misspecification are used to give an
approximation to the power function and the sample size which
ensures a desired power for a given alternative.\\
\item [4)] We extend confidence region (C.R.) estimation
techniques based on EL (see \cite{Owen1990}), providing a wide
range of such C.R.'s, each one depending upon a specific
divergence.
\end{enumerate}
\noindent From the point of view of the statistical criterion
under consideration, the main advantage, of using a divergence
based approach and duality, lays in the fact that it leads to
asymptotic properties of the estimates and test statistics under
the alternative, including misspecification, which cannot be
achieved through the classical EL context. In the case of
parametric models of densities, \cite{White1982} studied the
asymptotic properties of the parametric maximum likelihood
estimate and the parametric likelihood ratio statistic under
misspecification; \cite{Keziou2003} and \cite{Broniatowski_Keziou2009} stated the consistency and obtained the limiting distributions of the minimum
divergence estimates and the corresponding test statistics
(including the parametric likelihood ones) both under the null
hypotheses and the alternatives, from which they deduced an
approximation to the power function. In this paper, we extend the
above results for the proposed class of estimates and test
statistics (including the EL ones) in the context of
semiparametric models
(\ref{modele}). \\
\noindent The rest of the paper is organized as follows. Section 2
describes the statistical divergences used in the sequel. Section
3 is devoted to the description of the proposed estimation and
test procedures. In Section 3, we adapt the Lagrangian duality
formalism to the context of statistical divergence, and we use it
to give practical formulas (for the study and the numerical
computation) of the proposed estimates and test statistics.
Section 5 deals with the asymptotic properties of the estimates
and the test statistics under the model and under
misspecification. Simulations results are given in Section 6. All
proofs are postponed to the Appendix.
\section{Statistical divergences}
\noindent We first set some general definitions and notations. Let
$P$ be some p.m. on the measurable space
$(\mathbb{R}^m,\mathcal{B}(\mathbb{R}^m))$. Denote by $M$ the
space of all signed finite measures (s.f.m.) on
$(\mathbb{R}^m,\mathcal{B}(\mathbb{R}^m))$. Let $\varphi$ be a
convex function from $\mathbb{R}$ onto $[0,+\infty]$ with
$\varphi(1)=0$, and such that its domain, $\text{dom}\varphi
:=\left\{x\in\mathbb{R} \text{ such that }
\varphi(x)<\infty\right\}=:(a,b),$ is an interval, with endpoints
$a<1<b$, which may be bounded or unbounded, open or not. We assume
that $\varphi$ is closed\footnote{The closedness of $\varphi$
means that if $a$ or $b$ are finite then $\varphi(x)\to\varphi(a)$
when $x \downarrow a$, and $\varphi(x)\to\varphi(b)$
when $x \uparrow b$. Note that, this is equivalent to the fact that the level sets
$\{x\in\mathbb{R};~ \varphi(x)\leq \alpha \}$, $\forall \alpha\in\mathbb{R}$, are closed
in $\mathbb{R}$ endowed with the usual topology.}. For any
s.f.m. $Q\in M$, the $\varphi$-divergence between $Q$ and the p.m.
$P$, when $Q$ is absolutely continuous with respect to (a.c.w.r.t)
$P$, is defined through
\begin{equation}
D_{\varphi}(Q,P):=\int_{\mathbb{R}^m}\varphi\left(\frac{dQ}{dP}(x)\right)
~dP(x),\label{divRusch}
\end{equation}
in which $\frac{dQ}{dP}(\cdot)$ denotes the Radon-Nikodym
derivative. When $Q$ is not a.c.w.r.t. $P$, we set
$D_{\varphi}(Q,P):=+\infty$. For any p.m. $P$, the mapping $Q\in
M\mapsto D_{\varphi}(Q,P)$ is convex and takes nonnegative values.
When $Q=P$ then $D_{\varphi}(Q,P)=0$. Furthermore, if the function
$x\mapsto\varphi(x)$ is strictly convex on a neighborhood of
$x=1$, then
\begin{equation}
D_{\varphi}(Q,P)=0~\text{ if and only if }~Q=P.\label{p.f.}
\end{equation}
\noindent All the above properties are presented in
\cite{Csiszar1963}, \cite{Csiszar1967} and in Chapter 1 of
\cite{Liese-Vajda1987}, for $\varphi -$divergences defined on the
set of all p.m.'s $M^{1}$. When the $\varphi $-divergences are
extended to $M$, then the same arguments as developed on $M^{1}$
hold. When defined on $M^{1}$, the Kullback-Leibler $(KL)$,
modified Kullback-Leibler $(KL_{m})$, $\chi^{2}$, modified
$\chi^{2}$ $(\chi_{m}^{2})$, Hellinger $(H)$, and $L^{1}$
divergences are respectively associated to the convex functions
$\varphi(x)=x\log x-x+1$, $\varphi(x)=-\log x+x-1$, $\varphi
(x)=\frac{1}{2}{(x-1)}^{2}$, $\varphi(x)=\frac{1}{2}{(x-1)}^{2}/x
$, $\varphi(x)=2{(\sqrt{x}-1)}^{2}$ and $\varphi(x)=\left\vert
x-1\right\vert $. All these divergences except the $L^{1}$ one,
belong to the class of the so called power divergences introduced
in \cite{Cressie-Read1984} (see also \cite{Liese-Vajda1987} and
\cite{PardoLeandro2006}). They are defined through the class of
convex functions
\begin{equation}
x\in\mathbb{R}_{+}^{\ast}\mapsto\varphi_{\gamma}(x):=\frac{x^{\gamma}-\gamma
x+\gamma-1}{\gamma(\gamma-1)}\label{gamma convex functions}
\end{equation}
if $\gamma\in\mathbb{R}\setminus\left\{ 0,1\right\} $, $\varphi
_{0}(x):=-\log x+x-1$ and $\varphi_{1}(x):=x\log x-x+1$. So, the
$KL-$divergence is associated to $\varphi_{1}$, the $KL_{m}$ to
$\varphi_{0}$, the $\chi^{2}$ to $\varphi_{2}$, the $\chi_{m}^{2}$
to $\varphi_{-1}$ and the Hellinger distance to $\varphi_{1/2}$.
We extend the definition of the power divergences functions $Q\in
M^{1}\mapsto D_{\varphi_{\gamma}}(Q,P)$ onto the whole set of
signed finite measures $M$ as follows. When the function
$x\mapsto\varphi_{\gamma}(x)$ is not defined on $]-\infty,0[$ or
when $\varphi_{\gamma}$ is defined on $\mathbb{R}$ but is not
convex, we extend the definition of $\varphi_{\gamma}$ as follows
\begin{equation}
x\in\mathbb{R}\mapsto\varphi_{\gamma}(x)\mathds{1}_{[0,+\infty[}(x)+(+\infty
)\mathds{1}_{]-\infty,0[}(x).\label{gamma convex functions sur R}
\end{equation}
Note that for $\chi^2$-divergence, the corresponding $\varphi$
function $\varphi(x)=\frac{1}{2}(x-1)^2$ is convex and defined on
whole $\mathbb{R}$. In this paper, for technical considerations,
we assume that the functions $\varphi$ are strictly convex on
their domain $(a,b)$, twice continuously differentiable on
$]a,b[$, the interior of their domain. Hence, $\varphi'(1)=0$, and
for all $x\in ]a,b[$, $\varphi''(x)>0$. Here, $\varphi'$ and
$\varphi''$ are used to denote respectively the first and the
second derivative functions of $\varphi$. Moreover, we assume
that $\varphi$ is ``essentially smooth'' in the sense that
$\lim_{x\downarrow a}\varphi'(x) =-\infty$ if $a$ is finite and
$\lim_{x\uparrow b}\varphi'(x) =+\infty$ if $b$ is finite. Note
that the above assumptions on $\varphi$ are not restrictive, and
that all the power functions $\varphi_{\gamma}$, see (\ref{gamma
convex functions sur R}), satisfy the above conditions, including
all standard divergences.
\begin{definition}
Let $\Omega$ be some subset of $M$. The $\varphi-$divergence
between the set $\Omega$ and a p.m. $P$ is defined by
\[
D_{\varphi}(\Omega,P):=\inf_{Q\in\Omega}D_{\varphi}(Q,P).
\]
A finite measure
$Q^{\ast}\in\Omega$, such that $D_{\varphi}(Q^{\ast},P)<\infty$
and
\[
D_{\varphi}(Q^{\ast},P)\leq D_{\varphi}(Q,P)~\text{ for all
}~Q\in\Omega,
\]
is called a projection of $P$ on $\Omega$. This projection may not
exist, or may be not defined uniquely.
\end{definition}
\section{Minimum divergence estimates}
\noindent Let $X_{1},...,X_{n}$
denote an i.i.d. sample of a random vector $X\in\mathbb{R}^m$ with
distribution $P_{0}$. Let $P_{n}$ be the empirical measure
pertaining to this sample, namely
\[
P_{n}:=\frac{1}{n}\sum_{i=1}^{n}\delta_{X_{i}},
\]
where $\delta_{x}$ denotes the Dirac measure at point $x$, for all
$x$. We will endow our statistical approach in the global context
of s.f.m's with total mass $1$ satisfying $l$ linear constraints:
\begin{equation}
\mathcal{M}_{\theta}:=\left\{ Q\in M~\text{ such that
}~\int_{\mathbb{R}^m} dQ(x)=1\text{ and }\int_{\mathbb{R}^m}
g(x,\theta)~dQ(x)=0\right\} \label{modele signees finies}
\end{equation}
and
\begin{equation}\label{modele msf}
\mathcal{M}:=\bigcup_{\theta\in\Theta}\mathcal{M}_{\theta},
\end{equation}
sets of signed finite measures that replace
$\mathcal{M}_{\theta}^{1}$ and $\mathcal{M}^{1}$. Enhancing the
model (\ref{modele}) to the above one (\ref{modele msf}) bears a
number of improvements upon existing results; this is argued at
the end of the present Section; see also Remark \ref{remark
calcule des estim} below. The ``plug-in'' estimate of
$D_{\varphi}(\mathcal{M}_{\theta},P_{0})$ is
\begin{equation}
\widehat{D}_{\varphi}(\mathcal{M}_{\theta},P_{0}):=\inf_{Q\in\mathcal{M}_{\theta
}}D_{\varphi}(Q,P_{n})=\inf_{Q\in\mathcal{M}_{\theta}}\int_{\mathbb{R}^m}\varphi\left(
\frac {dQ}{dP_{n}}(x)\right)~dP_{n}(x).\label{plug-in est of phi}
\end{equation}
If the projection $Q^{(n)}_\theta$ of $P_n$ on
$\mathcal{M}_\theta$ exists, then it is clear that
$Q^{(n)}_\theta$ is a s.f.m. (or possibly a p.m.) a.c.w.r.t.
$P_n$; this means that the support of $Q^{(n)}_\theta$ must be
included in the set $\left\{X_1,\ldots,X_n\right\}$. So, define
the sets
\begin{equation}\label{m theta n}
\mathcal{M}_{\theta}^{(n)}:=\left\{ Q\in M~|~Q \text{ a.c.w.r.t. }
P_{n},~\sum_{i=1}^{n}Q(X_{i})=1 \text{ and } \sum_{i=1}^{n}Q(X_{i}
)g(X_{i},\theta)=0\right\},
\end{equation}
which may be seen as subsets of $\mathbb{R}^{n}$. Then, the
plug-in estimate (\ref{plug-in est of phi}) can be written as
\begin{equation}
\widehat{D}_{\varphi}(\mathcal{M}_{\theta},P_{0})=\inf_{Q\in\mathcal{M}_{\theta
}^{(n)}}\frac{1}{n}\sum_{i=1}^{n}\varphi\left(
nQ(X_{i})\right).\label{estim de phiMtheta cas cont}
\end{equation}
In the same way,
$D_{\varphi}(\mathcal{M},P_{0}):=\inf_{\theta\in\Theta}
\inf_{Q\in\mathcal{M}_{\theta}}D_{\varphi}(Q,P_{0})$ can be
estimated by
\begin{equation}
\widehat{D}_{\varphi}(\mathcal{M},P_{0}):=\inf_{\theta\in\Theta}\inf
_{Q\in\mathcal{M}_{\theta}^{(n)}}\frac{1}{n}\sum_{i=1}^{n}\varphi\left(
nQ(X_{i})\right).\label{estim de phiM cas cont}
\end{equation}
By uniqueness of
$\arg\inf_{\theta\in\Theta}D_{\varphi}(\mathcal{M}_{\theta},P_{0})$
and since the infimum is reached at $\theta=\theta_{0}$ under the
model, we estimate $\theta_{0}$ through
\begin{equation}\label{estim de theta0 cas cont}
\widehat{\theta}_{\varphi}
:=\arg\inf_{\theta\in\Theta}\inf_{Q\in\mathcal{M}_{\theta}^{(n)}}
\frac{1}{n}\sum_{i=1}^{n}\varphi\left( nQ(X_{i})\right).
\end{equation}
\noindent Enhancing $\mathcal{M}^{1}$ to $\mathcal{M}$ and
accordingly extensions in the definitions of the $\varphi$
functions on $\mathbb{R}$ and the $\varphi $-divergences on the
whole space of s.f.m's $M$, is motivated by the following
arguments:
\begin{enumerate}
\item [-] If the domain $(a,b)$ of the function $\varphi$ is
included in $[0,+\infty[$ then minimizing over $\mathcal{M}^{1}$
or over $\mathcal{M}$ leads to the same estimates and test
statistics. It is the case of the $KL_m$, $KL$, modified $\chi^2$
and Hellinger divergences.
\item[-] Let $\theta$ be a given value in $\Theta$. Denote
$Q^{(1,n)}_{\theta}$ and $Q^{(n)}_\theta$, respectively, the
projection of $P_{n}$ on $\mathcal{M}_{\theta}^{1}$ and on
$\mathcal{M}_{\theta}$. If $Q^{(1,n)}_{\theta}$ satisfies
$0<Q^{(1,n)}_\theta(X_i)<1$, for all $i=1,\ldots,n,$ then
$Q^{(1,n)}_{\theta}=Q^{(n)}_\theta$. Therefore, in this case, both
approaches leads also to the same estimates and test statistics.
\item[-] It may occur that for some $\theta$ in $\Theta$ and some
$i=1,\ldots,n,$ $Q^{(1,n)}_\theta(X_i)$ is a boundary value of
$[0,1]$, hence the first order conditions are not met which makes
a real difficulty for the calculation of the estimates over the
sets of p.m. $\mathcal{M}^1_\theta$ and $\mathcal{M}^1$. However,
when $\mathcal{M}^{1}$ is replaced by $\mathcal{M}$, then this
problem does not hold any longer in particular when
$\text{dom}\varphi=\mathbb{R}$, which is the case for the
$\chi^2$-divergence. Other arguments are given in Remark
\ref{remark calcule des estim} below.
\end{enumerate}
\noindent The empirical likelihood paradigm (see \cite{Owen1988},
\cite{Owen1990}, \cite{Qin-Lawless1994} and \cite{Owen2001}),
enters as a special case of the statistical issues related
to estimation and tests based on
$\varphi-$divergences with $\varphi(x)=\varphi_{0}(x):=-\log
x+x-1$, namely on $KL_{m}-$divergence. Indeed, it is
straightforward to see that the empirical log-likelihood ratio
statistic for testing $\mathcal{H}_0 : P_0\in\mathcal{M}$ against
$\mathcal{H}_1 : P_0\notin\mathcal{M}$, in the context of
$\varphi$-divergences, can be written as
$2n\widehat{D}_{KL_m}(\mathcal{M},P_{0})$; and that the EL
estimate of $\theta_0$ can be written as
$\widehat{\theta}_{KL_m}=\arg\inf_{\theta\in
\Theta}\widehat{D}_{KL_m}(\mathcal{M}_\theta,P_0)$; see Remark
\ref{remarque c_0} below. In the case of the power functions
$\varphi=\varphi_{\gamma}$, the corresponding estimates
(\ref{estim de theta0 cas cont}) belong to the class of GEL
estimates introduced by \cite{Smith1997} and
\cite{NeweySmith2004}, and (\ref{estim de phiMtheta cas cont}) in
this case are the empirical Cressie-Read statistics
introduced by \cite{Baggerly1998}
and \cite{Corcoran1998}; see Remark \ref{remark on GEL estimates} below.\\
\noindent The constrained optimization problems (\ref{estim de
phiMtheta cas cont}), (\ref{estim de phiM cas cont}) and
(\ref{estim de theta0 cas cont}) can be transformed into
unconstrained ones making use of some arguments of ``duality''
which we briefly state below from \cite{Rockafellar1970}. On the
other hand, the obtaining of asymptotic statistical results of the
estimates and the test statistics, under misspecification or under
alternative hypotheses, requires handle existence conditions and
characterization of the projection of $P_{0}$ on the submodel
$\mathcal{M}_{\theta}$ or on the model $\mathcal{M}.$ This also
will be considered through duality, along the following Section.
\section{Dual representation of $\varphi-$divergences under
constraints} \noindent This Section is central for our purposes.
Indeed, it provides the explicit form of the proposed estimates by
transforming the constrained problems (\ref{estim de phiMtheta cas
cont}) to unconstrained ones, using Lagrangian duality which is a
classical tool in optimization theory. This Section adapts this
formalism to the context of divergences and the present
statistical setting. The Lagrangian ``dual'' problem,
corresponding to the ``primal'' one
\begin{equation}
\inf_{Q\in\mathcal{M}_{\theta}}D_\varphi(Q,P_{0})\label{Primal
theo}
\end{equation}
and its empirical counterpart (\ref{estim de phiMtheta cas cont}),
make use of the so-called Fenchel-Legendre transform of
$\varphi$, defined through
\begin{equation}
\psi : t\in \mathbb{R}\mapsto
\psi(t):=\sup_{x\in\mathbb{R}}\left\{tx-\varphi(x)\right\}.
\end{equation}
The ``dual'' problems associated to $(\ref{Primal theo})$ and
(\ref{estim de phiMtheta cas cont}) are respectively
\begin{equation}\label{dual theo}
\sup_{t\in\mathbb{R}^{1+l}}\left\{ t_{0}-\int_{\mathbb{R}^m} \psi
( t_{0}+\sum_{j=1}^{l}t_{j}g_{j}(x,\theta))~ dP_{0}(x)\right\},
\end{equation}
and
\begin{equation}\label{dual empi}
\sup_{t\in\mathbb{R}^{1+l}}\left\{
t_{0}-\frac{1}{n}\sum_{i=1}^{n}\psi(
t_{0}+\sum_{j=1}^{l}t_{j}g_{j}(X_{i},\theta))\right\}.
\end{equation}
In the following Propositions \ref{proposition 1} and
\ref{proposition 2}, we state sufficient conditions under which
the primal problems (\ref{Primal theo}) and (\ref{estim de
phiMtheta cas cont}) coincide respectively with the dual ones
(\ref{dual theo}) and (\ref{dual empi}). First, recall some
properties of the convex conjugate $\psi$ of $\varphi$. For the
proofs, we can refer to Section 26 in \cite{Rockafellar1970}. The
function $\psi$ is convex and closed, its domain is an interval
with endpoints
\begin{equation}\label{domain de varphi*}
a^*:=\lim_{x\to -\infty}\frac{\varphi(x)}{x},\quad b^*:=\lim_{x
\to +\infty}\frac{\varphi(x)}{x}
\end{equation}
satisfying $a^*<0<b^*$ with $\psi(0)=0$. The strict convexity of
$\varphi$ on its domain $(a,b)$ is equivalent to the condition
that its conjugate $\psi$ is essentially smooth, i.e.,
differentiable with
\begin{equation}
\begin{array}
[c]{ccccc} \lim_{t\downarrow a^*}{\psi}^{\prime}(t) & = & -\infty
&
\text{ if } & a^* \text{ is finite},\\
\lim_{t\uparrow b^*}{\psi}^{\prime}(t) & = & +\infty & \text{ if }
& b^* \text{ is finite}.
\end{array}
\end{equation}
Conversely, $\varphi$ is essentially smooth on its domain $(a,b)$
if and only if $\psi$ is strictly convex on its domain
$(a^*,b^*)$. In all the sequel, we assume additionally that
$\varphi$ is essentially smooth. Hence, $\psi$ is strictly convex
on its domain $(a^*,b^*)$, and it holds that
$$a^*=\lim_{x\downarrow a}\varphi'(x),\quad \quad b^*=
\lim_{x\uparrow b}\varphi'(x),$$
and
\begin{equation}\label{forme explicite de psi}
\psi(t)=t{\varphi'}^{-1}(t)-\varphi\left(
{\varphi'}^{-1}(t)\right),\quad \text{for all } t\in ]a^*,b^*[,
\end{equation}
where ${\varphi'}^{-1}$ denotes the inverse function of
$\varphi'$. It holds also that $\psi$ is twice continuously
differentiable on $]a^*,b^*[$ with
\begin{equation}\label{derivee de psi}
\psi'(t)={\varphi'}^{-1}(t) \quad \text{and}\quad
\psi''(t)=\frac{1}{\varphi''\left({\varphi'}^{-1}(t)\right)}.
\end{equation}
In particular, $\psi'(0)=1$ and $\psi''(0)=1$. Obviously, since
$\varphi$ is assumed to be closed, we have
$$\varphi(a)=\lim_{x\downarrow a}\varphi(x)\quad \text{ and }\quad
\varphi(b)=\lim_{x\uparrow b}\varphi(x),$$
which may be finite or infinite. Hence, by closedness of $\psi$,
we have
$$\psi(a^*)=\lim_{t\downarrow a^*}\psi(x)\quad \text{ and }\quad
\psi(b^*)=\lim_{t\uparrow b^*}\psi(t).$$
Finally, the first and second derivatives of $\varphi$ in $a$ and $b$ are
defined to be the limits of $\varphi'(x)$ and $\varphi''(x)$ when $x\downarrow a$ and
when $x\uparrow b$. The first and second derivatives of $\psi$ in $a^*$ and $b^*$
are defined in
a similar way. In Table \ref{table convex conjugates}, we give
the convex conjugates $\psi$ of some standard functions $\varphi$, associated to
some standard divergences. We determine also their domains, $(a,b)$ and
$(a^*,b^*)$.
\begin{table}[h]
\caption{Convex conjugates for some standard divergences.}
\begin{tabular}{|l||l|l||l|l|}
\hline
$D_\varphi$ & $\varphi$ & $\text{dom}\varphi $ & $\text{dom}\psi $ & $\psi$ \\
\hline
\hline
$D_{KL_m}$ & $\varphi(x):=-\log x +x -1$ & $]0,+\infty[$ & $]-\infty,1[$ & $\psi(t)= - \log(1-t)$ \\
\hline
$D_{KL}$ & $\varphi(x):=x\log x -x +1$ & $[0,+\infty[$ & $\mathbb{R}$ & $\psi(t)=e^t-1$ \\
\hline
$D_{\chi^2_m}$ & $\varphi(x):= \frac{1}{2}\frac{\left(x-1\right)^2}{x}$
& $]0,+\infty[$ &
$\left]-\infty,\frac{1}{2}\right]$ & $\psi(t)=1-\sqrt{1-2t}$ \\
\hline
$D_{\chi^2}$ & $\varphi(x):= \frac{1}{2}\left(x-1\right)^2$ & $\mathbb{R}$ & $\mathbb{R}$ & $\psi(t)=\frac{1}{2}t^2+t$ \\
\hline
$D_H$ & $\varphi(x):=2(\sqrt{x}-1)^2$ & $[0,+\infty[$ & $]-\infty, 2[$ & $\psi(t)=\frac{2t}{2-t}$ \\
\hline
$D_{\varphi_\gamma}$ & $\varphi(x):=\frac{x^\gamma -\gamma x +\gamma -1}{\gamma (\gamma -1)}$ & $--$ & $--$ &
$\psi(t)=\frac{1}{\gamma}\left(\gamma t -t+1\right)^{\frac{\gamma}{\gamma -1}}-\frac{1}{\gamma}$
\\
\hline
\end{tabular}
\label{table convex conjugates}
\end{table}
\begin{proposition}\label{proposition 1}
\label{Propo carct empirique} Let $\theta$ be a given value in
$\Theta$. If there exists $Q_{0}$ in $\mathcal{M}_{\theta}^{(n)}$
such that
\begin{equation}\label{cond Owen}
a< Q_{0}(X_{i})<b, \quad \text{for all} \quad i=1,\ldots,n,
\end{equation}
then
\begin{equation}\label{egalite duale 1}
\inf_{Q\in\mathcal{M}_{\theta}^{(n)}}D_{\varphi}(Q,P_{n})=\sup_{t\in
\mathbb{R}^{1+l}}\left\{ t_{0}-\frac{1}{n}\sum_{i=1}^{n}\psi(
t_{0} +\sum_{j=1}^{l}t_{j}g_{j}(X_{i},\theta)) \right\}
\end{equation}
with dual attainment. Conversely, if there exists some dual
optimal solution
$\widehat{t}:=(\widehat{t}_0,\widehat{t}_1,\ldots,\widehat{t}_l)^{\top}\in\mathbb{R}^{1+l}$
such that
\begin{equation}\label{cond Owen 2}
a^*< \widehat{t_0}+\sum_{j=1}^l \widehat{t_j}g_j(X_i,\theta)<b^*, \quad \text{for
all}\quad
i=1,\ldots,n,
\end{equation}
then the equality (\ref{egalite duale 1}) holds, and the unique
optimal solution of the primal problem
$\inf_{Q\in\mathcal{M}_{\theta}^{(n)}}D_{\varphi}(Q,P_{n})$,
namely the projection of $P_n$ on $\mathcal{M}^{(n)}_\theta$, is
given by
\[
Q^{(n)}_\theta(X_i)=\frac{1}{n}
{\varphi'}^{-1}(\widehat{t_0}+\sum_{j=1}^{l}\widehat{t_j} g_j(
X_i,\theta)),\quad i=1,\ldots,n,
\]
where
$\widehat{t}:=(\widehat{t}_0,\widehat{t}_1,\ldots,\widehat{t}_l)^{\top}$
is solution of the system of equations
\begin{equation*}
\left\{\begin{array}{lll}
1-\frac{1}{n}\sum_{i=1}^{n} {\varphi'}^{-1}(\widehat{t_0}
+\sum_{j=1}^l\widehat{t_j} g_j (X_i,\theta)) & = & 0, \\
-\frac{1}{n}\sum_{i=1}^n g_{j}(X_i,\theta)
{\varphi'}^{-1}( \widehat{t_0}+\sum_{j=1}^{l}\widehat{t_j}g_{j}(
X_{i},\theta)) & = & 0, \quad j=1,\ldots,l.\\
\end{array} \right.
\end{equation*}\\
\end{proposition}
\begin{remark}{\rm
For the $\chi^{2}-$divergence, we have $a=-\infty$ and
$b=+\infty$. Hence, condition (\ref{cond Owen}) holds whenever
$\mathcal{M}_{\theta}^{(n)}$ is not void. More generally, the
above Proposition holds for any $\varphi$-divergence with
$\text{dom}\varphi=\mathbb{R}$.}
\end{remark}
\begin{remark}{\rm
Assume that $g(x,\theta):=(x-\theta)^\top$. So, for any divergence
$D_\varphi$ with $\text{dom}\varphi = ]0,+\infty[$, which is the
case of the modified $\chi^2$ divergence and the modified
Kullback-Leibler divergence (or equivalently EL method), condition
$(\ref{cond Owen})$ means that $\theta$ is an interior point of
the convex hull of the data $(X_{1},...,X_{n})$. This is precisely
what is checked in \cite{Owen1990}, p. 100, for the EL method; see
also \cite{Owen2001}.}
\end{remark}
\noindent For the asymptotic counterpart of the above results we
have; see Theorem 1 in \cite{Bronia_Kez2006_STUDIA}:
\begin{proposition}\label{proposition 2}
\label{Prop caract theo} Let $\theta$ be a given value in
$\Theta$. Assume that $\int |g_j(x,\theta)|~dP_0(x)<\infty$, for
all $j=1,\ldots,l$. If there exists $Q_{0}$ in
$\mathcal{M}_{\theta}$ with $D_\varphi(Q_0,P_0)<\infty$
and\footnote{The strict inequalities (\ref{condition de
qualification}) mean that
$P_0\left\{x\in\mathbb{R}^m~|~\frac{dQ_0}{dP_0}(x)\leq
a\right\}=P_0\left\{x\in\mathbb{R}^m~|~\frac{dQ_0}{dP_0}(x)\geq
b\right\}=0.$}
\begin{equation}\label{condition de qualification}
a<\inf_{x\in\mathbb{R}^m}\frac{dQ_{0}}{dP_{0}}(x)\leq\sup_{x\in\mathbb{R}^m}\frac{dQ_{0}}{dP_{0}
}(x)<b, \quad P_{0}-a.s.,
\end{equation}
then
\begin{equation}\label{egalite duale 2}
\inf_{Q\in\mathcal{M}_{\theta}}D_{\varphi}(Q,P_{0})=\sup_{t\in\mathbb{R}^{1+l}
}\left\{ t_{0}-\int_{\mathbb{R}^m} \psi(
t_{0}+\sum_{j=1}^{l}t_{j}g_{j}( x,\theta))~dP_{0}(x)\right\}
\end{equation}
with dual attainment. Conversely, if there exists some dual optimal solution
$t^*$ which is an interior point of the set
\begin{equation}\label{condition de qualification 2}
\left\{t\in\mathbb{R}^{1+l}~\text{such that}~ \int_{\mathbb{R}^m}
|\psi(t_{0}+\sum_{j=1}^{l}t_jg_{j}(x,\theta))| ~dP_0(x) <\infty
\right\},
\end{equation}
then the dual equality (\ref{egalite duale 2}) holds, and
the unique optimal solution $Q^*_\theta$ of the
primal problem
$\inf_{Q\in\mathcal{M}_{\theta}}D_{\varphi}(Q,P_{0})$, namely the
projection of $P_0$ on $\mathcal{M}_\theta$, is given by
\begin{equation*}
\frac{dQ^*_\theta}{dP_{0}}(x)={\varphi'}^{-1}(t_{0}^*+\sum_{j=1}^{l}t_j^*g_{j}(x,\theta)),
\end{equation*}
where $t^*:=(t^*_0,t^*_1,\ldots,t^*_l)^{\top}$ is solution of the
system of equations
\begin{equation}\label{systeme}
\left\{\begin{array}{lll}
1-\int {\varphi'}^{-1}( t_{0}^*+\sum_{j=1}^{l}t_j^*g_{j}( x,\theta)
)~ dP_{0}(x) & = & 0, \\
-\int g_{j}(x,\theta)
{\varphi'}^{-1}(t_{0}^*+\sum_{j=1}^{l}t_j^*g_{j}(x,\theta))~
dP_{0}(x) & = & 0,\quad j=1,\ldots,l. \\
\end{array} \right.
\end{equation}
Furthermore, $t^*$ is unique if the functions
$\mathds{1}_{\mathbb{R}^m}, g_1(.,\theta),\ldots,g_l(.,\theta)$
are linearly independent in the sense that
$P_0\left\{x\in\mathbb{R}^m~|~t_0+\sum_{j=1}^lt_jg_j(x,\theta)\neq
0\right\}>0$ for all
$t\in\mathbb{R}^m$ with $t\neq 0.$\\
\end{proposition}
\noindent For sake of brevity and clearness, we must introduce
some additional notations. In all the sequel, $\|x\|$ denotes the
norm of $x$ defined by $\|x\|:=\sup_{i}|x_i|$ for any vector
$x:=(x_1,\ldots,x_k)^\top\in\mathbb{R}^k$, and for any matrix $A$,
the norm of $A$ is defined by $\|A\|:=\sup_{i,j}|a_{i,j}|$. Denote
by $\overline{g}$ the vector valued function
$\overline{g}:=(\mathds{1}_{\mathbb{R}^m},g_1,\ldots,g_l)^{\top}\in\mathbb{R}^{1+l}$.
For any p.m. $P$ on $(\mathbb{R}^m,\mathcal{B}(\mathbb{R}^m))$ and
any real measurable function $f$ from
$(\mathbb{R}^m,\mathcal{B}(\mathbb{R}^m))$ to
$(\mathbb{R},\mathcal{B}(\mathbb{R}))$, denote
$$Pf:=\int_{\mathbb{R}^m}f(x)~dP(x).$$
Let
$$t^{\top} \overline{g}(x,\theta):=t_0+\sum_{j=1}^lt_jg_j(x,\theta)$$
and
\begin{equation}\label{m x theta t} m(x,\theta,t):= t_0-
\psi(t^{\top} \overline{g}(x,\theta)), \text{
for all } x\in\mathbb{R}^m,
\theta\in\Theta\subset\mathbb{R}^d,t\in \mathbb{R}^{1+l}.
\end{equation}
\noindent Note that the $\sup$ in (\ref{egalite duale 1}) and
(\ref{egalite duale 2}) can be restricted, respectively, to the
sets
\begin{equation}\label{lambda n theta} \Lambda_\theta^{(n)}:=\left\{t\in\mathbb{R}^{1+l}~|~
a^*<t^{\top}\overline{g}(X_i,\theta)<b^*, ~\text{ for all }
i=1,\ldots,n\right\}\end{equation} and
\begin{equation}\label{lambda theta}\Lambda_\theta:=\left\{t\in\mathbb{R}^{1+l}~|~
\int_{\mathbb{R}^m} |\psi(t_{0}+\sum_{j=1}^{l}t_jg_{j}(x,\theta))|
~dP_0(x) <\infty \right\}.\end{equation}
In view of
the above two Propositions \ref{proposition 1} and
\ref{proposition 2}, we redefine the estimates (\ref{estim de
phiMtheta cas cont}), (\ref{estim de phiM cas cont}) and
(\ref{estim de theta0 cas cont}) as follows
\begin{equation}\label{estimateur1}
\widehat{D}_\varphi\left(\mathcal{M}_\theta,P_0\right):=
\sup_{t\in\Lambda_\theta^{(n)}}\frac{1}{n}\sum_{i=1}^n
m(X_i,\theta,t):=\sup_{t\in\Lambda_\theta^{(n)}} P_nm(\theta,t),
\end{equation}
\begin{equation}\label{estimateur2}
\widehat{D}_\varphi\left(\mathcal{M},P_0\right):=
\inf_{\theta\in\Theta}\sup_{t\in\Lambda_\theta^{(n)}}\frac{1}{n}\sum_{i=1}^n
m(X_i,\theta,t):=\inf_{\theta\in\Theta}\sup_{t\in\Lambda_\theta^{(n)}}
P_nm(\theta,t)
\end{equation}
and
\begin{equation}\label{estimateur3}
\widehat{\theta}_\varphi:=
\arg\inf_{\theta\in\Theta}\sup_{t\in\Lambda_\theta^{(n)}}\frac{1}{n}\sum_{i=1}^n
m(X_i,\theta,t):=\arg\inf_{\theta\in\Theta}\sup_{t\in\Lambda_\theta^{(n)}}P_nm(\theta,t).
\end{equation}
\begin{remark}\label{remarque c_0}{\rm
When $\varphi(x)=-\log x+x-1$, then the estimate (\ref{estim de theta0 cas cont})
clearly coincides with the EL one, so
it can be seen as the value of the parameter which
minimizes the $KL_m$-divergence between the model $\mathcal{M}$ and the empirical measure
$P_n$ of the data $X_1,\ldots, X_n$. The statistic $2n\widehat{D}_{KL_m}(\mathcal{M},P_0)$, see (\ref{estim de phiM cas
cont}), coincides with the empirical likelihood ratio statistic associated
to the null hypothesis $\mathcal{H}_0:P_0\in\mathcal{M}$ against the alternative $\mathcal{H}_1:P_0\not\in\mathcal{M}$.
The dual representation of $\widehat{D}_{KL_m}(\mathcal{M},P_0)$, see (\ref{estimateur2}) and (\ref{egalite duale 1}), is
$$\widehat{D}_{KL_m}(\mathcal{M},P_0)=\inf_{\theta\in\Theta}\sup_{t\in\Lambda_\theta^{(n)}}\left\{t_0+\frac{1}{n}\sum_{i=1}^n
\log(1-t_0-
\sum_{j=1}^l t_jg_j(X_i,\theta))\right\}.$$
For $a$ given $\theta\in\Theta$, the $KL_m$-projection $Q^{(n)}_\theta$, of $P_n$
on $\mathcal{M}_{\theta}$, is given by (see Proposition \ref{proposition 1})
$$\frac{1}{Q^{(n)}_\theta(X_i)}=n\left(1-\widehat{t}_0-\sum_{j=1}^l\widehat{t}_jg(X_i,\theta)\right),\quad i=1,\dots,n,$$
which, multiplying by $Q^{(n)}_\theta(X_i)$ and summing upon $i=1,\ldots,n,$ yields
$\widehat{t}_0=0$. Therefore, $t_0$ can be omitted, and the above representation can be rewritten
as follows
$$\widehat{D}_{KL_m}(\mathcal{M},P_0)=\inf_{\theta\in\Theta}\sup_{t_1,\ldots,t_l}\left\{\frac{1}{n}\sum_{i=1}^n
\log(1+
\sum_{j=1}^l t_jg_j(X_i,\theta))\right\}$$
and then
\begin{equation}\label{representation duale de EL}
\widehat{\theta}_{KL_m}=\widehat{\theta}_{EL}=\arg\inf_{\theta\in\Theta}\sup_{t_1,\ldots,t_l}\left\{\frac{1}{n}\sum_{i=1}^n
\log(1+
\sum_{j=1}^l t_jg_j(X_i,\theta))\right\}
\end{equation}
in which the $\sup$ is taken over the set $$\left\{(t_1,\ldots,t_l)^\top\in\mathbb{R}^m ~ |~ -1<
\sum_{j=1}^lt_jg_j(X_i,\theta)<+\infty, \text{ for all }
i=1,\ldots,n
\right\}.$$
The formula (\ref{representation duale de EL}) is the ordinary dual representation of the EL estimate;
see \cite{Qin-Lawless1994} and \cite{Owen2001}.}\\
\end{remark}
\begin{remark}\label{remark on GEL estimates}{\rm Consider the power divergences, associated to the power
functions $\varphi_\gamma$; see (\ref{gamma convex functions}) and
(\ref{gamma convex functions sur R}). We will show that the
estimates $\widehat{\theta}_{\varphi_{\gamma}}$ belong to the
class of GEL estimators introduced by \cite{Smith1997} and
\cite{NeweySmith2004}. The projection $Q^{(n)}_\theta$ of $P_n$ on
$\mathcal{M}_\theta$ is given by
$$Q^{(n)}_\theta(X_i)=\left((\gamma-1) (\widehat{t}_0+\sum_{j=1}^l \widehat{t}_jg(X_i,\theta))
+1\right)^{1/(\gamma-1)},\quad i=1,\ldots,n.$$ Using the
constraint $\sum_{i=1}^n Q^{(n)}_\theta(X_i)=1$, we can explicit
$\widehat{t}_0$ in terms of $\widehat{t}_1,\ldots,\widehat{t}_l$,
and hence the $\sup$ in the dual representation
(\ref{estimateur3}) can be reduced to a subset of $\mathbb{R}^l,$
as in \cite{NeweySmith2004}. When $\varphi(x)=\frac{1}{2}(x-1)^2$,
it is straightforward to see that the corresponding estimate
$\widehat{\theta}_\varphi$ coincides with the continuous updating
estimator of
\cite{Hansen_Healton_Yaron1996}.}
\end{remark}
\begin{remark}{\rm (\textbf{Numerical calculation of the estimates and the specific role of the
$\chi^2$-divergence}). The computation of $\widehat{t}(\theta)$
for fixed $\theta\in\Theta$ as defined in (\ref{systeme}) is
difficult when handling a generic divergence. In the particular
case of $\chi^2$-divergence, i.e., when
$\varphi(x)=\frac{1}{2}(x-1)^2$, optimizing on all s.f.m's, the
system (\ref{systeme}) is linear; we thus easily obtain an
explicit form for $\widehat{t}(\theta)$, which in turn allows for
a single gradient descent when optimizing upon $\Theta$. This
procedure is useful in order to compute the estimates for all
other divergences (for which the corresponding system is non
linear) including EL, since it provides an easy starting point for
the resulting double gradient descent. Moreover,
\cite{Hjort_McKeague_VanKeilegom2009} extend the EL approach, to
more general moment condition models, allowing the number of
constraints to increase with growing sample size. In this case,
the computation of EL estimate is more complex, and the same idea
as above can help to solve the problem. \label{remark calcule des
estim}}
\end{remark}
\section{Asymptotic properties of the estimates of the parameter and
the divergences}
\subsection{Asymptotic properties under the model}
This Section addresses Problems 1 and 2, aiming at testing the
null hypothesis $\mathcal{H}_0 : P_{0} \in \mathcal{M}$ against
the alternative $\mathcal{H}_1: P_{0}\not\in \mathcal{M}$. We
derive the limiting distributions of the proposed test statistics
which are the estimated divergences between the model
$\mathcal{M}$ and $P_0$. We also derive the limiting distributions
of the estimates of $\theta_0$. The following two Theorems
\ref{theoreme 1} and \ref{theoreme 2} extend Theorems 3.1 and 3.2
in \cite{NeweySmith2004} to the context of divergence based
approach. The Assumptions which we will consider match those of
Theorems 3.1 and 3.2 in \cite{NeweySmith2004}.
\begin{assumption}
\label{assumption 1} a) $P_{0}\in \mathcal{M}$ and $\theta _{0}\in
\Theta $ is the unique solution of $\mathbb{E}\left[ g(X,\theta
)\right] =0$; b) $\Theta \subset \mathbb{R}^{d}$ is compact; c)
$g(X,\theta )$ is continuous at each $\theta \in \Theta $ with
probability one; d) $\mathbb{E}\left[ \sup_{\theta \in \Theta
}\Vert g(X,\theta )\Vert ^{\alpha }\right] <\infty $ for some
$\alpha >2$; e) the matrix $\Omega :=\mathbb{E}\left[ g(X,\theta
_{0})g(X,\theta _{0})^{\top}\right] $ is nonsingular.
\end{assumption}
\begin{theorem}\label{theoreme 1}
Under Assumption \ref{assumption 1}, with probability approaching one as $n\to\infty$,
the estimate $\widehat{\theta}_\varphi$
exists, and converges to
$\theta_0$ in probability. $\frac{1}{n}\sum_{i=1}^n
g(X_i,\widehat{\theta}_\varphi)=O_P(1/\sqrt{n})$,
$\widehat{t}(\widehat{\theta}_\varphi):=\arg\sup_{t\in\Lambda^{(n)}_{\widehat{\theta}_\varphi}}
P_nm(\widehat{\theta}_\varphi,t)$
exists and belongs to $\text{int}(\Lambda^{(n)}_{\widehat{\theta}_\varphi})$
with probability approaching one as $n\to\infty$, and
$\widehat{t}(\widehat{\theta}_\varphi)=O_P(1/\sqrt{n})$.
\end{theorem}
\noindent In order to obtain asymptotic normality, we need some
additional Assumptions. Denote by $G$ the matrix
$G:=\mathbb{E}\left[\partial g(X,\theta_0)/\partial
\theta\right]$.
\begin{assumption}
\label{assumption 2} a) $\theta _{0}\in \text{int}(\Theta )$; b)
with probability one, $g(X,\theta )$ is continuously
differentiable in a neighborhood $N_{\theta_0}$ of $\theta _{0}$,
and $\mathbb{E}\left[ \sup_{\theta \in N_{\theta_0}}\Vert
\partial g(X,\theta )/\partial \theta \Vert \right] <\infty $; c)
$\text{rank}(G)=d.$
\end{assumption}
\begin{theorem}\label{theoreme 2}
Assume that Assumptions \ref{assumption 1} and \ref{assumption 2} hold.
Then,
\begin{enumerate}
\item [1)]
$\sqrt{n}\left(\widehat{\theta}_\varphi-\theta_0\right)$
converges in distribution
to a centered normal random vector with covariance matrix $$V:=\left[G \Omega^{-1}G^{\top}\right]^{-1}.$$
\item [2)] If $l>d$, then the statistic $2n\widehat{D}_\varphi(\mathcal{M},P_0)$
converges in distribution to a $\chi^2$ random variable with
$(l-d)$ degrees of freedom.
\end{enumerate}
\end{theorem}
\begin{remark} \label{remarque test du model}{\rm The above Theorem allows to perform statistical tests (of the model)
with asymptotic level $\alpha\in ]0,1[$. Consider the null
hypothesis
\begin{equation}\label{test de model}\mathcal{H}_0 : P_0\in\mathcal{M} \quad
\text{against the alternative}\quad \mathcal{H}_1:P_0\not\in\mathcal{M}.
\end{equation}
The critical region is then
$$C_\varphi:=\left\{2n\widehat{D}_\varphi(\mathcal{M},P_0)>q_{(1-\alpha)}\right\}$$
where $q_{(1-\alpha)}$ is the $(1-\alpha)$-quantile of the
$\chi^2(l-d)$ distribution. When $\varphi(x)=-\log x +x -1$, it is straightforward to see that the
corresponding test is the empirical likelihood ratio one;
see \cite{Qin-Lawless1994}.}
\end{remark}
\subsection{Asymptotic properties of the estimates of the divergences for a
given value of the parameter}
For a given $\theta \in\Theta$, consider the test problem of the
null hypothesis $\mathcal{H}_0: P_{0}\in \mathcal{M}_{\theta}$
against two different families of alternative hypotheses:
$\mathcal{H}_1: P_{0}\notin \mathcal{M}_{\theta}$ and
$\mathcal{H}_1':P_{0}\in \mathcal{M}\setminus\mathcal{M}_\theta.$
Those two tests address different situations since $\mathcal{H}_1$
may include misspecification of the model. We give two different
test statistics each pertaining to one of the situations and
derive their limiting distributions both under $\mathcal{H}_0$ and
under the alternatives. As a by product, we also derive confidence
areas for the true value $\theta_0$ of the parameter. We will
first state the convergence in probability of
$\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)$ to
$D_\varphi(\mathcal{M}_{\theta},P_0)$, and then we obtain the
limiting distribution of
$\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)$ both when
$P_0\in\mathcal{M}_{\theta}$ and when $P_0\not\in
\mathcal{M}_{\theta}$. Obviously, when $P_0\in
\mathcal{M}_{\theta}$, this means that $\theta=\theta_0$ since the
true value $\theta_0$ of the parameter is assumed to be unique.
\begin{assumption}
\label{assumption 3} a) $P_0\in\mathcal{M}_{\theta}$ and $\theta$
is the unique solution of $\mathbb{E}\left[g(X,\theta)\right]=0$;
b) $\mathbb{E} \left[\|g(X,\theta)\|^\alpha\right]<\infty$ for
some $\alpha>2$; c) the matrix $\Omega:=\mathbb{E}\left[
g(X,\theta)g(X,\theta)^{\top}\right]$ is nonsingular.
\end{assumption}
\begin{theorem}\label{theoreme 3}
Under Assumption \ref{assumption 3}, we have
\begin{enumerate}
\item [1)]
$\widehat{t}(\theta):=\arg\sup_{t\in\Lambda_\theta^{(n)}}P_nm(\theta,t)$
exists and belongs to $\text{int}(\Lambda^{(n)}_\theta)$ with
probability approaching one as $n\to\infty$, and
$\widehat{t}(\theta)=O_P(1/\sqrt{n})$.
\item [2)] The statistic
$2n\widehat{D}_\varphi(\mathcal{M}_\theta,P_0)$ converges in
distribution to a $\chi^2(l)$ random variable.
\end{enumerate}
\end{theorem}
\noindent In order to obtain the limiting distribution of the test
statistic $2n\widehat{D}_{\varphi }\left( \mathcal{M}_{\theta
},P_{0}\right) $ under the alternative $\mathcal{H}_1:P_{0}\notin
\mathcal{M}_{\theta }$, including misspecification, the following
Assumption is needed.
\begin{assumption} \label{assumption 4}
a) $P_{0}\not\in \mathcal{M}_{\theta}$, and
$t^{\ast}(\theta ):=\arg \sup_{t\in
\Lambda_\theta}\mathbb{E}\left[ m(X,\theta,t)\right]$ exists and
is an interior point of $\Lambda_\theta$; b) $\mathbb{E}\left[
\sup_{t\in N_{t^*(\theta)}}|m(X,\theta ,t)|\right] <\infty $ for
some compact set $N_{t^*(\theta)}\subset \Lambda_\theta$ such that
$t^{\ast}(\theta)\in \text{int}(N_{t^*(\theta)})$; c) the
functions $\mathds{1}_{\mathbb{R}^{m}},g_{1},\ldots ,g_{l}$ are
linearly independent in the sense that : $P_{0}\left\{
x\in\mathbb{R}^m~|~t_{0}+\sum_{j=1}^{l}t_{j}g_{j}(x,\theta )\neq
0\right\}
>0$ ~ for all $t\in \mathbb{R}^{1+l}$ with $t\neq 0$.
\begin{remark}\label{remark 1 sur les conditions}{\rm
Assumption \ref{assumption 4}.c above ensures the strict concavity of the function
$t\in \Lambda_\theta\mapsto \mathbb{E}\left[m\left( X,\theta
,t\right)\right]$ on the convex set $\Lambda_\theta$, which
implies that $t^*(\theta)$ is unique. It can be replaced by the
following Assumption : there exists a neighborhood,
$N_{t^*(\theta)}\subset\Lambda_\theta$, of $t^*(\theta)$, such
that $\mathbb{E}\left[\sup_{t\in N_{t^*(\theta)}}\left\|\partial
m(X,{\theta},t)/\partial t\right\|\right]<\infty$,
$\mathbb{E}\left[\sup_{t\in
N_{t^*(\theta)}}\left\|\partial^2m(X,{\theta},t)/\partial t
^2\right\|\right]<\infty$ and the matrix
$\mathbb{E}\left[\partial^2m(X,{\theta},t^*(\theta))/\partial t
^2\right]$ is nonsingular; which implies also that $t^*(\theta)$
is unique.}
\end{remark}
\end{assumption}
\begin{theorem}\label{theoreme 4}
Under Assumption \ref{assumption 4}, when $P_0\not\in
\mathcal{M}_{\theta}$, we have
\begin{enumerate}
\item [1)] $\widehat{t}(\theta)$ converges in probability to $t^*(\theta)$.
\item [2)] $\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)$ converges
in probability to $D_\varphi(\mathcal{M}_{\theta},P_0).$
\end{enumerate}
\end{theorem}
\noindent We now give the limiting distribution of the test
statistic under $\mathcal{H}_1.$ We need the following additional
condition.
\begin{assumption}
\label{assumption 5} a) There exists
$N_{t^*(\theta)}\subset\Lambda_\theta$, some compact neighborhood
of $t^*(\theta)$, such that
$$\mathbb{E}[\sup_{t\in N_{t^*(\theta)}} \|\partial m(X,\theta,t^*(\theta))/\partial t\|]<\infty,\quad
\mathbb{E}[\sup_{t\in N_{t^*(\theta)}} \|\partial^2
m(X,\theta,t^*(\theta))/\partial t^2\|]<\infty;$$
b) as $\delta\to 0$,
$$\mathbb{E}\left\{\sup_{\{t; \|t-t^*(\theta)\|\leq \delta\}}\left\|\partial^2m(X,\theta,t)/\partial t^2-
\partial^2m(X,\theta,t^*(\theta))/\partial t^2\right\|\right\}\to 0;$$
c) $\mathbb{E}\left[m(X,{\theta},t^*(\theta))^2\right]<\infty$, $\mathbb{E}\left[\|\partial m(X,{\theta},
t^*(\theta))/\partial t\|^2\right]<\infty$\\
and the matrix
$\mathbb{E}\left[\partial^2m(X,{\theta},t^*(\theta))/\partial t
^2\right]$ is nonsingular.
\end{assumption}
\begin{remark}\label{remark 1 sur les conditions de 3 ordre}
Assumption \ref{assumption 5}.b is used here to relax the
condition on the third derivatives (in $t$) of the function $t\mapsto
m(X,\theta,t)$.
\end{remark}
\begin{theorem}\label{theoreme 5}
Under Assumptions \ref{assumption 4} and \ref{assumption 5}, we
have
\begin{enumerate}
\item [1)] $\sqrt{n}(\widehat{t}(\theta)-t^*(\theta))$ converges in
distribution to a centered normal random vector with covariance matrix
$$\left[\mathbb{E}\left[m''(X,{\theta},t^*)\right] \right]^{-1}
\mathbb{E}\left[m'(X,{\theta},t^*)m'(X,{\theta},t^*)^{\top}\right]
\left[\mathbb{E}\left[m''(X,{\theta},t^*)\right] \right]^{-1}.$$
\item [2)] $\sqrt{n}\left(\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)-D_\varphi(\mathcal{M}_{\theta},P_0)\right)$
converges in distribution to a centered normal random variable
with variance
$$\sigma^2(\theta):=\mathbb{E}\left[m(X,{\theta},t^*(\theta))^2\right]-\left[\mathbb{E}\left[m(X,{\theta},
t^*(\theta))\right]\right]^2.$$
\end{enumerate}
\end{theorem}
\begin{remark}{\rm \label{remarque 1}
Let $\theta$ be a given value in $\Theta$. Consider the test
of the null hypothesis
\begin{equation}\label{problem de test 1}
\mathcal{H}_0 : P_0\in\mathcal{M}_\theta \quad
\text{against}\quad \mathcal{H}_1 : P_0\notin \mathcal{M}_\theta.
\end{equation}
In view of Theorem \ref{theoreme 3} part 2, we reject
$\mathcal{H}_0$ against $\mathcal{H}_1$, at asymptotic level
$\alpha\in ]0,1[$, when $2n\widehat{D}_{\varphi
}\left(\mathcal{M}_{\theta },P_{0}\right)$ exceeds the
$(1-\alpha)$- quantile of the $\chi^{2}(l)$ distribution. Theorem
\ref{theoreme 5} part 2 is useful to give an approximation to the
power function
$$P_0\notin \mathcal{M}_\theta \mapsto \beta(P_0):=P_0\left[2n\widehat{D}_{\varphi }\left(\mathcal{M}_{\theta
},P_{0}\right)>q_{(1-\alpha)}\right].$$ We obtain then the
following approximation
\begin{equation}\label{power approxi 1}
\beta(P_0)\approx
1-F_\mathcal{N}\left(\frac{\sqrt{n}}{\sigma(\theta)}\left[\frac{q_{1-\alpha}}{2n}-
D_\varphi(\mathcal{M}_\theta,P_0)\right]\right),
\end{equation}
where $F_\mathcal{N}$ is the cumulative distribution function of
the standard normal distribution. From this approximation, we can
give the approximate sample size that ensures a desired power
$\beta$ for a given alternative $P_0\notin \mathcal{M}_\theta$.
Let $n_{0}$ be the positive root of the equation
\[
\beta =1-F_{\mathcal{N}}\left[ \frac{\sqrt{n}}{\sigma
\left(\theta\right) }\left( \frac{q_{(1-\alpha) }}{2n}-D_{\varphi
}\left( \mathcal{M}_\theta,P_0\right) \right) \right]
\]
i.e.,
\[
n_{0}=\frac{\left( a+b\right) -\sqrt{a\left( a+2b\right)
}}{2D_{\varphi }\left(\mathcal{M}_\theta,P_0\right)^{2}}
\]
with $a:=\sigma(\theta^*)^{2}\left[
F_{\mathcal{N}}^{-1}\left(1-\beta\right)\right]^2$ and
$b:=q_{(1-\alpha) } D_{\varphi }\left(
\mathcal{M}_\theta,P_0\right).$ The required sample size is then
$\left\lfloor n_{0}\right\rfloor +1$, where $\left\lfloor
n_{0}\right\rfloor$ denotes the integer part of $n_{0}.$}\\
\end{remark}
\begin{remark}{\rm (\textbf{Generalized empirical likelihood ratio
test}). For testing $\mathcal{H}_0:P_0\in\mathcal{M}_\theta$
against the alternative
$\mathcal{H}_1':\mathcal{M}\setminus\mathcal{M}_\theta$, we
propose to use the statistics
\begin{equation}\label{S n phi}
2nS_{n}^{\varphi}:=2n\left[\widehat{D}_{\varphi }\left(
\mathcal{M}_{\theta },P_{0}\right)
-\inf_{\theta\in\Theta}\widehat{D}_{\varphi }\left(
\mathcal{M}_{\theta },P_{0}\right)\right],
\end{equation}
which converge in distribution to a $\chi^{2}(d)$ random variable
under $\mathcal{H}_0$ when Assumptions \ref{assumption 1} and
\ref{assumption 2} hold. This can be proved using similar
arguments as in Theorems \ref{theoreme 2} and \ref{theoreme 3}. We
then reject $\mathcal{H}_0$ at asymptotic level $\alpha$ when
$2nS_{n}^{\varphi}>q_{(1-\alpha)}$, the $(1-\alpha)$-quantile of
the $\chi^2(d)$-distribution. Under $\mathcal{H}_1'$ and when
Assumptions \ref{assumption 1},\ref{assumption 2},\ref{assumption
4} and \ref{assumption 5} hold, as in Theorem \ref{theoreme 5}, it
can be proved that
\begin{equation}\label{S n phi alternative}
\sqrt{n}\left( S_{n}^{\varphi }-D_{\varphi }\left(
\mathcal{M}_{\theta },P_{0}\right) \right)
\end{equation}
converges to a centered normal random variable with variance
$$\sigma^{2}(\theta):=\mathbb{E}\left(m(X,\theta ,t^*(\theta))^2\right)-
\left(\mathbb{E}m(X,\theta,t^*(\theta))\right)^2.$$ So, as in the
above Remark, we obtain the following approximation
\begin{equation}\label{power approxi 2}
\beta(P_0)\approx
1-F_\mathcal{N}\left(\frac{\sqrt{n}}{\sigma(\theta)}\left[\frac{q_{1-\alpha}}{2n}-
D_\varphi(\mathcal{M}_\theta,P_0)\right]\right)
\end{equation}
to the power function $P_0\in\mathcal{M}/\mathcal{M}_\theta\mapsto
\beta(P_0):=P_0\left[2nS_n^\varphi>q_{(1-\alpha)}\right].$ The
approximated sample size required to achieve a desired power for a
given alternative can be obtained in a similar way.}
\end{remark}
\begin{remark}{\rm (\textbf{Confidence region for the parameter}).
For a fixed level $\alpha\in]0,1[$,
using convergence (\ref{S n phi}), the set
$$\left\{\theta\in\Theta\text{ such that } 2nS_n^\varphi\leq q_{(1-\alpha)}\right\}$$
is an asymptotic confidence region for $\theta_0$ where
$q_{(1-\alpha)}$ is the $(1-\alpha)$-quantile of the
$\chi^2(d)$-distribution. It is straightforward to see that the
confidence region obtained for the $KL_m$-divergence coincides
with that of \cite{Owen1991} and \cite{Qin-Lawless1994}.}
\end{remark}
\subsection{Asymptotic properties under misspecification}
We address Problem 1 stating the limiting distribution of the
proposed test statistics under the alternative $\mathcal{H}_1:
P_{0}\notin \mathcal{M}.$ This needs the introduction of
$Q_{\theta^*}^*$, the projection of $P_{0}$ on $\mathcal{M}$.
Assumption 6 below ensures the existence of the ``pseudo-true''
value $\theta^*$ as well as the existence of the projection
$Q_{\theta^*}^*$ of $P_0$ on $\mathcal{M}$, and states some
necessary other regularity conditions. Proposition
\ref{proposition 2} above states the existence and
characterization of the projection $Q_\theta^*$ of $P_0$ on
$\mathcal{M}_\theta$, for a given $\theta\in\Theta$.
\begin{assumption}
\label{assumption 6} a) $\Theta$ is compact,
$\theta^*:=\arg\inf_{\theta\in\Theta}\sup_{t\in\Lambda_\theta}
\mathbb{E}\left[m(X,\theta,t)\right]$ exists and is unique; b)
$g(X,\theta)$ is continuous at each $\theta\in\Theta$ with
probability one;\\ c) $\mathbb{E}\left[\sup_{\{\theta\in\Theta,
t\in N_{t^*(\theta)}\}}|m(X,\theta,t)|\right]<\infty,$ where
$N_{t^*(\theta)}\subset \Lambda_\theta$ is a compact set such
that $t^*(\theta)\in\text{int}\left(N_{t^*(\theta)}\right)$; d)
for all $\theta\in\Theta$, the functions
$\mathds{1}_{\mathbb{R}^m}, g_1,\ldots,g_l$ are linearly
independent in the sense that
$P_0\left\{x\in\mathbb{R}^m~|~t_0+\sum_{j=1}^lt_jg_j(x,\theta)
\neq 0\right\}>0$, for all $t\in\mathbb{R}^{1+l}$ with $t\neq 0.$
\end{assumption}
\begin{remark}\label{remark 2 sur les conditions}
Assumption \ref{assumption 6}.d ensures the strict concavity of the function
$t\in\Lambda_\theta \mapsto \mathbb{E}\left[m(X,\theta,t)\right]$
on the convex set $\Lambda_\theta$, which implies the uniqueness
of $t^*(\theta)$, for all $\theta\in\Theta$. This Assumption can
be replaced by the following one : for all $\theta\in\Theta$, there
exists a neighborhood $N_{t^*(\theta)}$ of $t^*(\theta)$ such that
$$\mathbb{E}[\sup_{t\in N_{t^*(\theta)}} \left\|\partial m(X,\theta,t)/\partial
t\right\|]<\infty, \quad\mathbb{E} [\sup_{t\in N_{t^*(\theta)}} \left\|\partial^2
m(X,\theta,t)/\partial
t^2\right\|]<\infty$$ and the matrix $\mathbb{E}\left[\partial^2 m(X,\theta,t^*(\theta))/\partial
t^2\right]<\infty$ is nonsingular, which implies the uniqueness
of $t^*(\theta)$, for all $\theta\in\Theta$.
\end{remark}
\begin{theorem}\label{theoreme 6}
Under Assumption \ref{assumption 6}, we have
\begin{enumerate}
\item [1)] $\|\widehat{t}(\theta)-t^*(\theta)\|$ converges in
probability to $0$ uniformly in $\theta\in\Theta$.
\item [2)] $\widehat{\theta}_\varphi$ converges in probability to
$\theta^*$;
\item [3)] $\widehat{D}_\varphi(\mathcal{M},P_0)$ converges in
probability to $D_\varphi(\mathcal{M},P_0)$.
\end{enumerate}
\end{theorem}
\noindent The asymptotic normality of the test statistics under
misspecification requires the following additional conditions.
\begin{assumption}
\label{assumption 7}
a) $\theta ^{\ast }\in \text{int}(\Theta )$;
b) there exists $\mathcal{N}\subset \Theta\times\Lambda_\Theta$,
some compact neighborhood of
$(\theta^*,t^*(\theta^*))$, such that with probability one $(\theta, t)\in \mathcal{N}\mapsto m(X,\theta,t)$
is $\mathcal{C}^2$ and
$$\mathbb{E}[ \sup_{(\theta,t)\in\mathcal{N}}\|\partial m(X,\theta,t)/\partial (\theta, t)\|]<\infty, \quad
\mathbb{E}[ \sup_{(\theta,t)\in\mathcal{N}}\|\partial^2m(X,\theta,t)/\partial (\theta, t)^2\|]<\infty;$$
c) as $\delta\to 0$,
$$\mathbb{E}\left\{\sup_{\{(t,\theta); \|(t,\theta)-(t^*(\theta^*),\theta^*)\|\leq
\delta\}}\left\|\partial^2m(X,\theta,t)/\partial (\theta,t)^2-
\partial^2m(X,\theta^*,t^*(\theta^*))/\partial (\theta,t)^2\right\|\right\}\to 0;$$
d) $\mathbb{E}\left[ m(X,\theta ^{\ast },t^{\ast
}(\theta ^{\ast }))^{2}\right] ,$ $\mathbb{E}\left[ \left\Vert
\partial m(X,\theta ^{\ast },t^{\ast }(\theta ^{\ast }))/\partial
t\right\Vert ^{2}\right] $ and $\mathbb{E}\left[ \left\Vert
\partial m(X,\theta ^{\ast },t^{\ast }(\theta ^{\ast })/\partial
\theta \right\Vert ^{2}\right] $ are finite, and the matrix
\[
S:=\left(
\begin{array}{cc}
S_{11} & S_{12} \\
S_{21} & S_{22} \\
\end{array}
\right) ,
\]
is nonsingular, where $S_{11}:=\mathbb{E}\left[
\partial ^{2}m(X,\theta ^{\ast },t^{\ast }(\theta ^{\ast
}))/\partial t^{2}\right] $,\\
$S_{12}={S_{21}}^{\top}:=\mathbb{E}\left[ \partial ^{2}m(X,\theta
^{\ast },t^{\ast }(\theta ^{\ast }))/\partial t\partial \theta
\right] $ and $S_{22}:=\mathbb{E}\left[ \partial ^{2}m(X,\theta
^{\ast },t^{\ast }(\theta ^{\ast }))/\partial \theta ^{2}\right].$
\end{assumption}
\begin{remark}\label{remark 2 sur les conditions de 3 ordre}
Assumption \ref{assumption 7}.c is used here to relax the
condition on the third derivatives (in $t$ and $\theta$) of the function $(\theta,t)\mapsto
m(X,\theta,t)$.
\end{remark}
\begin{theorem}\label{theoreme 7}
Under Assumptions \ref{assumption 6} and \ref{assumption 7}, we
have
\begin{enumerate}
\item [1)]
\begin{equation*}
\sqrt{n}\left(
\begin{array}{c}
\widehat{t}(\widehat{\theta}_\varphi)-t^*(\theta^*) \\
\widehat{\theta}_\varphi-\theta^* \\
\end{array}
\right)
\end{equation*}
converges in distribution to a centered normal random vector with
covariance matrix
\begin{equation*}\label{matrice limit a l'exter du model}
W := S^{-1}M{S^{-1}}
\end{equation*}
where
\begin{equation*}
M := \mathbb{E}\left[\left[
\begin{array}{c}
\frac{\partial}{\partial t}m\left(X,\theta^*,t^*(\theta^*)\right) \\
\frac{\partial }{\partial \theta}m\left(X,\theta^*,t^*(\theta^*)\right) \\
\end{array}
\right]\left[
\begin{array}{c}
\frac{\partial}{\partial t}m\left(X,\theta^*,t^*(\theta^*)\right) \\
\frac{\partial }{\partial \theta}m\left(X,\theta^*,t^*(\theta^*)\right) \\
\end{array}
\right]^{\top}\right];
\end{equation*}
\item [2)]
$\sqrt{n}\left(\widehat{D}_\varphi(\mathcal{M},P_0)-D_\varphi(\mathcal{M},P_0)\right)$
converges in distribution to a centered normal random variable
with variance
$$\sigma^2(\theta^*):=\mathbb{E}\left[m(X,\theta^*,t^*(\theta^*))^2\right]-
\left[\mathbb{E}\left[m(X,\theta^*,t^*(\theta^*))\right]\right]^2.$$
\end{enumerate}
\end{theorem}
\begin{remark}{\rm
In the case of EL, i.e., when $\varphi(x)=-\log x +x -1$,
Assumption \ref{assumption 6}.c implies that
\begin{equation}\label{bounded}
-\infty < \inf_{x\in\mathbb{R}^m} t_0+t^{{\top}} g(x,\theta)\leq
\sup_{x\in\mathbb{R}^m}t_0+t^{\top}g(x,\theta)<1
\end{equation}
for all $x\in\mathbb{R}^m-P_0$-a.s., for all $\theta\in \Theta$ and for all $t\in N_{t^*(\theta)}$. This
imposes a restriction on the model when the support of $P_0$ and the functions $g_j$
are unbounded. Indeed, when the support of $P_0$
is for example the whole space $\mathbb{R}^m$, the condition above
does not hold when $g$ is unbounded. In this case, the EL estimate may cease to be consistent
as it is stated by \cite{Schennach2007} under misspecification. This is a potential problem
for all divergences associated to $\varphi$-functions with domain
of the form $(a,+\infty[$, $]-\infty,b)$ or $(a,b),$ where $a$ and
$b$ are some finite real numbers; it is the case of modified $\chi^2$,
Hellinger, KL and modified $KL$ divergences.
At the contrary, Assumption 6.c may be satisfied for other divergences
associated to $\varphi$ functions with
$\text{dom}
\varphi=\mathbb{R}$ which
is the case of $\chi^2$ divergence for example.}
\end{remark}
\begin{remark} \label{remark power approximation}
{\rm Theorem \ref{theoreme 7} part 2 is useful for the computation
of the power function. For testing the null hypothesis
$\mathcal{H}_0 : P_0\in\mathcal{M}$ against the alternative
$\mathcal{H}_1:P_0\notin\mathcal{M}$, the power function is
\begin{equation}\label{power function model}
P_0\notin\mathcal{M}\mapsto
\beta(P_0):=P_0\left[2n\widehat{D}_{\varphi }\left(
\mathcal{M},P_0\right)
>q_{(1-\alpha)}\right].
\end{equation}
Using Theorem \ref{theoreme 7} part 2, we obtain the following
approximation to the power function (\ref{power function model}):
\begin{equation}\label{power approxi}
\beta(P_0)\thickapprox 1-F_{\mathcal{N}}\left[
\frac{\sqrt{n}}{\sigma\left( \theta ^{\ast }\right) }\left(
\frac{q_{(1-\alpha)}}{2n}- D_\varphi\left( \mathcal{M},P_0\right)
\right) \right]
\end{equation}
where $F_\mathcal{N}$ is the empirical cumulative distribution of
the standard normal distribution. From the proxy value of
$\beta(P_0)$ hereabove, the approximate sample size that ensures
a given power $\beta$ for a given alternative $P_0\not\in
\mathcal{M}$ can be obtained as follows. Let $n_{0}$ be the
positive root of the equation
\[
\beta =1-F_{\mathcal{N}}\left[ \frac{\sqrt{n}}{\sigma( \theta^*)
}\left( \frac{q_{(1-\alpha) }}{2n}-D_{\varphi }\left(
\mathcal{M},P_0\right) \right) \right]
\]
i.e.,
\[
n_{0}=\frac{\left( a+b\right) -\sqrt{a\left( a+2b\right)
}}{2D_\varphi\left( \mathcal{M},P_0\right) ^{2}},
\]
where $a:=\sigma(\theta^*) ^{2}\left[ F_{\mathcal{N}}^{-1}\left(
1-\beta \right) \right] ^{2}$ and $b:=q_{(1-\alpha) } D_{\varphi
}\left( \mathcal{M},P_0\right).$ The required sample size is then
$\left\lfloor n_{0}\right\rfloor +1$.}
\end{remark}
\section{Simulation results: Approximation of the power function of
the empirical likelihood ratio test}
\noindent We will illustrate by simulation the accuracy of the
power
approximation (\ref{power approxi}) in the case of EL method, i.e., when
$\varphi(x)=-\log x +x-1.$
Consider the test problem of the composite null hypothesis
$$\mathcal{H}_0: P_0\in\mathcal{M}\quad \text{against the alternative}
\quad \mathcal{H}_1:P_0\notin \mathcal{M},$$
where
$\mathcal{M}:=\bigcup_{\theta\in\mathbb{R}}\mathcal{M}_\theta$ and
$\mathcal{M}_\theta$ is the set of all s.f.m's satisfying the constraints $\int dQ(x)=1$
and $\int g(x,\theta)~dQ(x)=0$ with $g(x,\theta):=(x,
x^2-\theta)^{\top}$, namely
$$\mathcal{M}_\theta:=\left\{Q\in M ~ \text{ such that } ~ \int_\mathbb{R} dQ(x)=1 \text{ and }
\int_\mathbb{R} g(x,\theta)~dQ(x)=0 \right\},$$
where $\theta\in\mathbb{R}$ is the parameter of interest. We
consider the asymptotic level $\alpha=0.05$ and the alternatives
$P_0:=\mathcal{U}([-1,1+\epsilon])\not\in\mathcal{M}$ for different values of
$\epsilon$ in the interval $]0,1]$. Note that when $\epsilon=0$ then the
uniform distribution $\mathcal{U}([-1,1])$ belongs to
the model $\mathcal{M}$. For this model, we can show also that all Assumptions of Theorem \ref{theoreme 2}
are satisfied when $\epsilon=0$, and all Assumptions of Theorem
\ref{theoreme 7} are met under alternatives. In Figure
\ref{fig power approxim}, the power function (\ref{power function
model}) is plotted (with a continuous line),
with sample sizes $n=50, n=100$, $n=200$ and $n=500$, for different values
of $\epsilon$. Each power entry was obtained by Monte-Carlo from
$1000$ independent runs. The approximation (\ref{power approxi}) is
plotted (with a dashed line) as a function of $\epsilon$.
The estimates $\widehat{\theta}_\varphi$ and $\widehat{D}_\varphi(\mathcal{M},P_0)$
are calculated using the Newton-Raphson algorithm. We observe from
Figure \ref{fig power approxim} that the approximation is accurate
even for moderate sample sizes.
\begin{figure}[!h]
\caption[]{Approximation of the power function}
\centerline{
\begin{tabular}{ c c }
\includegraphics[width=.55\textwidth]{im50}
& \includegraphics[width=.55\textwidth]{im100}
\end{tabular} }
\par
\centerline{
\begin{tabular}{ c c }
\includegraphics[width=.55\textwidth]{im200}
& \includegraphics[width=.55\textwidth]{im500}
\end{tabular} }
\label{fig power approxim}
\end{figure}
\section{Concluding remarks and possible developments}
\noindent We have proposed new estimates and tests for model
satisfying linear constraints with unknown parameter through
divergence based methods which generalize the EL approach. This
leads to the obtaining of the limiting distributions of the test
statistics and the estimates under alternatives and under
misspecification. Consistency of the test statistics under the
alternatives is the starting point for the study of the optimality
of the tests through Bahadur approach; also the generalized
Neyman-Pearson optimality of EL test (as developed by
\cite{Kitamura2001}) can be adapted for empirical divergence based
methods. Many problems remain to be studied in the future such as
the choice of the divergence which leads to an optimal (in some
sense) estimator or test in terms of efficiency and/or robustness.
Preliminary simulation results show that Hellinger divergence
enjoys good properties in terms of efficiency-robustness; see
\cite{BK_2006_div_emp}. Also comparisons under local alternatives
should be developed.
\section{Appendix}
\noindent \textbf{Proof of Theorem \ref{theoreme 1}}.\\ The same
arguments, used for the proof of Theorem 3.1 in
\cite{NeweySmith2004}, hold when their criterion function
$(\theta,\lambda)\in\Theta\times\mathbb{R}^l\mapsto
\frac{1}{n}\sum_{i=1}^n \rho(\lambda^{\top}g(X,\theta))$ is
replaced by our function
$(\theta,t)\in\Theta\times\mathbb{R}^{1+l}\mapsto
\frac{1}{n}\sum_{i=1}^n m(t^{\top}\overline{g}(X,\theta)).$ In
particular, we have $$\max_{1\leq i\leq
n}\left|\widehat{t}(\widehat{\theta}_\varphi)^{\top}\overline{g}(X_i,\widehat{\theta}_\varphi)\right|\to
0$$ in probability, which implies that
$\widehat{t}(\widehat{\theta}_\varphi)\in\text{int}(\Lambda^{(n)}_{\widehat{\theta}_\varphi})$
with probability one as $n\to\infty$, since $a^*<0<b^*$.\\
\noindent \textbf{Proof of Theorem \ref{theoreme 2}}.\\ The proof
is similar to that of \cite{NeweySmith2004} Theorem 3.2. Hence, it
is
omitted.\\
\noindent \textbf{Proof of Theorem \ref{theoreme 3}}.\\
It is a particular case of Theorem \ref{theoreme 1} taking
$\Theta=\{\theta\}.$ Hence, the proof is omitted.\\
\noindent \textbf{Proof of Theorem \ref{theoreme 4}}.\\ 1) First,
note that $t^*(\theta)$ exists and is unique by Assumption
\ref{assumption 4}. By the uniform weal law of large numbers
(UWLLN), using continuity of $m(X,\theta,t)$ in $t$, and
Assumption \ref{assumption 4}.b, we obtain
\begin{equation}\label{eqn3}
\left|P_nm(\theta,t)-\mathbb{E}\left[m(X,\theta,t)\right]\right|\to
0,\end{equation} in probability uniformly in $t$ over the compact
set $N_{t^*(\theta)}$. Using this and the fact that
$t^*(\theta):=\arg\sup_{t\in\Lambda_\theta}P_0m(\theta,t)$ is
unique and belongs to $\text{int}(N_{t^*(\theta)})$ and the strict
concavity of $t\mapsto P_0m(\theta,t)$, we conclude that any
value
\begin{equation}\label{eqn t bare}\overline{t}:=\arg\sup_{t\in N_{t^*(\theta)}}P_nm(\theta,t)
\end{equation} converges in
probability to $t^*(\theta)$; see e.g. Theorem 5.7 in
\cite{vanderVaart1998}. We end then the proof by showing that
$\widehat{t}(\theta)$ belongs to $\text{int}(N_{t^*(\theta)})$
with probability one as $n\to \infty$, and therefore it converges
to $t^*(\theta).$ In fact, since for $n$ sufficiently large any
value $\overline{t}$ lies in the interior of $N_{t^*(\theta)}$,
concavity of $t\mapsto P_nm(\theta,t)$ implies that no other point
$t$ in the complement of $\text{int}(N_{t^*(\theta)})$ can
maximize $P_nm(\theta,t)$ over $t\in\mathbb{R}^{1+l}$, hence
$\widehat{t}(\theta)$ must belongs
to $\text{int}(N_{t^*(\theta)})$.\\
2) With probability tending to $1$ as $n\to \infty$, we have
$\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)=P_nm(\theta,\widehat{t})=P_nm(\theta,\overline{t})$.
Hence, we can write
$$\left|\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)-D_\varphi(\mathcal{M}_{\theta},P_0)\right|=
\left|P_nm(\theta,\overline{t})-P_0m(\theta,t^*(\theta))\right|=:|A|,$$
and
$$P_nm(\theta,t^*(\theta))-P_0m(\theta,t^*(\theta))\leq A\leq
P_nm(\theta,\overline{t})-P_0m(\theta,\overline{t}).$$
Both the RHS and the LHS in the above display tend to $0$ in
probability by (\ref{eqn3}). Hence,
$\left|\widehat{D}_\varphi(\mathcal{M}_{\theta},P_0)-D_\varphi(\mathcal{M}_{\theta},P_0)\right|$
tends to $0$ in probability as $n\to\infty$. This ends the proof.\\
\noindent \textbf{Proof of Theorem \ref{theoreme 5}}.\\ 1) For $n$
sufficiently large, by a Taylor expansion, there exists
$\overline{t}\in \mathbb{R}^{1+l}$ inside the segment that links
$\widehat{t}$ and $t^*(\theta)$ with
\begin{equation}\label{Taylor Expansion}
\begin{array}{ccl}
0 & = & P_nm'(\theta,\widehat{t})\\
& = &
P_nm'(\theta,t^*(\theta))+\left(P_n m''(\theta,\overline{t})\right)^{\top}
\left(\widehat{t}-t^*(\theta)\right).
\end{array}
\end{equation}
By Assumptions \ref{assumption 5}.a and \ref{assumption 5}.b,
using the fact that $\overline{t}=t^*(\theta)+o_P(1)$ and the
UWLLN, we can prove that $$P_n m''(\theta,\overline{t})= P_0
m''(\theta,t^*(\theta))+o_P(1).$$ Using this display, one gets
from (\ref{Taylor Expansion})
\begin{equation}\label{equation loi limite}
-P_nm'(\theta,t^*(\theta))=\left(P_0m''(\theta,t^*(\theta))+o_P(1)\right)
\left(\widehat{t}-t^*(\theta)\right).
\end{equation}
Assumptions \ref{assumption 4}.a and \ref{assumption 5}.a imply
that $P_0m'(\theta,t^*(\theta))=0$. Hence, by the central limit
theorem (CLT), we have
$$\sqrt{n}P_nm'(\theta,t^*(\theta))=O_P(1),$$
which by (\ref{equation loi limite}) implies that
$\sqrt{n}\left(\widehat{t}-t^*(\theta)\right)=O_P(1).$ Hence, from
(\ref{equation loi limite}), we get
\begin{equation}\label{equation qui donne la loi limite}
\sqrt{n}\left(\widehat{t}-t^*(\theta)\right)=
{\left[-P_0m''(\theta,t^*(\theta))\right]}^{-1}
\sqrt{n}P_nm'(\theta,t^*(\theta))+o_P(1).
\end{equation}
The CL and Slutsky theorems conclude the proof of part 1.\\
2) Using the fact that
$\left(\widehat{t}-t^*(\theta)\right)=O_P(1/\sqrt{n})$ and
$P_nm'(\theta,t^*(\theta))=P_0m'(\theta,t^*(\theta))+o_P(1)=0+o_P(1)=o_P(1)$,
we obtain
\begin{eqnarray}
\sqrt{n}\left(\widehat{D}_\varphi(\mathcal{M}_\theta,P_0)-
D_\varphi(\mathcal{M}_\theta,P_0)\right) & = & \sqrt{n}\left(
P_nm(\theta,\widehat{t})-P_0m(\theta,t^*(\theta))\right)\nonumber\\
& = &
\sqrt{n}\left(P_{n}m(\theta,t^*(\theta))-P_0m(\theta,t^*(\theta))\right)+o_{P}(1),\nonumber
\end{eqnarray}
and the CL and Slutsky theorems conclude the proof.\\
\noindent \textbf{Proof of Theorem \ref{theoreme 6}}.\\ 1) First
note that Assumption \ref{assumption 6}.d implies that the
function $t\in\Lambda_\theta\mapsto \mathbb{E} m(X,\theta,t)$ is
strictly concave for all $\theta\in\Theta$, which implies that
$t^*(\theta)$ is unique for all $\theta\in\Theta$. By the UWLLN,
using continuity of $m(X,\theta,t)$, in $\theta$ and $t$, and
Assumption \ref{assumption 6}.c, we obtain the uniform convergence
in probability, over the compact set
$\left\{(\theta,t)\in\Theta\times\mathbb{R}^{1+l};
~\theta\in\Theta, t\in N_{t^*(\theta)}\right\}$,
\begin{equation}\label{eqn 4}\sup_{\{\theta\in\Theta,t\in N_{t^*(\theta)}\}}\left|P_nm(\theta,t)
-P_0m(\theta,t)\right|\to 0.
\end{equation} We can then prove
the convergence in probability
$\sup_{\theta\in\Theta}\|\widehat{t}(\theta)-t^*(\theta)\|\to 0$
in two steps. Step 1: Let $\eta >0$. We will show that
$P_0\left[\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|\geq
\eta\right]\to 0$ for any value
\begin{equation}\label{eqn t bare theta}
\overline{t}(\theta):=\arg\sup_{t\in N_{t^*(\theta)}}P_nm(\theta,t).
\end{equation}
Step 2: To conclude the proof, we will show that
$\widehat{t}(\theta)$ belongs to $\text{int}(N_{t^*(\theta)})$
with probability one as $n\to\infty$ for all $\theta\in\Theta$.
Let $\eta
>0$ such that
$\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|\geq
\eta$. Sine $\Theta$ is a compact set, by continuity there exists
$\overline{\theta}\in\Theta$ such that
$\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|=
\|\overline{t}(\overline{\theta})-t^*(\overline{\theta})\|\geq
\eta$. Hence, there exists $\varepsilon >0$ such that
$P_0m(\overline{\theta},t^*(\overline{\theta}))-
P_0m(\overline{\theta},\overline{t}(\overline{\theta}))>\varepsilon$.
In fact, $\varepsilon$ may be defined as follows
$$\varepsilon := \inf_{\theta\in\Theta}
\sup_{\{t\in N_{t^*(\theta)};~ \|t-t^*(\theta)\|\geq \eta\}}
\mathbb{E}[m(X,\theta,t^*(\theta))]- \mathbb{E}[m(X,\theta,t)],$$
which is strictly positive by the strict concavity of
$\mathbb{E}[m(X,\theta,t)]$ in $t$ for all $\theta\in\Theta$, the
uniqueness of $t^*(\theta)\in\text{int}(N_{t^*(\theta)})$ and the
fact that $\Theta$ is compact. Hence the event
$\left[\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|\geq
\eta\right]$ implies the event
$$\left[P_0m(\overline{\theta},t^*(\overline{\theta}))-
P_0m(\overline{\theta},\overline{t}(\theta))\geq \varepsilon
\right],$$ from which we obtain
\begin{equation}\label{eqn 5}
P_0\left[\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|\geq
\eta\right]\leq P_0\left[
P_0m(\overline{\theta},t^*(\overline{\theta}))-
P_0m(\overline{\theta},\overline{t}(\theta))\geq
\varepsilon\right].
\end{equation}
On the other hand, by (\ref{eqn 4}), we have
\begin{eqnarray} P_0m(\overline{\theta},t^*(\overline{\theta}))-
P_0m(\overline{\theta},\overline{t}(\theta)) & = &
P_nm(\overline{\theta},t^*(\overline{\theta}))-
P_0m(\overline{\theta},\overline{t}(\theta))+o_P(1)\nonumber\\
& \leq & P_nm(\overline{\theta},\overline{t}(\overline{\theta}))-
P_0m(\overline{\theta},\overline{t}(\theta))+o_P(1)\nonumber\\
& \leq & \sup_{\{\theta\in\Theta, t\in N_{t^*(\theta)}\}}
|P_nm(\theta,t)-P_0m(\theta,t)|+o_P(1).\nonumber
\end{eqnarray}
Combining this with (\ref{eqn 5}) and (\ref{eqn 4}), we conclude
that
\begin{equation}\label{convergence en proba de t bare theta}
\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|\to 0
\end{equation}
in probability. In particular,
$\overline{t}(\theta)\in\text{int}(N_{t^*(\theta)})$ for
sufficiently large $n$, uniformly in $\theta\in \Theta$. Since
$t\mapsto P_nm(\theta,t)$ is concave, then the maximizer
$\widehat{t}(\theta)$ belongs to $\text{int}(N_{t^*(\theta)})$ for
sufficiently large $n$; hence the
same result (\ref{convergence en proba de t bare theta}) holds when
$\overline{t}(\theta)$ is replaced by $\widehat{t}(\theta)$.\\
2) From part 1, we have for large $n$,
\begin{eqnarray}
\sup_{\theta\in\Theta}|P_nm(\theta,\widehat{t}(\theta))-P_0m(\theta,t^*(\theta))|
& = &
\sup_{\theta\in\Theta}|P_nm(\theta,\overline{t}(\theta))-P_0m(\theta,t^*(\theta))|=:
\sup_{\theta\in\Theta}|B|.\nonumber
\end{eqnarray}
On the other hand, we have
$$P_nm(\theta,t^*(\theta))-P_0m(\theta,t^*(\theta))\leq B\leq P_nm(\theta,\overline{t}(\theta))
-P_0m(\theta,\overline{t}(\theta)).$$
By Assumption \ref{assumption 6}.c, and the convergence in
probability $\sup_{\theta\in\Theta}\|\overline{t}(\theta)-t^*(\theta)\|\to
0$,
both the RHS and LHS of the above display
tends to $0$ in probability uniformly in $\theta\in\Theta$, by the UWLLN. Hence,
$\sup_{\theta\in\Theta}|P_nm(\theta,\widehat{t}(\theta))-P_0m(\theta,t^*(\theta))|\to 0$
in probability. Now, since the minimizer $\theta^*$ of $\theta\mapsto
P_0m(\theta,t^*(\theta))$
over the compact set $\Theta$
is unique and interior point of $\Theta$, by continuity and the above uniform convergence, we
conclude that $\widehat{\theta}_\varphi$ tends in probability to
$\theta^*$; see e.g. Theorem 5.7 in
\cite{vanderVaart1998}.\\
3) This holds as a consequence of the uniform convergence in
probability
\begin{equation}\label{conv unif en theta}
\sup_{\theta\in\Theta}|P_nm(\theta,\widehat{t}(\theta))-P_0m(\theta,t^*(\theta))|\to 0
\end{equation}
proved in part 2 above. In fact, we have for $n$
sufficiently large
$$ |\widehat{D}_\varphi(\mathcal{M}, P_0)-D_\varphi(\mathcal{M}, P_0)|=
|P_nm(\widehat{\theta},\widehat{t}(\widehat{\theta}))-P_0m(\theta^*,t^*(\theta^*))|=:|C|,$$
with
$$ P_nm(\widehat{\theta},\widehat{t}(\widehat{\theta}))-P_0m(\widehat{\theta},t^*(\widehat{\theta}))
\leq C
\leq P_nm(\theta^*,\widehat{t}(\theta^*))-P_0m(\theta^*,t^*(\theta^*))$$
and both the RHS and LHS tend to $0$ in probability by (\ref{conv unif en
theta}). This concludes the proof.\\
\noindent \textbf{Proof of Theorem \ref{theoreme 7}}.\\
1) By the
first order conditions, with probability tending to one, we have
\begin{equation*}
\left\{
\begin{array}{lll}
P_n\frac{\partial}{\partial
t}m\left(\widehat{\theta},\widehat{t}(\widehat{\theta})\right)
& = & 0\\
P_n\frac{\partial}{\partial\theta}
m\left(\widehat{\theta},\widehat{t}(\widehat{\theta})\right)+P_n
\frac{\partial}{\partial t}
m\left(\widehat{\theta},\widehat{t}(\widehat{\theta})\right)
\frac{\partial}{\partial \theta}\widehat{t}(\widehat{\theta}) & =
& 0.
\end{array}
\right.
\end{equation*}
The second term in the LHS of the second equation is equal to $0$,
due to the first equation. Hence, $\widehat{t}(\widehat{\theta})$
and $\widehat{\theta}$ are solutions of the somehow simpler system
\begin{eqnarray}
P_n\frac{\partial}{\partial
t}m\left(\widehat{\theta},\widehat{t}(\widehat{\theta})\right)
& = & 0 \label{E1} \\
P_n\frac{\partial}{\partial\theta}
m\left(\widehat{\theta},\widehat{t}(\widehat{\theta})\right) & = &
0. \label{E2}
\end{eqnarray}
Using a Taylor expansion in (\ref{E1}) in
$(\widehat{\theta},\widehat{t})$ around $(\theta^*,t^*)$; there
exists $\left(\overline{\theta},\overline{t}\right)$ inside the
segment that links
$(\widehat{\theta},\widehat{t}(\widehat{\theta}))$ and
$(\theta^*,t^*(\theta^*))$ such that
\begin{eqnarray}\label{Taylor 1 E1}
0 & = & P_n\frac{\partial}{\partial t}m\left(\theta^*,t^*(\theta^*)\right)+
\left[\left(P_n\frac{\partial^2}{\partial t^2}m(\overline{\theta},\overline{c})\right)^{\top},
\left(P_n\frac{\partial^2}{\partial \theta\partial t}
m(\overline{\theta},\overline{c})\right)^{\top}\right]a_n\nonumber\\
\end{eqnarray}
with
\begin{equation}\label{a n }
a_n:={\left({\left(\widehat{t}(\widehat{\theta})-t^*(\theta^*)\right)}^{\top},
{\left(\widehat{\theta}-\theta^*\right)}^{\top}\right)}^{\top}.
\end{equation}
\noindent By Assumption \ref{assumption 7}, using the UWLLN, we
can write
$$\left[P_n\frac{\partial^2}{\partial
t^2}m(\overline{\theta},\overline{c}),
P_n\frac{\partial^2}{\partial\theta\partial t}
m(\overline{\theta},\overline{c})\right]=\left[P_0\frac{\partial^2}{\partial t^2}
m(\theta^*,t^*(\theta^*)),
P_0\frac{\partial^2}{\partial\theta\partial t}
m(\theta^*,t^*(\theta^*))\right]+o_P(1),$$
to obtain
from (\ref{Taylor 1 E1})
\begin{equation}\label{Equation E1 donne}
-P_n\frac{\partial}{\partial
t}m(\theta^*,t^*)=
\left[\left(P_0\frac{\partial^2}{\partial t^2}m(\theta^*,t^*)\right)^{\top}+o_P(1),
\left(P_0\frac{\partial^2}{\partial\theta\partial
t}m(\theta^*,t^*)\right)^{\top}
+o_P(1)\right]a_n.
\end{equation}
In the same way, using a Taylor expansion in (\ref{E2}), we obtain
\begin{equation}\label{Equation E2 donne}
-P_n\frac{\partial}{\partial
\theta}m(\theta^*,t^*)=
\left[\left(P_0\frac{\partial^2}{\partial t\partial\theta}m(\theta^*,t^*)
\right)^{\top}+o_P(1),
\left(P_0\frac{\partial^2}{\partial
\theta^2}m(\theta^*,t^*)\right)^\top+o_P(1)\right]a_n.
\end{equation}
From (\ref{Equation E1 donne}) and (\ref{Equation E2 donne}), we
get
\begin{eqnarray}\label{racine de n a n 1}
\sqrt{n}a_n & = & \sqrt{n}\left(
\begin{array}{cc}
P_0\frac{\partial ^2}{\partial t^2}m(\theta^*,t^*) & \left(P_0\frac{\partial^2}
{\partial\theta\partial t} m(\theta^*,t^*)\right)^{\top}\\
\left(P_0\frac{\partial^2}{\partial t\partial \theta} m(\theta^*,t^*)\right)^{\top} &
P_0\frac{\partial^2}{\partial \theta^2}m(\theta^*,t^*) \\
\end{array}
\right)^{-1}\times\nonumber\\
& & \times\left(
\begin{array}{c}
-P_n\frac{\partial}{\partial t}m(\theta^*,t^*) \\
-P_n\frac{\partial }{\partial \theta} m(\theta^*,t^*)\\
\end{array}
\right)+o_P(1).
\end{eqnarray}
Denote $S$ the $(1+l+d)\times(1+l+d)-$matrix defined by
\begin{equation}\label{S}
S:=\left(
\begin{array}{cc}
S_{11} & S_{12} \\
S_{21} & S_{22} \\
\end{array}
\right):=\left(
\begin{array}{cc}
P_0\frac{\partial ^2}{\partial t^2}m(\theta^*,t^*) & \left(P_0\frac{\partial^2}
{\partial\theta\partial t} m(\theta^*,t^*)\right)^{\top}\\
\left(P_0\frac{\partial^2}{\partial t\partial \theta} m(\theta^*,t^*)\right)^{\top} &
P_0\frac{\partial^2}{\partial \theta^2}m(\theta^*,t^*) \\
\end{array}
\right).
\end{equation}
Hence, we obtain
\begin{equation*}
\sqrt{n}\left(
\begin{array}{c}
\widehat{t}(\widehat{\theta})-t^*(\theta^*) \\
\widehat{\theta} -\theta^*\\
\end{array}
\right)=\sqrt{n}S^{-1}\left(
\begin{array}{c}
-P_n\frac{\partial}{\partial t}m(\theta^*,t^*) \\
-P_n\frac{\partial}{\partial \theta}m(\theta^*,t^*)\\
\end{array}
\right)+o_P(1),
\end{equation*}
and the CL and Slutsky theorems conclude the proof.\\
2) Using the fact that
$$\widehat{t}(\widehat{\theta})-t^*(\theta^*)=O_P(1/\sqrt{n}),~
P_n\partial m(\theta^*,t^*(\theta^*))/\partial t=
P_0\partial m(\theta^*,t^*(\theta^*))/\partial t
+o_P(1)=o_P(1)$$
and $$\widehat{\theta}-\theta^*=O_P(1/\sqrt{n}),~
P_n\partial m(\theta^*,t^*(\theta^*))/\partial \theta=
P_0\partial m(\theta^*,t^*(\theta^*))/\partial \theta
+o_P(1)=o_P(1),$$
we can write
\begin{eqnarray}
\sqrt{n}\left(\widehat{D}_\varphi(\mathcal{M},P_0)-D_\varphi(\mathcal{M},P_0)\right)
& = &
\sqrt{n}\left(P_nm(\widehat{\theta},\widehat{t}(\widehat{\theta}))-P_0m(\theta^*,t^*(\theta^*))\right)\nonumber\\
& = & \sqrt{n}\left(P_nm(\theta^*,t^*(\theta^*))-P_0m(\theta^*,t^*(\theta^*))\right)+o_P(1),\nonumber
\end{eqnarray}
and the CL and Slutsky theorems end the proof.
\bibliographystyle{natbib}
|
\section{Introduction}
Before we will present our C++ Interface of TAUOLA,
for LHC era generators, let us first list projects and applications which helped us
to gain experience and to define the main tasks our interface
for a $\tau$ decay generator has to resolve.
In the present day experiments at High Energy Physics accelerators,
interpretation of results becomes increasingly involved.
Not only the detector
response is complex, but some theoretical effects need to be
removed. Otherwise results are difficult to interpret for the non-specialist.
For that purpose the concept of work with realistic and idealized observables was introduced as well
as finally with pseudo-observables which can be easily
understood by theorists, such as $W$, $Z$ masses or couplings.
Good examples of
such approaches were measurements of the two-fermion final states at LEP.
Because of increasing precision of the experimental measurements, definitions of
quantities to be measured, simple at first, later evolved into
several options \cite{Kobel:2000aw},
each based on the properties of individual detectors and each requiring individual discussion of the systematic
error.
One could assume, that if all theoretical effects are embodied into one
theoretical black-box
and, experiments while using it tune parameters (representing pseudo-observables) to the data, interpretation of the observed effects
could be separated into theoretical and experimental components.
Unfortunately this strategy is limited, as it leaves little room
for cases where
theory and experimental effects are convoluted: size and even nature of the theoretical corrections
depend on the experimental conditions. Such discussion on observables involving $\tau$ decays can be found in \cite{RichterWas:2004jf}.
Recently, discussion for the physics of
$\tau$ production and decay at low energies where similar aspects are addressed,
was presented in \cite{Actis:2009gg}.
For LHC experiments $\tau$
decays are not of primary interest in themselves, but rather
will be used to measure properties of $\tau$ production processes.
Let us explain this using examples. Physics effects necessary for
the prediction of hard processes
at the LHC experiments can be separated into several parts, among them:
parton showers,
the underlying event and structure functions, final
state QED bremsstrahlung, QED bremsstrahlung interference between initial and
final states and finally the hard process including electroweak corrections.
Such separation
is not only for the convenience of organizing theoretical work but
provides efficient and flexible
component in the framework for experimental data analysis
strategy (see eg. \cite{Aad:2009wy}).
Some
of such building blocks are of genuine theoretical interests, some others
are not so much. The
hard process
usually depends on the parameters intended for the measurement, eg.
$W$ or Higgs mass, or new coupling constants.
Other building blocks may be less interesting, nonetheless they may affect
the results of measurements. This is certainly true in the case of the underlying event
or missing transverse energy or $p_T$ distributions generated from parton
showers \cite{RichterWas:2004jf}. It may also be the case for QED final
state bremsstrahlung or
initial-final state interference (where potential difficulties may be expected
\cite{Abdallah:2005wn} and predictions may need to be fixed with the
help of experimental data).
The black-box approach, where all simulation segments are put together by
theorists, may look advantageous to the experimental user.
However in such a case one has less flexibility to distinguish experimental
effect from the theoretical ones, thus limiting control on the systematic errors.
The particular problems may be left unnoticed. Typically
the difficulties will not affect all
observables. Unfortunately, complications tend to show up only when more
detailed discussion on the systematic errors of experimental analysis is
performed.
In the present document we discuss the implementation of $\tau$ decays
into a simulation chain as a separate module which can be configured by the end
user. For the purpose of generation of $\tau$ decays themselves,
the {\tt TAUOLA} library, as described
in \cite{Jadach:1990mz,Jezabek:1991qp,Jadach:1993hs}
is used. This part of the code is expected to be
a black-box for the High Energy experimental user.
At present, from the technical side, the black-box consists of
the same {\tt FORTRAN} code as described
in \cite{Golonka:2003xt}. We will call it {\tt TAUOLA FORTRAN}. Such organization
makes it
easy for low energy phenomenologists or experimentalists to work on this part of the code,
such as the activities described in \cite{Actis:2009gg}, leading to the
new parametrization of hadronic $\tau$ decay currents becoming available for
High Energy experimental users in a rather straightforward way.
The role of the interface is to prepare information on the $\tau$
(four-momentum, spin state) in a format which is understood by
{\tt TAUOLA FORTRAN}, and as a post processing step to return (insert) $\tau$
decay products to the primary event record.
Finally, role of such interfacing code is to calculate dedicated
weights from the production process information as well as from the decay,
and unweight accordingly to standard MC procedures. Spin effects,
electroweak corrections and also effects of anomalous couplings can be
introduced in this way.
A rather modest version of {\tt TAUOLA Universal Interface} based on
{\tt FORTRAN HEPEVT}
event record
is described in \cite{Golonka:2003xt}, we will call it
{\tt TAUOLA Fortran Interface}. A new version of {\tt TAUOLA Universal Interface}
based on {\tt HepMC} \cite{Dobbs:2001ck}, the most popular event record
of C++, will be documented here. It also includes new
functionalities.
We will call it {\tt TAUOLA C++ Universal Interface}, or, if no ambiguity
could arise simply {\tt TAUOLA C++ Interface} or just
{\tt TAUOLA Interface} if it is clear that the C++ not the
FORTRAN version is in mind\footnote{The main class of
{\tt TAUOLA C++ Universal Interface} is called {\tt Tauola}.}.
The {\tt PHOTOS} generator for
QED bremsstrahlung in decays, which was previously distributed together with {\tt TAUOLA FORTRAN}, is not discussed in the present
paper\footnote{{\tt PHOTOS} was distributed together with
{\tt TAUOLA} since ref. \cite{Golonka:2003xt}. That is why it is present in our
{\tt TAUOLA FORTRAN} but will be not used here. }. It is now embodied in a separate module
\cite{Barberio:1990ms,Barberio:1994qi} of the Monte Carlo simulation chain,
as the {\tt PHOTOS} generator has found significant applications outside the domain of $\tau$ decays.
That is why, in the future, we plan to distribute the C++ version of {\tt PHOTOS}
separately \cite{photosC++}.
Important physics improvements, with respect to the {\tt TAUOLA FORTRAN Interface} implementation,
described in \cite{Golonka:2003xt}, comes with the code presented here. In
particular, transverse spin correlations have been implemented for
processes mediated by $Z/\gamma^*$ and genuine electroweak corrections
are now available for such processes.
With the new interface, it is rather straightforward to implement effects beyond
standard
model physics. Only read-in data-tables should be replaced,
no modification
to the code itself is needed. Further minor extensions
include an algorithm to
decay a single $\tau$ with a user defined polarization, or the availability of methods to
access generated $\tau$ leptons helicity states.
\section{Requirements for {\tt TAUOLA Interface} }
\subsection{General Requirements}
For a $\tau$ decay to be generated it is enough to know its spin
state and define the frame in which the decay should be performed. In case there are
more than one $\tau$ lepton in the final state, the quantum spin state of
both (or more) $\tau$ leptons must be provided in the form of a density matrix.
The exact algorithm for the generation of spin correlations exist
since the papers \cite{Jadach:1990mz,Jezabek:1991qp,Jadach:1993hs}.
However, for the algorithm functioning, the density matrix must be known exactly as well.
In practice, the actual form of the spin density matrix
(in our present paper it will be called $R_{ij}$, exactly as in the original
{\tt TAUOLA FORTRAN} and its documentation) is often available with some
approximations only.
With the increasing precision of experiments,
one may need to remove certain approximations introduced into $R_{ij}$.
Already now, our program features, as an option,
complete spin effects in decays of $\tau$ pairs originating from the annihilation
of quarks. Effects of genuine weak corrections are included, and an extension
for the implementation of new physics signatures is straightforward.
One should not forget that the density matrix itself is not the only place
approximations occur. Effects of higher
order QCD
corrections need to be taken into account to define the kinematical configuration
of initial partons
used in spin density calculations (otherwise the density matrix for each individual
process would have to be provided). At present, this is available in the
leading (collinear) approximation only.
Before we will discuss details of specific implementation, let us recall
first, the minimal list of steps the interface has to perform, independent
of the programming language and data structure used.
\begin{enumerate}
\item The $\tau$ lepton or lepton pair(s) have to be localized in the
event record. For processes mediated by $W$ bosons (or charged Higgs)
$\nu_\tau$ has to be localized as well.
\item
If possible, the hard process leading to $\tau$ production has to be
determined. This is necessary to control transverse spin correlations.
Preferably minimal information from the host
generators should be used. This is to reduce dependence on the host program\footnote{We
target also applications when event records will be filled by
measured data. For example,
measured $\mu^+ \mu^-$ events can be modified and final state muons replaced
appropriately with $\tau^+ \tau^-$ pairs. Then generation of $\tau$ lepton decay is necessary.}.
\item Flavours and orientation of fields entering the production vertex for intermediate states, such as
$Z/\gamma^*$,
have to be reconstructed too. This orientation (with respect to $\tau^\pm$
rest-frames) is necessary for calculation of the $\tau^\pm$ spin density matrix,
if spin of intermediate state is different from zero.
\item The relative orientation of $\tau^+$ and $\tau^-$ rest-frames
should be established and respected by Lorentz transformations.
\item Transformation of $\tau$ decay products from the $\tau$ rest-frame to
lab frame has to be performed and the event record has to be completed with $\tau$ decay products.
\end{enumerate}
\subsection{C++ Specific Requirements}
The C++ version of the {\tt TAUOLA Interface} implements all functionalities
of its predecessor, the {\tt TAUOLA Interface} coded in FORTRAN \cite{Golonka:2003xt}.
It can be attached to any Monte-Carlo program where $\tau$'s
are generated, provided its output is available through a {\tt HepMC} \cite{Dobbs:2001ck} event record.
This condition is not very restrictive, is seems that {\tt HepMC} will
remain a generally accepted standard for the near future. However,
already now several different options for how {\tt HepMC} is used are
widespread. Possibility of the flexible adaptation of our event record
interface to different
options has been considered in the design, drawing experience
from {\tt MC-TESTER} \cite{Golonka:2002rz,Davidson:2008ma}.
We have also
envisaged the possibility that {\tt HepMC} may one day be replaced by another
standard of event record, and we have provided an easy way to extend
the interface to a possible new event record standard.
\subsection{Object Oriented Event Records -- The Case of {\tt HepMC}}
In adopting {\tt TAUOLA Interface} to the C++ event record format
the difference between the {\tt HEPEVT} event record used in the FORTRAN
version of {\tt TAUOLA Interface} and {\tt HepMC} event record
which is used for the C++ based interface
has to be taken into account.
In the first case the whole event was represented
by a common block containing a list of particles with their properties and
with integer
variables denoting pointers to their origins and descendants.
The {\tt HepMC} event structure
is built from vertices, each of them having pointers to their origins and descendants. Links between vertices represent particles or fields.
In both, {\tt FORTRAN} and C++
cases, the event is structured as
a tree\footnote{At least in principle, because in practice its properties
may be rather of the graph of not universally defined properties.
This makes our task challenging.}, the necessary algorithms are analogous,
but nonetheless different.
In {\tt HepMC version 2.04}, an event is represented by a {\tt GenEvent} object,
which contains all information regarding itself, including event id,
units used for dimensional quantities in the event and the list of produced particles. The particles
themselves are grouped into {\tt GenVertex} objects allowing access to mother
and daughter particles of a single decay. Vertices provide an easy way
to point to the whole branch in a decay tree that needs to be accessed,
modified or deleted if needed. The information of a particle itself is stored
in a {\tt GenParticle} object containing the particle id, status and momentum
as well as information needed to locate its position in the decay tree.
This approach allows traversing the event record structure in several different
ways.
The {\tt HepMC} event record format is evolving with time, making it necessary
to adapt
the code to the new versions.
{\tt HepMC} version 2.05 is used as a reference. In the case of version 2.03 restrictions on methods for
units
conversion have to be taken into account, for details see Appendix~\ref{subs:Issues}.
One should keep in mind that future adaptations to {\tt HepMC} changes may restrain
backward compatibility.
Evolution of the {\tt HepMC} format itself is not a crucial problem.
In contrary, conventions how physics information is filled into {\tt HepMC}
represent the source of main technical and also physics
challenge for our interface.
This is quite similar to the previous
{\tt HEPEVT - FORTRAN} case. Let us discuss this point in more detail now.
\subsubsection{Event Record Structure Scenarios}
While many Monte-Carlo generators (eg. {\tt PYTHIA 8.1} \cite{Sjostrand:2007gs},
HERWIG++ \cite{Bahr:2008pv})
store events in {\tt HepMC} format, the representations of
these events are not subject to strict standards, which can therefore
vary between Monte-Carlo generators or even physics processes. Some examples
of these variations include the conventions of status codes, the way
documentary information on the event is added, the direction of pointers at a vertex
and the conservation (or lack of conservation) of energy-momentum at a vertex.
Below is a list of scenarios we have observed in Monte-Carlo
generators used for testing the code.
This list will serve as a declaration for convention of {\tt HepMC} filling, which the
interface should be able to interpret correctly.
\begin{itemize}
\item \textbf{4-momentum conservation} is assumed for all vertices in the event record.
\item \textbf{Status codes:} only information whether given particle is incoming, outgoing or intermediate will be used,
\item \textbf{Pointers at a vertex} are assumed to be bi-directional. That is, it is possible to traverse the record structure from mother to daughter and from daughter to mother along the same path.
\end{itemize}
\noindent
\textbf{ Extensions/Exceptions} to this specifications are handled in some cases. We will call them
options for conventions of event record filling.
\begin{itemize}
\item Vertices like $\tau^\pm \rightarrow \tau^\pm$ and $\tau^\mp \rightarrow \tau^\mp \gamma$
where 4-momentum conservation is not preserved, but this
non-conservation is balanced, for example, between the two branches
outgoing from $Z$.
\item Lines representing
intermediate bosons may be missing. In fact this may be unavoidable, if several
diagrams contribute simultaneously. In that case, our algorithm makes a choice
based on an approximation that
only the dominant single diagram is considered
and an intermediate boson state is defined accordingly on the fly. Other possible treatments:
statistical choice of the dominant process, or calculations based on higher order
matrix elements for the hard process, are not available at present.
\item
As in the FORTRAN cases, we expect that new types of
conventions for filling the event record
will appear, because of physics motivated requirements.
Unfortunately, the resulting options do not always guarantee
an algebraically closed structure.
Host program specific patches may need to be defined for the
{\tt TAUOLA interface}.
Debugging it could be time consuming, and will need to be repeated for every new
case.
\end{itemize}
Detailed
conventions for the actual filling of physics information into {\tt HepMC} format
is defined by authors of each Monte Carlo program.
In future, an important special case of event records filling with
information extracted from experimentally observed event (eg. $Z\to \mu^+\mu^-$
modified later to $Z\to \tau^+\tau^-$) should be allowed.
Obviously, a new type (or types) of {\tt HepMC} filling will then appear.
\section{Design}
The structure of our code is documented using Doxygen standards \cite{Doxygen} and is
presently available from the project web page \cite{tauolaC++}. The
source code for this web page is also available in our package
distribution. Doxygen documentation can be thus compiled on a users
platform, and hence provide documentation which matches the actual version of the
distribution.
Let us present here briefly the directory structure and list the main classes
with a short description of their functionality.
\subsection{Interface Structure and Responsibilities}
The choice of splitting the source code into three main modules,
see Fig.~\ref{fig:design} (blue part),
allows to separate the FORTRAN related code from the abstract C++ interface
and the concrete implementation of the interface created for the appropriate
event record.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.6]{tauola_interface_design.eps}
\label{fig:design}
\caption{{\tt TAUOLA C++ Interface} class relation diagram}
\end{figure}
\begin{itemize}
\item {\bf Tauola Fortran Interface}\\
This part of the code provides an interface
to the FORTRAN library of {\tt TAUOLA}. In particular, routines
necessary for library initialization and wrapper for routine invoking the decay of a single
$\tau$. Parts of the interface code are still left in FORTRAN, but can be rather easily
rewritten to C++.
The most important method, {\tt filhep\_},
is already implemented in C++. Its FORTRAN predecessor writes single particles to
the {\tt HEPEVT} common block. At present the method {\tt filhep\_} inserts the particle
into the {\tt HepMC} event record but remains to be called from the FORTRAN library.
For further details see Appendix \ref{Interface to TAUOLA}.
\item {\bf Tauola C++ Interface} \\
The abstract part of the interface to the event record.
The class {\tt TauolaEvent} contains information regarding the whole event
structure, while {\tt TauolaParticle} stores all information regarding a single particle.
All particles used by the interface are located in the event in the form of
a list of {\tt TauolaParticle} objects.
The last class located here, {\tt TauolaParticlePair}, is the core of all polarization
and decay algorithms. They are independent from the event record used by the interface
as they operate on these two abstract classes presented above.
\item {\bf Event Record Interface} \\
The event record implementation classes. All classes stored here represent
the implementation of specific event record interfaces and are responsible for reading,
traversing and writing to the event record structure.
Only {\tt TauolaEvent} and {\tt TauolaParticle} classes must be implemented.
The {\tt HepMC} event record interface is implemented
through {\tt TauolaHepMCEvent} and {\tt TauolaHepMCParticle}.
\end{itemize}
\subsection{Directory Structure}
\begin{itemize}
\item {\bf src/eventRecordInterfaces/ } - source code for classes which interface with {\tt HepMC}.\\
Classes:
\begin{itemize}
\item { \bf TauolaHepMCEvent} - Interface to HepMC::GenEvent objects.
\item { \bf TauolaHepMCParticle} - Interface to HepMC::GenParticle objects.
\end{itemize}
\item {\bf src/tauolaCInterfaces/ } - source code for general {\tt TAUOLA Interface} classes, such as those
responsible for spin correlations and boosting. \\
Classes:
\begin{itemize}
\item { \bf DecayList } - Storage class for keeping track of TauolaParticles and their indices
\item { \bf Tauola } - Controls the configuration and initialization of {\tt TAUOLA FORTRAN}
\item { \bf TauolaEvent } - Abstract base class for event information.
\item { \bf TauolaParticle } - Abstract base class for particles in the event. This class also handles particle boosting.
\item { \bf TauolaParticlePair } - Contains two objects of type TauolaParticle that are related by the same mother. Spin correlations and other minor algorithms are handled here.
\end{itemize}
\item {\bf src/tauolaFortranInterfaces/ } - interface to {\tt TAUOLA FORTRAN} routines and common blocks. \\
Files:
\begin{itemize}
\item { \bf f\_Decay} - contains a wrapper for the {\tt TAUOLA FORTRAN} routine for decaying
$\tau$'s (DEKAY).
\item { \bf f\_FilHep } - provides a method which {\tt TAUOLA FORTRAN} calls to fill a $\tau$ decay product into the
event record.
\item { \bf f\_Init } - contains a wrapper for the {\tt TAUOLA FORTRAN} routines for tauola initialization.
\item { \bf f\_Variables } - contains definitions of {\tt TAUOLA FORTRAN} routines and common blocks used by other methods in tauolaFortranInterfaces.
\item { \bf tauola\_extras.f } - contains extra FORTRAN routines (taken from the {\tt TAUOLA Interface} in {\tt FORTRAN}) which should ultimately be migrated to C++.
\end{itemize}
\item {\bf src/utilities/ } - source code for utilities that help in debugging and plotting distributions.\\
Classes:
\begin{itemize}
\item { \bf Log} - General purpose logging class that allows filtering out output messages
of {\tt TAUOLA C++ Interface} and keeps statistics regarding tauola run.
\item { \bf Plot} - Simple class that gathers data for some useful debug plots.
\end{itemize}
\item {\bf examples/ } - examples of different {\tt TAUOLA C++ Interface} uses
\begin{itemize}
\item {\bf taumain\_stand\_alone\_example} - stand alone example with a simple
$e^+e^- \rightarrow \tau^+\tau^-$ event in {\tt HepMC} format
and then $\tau$'s decayed by {\tt TAUOLA}.
\item {\bf single\_tau\_gun\_example } - example of {\tt TAUOLA} linked with pythia 8.1 and used to decay single $\tau$ selected from the event record.
\item {\bf taumain\_pythia\_example } - example of {\tt TAUOLA} linked with pythia 8.1, and decay chain analysed with {\tt MC-TESTER}.
\end{itemize}
\item {\bf SANC/} - code for the computation of electroweak corrections.
\item {\bf include/} - directory for the header files.
\item {\bf lib/ } - directory for the compiled libraries.
\item {\bf documentation/ } - contains doxygen documentation and this latex file.
\item {\bf tauola-fortran/ } standard {\tt TAUOLA FORTRAN} distribution
exactly as described in \cite{Golonka:2003xt}. It is kept intact and
is prepared for future updates, see
ref.~\cite{Actis:2009gg} for details of that project.
\end{itemize}
\subsection{Algorithm Outline}
\label{sect:Outline}
An overview of the algorithm for the {\tt TAUOLA Universal Interface} is
given below, for more detail the reader
should refer to the project's Doxygen documentation ~\cite{tauolaC++}.
Documentation of the {\tt TAUOLA FORTRAN Interface}
~\cite{Golonka:2003xt} describes some aspects of the spin
correlation algorithm which are also relevant to this interface.
The first step is creation of a {\tt TauolaHepMCEvent} object from
a {\tt HepMC::GenEvent} event record. At this step the units for dimensional
quantities (four-momenta, masses, etc.) are checked, and if needed the {\tt HepMC} event record is
reset to use GEV and MM ensuring proper execution of $\tau$ decay library.
After a {\tt TauolaHepMCEvent} is created the
{\tt decayTaus()} method should be executed by the user's code\footnote{Prior to this step
the user may want to execute {\tt Tauola::decayOne(...)} for $\tau$ leptons, where {\tt TAUOLA Universal Interface} is
expected not to work properly. For details see Appendix~\ref{TAUOLAgun}.
},
invoking the following process:
\begin{enumerate}
\item The {\tt HepMC} event record is traversed and a list of all stable $\tau$'s in the event is created.
\item From each found $\tau$ location, the tree is traversed backwards so
that information about the production process can be extracted
and used for the calculation of the spin density matrix.
\item The siblings of the $\tau$ are identified through common
parents, i.e. requiring that they are produced at the same {\tt HepMC} vertex. In
cases such as $\tau \rightarrow \gamma \tau$, the parent(s) are defined
as the particle(s) which produced the first $\tau$;
$\tau$ and $\nu_\tau$ siblings are paired to the $\tau$.
\item The density matrix is set-up using information about the
$\tau$-pair and their parent type (for $Z/\gamma$ processes,
grandparent information is also required). This is described in
detail in Sec. \ref{dwaTrzy}. The density matrix assumes
a center-of-mass frame for the $\tau$-pair, with the $\tau$'s and their
grandparents orientated as shown in Figure \ref{fig:designKB}.
\item The pair is then decayed by executing the {\tt DEKAY} routine
for each $\tau$ in the pair. The {\tt DEKAY} routine
is stored in the {\tt tauola-fortran} directory, for details see Appendix
\ref{subsubsec:routines}.
\item A spin weight is calculated using the polarimetric vectors
returned from {\tt TAUOLA FORTRAN} and the density matrix previously set-up (described in
Sec. \ref{dwaTrzy}).
\item If the decays are rejected, the pair is decayed anew and the
process is repeated until the decays are accepted. In this way unweighting of spin
effects is performed.
\item Once accepted, the decay products are added into the event record with the procedure
as follows:
\begin{enumerate}
\item As the density matrix is only valid in the special reference
frame of Figure \ref{fig:designKB}, the $\tau$-pair are boosted and rotated into this hard process frame.
\item The DEKAY routine of {\tt TAUOLA FORTRAN} is executed with state = 11 or 12
(write). This initiates {\tt TAUOLA FORTRAN} to return the daughter
information via the {\tt filhep\_} routine (see Section \ref{subsubsec:routines}).
\item The $\tau$'s status code is changed from ``1'' (stable particle)
to ``2'' (intermediate particle).
\item A new object HepMC::GenParticle is created for each daughter and the
appropriate tree structure is created and added into the event.
\item Each daughter is boosted using the $\tau$'s 4-momentum
(as {\tt TAUOLA} constructs a decay for a $\tau$ at rest) to the hard process frame.
\item The $\tau$'s and their decay products are boosted back into the
laboratory frame.
\end{enumerate}
\item As the final step, the position of vertices containing the $\tau$'s and their decay products
is set according to the $\tau$'s momentum and lifetime.
\end{enumerate}
The underlying HepMC::GenEvent is hence modified with insertion of $\tau$ decay products.
All that remains is to convert the event back to its initial units, which is done
via the {\tt eventEndgame() } routine of the {\tt TauolaHepMCEvent} class.
\section{Calculation of Spin Correlations}
\label{dwaTrzy}
If more than one $\tau$ lepton is present in a final state, then not only is the individual
spin state for each $\tau$
necessary for proper generation, but the complete correlation matrix
of all $\tau$ leptons must be taken into account as well. In the case
of $\tau$-pair production, the standard algorithm explained
in \cite{Jadach:1990mz,Jadach:1993hs}
can be used without much modification.
For the single $\tau$ produced in a $\tau,\;\nu_\tau$ pair,
it is convenient to use the same algorithm as well,
even though it is not necessary from the physics point of view.
Let us describe now the algorithm given in
Refs.~\cite{Jadach:1990mz,Jadach:1993hs}. We will
use definitions and notations from that papers as well.
Spin correlations and spin polarization effects can be simulated by accepting or rejecting
a pair of generated $\tau$ decays based on a weighting factor {\tt wt}.
\begin{equation}
wt=\displaystyle \frac{1}{4} \sum_{i,j=0}^{4} h^1_i h^2_j R_{ij}
\label{weight}
\end{equation}
where $h^1$ and $h^2$ are the polarimetric vectors
for the $\tau^+$ and $\tau^-$ respectively and $R_{ij}$ is the
density matrix associated with the $\tau$ production vertex.
The matrix $R_{ij}$ depends on the mechanism
and particular kinematical configuration of
$\tau$ pair production. The $h^1_i$, $h^2_j$ depend on the respective
decays of $\tau^+$ and $\tau^-$.
The solution can be used for $\tau-\nu_{\tau}$ production as well.
In this case $\nu_{\tau}$ decay is not performed and its polarimetric vector is set to $h=(2,0,0,0)$.
A pair of $\tau$ decays should be accepted if the weight is greater than
a randomly generated number between 0 and 1.
If this condition fails, the $\tau$ pair decays should either be rejected and
regenerated, or rotated\footnote{Rotation instead of rejection increases
efficiency by a factor of 4. This however only affects the generation of $\tau$
lepton decays and represents a small fraction of the total time of
constructing the event.},
and the weight recalculated. The production process does not need to be
reprocessed.
The density matrices, $R_{ij}$, for the most standard
processes of $\tau$-pair production, are documented below.
The following frame
convention\footnote{Fig.~\ref{fig:designKB} illustrates our choice too.
There however the reaction frame is rotated by angle $\theta$ around x axis.}
was adopted:
\begin{itemize}
\item The $\tau$-pair's center of mass system is used
\item The $\tau^{+}$ (if present) lies along the positive z axis
\item The $\tau^{-}$ (if present) lies along the negative z axis
\item The incoming beams (if present) lie in the z-y plane.
\item If applicable, the incoming antiquark (or antilepton) y momentum component is positive.
\end{itemize}
$h$ is defined such that $h_0$=1 and $h_{1,2,3}$ form the polarimetric vertex
returned from {\tt TAUOLA FORTRAN} (see DEKAY in Appendix \ref{subsubsec:routines}).
The $h$ is defined in the rest frame of the $\tau$ it
was calculated for. One should stress that formally speaking $R_{ij}$
does not represent a Lorentz invariant object. Its first index is defined in the
rest frame of the $\tau^+$, whereas the second index
is in the frame of the $\tau^-$.
In the following subsections we will list the form of $R_{ij}$ for the most
commonly used processes of $\tau$-pair (or $\tau$, $\nu$) production.
\subsection{Form of $R_{ij}$ for Standard Processes}
\subsubsection{$Z/\gamma \rightarrow \tau^+ \tau^-$}
\label{subsubsection:z/gamma}
\[R=\begin{pmatrix}
1 & 0 & 0 & 2P_z(cos\theta)-1 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
2P_z(cos\theta)-1 & 0 & 0 & 1
\end{pmatrix}\]
where $P_z$ is calculated from the square of the matrix elements of the
Born-level $2 \rightarrow 2$ process $f \bar{f} \rightarrow \tau^+ \tau^-$.
\[P_z(s,\theta) = \frac{\frac{d\sigma(s,\theta,+,+)}{d\Omega}}{\frac{d\sigma(s,\theta,+,+)}{d\Omega}+\frac{d\sigma(s,\theta,-,-)}{d\Omega}} \]
$\theta$ is the angle between the incoming antiparticle beam and outgoing $\tau^+$. If the incoming beam
cannot be reconstructed from the event record the average of $P_z$ should be used
(which is equal to $P_z(cos\theta=0)$). The ``$+$'' denotes that
the spin states of $\tau^+$ and $\tau^-$ are parallel to the $\tau^+$ momentum.
For ``$-$'' it is placed in the opposite direction.
The spin correlation matrix explained above is approximate and is valid only for
longitudinal spin effects. The object $R_{i,j}$ is nonetheless prepared to
host complete spin effects. For that purpose information available from the module based on {\tt SANC}
can be used, see Appendix \ref{section:SANC}.
Further advantage is that genuine weak effects can be calculated and
included as a weight\footnote{In a similar way one can implement effects of new physics,
such as Z'
into the program. With the help of our interface effects of weak corrections
on the cross section, and not only on polarization can be installed with additional
weights.} not only on polarization but on the cross section as well.
This was found to be numerically important \cite{Adam:2008pc,Adam:2008ge} for final states
of virtuality largely surpassing the $Z$ mass and should be taken into account prior to the implementation of new physics effects.
In the formulas above, the hard process kinematical variables $s$ and $\theta$ have to be known
for each event. Those variables, together with the flavour of the incoming beam are used by a module for calculating
electroweak corrections or the function $P_z$.\footnote{
The principle behind our
solution, is quite similar to the one used in {\tt PHOTOS} Monte Carlo
where it is was shown \cite{Golonka:2006tw} to be valid up to NLO (QED) precision level.
It relies on factorization properties of fully differential distributions into the appropriately chosen
Born level ones and emission factors.
To achieve such precision in the case of spin correlations in proton-proton
collisions, rather challenging work on QCD matrix elements would be necessary.
At present,
our predictions will not be better than LL {\it on spin effects}. It is known \cite{Kleiss:1990jv},
that the solution can not be constructed beyond NLO.}
To apply the method we need to identify the four momenta of the $\tau^+$ and $\tau^-$ pair first. In the rest frame of the pair the two effective partons leading to the hard process are not necessarily back to back. Two scattering angles $\theta_1$ and $\theta_2$ can be thus reconstructed. The angle $\theta_1$ is between the $\tau^+$ and the first incoming state,
$\theta_2$ is between $\tau^-$ and the second one\footnote{We choose the first incoming state to be antiparticle.
}. Both angles are
calculated in the rest frame of the $\tau$ pair.
The average angle $\theta^\bullet$ accordingly to the description given in \cite{Was:1989ce} is taken:
$\cos\theta^\bullet=\frac{\sin\theta_1\cos\theta_2+\sin\theta_2\cos\theta_1}{\sin\theta_1+\sin\theta_2}$.
If events originate from generator such as {\tt PYTHIA},
the flavour of incoming partons is explicitly given or it
can be calculated using information encoded in the event record.
In the generic case this information is not available, and one will have to rely on measured structure
functions and statistical choice. This is the method, for example, to be applied in the analysis where one
reconstructs
from experimental data decays $Z\to \mu^+ \mu^-$ (or $W \to \mu \nu_\mu$)
and then replaces $\mu$ by the
Monte Carlo generated
$\tau$ object.
The resulting
uncertainty on one hand will not be large; angular polarization dependence of $d$ and $u$ quarks
couplings to a $Z$ is not that different. On the other hand, a mismatch between choosing quark
and antiquark may have larger effect.
The density matrix presented above features only longitudinal spin correlations.
Once the matrix $R$ is replaced with the one featuring complete spin effects,
see Appendix \ref{section:SANC}, this is
passed to the results of simulation as well, without any need of further changes.
\subsubsection{$H^0 \rightarrow \tau^+ \tau^-$ and $A^0 \rightarrow \tau^+ \tau^-$ and mixed $A^0/H^0 \rightarrow \tau^+ \tau^-$}
The complete density matrix for a scalar neutral Higgs boson $H^0$ is rather simple,
\[R=\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 \\
0 & 0 & 0 & -1
\end{pmatrix}\]
it is also true for a pseudoscalar neutral Higgs boson $A^0$
\[R=\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & -1 & 0 & 0 \\
0 & 0 & -1 & 0 \\
0 & 0 & 0 & -1
\end{pmatrix}\]
A mixed $A^0/H^0 \rightarrow \tau^+ \tau^-$ represents only a slightly more complicated case:
\[R=\begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & \frac{(\beta cos\phi)^2 - sin\phi^2 }{(\beta cos\phi)^2+sin\phi^2} & -\frac{2\beta cos\phi sin\phi}{(\beta cos\phi)^2+sin\phi^2} & 0 \\
0 & \frac{2\beta cos\phi sin\phi}{(\beta cos\phi)^2+sin\phi^2} & \frac{(\beta cos\phi)^2 - sin\phi^2 }{(\beta cos\phi)^2+sin\phi^2} & 0 \\
0 & 0 & 0 & -1
\end{pmatrix}\]
where $\beta = \sqrt{1-(\frac{2M_{\tau}}{M_{A^0/H^0}})^2}$
and $\phi$ is the scalar-pseudoscalar mixing angle.
\subsubsection{$W^{\pm} \rightarrow \tau^{\pm} \nu$}
\label{subsubsection:W}
For $W$ the matrix $R_{ij}$ takes the following form:
\[R=\begin{pmatrix}
1 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0
\end{pmatrix}\]
\subsubsection{$H^{\pm} \rightarrow \tau^{\pm} \nu$}
For charged Higgs decay the matrix $R_{ij}$ differs from the $W$ case by signs only:
\[R=\begin{pmatrix}
1 & 0 & 0 & -1 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
-1 & 0 & 0 & 0
\end{pmatrix}\]
\subsection{Cases of Partly defined Hard Processes}
\subsubsection{$\tau^+ \tau^-$ Pair with Multiple Parents and Sisters}
If a $\tau^+ \tau^-$ pair is found with multiple parents rather
than a single parent, the parent type is assumed to be a
$Z/\gamma$ with the 4-momentum reconstructed from
the 4-momentum of the $\tau$ pair. The density matrix
\ref{subsubsection:z/gamma} is used. For more details on the construction
of effective incoming beams see Section~\ref{dwaCztery}.
\subsubsection{$\tau \nu_{\tau}$ Pair with Multiple Parents and Sisters}
If a $\tau^+ \nu$ pair is found with multiple parents rather
than a single parent, the parent type is assumed to be a
$W^{\pm}$ with the 4-momentum reconstructed from
the 4-momentum of the $\tau\; \nu_\tau$ pair. The density matrix
\ref{subsubsection:W} is used.
\subsubsection{Single $\tau$ and Multiple Unpaired $\tau$'s}
By default a single $\tau$, with no $\tau$ or $\nu_{\tau}$ pairing, is treated
as unpolarized and spin effects are ignored. However, in such a case the {\tt TAUOLA decayOne} method,
see Appendix~\ref{TAUOLAgun}, can be used if including spin effects is requested. This method can be applied
by the user, in such a case the spin state of a single $\tau$ can be imposed. The method can be used even for
multiple $\tau$ final state and exact spin correlations.
In those cases a user defined quantization frame for each $\tau$ may be necessary. For each $\tau$ a distinct routine for boosting from its rest-frame to the lab-frame
may be necessary. The appropriate method is explained in the Appendix mentioned above.
Another user defined option (which may become in the future a default part of the {\tt TAUOLA Interface}) is when pairing can not be done on the basis of inspecting $\tau$
mother(s) but
can be performed
according to the closeness of the reconstructed invariant
masses of the pairs
to the masses of $W's$ or $Z's$ (if the appropriate Standard Model processes are under considerations). In general, as such configurations will often
appear for processes of new physics, again a user defined and hard process dependent solutions based on the {\tt TAUOLA decayOne} method,
might be the only option.
\subsection{Quantum Entanglement and Helicity States}
One of the convenient feature of Monte Carlo simulation is the availability of
variables used in the generation of hard processes. However, it is often not possible to define
in an exact manner what is the energy transfer, eg. in the Z propagator, if several
diagrams contribute simultaneously. Nonetheless use of such
information is tempting as it would be helpful to validate algorithms used in experimental analysis
for defining observables aimed
at the measurement of the Z lineshape.
Another example of such useful variables are the helicities of $\tau^+$ and $\tau^-$.
Even though, it is possible to attribute such variables only in the ultrarelativistic limit
and its use can be restricted to a downgraded physics approximation only (quantum entanglement
\cite{Einstin:1935rr} ignored), helicity was offering invaluable help at LEP time,
for the measurement of $\tau$ polarization \cite{Heister:2001uh}.
Our program provides helicity states of $\tau^+$ and $\tau^-$ even in cases
when exact spin effects are taken into account in the generation, as is the case of processes mediated by
intermediate $Z/\gamma^*$ state or in Higgs boson decay.
We just attribute helicity states for the $\tau$ lepton after decays
are already generated and accepted.
The approximation used in calculation of these helicities is explained in Section~\ref{Exact}
and technical aspects of our solution
are given in Appendix~\ref{subsection:eHelicity-ew}. This information
can be used for solutions similar to the ones in \cite{Heister:2001uh},
but this time for LHC purposes
\subsection{Handling of Events with Bremsstrahlung or Parton Shower Activity}
\label{dwaCztery}
Obviously, there are cases, when spin correlations calculated from Born level
processes can not be applied directly. Good examples are:
(a) $Z \to \tau^+ \tau^- \gamma$,
(b) $f1 + f2 \to Z + X$, $Z \to \tau^+ \tau^-$ where the intermediate
state $Z$ is explicitly stored,
(c) $f1 + f2 \to \tau^+ \tau^- X$ or $f1 + f2 \to \tau^+ \nu_\tau X$ where the $Z$ state is not
available. Here $X$ represents parton shower and/or final state bremsstrahlung.
The fist step, necessary eg. for calculation of the spin correlation
matrix, is to reconstruct effective Born level variables $s$ and $\theta$ and the incoming state flavors; the arguments of
the function $P_Z$. This is equivalent to the construction of effective incoming and outgoing $\tau$ fields.
Let us discuss now our cases:
\begin{itemize}
\item[a)]
An additional photon is added to the $\tau$ with which it forms
smaller virtuality. For construction of the transformation between the laboratory frame and rest frame
of $\tau$'s the effective state $Z-\gamma$ frame is used instead of intermediate $Z$.
This choice is motivated by inspection of properties of spin amplitudes.
In such a frame the effective incoming states f1 f2 will not necessarily be back to back. The average of the two directions ($\theta^*$, $\theta^\bullet$ as explained earlier in this subsection)
can be used. The virtuality of the $Z$ is nonetheless used in the effective Born calculation.
\item[b)] Additional fields $X$ representing parton showers should be subtracted from $f1$ or $f2$,
preferably from the one with which it forms smaller virtuality.
This is motivated by inspection of the spin amplitudes.
Once the effective incoming states are constructed, the definition of
boosting routines is straightforward.
\item[c)] This case is a combination of the above two. Additional fields should be subtracted from f1 f2 or considered
as originating from the intermediate $Z$ reconstructed on flight together with
$\tau^+$, $\tau^-$. The minimalization of virtuality should be used as a guide whenever combinatorial
choice has to be made. However electromagnetic
charge or colour charge should not be neglected. Obviously photons should not be combined with neutrinos nor gluons with leptons.
\end{itemize}
The approach presented above is explained in ref \cite{Golonka:2003xt} in more detail. The principle
is based on the simplest factorization properties of SM/QCD matrix elements.
In particular the assumption is made that photon(s) can be treated as (nearly) collinear with one of the final state
$\tau$ leptons. Then the kinematics can be built on the
$\tau^+ \tau^-$ pair rest frame and the effective incoming states. The
z axis is taken along effective
antiparticle incoming state and y axis is of half plane including $\tau^-$.
The second effective beam is placed then in zy plane as well, even though it is not back to back
with the first one.
For calculation of $R_{ij}$ the variable $s$ should be calculated from effective incoming states.
The scattering angle can be calculated in the $Z$ frame using formula for $\theta^*$ or $\theta^\bullet$, see ref. \cite{Was:1989ce}.
In the collinear limit for photon emissions, Lorentz transformation between the $\tau^+$, $\tau^-$ and $Z$ rest
frames is reduced to a simple boost along the $\tau^+-\tau^-$ flight directions in the $Z$ rest-frame.
Further improvements with respect to that description
require explicit use of higher order matrix elements. The described approximation is already quite good
and works up to $\alpha_{QED} / \pi \simeq$ 0.1 - 0.2 \% precision level,
for observables where it is not requested explicitly that high $p_T$ photons are present.
Situation with initial state parton shower emissions is similar, however in this case
the omitted terms for $R$ calculation may be of the order of $\alpha_{QCD} / \pi$. Thus significantly larger,
but still at the level of ten percent or so.
If the event record under
study is originating from experimental data, the flavours of the incoming partons
can not be known from the event record itself.
Instead, information in PDFs can be used to attribute such flavours on a
statistical basis and then used in the calculation of the matrix $R_{ij}$.
\subsection{Exact Spin Effects and Helicity States}
\label{Exact}
Earlier in this Section we have listed some examples of spin polarization,
spin correlation and density matrices. Those matrices can be easily
changed/replaced by the externally calculated ones
for the sake of studies on new physics phenomena, like
a consequence of certain types of spin correlations on the signal/background
separation. In that way, effects of
new physics or ad hoc modification of spin effects component by component,
can be easily included.
For calculating complete spin correlations, including the effects
of genuine weak corrections in the case of the single boson mechanism of $\tau$-pair production,
another solution
is also prepared, see Section \ref{Sect:electroweak}.
It works in the case of an intermediate $\gamma^*/Z$ state
produced in the annihilation of a pair of quarks. This can serve as an example for other processes
which can be implemented in a similar manner.
Let us stress that in the case of including complete spin effects
it is not possible to attribute helicity states to the produced $\tau$ leptons.
This can be done only in an approximate way. For that purpose,
a modified formula (\ref{weight}) can be used.
The weight
for each helicity configuration is:
\begin{equation}
weight(\;,\;)=\displaystyle \frac{1}{4} \sum_{i,j,i'j'=0}^{4} h^1_{i'} h^2_{j'} R_{ij} P^1_{ii'} P^2_{jj'}
\label{weights}
\end{equation}
The matrices $P^{1,2}$ (spin projection operators) read
\[P^{1,2}=\begin{pmatrix}
1 & 0 & 0 & \pm 1 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
\pm 1 & 0 & 0 & 1
\end{pmatrix}\]
As a consequence for {\tt Hel1=}$\pm 1$ and {\tt Hel2=}$\pm 1$
\[\begin{pmatrix}
weight(+,+)= & (h^1_{0}+ h^1_{3}) & (h^2_{0}+ h^2_{3}) & (R_{00}+R_{03}+R_{30}+R_{33}) \\
weight(+,-)= & (h^1_{0}+ h^1_{3}) & (h^2_{0}- h^2_{3}) & (R_{00}-R_{03}+R_{30}-R_{33}) \\
weight(-,+)= & (h^1_{0}- h^1_{3}) & (h^2_{0}+ h^2_{3}) & (R_{00}+R_{03}-R_{30}-R_{33}) \\
weight(-,-)= & (h^1_{0}- h^1_{3}) & (h^2_{0}- h^2_{3}) & (R_{00}-R_{03}-R_{30}+R_{33})
\end{pmatrix}\]
The actual helicities can be then attributed by unweighting the above helicity weight.
It is no omission that we are not explicit on sign conventions
for the generated helicity variables {\tt Hel1} and {\tt Hel2}.
This depend on the particular choice of boosting routine.
Our choice adopted here is, that for the $\tau$ pair produced at the $Z$
peak on average both {\tt Hel1} and {\tt Hel2} will be negative and
{\tt Hel1 = Hel2}.
\section{Electroweak Corrections and Refined Spin Effects}
\label{Sect:electroweak}
\subsection{External Calculation of the Spin Density Matrix $R_{ij}$}
Our program is equipped with methods for calculating simple spin density matrices
for most of the interesting hard processes. These methods are explained in the text of the paper,
see Section \ref{dwaTrzy}.
In some cases, notably in the case of $\tau^+ \tau^-$ produced from the annihilation of a pair of quarks,
the standard density matrices may not be sufficient for some applications. A more exact solution
is also available. Instead of a native $R_{ij}$ density matrix, an externally calculated one can be used.
The solution is based on {\tt SANC} library ~\cite{Andonov:2008ga,Andonov:2004hi}
for calculation of electroweak correction
\footnote{It may serve as an example of how other
calculations featuring heavy $Z'$ boson, for example, may be used in our interface.}.
With its help the density matrix $R_{ij}$ for $q \bar q \to \tau^+ \tau^-$ process
can be calculated as a function of the incoming state flavour and Born level variables (Mandelstam
$s$ and scattering angle $\theta$).
Additional two weights are also provided, which include
the matrix elements squared and averaged over the spin.
For additional weights genuine weak corrections are respectively switched on and off.
This may be helpful for the evaluation of genuine weak corrections for states
of large $s$, significantly above the $Z$ peak, where they become sizable.
See eg. refs. \cite{Adam:2008pc,Adam:2008ge}.
For better modularity of the interface and to speed up execution of the program, pretabulation is used.
At first, a dedicated module has to be invoked, as will be explained later.
In such dedicated runs, $R_{ij}$ is calculated and stored in a lattice of
($s$ and $\cos\theta$) points.
Later, in the actual execution of our interface,
these pretabulated values of $R_{ij}$ are interpolated to the actual phase space point.
For this purpose, the standard bilinear interpolation algorithm
is used. Additionally, in order to avoid numerical errors, for $\cos\theta$ values near -1 and 1 we're using the linear extrapolation algorithm.
Pretabulation is prepared for 3 domains of $s$: around the $Z$ peak, close to the $WW$
pair production threshold and over a broad energy range. The actual choice for
pretabulation zones is $ 85\; $GeV$ < \sqrt{s} < 110\; $GeV,
$160\; $GeV$< \sqrt{s}< 220\; $GeV\footnote{ In the case when the application featuring $Z'$ is used this
pretabulation zone should be replaced, for example by
$ M_{Z'}- 3\Gamma_{Z'}< \sqrt{s} < M_{Z'}+ 3\Gamma_{Z'}$.},
$6\; $GeV$< \sqrt{s}< 17\; $TeV. For $s$ below $36\; $GeV$^2$ the analytic form taken from
ref. \cite{Jadach:1985ac} is used.
It features all spin and mass effects, but electroweak corrections, and even
$Z$ exchange, are not taken into account. This is reasonable for $s<36$ GeV$^2$
(up to, say, 100 GeV$^2$).
The advantage of this solution is that results of {\tt SANC} library calculation can be modified by the user before it is
loaded into our interface without intervention into the code of the interface itself.
\subsection {Conventions of Frames: KORALB, SANC and TAUOLA Interface}
\label{subsection:KORALBSANC}
It is not essential for {\tt TAUOLA Interface} and the segment calculating electroweak
corrections to follow exactly the same conventions for spin quantization.
In {\tt TAUOLA Interface} we
follow the frame orientation exactly as in paper
\cite{jadach-was:1984}. The adopted frame orientation is shown in Fig.~\ref{fig:designKB}.
In {\tt SANC}, the orientation of axes is different, see Fig.~\ref{fig:designS}
In the case of our interface, the beam momenta are laid along the z axis. The anti-particle beam is parallel, particle beam is antiparallel.
The $y$ component of the $\tau^-$ is always positive.
The $\theta$ angle to be used for calculation of the density matrix is between
the directions of the antiparticle beam and the $\tau^+$.
In the case of the {\tt SANC} module, the $\tau$ momenta $p_{\tau^+}, p_{\tau^-}$ are laid on the $xz$ plane. The $xz$ plane is the reaction plane: a beam of particles (quarks or leptons) is parallel to the $z$ axis.
The $x$ component of $p_{\tau^-}$ is always negative.
The $y^{'}$ and $y^{''}$ axes of the $\tau^+$ and $\tau^-$ spin frames correspondingly have opposite direction to each other. The $y^{''}$ axis
is parallel to the $y$ axis of the hard process frame.
Appropriate rotations and other convention adjustments are performed by the program in preparation of the $R_{ij}$ tables:
{\tt SANCtable.cxx}.
\begin{figure}[h!]
\centering
\subfigure[KORALB-like orientation]{
\includegraphics[scale=0.25]{R-frame.eps}
\label{fig:designKB}
}
\subfigure[SANC module orientation]{
\includegraphics[scale=0.25]{R-frame-sancn.eps}
\label{fig:designS}
}
\caption{
The relative orientation of reference frames for the spin states
of $\tau^+$, $\tau^-$ and for hard processes as used in our interface (Fig. (a)) and in {\tt SANC} module (Fig. (b)) are shown. In Fig (a) the axes $x$ (not shown explicitly), $x'$ and $x''$
are parallel to each other
and point behind the picture, axis $z$ is parallel to the direction
of the anti-particle beam. In Fig (b) the axis $z$ is parallel to the direction of the particle beam, the axis
$y'$ points behind the picture. Axes $y$ (not shown explicitly) and $y''$ point toward the reader and
are antiparallel to $y'$.
}
\end{figure}
\subsection {Numerical Significance of Electroweak Corrections }
One may wonder whether the numerical results induced by electroweak corrections are
of any practical purpose. They are expected to be of the order of
1\% and indeed are not that large for
the intermediate state virtuality of up to 100 GeV above the $Z$ boson mass.
The situation changes however significantly at higher energies. As can be
seen from Figures \ref{fig:sigmaUp} and \ref{fig:sigmaDown} the effect may
be of the order of even 50\% at virtualities of several TeV.
This is quite in agreement with the results of refs.
\cite{Adam:2008ge,Adam:2008pc}. In Figures
\ref{fig:polUp} and \ref{fig:polDown} we collect results for
$\tau$ polarization calculated at $\cos \theta=-0.2$. Again,
the effects are small up to the energy scale of about 500 GeV. At larger scales
corrections become sizable. The electroweak corrections
should be therefore considered in studies aiming for new physics phenomena
such as $Z'\to \tau^+\tau^-$ decays.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.76]{2u2tau_sigma.eps}
\caption{\label{fig:sigmaUp}
The integrated cross section of $\tau$ pair production from up quarks
calculated with and without NLO
EW corrections (red and blue lines) is shown in the left hand side plot.
The ratio of the two distributions
is given on the right hand plot.
We are using the alpha scheme for electroweak corrections. That is why
light fermion loops contribute to the difference between the two lines.
The differences between the alpha scheme Born predictions and expressions
used in the host program
must be understood before the correcting weight (see Appendix \ref{subsection:eHelicity-ew}) is used.
}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.76]{2d2tau_sigma.eps}
\caption{\label{fig:sigmaDown}
The integrated cross section of $\tau$ pair production from down quarks
calculated with and without NLO
EW corrections (red and blue lines) is shown in the left hand side plot.
The ratio of the two distributions
is given on the right hand plot.
We are using the alpha scheme for electroweak corrections. That is why
light fermion loops contribute to the difference between the two lines.
The differences between the alpha scheme Born predictions and expressions
used in host program
must be understood before the correcting weight (see Appendix \ref{subsection:eHelicity-ew}) is used.
}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[Up quarks]{
\includegraphics[scale=0.35]{ew_spin_up_cos0.eps}
\label{fig:polUp}
}
\subfigure[Down quarks]{
\includegraphics[scale=0.35]{ew_spin_down_cos0.eps}
\label{fig:polDown}
}
\caption{ Polarization for $\tau$ leptons produced from up quarks
(Fig. (a)) and down quarks (Fig. (b)) at $\cos\theta=-0.2$. The red
line is with electroweak corrections, the black is Standard Born as
is default in the interface. The blue line is Born according to
alpha scheme. The main purpose of these results is a technical test
of the software installation. Note however the inadequateness of the alpha
scheme Born, which is significantly different from the other two
results even at relatively low energies. The small bump on the red
line on Fig. (a) is due to the WW threshold. It is insignificant for
positive $\cos\theta$. }
\end{figure}
\section{Tests of Spin Correlations and Numerical Results}
There are two purposes of the presented results in this section. On
one hand these results complement the technical tests described in
Appendix ~\ref{sect:elem} with the ones oriented toward a particular
hard processes. The technical tests should be repeated for every new
program installation or configuration. On the other hand, results of
the present section are of potential physics interest as well. They
illustrate the dominant spin effects on idealized distributions for LHC
measurements.
\label{sec:results}
Tests presented here
were conducted using {\tt MC-TESTER}\cite{Golonka:2002rz,Davidson:2008ma}.
{\tt MC-TESTER} allows semi-automated comparisons of invariant mass
distributions of each sub-group of eg. $\tau$ or $Z$ stable decay products.
The results of these tests were also compared to the
results obtained with the {\tt FORTRAN Interface} (which
has been well validated by comparison with analytical and numerical
calculations for $\tau$ pair production\footnote{This represents
tests of interface. In all cases $\tau$ decays are generated
with the help of {\tt TAUOLA FORTRAN}.
For a review of physics oriented tests of $\tau$ decays themselves, and
projects for future improvements based on low energy $e^+e^-$ data,
see ref. \cite{Actis:2009gg}.}).
In addition to this, we created custom {\tt MC-TESTER }
macros for plotting other spin sensitive quantities and compared these to
published results. Numerical results are presented later in the section, see Figs.~\ref{fig:H_spin_3}, \ref{fig:H_spin_4} and \ref{fig:H_spin_C}, \ref{fig:H_spin_D}.
\subsection{$Z/\gamma \rightarrow \tau^+ \tau^-$}
\label{mctester_results_z}
The longitudinal spin effects for $Z$ decay into $\tau$'s was tested
by restricting the $\tau$ decay mode to $\tau^{\pm} \rightarrow \pi^{\pm} \nu_{\tau}$
and examining the invariant mass of the $\pi^+\pi^-$ pair, $M_{\pi^+\pi^-}$
(see Fig. \ref{fig:Z_spin_1}) and the $\pi$ energy distribution in the rest frame of
the $Z$ (see Fig. \ref{fig:Z_spin_2}). The effect of $Z$ polarization on these distributions
was studied in \cite{Pierzchala:2001gc} and we obtained consistent results with the new
C++ implementation of the {\tt TAUOLA Interface}.
\begin{figure}[h!]
\centering
\subfigure[$M_{\pi^+\pi^-}$]{
\includegraphics[scale=0.35]{plot_Z_on_off_1.eps}
\label{fig:Z_spin_1}
}
\subfigure[$1-2\frac{E_{\pi^+}}{M_Z}$]{
\includegraphics[scale=0.35]{plot_Z_on_off_2.eps}
\label{fig:Z_spin_2}
}
\caption{Longitudinal spin observables for the Z boson ($e^+e^- \rightarrow Z$ at 500 GeV).
Distributions are shown for spin effects switched on (red), spin effects switched off (green), and their ratio (black).
}
\end{figure}
\subsection{$H^0/A^0 \rightarrow \tau^+ \tau^-$}
\label{subsection:HA}
As was done for $Z$ decay in Section \ref{mctester_results_z}, longitudinal spin
effects for Higgs decay into $\tau$'s was tested using $M_{\pi^+\pi^-}$ (Fig. \ref{fig:H_spin_1})
and the $\pi$ energy distribution in the rest frame of the $H^0$ (see Fig. \ref{fig:H_spin_2}),
which was flat as expected.
\begin{figure}[h!]
\centering
\subfigure[$M_{\pi^+\pi^-}$]{
\includegraphics[scale=0.35]{plot_H_on_off_1.eps}
\label{fig:H_spin_1}
}
\subfigure[$1-2\frac{E_{\pi^+}}{M_Z}$]{
\includegraphics[scale=0.35]{plot_H_on_off_2.eps}
\label{fig:H_spin_2}
}
\caption{Longitudinal spin observables for the H boson for $\tau^{\pm}\rightarrow \pi^{\pm} \nu_{\tau}$.
Distributions are shown for spin effects switched on (red), spin effects switched off (green), and their ratio (black). }
\end{figure}
Let us now turn to transverse spin correlations. In Fig. \ref{fig:H_spin_3}
the benchmark histogram as produced by our {\tt FORTRAN Interface} and given in Fig.3 of
reference \cite{Was:2002gv} is reproduced\footnote{ In the plot
the case of non zero scalar-pseudoscalar mixing was chosen. This is the origin of the difference with
ref. \cite{Was:2002gv}.}.
It features acollinearity of the $\pi^+$, $\pi^-$ pair in
the Higgs boson rest frame, both $\tau$'s decay to $\pi^\pm \nu$.
For the same decay set up, Fig. \ref{fig:H_spin_4} features acoplanarity of the planes built respectively on decay products of $\tau^+$ and $\tau^-$. The spin effect is indeed large. However, it requires
use of unobservable neutrino momenta. It is difficult or even impossible
to achieve sufficient experimental precision in reconstruction of the reaction frame necessary for this
observable. Also, the first observable presented on
Fig. \ref{fig:H_spin_3} suffers from the same limitation.
The two other tests, Figures \ref{fig:H_spin_C} and \ref{fig:H_spin_D} present distribution of acoplanarity
angle for the two planes built respectively on the momenta of $\pi^+\pi^0$ and
$\pi^-\pi^0$; the decay products of $\rho^+$ and
$\rho^-$. All in the rest frame of the $\rho$-pair. It is directly based
on measurable quantities. The $\rho^\pm$ originate respectively
from $\tau^\pm \to \nu \rho^\pm$ decays.
There is no need for Higgs rest frame reconstruction in this case. Events are divided
into two categories. If the energy difference between
charged and neutral pions coming from the two $\tau$'s are of the same sign, they
contribute to Fig. \ref{fig:H_spin_C},
otherwise they contribute to Fig. \ref{fig:H_spin_D}. For details of the
definition and for more numerical results obtained with the {\tt FORTRAN} Interface,
see \cite{Desch:2003rw}.
\begin{figure}[h!]
\centering
\subfigure[$\pi^+ \pi^-$ acollinearity distribution ($\approx \pi$)) ]{
\includegraphics[scale=0.35]{plot_H_delta.eps}
\label{fig:H_spin_3}
}
\subfigure[$\pi^+ \pi^-$ acoplanarity distribution]{
\includegraphics[scale=0.35]{plot_H_theta_pi.eps}
\label{fig:H_spin_4}
}
\caption{Transverse spin observables for the Higgs boson for $\tau^{\pm}\rightarrow \pi^{\pm} \nu_{\tau}$.
Distributions are shown for scalar Higgs (red), scalar-pseudoscalar Higgs with mixing angle $\frac{\pi}{4}$. For the definition of angles see Section \ref{subsection:HA}.}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[acoplanarity distribution ($y_1y_2>0$)]{
\includegraphics[scale=0.35]{plot_H_coplanarity-angle-C.eps}
\label{fig:H_spin_C}
}
\subfigure[acoplanarity distribution ($y_1y_2<0$)]{
\includegraphics[scale=0.35]{plot_H_coplanarity-angle-D.eps}
\label{fig:H_spin_D}
}
\caption{Transverse spin observables for the Higgs boson for
$\tau^{\pm}\rightarrow \pi^{\pm} \pi^0 \nu_{\tau}$.
Distributions are shown for scalar Higgs (red), scalar-pseudoscalar
Higgs with mixing angle $\frac{\pi}{4}$ (green).
For the definition of angles see Section \ref{subsection:HA}.}
\end{figure}
\subsection{$W^{\pm} \rightarrow \tau^{\pm} \nu_{\tau}$ and $H^{\pm} \rightarrow \tau^{\pm} \nu_{\tau}$}
\label{sect:WH}
\begin{figure}[h!]
\centering
\subfigure[$W^+ \to \nu_\tau \tau^+, \; \tau^+ \to \pi^+ \bar \nu_\tau$]{
\includegraphics[scale=0.35]{plot_W_on_off_2.eps}
\label{fig:W}
}
\subfigure[$H^+ \to \nu_\tau \tau^+, \; \tau^+ \to \pi^+ \bar \nu_\tau$]{
\includegraphics[scale=0.35]{plot_H+_on_off_2.eps}
\label{fig:H}
}\caption{
Pion energy
spectrum in the rest frame of $W$ (left hand side) and $H^+$ (right-hand side).
Spin effects included (red line) and neglected
(green line) are plotted. The variable $1-2\frac{E_{\pi^+}}{M_{W^+}}$ or
$1-2\frac{E_{\pi^+}}{M_{H^+}}$ is used respectively.}
\end{figure}
For the simplest decay mode $\tau^{\pm}\rightarrow \pi^{\pm} \nu_{\tau}$,
as was already discussed in ref.~\cite{Pierzchala:2001gc},
the pion energy spectrum
should be softer in the case of $W^\pm$ decays and harder in the case of charged Higgs decay.
This is indeed reproduced in Figs. \ref{fig:W} and \ref{fig:H}
and the spectra are reversed for the two cases.
\newpage
\section{Outlook}
Let us summarise briefly the next steps which are planned for the work on TAUOLA
Universal Interface.
Further extension of our work will be focused on generating spin states
in the production processes from quantities which can be
experimentally measured for real data events. Discussion of
systematic error for such reconstructed spin states will require discussion of QCD corrections.
At present, our interface is designed to work with the {\tt HepMC} event record.
However, following the design we tested in
{\tt MC-TESTER} \cite{Golonka:2002rz,Davidson:2008ma} it will be rather easy
to adapt to any other event structure, by writing the appropriate event record
interface.
{\tt TAUOLA Universal Interface} as well as {\tt TAUOLA} itself are expected to remain
framework-like code where the user is supposed to modify some of the parts according
to her/his particular purposes.
The segment of code for analyzing the hard process and generating spin
states is now becoming a significant component of the project and
already exceeds by far the category of peripheric methods related to
the {\tt TAUOLA Interface}. In the future it should be moved to a
separate class.
We expect that in the next few years better parametrization of
hadronic form-factors based on $\tau$ data from the Belle and BaBar collaborations
and refined models of decays will become available. Then the directory
{\tt /tauola-fortran} will be
replaced by the new version incorporating those achievements.
\vskip 5 mm
\centerline{\Large\bf Acknowledgments}
\vskip 5 mm
Useful discussions with P. Golonka during the early stage of project development and discussions
with members of the ATLAS and CMS collaborations, and the LCG group are acknowledged.
We are specially indebted to the pilot users of the interface, in particular to: Julia Yarba, Sami Lehti, Eric Torrence,
Marcin Wolter
and Anna Kaczmarska.
Partial support of Polish-French collaboration
no. 06-124 within IN2P3 through LAPP Annecy during final completion of this work is
also acknowledged.
\providecommand{\href}[2]{#2}\begingroup |
\section{Introduction}
\global\long\def\operatorname{Aut}{\operatorname{Aut}}
This note arises from a simple observation deriving from the computation
of the Hausdorff dimension $\dim X$ of a self-similar space $X$
that satsifies the open set condition and has similarity ratios $r_{1},\dots,r_{n}\in(0,1)$;
$\dim X$ is the unique positive number $s$ satisfying\[
1=r_{1}^{s}+\dots+r_{n}^{s}\]
(see \cite[Section 6.4]{Edg_MTFG2}). But this agrees with the condition that ensures
that a certain one-parameter automorphism group on the Cuntz algebra
$\mathcal{O}_{n}$ has a KMS state. Let $\alpha:\mathbb{R}\to\operatorname{Aut}(\mathcal{O}_{n})$
be defined such that $\alpha_{t}(S_{j})=e^{it\lambda_{j}}S_{j}$ for
$j=1,2,\dots,n$; then by \cite[Proposition 2.2]{Eva_80} there is
a KMS state for $\alpha$ at inverse temperature $\beta$ iff\[
1=e^{-\beta\lambda_{1}}+\dots+e^{-\beta\lambda_{n}}.\]
Setting $r_{j}=e^{-\lambda_{j}}$ for $j=1,\dots,n$ and $s=\beta$,
the two conditions coincide (see Example \ref{sub:Cuntz-algebras}
for more details). It is the purpose of this note to account for this
coincidence and explore more examples. To accomplish this we will apply
groupoid methods from \cite{Ren_LNM793} and
\cite{Ku_Re_PAMS06} (see also \cite{Ren_2009}).
In 1967 Haag, Hugenholtz and Winnink (see \cite{HaHuWi_CMP67}) discovered the
relevance of the KMS condition in the $C^*$-algebraic formulation of equilibrium states
in quantum statistical mechanics. Thereafter the KMS states of a $C^*$-algebra with
respect to a natural one-parameter automorphism group (regarded as time evolution)
have played a key role in the development of the theory. Many of the basic facts
may be found in the text by Bratteli and Robinson (see \cite{Br_Ro_OAQS2}).
Given a compact metric space $X$, a local homeomorphism $T:X\to X$ is said to satisfy
the local scaling condition (see Definition \ref{def:loc-scaling} and
Proposition \ref{rem:equivdef})
if $(x, y) \mapsto \frac{\rho(Tx,Ty)}{\rho(x,y)}$ extends to a continuous
function on $X \times X$ that is strictly positive on the diagonal.
In this case, there is a continuous function
$\varphi\in C(X,\mathbb{R})$ such that for all $x\in X$,
\[
e^{\varphi(x)} = \lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)}.
\]
This function defines a continuous real-valued cocycle $c_\varphi$ on the
Deaconu-Renault groupoid $G$ which in turn defines a one-parameter
action $\alpha$ on the associated $C^{*}(G)$.
Our main result (see Theorem \ref{thm:Main}) asserts that if $T:X\to X$ satisfies the local scaling condition,
then the Hausdorff measure may be used to define an $(\alpha, \beta)$-KMS state
where $\beta$ is the Hausdorff dimension of $X$. We next obtain condtions
for the uniqueness of the KMS state (see Proposition \ref{pro:uniqueness}).
If $\varphi$ is constant, we derive a simple equation involving the topological
entropy of $T$, the Hausdorff dimension of $X$ and the scaling constant
(see Proposition \ref{pro:Entropy}).
The remainder of the paper is devoted to applications of our results to a number
of examples. The last example, which is based on the Sierpinski octafold, is perhaps
the most interesting. Although it does not satisfy the hypotheses of our main result,
the conclusions hold. This suggests that it should be possible to weaken our hypotheses.
\section{Preliminaries}\label{sec:prelim}
Let $(X,\rho)$ be a metric space and let $s>0$. For a set $F\subset X$
and $\varepsilon>0$, a countable cover $\mathcal{A}$ of $F$ is
called an $\varepsilon$-\emph{cover} of $F$ iff \global\long\def\operatorname{diam}{\operatorname{diam}}
$\operatorname{diam} A\le\varepsilon$ for all $A\in\mathcal{A}$. Define\[
\overline{\mu}_{\varepsilon}^{s}(F)=\inf\sum_{A\in\mathcal{A}}(\operatorname{diam} A)^{s},\]
where the infimum is over all countable $\varepsilon$-covers $\mathcal{A}$
of the set $F$ \cite[Section 6.1]{Edg_MTFG2}. Then $\overline{\mu}_{\varepsilon}^{s}$
is decreasing with respect to $\varepsilon$ and\[
\overline{\mu}^{s}(F)=\lim_{\varepsilon\to0}\overline{\mu}_{\varepsilon}^{s}(F)\]
is a metric outer measure on $X$. Let $\mu^{s}$ be the Borel measure
defined by $\overline{\mu}^{s}$.
Let $s,t>0$ such that $s<t$. It is well known (\cite[Theorem 6.1.6]{Edg_MTFG2})
that if $\mu^{t}(F)>0$ then $\mu^{s}(F)=\infty$ and if $\mu^{s}(F)<\infty$
then $\mu^{t}(F)=0$. The \emph{Hausdorff dimension }$\dim F$ of
a set $F$ is the unique number $s_{0}\in[0,\infty]$ such that $\mu^{s}(F)=\infty$
for all $s<s_{0}$ and $\mu^{s}(F)=0$ for all $s>s_{0}$. If $s=\dim X$,
then we call $\mu^{s}$ the \emph{Hausdorff measure} on $X$. In the
following we assume that $0<s<\infty$ and $0<\mu^{s}(X)<\infty$.
Assume that $T:X\to X$ is a local homeomorphism. The \emph{Deaconu-Renault}
groupoid (see \cite{Ren_CLA00}, \cite{Dea_TAMS95}) is defined via\[
G:=\{(x,m-n,y)\,:\, T^{m}(x)=T^{n}(y)\}\subset X\times\mathbb{Z}\times X.\]
Two triples $(x_{1},n_{1},y_{1})$ and $(x_{2},n_{2},y_{2})$ are
composable if and only if $y_{1}=x_{2}$, and, in this case,
$(x_{1},n_{1},y_{1})(y_{1},n_{2},y_{2})=(x_{1},n_{1}+n_{2},y_{2})$.
The inverse of $(x,n,y)$ is $(y,-n,x)$. A basis of topology for
$G$ is given by the sets\[
Z(U,V,k,l)=\{(x,k-l,y)\in G\,:\, x\in U,y\in V\},\]
where $U$ and $V$ are open subsets of $X$ such that $T^{k}|_{U}$
and $T^{l}|_{V}$ are homeomorphisms onto their images and $T^{k}(U)=T^{l}(V)$.
The groupoid $G$ is \'etale so it admits a Haar system consisting
of counting measures.
Let $A$ be a $C^{*}$-algebra, let $\alpha_{t}:\mathbb{R}\to\operatorname{Aut}(A)$
be a strongly continuous action, and let $\beta\in\mathbb{R}$. Recall
(\cite[Section 5.3]{Br_Ro_OAQS2}) that a state $\varphi$ on $A$
is said to be an $(\alpha,\beta)$-KMS state if\[
\varphi(b\alpha_{i\beta}(a))=\varphi(ab)\]
for all $a,b$ entire for $\alpha$. If $\beta=0$, then $\varphi$
is an $\alpha$-invariant tracial state. The parameter $\beta$ is
called the \emph{inverse temperature}.
Let $G$ be a locally compact \'etale groupoid and let $c$ be a
real-valued continuous cocycle. Then $c$ defines a one-parameter
automorphism group $\alpha^{c}$ of $C^{*}(G)$ via\[
\alpha_{t}^{c}(f)(\gamma)=e^{itc(\gamma)}f(\gamma),\]
for all $t\in\mathbb{R}$, $\gamma\in G$, and $f\in C_{c}(G)$ (\cite[Section II.5]{Ren_LNM793}).
Each probability measure $\mu$ on $G^{(0)}$ defines a state $\omega_{\mu}$
on $C^{*}(G)$ via\begin{equation}
\omega_{\mu}(f)=\int_{G^{(0)}}E(f)d\mu,\label{eq:KMSstate}\end{equation}
for all $f\in C^{*}(G)$, where $E$ is the canonical expectation
onto $C_{0}(G^{(0)})$. A probability measure $\mu$ on $G^{(0)}$
is quasi-invariant under $G$ with Radon-Nikod\'ym derivative $dr^{*}\mu/ds^{*}s=e^{-\beta c}$
if and only if the state $\omega_{\mu}$ of $C^{*}(G)$ is an $(\alpha^{c},\beta)$-KMS
state (\cite[Proposition II.5.4]{Ren_LNM793}). Moreover, if $c^{-1}(0)$
is principal then every $(\alpha^{c},\beta)$-KMS state of $C^{*}(G)$
is of the form $\omega_{\mu}$ for some quasi-invariant probability
measure $\mu$ on $G^{(0)}$ with Radon-Nikod\'ym derivative $dr^{*}\mu/ds^{*}\mu=e^{-\beta c}$
(\cite[Proposition 3.2]{Ku_Re_PAMS06}).
If $G$ is the Deaconu-Renault groupoid associated with a local homeomorphism
$T$ on a compact metric space $(X,\rho)$, then every real valued
continuous function $\varphi$ on $X$ defines a continuous one-cocycle
on $G$ via the formula\[
c_{\varphi}(x,m-n,y)=\sum_{i=0}^{m-1}\varphi(T^{i}x)-\sum_{i=0}^{n-1}\varphi(T^{i}y).\]
Moreover, every continuous one-cocycle on $G$ is of this form (see
\cite[Lemma 2.1]{De_Ku_Mu_JOT01}). In this case, the condition that
$c_{\varphi}^{-1}(0)$ is principal means that (\cite[page 2073]{Ku_Re_PAMS06})\[
T^{n}(x)=x\;\text{and}\: n\ge1\;\implies\;\varphi(x)+\varphi(Tx)+\dots+\varphi(T^{n-1}x)\ne0.\]
This condition is trivially satisfied when $\varphi$ is strictly
positive or strictly negative.
\section{The Main Result}
In this section we show that if $(X,\rho)$ is a compact metric space
and $T:X\to X$ is a local homeomorphism which satisfies additional
conditions as described below, then the Hausdorff measure gives rise
to a KMS-state with inverse temperature the Hausdorff dimension. We
will also show that the KMS state is unique under some mild hypotheses.
\begin{defn}\label{def:loc-scaling}
Let $(X,\rho)$ be a compact metric space and let $T:X\to X$ be a
local homeomorphism. We say that $T$ satisfies the \emph{local scaling
condition} if there is a continuous function $\varphi\in C(X,\mathbb{R})$
such that:
\begin{itemize}
\item[i.]
for all $x\in X$,
\[
e^{\varphi(x)} = \lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)};
\]
\item[ii.]
for any $\varepsilon>0$ there is $\delta>0$
such that if $0<\rho(x,y)<\delta$, then
\begin{equation}
\left\vert \frac{\rho(x,y)}{\rho(Tx,Ty)}-e^{-\varphi(x)}\right\vert <\varepsilon. \label{eq:0}
\end{equation}
\end{itemize}
\end{defn}
\begin{rem}
Notice that since $T$ is assumed to be a local homeomorphism and $X$
is compact, then there is $\delta_1>0$ such that if
$0<\rho(x,y)<\delta_1$ then $Tx\ne Ty$. In the following we will
implicitly assume that $\delta$ above
satisfies $\delta \le \delta_1$. This will ensure
that inequality (\ref{eq:0}) makes sense.
\end{rem}
The next proposition provides an equivalent formulation of the local
scaling property.
\begin{prop}\label{rem:equivdef}
$T$ satisfies the local scaling condition if and only if \[(x, y) \mapsto \frac{\rho(Tx,Ty)}{\rho(x,y)}\]
extends to a continuous function $f$ on $X \times X$ that is strictly
positive on the diagonal $\Delta_X = \{(x,x)\;|\; x\in X\}$.
\end{prop}
\begin{proof}
Suppose that $f$ satisfies these conditions. Then the function $\varphi: X \to \mathbb{R}$,
given by $\varphi(x)=\log f(x,x)$ for $x\in X$,
is continuous on $X$; since $f$ is uniformly continuous on
$X\times X$, it follows easily that the two conditions of Definition
\ref{def:loc-scaling} are satisfied.
Conversely, suppose that $T$ satisfies the local scaling condition. We then define
$f:X\times X\mapsto \mathbb{R}$ via
\[
f(x,y)=\begin{cases}\frac{\rho(Tx,Ty)}{\rho(x,y)}& \text{ if
}x\ne y \\
e^{\varphi(x)} & \text{ if }x=y.
\end{cases}
\]
Then $f(x,x) > 0$ for all $x\in X$ and it suffices to
show that $f$ is continuous. Since $f$ is clearly continuous at $(x,y)$ if
$x\ne y$, we need only prove that $f$ is continuous at $(x, x)$ for $x \in X$.
We prove first that $f$ is bounded on $X\times X$. Assume, by
contradiction, that for each $n\in\mathbb{N}$ there are $x_n$ and
$y_n$ in $X$
such that $\vert f(x_n,y_n)\vert >n$. Then there is a subsequence
$(x_{n_k},y_{n_k})$ that converges to $(x,y)\in X\times X$. If $x\ne y$
then if follows that $x_{n_k}\ne y_{n_k}$ eventually, and that $\lim_{k\to
\infty}f(x_{n_k},y_{n_k})=f(x,y)$.
But this contradicts the fact that the sequence $|f(x_{n_k},y_{n_k})|$
is unbounded.
Assume now that
$x=y$. Since $\varphi$ is continuous on $X$ and thus bounded,
we may assume (by passing to a subsequence if necessary) that $x_{n_k}\ne y_{n_k}$ for all $k$.
Given $\varepsilon>0$, choose $\delta>0$ such that (\ref{eq:0}) holds.
Then there is an integer $N \ge 1$ such that for all $k \ge N$
\begin{alignat*}{3}
\rho(x,x_{n_k})&<\delta/2,\ & \rho(x,y_{n_k})&<\delta/2 &\qquad\text{and}\\
& \vert 1/f(x, x)&-1/f(x_{n_k},x_{n_k})\vert &< \varepsilon
\end{alignat*}
(by the continuity of $e^{-\varphi}$).
It follows then for such $k$ that $\rho(x_{n_k},y_{n_k})<\delta$ and so
\begin{align*}
\left\vert \frac{1}{f(x_{n_k},y_{n_k})}-\frac{1}{f(x,x)} \right\vert &\le
\left\vert \frac{1}{f(x_{n_k},y_{n_k})}-\frac{1}{f(x_{n_k},x_{n_k})} \right\vert\ +
\left\vert \frac{1}{f(x_{n_k},x_{n_k})}-\frac{1}{f(x,x)} \right\vert\
\\
&< 2\varepsilon.
\end{align*}
Since $\vert 1/f(x_n,y_n)\vert <1/n$ it follows that $1/f(x,x)=0$,
which is a contradiction.
Thus $f$ is bounded on $X\times X$; let $M>0$ be a bound
of the function.
We return now to the continuity of $f$ for $x=y$. Let $x_n\to x$ and
$y_n\to x$. If $x_n=y_n$ for all $n$ then
$\lim_{n\to \infty}f(x_n,x_n)=f(x,x)$ because $\varphi$ is assumed to be continuous. So,
by passing to a subsequence, if necessary,
we may assume that $x_n\ne y_n$ for all $n$. Let $\varepsilon>0$ be given;
choose $\delta>0$ such that (\ref{eq:0}) holds.
There is an integer $N$ such that for all $n\ge N$
\begin{alignat*}{3}
\rho(x,x_n) &<\delta/2, & \rho(x,y_n)&<\delta/2 &\qquad\text{and}\\
&\vert 1/f(x,x)&-1/f(x_{n},x_{n})\vert&< \varepsilon.
\end{alignat*}
Then, for such $n$,
\begin{align*}
\vert f(x_n,y_n)-f(x,x)\vert &= \vert f(x_n,y_n)f(x,x)\vert \left\vert \frac{1}{f(x,x)} -
\frac{1}{f(x_n,y_n)}\right\vert \\
&\le M^2\left(\left\vert \frac{1}{f(x,x)}-\frac{1}{f(x_n,x_n)} \right\vert
+ \left\vert \frac{1}{f(x_n,x_n)} -\frac{1}{f(x_n,y_n)}\right\vert\right)\\
&< 2M^2\varepsilon.
\end{align*}
Therefore, $f$ is continuous on $X\times X$.
\end{proof}
\begin{thm}
\label{thm:Main}Let $(X,\rho)$ be a compact metric space and let
$T:X\to X$ be a local homeomorphism that satisfies the local scaling
condition. Let $\varphi\in C(X,\mathbb{R})$ be such that\[
e^{\varphi(x)}=\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)},\]
and let $c_{\varphi}$ be the associated one-cocycle on $G$. Then
the state $\omega_{\mu}$ given by the equation \eqref{eq:KMSstate}
is an $(\alpha,\beta)$-KMS state where $\alpha=\alpha^{c_{\varphi}}$,
$\beta=\dim X$, and $\mu=\mu^{\beta}/\mu^{\beta}(X)$.\end{thm}
\begin{proof}
Let $G$ be the Deaconu-Renault groupoid associated with $T$. As
noted above in Section \ref{sec:prelim}, by
\cite[Proposition II.5.4]{Ren_LNM793}, it suffices to show that $\mu$
is a quasi-invariant measure on $G^{(0)}$ and
\begin{equation}
\frac{dr^{*}\mu}{ds^{*}\mu}=e^{-\beta c_{\varphi}}. \label{eq:RN}
\end{equation}
This is equivalent with\begin{equation}
\mu(U)=\int_{TU}e^{-\beta\varphi\bigl((T|_{U})^{-1}x\bigr)}d\mu(x)\label{eq:qinv}\end{equation}
for all open sets $U$ such that $T|_{U}$ is a homeomorphism onto
$TU$ (\cite[Page 6]{Ren_ETDN_05}; see also \cite{Ren_OAMP_03}).
We will call such sets \emph{sections} of $T$.
By the local scaling property (see Definition \ref{def:loc-scaling}) and
since $X$ is compact, for any $\varepsilon>0$ there is $\delta>0$
such that if $0<\rho(x,y)<\delta$ then\begin{equation}
\left\vert \frac{\rho(x,y)}{\rho(Tx,Ty)}-e^{-\varphi(x)}\right\vert <\varepsilon\label{eq:1}\end{equation}
and\begin{equation}
\vert e^{-\varphi(x)}-e^{-\varphi(y)}\vert<\varepsilon.\label{eq:2}\end{equation}
Fix $\varepsilon>0$ such that $2\varepsilon<\min_{x\in X}e^{-\varphi(x)}$.
Suppose, first, that $U$ is a section of $T$ such that $\operatorname{diam} U<\delta$,
where $\delta$ is chosen such that \eqref{eq:1} and \eqref{eq:2}
hold. Let $\underline{x},\overline{x}\in\overline{U}$ such that $e^{-\varphi(\underline{x})}\le e^{-\varphi(x)}\le e^{-\varphi(\overline{x})}$
for all $x\in U$. Let $\varepsilon_{1}>0$ and suppose that $\mathcal{A}$
is an $\varepsilon_{1}$-cover of $U$ consisting of subsets of $U$
(see \cite[page 167]{Edg_MTFG2}). Then, for any $A\in\mathcal{A}$
we have that\[
\operatorname{diam} A=\sup_{x,y\in A,x\ne y}\rho(x,y)=\sup_{x,y\in A,x\ne y}\rho(Tx,Ty)\frac{\rho(x,y)}{\rho(Tx,Ty)}.\]
Now\[
\left\vert \frac{\rho(x,y)}{\rho(Tx,Ty)}-e^{-\varphi(\underline{x})}\right\vert <2\varepsilon\]
and\[
\left\vert \frac{\rho(x,y)}{\rho(Tx,Ty)}-e^{-\varphi(\overline{x})}\right\vert <2\varepsilon\]
for all $x,y\in A$ with $x\ne y$. Therefore\[
\bigl(e^{-\varphi(\overline{x})}-2\varepsilon\bigr)\operatorname{diam} TA\le\operatorname{diam} A\le\bigl(e^{-\varphi(\underline{x})}+2\varepsilon)\operatorname{diam} TA.\]
Hence \[
\bigl(e^{-\varphi(\overline{x})}-2\varepsilon\bigr)^{\beta}\overline{\mu}_{\varepsilon_{1}}^{\beta}(TU)\le\overline{\mu}_{\varepsilon_{1}}^{\beta}(U)\le\bigl(e^{-\varphi(\underline{x})}+2\varepsilon\bigr)^{\beta}\overline{\mu}_{\varepsilon_{1}}^{\beta}(TU)\]
and, by taking the limit when $\varepsilon_{1}$ goes to $0$,\begin{equation}
\bigl(e^{-\varphi(\overline{x})}-2\varepsilon\bigr)^{\beta}\mu(TU)\le\mu(U)\le\bigl(e^{-\varphi(\underline{x})}+2\varepsilon\bigr)^{\beta}\mu(TU).\label{eq:3}\end{equation}
Let now $U$ be an arbitrary section of $T$. Then there is a finite disjoint
family $\{U_{1}, \dots , U_{N}\}$ of sections of $T$ such that $\overline{U}=\bigcup_{n=1}^N\overline{U}_{n}$
and $\operatorname{diam} U_{n}<\delta$ for $n = 1, \dots, N$. Indeed, by the compactness of $\overline{U}$, there are finitely
many open sets $\{V_{1}, \dots , V_{N}\}$ such that $\overline{U} \subset \bigcup_{n=1}^N V_{n}$ and
$\operatorname{diam} V_{n}<\delta$ for $n = 1, \dots, N$. Now set $U_1 =: U \cap V_1$ and
\[
U_{n+1} =: U \cap V_{n+1} \setminus (\overline{V}_{1} \cup \cdots \cup \overline{V}_{n}).
\]
It is routine to check that the $U_n$ satisfy the requisite conditions.
Let $\underline{x}_{n},\overline{x}_{n}\in\overline{U}_{n}$
such that $e^{-\varphi(\underline{x}_{n})}\le e^{-\varphi(x)}\le e^{-\varphi(\overline{x}_{n})}$
for all $x\in U_{n}$. Then, by inequality \eqref{eq:3},\[
\mu(U)=\sum_{n}\mu(U_{n})\le\sum_{n}\int_{TU_{n}}\bigl(e^{-\varphi(\underline{x}_{n})}+2\varepsilon)^{\beta}d\mu(x).\]
For all $x\in TU_{n}$ we have that $e^{-\varphi(\underline{x}_{n})}\le e^{-\varphi\bigl((T|_{U})^{-1}(x)\bigr)}$.
Thus\[
\mu(U)\le\int_{TU}\left(e^{-\varphi\bigl((T|_{U})^{-1}(x)\bigr)}+2\varepsilon\right)^{\beta}d\mu(x).\]
Similarly, since $e^{-\varphi\bigl((T|_{U})^{-1}(x)\bigr)}\le e^{-\varphi(\overline{x}_{n})}$
for all $x\in TU_{n}$, we have that\[
\int_{TU}\left(e^{-\varphi\bigl(T|_{U})^{-1}(x)\bigr)}-2\varepsilon\right)^{\beta}d\mu(x)\le\mu(U).\]
Since $\varepsilon$ was arbitrarily chosen it follows that \eqref{eq:qinv}
holds and the conclusion follows.
\end{proof}
To ensure the uniqueness of the $(\alpha,\beta)$-KMS state we need
to impose some conditions on the local homeomorphism $T$ and the
map $\varphi$. We say that $T$ is \emph{positively expansive} if
there is an $\varepsilon>0$ such that for all $x\ne y$ there is
an $n\in\mathbb{N}$ with $\rho(T^{n}x,T^{n}y)\ge\varepsilon$. We
say that $T$ is \emph{exact} if for every non-empty open set $U\subset X$
there is an $n>0$ such that $T^{n}(U)=X$.
We say that a real-valued continuous function $\varphi$ on $X$ satisfies the
\emph{Bowen condition} with respect to $T$
(see \cite[Definition 2.7]{Ku_Re_PAMS06})
if there are $\delta, C>0$ such that
\[
\sum_{i=0}^{n-1}\varphi(T^{i}x)-\varphi(T^{i}y)\le C,
\]
for all $x, y \in X$ and $n > 0$ such that $\rho(T^ix, T^iy) \le \delta$
for $0 \le i \le n-1$.
Note that if $T$ is positively expansive and
$\varphi$ is H\"older, that is, $\vert\varphi(x)-\varphi(y)\vert\le k\rho(x,y)^{l}$
for some positive constants $k$ and $l$, then $\varphi$ satisfies
the Bowen condition
(see the discussion following \cite[Definition 2.7]{Ku_Re_PAMS06}).
\begin{prop}
\label{pro:uniqueness}Assume that the local homeomorphism $T:X\to X$
is positively expansive and exact, and assume that $\varphi$ satisfies
the Bowen condition. Let $\alpha$ be the action on $C^{*}(G)$ determined
by $c_{\varphi}$ and let $\beta$ the Hausdorff dimension of $X$.
If $T$ satisfies the local scaling condition and $c_{\varphi}^{-1}(0)$
is principal then $\beta$ is the unique inverse temperature which
admits a KMS state for $\alpha$. Moreover, the $(\alpha,\beta)$-KMS
state $\omega_{\mu}$ is unique.\end{prop}
\begin{proof}
Theorem \ref{thm:Main} implies that there is a $(\alpha,\beta)$-KMS
state, namely $\omega_{\mu}$. By Walters\textquoteright{} version
of the Ruelle-Perron-Frobenius Theorem (\cite[Theorem 2.8]{Ku_Re_PAMS06},
see also \cite[Theorem 6.1]{Ren_ETDN_05}, \cite[Theorem 8]{Wal_78},
\cite[Theorem 2.16]{Wal_01}) there are unique $\lambda>0$ and probability
measure $\nu$ such that\[
\mathcal{L}_{-\varphi}^{*}(\nu)=\lambda\nu,\]
where $\mathcal{L}_{-\varphi}(f)(x)=\sum_{Ty=x}e^{-\varphi(y)}f(y)$
is the transfer operator associated with $\varphi$. Proposition 4.2
of \cite{Ren_OAMP_03} (see also \cite[Proposition 3.4.1]{Ren_2009})
implies that $\beta$ is the unique inverse temperature which admits
a KMS state for $\alpha$ and $\nu=\mu$.
Theorem \ref{thm:Main} and \cite[Theorem 3.5 i)]{Ku_Re_PAMS06} imply
that $P(T,-\beta\varphi)=0$, where $P(T,\cdot)$ is the topological
pressure(\cite[Section 9.1]{Wal_GTM82}). Since $c_{\varphi}^{-1}(0)$
is principal, \cite[Theorem 3.5 ii)]{Ku_Re_PAMS06} implies that the
$(\alpha,\beta)$-KMS state $\omega_{\mu}$ is unique.
\end{proof}
In the last result of this section we show how one can compute the
topological entropy of the local endomorphism $T$ under some suitable
hypothesis. Recall that the topological entropy of $T$, $h(T)$,
equals $P(T,0)$.
\begin{prop}
\label{pro:Entropy}Let $T:X\to X$ be a local homeomorphism. Suppose
that $T$ is positively expansive, exact, and satisfies the local
scaling condition. Assume that the map $\varphi$ is constant,
that is, there is a constant $\tau>1$ such that\[
\tau=\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)},\]
for all $x\in X$. Then $h(T)=\beta\ln\tau$, where $\beta$ is the
Hausdorff dimension of $X$.\end{prop}
\begin{proof}
By Theorem \ref{thm:Main} and Proposition \ref{pro:uniqueness} there
is a unique $(\alpha,\beta)$-KMS state $\omega_{\mu}$. Moreover,
Theorem 3.5 of \cite{Ku_Re_PAMS06} implies that $P(T,-\beta\ln\tau)=0$.
Since $P(T,-\beta\ln\tau)=h(T)-\beta\ln\tau$, it follows that $h(T)=\beta\ln\tau$.
\end{proof}
\section{Examples}
\subsection{\label{sub:Cuntz-algebras}Cuntz algebras}
Fix $n\in\mathbb{N}$ with $n>1$ and let $E=\{1,\dots,n\}$ be the
alphabet with $n$ letters. Let $(r_{1},r_{2},\dots,r_{n})$ be a
list of positive numbers such that $r_{i}<1$ for all $i\in\{1,\dots,n\}$.
We call such a list a \emph{contractive ratio list} (see \cite[Chapter 4]{Edg_MTFG2}).
We set $X$ to be the infinite path space over the alphabet $E$,
that is\[
X=E^{\infty}=\{(x_{k})_{k\in\mathbb{N}}\,:\, x_{k}\in E\}.\]
We write $E^{k}$ for the set of paths of length $k$ over $E$ and
we set $E^{*}=\bigcup_{k}E^{k}$. We define a metric on $X$ based
on the given ratio list. For $\sigma\in E^{k}$ the cylinder of $\sigma$
is\begin{equation}
Z(\sigma)=\{(x_{m})\in X\,:\, x_{0}=\sigma_{0},\dots,x_{k-1}=\sigma_{k-1}\}.\label{eq:cylinder}\end{equation}
We specify the metric on $X$ by requiring that the diameter of a
cylinder $Z(\sigma)$, with $\sigma\in E^{k+1}$ for some k, to equal
$r_{\sigma}:=r_{\sigma_{0}}r_{\sigma_{1}}\dots r_{\sigma_{k}}$. Thus
for $x\ne y$\begin{eqnarray*}
\rho(x,y) & = & \begin{cases}
\inf\{r_{\sigma}\,:\, x,y\in Z(\sigma)\} & \text{if }x_{0}=y_{0}\\
1 & \text{if }x_{0}\ne y_{0}\end{cases}.\end{eqnarray*}
Then the left shift, $T:X\to X$, $T((x_{k}))=(x_{k+1})$, is a local
homeomorphism on $X$. If $G$ is the corresponding Deaconu-Renault
groupoid then $C^{*}(G)$ is isomorphic to the Cuntz algebra $\mathcal{O}_{n}$
(\cite[Section III.2]{Ren_LNM793}, \cite{Dea_TAMS95}); under this identification
the generating isometries of $\mathcal{O}_{n}$ are given by
$S_{j}=1_{\Gamma_{j}}$, where $\Gamma_{j}=\{(jx,1,x)\,:\, x\in X\}$,
for $j=1,\dots,n$. A routine computation shows that $T$ satisfies the
local scaling property and
\[
\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)}=\frac{1}{r_{x_{0}}},
\]
that is, $\varphi(x)=-\log r_{x_{0}}$. Thus the automorphism
$\alpha$ on $\mathcal{O}_{n}$ is determined by
\begin{equation}
\alpha_{t}(S_{j})=e^{-it\log r_{j}}S_{j}.\label{eq:gauge_action}
\end{equation}
The Hausdorff dimension $\beta$ of $X$ is the
unique number that satisfies the equation (\cite[Theorem 6.4.3]{Edg_MTFG2})
\begin{equation}
\sum_{i=1}^{n}r_{i}^{s}=1.\label{eq:Hausdorff_Cantor}\end{equation}
The Hausdorff measure on $X$ is the unique Borel measure that satisfies
$\mu^{\beta}(Z(\sigma))=r_{\sigma}^{\beta}$ for all $\sigma\in E^{*}$.
Theorem \ref{thm:Main} and Proposition \ref{pro:uniqueness} imply
that there exists a KMS-state for $\alpha$ at temperature $\beta$
if and only if $\beta$ satisfies \eqref{eq:Hausdorff_Cantor}. The
KMS state is unique in this case. We recover, thus, Theorem 2.2 of
\cite{Eva_80}.
If $r_{1}=r_{2}=\dots=r_{n}=1/e$, then the inverse temperature is
$\log n$, which is the main result in \cite{Ole_Ped_MS78}. Moreover
the topological entropy of $T$ is $\log n$. More generally, let
$s>0$ be arbitrary and set $r_{1}=r_{2}=\dots=r_{n}=n^{-1/s}$. Then
$(r_{1},\dots,r_{n})$ is a contractive ratio list. The Hausdorff
dimension of $X$ is $\beta=s$ and $\varphi(x)=\frac{\log n}{s}$.
The topological entropy of $T$ is still $\log n$.
\subsection{\label{sub:Generalized-gauge-actions-Cuntz}Generalized gauge actions
on Cuntz algebras}
We still assume that $E=\{1,\dots,n\}$ is a finite alphabet and we
let $X=E^{\infty}$ endowed with the product topology. Then $X$ is
a compact topological space. Let $T$ be the left shift, as in the
previous example. Fix now a continuous function $f$ on $X$ such
that $f(x)>1$ for all $x\in X$. We will define next a metric $\rho_{f}$
on $X$ such that\begin{equation}
\lim_{y\to x}\frac{\rho_{f}(Tx,Ty)}{\rho_{f}(x,y)}=f(x).\label{eq:f_metric}\end{equation}
Let $\sigma\in E^{*}$ and let $Z(\sigma)$ be the cylinder of $\sigma$
as defined in \eqref{eq:cylinder}. We will write $\vert\sigma\vert$
for the length of the finite word $\sigma$. Define
\begin{equation}\label{eq:w_sigma}
w_{\sigma}:=\max_{z,w\in
Z(\sigma)}\left(\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}z)\cdot\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}w)\right)^{-1/2}.
\end{equation}
Let $\alpha,\beta\in E^{*}$ such that $\vert\alpha\vert<\vert\beta\vert$
and $\alpha_{i}=\beta_{i}$ for $i=0,\dots,\vert\alpha\vert-1$; in this case,
write $\alpha<\beta$ (\cite{Edg_MTFG2}). Since $f(x)>1$
for all $x\in X$, it follows that if $\alpha<\beta$ then
$w_{\beta}>w_{\alpha}$ and $\lim_{n\to \infty}w_{\sigma|n}=0$ for
$\sigma\in E^\infty$.
Proposition 2.6.5 of \cite{Edg_MTFG2} implies that there is a metric
$\rho_{f}$ on $X$ such that the diameter of $Z(\alpha)=w_{\alpha}$
for all $\alpha\in E^{*}$ and, if $x,y\in X$ such that
$x_{i}=y_{i}$ for $i=0,\dots,k-1$ and $x_{k}\ne y_{k}$ then
$\rho_{f}(x,y)=w_\sigma$, where $\sigma\in E^k$ such that $\sigma_i=x_i$,
$i=0,\dots, k-1$. Thus if $x,y\in X$ and the length of their common
longest prefix $\sigma$ is at least one, then
\begin{equation}
\label{eq:4}
\rho_f(x,y)=w_\sigma:=\max_{z,w\in
Z(\sigma)}\left(\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}z)\cdot\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}w)\right)^{-1/2}.
\end{equation}
Note that the maximum in (\ref{eq:4}) is attained for some
$\overline{z}$ and $\overline{w}$ in $Z(\sigma)$ since the function
$f$ is continuous. Also, if $m=\min_{x\in X}\vert f(x)\vert$ and
$M=\max_{x\in X} \vert f(x)\vert$, then
\[
\frac{1}{M^{\vert\sigma\vert}}\le \rho_f(x,y) \le \frac{1}{m^{\vert\sigma\vert}},
\]
where $\sigma$ is the longest common prefix of $x$ and $y$. Thus,
since $m>1$ by hypothesis, $\rho_f(x,y)<\varepsilon$ if and only if
there is a large $N$ and $\sigma\in E^N$ such that $x,y\in
Z(\sigma)$. Thus the metric $\rho_f$ generates the topology on $X$.
If $(r_{1},\dots,r_{n})$ is a contractive ratio list and $f(x)=1/r_{x_{0}}$
we recover the metric from the previous example.
We prove next that $T$ satisfies the local scaling condition
(see Definition \ref{def:loc-scaling}) with respect to
the metric $\rho_{f}$ with $\varphi = \log f$.
For this we need
first a lemma.
\begin{lem}\label{lem:generalized_metric}
Let $x,y\in X$ such that their longest common prefix $\sigma$ has
length at least $2$. Then there are points $z_i,w_i\in Z(\sigma)$,
$i=1,2$ such that
\begin{equation}\label{eq:ineq_distance}
\bigl(f(z_1)f(w_1)\bigr)^\frac{1}{2} \le
\frac{\rho_f(Tx,Ty)}{\rho_f(x,y)}\le \bigl(f(z_2)f(w_2)\bigr)^{\frac{1}{2}}.
\end{equation}
\end{lem}
\begin{proof}
Observe $w_\sigma=\rho_f(x,y)$ and $w_{\sigma^\prime
}=\rho_f(Tx,Ty)$ (Equation~(\ref{eq:4})), where $\sigma^\prime
:=\sigma_1\dots\sigma_{\vert\sigma\vert-1}$. Then
\[
\frac{\rho_f(Tx,Ty)}{\rho_f(x,y)}=\frac{w_{\sigma^\prime}}{w_\sigma}.
\]
Note that
if $z\in Z(\sigma)$ then $Tz\in Z(\sigma^\prime)$. Conversely, if
$z^\prime \in Z(\sigma^\prime)$ then $\sigma_0z^\prime \in
Z(\sigma)$. Let $z_1,w_1\in Z(\sigma)$ the points for which the maximum in
(\ref{eq:4}) is attained. That is,
\[
w_\sigma=\left(\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}z_1)\cdot\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}w_1)\right)^{-1/2}.
\] Then $Tz_1,Tw_1\in Z(\sigma^\prime)$ and
\begin{eqnarray*}
w_\sigma&=&\bigl(f(z_1)f(w_1)\bigr)^{-\frac{1}{2}}\cdot
\left(\prod_{i=1}^{\vert\sigma\vert-1}f(T^{i}z_1)\cdot\prod_{i=1}^{\vert\sigma\vert-1}f(T^{i}w_1)\right)^{-1/2}\\
&\le&\bigl(f(z_1)f(w_1)\bigr)^{-\frac{1}{2}}\cdot w_{\sigma^\prime}.
\end{eqnarray*}
Thus the first part of (\ref{eq:ineq_distance}) holds. Let now
$\overline{z}_2,\overline{w}_2\in Z(\sigma^\prime)$ such that
\[
w_{\sigma^\prime}=\left(\prod_{i=0}^{\vert\sigma\vert-2}f(T^{i}\overline{z}_2)\cdot\prod_{i=0}^{\vert\sigma\vert-2}f(T^{i}\overline{w}_2)\right)^{-1/2}.
\] Then $z_2=\sigma_0\overline{z}_2$ and $w_2=\sigma_0\overline{w}_2$
belong to $Z(\sigma)$ and
\[
\left(\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}z_2)\cdot\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}w_2)\right)^{-1/2}
\le w_\sigma.
\]
Thus
\[
\bigl(f(z_2)f(w_2)\bigr)^{-\frac{1}{2}} w_{\sigma^\prime}\le w_\sigma,
\] which is the second part of (\ref{eq:ineq_distance}).
\end{proof}
Let $\varepsilon >0$ be given. Since $f$ is uniformly continuous on $X$
there
is a $\delta >0$ such that if $x$ and $y$ are points in $X$ such that
the length of their longest common prefix $\sigma$ satisfies
$1/m^{\vert\sigma\vert}<\delta$ then $\vert
f(z)-f(w)\vert<\varepsilon$ for all $z,w\in Z(\sigma)$. Inequality
(\ref{eq:ineq_distance}) implies then that
\[
f(x)-\varepsilon \le
\frac{\rho_f(Tx,Ty)}{\rho_f(x,y)}\le f(x)+\varepsilon
\]
Therefore~(\ref{eq:f_metric}) holds and (\ref{eq:0}) also holds with $\varphi = \log f$;
hence, $T$ satisfies the local scaling condition (see Definition \ref{def:loc-scaling}).
The associated automorphism $\alpha_{t}$ of $C^{*}(G)\simeq\mathcal{O}_{n}$
is given via\[
\alpha_{t}(a)(x,m-k,y)=e^{itc_{f}(x,m-k,y)}a(x,m-k,y),\]
for all $(x,m-k,y)\in G$ and $a\in C_{c}(G)$, where the cocycle
$c_{f}$ is given by\begin{equation}
c_{f}(x,m-k,y)=\log\left(\prod_{i=0}^{m-1}f(T^{i}x)/\prod_{i=0}^{k-1}f(T^{i}y)\right).\label{eq:cocycle_gen}\end{equation}
These actions generalize the gauge actions on $\mathcal{O}_{n}$ described
in the previous example. To see this, recall that $S_{j}=1_{\Gamma_{j}}$,
where $\Gamma_{j}=\{(jx,1,x)\,:\, x\in X\}$, for $j=1,\dots,n$.
If $f(x)=1/r_{x_{0}}$ we have that\[
e^{itc_{f}(x,m-k,y)}S_{j}(x,m-k,y)=e^{-it\log r_{j}}S_{j}(x,m-k,y),\]
and we recover \eqref{eq:gauge_action}. Assuming that the Hausdorff
dimension $\beta$ of $X$ is strictly positive and finite, Theorem
\ref{thm:Main} implies that there exists a KMS-state $\omega_{\mu}$
on $\mathcal{O}_{n}$ for $\alpha$ at inverse temperature $\beta$.
If $\log f(x)$ satisfies the Bowen condition, then $\beta$ and $\mu$
are the unique solutions of the equation $\mathcal{L}_{f,\beta}^{*}(\mu)=\mu$,
where\[
\mathcal{L}_{f,\beta}(a)(x)=\sum_{j=1}^{n}f(jx)^{-\beta}a(jx),\]
for all $x\in X$ and $a\in C(X)$ (\cite[Proposition 7.1]{Ren_ETDN_05},
\cite[Theorem 2.8 and proof of Proposition 3.5]{Ku_Re_PAMS06}). In this case, $\omega_{\mu}$
is the unique KMS state for $\alpha$ on $\mathcal{O}_{n}$ by Proposition
\ref{pro:uniqueness}.
\subsection{Graph $C^{*}$-algebras}
Suppose that $E=(E^{0},E^{1},r,s)$ is a finite directed graph, where
$r$ and $s$ are the range and source maps. Suppose that $\{r_{e}\}_{e\in E^{1}}$
is a list of positive numbers such that $r_{e}<1$ for all $e\in E^{1}$.
We say that $\{r_{e}\}_{e\in E^{1}}$ is a \emph{contractive ratio
list }for the graph $E$. A path of length $n$ in the graph $E$
is a finite sequence $\sigma=\sigma_{0}\sigma_{1}\dots\sigma_{n-1}$
such that $\sigma_{i}\in E^{1}$ for all $i\in\{0,\dots,n-1\}$ and
$s(\sigma_{i})=r(\sigma_{i+1})$ for all $i\in\{0,\dots,n-2\}$. We
write $E^{n}$ for the set of paths of length $n$ and $E^{*}=\bigcup_{n}E^{n}$
for the set of \emph{finite paths} in the graph $E$. We extend the
definition of $r$ and $s$ to $E^{*}$ via $r(\sigma)=r(\sigma_{0})$
and $s(\sigma)=s(\sigma_{n-1})$ if $\sigma\in E^{n}$. The \emph{infinite
path space} $X$ is defined via\[
X:=E^{\infty}:=\{(x_{n})_{n\in\mathbb{N}}\,:\, x_{n}\in E^{1}\text{ and }s(x_{n})=r(x_{n+1})\}.\]
For an element $x\in E^{\infty}$ we set $r(x)=r(x_{0})$. We say
that the graph $E$ is \emph{irreducible }(or strongly connected)
if for every two elements $v,w\in E^{0}$ there is a path $\sigma\in E^{*}$
such that $r(\sigma)=v$ and $s(\sigma)=w$. We assume that the graph
$E$ is irreducible and it satisfies condition (L) from \cite{KuPaRa_PJM98}
(or, equivalently since $E$ is finite, condition (I) from \cite{Cu_Kr_80}).
That is, we assume that every loop in $E$ has an exit.
If $s$ is a positive real number, then $s$-dimensional \emph{Perron
numbers} for the graph $E$ are positive numbers $q_{v}$, one for
each vertex $v\in E^{0}$, such that\begin{equation}
q_{v}^{s}=\sum_{w\in E^{0},e\in vE^{1}w}r(e)^{s}q_{w}^{s}\label{eq:Hausdorff_graph}\end{equation}
for all $v\in E^{0}$ (\cite[Section 6.6]{Edg_MTFG2}). For an irreducible
graph $E$ there is a unique number $s\ge0$ such that Perron numbers
exist (\cite[Theorem 6.9.6]{Edg_MTFG2}).
For $v\in E^{0}$, we define $vE^{\infty}:=\{x\in E^{\infty}\,:\, r(x)=v\}$.
The metric $\rho$ on $E^{\infty}$ is defined such that \[
\rho(ex,ey)=r_{e}\rho(x,y)\text{ for all }x,y\in s(e)E^{\infty}.\]
We define it such that the diameter of a cylinder $Z(\sigma)$ of
$\sigma$ equals $r_{\sigma}q_{s(\sigma)}$, where, for $\sigma\in E^{n}$,\[
Z(\sigma)=\{x\in E^{\infty}\,:\, x_{i}=\sigma_{i},i=0,\dots,n-1\}.\]
Provided that the number $s$ which satisfies condition \eqref{eq:Hausdorff_graph}
is positive, the Hausdorff dimension $\beta$ of $X$ equals $s$
and the Hausdorff measure on $X$ is the unique Borel measure such
that $\mu^{\beta}(Z(\sigma))=q_{s(\sigma)}^{\beta}r_{\sigma}^{\beta}$
for all $\sigma\in E^{*}$ (\cite[Section 6.6]{Edg_MTFG2}).
The left shift $T:X\to X$ defined by $T(x_{n})=(x_{n+1})$ is a local
homeomorphism on $X$. If $G$ is the associated Deaconu-Renault groupoid,
then $C^{*}(G)$ is isomorphic to the graph $C^{*}$-algebra $C^{*}(E)$
(\cite{KuPaRaRe_JFA97},\cite{KuPaRa_PJM98}; see also \cite{Rae_CBMS05}).
The local homeomorphism $T$ satisfies the local scaling property
with\[
\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)}=\frac{1}{r_{x_{0}}},\]
that is, $\varphi(x)=-\log r_{x_{0}}$. The corresponding automorphism
of $C^{*}(G)$ is defined then via\[
\alpha_{t}(P_{v})=P_{v}\;\text{for all }v\in E^{0}\]
and\[
\alpha_{t}(S_{e})=e^{-it\log r_{e}}S_{e},\]
where $(\{P_{v}\}_{v\in E^{0}},\{S_{e}\}_{e\in E^{1}}\}$ is a \emph{universal}
Cuntz-Krieger family generating $C^{*}(G)$ (\cite{KuPaRa_PJM98,Rae_CBMS05}).
Theorem \ref{thm:Main} and \ref{pro:uniqueness} imply that there
exists a KMS-state for $\alpha$ at inverse temperature $\beta$ if
and only if $\beta$ is the unique positive number for which Perron
numbers exist. In this case, there is a unique $(\alpha,\beta)$-KMS
state, namely\[
\omega_{\mu}(f)=\int E(f)d\mu,\]
where $\mu$ is the normalized Hausdorff measure on $X$ and $E$ is the
canonical expectation onto $C(X)$.
If $r_{e}=r\in(0,1)$ for all $e\in E^{1}$, then the Hausdorff dimension
satisfies $\lambda=r^{-\beta}$, where $\lambda$ is the Perron-Frobenius
eigenvalue of the vertex matrix $A_{E}$ of the graph $E$ (see, for
example, \cite[Theorem 6.9.6]{Edg_MTFG2}). Proposition
\ref{pro:Entropy} implies that the topological entropy of $T$ is
$\log\lambda$.
\subsection{Generalized gauge actions on graph $C^{*}$-algebras}
The construction of Example \ref{sub:Generalized-gauge-actions-Cuntz}
extends to the graph $C^{*}$-algebras described in the previous example.
Suppose that $X$ is the infinite path space of a finite directed
graph $E=(E^{0},E^{1},r,s)$ which is irreducible and satisfies condition
(L) of \cite{Ku_Re_PAMS06}. We endow $X$ with the product topology
so that it is a compact topological space and we let $T$ be the left
shift map on $X$. Fix a continuous function $f$ on $X$ such that
$f(x)>1$ for all $x\in X$ and fix a list of positive number
$\{q_v\}_{v\in E^0}$ such that
\begin{equation}
\label{eq:5}
f(x)>\frac{q_{s(x_0)}}{q_{r(x_0)}}\;\text{ for all}\;x\in E^\infty.
\end{equation}
For example, if $q_v=1$ for all $v\in V$ then the above condition is
trivially satisfied. We define a metric $\rho_{f}$ on $vX=vE^{\infty}$
for each $v\in E^{0}$ by requiring that the diameter of each cylinder
$Z(\sigma)$, $\sigma\in vE^{*}$, to equal $w_\sigma\cdot
q_{s(\sigma)}$, where
\begin{equation}\label{eq:wsigma_graph}
w_{\sigma}:=\max_{z,w\in Z(\sigma)}\left(\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}z)\cdot\prod_{i=0}^{\vert\sigma\vert-1}f(T^{i}w)\right)^{-1/2}.
\end{equation}
Since $f(x)>1$ and the numbers $q_v$ satisfy (\ref{eq:5}), it follows
that if $\alpha<\beta$ with $\alpha,\beta\in E^*$ then $\operatorname{diam}
Z(\alpha)<\operatorname{diam} Z(\beta)$ and $\lim_{n\to \infty}\operatorname{diam} Z(\sigma|_n)=0$
for all $\sigma\in E^{\infty}$. Then \cite[Proposition
2.6.5]{Edg_MTFG2} implies that there
is a metric $\rho_{f}$ on $vX$ with the desired property for each
$v\in E^{0}$. Namely, if
$x,y\in vE^\infty$ and $\sigma\in E^*$ is their longest common prefix,
then
\[
\rho_f(x,y)=w_\sigma\cdot q_{s(\sigma)},
\]
where $w_\sigma$ is defined in (\ref{eq:wsigma_graph}).
Since
$X$ is the finite disjoint union of $vX$ we can extend $\rho_{f}$
to a metric on $X$.
If $f(x)=1/r_{x_{0}}$ for a contractive ratio
list $\{r_{e}\}_{e\in E^{1}}$ and $\{q_v\}_{v\in E^{0}}$ are $s$-dimensional
Perron numbers for the graph $E$ then the condition (\ref{eq:5}) is
implied by (\ref{eq:Hausdorff_graph}). Thus we recover the metric
from the previous example.
One can easily see that if $x$ and $y$ have a common prefix $\sigma$
of length at least $2$ then
\[
\frac{\rho(Tx,Ty)}{\rho(x,y)}=\frac{w_{\sigma^\prime}}{w_\sigma},
\] because $s(\sigma^\prime)=s(\sigma)$, where $\sigma^{\prime}$ is
obtained from $\sigma$ by removing the first entry
($\sigma^\prime=\sigma_1\dots \sigma_{\vert \sigma\vert -1}$). Then the equivalent of Lemma
\ref{lem:generalized_metric} holds and $T$ satisfies the local
scaling property (see Definition \ref{def:loc-scaling}) with respect to
the metric $\rho_{f}$ with\[
\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)}=f(x),\]
for all $x\in X$, and condition (\ref{eq:0}) holds with
$\varphi =\log f$. The associated
action $\alpha_{t}$ on $C^{*}(G)\simeq C^{*}(E)$, where $G$ is
the Deaconu-Renault groupoid defined by $T$, is given via\[
\alpha_{t}(a)(x,m-n,y)=e^{itc_{f}(x,m-n,y)}a(x,m-n,y),\]
for all $(x,m-n,y)\in G$ and $a\in C(G)$, where the cocycle $c_{f}$
is defined as in \eqref{eq:cocycle_gen}. These actions generalize
the gauge actions described in the previous example because a universal
Cuntz-Krieger family on $C^{*}(E)$ is given by $\bigl(\{1_{\Delta_{v}}\}_{v\in E^{0}},\{1_{\Gamma_{e}}\}_{e\in E^{1}}\bigr)$,
where\[
\Delta_{v}=\{(x,0,x)\in G^{0}\,:\: r(x)=v\}\]
and\[
\Gamma_{e}=\{(ex,1,x)\in G\,:\, x\in s(e)E^{\infty}\}.\]
Then\begin{eqnarray*}
e^{itc_{f}(x,m-n,y)}1_{\Delta_{v}}(x,m-n,y) & = & 1_{\Delta_{v}}(x,m-n,y),\;\text{and}\\
e^{itc_{f}(x,m-n,y)}1_{\Gamma_{e}}(x,m-n,y) & = & e^{it\log r_{e}}1_{\Gamma_{e}}(x,m-n,y).\end{eqnarray*}
Let $\beta$ be the Hausdorff dimension of $X$ with respect to the
metric $\rho_{f}$ and assume that $0<\beta<\infty$. Theorem \ref{thm:Main}
implies that there exists a KMS state $\omega_{\mu}$ on the graph
$C^{*}$-algebra $C^{*}(E)$ for $\alpha$ at inverse temperature
$\beta$. Moreover, if $\log f$ satisfies the Bowen condition, then
$\beta$ is the unique inverse temperature for which KMS states exist
and $\omega_{\mu}$ is the unique $(\alpha,\beta)$-KMS state. The
Hausdorff dimension $\beta$ and the measure $\mu$ are the unique
solutions of the equation $\mathcal{L}_{f,\beta}^{*}(\mu)=\mu$ (\cite{Ren_ETDN_05,Ku_Re_PAMS06}),
where\[
\mathcal{L}_{f,\beta}(a)(x)=\sum_{e\in E^{1},s(e)=r(x)}f(ex)^{-\beta}a(ex).\]
We recover, thus, some of the results of \cite{Exe_BBMS04}.
\subsection{Coverings of $\mathbb{T}$}
Let $f:[0,1]\to\mathbb{R}$ be a continuous positive function such
that $f(0)=f(1)$
and $n:=\int_{0}^{1}f(t)dt$ is a positive integer such that $n\ge2$.
For example, the constant function $n$ satisfies these requirements.
Let $X=\mathbb{R}/\mathbb{T}$ and let $\rho$ the induced metric
on $X$. We define $T:X\to X$ by\[
T(x)=\int_{0}^{x}f(t)dt.\]
Then $T$ is a local homeomorphism. Moreover, $T$ is an $n$-fold
covering map. One can easily compute that\[
\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)}=f(x).\]
Thus $T$ satisfies the local scaling condition and $\varphi(x)=\log f(x)$.
Since the Hausdorff dimension $\beta=1$, Theorem \ref{thm:Main}
implies that $\omega_{\mu}$ is an $(\alpha,1)$-KMS state, where
$\mu=\mu^{1}$. If $f$ is continuously differentiable and $f(x)>1$,
then $\varphi=\log f$ is both positive and H\"older. Thus $c_{\varphi}^{-1}(0)$
is principal and $\varphi$ satisfies the Bowen condition. Moreover,
$T$ is positively expansive and exact. Therefore, by Proposition
\ref{pro:uniqueness}, $\omega_{\mu}$ is the unique KMS state on
$C^{*}(G)$.
Suppose now that $f(x)=n$ for all $x\in[0,1]$. Then $f$ is continuously
differentiable and $f(x)>1$. Moreover, Proposition \ref{pro:Entropy}
implies that $h(T)=\log n$.
\subsection{The Sierpinski octafold}
Consider the triangle $Y$ with vertices $v_{1}=(1,0,0)$, $v_{2}=(0,1,0)$,
and $v_{3}=(0,0,1)$ in $\mathbb{R}^{3}$. Let $f_{1},f_{2},f_{3}$
be the the restrictions to $Y$ of the linear maps defined by the
matrices
\[
A_{1}=\left[\begin{array}{ccc}
1 & 1/2 & 1/2\\
0 & 1/2 & 0\\
0 & 0 & 1/2\end{array}\right],\, A_{2}=\left[\begin{array}{ccc}
1/2 & 0 & 0\\
1/2 & 1 & 1/2\\
0 & 0 & 1/2\end{array}\right],\, A_{3}=\left[\begin{array}{ccc}
1/2 & 0 & 0\\
0 & 1/2 & 0\\
1/2 & 1/2 & 1\end{array}\right].
\]
Note that for $i = 1, 2, 3$, we have $f_i(Y) \subset Y$ and
$f_{i}$ is a similarity with ratio $1/2$.
So $(f_{1},f_{2},f_{3})$ is an iterated function system (\cite{Hut_81})
and it admits a unique invariant set $K$ (\cite{Hut_81,Edg_MTFG2})\[
K=f_{1}(K)\bigcup f_{2}(K)\bigcup f_{3}(K).\]
The set $K$ is a copy of the Sierpinski gasket. Using the terminology from
\cite{Kig_CUP01} and \cite{Str_Prin06}, we call the sets $f_{i}(K)$,
$i=1,2,3$, $1$-cells.
Now consider the octahedron in $\mathbb{R}^{3}$ with vertices at
$(\pm 1, 0, 0)$, $(0, \pm 1, 0)$ and $(0, 0, \pm 1)$; note that each face
is isometric to $Y$ and $Y$ is one of the faces. We consider four copies
of the Sierpinski gasket $K$ (one of which is $K$ itself)
on alternating faces of the octahedron; so any two of the Sierpinski
gaskets intersect only at a single point.
Let $X$ be the union of the four Sierpinski gaskets.
Note that each point in $X$ has a neighborhood which is similar to
a neighborhood of the Sierpinski gasket. Thus $X$ is what is called
in \cite{Str_TAMS03} a fractafold (see also \cite{Str_Prin06} and
\cite{Str_CJM98}). We call $X$ the \emph{Sierpinski octafold}.
One can also think of the octafold as four copies of the Sierpinski
gasket such that any two of them are glued at one of the vertices
as in Figure \ref{Flo:figure}, where the dotted curves mean that the
two endpoints of the curve are identified.
\begin{figure}
\caption{The Sierpinski octafold}
\label{Flo:figure}
\includegraphics[scale=0.5]{Sierpinski_fractafold}
\end{figure}
The metric $\rho$ on $X$ is the restriction of the Euclidean metric
to $X$. Therefore the Hausdorff dimension of $X$ is $\log3/\log2$
(see \cite[Theorem 6.1.7]{Edg_MTFG2} and \cite[Theorem 6.5.4]{Edg_MTFG2}).
To describe the local homeomorphism, label the four copies of the
Sierpinski gasket, which we will call $0$-cells, as $1,2,3,$ and
$4$. Let $11,12,13$ be the three $1$-cells of the $0$-cell $1$,
and similarly for the other three $0$-cells (see Figure \ref{Flo:figure}).
The local homeomorphism $T$ we define is uniquely determined by the
following two properties
\begin{enumerate}
\item The vertices of the $0$-cells are left fixed.
\item The restriction to each $1$-cell defines an affine homeomorphism
onto the $0$-cell that shares exactly one vertex with the original
$1$-cell.
\end{enumerate}
\begin{comment}
We define the local homeomorphism $T$ such that the $1$-cell $11$
is mapped onto the $0$-cell $3$, the $1$-cell $12$ is mapped onto
the $0$-cell $4$, the $1$-cell $13$ is mapped onto the $0$-cell
$2$, the $1$-cell $21$ is mapped onto the $0$-cell $1$, the $1$-cell
$22$ is mapped onto the $0$-cell $4$, the $1$-cell $23$ is mapped
onto the $0$-cell $3$, the $1$-cell $31$ is mapped onto the $0$-cell
$1$, the $1$-cell $32$ is mapped onto the $0$-cell $2$, the $1$-cell
$33$ is mapped onto the $0$ cell $4$, the $1$-cell $41$ is mapped
onto the $0$-cell $1$, the $1$-cell $42$ is mapped onto the $0$-cell
$3$, and the $1$-cell $43$ is mapped onto the $0$-cell $2$. We
assume the all these mappings are similarities and the $6$ vertices
of the $4$ gaskets are fixed.
\end{comment}
{}It is enough to indicate where the points that are the vertices of
the $1$-cells and are not vertices of the $0$-cells map. We call
these points \emph{midpoints}. Each midpoint may be written $.5(\epsilon_{1}e_{\sigma(1)}+\epsilon_{2}e_{\sigma(2)})$
where $\epsilon_{i}=\pm1$, $e_{i}$ is a canonical basis element
of $\mathbb{R}^{3}$ and $\sigma$ is in the permutation group $S_{3}$.
Then \[
T\bigl(.5(\epsilon_{1}e_{\sigma(1)}+\epsilon_{2}e_{\sigma(2)})\bigr)=-\epsilon_{1}\epsilon_{2}e_{\sigma(3)}.\]
It is easy to see that this is well defined and $T$ is a local
homeomorphism
such that
\[
\lim_{y\to x}\frac{\rho(Tx,Ty)}{\rho(x,y)}=2.\]
It does not, however, satisfy condition (\ref{eq:0}) of Definition
\ref{def:loc-scaling} and, thus, it does not satisfy the local scaling
condition. To see this, consider a point $y$ in the $1$-cell $11$ and
another point $z$ in the $1$-cell $12$ such that they lie on the line
segments, that form an angle of $\pi/3$ and intersect at the midpoint
$x$ (where the two $1$-cells meet). We may also assume that they
are equidistant from the midpoint $x$ (see Figure \ref{Flo:figure}).
Their images lie on adjacent edges of a square, so the Pythagorean Theorem
implies that $\rho(Ty, Tz)=2\sqrt{2}\rho(y,z)$, while $\rho(Tx,Ty)=2\rho(x,y)$
and $\rho(Tx,Tz)=2\rho(x,z)$. Since we can take $y$ and $z$ as close as we
want to $x$, it follows that $T$ does not satisfy the local scaling
property.
We claim, however, that the conclusion of Theorem \ref{thm:Main} is still valid
for this example with $\varphi=\log 2$. Since the proofs of Proposition
\ref{pro:uniqueness} and Proposition \ref{pro:Entropy} do not depend on
the local scaling property and only on the existence of the KMS-state
$\omega_\mu$, it follows that they are also valid, once we prove the claim.
Let $G$ be the associated Deaconu-Renault groupoid. The action on
$C^{*}(G)$ is given via\[
\alpha_{t}(f)(x,m-n,y)=e^{-it(m-n)\log2}f(x,m-n,y).\]
Recall from \cite[Proposition
II.5.4]{Ren_LNM793} that in order for the Hausdorff measure $\mu$ to be a quasi-invariant
measure for $G$ the equality (\ref{eq:RN}) need only hold
almost everywhere. One can modify the proof of Theorem \ref{thm:Main}
to show that this is the case for this example. Let $C$ be any of the $1$-cells of the
octafold. Then the restriction of $T$ to $C$ is a similarity with ratio
$2$. Theorem 6.1.9 of \cite{Edg_MTFG2} implies that $\mu(TU)=2^\beta\mu(U)$ for
all subsets of $C$, where $\beta=\log 3/\log 2$ is the Hausdorff dimension. In
particular the equality is true for all open sections $U$ of $C$.
It follows
that the equation (\ref{eq:RN}) is true with $\varphi =\log 2$ for all points
except possibly the
midpoints and the conclusion of Theorem \ref{thm:Main} is true
for the octafold.
Thus, $\beta=\log3/\log2$
is the unique inverse temperature which admits a KMS state on $C^{*}(G)$
for $\alpha$, the $(\alpha,\beta)$-KMS state $\omega_{\mu}$ given
by \eqref{eq:KMSstate} is unique, and the topological entropy is $h(T)=\log3$.
\bibliographystyle{amsplain}
|
\section{Introduction}
Let $A$ be a Noetherian local ring with maximal ideal $\mathfrak m$ and $d
= \operatorname{dim} A > 0$. Let $\ell_A(M)$ denote, for an
$A$-module $M$, the length of $M$. Then, for each $\mathfrak m$-primary ideal
$I$ in $A$, we have integers $\{e_i(I)\}_{0 \le i \le d}$ such that
the equality
$$\ell_A(A/I^{n+1})={e}_0(I)\binom{n+d}{d}-
e_1(I)\binom{n+d-1}{d-1}+\cdots+(-1)^d{e}_d(I)$$
holds true for all integers $n \gg 0$. We call $\{e_i(I)\}_{0 \le i \le d}$ the Hilbert
coefficients of $A$ with respect to $I$.
These integers carry a great deal of information about the ideal $I$.
We will argue that $e_1(Q)$, for parameter ideals $Q$, codes
structural information about the ring $A$ itself. Noteworthy properties of $A$ associated to values of $e_1(Q)$ are the Cohen--Macaulay, the generalized Cohen--Macaulay and the Buchsbaum conditions.
We say that $A$ is unmixed,
if $\operatorname{dim} \widehat{A}/\mathfrak p = d$ for every $\mathfrak p \in
\operatorname{Ass} (\widehat{A})$, where $\widehat{A}$ denotes the
$\mathfrak m$-adic completion of $A$. With this notation Vasconcelos, exploring the vanishing of $e_1(Q)$ for parameter ideals
$Q$, posed the
following conjecture in his lecture at the conference in Yokohama in March 2008.
\begin{conj}[\cite{V}]\label{conj}
Assume that $A$ is unmixed. Then $A$ is a Cohen--Macaulay local ring,
once $e_1(Q) = 0$ for some parameter ideal $Q$ of $A$.
\end{conj}
In Section 2 of the present paper we shall settle Conjecture
\ref{conj} affirmatively (Theorem 2.1). Here we should note that Conjecture
\ref{conj} is already solved partially by \cite{GhHV} and \cite{MSV}.
In fact, Ghezzi, Hong, and Vasconcelos \cite[Theorem 3.3]{GhHV}
proved that the conjecture holds true, if $A$ is an integral domain
which is a homomorphic image of a Cohen--Macaulay ring. Mandal, Singh and
Verma \cite{MSV} proved that $e_1(Q) \le
0$ for every parameter ideal $Q$ in an arbitrary Noetherian local
ring $A$ and showed that $e_1(Q) < 0$, if $\operatorname{depth} A
= d-1$.
\medskip
\noindent {\bf Theorem 2.1} \;
Let $A$ be a Noetherian local ring with $d = \dim A > 0$ and let $Q$ be a parameter ideal in $A$. Then following are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $A$ is Cohen--Macaulay{\rm ;}
\item[{\rm (b)}] $A$ is {unmixed} and $e_1(Q)=0${\rm ;}
\item[{\rm (c)}] $A$ is {unmixed} and $e_1(Q) \geq 0$.
\end{enumerate}
\medskip
Let $\operatorname{Assh} (A) =\{ \mathfrak p \in \operatorname{Ass} (A) \mid
\operatorname{dim} A/\mathfrak p = d\}$ and let $(0) = \bigcap_{\mathfrak p \in
\operatorname{Ass} (A)}I(\mathfrak p)$ be a primary decomposition of
$(0)$ in $A$ with $\mathfrak p$-primary ideals $I(\mathfrak p)$ in $A$. We put
$$U_A(0) = \bigcap_{\mathfrak p \in \operatorname{Assh} (A)}
I(\mathfrak p)$$ and call it the {\it unmixed component} of $(0)$ in $A$.
Let us call those local rings $A$ with $e_1(Q) = 0$ for some
parameter ideal $Q$ of $A$ {\it Vasconcelos}\footnote{The terminology
is due to the first five authors} rings. In Section 3 we shall explore
basic properties of Vasconcelos rings. Certain sequentially
Cohen--Macaulay rings are good examples of Vasconcelos rings. A basic
characterization of some of these rings is:
\medskip
\noindent {\bf Theorem 2.7} \; Let $A$ be a Noetherian
local ring of dimension $d \geq 2$. Let $U=U_A(0)$ and $Q$ a parameter
ideal of $A$. Suppose that $A$ is a homomorphic image of a
Cohen--Macaulay ring. Then the following are equivalent:
\begin{enumerate}
\item[{\rm (a)}] $e_1(Q)=0$;
\item[{\rm (b)}] $A/U$ is Cohen--Macaulay and $\dim U \leq d-2$.
\end{enumerate}
Notice that unless $A$ is a homomorphic image of a Cohen--Macaulay
ring, the implication (a) $\Rightarrow$ (b) is not true in general
(Remark \ref{lech}).
In Section 4 we will study the problem of when $e_1(Q)$ is
independent of the choice of the parameter ideal $Q$ in
$A$. We shall show that $A$ is a quasi-Buchsbaum ring, if $A$ is
unmixed and $e_1(Q)$ is constant (Corollary \ref{q-Bbm}). The
authors conjecture that $A$ is furthermore a Buchsbaum ring, if $A$ is
unmixed and $e_1(Q)$ is independent of the choice of parameter
ideals $Q$ of $A$. We will show that this is the case, at least when
$e_1(Q)=-1$ or $-2$ (Theorem~\ref{constant}, Theorem~\ref{constant2}).
Goto and Ozeki \cite{GO} recently solved the conjecture affirmatively.
Another important issue is that of the variability of $e_1(Q)$,
sometimes for $Q$ in a same integral closure class, and its role in
the structure of the ring. This will be pursued in a sequel paper.
In what follows, unless otherwise specified, let $(A,\mathfrak m)$ denote a
Noetherian local ring with maximal ideal $\mathfrak m$ and $d =
\operatorname{dim} A$. Let $\{\mathrm{H}_{\mathfrak m}^i(*)\}_{i \in \mathbb Z}$ be the
local cohomology functors of $A$ with respect to the maximal ideal
$\mathfrak m$.
\section{The vanishing conjecture}
The purpose of this section is to prove the following, which settles Conjecture \ref{conj} affirmatively.
Throughout let $A=(A, \mathfrak m)$ be a Noetherian local ring with maximal ideal $\mathfrak m$ and $d=\dim A$.
\begin{thm}\label{2.1} Let $A$ be a Noetherian local ring with $d = \dim A > 0$ and let $Q$ be a parameter ideal in $A$. Then following are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $A$ is Cohen--Macaulay{\rm ;}
\item[{\rm (b)}] $A$ is {unmixed} and $e_1(Q)=0${\rm ;}
\item[{\rm (c)}] $A$ is {unmixed} and $e_1(Q) \geq 0$.
\end{enumerate}
\end{thm}
In our proof of Theorem \ref{2.1} the following facts are the key. See {\cite[Section 3]{GNa}} for the proof.
\begin{prop}[\cite{GNa}]\label{GNa}
Let $(A,\mathfrak{m})$ be a Noetherian local ring with $d = \dim A \ge 2$, possessing the canonical module $\mathrm{K}_A$. Assume that $\dim A/\mathfrak{p} = d$ for every $\mathfrak{p} \in \mathrm{Ass} (A) \setminus \{\mathfrak{m} \}$. Then the following assertions hold true.
\begin{enumerate}
\item[{\rm (a)}] The local cohomology module $\mathrm{H}_{\mathfrak{m}}^1(A)$ is finitely generated.
\item[{\rm (b)}] The set $\mathcal{F} = \{ \mathfrak{p} \in \mathrm{Spec} A \mid \dim A_{\mathfrak{p}} > \mbox{\rm depth} (A_{\mathfrak{p}}) = 1 \}$ is finite.
\item[{\rm (c)}] Suppose that the residue class field $k=A/\mathfrak{m}$ of $A$ is infinite and let $I $ be an $\mathfrak{m}$-primary ideal in $A$. Then one can choose an element $a \in I$ so that $a$ is superficial for $I$ and $\dim A/\mathfrak{p} = d - 1$ for every $\mathfrak{p} \in \mathrm{Ass}_A (A/aA) \setminus \{\mathfrak{m} \}$.
\end{enumerate}
\end{prop}
\begin{rem}\label{U(A)}
Let $(A, \mathfrak m)$ be a Noetherian local ring with $\dim A=d>0$. Recall that the unmixed component of $(0)$ in $A$ is
$U_A(0) = \bigcap_{\mathfrak p \in \operatorname{Assh} (A)}
I(\mathfrak p)$. Since $\mathrm{H}^0_{\mathfrak m}(A)= \bigcap_{\mathfrak p \in \mathrm{Ass}(A) \setminus \{ \mathfrak m \}}I(\mathfrak p)$, we have that $\mathrm{H}^0_{\mathfrak m} (A) \subseteq U_A(0)$. If $\mathrm{Ass}(A) \setminus \{\mathfrak m\} = \mathrm{Assh}(A)$,
then $\mathrm{H}^0_{\mathfrak m} (A) = U_A(0)$.
\end{rem}
\begin{proof}[Proof of Theorem $\ref{2.1}$] (a) $\Rightarrow$ (b) $\Rightarrow$ (c) are clear. In order to show (c) $\Rightarrow$ (a), we may assume that $A$ is a complete unmixed local ring with $d\ge 2$ and infinite residue field. Let $Q=(a_1, \ldots, a_d)$.
We use induction on $d$.
\medskip
\noindent Let $d=2$. Then $Q=(a_1, a_2)$, where we may assume that $a=a_1$ is a superficial element.
Let $S=A/aA$ and let $q = QS$. Then $\dim S=1$ and by \cite[Lemma 2.2]{GNi} we have
\[ - \ell_A(\mathrm{H}^0_{\mathfrak m}(S)) = e_1(q) = e_1(Q)- \ell_A(0:_A a )= e_1(Q) \geq 0. \]
Hence $\mathrm{H}^0_{\mathfrak m}(S)=(0)$. Therefore $S$ is Cohen--Macaulay and so is $A$.
\medskip
\noindent Suppose $d \geq 3$. Then there exists $a \in Q$ such that $\mathrm{Ass}(A/aA) \subseteq \mathrm{Assh}(A/aA) \cup \{ \mathfrak m \}$ (Proposition~\ref{GNa} (c)). Let $S=A/aA$ and $q=QS$. Note that $S$ is not necessarily unmixed. Let $U=U_S(0)$,
$\overline{S}=S/U$ and $\overline{q}=q \overline{S}$. Then $\overline{S}$ is unmixed of dimension $d-1$.
Since $e_1(\overline{q})=e_1(q)=e_1(Q) \geq 0$, by the induction hypothesis $\overline{S}$ is Cohen--Macaulay, i.e., $\mathrm{H}_{\mathfrak m}^{i}(\overline{S}) =(0)$ for all $0 \leq i \leq d-2$.
\smallskip
\noindent From the exact sequence $0 \rightarrow U \rightarrow S \rightarrow \overline{S} \rightarrow 0$, we get a long exact sequence
\[ \begin{array}{lll}
0 &\longrightarrow & \mathrm{H}^{0}_{\mathfrak m}(U) \longrightarrow \mathrm{H}^{0}_{\mathfrak m}(S) \longrightarrow \mathrm{H}^{0}_{\mathfrak m}(\overline{S}) \longrightarrow \mathrm{H}^{1}_{\mathfrak m}(U) \longrightarrow \mathrm{H}^{1}_{\mathfrak m}(S) \longrightarrow \mathrm{H}^{1}_{\mathfrak m}(\overline{S}) \longrightarrow \cdots \\ && \\
&\longrightarrow & \mathrm{H}^{d-2}_{\mathfrak m}(U) \longrightarrow \mathrm{H}^{d-2}_{\mathfrak m}(S) \longrightarrow \mathrm{H}^{d-2}_{\mathfrak m}(\overline{S}) .
\end{array}\]
Since $\mathrm{H}_{\mathfrak m}^{0}(S)=U$ (Remark~\ref{U(A)}), we have $\mathrm{H}^{i}_{\mathfrak m}(U)=(0)$ for all $i \geq 1$. Therefore
\[\mathrm{H}^{i}_{\mathfrak m}(S)=(0), \quad \mbox{\rm for all} \;\; 1 \leq i \leq d-2. \]
\smallskip
\noindent From the exact sequence $0 \rightarrow A \stackrel{\cdot a}{\longrightarrow} A \rightarrow S \rightarrow 0$, we get
\[\begin{array}{lll}
0 &\longrightarrow & \mathrm{H}^{0}_{\mathfrak m}(A) \longrightarrow \mathrm{H}^{0}_{\mathfrak m}(A) \longrightarrow \mathrm{H}^{0}_{\mathfrak m}(S) \longrightarrow \mathrm{H}^{1}_{\mathfrak m}(A) \stackrel{\cdot a}{\longrightarrow}
\mathrm{H}^{1}_{\mathfrak m}(A) \longrightarrow 0 \longrightarrow \cdots \\ && \\
&\longrightarrow & 0 \rightarrow \mathrm{H}^{i}_{\mathfrak m}(A) \stackrel{\cdot a}{\longrightarrow} \mathrm{H}^{i}_{\mathfrak m}(A) \rightarrow 0 \rightarrow \cdots \rightarrow \mathrm{H}^{d-2}_{\mathfrak m}(S)=0 \rightarrow \mathrm{H}^{d-1}_{\mathfrak m}(A) \stackrel{\cdot a}{\longrightarrow} \mathrm{H}^{d-1}_{\mathfrak m}(A) .
\end{array}\]
\smallskip
\noindent Since $A$ has a nonzero divisor,
$\mathrm{H}^{0}_{\mathfrak m}(A)=(0)$.
The epimorphism $\mathrm{H}^{1}_{\mathfrak m}(A) \rightarrow \mathrm{H}^{1}_{\mathfrak m}(A) \rightarrow 0$ implies $\mathrm{H}^{1}_{\mathfrak m}(A) =a \mathrm{H}^{1}_{\mathfrak m}(A) $. Since $\mathrm{H}^{1}_{\mathfrak m}(A) $ is finitely generated (Proposition~\ref{GNa} (a)), $\mathrm{H}^{1}_{\mathfrak m}(A) =(0)$.
For $2 \leq i \leq d-1$, we obtain
$\mathrm{H}^{i}_{\mathfrak m}(A) =(0)$ because for every $x \in \mathrm{H}^{i}_{\mathfrak m}(A)$ some power of $a$ annihilates $x$.
\end{proof}
Let us discuss some consequences of Theorem \ref{2.1}.
\begin{lem}\label{lemma}
Let $A$ be a Noetherian local ring of dimension $d > 0$. Let $Q$ be a parameter ideal of $A$.
Suppose that $U=U_A(0) \neq (0)$.
Let $C = A/U$. Then the following assertions hold true.
\begin{enumerate}
\item[{\rm (a)}] $\dim U < \dim A$.
\item[{\rm (b)}] We have
\[
e_1(Q) = \left\{\begin{array}{ll}
e_1(QC) &\quad \mbox{\rm if} \;\; \dim U \leq d-2 \\& \\
e_1(QC)-s_0 &\quad \mbox{\rm if} \;\; \dim U=d-1,
\end{array}
\right.
\]where $s_0 \geq 1$ is the multiplicity of ${\displaystyle \bigoplus_n U/(Q^{n+1}\cap U)}$.
\item[{\rm (c)}] $e_1(Q) \leq e_1(QC)$ with equality if and only if $\dim U \leq d-2$.
\end{enumerate}
\end{lem}
\begin{proof} (b) We write
$$\ell_A(U/(Q^{n+1}\cap U)) = s_0\binom{n + t }{t} -s_1\binom{n + t-1}{t-1} + \cdots + (-1)^ts_t$$
for $n \gg 0$ with integers $\{s_i\}_{0 \le i \le t}$, where $t=\dim U$.
Then the claim follows from
the exact sequence $0 \rightarrow U \rightarrow A \rightarrow C \rightarrow 0$ of $A$--modules, which gives
\[ \ell_A(A/Q^{n+1}) = \ell_A(C/Q^{n+1}C) + \ell_A(U/(Q^{n+1}\cap U)) \]
for all $n \ge 0$.
\medskip
\noindent (c) It follows from (b) and the fact that $s_0 \ge 1$.
\end{proof}
The following results are due to \cite{MSV}. We include an independent proof.
\begin{cor}[{\cite{MSV}}]\label{cor} Let $A$ be a Noetherian local ring of dimension $d >0$. Let $Q$ be a parameter ideal of $A$. Then the following assertions hold true.
\begin{enumerate}
\item[{\rm (a)}] $e_1(Q) \le 0$.
\item[{\rm (b)}] If $\mbox{\rm depth}(A) =d-1$, then $e_1(Q) < 0$.
\end{enumerate}
\end{cor}
\begin{proof}
(a) We may assume that $A$ is complete. Let $U=U_A(0)$ and $C=A/U$. Then by Lemma~\ref{lemma},
$e_1(Q) \le e_1(QC)$. Hence we may also assume that $A$ is unmixed. Hence the claim follows from Theorem \ref{2.1}.
\medskip
\noindent (b) We may assume that the residue field $A/\mathfrak m$ is infinite. If $d=1$, by \cite[Lemma 2.4 (1)]{GNi}, we have
$e_1(Q)= - \ell_A(H^0_{\mathfrak m}(A)) < 0$,
where the last inequality follows from the fact that $\mbox{\rm depth}(A)=0$.
Suppose that $d \geq 2$. Let $a_1,\dots, a_{d-1}\in Q$ be a superficial sequence and let $\overline{A}=A/(a_1,\dots, a_{d-1})$, $\overline{Q}=Q/(a_1,\dots, a_{d-1})$. Since $\mbox{\rm depth}(A)=d-1$, we have that
$e_1(Q)= e_1(\overline{Q})=- \ell_A(H^0_{\mathfrak m}(\overline{A})) < 0$.
\end{proof}
\begin{prop}\label{1.2-1}
Let $A$ be a Noetherian local ring of dimension $d \geq 2$. Let $U=U_A(0)$.
Suppose that $A/U$ is Cohen--Macaulay and $\dim U \leq d-2$. Then
$e_1(Q)=0$ for every parameter ideal $Q$ of $A$.
\end{prop}
\begin{proof} Let $C=A/U$. Since $\dim U \leq d-2$, we get $e_1(Q)=e_1(QC)$ (Lemma~\ref{lemma}). Since $C$ is Cohen--Macaulay, we have $e_1(QC)=0$ (Theorem~\ref{2.1}), which completes the claim.
\end{proof}
\begin{thm}\label{1.2-2}
Let $A$ be a Noetherian local ring of dimension $d \geq 2$. Suppose that $A$ is a homomorphic image of a Cohen--Macaulay ring. Let $U=U_A(0)$ and let $Q$ be a parameter ideal of $A$.
Then the following are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $e_1(Q)=0${\rm ;}
\item[{\rm (b)}] $A/U$ is Cohen--Macaulay and $\dim U \leq d-2$.
\end{enumerate}
\end{thm}
\begin{proof
It is enough to prove that (a) $\Rightarrow$ (b). Let $C = A/U$. Then $C$ is an unmixed local ring because $C$ is a homomorphic image of a Cohen--Macaulay ring and $\operatorname{dim} C/P = d$ for all $P \in \operatorname{Ass} (C)$. If $U =(0)$, then $A$ is unmixed. Therefore $A$ is Cohen--Macaulay by Theorem~\ref{2.1}.
Suppose that $U \ne (0)$. We show that $\dim U \leq d-2$. Suppose not, i.e., $\dim U = d-1$. By Lemma~\ref{lemma}, we get
\[ e_1(Q) = e_1(QC) - s_0, \]where $s_0 \geq 1$. On the other hand, $e_1(QC) \leq 0$ by Corollary~\ref{cor} (a). Hence $e_1(Q) <0$, which is a contradiction. Therefore $\dim U \leq d-2$. Now Lemma~\ref{lemma} implies that $e_1(QC) = e_1(Q) = 0$. By Theorem~\ref{2.1}, $C$ is a Cohen--Macaulay ring.
\end{proof}
The following corollary gives a characterization of Cohen-Macaulayness.
\begin{cor}
Let $A$ be a Noetherian local ring of dimension $d > 0$. Let $Q$ be a parameter ideal in $A$. Suppose that $e_i(Q) = 0$ for all $1 \le i \le d$. Then $A$ is a Cohen--Macaulay ring.
\end{cor}
\begin{proof}
We may assume that $A$ is complete. Let $U = U_A(0)$. By Theorem~\ref{2.1} it is enough to show that $U=(0)$.
Suppose that $U \neq (0)$ and let $C=A/U$. Since $e_1(Q)=0$ we have that $C$ is Cohen--Macaulay and $\dim U \leq d-2$ (Theorem \ref{1.2-2}).
From the exact sequence $0 \rightarrow U \rightarrow A \rightarrow C \rightarrow 0$ of $A$-modules,
we get
\[ \ell_A(A/Q^{n+1}) = \ell_A(C/Q^{n+1}C) + \ell_A(U/(Q^{n+1}\cap U)) \]
for all $n \ge 0$. By assumption, for $n \gg 0$ we have
\[ \ell_A(A/Q^{n+1}) = e_0(Q) \binom{n+d}{d}.\] Also since $C$ is Cohen--Macaulay, for $n \gg 0$ we have
\[ \ell_A(C/Q^{n+1}C) = e_0(QC) \binom{n+d}{d}.\] Since $e_0(Q)=e_0(QC)$, for $n \gg 0$ we obtain
\[\ell_A(U/(Q^{n+1}\cap U))=0,\] i.e., $U \subseteq Q^{n+1}$ for $n \gg 0$.
Thus $U = (0)$, which is a contradiction.
\end{proof}
The implication (a) $\Rightarrow$ (b) in Theorem \ref{1.2-2} is not true in general without the assumption that $A$ is a homomorphic image of a Cohen--Macaulay ring.
\begin{rem}\label{lech}
Let $R$ be a Noetherian equi-characteristic complete local ring and assume that $\operatorname{depth} (R) >0$.
Then one can construct a Noetherian local integral domain $(A,\mathfrak m)$ such that $R = \widehat{A}$, where $\widehat{A}$ denotes the $\mathfrak m$-adic completion of $A$ (\cite{L}). For example, let $k[[X,Y,Z,W]]$ be the formal power series ring over a field $k$ and consider the local ring $$R = k[[X, Y, Z, W]]/(X) \cap (Y,Z,W).$$ We can choose a Noetherian local integral domain $(A,\mathfrak m)$ so that $R = \widehat{A}$. Then $e_1(Q)= 0$ for every parameter ideal $Q$ in $A$, since $e_1(Q) = e_1(QR) = 0$. But $A$ is not Cohen--Macaulay because $R= \widehat{A}$ is not Cohen--Macaulay.
\end{rem}
\section{Vasconcelos rings}
Throughout this section let $A=(A, \mathfrak m)$ be a Noetherian local ring with maximal ideal $\mathfrak m$ and $d = \dim A$.
Our purpose is to develop a theory of Vasconcelos rings. Let us begin with the definition.
\begin{defn}
A Noetherian local ring $A$ is a {\em Vasconcelos} ring, if either $\dim A = 0$, or $\dim A > 0$ and $e_1(Q) = 0$ for some parameter ideal $Q$ in $A$.
\end{defn}
A Cohen--Macaulay local ring is a Vasconcelos ring.
\begin{prop}\label{Vring} Let $(A, \mathfrak m)$ be a Noetherian local ring with $\dim A =d$.
\begin{enumerate}
\item[{\rm (a)}] A $1$--dimensional Vasconcelos ring is Cohen--Macaulay.
\item[{\rm (b)}] An unmixed Vasconcelos ring is Cohen--Macaulay.
\item[{\rm (c)}] Suppose that $d \geq 2$. Then $A$ is a Vasconcelos ring if and only if $A/\mathrm{H}^0_{\mathfrak m}(A)$ is a Vasconcelos ring.
\end{enumerate}
\end{prop}
\begin{proof} (a) There exists a parameter ideal $Q$ of $A$ such that $e_1(Q)=0$. By \cite[2.4 (1)]{GNi}, we have
$0= e_1(Q) = -\ell_A(\mathrm{H}^0_{\mathfrak m}(A))$,
which shows that $A$ is Cohen--Macaulay.
\noindent (b) It follows from Theorem~\ref{2.1}.
\noindent (c) Let $B = A/\mathrm{H}_\mathfrak m^0(A)$. Then $e_1(QB) = e_1(Q)$ for every parameter ideal $Q$ in $A$. Hence $A$ is a Vasconcelos ring if and only if $B$ is a Vasconcelos ring.
\end{proof}
Let $A$ be a Noetherian local ring of dimension $d \geq 1$ and let $M$ be a finite $A$--module.
We denote
\[\mathrm{Assh}(M) = \{ \mathfrak p \in \mathrm{Ass}(M) \mid \dim A/\mathfrak p = \dim M \}. \]
Let ${ (0_M) = \bigcap_{\mathfrak p \in \mathrm{Ass}(M)} M(\mathfrak p) }$ be a primary decomposition, where $M(\mathfrak p)$ is $\mathfrak p$--primary.
The {\em unmixed component} $U_M(0)$ of $(0)$ in $M$ is
\[ U_M(0) = \bigcap_{\mathfrak p \in \mathrm{Assh}(M)} M(\mathfrak p). \]
Note that for any $\mathfrak p \in \mathrm{Assh}(M)$, we have $U_M(0)_{\mathfrak p} = (0)$.
\begin{lem}\label{U(M)}
Let $A$ be a Noetherian local ring of dimension $d \geq 1$. Let $M$ be a finite $A$--module and $L$ a submodule of $M$. Suppose that $\dim L \leq \dim M -1$ and $\mathrm{Ass}(M/L) \subseteq \mathrm{Assh}(M)$.
Then $L=U_M(0)$.
\end{lem}
\begin{proof}
Let $\dim M =g$ and $U=U_M(0)$. Since $\dim L \leq g-1$, we have $L_{\mathfrak p}=(0)$ for every $\mathfrak p \in \mathrm{Assh}(M)$.
Notice that for each $\mathfrak p \in \mathrm{Assh}(M)$ there exists $s \in A \setminus \mathfrak p$ such that $s L =(0)$.
Then $L \subseteq M(\mathfrak p)$ for every $\mathfrak p \in \mathrm{Assh}(M)$ because $s$ is a non zero divisor on $M/M(\mathfrak p)$ and
$M(\mathfrak p)$ is $\mathfrak p$--primary. This means that
\[ L \subseteq \bigcap_{\mathfrak p \in \mathrm{Assh}(M)} M(\mathfrak p) = U.\]
Now suppose that $L \subsetneq U$. Then exists $\mathfrak q \in \mathrm{Ass}(U/L)$. By assumption, $\mathfrak q \in \mathrm{Assh}(M)$. Hence $L_{\mathfrak q} =(0)=U_{\mathfrak q}$, so that $(U/L)_{\mathfrak q}=(0)$, which is a contradiction.
\end{proof}
Here is a basic characterization of Vasconcelos rings.
\begin{thm}\label{2.5} Let $(A, \mathfrak m)$ be a Noetherian local ring of dimension $d \geq 2$. Let $\widehat{A}$ be the $\mathfrak m$-adic completion of $A$.
The following are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $A$ is a Vasconcelos ring{\rm ;}
\item[{\rm (b)}] $e_1(Q) = 0$ for every parameter ideal $Q$ in $A${\rm ;}
\item[{\rm (c)}] $\widehat{A}/U_{\widehat{A}}(0)$ is a Cohen--Macaulay ring and $\operatorname{dim}_{\widehat{A}} U_{\widehat{A}}(0) \le d-2${\rm ;}
\item[{\rm (d)}] There exists a proper ideal $I$ of $\widehat{A}$ such that $\widehat{A}/I$ is a Cohen--Macaulay ring of dimension $d$ and $\operatorname{dim}_{\widehat{A}}I \le d-2$.
\end{enumerate}
When this is the case, $\widehat{A}$ is a Vasconcelos ring, $\mathrm{H}_{\mathfrak m}^{d-1}(A) = (0)$, and the canonical module $K_{\widehat{A}}$ of $\widehat{A}$ is a Cohen--Macaulay $\widehat{A}$--module.
\end{thm}
\begin{proof} (c) $\Rightarrow$ (b) follows from Proposition~\ref{1.2-1} and (a) $\Rightarrow$ (c) follows from Theorem~\ref{1.2-2}. The implications (b) $\Rightarrow$ (a) and (c) $\Rightarrow$ (d) are trivial.
Finally (d) $\Rightarrow$ (c) follows from Lemma~\ref{U(M)}.
To see the last assertions,
let $U=U_{\widehat{A}}(0)$ and let $\widehat{\mathfrak m} = \mathfrak m \widehat{A}$ be the maximal ideal of $\widehat{A}$. Then $\widehat{A}/U$ is a Cohen--Macaulay ring and $\operatorname{dim}_{\widehat{A}} U \le d-2$. In particular, $\mathrm{H}_{\widehat{\mathfrak m}}^{d-1}(\widehat{A}/U) =(0) = \mathrm{H}_{\widehat{\mathfrak m}}^{d-1}(U)$. Hence $\mathrm{H}_{\widehat{\mathfrak m}}^{d-1}(\widehat{A}) = (0)$. Moreover we have $\mathrm{H}_{\widehat{\mathfrak m}}^{d}(\widehat{A}) \cong \mathrm{H}_{\widehat{\mathfrak m}}^{d}(\widehat{A}/U)$, which means that $\mathrm{K}_{\widehat{A}} \cong \mathrm{K}_{\widehat{A}/U}$. Since $\mathrm{K}_{\widehat{A}/U}$ is Cohen--Macaulay, so is $\mathrm{K}_{\widehat{A}}$.
\end{proof}
\begin{cor}\label{cor4.3}
Let $A$ be a Vasconcelos ring of dimension $d \geq 1$. If $x$ is a nonzerodivisor in $A$, then $A/xA$ is a Vasconcelos ring.
\end{cor}
\begin{proof}
We may assume that $A$ is complete and that $d\geq 2$. Let $U = U_A(0)$ and $C = A/U$.
By Theorem~\ref{2.5}, $C$ is a $d$--dimensional Cohen--Macaulay ring and $\dim U \leq d-2$. Then $\dim U/xU \leq d-3$. Since $x$ is $C$--regular, $C/xC$ is Cohen--Macaulay. By Theorem~\ref{2.5}, $A/xA$ is a Vasconcelos ring.
\end{proof}
\begin{ex}\label{2.7}
Let $(R, \mathfrak n)$ be a Cohen--Macaulay local ring of dimension $d \geq 3$. Let $x$ be a nonzerodivisor of $R$ and let $\mathfrak a$ be an $R$--ideal such that $\mathrm{ht}(\mathfrak a) \ge 3$ and $x \not\in \mathfrak a$. Let $A= R/(xR \cap \mathfrak a)$.
Then $A$ is a non--Cohen--Macaulay Vasconcelos ring of dimension $d-1 \geq 2$.
\end{ex}
\begin{proof}
Consider the exact sequence $0 \to R/(\mathfrak a: x) \to A \to R/xR \to 0$. Then $A/xA$ is Cohen--Macaulay. Also we have
\[\dim_A(xA) = \dim R/(\mathfrak a : x) \leq \dim R/\mathfrak a \leq d-3 = \dim A-2.\]
By Theorem~\ref{2.5}, $A$ is a non-Cohen--Macaulay Vasconcelos ring with $\dim A = d-1 \geq 2$.
\end{proof}
\begin{ex}\label{ex2.7}
Let $(R, \mathfrak n)$ be a Cohen--Macaulay local ring with $\dim R = d \geq 2$. Let $M$ be a finitely generated $R$-module with $\dim_RM \le d-2$. Then the idealization $A= R \ltimes M$ is a Vasconcelos ring with $\dim A = d$. Also we have $\mathrm{Assh}(A) = \mathrm{Min}(A)$. Moreover if $R=k [[x,y]]$ and $M=R/(x,y)^2$, then $A=R \ltimes M$ is not quasi--Buchsbaum.
\end{ex}
\begin{proof}
Let $N =(0) \times M$. Consider the exact sequence $0 \to N \to A \overset{\varepsilon}{\to} R \to 0$, where
$\varepsilon (x) = r$ for each $x=(r,m) \in A$. Then $A/N \simeq R$ is Cohen--Macaulay. Also $\dim_AN = \dim_RM \le d-2$. By Theorem~\ref{2.5}, $A$ is a Vasconcelos ring with $\dim A = d$. Moreover we have
\[ \mathrm{Assh}(A) = \{\mathfrak p \times M \mid \mathfrak p \in \mathrm{Assh}(R) \} = \mathrm{Min}(A).\]
\end{proof}
We are now interested in the question of how Vasconcelos rings are preserved under flat base changes.
\begin{thm}\label{thm4.6}
Let $(A, \mathfrak m)$ and $(B, \mathfrak n)$ be Noetherian local rings and let $\varphi : A \to B$ be a flat local homomorphism.
Then the following assertions hold true.
\begin{enumerate}
\item[{\rm (a)}] Suppose that $A$ is a Vasconcelos ring and $B/\mathfrak m B$ is a Cohen--Macaulay ring. Then $B$ is a Vasconcelos ring.
\item[{\rm (b)}] Suppose that $B$ is a homomorphic image of a Cohen--Macaulay ring. If $B$ is a Vasconcelos ring and $\mathrm{Ass}_B (B/\mathfrak p B) = \mathrm{Assh}_B(B/\mathfrak p B)$ for every $\mathfrak p \in \mathrm{Assh}(A)$, then
$A$ is a Vasconcelos ring and $B/\mathfrak m B$ is a Cohen--Macaulay ring.
\end{enumerate}
\end{thm}
\begin{proof}
Let $U=U_A(0)$.
Notice that $(A/U) \otimes_A B$ is Cohen--Macaulay if and only if both $A/U$ and $B/\mathfrak m B$ are Cohen--Macaulay.
\medskip
\noindent (a)
We may assume that $A$ is complete and $\dim A=d>0$. Since $A$ is a Vasconcelos ring, by Theorem~\ref{2.5} $A/U$ is Cohen--Macaulay and $\dim_A U \leq d-2$. Since $B/\mathfrak m B$ is also Cohen--Macaulay by assumption, $B/UB \simeq (A/U) \otimes_A B$ is Cohen--Macaulay. Moreover we have
\[ \dim_B (UB) = \dim_B (U \otimes_A B) = \dim_A U + \dim_B (B/\mathfrak m B) \leq d-2 + \dim_B (B/\mathfrak m B) = \dim B-2.\]
Hence by Proposition~\ref{1.2-1}, $B$ is a Vasconcelos ring.
\medskip
\noindent (b) Suppose that $A$ is Cohen--Macaulay. It is enough to show that $B$ is Cohen--Macaulay.
Let $\mathfrak P \in \mathrm{Ass}(B)$ and $\mathfrak p = \mathfrak P \cap A$.
Then $\mathfrak p \in \mathrm{Ass}(A) = \mathrm{Assh}(A)$. We have
$\mathfrak P \in \mathrm{Ass}(B/\mathfrak p B) = \mathrm{Assh}(B/\mathfrak p B)$ by assumption. This means that
\[
\dim B/\mathfrak P = \dim B/\mathfrak p B = \dim_A (A/ \mathfrak p) + \dim_B (B/\mathfrak m B) = d+ \dim_B (B/\mathfrak m B)= \dim B.
\]
Therefore $ \mathrm{Ass}(B) = \mathrm{Assh}(B)$, which shows that $B$ is an unmixed Vasconcelos ring.
By Proposition~\ref{Vring} (b), $B$ is Cohen--Macaulay.
Now suppose that $A$ is not Cohen--Macaulay. We claim that $UB=U_B(0)$, the unmixed component of $(0)$ in $B$. Let $\mathfrak P \in \mathrm{Ass}(B/ UB)$. Then $\mathfrak P \in \mathrm{Ass}(B/\mathfrak p B)$ for some $\mathfrak p \in \mathrm{Ass}(A/U)=\mathrm{Assh}(A)$. By assumption, we have $\mathfrak P \in \mathrm{Assh}(B/\mathfrak p B)$. This means that $\dim B/\mathfrak P = \dim B/\mathfrak p B = \dim B$. Hence $\mathrm{Ass}(B/UB)= \mathrm{Assh}(B)$, which proves the claim.
Now since $B$ is a Vasconcelos ring, by Theorem~\ref{1.2-2} we have that $B/UB$ is Cohen--Macaulay and $\dim_B UB \leq \dim B-2$. Therefore both $A/U$ and $B/\mathfrak m B$ are Cohen--Macaulay because $(A/U) \otimes_A B \simeq B/UB$ is Cohen--Macaulay. Also $\dim_B UB \leq \dim B-2$ implies that $\dim U \leq d-2$. In particular, $A$ is a Vasconcelos ring by Proposition~\ref{1.2-1}.
\end{proof}
\begin{cor}
Let $(A,\mathfrak m)$ be a Vasconcelos ring and let $R = A[X_1, X_2, \cdots, X_n]$ be the polynomial ring. Then $R_P$ is a Vasconcelos ring for every $P \in \mathrm{V}(\mathfrak m R)$.
\end{cor}
\begin{cor} Let $(A,\mathfrak m)$ and $(B,\mathfrak n)$ be Noetherian local rings which are homomorphic images of Cohen--Macaulay rings. Let $\varphi : A \to B$ be a flat local homomorphism. Then the following two conditions are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $A$ is a Vasconcelos ring and $B/\mathfrak m B$ is a Cohen--Macaulay ring{\rm ;}
\item[{\rm (b)}] $B$ is a Vasconcelos ring and $\operatorname{Ass}_B(B/\mathfrak p B) = \operatorname{Assh}_B(B/\mathfrak p B)$ for every $\mathfrak p \in \operatorname{Assh} (A)$.
\end{enumerate}
\end{cor}
\begin{proof}
(a) $\Rightarrow$ (b) Let $\mathfrak p \in \operatorname{Assh}(A)$. Then, since $\mathfrak p \in \operatorname{Ass}(A/U)$, we have that $\operatorname{Ass}_B (B/\mathfrak p B) \subseteq \operatorname{Ass}_B (B \otimes_AA/U)$. Hence $\operatorname{dim} B/\mathfrak P = \operatorname{dim} B$ for every $\mathfrak P \in \operatorname{Ass}_B(B/\mathfrak p B)$, since $B \otimes_AA/U$ is a Cohen--Macaulay ring with $\operatorname{dim} (B\otimes_AA/U) = \operatorname{dim} B$.
\end{proof}
Next we show that quasi-unmixed Vasconcelos rings behave well under localization.
\begin{prop}
Let $A$ be a Noetherian local ring of dimension $d$. Suppose that $A$ is a homomorphic image of a Cohen--Macaulay ring and $\operatorname{Assh} (A) = \operatorname{Min} (A)$. If $A$ is a Vasconcelos ring, then $A_\mathfrak p$ is a Vasconcelos ring for every $\mathfrak p \in \operatorname{Spec} A$.
\end{prop}
\begin{proof}
We may assume that $A$ is not a Cohen--Macaulay ring. Let $U = U_A(0)$. Then $U \ne (0)$ by Theorem \ref{1.2-2}. Let $\mathfrak p \in \operatorname{Spec} A$. Notice that $U\subseteq \mathfrak p$. Since $(A/U)_\mathfrak p$ is a Cohen--Macaulay ring we may assume that $U_\mathfrak p \ne (0)$. Let $\mathfrak a = (0) : U$. Then $\mathfrak a \ne A$ and $\operatorname{ht}_A\mathfrak a \ge 2$, since $\dim U \le d-2$ and $A$ is catenary and equidimensional. Hence $\operatorname{ht}_{A_\mathfrak p}\mathfrak a A_\mathfrak p \ge 2$ and $\operatorname{dim} A_\mathfrak p / \mathfrak a A_\mathfrak p \le \operatorname{dim}A_\mathfrak p - 2.$ Therefore $A_\mathfrak p$ is a Vasconcelos ring by Theorem \ref{1.2-2}.
\end{proof}
\begin{cor}
Suppose that $A$ is a Vasconcelos ring and $\operatorname{Assh}(\widehat{A}) = \operatorname{Min}(\widehat{A})$. Then $\widehat{A}_\mathfrak p$ is a Vasconcelos ring for every $\mathfrak p \in \operatorname{Spec}\widehat{A}$.
\end{cor}
Suppose that $d > 0$ and let $Q$ be a parameter ideal in $A$. We
denote by $R = \mathcal{R} (Q)= A[Qt]$ (resp. $G = \mathrm{G}(Q))$ the Rees
algebra (resp. the associated graded ring) of $Q$. Let $\mathfrak{M} = \mathfrak m R + R_+$ be the graded maximal ideal in $R$.
\begin{thm}
Let $A$ be a Noetherian local ring with dimension $d>0$. With the above notation we have the following.
\begin{enumerate}
\item[{\rm (a)}] $A$ is a Vasconcelos ring if and only if $G_\mathfrak{M}$ is a Vasconcelos ring.
\item[{\rm (b)}] Suppose that $A$ is a homomorphic image of a Cohen--Macaulay ring. If $A$ is a Vasconcelos ring, then $R_\mathfrak{M}$ is a Vasconcelos ring.
\end{enumerate}
\end{thm}
\begin{proof}
(a) Let $Q = (a_1, a_2, \cdots, a_d)$ be a parameter ideal in $A$, and let $f_i = a_i t$ for each $1 \le i \le d$. Then $G_+ = (f_1, f_2, \cdots, f_d)G$ and $f_1, f_2, \cdots, f_d$ forms a linear system of parameters in the graded ring $G$. We furthermore have $$\ell_G(G/[(f_1, f_2, \cdots, f_d)G]^{n+1}) = \ell_A(A/Q^{n+1})$$ for all $n \ge 0$. Hence $e_1(Q) = e_1((f_1, f_2, \cdots, f_d)G_\mathfrak{M})$ and the conclusion follows from Theorem \ref{2.5}.
(b) We may assume that $A$ is not a Cohen--Macaulay ring. Let $U =
U_A(0)$ and $B = A/U$. Then by Theorem \ref{1.2-2} $B$ is a Cohen--Macaulay ring and $\operatorname{dim}U \le d-2$. Consider the canonical epimorphism $\varphi : R \to
\mathcal{R}(QB)$ and let $K = \mathrm{Ker}~\varphi$. Then $K \subseteq
U A[t]$.
Let $\mathfrak a = (0) : U$. Let $P \in \operatorname{Ass}_R(K)$ and let $\mathfrak p = P \cap A$. Then $\mathfrak a \subseteq \mathfrak p$, since $\mathfrak a K = (0)$. Notice that $P$ is the kernel of the canonical epimorphism $\phi : R \to \mathcal{R}([Q + \mathfrak p]/\mathfrak p)$.
If $Q \subseteq \mathfrak p$, then $\mathfrak p = \mathfrak m$ and $\operatorname{dim} R/P = \operatorname{dim} \mathcal{R}([Q+\mathfrak p]/\mathfrak p) = 0$. If $Q \not\subseteq \mathfrak p$ we have $\operatorname{dim} R/P = \operatorname{dim} A/\mathfrak p + 1 \le \operatorname{dim} A/\mathfrak a + 1 \le d-1$. Therefore $\operatorname{dim} R/P \le d-1$ for every $P \in \operatorname{Ass}_R(K)$, which shows that $\operatorname{dim}_RK \le d-1 = (d+1) - 2$. Since $\mathcal{R}(QB)$ is a Cohen--Macaulay ring, we have that $R_\mathfrak{M}$ is a Vasconcelos ring by Theorem \ref{1.2-2}.
\end{proof}
We close this section with an application to sequentially Cohen--Macaulay rings. We refer the reader to \cite{GHS} for the definition and details. See also \cite{CC}, \cite{Sch2}, \cite{St}. We use the following characterization.
\begin{prop}[\cite{Sch2}]\label{seqCM2}
Let $A$ be a Noetherian local ring.
Then M is a sequentially Cohen--Macaulay $A$-module if and only if
$M$ admits a Cohen--Macaulay filtration, that is,
a family $\mathcal{M}=\{M_i\}_{0\leq i\leq t}$ $(t>0)$
of $A$-submodules of $M$ with
$$M_0=(0)\subsetneq M_1 \subsetneq M_2 \subsetneq \cdots
\subsetneq M_t=M$$ such that
\begin{enumerate}
\item[$(\mathrm{i})$] $\dim_AM_{i-1}<\dim_AM_i$ and
\item[$(\mathrm{ii})$] $M_i/M_{i-1}$ is a Cohen--Macaulay $A$-module
for all $1\leq i\leq t$.
\end{enumerate}
\end{prop}
The result below follows from Proposition~\ref{1.2-1}.
\begin{cor}\label{seqCM}
Let $A$ be a Noetherian local ring with dimension $d>0$. Suppose that $A$ is a sequentially Cohen--Macaulay ring. If $\operatorname{dim} A/\mathfrak p \ne d-1$ for every $\mathfrak p \in \operatorname{Ass} (A)$,
then $A$ is a Vasconcelos ring.
\end{cor}
\begin{proof} We may assume that $A$ is not a Cohen--Macaulay ring. Let $U=U_A(0)$. In the notation of Proposition~\ref{seqCM2}, we have that $t\geq 2$ and $M_{t-1}=U$ (\cite[Theorem 2.3]{GHS}). Therefore $\dim U \le d-2$, since $\operatorname{dim} A/\mathfrak p \ne d-1$ for all $\mathfrak p \in \operatorname{Ass} (A)$. Thus $A$ is a Vasconcelos ring by Proposition \ref{1.2-1}, because the ring $A/U$ is Cohen--Macaulay.
\end{proof}
Last we show that the
converse of Corollary \ref{seqCM} holds true, when $\operatorname{dim} A = 3$ and $A$ is a homomorphic image of a Cohen--Macaulay ring.
\begin{prop}
Let $A$ be a Noetherian local ring of dimension $3$, which is a homomorphic image of a Cohen--Macaulay ring. If $A$ is a Vasconcelos ring, then $A$ is a sequentially Cohen--Macaulay ring with $\operatorname{dim} A/\mathfrak p \ne 2$ for every $\mathfrak p \in \operatorname{Ass} (A)$.
\end{prop}
\begin{proof}
Let $U = U_A(0)$. We may assume that $U \ne (0)$. Then $A/U$ is a Cohen--Macaulay ring and $\dim U \le 1$ by Theorem \ref{1.2-2}. Hence $\operatorname{dim} A/\mathfrak p \ne 2$ for all $\mathfrak p \in \operatorname{Ass} (A)$. Let $W = \mathrm{H}_\mathfrak m^0(A)$. Then $W \subseteq U$ and $W = \mathrm{H}_\mathfrak m^0(U)$, because $\operatorname{depth} (A/U) >0 $. Therefore, if $\dim U = 1$ and $\operatorname{depth}(A) = 1$, then $(0) \subsetneq U \subsetneq A$ is a Cohen--Macaulay filtration of $A$. If $\dim U = 1$ but $\operatorname{depth} (A) = 0$, then $(0) \subsetneq W \subsetneq U \subsetneq A$ is a Cohen--Macaulay filtration of $A$. If $\dim U = 0$, then $(0) \subsetneq U \subsetneq A$ is a Cohen--Macaulay filtration of $A$. Thus $A$ is a sequentially Cohen--Macaulay ring by Proposition \ref{seqCM2}.
\end{proof}
\begin{cor}
Suppose that $A$ is a Vasconcelos ring of dimension $3$. Then the completion $\widehat{A}$ of $A$
is a sequentially Cohen--Macaulay ring.
\end{cor}
\section{Rings with $e_1(Q)$ constant}
In this section we study the
problem of when $e_1(Q)$ is independent of the choice of parameter
ideals $Q$ in $A$. Part of the motivation comes from the fact that
Buchsbaum rings have this property. We establish here that when
$e_1(Q)=-1$ or $e_1(Q)=-2$ and $A$ is unmixed, then $A$ is indeed
Buchsbaum. (The question of the variability of $e_1(Q)$ will be
considered in another paper.)
Let $(A,\mathfrak m)$ be a Noetherian local ring with maximal ideal $\mathfrak m$ and $d=\operatorname{dim}A > 0$. Assume that $A$ is a homomorphic image of a Gorenstein ring. Then $A$ contains a system of parameters $x_1, x_2, \cdots, x_d$ which forms a strong $d$-sequence in $A$, that is, the sequence $x_1^{n_1}, x_2^{n_2}, \cdots, x_d^{n_d}$ is a $d$-sequence in $A$ for all integers $n_1, n_2, \cdots, n_d \ge 1$ (see \cite[Theorem 2.6]{Cu} or \cite[Theorem 4.2]{Kw} for the existence of such systems of parameters). For each integer $q \ge 1$
let $\Lambda_q(A)$ be the set of values $e_1(Q)$, where $Q$ runs over the parameter ideals of $A$ such that $Q \subseteq \mathfrak m^q$ and $Q=(a_1, a_2, \cdots, a_d)$ with $a_1, a_2, \cdots, a_d$ a $d$-sequence. We then have $\Lambda_q(A) \ne \emptyset$, $\Lambda_{q+1}(A) \subseteq \Lambda_q(A)$ for all $q \ge 1$, and $\alpha \le 0$ for every $\alpha \in \Lambda_q(A)$ (Corollary \ref{cor} (a)).
\begin{lem}\label{key}
Let $(A,\mathfrak m)$ be a Noetherian local ring of dimension $d\ge 2$, which is a homomorphic image of a Gorenstein ring. Assume that $\Lambda_q(A)$ is a finite set for some integer $q \ge 1$ and let $\ell = -\operatorname{min}\Lambda_q(A)$. Suppose that $\operatorname{Ass} (A)= \operatorname{Assh} (A)$. Then $\mathfrak m^{\ell}\mathrm{H}_\mathfrak m^i(A) = (0)$ for all $i \ne d$, whence all the local cohomology modules $\{\mathrm{H}_\mathfrak m^i(A)\}_{0 \le i < d}$ of $A$ are finitely generated.
\end{lem}
\begin{proof}
Let $C = \mathrm{H}_\mathfrak m^1(A)$. Then $C$ is a finitely generated $A$-module (Proposition \ref{GNa} (a)). Suppose that $d = 2$ and let $\ell' = \ell_A(C)$. Let $a, b$ be a system of parameters of $A$ and assume that $a,b$ is a $d$-sequence in $A$. Then the element $a$ is superficial for the ideal $Q = (a,b)$, so that $e_1(Q) = e_1(Q/(a)) = -\ell_A((0):_C a )$. Therefore, choosing $a, b \in \mathfrak m^q$ with $aC = (0)$, we get $-\ell' =e_1(Q) \in \Lambda_q(A)$, whence $\ell' \le \ell$. Thus $\mathfrak m^{\ell}C = (0)$, because $\mathfrak m^{\ell'}C = (0)$.
Suppose now that $d \ge 3$ and that our assertion holds true for $d-1$. Let $$\mathcal{F}' = \{\mathfrak p \in \operatorname{Spec} A \mid \mathfrak p \ne \mathfrak m, \operatorname{dim}A_\mathfrak p > \operatorname{depth}(A_\mathfrak p) = 1\}.$$ Then $\mathcal{F}'$ is a finite set (Proposition \ref{GNa} (b)). We choose $x \in \mathfrak m$ so that $$x \not\in \bigcup_{\mathfrak p \in \operatorname{Ass}(A)}\mathfrak p \cup \bigcup_{\mathfrak p \in \mathcal{F}'} \mathfrak p.$$ Let $n \ge q$ be an integer such that $x^n\mathrm{H}_\mathfrak m^1(A) = (0)$ and put $y = x^n$. Let $B =A/yA$. Then $\operatorname{dim} B = d-1$ and $\operatorname{Ass}_A (B) \setminus \{\mathfrak m \} = \operatorname{Assh}_A(B)$. It follows that $U_B(0) = \mathrm{H}_\mathfrak m^0(B)$ (see Remark~\ref{U(A)}). Let $\widetilde{B}=B/\mathrm{H}_\mathfrak m^0(B)$.
Let $q' \ge q$ be an integer such that ${\mathfrak n}^{q'} \cap \mathrm{H}_\mathfrak m^0(B) = (0)$, where $\mathfrak n$ denotes the maximal ideal of $B$. Let $y_2, y_3, \cdots, y_d \in \mathfrak m^{q'}$ be a system of parameters for the $A$-module $\widetilde{B}$ and assume that $y_2, y_3, \cdots, y_d$ is a $d$-sequence in $\widetilde{B}$. Since $(y_2, y_3, \cdots, y_d)B \cap \mathrm{H}_\mathfrak m^0(B) = (0)$, we have that $y_2, y_3, \cdots, y_d$ forms a $d$-sequence in $B$ also. Then, because $y$ is $A$-regular, the sequence $y_1=y, y_2, \cdots, y_d$ forms a $d$-sequence in $A$, whence $y_1$ is superficial for the parameter ideal $Q = (y_1, y_2, \cdots, y_d)$ of $A$. Consequently
$$e_1((y_2, y_3, \cdots, y_d)\widetilde{B})=e_1((y_2, y_3, \cdots, y_d)B) = e_1(Q) \in \Lambda_q(A),$$ so that $\Lambda_{q'}(\widetilde{B}) \subseteq \Lambda_q(A).$
Therefore $\Lambda_{q'}(\widetilde{B})$ is a finite set, whence the hypothesis of induction on $d$ yields that $\mathfrak m^{\ell''}\mathrm{H}_\mathfrak m^i(\widetilde{B}) = (0)$ for all $i \ne d-1$, where $\ell'' = -\operatorname{min} \Lambda_{q'}(\widetilde{B})$.
Hence $\mathfrak m^{\ell}\mathrm{H}_\mathfrak m^i(\widetilde{B}) = (0)$ for all $i \ne d-1$, because $\ell'' \le \ell$.
Now consider the exact sequence
$$\cdots \to \mathrm{H}_\mathfrak m^1(A) \overset{x^n}{\to} \mathrm{H}_\mathfrak m^1(A) \to \mathrm{H}_\mathfrak m^1(B) \to \cdots \to \mathrm{H}_\mathfrak m^i(B) \to \mathrm{H}_\mathfrak m^{i+1}(A) \overset{x^n}{\to} \mathrm{H}_\mathfrak m^{i+1}(A) \to \cdots $$
of local cohomology modules. We then have
$$
\mathfrak m^{\ell}\left[(0):_{\mathrm{H}_\mathfrak m^{i+1}(A)}x^n\right] = (0)
$$
for all integers $1 \le i \le d-2$ and $n \ge q$, because $\mathfrak m^{\ell}\mathrm{H}_\mathfrak m^i(B) = (0)$ for all $1 \le i \le d - 2$. Hence $\mathfrak m^{\ell}\mathrm{H}_\mathfrak m^{i+1}(A)= (0)$, because $$\mathrm{H}_\mathfrak m^{i+1}(A) = \bigcup_{n \ge 1}\left[(0):_{\mathrm{H}_\mathfrak m^{i+1}(A)}\mathfrak m^n\right].$$
On the other hand we have the embedding $\mathrm{H}_\mathfrak m^1(A) \subseteq \mathrm{H}_\mathfrak m^1(B)$, since $x^n\mathrm{H}_\mathfrak m^1(A) = (0)$. Thus
$\mathfrak m^{\ell}\mathrm{H}_\mathfrak m^{1}(A) = (0)$, which completes the proof of the lemma.
\end{proof}
Now let $A$ be a Noetherian local ring with maximal ideal $\mathfrak m$ and $d = \operatorname{dim}A > 0$. Let $\Lambda (A) = \{e_1(Q) \mid Q ~\operatorname{is ~a ~parameter ~ideal ~in} A \}$. Passing to the completion $\widehat{A}$ of $A$ and applying Lemma~\ref{key} we obtain the following.
\begin{prop}\label{flc}
Let $(A,\mathfrak m)$ be an unmixed Noetherian local ring of dimension $d\ge 2$. Assume that $\Lambda (A)$ is a finite set and put $\ell = -\operatorname{min} \Lambda (A)$. Then $\mathfrak m^{\ell}\mathrm{H}_\mathfrak m^i(A) = (0)$ for every $ i \ne d$, whence $\mathrm{H}_\mathfrak m^i(A)$ is a finitely generated $A$-module for every $i \ne d$.
\end{prop}
A system of parameters $a_1, a_2, \cdots, a_d$ of $A$ is said to be {\it standard}, if it forms a $d^+$-sequence, that is, $a_1, a_2, \cdots, a_d$ forms a strong $d$-sequence in any order. We have that $A$ possesses a standard system of parameters if and only if $A$ is a generalized Cohen--Macaulay ring, i.e., all the local cohomology modules $\{\mathrm{H}_\mathfrak m^i(A) \}_{0 \le i < d}$ are finitely generated (see \cite{T}).
We say that a parameter ideal $Q$ of $A$ is standard, if it is
generated by a standard system of parameters. We have that $Q$ is
standard if and only if the equality $$\ell_A(A/Q) - e_0(Q) =
\sum_{i=0}^{d-1}\binom{d-1}{i}h^i(A):={\mathbb I}(A)$$ holds true,
where $h^i(A) = \ell_A(\mathrm{H}_\mathfrak m^i(A))$ for each $i \in \mathbb Z$ (cf. \cite[Theorem 2.1]{T}). See \cite{STC, T} for details, where the notion of generalized Cohen--Macaulay module is also given and various basic properties of generalized Cohen--Macaulay rings and modules are explored.
\medskip
Assume that $A$ is a generalized Cohen--Macaulay ring with $d \ge 2$ and let $$s=\sum_{i=1}^{d-1}\binom{d-2}{i-1}h^i(A).$$ If $Q$ is a parameter ideal of $A$, by \cite[Lemma 2.4]{GNi} we have that
$$
e_1(Q) \ge -s,
$$
where the equality holds true, if $Q$ is standard (\cite[Korollar 3.2]{Sch1}).
Therefore, if $A$ is unmixed, $d \ge 2$, and $\Lambda (A)$ is a finite set, by Proposition~\ref{flc} we have that
$$
\mathfrak m^s \mathrm{H}_\mathfrak m^i(A)= (0)
$$
for all $i \ne d$. This exponent is, however, never the best possible, as we show in the following.
\begin{cor}\label{q-Bbm}
Let $(A,\mathfrak m)$ be an unmixed Noetherian local ring of dimension $d\ge 2$. If $\# \Lambda (A) = 1$, then $A$ is a quasi-Buchsbaum ring, that is, $\mathfrak m \mathrm{H}_\mathfrak m^i(A) = (0)$ for all $i \ne d$.
\end{cor}
To prove Corollary \ref{q-Bbm} we need the lemmas below.
\begin{lem}\label{d=2}
Suppose that $(A,\mathfrak m)$ is a Noetherian local ring of dimension $d= 2$ and $\operatorname{depth} (A) = 1$. Assume that $\mathrm{H}_\mathfrak m^1(A)$ is finitely generated. Let $Q$ be a parameter ideal of $A$. Then the following three conditions are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $e_1(Q) = -\ell_A(\mathrm{H}_\mathfrak m^1(A))${\rm ;}
\item[{\rm (b)}] $Q \mathrm{H}_\mathfrak m^1(A) = (0)${\rm ;}
\item[{\rm (c)}] $Q$ is standard.
\end{enumerate}
\end{lem}
\begin{proof} (c) $\Rightarrow$ (a) See \cite[Lemma 2.4 (2)]{GNi}.
(b) $\Leftrightarrow$ (c) See \cite[Corollary 3.7]{T}.
(a) $\Rightarrow$ (b) We may assume that the field $A/\mathfrak m$ is infinite. Let $Q = (a,b)$ be such that each one of $a,b$ is a superficial element of $Q$. Then $$-\ell_A(\mathrm{H}_\mathfrak m^1(A)) = e_1(Q) = e_1(Q/(a)) = -\ell_A(\mathrm{H}_\mathfrak m^0(A/(a))$$ by \cite[Lemma 2.1 (1)]{GNi}. Hence $\ell_A(\mathrm{H}_\mathfrak m^1(A))=\ell_A((0) :_{\mathrm{H}_\mathfrak m^1(A)} a)$, and so $a\mathrm{H}_\mathfrak m^1(A) = (0)$. Similarly we get $b\mathrm{H}_\mathfrak m^1(A) = (0)$, whence $Q\mathrm{H}_\mathfrak m^1(A) = (0)$.
\end{proof}
\begin{lem}\label{d=2generalized}
Suppose that $(A,\mathfrak m)$ is a generalized Cohen--Macaulay local ring of dimension $d \ge 2$ and $\operatorname{depth} (A) > 0$. Let $Q$ be a parameter ideal of $A$ such that $e_1(Q)= - \sum_{i = 1}^{d-1}\binom{d - 2}{i - 1}h^i(A)$. Then $Q \mathrm{H}_\mathfrak m^i(A) = (0)$ for all $i \ne d$.
\end{lem}
\begin{proof}
If $d=2$ the conclusion follows from Lemma~\ref{d=2}. Assume that $d\geq 3$. Let $Q = (a_1, a_2, \cdots, a_d)$, where each $a_i$ is superficial for the ideal $Q$, and let $a = a_i$. Let $B=A/aA$. We have that $e_1(QB)=e_1(Q)$.
Consider the exact sequence of local cohomology modules $$
\cdots \to \mathrm{H}_\mathfrak m^i(A) \overset{a}{\to} \mathrm{H}_\mathfrak m^i(A) \to \mathrm{H}_\mathfrak m^i(B) \to \mathrm{H}_\mathfrak m^{i+1}(A) \overset{a}{\to} \mathrm{H}_\mathfrak m^{i+1}(A) \to \mathrm{H}_\mathfrak m^{i+1}(B) \to \cdots.
$$ We then have
$$
h^i(B) = \ell_A(\mathrm{H}_\mathfrak m^i(A)/a\mathrm{H}_\mathfrak m^i(A)) + \ell_A((0) :_{\mathrm{H}_\mathfrak m^{i+1}(A)}a)
\le h^i(A) + h^{i+1}(A)
$$
for all $0 \le i \le d-2$. Hence
\begin{eqnarray*}
e_1(QB)\geq -\sum_{i=1}^{d-2}\binom{d-3}{i-1}h^i(B) &\ge&-\sum_{i=1}^{d-2}\binom{d-3}{i-1}\left[h^i(A) + h^{i+1}(A)\right]\\
&=& -\sum_{i=1}^{d-1}\binom{d-2}{i-1}h^i(A)\\
&=& e_1(QB).
\end{eqnarray*}
It follows that $h^i(B) = h^i(A) + h^{i+1}(A)$ for every $1 \le i \le d-2$, whence $a\mathrm{H}_\mathfrak m^i(A) = (0)$ for all $1 \le i \le d - 1$. Thus $Q\mathrm{H}_\mathfrak m^i(A) = (0)$, if $i \ne d$.
\end{proof}
\begin{proof}[Proof of Corollary $\ref{q-Bbm}$]
By Proposition \ref{flc} $A$ is a generalized Cohen--Macaulay ring. Hence we have $\Lambda (A) = \{-\sum_{i=1}^{d-1}\binom{d-2}{i-1}h^i(A)\}$ by \cite[Korollar 3.2]{Sch1}.
Let $a \in \mathfrak m$ such that $\operatorname{dim} A/aA = d-1$. It is enough to show that $a\mathrm{H}_\mathfrak m^i(A) = (0)$ for all $i \ne d$. This follows from Lemma~\ref{d=2generalized}.
\end{proof}
Since every quasi-Buchsbaum ring $A$ is Buchsbaum once $\operatorname{depth} A \ge d-1$ (\cite[Corollary 1.1]{SV}), we readily get the following.
\begin{cor} Suppose that $A$ is an unmixed Noetherian local ring of dimension $d \ge 2$ and $\operatorname{depth} A \ge d-1$. Then $\# \Lambda (A) = 1$ if and only if $A$ is a Buchsbaum ring.
\end{cor}
The authors expect that $A$ is a Buchsbaum ring, if $A$ is unmixed and $e_1(Q)$ is independent of the choice of parameter ideals $Q$ of $A$. We will show that this is the case, when $e_1(Q) = -1$ and when $e_1(Q) = -2$.
\begin{prop}\label{dimU}
Let $A$ be a Noetherian local ring of dimension $d \ge 2$ and suppose that for all parameter ideals $Q$ of $A$, $e_1(Q)=-t$ for some $t\geq 0$. Let $U=U_A(0)$. Then $\dim U \le d-2$, and for all parameter ideals $\mathfrak q$ of $A/U$ we have that $e_1(\mathfrak q)=-t$ .
\end{prop}
\begin{proof}
Let $B=A/U$. Assume that $\dim U = d-1$. Choose a system of parameters $a_1, a_2, \cdots, a_d$ of $A$ so that $(a_d) \cap U = (0)$ (cf. \cite{CC}). Let $\ell > t$ be an integer and let $Q = (a_1^\ell, a_2, \cdots, a_d)$. For all $n \ge 0$ we have the exact sequence of $A$-modules
$$0 \to U/(Q^{n+1} \cap U) \to A/Q^{n+1} \to B/Q^{n+1}B \to 0.$$ Let $k \ge 0$ be an integer such that
$$Q^{n} \cap U =
Q^{n-k}(Q^k \cap U)$$
for all $n \ge k$. We put $U' = Q^k \cap U$ and $\mathfrak q = (a_1^{\ell}, a_2, \cdots, a_{d-1})$. Then $Q^{n-k}U' = \mathfrak q^{n-k}U'$, because $a_dU=(0)$. Therefore, for all $n \ge k$ we have
$$\ell_A(A/Q^{n+1}) = \ell_A(B/Q^{n+1}B) + \ell_A(U'/\mathfrak q^{n-k+1}U') + \ell_A(U/U'),$$
which implies
$$
-t = e_1(Q) = e_1(QB) -e_0(\mathfrak q U').
$$
Consequently, since $e_1(QB) \le 0$, we have $$
\ell \le \ell e_0((a_1, a_2, \cdots, a_{d-1})U') = e_0(\mathfrak q U') = e_1(QB) + t \le t,
$$
which is impossible. Thus $\dim U \le d-2$.
To see the second assertion, let $\mathfrak q$ be a parameter ideal of $B$. Then, choosing a parameter ideal $Q$ of $A$ so that $QB = \mathfrak q$, we get $e_1(\mathfrak q) = e_1(Q)=-t$, since $\dim U \le d-2$.
\end{proof}
\begin{thm}\label{constant}
Let $(A,\mathfrak m)$ be a Noetherian local ring of dimension $d \ge 2$. Then the following two conditions are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $e_1(Q) = -1$ for every parameter ideal $Q$ of $A${\rm ;}
\item[{\rm (b)}] Let $U = U_{\widehat{A}}(0)$ be the unmixed component of $(0)$ in the $\mathfrak m$-adic completion $\widehat{A}$ of $A$. Then $\operatorname{dim}_{\widehat{A}}U \le d-2$ and $\widehat{A}/U$ is a Buchsbaum ring such that either
\begin{enumerate}
\item[$(\mathrm{i})$] $\mathrm{H}_{\widehat{\mathfrak m}}^i(\widehat{A}/U) = (0)$ for all $i \ne 1, d$ and $\ell_{\widehat{A}}(\mathrm{H}_{\widehat{\mathfrak m}}^1(\widehat{A}/U)) = 1$, or
\item[$(\mathrm{ii})$] $\mathrm{H}_{\widehat{\mathfrak m}}^i(\widehat{A}/U) = (0)$ for all $i \ne d-1, d$ and $\ell_{\widehat{A}}(\mathrm{H}_{\widehat{\mathfrak m}}^{d-1}(\widehat{A}/U)) = 1$,
\end{enumerate}
where $\widehat{\mathfrak m}$ denotes the maximal ideal of $\widehat{A}$.
\end{enumerate}
\end{thm}
\begin{proof}
We may assume that $A$ is complete.
(b) $\Rightarrow$ (a)
Let $Q$ be a parameter ideal of $A$, and let $B = A/U$. Then $e_1(Q) = e_1(QB)$, since $\operatorname{dim}_{A}U \le d-2$. Consequently $e_1(Q) = -1$, because $e_1(QB) = -\sum_{i=1}^{d-1}\binom{d-2}{i-1}h^i(B) = -1$ by \cite[Korollar 3.2]{Sch1} and condition (i) or (ii).
(a) $\Rightarrow$ (b)
By Proposition~\ref{dimU} we may assume that $A$ is unmixed. Then $A$ is a quasi-Buchsbaum ring by Corollary~\ref{q-Bbm}. We have condition (i) or (ii), because $e_1(Q) = -\sum_{i=1}^{d-1}\binom{d-2}{i-1}h^i(A)$. Hence $A$ is a Buchsbaum ring, because $A$ is a quasi-Buchsbaum ring with $\mathrm{H}_\mathfrak m^i(A) = (0)$ for all $i \ne \operatorname{depth} A, \operatorname{dim} A$ (see \cite[Corollary 1.1]{SV}).
\end{proof}
To treat the case where $e_1(Q) = -2$ we need the following.
\begin{lem}\label{-2}
Suppose that $(A,\mathfrak m)$ is a generalized Cohen--Macaulay local ring of dimension $d \ge 2$ and $\operatorname{depth} (A) > 0$. Assume that $h^1(A) = 1$ and $h^i(A) = 0$ for all $2 \le i \le d-2$. Let $Q$ be a parameter ideal in $A$ such that $e_1(Q)= - \sum_{i = 1}^{d-1}\binom{d - 2}{i - 1}h^i(A)$. Then $Q$ is standard.
\end{lem}
\begin{proof}
We may assume that the field $A/\mathfrak m$ is infinite. If $d=2$, the conclusion follows from Lemma~\ref{d=2}.
Let $d \ge 3$ and let $Q = (a_1, a_2, \cdots, a_d)$, where each $a_i$ is superficial for the ideal $Q$. Let $a = a_i$, $\mathrm{U}(a) = (a) : \mathfrak m$, and put $B = A/aA$, $\widetilde{B} = B/\mathrm{H}_\mathfrak m^0(B)$. Then
$$e_1(Q) = e_1(QB) = e_1(Q\widetilde{B}).$$
Consider the exact sequence of local cohomology modules $$
\cdots \to \mathrm{H}_\mathfrak m^i(A) \overset{a}{\to} \mathrm{H}_\mathfrak m^i(A) \to \mathrm{H}_\mathfrak m^i(B) \to \mathrm{H}_\mathfrak m^{i+1}(A) \overset{a}{\to} \mathrm{H}_\mathfrak m^{i+1}(A) \to \mathrm{H}_\mathfrak m^{i+1}(B) \to \cdots.
$$
Since $h^1(A) = 1$, we have $h^0(B)=1$, whence $\mathrm{H}_\mathfrak m^0(B)=\mathrm{U}(a)/(a)$ and $\widetilde{B} = A/\mathrm{U}(a)$.
By Lemma~\ref{d=2generalized} we have that $a\mathrm{H}_\mathfrak m^{d-1}(A)=(0)$. Therefore, if $d=3$ we get an exact sequence
$$0 \to \mathrm{H}_\mathfrak m^1(A) \to \mathrm{H}_\mathfrak m^1(B) \to \mathrm{H}_\mathfrak m^2(A) \to 0,$$ whence $h^1(\widetilde{B}) = h^1(B) = 1 + h^2(A)$.
If $d\geq 4$ we get $\mathrm{H}_\mathfrak m^1(B) \cong \mathrm{H}_\mathfrak m^1(A)$, $\mathrm{H}_\mathfrak m^i(B) = (0)$ if $2 \le i \le d-3$, and $\mathrm{H}_\mathfrak m^{d-2}(B) \cong \mathrm{H}_\mathfrak m^{d-1}(A)$.
Consequently, if $d\geq 3$ we have that $e_1(Q\widetilde{B}) = - \sum_{i = 1}^{d-2}\binom{d - 3}{i - 1}h^i(\widetilde{B})$, and so by induction on $d$ the parameter ideal $Q\widetilde{B}$ is standard.
Therefore
$$
\ell_{\widetilde{B}}(\widetilde{B}/Q\widetilde{B}) - e_0(Q\widetilde{B}) = {\mathbb I}(\widetilde{B})=
(d-2) + h^{d-1}(A).
$$
Now assume by contradiction that $Q$ is not a standard parameter ideal in $A$. Then
\begin{eqnarray*}
(d-2) + h^{d-1}(A) &=& \ell_{\widetilde{B}}(\widetilde{B}/Q\widetilde{B}) - e_0(Q\widetilde{B})\\ &=& \ell_A(A/(\mathrm{U}(a) + Q)) - e_0(A)\\
&=& \left[\ell_A(A/Q) - e_0(A) \right] - \ell_A((Q + \mathrm{U}(a))/Q)\\
&<&{\mathbb I}(A) - \ell_A((Q + \mathrm{U}(a))/Q)\\
&=& \left[(d-1) + h^{d-1}(A)\right] - \ell_A((Q + \mathrm{U}(a))/Q).
\end{eqnarray*}
Consequently, $\ell_A((Q + \mathrm{U}(a))/Q)=0$, whence $\mathrm{U}(a) \subseteq Q$.
Therefore $$\sum_{i = 1}^d\mathrm{U}(a_i) = Q$$ by the symmetry among the elements $a_i$. Let $\widetilde{A}$ denote the $(\mathrm{S}_2)$-fication of $A$ and look at the canonical exact sequence
$$0 \to A \to \widetilde{A} \to \mathrm{H}_\mathfrak m^1(A) \to 0$$
(\cite[Theorem 1.6]{AG}). Then $\operatorname{dim} \widetilde{A} = d$, $\operatorname{depth} \widetilde{A} \ge d-1$, and $\mathrm{H}_\mathfrak m^{d-1}(\widetilde{A}) \cong \mathrm{H}_\mathfrak m^{d-1}(A)$. Hence $Q$ is a standard parameter ideal for the generalized Cohen--Macaulay $A$-module $\widetilde{A}$ by \cite[Corollary 3.7]{T}, because $Q\mathrm{H}_\mathfrak m^{d-1}(\widetilde{A}) = (0)$. Therefore, since $\operatorname{depth}_A\widetilde{A} \ge 2$, any two of $a_1, a_2, \cdots, a_d$ form an $\widetilde{A}$-regular sequence, whence $\mathrm{U}(a_i) \subseteq a_i\widetilde{A}$ for all $1 \le i \le d$. Consequently $\mathrm{U}(a_i) = a_i \widetilde{A}$, because $a_iA \subsetneq \mathrm{U}(a_i) \subseteq a_i \widetilde{A}$ and $\ell_A(a_i \widetilde{A}/a_iA) =\ell_A(\widetilde{A}/A) = 1$. Thus $Q = Q\widetilde{A}$, so that we have $Q^{n+1} = Q^{n+1}\widetilde{A}$ for all $n \ge 0$. Since $$\ell_A(A/Q^{n+1}) = \ell_A(\widetilde{A}/Q^{n+1}\widetilde{A}) - 1,$$ we get $$e_1(Q) = e_1(Q\widetilde{A}) = -h^{d-1}(\widetilde{A}) = -h^{d-1}(A)$$ by \cite[Korollar 3.2]{Sch1}, a contradiction. Thus the parameter ideal $Q$ is standard in $A$.
\end{proof}
\begin{thm}\label{constant2}
Let $(A,\mathfrak m)$ be a Noetherian local ring of dimension $d \ge 2$. Then the following two conditions are equivalent{\rm :}
\begin{enumerate}
\item[{\rm (a)}] $e_1(Q) = -2$ for every parameter ideal $Q$ of $A${\rm ;}
\item[{\rm (b)}] Let $U = U_{\widehat{A}}(0)$ be the unmixed component of $(0)$ in the $\mathfrak m$-adic completion $\widehat{A}$ of $A$. Then $\operatorname{dim}_{\widehat{A}}U \le d-2$ and $\widehat{A}/U$ is a Buchsbaum ring with $$\sum_{i = 1}^{d-1}\binom{d-2}{i-1}h^i(\widehat{A}/U) = 2.$$
\end{enumerate}
\end{thm}
\begin{proof}
(b) $\Rightarrow$ (a) The assertion follows from \cite[Korollar 3.2]{Sch1}.
(a) $\Rightarrow$ (b)
By Proposition~\ref{dimU}, we may assume that $A$ is an unmixed complete local ring. Hence $A$ is a quasi-Buchsbaum ring by Corollary \ref{q-Bbm} and $\sum_{i = 1}^{d-1}\binom{d-2}{i-1}h^i(A) = 2$ by \cite[Korollar 3.2]{Sch1}. By \cite[Corollary 1.1]{SV} we may assume that $d \ge 3$, $h^1(A) = h^{d-1}(A) = 1$, and $h^i(A) = 0$ if $2 \le i \le d-2$. Then by Lemma \ref{-2} every parameter ideal $Q$ of $A$ is standard, so that $A$ is a Buchsbaum ring.
\end{proof}
|
\section{Introduction} \lb{s1}
In this paper, we'll present and exploit a powerful analogy between spectral theory and the question of
when a power series $f(z)=\sum_{n=0}^\infty a_n z^n$ defining an analytic function on $\bbD=\{z\mid\abs{z}<1\}$ has
a natural boundary on $\partial\bbD$, in that for no $z_0=e^{i\theta}$ does $f$ have an analytic continuation to $\{z\mid
\abs{z-z_0}<\delta\}$ for some $\delta >0$. In particular, we shall import two notions (``reflectionless'' and ``right
limit'') from the spectral theory of Jacobi matrices and obtain a general theorem regarding the consequence of the
possibility to analytically continue $f$ across an arc. While spectral theory ideas motivated our approach to natural
boundaries, what we develop doesn't require any spectral theory. The reader not knowledgeable in spectral theory and not
interested in the background should skip to the paragraph containing \eqref{1.4}.
As it turns out, some of the results presented in this paper are not new.
In particular, in \cite{Agmon1} Agmon treated natural boundaries for general series with radius of convergence $1$ using a related approach. Our Theorem
\ref{T1.3} is a special case of his results. To the best of our knowledge, however, our treatment of \emph{strong} natural boundaries using this approach
(see Theorem \ref{T1.4} below) is completely new.
Of course, since the relevant spectral theoretic notions were not yet defined at the time of publication of \cite{Agmon1}, Agmon was unaware of
the analogy. Surprisingly, \cite{Agmon1} is rarely quoted in the relevant literature and seems to be little known.
Thus, a secondary aim of our paper is to draw attention to Agmon's work.
Moreover, we believe the significance of the analogy presented in this paper is
more than merely anecdotal. As an example, our knowledge of the applications of these ideas to random potentials in the spectral theoretic setting, led us
to apply these methods to random series. In this context we have obtained results for series where only a subsequence is random and for general bounded ergodic nondeterministic series. Another example is Theorem \ref{T1.8} that has Hecke's famous example \eqref{1.13} as a special case.
Thus, Theorems \ref{T1.3} and \ref{T1.4} provide a general framework in which many existing and new results concerning natural boundaries can be derived.
To set the stage, consider a Jacobi matrix,
\begin{equation} \lb{1.1}
J=
\begin{pmatrix}
b_1 & a_1 & 0 & \cdots \\
a_1 & b_2 & a_2 & \cdots \\
0 & a_2 & b_3 & \cdots \\
\vdots & \vdots & \vdots & \ddots
\end{pmatrix}
\end{equation}
where $\{a_n,b_n\}_{n=1}^\infty$ are bounded. By a right limit of $J$, we mean a two-sided Jacobi matrix, $J^{(r)}$, with
parameters $\{a_n^{(r)}, b_n^{(r)}\}_{n=-\infty}^\infty$ given by
\begin{equation} \lb{1.2}
a_n^{(r)} = \lim_{k\to\infty}\, a_{n+n_k} \qquad b_n^{(r)} = \lim_{k\to\infty}\, b_{n+n_k}
\end{equation}
where $n_k\to\infty$ is a subsequence. By compactness, right limits exist. This definition is from Last--Simon \cite{LS},
who used it in
\begin{theorem}[\cite{LS}] \lb{T1.1} If $\Sigma_\ac(A)$ is the essential support of the a.c.\ spectrum of an operator, $A$, then
\begin{equation} \lb{1.3}
\Sigma_\ac(J) \subset \Sigma_\ac(J^{(r)})
\end{equation}
for every right limit $J^{(r)}$ of $J$.
\end{theorem}
Right limits are also relevant to essential spectrum where they have been exploited by several authors (e.g.,
\cite{CWL08,CWLppt,GI1,LS,S304,M1,Rab05}), but their relevance for a.c.\ spectrum goes back to Last--Simon \cite{LS}.
Recently, Remling \cite{Rem} found a much stronger property of right limits when there is a.c.\ spectrum. It depends on the
notion of reflectionless two-sided Jacobi matrix. The precise definition is irrelevant for our discussion here (see, e.g.,
\cite{BRS,Rem,Rice,SY}), but we note that if $\ti J$ is a two-sided Jacobi matrix, reflectionless on some $\fre\subset\bbR$
with the Lebesgue measure, $\abs{\fre}$, of $\fre$ nonzero, then $\fre\subset\Sigma_\ac (\ti J)$ and $\{\ti a_n,
\ti b_n\}_{n=-\infty}^{-1}$ determine $\{\ti a_n, \ti b_n\}_{n=0}^\infty$. Remling \cite{Rem} proved
\begin{theorem}[\cite{Rem}] \lb{T1.2} If $\fre$ is $\Sigma_\ac(J)$ for a Jacobi matrix, then any right limit, $J^{(r)}$, is
reflectionless on $\fre$.
\end{theorem}
With this result, Remling was able to recover and extend virtually every result on the absence of a.c.\ spectrum for situations
where $\{a_n,b_n\}_{n=1}^\infty$ are bounded, and the lack of a.c.\ spectrum is due to a part of the $a$'s and $b$'s that is
dominant at infinity (rather than a perturbation that goes to zero at infinity).
Our work here began by our noticing that the major classes of Jacobi matrices with no a.c.\ spectrum have analogs in
the major classes of results on the occurrence of natural boundaries (see Remmert \cite[Ch.~11]{Remm} for a summary of
classical results on natural boundaries), as seen in
\begin{itemize}
\item Gap theorems \cite{Had,Fabry,Faber,Remm} $\sim$ sparse potentials \cite{Pear78,Rem}
\item Finite-valued power series \cite{Sz1922,Remm} $\sim$ finite-valued Jacobi matrices \cite{Kot89,Rem}
\item Random power series \cite{Stei,PZyg,Kah} $\sim$ Anderson localization \cite{CarLac,PasFig}
\end{itemize}
Let us describe our major abstract results on natural boundaries motivated by Last--Simon \cite{LS} and Remling \cite{Rem}.
Given a power series $f(z)=\sum_{n=0}^\infty a_n z^n$ with
\begin{equation} \lb{1.4}
\sup_n\, \abs{a_n} <\infty
\end{equation}
we define a right limit of $\{a_n\}_{n=0}^\infty$ to be a two-sided sequence $\{b_n\}_{n=-\infty}^\infty$ with
\begin{equation} \lb{1.5}
b_n =\lim_{j\to\infty}\, a_{n+n_j}
\end{equation}
for some $n_j\to\infty$. By compactness and \eqref{1.4}, right limits exist.
Given a two-sided bounded sequence, $\{b_n\}_{n=-\infty}^\infty$, we consider two functions, $f_+(z)$ on $\bbD$ and
$f_-(z)$ on $\bbC\cup\{\infty\}\setminus\ol{\bbD}$, defined by
\begin{equation} \lb{1.6}
f_+(z) =\sum_{n=0}^\infty b_n z^n \qquad f_-(z) =\sum_{n=-\infty}^{-1} b_n z^n
\end{equation}
where the series are guaranteed to converge on the indicated sets. Let $I$ be an open interval in $\partial\bbD$. We say
$\{b_n\}_{n=-\infty}^\infty$ is {\it reflectionless\/} on $I$ if and only if $f_+$ has an analytic continuation from
$\bbD$ to $\bbC\cup\{\infty\}\setminus(\partial\bbD\setminus I)$, so that on $\bbC\cup\{\infty\}\setminus\ol{\bbD}$, we
have that
\begin{equation} \lb{1.6x}
f_+(z)+f_-(z)=0
\end{equation}
Obviously, it suffices that $f_+$ have a continuation to a neighborhood of $I$ so that \eqref{1.6x} holds in the
intersection of that neighborhood and $\bbC\setminus\ol{\bbD}$.
\begin{ex} Let $b_n\equiv 1$. Then $f_+(z)=(1-z)^{-1}$ and $f_-(z)=-(1-z)^{-1}$. This series is reflectionless on
$I=\{e^{i\theta}\mid 0<\theta<2\pi\}$. Similarly, it is easy to see that a periodic $b_n$ of period $p$ is
reflectionless on any $I$ in $\partial\bbD$ with all the $p$-th roots of unity removed.
\end{ex}
Our first main theorem is:
\begin{theorem} \lb{T1.3} Let $f(z)=\sum_{n=0}^\infty a_n z^n$ be a power series with \eqref{1.4}. Suppose $I\subset
\partial\bbD$ is an open interval so that $f(z)$ has an analytic continuation to a neighborhood of $I$. Then every right
limit of $\{a_n\}_{n=0}^\infty $ is reflectionless on $I$.
\end{theorem}
As remarked above, this theorem is a special case of the results in \cite{Agmon1}. While our treatment is limited to bounded sequences and produces interesting results only when $a_n \nrightarrow 0$, Agmon treats general (not necessarily bounded) series. He introduces a way to renormalize the coefficients so that any series with radius of convergence 1 has, after renormalization, nontrivial right limits. He then shows that if the corresponding function has an analytic continuation to an arc on $\partial \bbD$, then the renormalized right limits are reflectionless across that arc.
For bounded series, we will actually prove a stronger result (we state Theorem~\ref{T1.3} both for conceptual reasons and because we need it in the proof of the stronger result):
\begin{theorem} \lb{T1.4} Let $f(z)=\sum_{n=0}^\infty a_n z^n$ be a power series with \eqref{1.4}. Suppose $I\subset
\partial\bbD$ is an open interval so that
\begin{equation} \lb{1.7}
\sup_{0<r<1} \int_{e^{i\theta}\in I} \abs{f(re^{i\theta})}\, \f{d\theta}{2\pi} <\infty
\end{equation}
Then every right limit of $\{a_n\}$ is reflectionless on $I$.
\end{theorem}
These theorems imply natural boundaries. We say $f$ has a {\it strong natural boundary\/} on $\partial\bbD$ if
\eqref{1.7} fails for every $I\subset\partial\bbD$. In particular, in that case, $f$ is unbounded in every sector
$\{re^{i\theta}\mid 0<r<1,\, e^{i\theta}\in I\}$.
\begin{corollary}\lb{C1.5} Let $f(z)=\sum_{n=0}^\infty a_n z^n$ be such that for any open interval $I\subset\partial\bbD$,
there is a right limit of $\{a_n\}_{n=0}^\infty$ which is not reflectionless on $I$. Then $f$ has a strong natural boundary
on $\partial\bbD$.
\end{corollary}
Of course, if $a_n\to 0$, it can happen that there are
natural boundaries which are not strong natural boundaries. For example, by the Hadamard gap theorem \cite{Remm},
\begin{equation} \lb{1.8}
f(z)=\sum_{n=1}^\infty \, \f{z^{n!}}{(n!)^n}
\end{equation}
has a natural boundary on $\partial\bbD$ but, for all $k$, $f^{(k)}(z)$ is bounded on $\bbD$.
We note that Duffin--Schaeffer \cite{DuSc} and Boas \cite{Boas} long ago had results on what we call strong natural
boundaries. These results are all for series with gaps or finite-valued series. These authors and also Agmon \cite{Agmon1} have results that prove that certain functions rather than merely having classical natural boundaries are unbounded in every sector (we call this $L^\infty$ natural boundaries). We note that our approach for going from classical to strong natural boundaries is very close to the method that Agmon \cite{Agmon1} uses to go from classical natural boundaries to unboundedness in every sector.
Theorem \ref{T1.4} is, to the best of our knowledge, the first general theorem concerning strong natural boundaries.
For us, the point of these theorems is conceptual: they provide a unified framework that makes many theorems transparent.
That said, the following results seem to be new:
\begin{theorem}\lb{T1.6} Suppose $\{a_n\}_{n=0}^\infty$ obeys \eqref{1.4} and there exists $n_j\to\infty$ so that for all
$k<0$,
\begin{align}
\lim_{j\to\infty}\, a_{n_j+k} & =0 \lb{1.9} \\
\liminf\abs{a_{n_j}} &> 0 \lb{1.10}
\end{align}
Then $f(z)=\sum_{n=0}^\infty a_n z^n$ has a strong natural boundary on $\partial\bbD$.
\end{theorem}
\begin{remarks} 1. This result (as well as its extension---Theorem \ref{T4.3}), for $L^\infty$ natural boundaries, appears in Agmon \cite{Agmon1}.
\smallskip
2. The result is true if \eqref{1.9} is replaced by the assumption for all $k>0$ or for all $k<K$\!, some $K<0$ or
for all $k>K$\!, some $K>0$. The proofs are essentially identical. In addition, \eqref{1.9} can be replaced by an
exponential decay condition on the limit; see Theorem~\ref{T4.3}.
\smallskip
3. This includes the famous examples $\sum_{n=1}^\infty z^{n!}$ of Weierstrass and $\sum_{n=0}^\infty z^{n^2}$ of Kronecker.
\smallskip
4. This allows gaps where the set of zeros has zero density. At first sight, this seems a violation of the result of P\'olya
\cite{Polya} and Erd\"os \cite{Erdos} that the Fabry gap theorem is optimal. We'll explain this apparent discrepancy in
Section~\ref{s4}.
\end{remarks}
\begin{theorem} \lb{T1.7} Let $\{a_n(\omega)\}_{\omega\in\Omega}$ be a translation invariant, ergodic, stochastic process,
which is nondeterministic, so that
\begin{equation} \lb{1.11}
\sup_{n,\omega}\, \abs{a_n(\omega)} <\infty
\end{equation}
Then for a.e.\ $\omega$, $\sum_{n=0}^\infty a_n(\omega) z^n$ has a strong natural boundary.
\end{theorem}
\begin{remarks} 1. So far as we know, all previous results on random power series rely on independence and only obtain natural
boundaries, not strong natural boundaries.
\smallskip
2. Recall that a stochastic process $\{a_n(\omega)\}$ is deterministic if $a_0(\omega)$ is a measurable function of
$\{a_n(\omega)\}_{n=-\infty}^{-1}$ for a.e.\ $\omega$, and nondeterministic if it is not deterministic.
\smallskip
3. By taking an average of a deterministic and a nondeterministic process, it is easy to see that ergodicity is essential
for this theorem to be true.
\end{remarks}
\begin{theorem}\lb{T1.8} Let $f\colon\partial\bbD\to\bbC$ be a bounded and piecewise continuous function with only a
finite number of discontinuities, at one of which the one-sided limits exist and are unequal. Then for any irrational
number $q$ and every
$\theta\in\bbR$, we have that
\begin{equation} \lb{1.12}
\sum_{n=0}^\infty f(e^{2\pi i (qn+\theta)}) z^n
\end{equation}
has a strong natural boundary.
\end{theorem}
\begin{remark} This includes Hecke's famous example \cite{Hecke},
\begin{equation} \lb{1.13}
f(z)=\sum_{n=0}^\infty \{nq\} z^n
\end{equation}
where $\{x\}$ is the fractional part of $x\in\bbR$.
\end{remark}
The following is elementary and does not require our full machinery. It was suggested to us by the Wonderland theorem of
Simon \cite{Wonder} and seems to be new.
\begin{theorem} \lb{T1.9} Let $\Omega\subset\bbC$ be a compact set with more than one point. Let $\Omega^\infty$ be a countable
product of copies of $\Omega$ in the weak topology. Then $\{\{a_n\}\}\in\Omega^\infty \mid\sum_{n=0}^\infty a_n z^n$ has a
natural boundary on $\partial\bbD\}$ is a dense $G_\delta$ in $\Omega^\infty$.
\end{theorem}
In Section~\ref{s2}, we prove Theorem~\ref{T1.3}. The key is a lemma of M.~Riesz whose proof we include for completeness. In Section~\ref{s3}, we prove Theorem~\ref{T1.4} using the theory of $H^p$ spaces on
a sector (Duren \cite{Duren}). In Section~\ref{s4}, we discuss gap theorems, including Theorem~\ref{T1.6}. In Section~\ref{s5}, we
discuss Szeg\H{o}'s theorem on finite-valued power series using Theorem~\ref{T1.4}. In Section~\ref{s6}, we discuss random
power series, including Theorem~\ref{T1.7}. In Section~\ref{s7}, we prove Theorem~\ref{T1.8}, following a spectral theory analysis
of Damanik--Killip \cite{DK}. In Section~\ref{s8}, we prove Theorem~\ref{T1.9}.
We believe our work here opens up numerous new directions in the study of power series and of spectral theory. In particular,
there is a dynamical view of reflectionless in spectral theory (see \cite{BRS}) and there is the distinction in spectral theory
between pure point and singular continuous spectra. What are the analogs for power series?
\medskip
We would like to thank Shmuel Agmon, John Garnett, Jean-Pierre Kahane, Rowan Killip, and Genadi Levin for useful discussions.
\section{Classical Natural Boundaries} \lb{s2}
In this section, we prove Theorem~\ref{T1.3}. The key will be a lemma of M.~Riesz \cite{MRie} used by many other authors
in the study of natural boundaries. The use of right limits and reflectionless power series can be viewed as a tool for
squeezing maximum benefit from Riesz's lemma.
Given a power series $f(z)=\sum_{n=0}^\infty a_n z^n$, where the $a_n$ obey \eqref{1.4}, we define
\begin{equation} \lb{2.1}
f_+^{(N)}(z)=\sum_{n=0}^\infty a_{n+N} z^n \qquad
f_-^{(N)}(z)=\sum_{n=-N}^{-1} a_{n+N} z^n
\end{equation}
so for $z\in\bbD\setminus\{0\}$,
\begin{equation} \lb{2.2}
f_+^{(N)}(z) + f_-^{(N)}(z)=z^{-N} f(z)
\end{equation}
Clearly, $f_-$ is defined and analytic on $\bbC\setminus\{0\}$, $f_+$ is defined initially on $\bbD$, but by \eqref{2.2},
has analytic continuation to any region that $f$ does.
\begin{theorem}[M.~Riesz's Lemma]\lb{T2.1} Suppose $\{a_n\}_{n=0}^\infty$ obeys \eqref{1.4} and that $f$ has an analytic
continuation to a neighborhood of $\bbD\cup S$ where
\begin{equation} \lb{2.3}
S=\{re^{i\theta}\mid 0<r\leq R,\, \alpha\leq\theta\leq\beta\}
\end{equation}
for some $R>1$ and $\alpha <\beta$. Then
\begin{equation} \lb{2.4}
\sup_{\substack{z\in S \\ N=0,1,2,\dots}} \abs{f_+^{(N)}(z)} <\infty
\end{equation}
\end{theorem}
\begin{remark} Riesz's lemma is usually in terms of $z^{-N} (f(z)-\sum_{n=0}^{N-1} a_n z^n)$, but this is eactly
$f_+^{(N)}(z)$.
\end{remark}
\begin{proof} While the result is classical and appears in many places (e.g., \cite{Remm}), we sketch the proof for completeness.
By comparing with a geometric series, for $z\in\bbD$ and all $N$\!,
\begin{equation} \lb{2.5}
\abs{f_+^{(N)}(z)}\leq (1-\abs{z})^{-1} \sup_n\, \abs{a_n}
\end{equation}
and similarly, for $z\in\bbC\setminus\ol{\bbD}$,
\begin{equation} \lb{2.6}
\abs{f_-^{(N)}(z)} \leq (1-\abs{z}^{-1})^{-1} \sup_n\, \abs{a_n}
\end{equation}
Let $\ti S$ have the form of $S$ with $\alpha,\beta$ replaced by $\ti\alpha,\ti\beta$ and $\ti\alpha <\alpha <\beta <\ti\beta$
so that $\ti S$ lies in the neighborhood of analyticity of $f$. Let $z_1=e^{i\ti\alpha}$, $z_2=e^{i\ti\beta}$, and define
\begin{equation} \lb{2.7}
g_\pm^{(N)}(z) = (z-z_1)(z-z_2) f_\pm^{(N)}(z)
\end{equation}
Clearly, \eqref{2.4} is implied by
\begin{equation} \lb{2.8}
\sup_{\substack{z\in\ti S \\ N=0,1,2,\dots}} \abs{g_+^{(N)}(z)}<\infty
\end{equation}
By the maximum principle, we need only check this on $\partial\ti S\setminus\{z_1,z_2\}$.
Because of \eqref{2.5} and the zeros of $g_+^{(N)}$, we have
\begin{equation} \lb{2.9}
\sup_{\substack{z\in\partial\ti S\cap\bbD \\ N=0,1,2,\dots}} \abs{g_+^{(N)}(z)} <\infty
\end{equation}
Similarly, by \eqref{2.6},
\begin{equation} \lb{2.10}
\sup_{\substack{z\in\partial\ti S\cap\bbC\setminus\ol{\bbD} \\ N=0,1,2,\dots }} \abs{g_-^{(N)}(z)} <\infty
\end{equation}
Since $\abs{z^{-N}}<1$ on $\bbC\setminus\ol{\bbD}$, \eqref{2.2} and \eqref{2.10} imply
\begin{equation} \lb{2.11}
\sup_{\substack{z\in\partial\ti S\cap\bbC\setminus\ol{\bbD} \\ N=0,1,2,\dots}} \abs{g_+^{(N)}(z)}<\infty
\end{equation}
proving \eqref{2.8}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T1.3}] Suppose $a_{N_j+n}\to b_n$ for $n\in\bbZ$ and let $f_\pm$ be the functions in
\eqref{1.6}. Then, by estimating Taylor series, we have
\begin{equation} \lb{2.12}
f_+^{(N_j)}(z)\to f_+(z) \qquad
f_-^{(N_j)}(z) \to f_-(z)
\end{equation}
uniformly on compact subsets of $\bbD$ and $\bbC\cup\{\infty\}\setminus\ol{\bbD}$, respectively.
If $f$ has an analytic continuation to a neighborhood of $I=\{e^{i\theta}\mid\alpha_0 <\theta <\beta_0\}$, we can apply
Riesz's lemma for any $S$ of the form \eqref{2.3} with $\alpha_0 <\alpha<\beta<\beta_0$ and some suitable $R>1$ (depending
on $\alpha,\beta)$. Thus, by the Vitali convergence theorem, $f_+^{(N_j)}$ converges uniformly on $S$, so $f_+(z)$
has an analytic continuation to $S$. Moreover, by the analytic continuation of \eqref{2.2} to $S$ and $z^{-N_j}\to 0$
on $\bbC\setminus\bbD$, we see on $S\setminus\ol{\bbD}$,
\begin{equation} \lb{2.13}
f_+(z)+f_-(z)=0
\end{equation}
Thus, $\{b_n\}_{n=-\infty}^\infty$ is reflectionless across $I$.
\end{proof}
The proof shows that if $\Omega_I=\bbC\cup\{\infty\}\setminus(\partial\bbD\setminus I)$ and if $\calR$ is the family of
all right limits and $f_b(z)$ the function on $\Omega_I$ equal to $\sum_{n=0}^\infty b_n z^n$ on $\bbD$, then
\begin{theorem} \lb{T2.2} Fix $I$ an interval in $\partial\bbD$ and $A\in (0,\infty)$. Let $\calR$ be the set of
two-sided sequences reflectionless across $I$ with
\begin{equation} \lb{2.14}
\sup_{-\infty < n < \infty} \abs{b_n} \leq A
\end{equation}
Then, one has that, for any compact $K\subset\Omega_I$,
\begin{equation} \lb{2.15a}
\sup_{b\in\calR}\, \sup_{z\in K}\, \abs{f_b(z)} <\infty
\end{equation}
Moreover, $\calR$ is compact in the topology of uniform convergence on compacts of $\Omega_I$ and on
\begin{equation} \lb{2.15}
\{\{b_n\}_{n=-\infty}^{-1}\mid\{b_n\}_{n=-\infty}^\infty\in\calR\}
\end{equation}
$b_0,b_1,\dots$ are continuous functions of $\{b_n\}_{n=-\infty}^{-1}$. In fact, for any $k<\ell$, there is a homeomorphism
of $\{b_n\}_{n=-\infty}^k$ to $\{b_n\}_{n=\ell}^\infty$ by associating the two ends of a two-sided sequence.
\end{theorem}
\begin{proof} If $K\subset\bbD$ or $K\subset\bbC\setminus\ol{\bbD}$, one can use \eqref{2.5} or \eqref{2.6} for
$f_b$ and $\sup_n \abs{b_n}$. For $K$ straddling $I$, use the argument in the proof of Riesz's lemma and the fact that
$f_b$ is $\sum_{n=0}^\infty b_n z^n$ on $\bbD$ and $-\sum_{n=-\infty}^{-1} b_nz^n$ on $\bbC\setminus\ol{\bbD}$. This
proves \eqref{2.15a}.
Compactness then follows from Montel's theorem and the fact that, by these bounds, the set of reflectionless functions
is closed. The continuity is immediate if one notes that $\{b_n\}_{n=-\infty}^{-1}$ determines $f$ near $\infty$ and
$f$ near $0$ determines an $\{b_n\}_{n=0}^\infty$. The ``in fact'' statement uses that if $\{b_n\}_{n=0}^\infty$ is
reflectionless, so are its translates.
\end{proof}
Since Theorem~\ref{T1.3} says that if some right limit is not reflectionless, then $f$ has a natural boundary, and
since reflectionless $\{b_n\}_{n=-\infty}^\infty$ have one half determining the other half, we have the following,
which implies almost all the natural boundary results of this paper:
\begin{theorem}\lb{T2.3} Let $\{a_n\}_{n=0}^\infty$ be a bounded sequence with two right limits, $\{b_n\}_{n=-\infty}^\infty$
and $\{c_n\}_{n=-\infty}^\infty$, that obey
\begin{equation} \lb{2.14x}
b_0\neq c_0
\end{equation}
and either for some $k>0$ and all $j\geq k$ or for some $k<0$ and all $j\leq k$,
\begin{equation} \lb{2.15x}
b_j=c_j
\end{equation}
Then $\sum_{n=0}^\infty a_n z^n$ has a natural boundary.
\end{theorem}
\begin{remarks} 1. For example, if $k>0$, we look at the right limits $\{b_{n-k}\}_{n=-\infty}^\infty$ and
$\{c_{n-k}\}_{n=-\infty}^\infty$ which have the same $f_+$'s but unequal $f_-$'s.
\smallskip
2. In the next section, we extend this to conclude strong natural boundaries.
\end{remarks}
\section{Strong Natural Boundaries} \lb{s3}
In this section, we prove Theorem~\ref{T1.4}. We suppose that $f$ obeys \eqref{1.7} where $I=(\alpha,\beta)$.
Since \eqref{1.7} implies $\int_0^1 (\int_\alpha^\beta \abs{f(re^{i\theta})} \f{d\theta}{2\pi})\, dr <\infty$,
for a.e.\ $\theta_0$ in $(\alpha,\beta)$, we have $\int_0^1 \abs{f(re^{i\theta})}\, dr <\infty$. So, for a sequence
$\veps_n\downarrow 0$, we have $f\in E(S_n)$, where $S_n=\{re^{i\theta}\mid 0<r<1,\, \alpha+\veps_n <\theta <
\beta-\veps_n\}$ and $E(S_n)$ is the space introduced in \cite[Sect.~10.1]{Duren}. It follows that:
\begin{SL}
\item[(a)] $\lim_{r\uparrow 1} f(re^{i\theta}) = f(e^{i\theta})$ exists for a.e.\ $\theta\in(\alpha,\beta)$
(\cite[Thm.~10.3]{Duren}).
\item[(b)] For all $\veps$, $\int_{\alpha+\veps}^{\beta-\veps} \abs{f(e^{i\theta})}\f{d\theta}{2\pi} <\infty$
(\cite[Thm.~10.3]{Duren}).
\item[(c)] For every $\veps>0$,
\begin{equation} \lb{3.1}
\lim_{r\uparrow 1} \int_{\alpha+\veps}^{\beta-\veps} \abs{f(re^{i\theta}) - f(e^{i\theta})}\, \f{d\theta}{2\pi} =0
\end{equation}
\item[(d)] If we define
\begin{equation} \lb{3.2}
F(z)=\int_{\alpha+\veps}^{\beta-\veps} f(e^{i\theta}) (e^{i\theta}-z)^{-1}\, \f{d\theta}{2\pi i}
\end{equation}
then $F$ is analytic in $\bbC\setminus\{e^{i\theta}\mid \alpha+\veps <\theta < \beta-\veps\}$ and
\begin{equation} \lb{3.3}
\lim_{r\uparrow 1}\, F(re^{i\theta}) -\lim_{r\downarrow 1}\, F(re^{i\theta}) =f(e^{i\theta})
\end{equation}
(follows from \cite[Thm.~10.4]{Duren}).
\end{SL}
We also need the following Painlev\'e-type theorem, which follows easily from Morera's theorem:
\begin{SL}
\item[(e)] If $f_+$ is analytic in $\bbD$, $f_-$ in $\bbC\setminus\ol{\bbD}$, and
\begin{equation} \lb{3.4}
\sup_{0<r<1} \int_\alpha^\beta \abs{f_+(re^{i\theta})}\, \f{d\theta}{2\pi} +
\sup_{1<r<2} \int_\alpha^\beta \abs{f_-(re^{i\theta})}\, \f{d\theta}{2\pi} <\infty
\end{equation}
and if for a.e.\ $\theta\in(\alpha,\beta)$,
\begin{equation} \lb{3.5}
f_+ (e^{i\theta}) = f_-(e^{i\theta})
\end{equation}
then there is $G$ analytic in $\bbC\setminus [\partial\bbD\setminus\{e^{i\theta}\mid\alpha<\theta<\beta\}]$ so that
$G=f_+$ on $\partial\bbD$ and $f_-$ on $\bbC\setminus\ol{\bbD}$.
\end{SL}
\begin{proof}[Proof of Theorem~\ref{T1.4}] Let $F$ be given by \eqref{3.2} and define
\begin{equation} \lb{3.6}
b_n = \int_{\alpha+\veps}^{\beta-\veps} e^{-in\theta} f(e^{i\theta})\, \f{d\theta}{2\pi}
\end{equation}
Then, by expanding $(e^{i\theta}-z)^{-1}$ in suitable geometric series, the Taylor expansion of $F$ near zero is
$\sum_{n=0}^\infty b_n z^n$ and near $\infty$ is $\sum_{n=-\infty}^{-1} b_n z^n$. Moreover, by the Riemann--Lebesgue
lemma and (b) above,
\begin{equation} \lb{3.7}
\lim_{\abs{n}\to\infty}\, b_n =0
\end{equation}
Let $c_n=a_n-b_n$ and
\begin{equation} \lb{3.8}
f_+(z)=\sum_{n=0}^\infty c_n z^n
\end{equation}
Then, by (a), for a.e.\ $\theta\in(\alpha+\veps,\beta-\veps)$,
\begin{equation} \lb{3.9}
\lim_{r\uparrow 1} f_+ (re^{i\theta}) = f(e^{i\theta}) -\lim_{r\uparrow 1}\, F(re^{i\theta})
\end{equation}
and if
\begin{equation} \lb{3.10}
f_-(z) = -\sum_{n=-\infty}^{-1} b_n z^n
\end{equation}
then, for a.e.\ $\theta\in(\alpha+\veps,\beta-\veps)$,
\begin{equation} \lb{3.11}
\lim_{r\downarrow 1}\, f_-(z) =-\lim_{r\downarrow 1}\, F(re^{i\theta})
\end{equation}
It follows that \eqref{3.5} holds, so $f_+$ has a classical analytic continuation across $(\alpha+\veps,
\beta-\veps)$. By Theorem~\ref{T1.3}, every right limit of $c_n$ is reflectionless on $(\alpha+\veps, \beta-\veps)$.
But, by \eqref{3.7}, the right limits of $c_n$ and $a_n$ are the same! Thus, each right limit of $a_n$ is reflectionless
on each $(\alpha+\veps,\beta-\veps)$, and so on $(\alpha,\beta)$.
\end{proof}
\begin{theorem}\lb{T3.1} If the hypotheses of Theorem~\ref{T2.3} hold, then $\sum_{n=0}^\infty a_n z^n$ has a
strong natural boundary on $\partial\bbD$.
\end{theorem}
\begin{example} \lb{E3.2} The Rudin--Shapiro \cite{Rudin,Shap} sequence is defined by defining polynomials $P_n$ and
$Q_n$ recursively by $P_0(z)=Q_0(z)=1$, $P_{n+1}(z) = P_n(z)+z^{2^n} Q_n(z)$, $Q_{n+1}(z)=P_n(z)- z^{2^n} Q_n(z)$. As
power series with bounded coefficients, $\lim\, P_n(z)$ exists and defines a series $f(z)=\sum_{n=0}^\infty a_n z^n$,
where each $a_n$ is $=1$ or $-1$. By using Szeg\H{o}'s theorem, Brillhart \cite{Brill} proved this function had a natural
boundary. Here is an elementary direct proof. $P_{n+1}$ is $P_{n-1} Q_{n-1} P_{n-1} (-Q_{n-1})$. Taking right limits at
the end of the $P_{n-1}$ in $P_{n-1} Q_{n-1}$ and in $P_{n-1}(-Q_{n-1})$ yields right limits which agree at negative
index but have opposite signs at positive index.
\qed
\end{example}
\section{Gap Theorems} \lb{s4}
In this section, we'll prove Theorem~\ref{T1.6} and resolve the apparent contradiction to P\'olya \cite{Polya}--Erd\"os
\cite{Erdos}. The following proof shows the power of reflectionless theorems.
\begin{proof}[Proof of Theorem~\ref{T1.6}] By compactness, we can find a subsubsequence, $n_{j_\ell}$, so $\lim_{\ell\to\infty}
a_{n_{j_\ell}+k}=b_k$ exists for all $k$ and $b_0\neq 0$, $b_k=0$ for $k<0$. For such a right limit, $f_-(z)=0$, but $f_+(0)
=b_0\neq 0$. Thus, $f_-$ cannot be an analytic continuation of $f_+$ through any $I$, and this $\{b_n\}_{n=-\infty}^\infty$
is not reflectionless on any $I$. Theorem~\ref{T1.4} completes the proof.
\end{proof}
It is simplest to resolve the apparent contradiction with \cite{Polya,Erdos} in the context of an example.
\begin{example}\lb{E4.1} Let $U=\cup_{j=2}^\infty \{n\mid j!\leq n\leq j! + j\}$. We want to consider bounded power series
with $a_n=0$ if $n\in U$\!. If $a_{j!+j+1}=1$ for $j\geq 2$ and $a_n$ is arbitrary but bounded for $n\notin U\cup
\{j!+j+1\}_{j=2}^\infty$, then $f(z)$ has a strong natural boundary on $\partial\bbD$. The zero values may be only on
$U$\!, which is a set with zero density.
On the other hand, \cite{Polya,Erdos} say that since $U$ does not have density $1$, there must be $\{a_n\}_{n=1}^\infty$
with $a_n=0$ for $n\in U$ so that $\partial\bbD$ is not a natural boundary.
The resolution is that our natural boundary examples have hard edges, that is, $a_n$ jumps at the edges of $U$\!, while
the examples of \cite{Erdos} have soft edges. $\bbZ_+\setminus U$ has longer and longer nonzero intervals and Erd\"os'
examples ramp up slowly and down slowly to be $1$ in the center of these intervals. It is easy to see that these examples
have right limits which are constant, and so reflectionless!
These examples of Erd\"os are reminiscent of the sparse potentials of Molchanov \cite{Mol} and Remling \cite{Rem}, where
approximate solitons are placed in between long gaps. In fact, given the chronology, we should say the examples of \cite{Mol,Rem}
are reminiscent of Erd\"os \cite{Erdos}!
\qed
\end{example}
As noted by Agmon \cite{Agmon1, Agmon}, it isn't important that $a_{N_j+n}\to 0$ as $j\to 0$ from gaps, only that its $\limsup$
decays exponentially fast. Using right limits, this is easy to see since
\begin{theorem}\lb{T4.2} Let $\{b_n\}_{n=-\infty}^\infty$ be a two-sided series which is reflectionless on some interval,
$I\subset\partial\bbD$. Suppose that for some $C,D>0$, we have that
\begin{equation} \lb{4.1}
\abs{b_n} \leq Ce^{-Dn} \qquad\text{for } n>0
\end{equation}
Then $b_n\equiv 0$.
\end{theorem}
\begin{proof} By \eqref{4.1}, $f_+(z)$ has an analytic continuation to the circle $\{z\mid\abs{z} <e^D\}$. Since $f_+ =
-f_-$ in a neighborhood of $I$, we conclude that $f_+$ defines an entire function. Since $\abs{f_-(z)}\to 0$ as $\abs{z}\to
\infty$, $f_+\equiv 0$, so $f_-\equiv 0$ also, and then $b_n\equiv 0$.
\end{proof}
This immediately implies the following extension of Theorem~\ref{T1.6}:
\begin{theorem} \lb{T4.3} Suppose $\{a_n\}_{n=0}^\infty$ obeys \eqref{1.4} and there exists $n_j\to\infty$ so that for
some $C,D>0$ and for all $k<0$,
\begin{gather}
\limsup_{j\to\infty}\, \abs{a_{n_j+k}} \leq Ce^{-D\abs{k}} \lb{4.2} \\
\liminf \, \abs{a_{n_j}} >0 \lb{4.3}
\end{gather}
Then $f(z)=\sum_{n=0}^\infty a_n z^n$ has a strong natural boundary on $\partial\bbD$.
\end{theorem}
\begin{proof} By compactness, there exists a right limit, so \eqref{4.1} holds for $n<0$ and $b_0\neq 0$. If
$\{b_n\}_{n=-\infty}^\infty$ is reflectionless on $I$, then $\{b_{-n}\}_{n=-\infty}^\infty$ is reflectionless on
$\bar I=\{z\mid\bar z\in I\}$, so Theorem~\ref{T4.2} implies $b_n\equiv 0$ if this right limit is reflectionless.
But $b_0\neq 0$.
\end{proof}
\section{Szeg\H{o}'s Theorem} \lb{s5}
In this section, we'll prove
\begin{theorem}\lb{T5.1} Let $\sum_{n=0}^\infty a_n z^n$ where the values of $\{a_n\}$ lie in a finite set, $F$\!. Then
either $f(z)=\sum_{n=0}^\infty a_n z^n$ has a strong natural boundary or $a_n$ is eventually periodic, in which case
$f$ is a rational function with poles at roots of unity.
\end{theorem}
\begin{remark} This result for ordinary natural boundaries is due to Szeg\H{o} \cite{Sz1922}. That $f$ is unbounded on any
sector is due to Duffin--Schaeffer \cite{DuSc}, and that there is a strong natural boundary is a result of Boas \cite{Boas}.
This is an analogy of spectral theory results of Kotani \cite{Kot89} and Remling \cite{Rem}. One could use an argument of Kotani
and our reflectionless machinery to prove Theorem~\ref{T5.1}, but instead we'll borrow part of Boas' argument and note
one could use that to find an alternate proof of the spectral theory results.
\end{remark}
\begin{proof} Let $V$ be the finite set of possible values of $a_n$. Suppose $\{a_n\}$ is not eventually periodic.
Fix $p=1,2,\dots$ and consider $p$ blocks, $\{a_j\}_{j=\ell p+1}^{(\ell +1)p}$ for $\ell=0,1,2,\dots$. Since there are
only $(\#V)^p$ possible $p$ blocks, some value must recur, that is, there exist $Q_p >P_p$ so that
\begin{equation}\lb{5.1}
a_{Q_p+j} = a_{P_p+j} \qquad j=1, \dots, p
\end{equation}
If \eqref{5.1} holds for all $j\geq 1$, then for $k=P_p + j\geq P_p+1$, we have
\[
a_{(Q_p-P_p)+k} = a_k \qquad k=1, \dots
\]
that is, $a$ is eventually periodic. Since we are assuming the contrary, there is $L_p\geq p+1$, so $a_{Q_p + L_p} \neq
a_{P_p+L_p}$.
Let $N_p=P_p + L_p\to\infty$ as $p\to\infty$ since $L_p\geq p$ and $M_p =Q_p + L_p >N_p+L$. Then
\begin{align}
a_{N_p}+j &= a_{M_p}+j \qquad j=-p, \dots, -1 \lb{5.2} \\
a_{N_p} &\neq a_{M_p} \lb{5.3}
\end{align}
By compactness, we get right limits, $b,c$, obeying \eqref{2.15}/\eqref{2.14x}, so by Theorem~\ref{T3.1}, $\sum_{n=0}^\infty
a_n z^n$ has a strong natural boundary.
\end{proof}
We note the following extension of Szeg\H{o}'s theorem which appears in Bieberbach \cite{Bie} (who only proved classical natural boundary):
\begin{theorem} \lb{T5.1} Let $\{a_n\}_{n=0}^\infty$ be a bounded sequence with finitely many limit points. Then either
$\sum_{n=0}^\infty a_n z^n$ has a strong natural boundary or there is a periodic sequence, $c_n$, with
\begin{equation} \lb{5.4}
\abs{a_n-c_n} \to 0 \qquad\text{as } n\to\infty
\end{equation}
\end{theorem}
\begin{proof} Let $V$ be the finite set of limit points. Let $\gamma=\f12 \min_{x,y\in V,\, x\neq y} \abs{x-y}$.
Eventually, for all $n$, there is $c_n\in V$ with $\abs{a_n-c_n}\leq \gamma$. It follows that \eqref{5.4} holds. If $c_n$
is eventually periodic, it can be modified for small $n$ to be periodic and \eqref{5.4} still holds.
If $c_n$ is not eventually periodic, by the above, it has a right limit which is not reflectionless. But $a_n$ and $c_n$
have the same right limits.
\end{proof}
\section{Random Power Series} \lb{s6}
In this section, we'll prove Theorem~\ref{T1.7} as well as
\begin{theorem}\lb{T6.1} Let $\{a_n(\omega)\}_{n=0}^\infty$ be a sequence of independent random variables so that
\begin{SL}
\item[{\rm{(i)}}] $\sup_{n,\omega} \abs{a_n(\omega)} =K<\infty$
\item[{\rm{(ii)}}] For some sequence $n_j\to\infty$,
\begin{equation} \lb{6.1}
\limsup_{j\to\infty}\, [\bbE(\abs{a_{n_j}(\omega)}^2) - \abs{\bbE(a_{n_j}(\omega))}^2] >0
\end{equation}
Then for a.e.\ $\omega$, $\sum_{n=0}^\infty a_n(\omega) z^n$ has a strong natural boundary.
\end{SL}
\end{theorem}
\begin{remark} $\bbE$ is expectation. We'll use $\bbP$ for probability and $\bbV$ for variation, so \eqref{6.1} is
$\limsup_{j\to\infty} \bbV(a_{n_j}(\omega)) >0$.
\end{remark}
\begin{lemma}\lb{L6.2}
\begin{SL}
\item[{\rm{(i)}}] For any $K$ and $m$, there exists $K_m >0$ so that for any random variable $f$ with $\norm{f}_\infty \leq K$\!,
there exists $z\in\bbC$ so that
\begin{equation} \lb{6.2}
\bbP\biggl( \abs{f(\omega)-z}\leq \f{1}{m}\biggr) \geq K_m
\end{equation}
\item[{\rm{(ii)}}] For any $K$\!, $m$, and $\sigma >0$, there exists $\wti K_m >0$ so that for any random variable $f$ with
$\norm{f}_\infty \leq K$ and $\bbV(f)\geq\sigma$, there exists $z,w\in\bbC$ with
\begin{equation} \lb{6.3}
\abs{z-w} \geq\biggl( \f{\sigma}{2}\biggr)^{1/2}
\end{equation}
so that
\begin{equation} \lb{6.4}
\bbP\biggl(\abs{f(\omega)-z} \leq \f{1}{m}\biggr) \geq \wti K_m \qquad
\bbP\biggl( \abs{f(\omega)-w}\leq \f{1}{m}\biggr) \geq \wti K_m
\end{equation}
\end{SL}
\end{lemma}
\begin{proof} (i) Cover $\{z\mid\abs{z}\leq K\}$ by disks of radius $\f1m$. By compactness (or simple geometry), one can arrange
for a finite number $N_m$. One of these disks must have probability at least $K_m=(N_m)^{-1}$. \eqref{6.2} holds for the center
of that disk.
\smallskip
(ii) For any $c>0$ and bounded random variable $g$,
\begin{equation} \lb{6.5}
\bbE(\abs{g}^2) \leq c^2\bbP(\abs{g}\leq c) + \norm{g}_\infty^2 \bbP(\abs{g}\geq c)
\end{equation}
Thus, for any $\alpha$ with $\abs{\alpha}\leq\norm{f}_\infty$,
\begin{equation} \lb{6.6}
\var(f) \leq c^2 + (2\norm{f}_\infty)^2 \bbP(\abs{f-\alpha}>c)
\end{equation}
Picking $c\leq (\f{\sigma}{2})^{1/2}$ with $\sigma=\var(f)$, we see that
\begin{equation} \lb{6.7}
\bbP(\abs{f-\alpha}>c)\geq \f{1}{8\norm{f}_\infty^2} \var (f)
\end{equation}
Use (i) to find $z$ so $\bbP(\abs{f(\omega)-z}\leq \f{1}{m})\geq K_m$ where, if necessary, $m$ is increased so
$\f{1}{m} \leq (\f{\sigma}{2})^{1/2}$. By repeating (i) using $\bbP(\abs{f-z}\geq (\f{\sigma}{2})^{1/2}) \geq
\f{1}{8\norm{f}_\infty} \sigma$, we get a $w$ outside the disk $\{\zeta\mid\abs{\zeta-z}\leq(\f{\sigma}{2})^{1/2}\}$
so that $\bbP(\abs{f(\omega)-w}\leq\f{1}{m}) \geq \wti K_m$ for some $\wti K_m\leq K_m$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{T6.1}] Pick $\sigma >0$ and $n_j\to\infty$ so that $\bbV(a_{n_j}(\omega))\geq\sigma$
for all $j$ and so that $n_{j+1} \geq n_j$ and $n_{j+1}-n_j\to\infty$. For each $n\neq n_j$, for all $j$, use (i) of
Lemma~\ref{L6.2} to pick $z_n^{(m)}$ so that
\begin{equation} \lb{6.8}
\bbP\biggl(\abs{a_n(\omega)-z_n^{(m)}} \leq \f{1}{m}\biggr) \geq K_m
\end{equation}
By compactness and using the diagonalization trick, one can pass to a subsequence of the $n_j$, which we'll still
denote by $n_j$,
so for all $m$ and all $k\neq 0$, $z_{n_j+k}^{(m)}\to z_k$ as $j\to\infty$.
By using (ii) of Lemma~\ref{L6.2}, find $w_j^{(m)}, \zeta_j^{(m)}$ so that
\begin{equation} \lb{6.9}
\begin{aligned}
\bbP\biggl( \abs{a_{n_j}(\omega)-w_j^{(m)}} \leq \f{1}{m}\biggr) &\geq\wti K_m \\
\bbP\biggl( \abs{a_{n_j}(\omega)-\zeta_j^{(m)}} \leq \f{1}{m}\biggr) & \geq \wti K_m
\end{aligned}
\end{equation}
and $\abs{w_j^{(m)} - \zeta_j^{(m)}} \geq (\f{\sigma}{2})^{1/2}$. Again, by passing to a subsequence, we may assume
that $w_j^{(m)}\to w$, $\zeta_j^{(m)}\to \zeta$ where, of course, $\abs{\zeta-w} \geq (\f{\sigma}{2})^{1/2}$.
By the Borel--Cantelli lemma, for each $m$ and each $Q$,
\begin{equation} \lb{6.10}
\abs{a_{n_j}(\omega) - w_j^{(m)}} \leq \f{1}{m} \qquad \abs{a_{n_j+k}(\omega) - z_{n_j+k}^{(m)}} \leq \f{1}{m}
\end{equation}
for all $k$ with $0 < \abs{k}\leq Q$, occurs infinitely often for a.e.\ $\omega$, and the same for $\zeta_j^{(m)}$. Thus, for
a.e.\ $\omega$, the right limits include $b,c$ with
\begin{equation} \lb{6.11}
b_k=c_k=z_k \text{ for } k\neq 0 \qquad b_0 =w \qquad c_0 = \zeta\neq w
\end{equation}
It follows from Theorem~\ref{T3.1} that $a_n(\omega)$ has a strong natural boundary for all $\omega$.
\end{proof}
\begin{remark} The use of the Borel--Cantelli lemma to get nonreflectionless limits (in the context of spectral theory of
CMV matrices) is taken from work of Breuer--Ryckman--Zinchenko \cite{BRZ}.
\end{remark}
For our proof of Theorem~\ref{T1.7}, we need two lemmas whose proof we defer to the end of the section.
\begin{lemma}\lb{L6.3} \ Let $\alpha <\beta$ in $(-\pi,2\pi)$ with $\abs{\beta-\alpha}<2\pi$. Then $\{\{a_n\}_{n=0}^\infty
\mid \sup_n \abs{a_n}\leq A$ for all $\veps >0$; $\sup_{0<r<1} (\int_{\alpha+\veps}^{\beta-\veps} \abs{\sum_{n=0}^\infty
a_n r^n e^{in\theta}} \f{d\theta}{2\pi}) <\infty \}$ is a measurable set {\rm{(}}in the product topology{\rm{)}} invariant
under $a_n\to a_{n+1}$.
\end{lemma}
\begin{lemma}\lb{L6.4} Let $\mu$ be a probability measure on $\{\{a_n\}_{n=-\infty}^\infty \mid \sup_n \abs{a_n}\leq A\}$
that defines an ergodic invariant stochastic process. Let $\calR_\mu$ be the set of two-sided series $\{b_n\}_{n=-\infty}^\infty$
so that $\{b_n\}_{n=-\infty}^\infty$ is a right limit of $\{a_n\}_{n=0}^\infty$ with positive probability. Then $\calR_\mu$ is
the support of the measure $\mu$.
\end{lemma}
\begin{proof}[Proof of Theorem~\ref{T1.7}] By Lemma~\ref{L6.3}, Theorem~\ref{T1.4}, and ergodicity (letting $\alpha,\beta$
run through rational multiples of $2\pi$), either $\sum_{n=0}^\infty a_n z^n$ has a strong natural boundary with probability
$1$ or there is an interval, $I$, in $\partial\bbD$ so that with probability $1$, all right limits of $\sum_{n=0}^\infty
a_n z^n$ are reflectionless across $I$. In that case, by Lemma~\ref{L6.4}, all $\{b_n\}_{n=-\infty}^\infty$ in the support
of $\mu$ are reflectionless across $I$. By Theorem~\ref{T2.2}, on $\supp(\mu)$, $b_0$ is a continuous, and so measurable, function
of $\{b_n\}_{n=-\infty}^{-1}$. Thus, the process is deterministic (in the strong sense of there being a continuous, rather than
merely a measurable function).
\end{proof}
\begin{proof}[Proof of Lemma~\ref{L6.3}] By Theorem~10.1 of Duren \cite{Duren}, the space $E$ (sector) can be defined by a
countable family of approximating curves. Thus, letting $z_1^{(n)}=\exp(i(\alpha+\f{1}{n}))$, $z_2^{(n)}=\exp(i(\beta-\f{1}{n}))$,
the $\sup$ condition can be replaced by
\begin{equation} \lb{6.12}
\sup_{m,n}\left( \int_{\alpha+\f{1}{n}}^{\beta-\f{1}{n}} \left. \biggl| (z-z_1^{(n)})(z-z_2^{(n)})
\sum_{n=0}^\infty a_n z^n \biggr| \right|_{z=(1-\f{1}{m})e^{i\theta}} \f{d\theta}{2\pi} \right) <\infty
\end{equation}
Since it is described by a countable $\sup$ of uniformly convergent sums, this set is clearly measurable.
If $f(z)=\sum_{n=0}^\infty a_n z^n$ and $\ti f(z)=\sum_{n=0}^\infty a_{n+1} z^n$, then
\begin{equation} \lb{6.13}
\ti f(z)=z^{-1} (f(z)-a_0 z)
\end{equation}
and finiteness of the $\sup$ in \eqref{6.12} implies finiteness of the $\sup$ for $\ti f$ replacing $f$. This proves the
claimed invariance.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{L6.4}] If $b_n$ is in $\supp(d\mu)$, then for any $\veps$, $N,\mu(\{a_n\mid
\abs{a_{N+n}-b_n}<\veps,\, n=0,\pm 1, \dots, \pm N\})>0$. By the ergodic theorem for a.e.\ $\omega$,
\begin{equation} \lb{6.14}
\lim_{M\to\infty} \, \f{1}{M}\, \#\{j<M\mid \abs{a_{n+j+N}-b_j} <\veps, \, n=0,\pm 1, \dots, \pm N\} >0
\end{equation}
so for a.e.\ $\omega$, there exists $N_\ell \to\infty$ with $\abs{a_{n+N_\ell}-b_n}<\veps$ for all $n$ with $\abs{n}<N$.
By a diagonalization trick, $b_n$ is a right limit for a.e.\ $\omega$.
If $b_n$ is not in the support of $\mu$, pick $N$ and $\veps$ so that $\mu(\{a_n\mid \abs{a_{N+n}-b_n}\leq \veps, \,
\abs{n}\leq N\})=0$. By translation invariance for all $k$,
\begin{equation} \lb{6.15}
\mu(\{a_n\mid \abs{a_{N+k+n}-b_n} \leq \veps, \, n\leq N\})=0
\end{equation}
which implies $b_n$ is not a right limit with probability $1$.
\end{proof}
\section{Hecke's Example} \lb{s7}
Motivated by a spectral theory result of Damanik--Killip \cite{DK}, we prove the following that includes Theorem~\ref{T1.8}
and so, Hecke's example.
\begin{theorem} \lb{T7.1} Let $T\colon\partial\bbD\to\partial\bbD$ be a homeomorphism so that for any $e^{i\theta}\in
\partial\bbD$, $\{T^k (e^{i\theta})\mid k=0,1,\dots\}$ is dense in $\partial\bbD$. Let $f\colon\partial\bbD\to\bbC$
be a bounded and piecewise continuous function with a finite number of discontinuities so that at, at least, one discontinuity,
the right and left limits exist and are not equal. Then for any $e^{i\theta}\in\partial\bbD$, $\sum_{n=0}^\infty
f(T^n (e^{i\theta}))z^n$ has a strong natural boundary on $\partial\bbD$.
\end{theorem}
\begin{remark} For other papers on other extensions of Hecke's example, see \cite{Salem,Newman,Mordell,Schwarz,Meijer,
Popken,CarKem}.
\end{remark}
\begin{proof} By rotating, we suppose $e^{i\theta}=1$ is a point of a discontinuity with $\lim_{\theta\downarrow 0}
f(e^{i\theta})=r\neq s=\lim_{\theta\uparrow 0} f(e^{i\theta})$. By the density of any orbit for any $e^{i\theta}$,
we can find $n_j\to\infty$, so $T^{n_j}(e^{i\theta})\to 1$ with $T^{n_j}(e^{i\theta})=e^{i\psi_j}$ with $-\pi < \psi_j <0$
and $m_j\to\infty$, so $T^{m_j}(e^{i\theta})\to 1$ and $T^{m_j}(e^{i\theta})=e^{i\eta_j}$ with $\pi >\eta_j>0$ (for find
$n_{j+1}>n_j$ so $\abs{T^{n_{j+1}}(e^{i\theta})-e^{-i/(j+1)}}\leq (j+1)^{-2}$).
Thus, $a_{n_j}\to s$ and $a_{m_j}\to r$. On the other hand, look at the orbit $\{T^\ell(1)\mid\ell=1,2,\dots\}$.
Since this orbit is dense, the $T^\ell(1)$ must be distinct, so for some $L$ and all $\ell >L$, $T^\ell(1)$ must be
a point of continuity. Thus, for $\ell >L$, $a_{n_j+\ell}-a_{m_j+\ell}\to 0$. The hypotheses of Theorem~\ref{T2.3}
hold for the right limits defined by $a_{n_j+n}$ and $a_{m_j+n}$. So, by Theorem~\ref{T3.1}, we have a strong natural
boundary.
\end{proof}
\section{Baire Genericity} \lb{s8}
\begin{proof}[Proof of Theorem~\ref{T1.9}] Let $\{a_n\}_{n=0}^\infty\in\Omega^\infty$. Pick two distinct points,
$z_0,z_1\in\Omega$. For any $\ell$, define
\[
a_n^{(\ell)} = \begin{cases}
a_n & n\leq \ell \\
z_0 & n=\ell+k! \text{ for } k=1,2,\dots \\
z_1 & n=\ell +m, \,\, m\geq 1,\,\, m\neq k! \text{ for any } k
\end{cases}
\]
Then, by the gap theorem, $\{a_n^{(\ell)}\}$ has a natural boundary on $\partial\bbD$, indeed, by Weierstrass' original
direct arguments. But $\lim_{\ell\to\infty} a_n^{(\ell)}=a_n$ for each fixed $n$, so the set in the theorem is dense in
the weak topology.
For every rational multiple $\alpha,\beta$ of $2\pi$ with $\alpha <\beta$, and every $K=1,2,\dots$, and any $n=1,\dots$,
let $F_{\alpha,\beta,K,n} = \{\{a_n\}_{n=1}^\infty\in\Omega^\infty \mid f(z)=\sum_{n=0}^\infty a_n z^n$ has an analytic
continuation to $\{z\mid \abs{z}<1+n^{-1},\, \alpha <\arg(z)<\beta\}$ with $\abs{f(z)}\leq K$ there$\}$. It is easy to see that
\begin{SL}
\item[(i)] Each $F_{\alpha,\beta,K,n}$ is closed in the topology of pointwise convergence of the $a_n$'s (by the Vitali
theorem).
\item[(ii)] $\cup F_{\alpha,\beta,K,n}$ is the set of power series for which $\partial\bbD$ is not a natural boundary.
\end{SL}
Thus, the complement of the set in the theorem is an $F_\delta$, so the set is a $G_\delta$.
\end{proof}
\bigskip
|
\section{Introduction}
\vspace{.3cm}
\subsection{A Brief History}\label{subsect-history}
In 1981, Gromov introduced the Gromov-Hausdorff distance between Riemannian
manifolds as an intrinsic version of the Hausdorff distance. Recall that the Hausdorff distance measures distances between subsets in a common metric space \cite{Gromov-metric}.
To measure the distance between Riemannian manifolds, Gromov isometrically
embeds the pair of manifolds into a common metric
space, $Z$, then measures the Hausdorff distance between them in $Z$, and
then takes the
infimum over all isometric embeddings into all common metric spaces, $Z$. Two
compact Riemannian manifolds have $d_{GH}(M_1,M_2)=0$ if and only if they are
isometric. This notion of distance enables Riemannian geometers to study
sequences of Riemannian manifolds which are not diffeomorphic to their limits
and have no uniform lower bounds on their injectivity radii.
The limits of converging sequences of
compact Riemannian manifolds with a uniform upper bound on diameter need not
be Riemannian manifolds at all. However they are
compact geodesic metric spaces.
Gromov's compactness theorem states that a sequence of compact metric
spaces, $X_j$, has a Gromov-Hausdorff converging subsequence to a compact metric
space, $X$, if and only if there is a uniform upper bound on diameter and a uniform upper
bound on the function, $N(r)$, equal to the number of disjoint balls of radius $r$
contained in the metric space. He observes that manifolds with nonnegative
Ricci curvature, for example, have a uniform upper bound on $N(r)$
and thus have converging subsequences \cite{Gromov-metric}.
Such sequences need not have uniform lower bounds on their injectivity radii
(c.f. \cite{Perelman-example}) and their limit spaces can have locally infinite topological type
\cite{Menguy-inf-top-type}. Nevertheless Cheeger-Colding proved these limit
spaces have many intriguing properties which has lead to a wealth of further
research. One particularly relevant result states
that when the sequence also has a uniform lower bound on volume,
then the limit spaces are countably $\mathcal{H}^m$ rectifiable of the
same dimension as the sequence \cite{ChCo-PartIII}.
However, Gromov-Hausdorff convergence does not apply well to
sequences with positive scalar curvature.
In 2004, Ilmanen described the following example of a sequence of
three dimensional spheres with positive scalar curvature which has
no Gromov-Hausdorff converging subsequence. He felt the sequence should
converge in some weak sense to a standard sphere [Figure~\ref{fig-hairy-sphere}].
\begin{figure}[h]
\centering
\includegraphics[width=4.3in]{sw-hairy-sphere-hor.jpg}
\caption{Ilmanen's sequence of increasingly hairy spheres}
\label{fig-hairy-sphere}
\end{figure}
Viewing the Riemannian manifolds in Figure~\ref{fig-hairy-sphere}
as submanifolds of Euclidean space,
they are seen to converge in Federer-Fleming's flat sense
as integral currents to the standard sphere. One of the beautiful properties of limits
under Federer-Fleming's flat convergence is that they are countably $\mathcal{H}^m$
rectifiable with the same dimension as the sequence. In light of Cheeger-Colding's
work, it seems natural, therefore, to look for an intrinsic flat convergence whose
limit spaces would be countably $\mathcal{H}^m$ rectifiable metric spaces.
The intrinsic flat distance introduced in this paper leads to exactly this kind of convergence.
The sequence of $3$ dimensional manifolds depicted in
Figure~\ref{fig-hairy-sphere} does in fact converge to the
sphere in this intrinsic flat sense [Example~\ref{ex-hairy-sphere}].
Ambrosio-Kirchheim's 2000 paper \cite{AK} developing the theory of currents on
arbitrary metric spaces is an essential ingredient for this paper. Without it we
could not define the intrinsic flat distance, we could not define an integral current
space and we could not explore the properties of converging sequences.
Other important
background to this paper is prior work of the second author, particularly
\cite{Wenger-flat}, and a coauthored piece \cite{SorWen1}.
Riemannian geometers may not have read these papers (which are aimed
at geometric measure theorists); so we review key
results as they are needed within.
\subsection{An Overview} \label{subsect-intro}
In this paper, we view a compact oriented Riemannian manifold with boundary, $M^m$,
as a metric space, $(X,d)$, with an integral current, $T \in \intcurr_m(M)$,
defined by integration over $M$: $T(\omega) :=\int_M \omega$. We write
$M=(X,d,T)$ and refer to $T$ as the integral current structure.
Using this structure we can define an intrinsic flat distance
between such manifolds and study the intrinsic flat limits of sequences of
such spaces. As an immediate consequence of
the theory of Ambrosio-Kirchheim,
the limits of converging sequences of such spaces are countably ${\mathcal H}^m$
rectifiable metric spaces, $(X,d)$, endowed with a current structure,
$T \in \intcurr_m(Z)$,
which represents an orientation and a multiplicity on $X$.
In Section 2 we describe these spaces in more detail referring to them as
{\em $m$ dimensional integral current spaces} [Defn~\ref{defn-current-space}]
[Defn~\ref{defn-integral-current-space}].
The class of such spaces is denoted ${\mathcal M}^m$ and includes the
zero current space, denoted $\bf{0}=(0,0,0)$.
Given an integral current space $(X,d,T)$, we define its boundary using the boundary,
$\partial T$, of the integral current structure
[Defn~\ref{defn-integral-current-space}]. We also define the mass
of the space using the mass, ${\mathbf M}(T)$, of the current structure
[Defn~\ref{defn-space-mass}]. When
$(X,d,T)$ is an oriented Riemannian manifold, the boundary is
just the usual boundary and the mass is just the volume.
Recall that the flat distance between $m$ dimensional integral currents
$S,T\in\intcurr_m\left(Z\right)$ is given by
\begin{equation} \label{eqn-Federer-Flat}
d^Z_{F}\left(S,T\right):=
\inf\{{\mathbf M}\left(U\right)+{\mathbf M}\left(V\right):
S-T=U+\partial V \}
\end{equation}
where $U\in\intcurr_m\left(Z\right)$ and $V\in\intcurr_{m+1}\left(Z\right)$.
This notion of a flat distance was first introduced by Whitney
in \cite{Whitney} and later adapted to rectifiable currents by Federer-Fleming \cite{FF}.
The flat distance between integral currents on an arbitrary metric space was
introduced by the second author in \cite{Wenger-flat}.
Our definition of the intrinsic flat distance between elements
of ${\mathcal M}^m$ is modeled after Gromov's intrinsic Hausdorff distance \cite{Gromov-metric}:
\begin{defn} \label{def-flat1}
For $M_1=\left(X_1,d_1,T_1\right)$ and $M_2=\left(X_2,d_2,T_2\right)\in\mathcal M^m$ let the
intrinsic flat distance be defined:
\begin{equation}\label{equation:def-abstract-flat-distance}
d_{{\mathcal F}}\left(M_1,M_2\right):=
\inf d_F^Z
\left(\varphi_{1\#} T_1, \varphi_{2\#} T_2 \right),
\end{equation}
where the infimum is taken over all complete metric spaces
$\left(Z,d\right)$ and isometric embeddings
$\varphi_1 : \left(\bar{X}_1,d_1\right)\to \left(Z,d\right)$ and $\varphi_2: \left(\bar{X}_2,d_2\right)\to \left(Z,d\right)$
and the flat norm $d_F^Z$ is taken in $Z$.
Here $\bar{X}_i$ denotes the metric completion of $X_i$ and $d_i$ is the extension
of $d_i$ on $\bar{X}_i$, while $\phi_\# T$ denotes the push forward of $T$.
\end{defn}
All notions from Ambrosio-Kirchheim's work needed to understand this definition are reviewed in detail in Section 2. As in Gromov, an isometric embedding is a map
$\phi: A \to B$ which preserves distances not just the Riemannian metric tensors:
\begin{equation} \label{def-isom-embed}
d_B\left(\phi\left(x\right),\phi\left(y\right)\right)=d_A\left(x,y\right) \qquad \forall x,y \in A.
\end{equation}
For example a map $f: S^1 \to D^2$ mapping the circle to the boundary of a flat
disk is not an isometric embedding while the map $\varphi:S^1 \to S^2$
mapping the circle to a great circle in the sphere is an isometric embedding.
If the infimum in (\ref{equation:def-abstract-flat-distance}) were taken over maps
preserving the Riemannian metric tensors rather than isometric embeddings in
the sense of Gromov, then the value would not be positive.
It is fairly easy to estimate the
intrinsic flat distances between
compact oriented Riemannian manifolds using standard
methods from Riemannian geometry. If $M^m_1$ and $M^m_2$
are $m$ dimensional Riemannian manifolds which isometrically
embed into an $m+1$ dimensional Riemannian manifold, $V$, such that
the boundary, $\partial V= \varphi(M_1) \sqcup \varphi(M_2) \sqcup U$,
then by (\ref{eqn-Federer-Flat}) we have
$$
d_{\mathcal F}(M_1, M_2) \le \operatorname{Vol}_{m}(U) +\operatorname{Vol}_{m+1}(V).
$$
This technique and others are applied in the Appendix to
explicitly compute the intrinsic flat limits of converging sequences of
manifolds depicted here.
It should be noted that $d_{\mathcal{F}}(M,\bf{0})$ is related
to Gromov's filling volume of a manifold \cite{Gromov-filling}
via \cite{Wenger-flat} and \cite{SorWen1}.
DePauw and Hardt have recently defined a flat norm
a la Gromov for chains in a metric space. When the chain is an isometrically
embedded Riemannian manifold, $M$, then their "flat norm" of $M$
seems to take on the same value as $d_{\mathcal{F}}(M,\bf{0})$
\cite{DH-chains}. \footnote{See \cite{DH-chains} page 20 and page 26.}
In Section 3 we explore the properties of our intrinsic flat distance, $d_\mathcal{F}$.
It is always finite
and, in particular,
satisfies $d_{{\mathcal F}}\left(M_1, M_2\right) \le \operatorname{Vol}\left(M_1\right)+\operatorname{Vol}\left(M_2\right)$
when $M_i$ are compact oriented Riemannian manifolds [Remark~\ref{finite}].
We prove $d_{{\mathcal F}}$ is a distance on
$\mathcal{M}^m_0$, the space of precompact
integral current spaces
[Theorem~\ref{zero-mani} and Theorem~\ref{triangle}].
In particular, for compact oriented Riemannian manifolds,
$M$ and $N$, $d_{{\mathcal F}}\left(M,N\right)=0$ iff there is an orientation
preserving isometry from $M$ to $N$.
Applying the Compactness Theorem of Ambrosio-Kirchheim, we see that when a
sequence of Riemannian manifolds, $M_j$, has volume uniformly bounded above
and converges in the Gromov-Hausdorff sense
to a compact metric space, $Y$, then a subsequence of the $M_j$ converges to an integral
current space, $X$, where $X \subset Y$ [Theorem~\ref{GH-to-flat}].
Example~\ref{example-one-hair} depicted in Figure~\ref{figure-one-hair},
demonstrates that the intrinsic flat and Gromov-Hausdorff limits
need not always agree: the Gromov-Hausdorff
limit is a sphere with an interval attached while the intrinsic flat limit is just the sphere.
\begin{figure}[h]
\centering
\includegraphics[width=4.7in]{sw-straight-hair.jpg}
\caption{A sphere with a disappearing hair [Ex~\ref{example-one-hair}].}
\label{figure-one-hair}
\end{figure}
Gromov-Hausdorff limits of Riemannian manifolds are geodesic spaces. Recall that
a geodesic space is a metric space such that
\begin{equation}
d(x,y)=\inf\{ L(c): \,\, c \textrm{ is a curve s.t. }c(0)=x, c(1)=y\}
\end{equation}
and the infimum is attained by a curve called a geodesic segment.
In Example~\ref{example-not-length}
depicted in Figure~\ref{figure-not-length},
we show that the intrinsic flat limit of Riemannian manifolds need not
be a geodesic space.
In fact the intrinsic flat limit is not even path connected.
\begin{figure}[h]
\centering
\includegraphics[width=4.7in]{sw-not-length.jpg}
\caption{The intrinsic flat limit is a disjoint pair of spheres [Ex~\ref{example-not-length}].}
\label{figure-not-length}
\end{figure}
While the limit spaces are not geodesic spaces, they are countably $\mathcal{H}^m$
rectifiable metric spaces of the same dimension.
These spaces, introduced and studied by Kirchheim in \cite{Kirchheim}, are
covered almost everywhere by the bi-Lipschtiz charts of Borel sets in $\mathbb{R}^m$.
Gromov-Hausdorff limits do not in general have rectifiability properties.
An interesting example
of such a space is depicted in Figure~\ref{figure-gym2} [Example~\ref{example-gym2}].
The intrinsic flat limit depicted here is the disjoint collection of spheres while the
Gromov-Hausdorff limit has line segments between them.
\begin{figure}[h]
\centering
\includegraphics[width=4.7in]{sw-gym2.jpg}
\caption{The limit is a countably ${\mathcal H}^m$ rectifiable space
[Ex~\ref{example-gym2}].}
\label{figure-gym2}
\end{figure}
If a sequence of Riemannian manifolds, $M^m_j$,
has volume converging to 0 or has a Gromov-Hausdorff
limit whose dimension is less than $m$, then the intrinsic flat limit is the zero space
[Remark~\ref{vol-to-zero} and Corollary~\ref{lower-dim}].
See Figure~\ref{figure-tori-GH} [Example~\ref{example-tori-GH}]. Such
sequences are referred to as collapsing sequences.
\begin{figure}[h]
\centering
\includegraphics[width=4.7in]{tori-GH.jpg}
\caption{The Gromov-Hausdorff limit is lower dimensional and the intrinsic flat limit is the zero space [Example~\ref{example-tori-GH}].}
\label{figure-tori-GH}
\end{figure}
Sequences may also converge to the zero integral current space due to an effect called
cancellation. With significantly growing local topology,
a sequence of $M^m_j$ which Gromov-Hausdorff converges to a Riemannian manifold,
$X$, of the same dimension might cancel with itself so that $Y=0$.
In \cite{SorWen1}, the authors gave an example of two
standard three dimensional spheres joined together by increasingly dense tunnels,
providing a sequence of compact manifolds of positive scalar curvature
which converges in the Gromov-Hausdorff sense to a standard sphere.
However the sequence could be isometrically embedded into a common space
$\varphi_j: M_j\to Z$ such that $\varphi_{j\#}M_j$ converges in the flat sense to
$0$ due to cancellation. Thus $M_j \stackrel {\mathcal{F}}{\longrightarrow} \bf{0}$.
In Figure~\ref{figure-cancels} we depict a two dimensional example. Here
two sheets are joined together by many tunnels so that they isometrically
embed into the boundary of a Riemannian manifold of arbitrarily
small volume.
\begin{figure}[h]
\centering
\includegraphics[width=4.7in]{sw2-cancellation-bw.jpg}
\caption{A sequence converging in the intrinsic flat sense to the zero space
due to cancellation [Example~\ref{example-cancels}].}
\label{figure-cancels}
\end{figure}
It is also possible for a sequence of Riemannian manifolds with increasing
local topology to overlaps with itself so that the limit $Y=2X$
[Example~\ref{example-doubles}]. If one provides a twist in the middle of
each tunnel in Figure~\ref{figure-cancels} so as to flip the orientation
of one of the two sheets, then the sequence of manifolds doesn't cancel in
the limit but instead doubles. We say the limit space has weight or multiplicity
$2$. In general, intrinsic flat limit spaces have a weight function, which is an
integer valued Borel measurable function, just like integral currents
[Defn~\ref{def-weight}].
In Section 4 we examine the properties of intrinsic flat convergence.
We first have a section proving that converging and Cauchy sequences
embed into a common metric space. This allows us to then immediately
extend properties of weakly converging sequences of integral currents
to integral current spaces. In particular the mass is lower semicontinuous
as in Ambrosio-Kirchheim \cite{AK} and the the filling volume
is continuous as in \cite{Wenger-flat}.
When $M_j^m$ have nonnegative Ricci curvature, the intrinsic flat limits and
Gromov-Hausdorff limits agree \cite{SorWen1}. In this sense one may think of
intrinsic flat convergence as a means of extending to a larger class of manifolds the
rectifiability properties already proven by Cheeger-Colding to hold on Gromov-Hausdorff
limits of noncollapsing sequences of such manifolds \cite{ChCo-PartI}.
When $M^m_j$
have a common lower bound on injectivity radius or a uniform linear local
contractibility radius, then work of Croke applying Berger's volume estimates
and work of Greene-Petersen applying Gromov's filling volume inequality imply that a subsequence
of the $M^m_j$ converge in the Gromov-Hausdorff sense
\cite{Croke-inj}\cite{Greene-Petersen}. In \cite{SorWen1}, the
authors proved cancellation does not occur in that setting either, so that the
Gromov-Hausdorff limit $X$ agrees with the flat limit $Y$ and is countable ${\mathcal H}^m$
rectifiable.
The second author has
proven a compactness theorem:
{\em Any sequence of oriented Riemannian manifolds with
boundary, $M_j^m$, with a uniform upper bound on $\operatorname{diam}\left(M_j^m\right)$,
$\operatorname{Vol}_m\left(M_j^m\right)$ and
$\operatorname{Vol}_{m-1}\left(\partial M_j^m\right)$ always has a subsequence which converges in the intrinsic flat
sense to an integral current space} \cite{Wenger-compactness}.
In fact Wenger's compactness theorem holds for integral current spaces.
We do not apply this theorem in this paper except
for a few immediate corollaries given in Subsection 4.5 and occasional footnotes.
Unlike the limits in Gromov's compactness theorem, the sequences in Wenger's
compactness theorem need not
converge to a compact limit space. In Figure~\ref{fig-many-tips}
we see that the limit space need not be precompact even when the
sequence of manifolds has a uniform upper bound
on volume and diameter.
\begin{figure}[h]
\centering
\includegraphics[width=4.7in]{sw-many-tips.jpg}
\caption{Spheres with increasingly thin extra bumps converging to a bounded noncompact limit [Ex~\ref{ex-many-tips}].}
\label{fig-many-tips}
\end{figure}
In Section 5, we describe the relationship between the intrinsic flat convergence
of Riemannian manifolds and other forms of convergence including $C^\infty$
convergence, $C^{k,\alpha}$ convergence, and
Gromov's Lipschitz convergence.
In the Appendix by the first author,
we include many examples of sequences explicitly proving they
converge to their limits. Although the examples are referred to throughout
the textbook, they are deferred to the final section so that proofs of convergence
may apply any or all lemmas proven in the paper.
While we do not have room in this introduction to refer to all the results presented here,
we refer the reader to the contents at the beginning of the paper and we introduce each
section with a more detailed description of what is contained within it. Some sections
mention explicit open problems and conjectures.
\subsection{Recommended Reading}
For Riemannian geometry recommended background is a standard
one semester graduate course. For metric
geometry background, the beginning of Burago-Burago-Ivanov \cite{BBI}
is recommended or Gromov's classic \cite{Gromov-metric}. For geometric measure
theory a basic guide to Federer is provided in Morgan's textbook \cite{Morgan}.
One may also consult Lin-Yang \cite{Lin-Yang-text}.
We try to cover what is needed from Ambrosio-Kirchheim's seminal
paper \cite{AK}, but we recommend that paper as well.
\subsection{Acknowledgements}
The first author would like to thank Columbia for its
hospitality in Spring-Summer 2004 and Ilmanen for many
interesting conversations at that time regarding the necessity of
a weak convergence of Riemannian manifolds and what properties
such a convergence ought to have. She would also like to
thank Courant Institute for its hospitality in Spring 2007 and
Summer 2008 enabling the two authors first to develop the
notion of the intrinsic flat distance between Riemannian manifolds
and later to develop the notion of an integral current space in
general extending their prior results to this setting.
The second author would like to thank Courant Institute for
providing such an excellent research environment. The first
author would also like to thank Paul Yang, Blaine Lawson, Steve Ferry
and Carolyn Gordon for their comments on the 2008 version of the
paper, as well as the participants in the CUNY 2009 Differential
Geometry Workshop\footnote{Marcus Khuri, Michael Munn, Ovidiu Munteanu,
Natasa Sesum, Mu-Tao Wang, William Wylie} for suggestions
leading to many of the examples added as an appendix that summer.
\section{{\bf Defining Current Spaces}} \label{Sect-Def-Current-Space}
In this section we introduce current spaces $\left(X,d,T\right)$.
Everything in this section is a reformulation of Ambrosio-Kirchheim's theory
of currents on metric spaces, so that we may clearly define the new notions an
integer rectifiable current space [Defn~\ref{defn-current-space}] and an integral
current space [Defn~\ref{defn-integral-current-space}].
Experts in the theory of Ambrosio-Kirchheim
may wish to skip to these definitions. In Section~\ref{sect-flat-distance} we will discuss the
intrinsic flat distance between such spaces. This section is aimed at Riemannian Geometers
who have not yet read Ambrosio-Kirchheim's work \cite{AK}.
In Subsection~\ref{Subsect-Def-Current-Space-1}, we provide a description of these spaces as weighted oriented countably
$\mathcal{H}^m$-rectifiable metric spaces. Our
spaces need not be complete but must be "completely settled" as defined in Definition~\ref{def-settled}.
In Subsections~\ref{Subsect-Def-Current-Space-2} and~\ref{subsect-param}, we review Ambrosio-Kirchheim's integer rectifiable currents
on complete metric spaces, emphasizing a parametric perspective and proving a
couple lemmas regarding
this parametrization.
In Subsection~\ref{Subsect-Def-Current-Space-3}, we introduce the notion of an integer rectifiable current structure on a metric space [Definition~\ref{defn-current-space}]
and prove in Proposition~\ref{prop-current-spaces} that metric spaces with such current structures are exactly the completely settled
weighted oriented rectifiable metric spaces defined in the first subsection.
In Subsection~\ref{Subsect-Def-Current-Space-4}, we introduce the notion of the boundary of a current space
and define integral current spaces [Definition~\ref{defn-integral-current-space}].
\subsection{Weighted Oriented Countably $\mathcal{H}^m$ Rectifiable Metric Spaces} \label{Subsect-Def-Current-Space-1}
We begin with the following standard definition (\cite{Federer} c.f. \cite{AK}):
\begin{defn} \label{def-rectspace}
A metric space $X$ is called
countably $\mathcal{H}^m$ rectifiable iff there exists countably many Lipschitz maps $\varphi_i$ from Borel measurable
subsets $A_i \subset \mathbb{R}^m$ to $X$ such that the Hausdorff measure
\begin{equation}
\mathcal{H}^m\left( X\setminus \bigcup_{i=1}^\infty \varphi_i\left(A_i\right) \right)=0.
\end{equation}
\end{defn}
\begin{rmrk} \label{defn-metric-differential}
Note that Kirchheim \cite{Kirchheim} defined a metric differential for Lipschitz maps $\varphi:A\subset \mathbb{R}^k \to Z$ where $Z$ is a metric space. When $A$ is open,
\begin{equation} \label{eqn-metric-differential}
md\varphi_y\left(v\right) :=\lim_{r\to 0} \frac {d\left(\varphi\left(y+rv\right),\varphi\left(y\right)\right)}{r},
\end{equation}
if the limit exists.
In fact Kirchheim has proven that for almost every $y \in A$,
$md\varphi_y\left(v\right)$ is defined for all $v \in \mathbb{R}^m$ and $md\varphi_y$ is a seminorm.
On a Riemannian manifold $Z$ with a smooth map $f$, $md f_y\left(v\right)=|df_y\left(v\right)|$. See also Korevaar-Schoen \cite{Korevaar-Schoen}.
\end{rmrk}
In \cite{Kirchheim}, Kirchheim proved
this collection of charts can be chosen so that the maps $\varphi_i$ are bi-Lipschitz.
So we may extend the Riemannian notion of an atlas to this setting:
\begin{defn} \label{defn-atlas-of-X}
A bi-Lipschitz collection of charts, $\{\varphi_i\}$, is called an {\bf atlas} of $X$.
\end{defn}
\begin{rmrk} \label{rmrk-tan-space-is-norm-space} {\em \bf }
Note that when $\varphi: A \subset \mathbb{R}^m \to X$ is bi-Lipschitz, then
$md\varphi_y$ is a norm on $\mathbb{R}^m$. In fact there is a notion of
an approximate tangent space at almost every $y\in X$ which is
a normed space of dimension $m$ whose norm is defined by the metric
differential of a well chosen bi-Lipscitz chart. (c.f. \cite{Kirchheim})
\end{rmrk}
Recall that by Rademacher's Theorem
we know that given a Lipschitz function $f: \mathbb{R}^m \to \mathbb{R}^m$, $\nabla f$ is defined $\mathcal{H}^m$ almost everywhere.
In particular given two bi-Lipschitz charts, $\varphi_i, \varphi_j$, $\det [\nabla\left(\varphi_i^{-1} \circ {\varphi}_j\right)]$ is
defined almost everywhere. So we can extend the Riemannian definitions
of an atlas and an oriented atlas to countably $\mathcal{H}^m$ rectifiable spaces:
\begin{defn} \label{def-orient-atlas}
An atlas on a countably $\mathcal{H}^m$ rectifiable space
$X$ is called an {\bf oriented atlas} if the orientations agree on all overlapping charts:
\begin{equation} \label{eqn-def-oriented}
\det \left[\nabla\left(\varphi_i^{-1} \circ {\varphi}_j\right)\right] >0
\end{equation}
almost everywhere
on $A_j \cap \varphi^{-1}_j \left( \varphi_i\left(A_i\right) \right)$.
\end{defn}
\begin{defn} \label{def-orientation}
An {\bf orientation} on a countably $\mathcal{H}^m$ rectifiable space
$X$ is an equivalence class of atlases where two atlases, $\{\varphi_i\}, \{\bar{\varphi}_j\}$ are considered to be equivalent if
their union is an oriented atlas.
\end{defn}
\begin{rmrk}
Given an orientation $[\{\varphi_i\}]$, we can choose a representative atlas such that the charts are pairwise disjoint,
$\varphi_i(A_i)\cap\varphi_j(A_j)=\emptyset$,
and the domains $A_i$ are precompact. We call such an oriented atlas
a preferred oriented atlas.
\end{rmrk}
\begin{rmrk} \label{Lip-mani-charts}
Orientable Riemannian manifolds and, more generally, connected orientable
Lipschitz manifolds have only two
standard orientations because they are connected metric spaces and
their charts overlap.
Countably $\mathcal{H}^m$ rectifiable
spaces may have uncountably many orientations as each disjoint
chart may be flipped on its own.
Recall that a Lipschitz manifold is a metric space, $X$, such that for all $x \in X$ there
is an open set $U$ about $x$ with a bi-Lipschitz homeomorphism to the open unit ball in Euclidean
space.
A Lipschitz
manifold is said to be orientable when the bi-Lipschitz maps can be chosen so that
(\ref{eqn-def-oriented}) holds for all pairs of charts.
When we say "oriented", we will mean that the orientation has been provided, and
we will always orient Riemannian manifolds and Lipschitz manifolds according to one
of their two standard orientations, and we will always assign them an atlas restricted
from the standard charts used to define them as manifolds.
\end{rmrk}
\begin{defn} \label{def-weight}
A {\bf multiplicity} function (or weight) on a countably $\mathcal{H}^m$ rectifiable space
$X$ with $\mathcal{H}^m(X)<\infty$
is a Borel measurable function $\theta: X \to \mathbb{N}$ whose weighted volume,
\begin{equation}
Vol\left(X, \theta\right):=\int_X \theta d\mathcal{H}^m,
\end{equation}
is finite.
\end{defn}
Note that on a Riemannian manifold, with multiplicity $\theta=1$,
the weighted volume is the volume. Later we will
define the mass of these spaces which will agree with the weighted volume on Riemannian manifolds with
arbitrary multiplicity functions but will not be equal to the weighted volume for
more general spaces.
\begin{rmrk}
Given a multiplicity function and an atlas, one may refine the atlas so that the multiplicity function is constant on the image of
each chart.
\end{rmrk}
Recall the notion of the lower $m$-dimensional density,
$\theta_{*m}(\mu,p)$, of a Borel measure $\mu$ at $p\in X$ is defined by
\begin{equation} \label{eqn-lower-density}
\Theta_{*m}\left(\mu, p\right):= \liminf_{r\to 0} \frac{\mu(B_p(r))}{\omega_m r^m}.
\end{equation}
We introduce the following new concept:
\begin{defn} \label{def-settled}
A weighted oriented countably $\mathcal{H}^m$ rectifiable metric space, $\left(X, d, [\{\phi_i\}], \theta\right)$, is
called {\bf completely settled} iff
\begin{equation}
X=\{p\in \bar{X}: \,\, \Theta_{*m}\left(\theta \mathcal{H}^m, p\right)>0\}.
\end{equation}
\end{defn}
\begin{example} \label{example-settled}
An oriented Riemannian manifold with a conical singular point and constant
multiplicity $\theta=1$, which includes the singular point, is a completely
settled space. An oriented Riemannian manifold with a cusped singular point and
constant multiplicity $\theta=1$, which does not
include the singular point is a completely settled space. In particular a completely settled space need not be complete.
An oriented Riemannian manifold with a cusped singular point $p$
and a multiplicity function, $\theta$, approaching infinity at $p$ such that
\begin{equation}
\lim_{r\to 0} \frac{1}{r^m}\int_{B_p\left(r\right)} \theta \, d\mathcal{H}^m >0
\end{equation}
is completely settled only if it includes $p$.
\end{example}
In Subsection~\ref{Subsect-Def-Current-Space-3} we will define our
current spaces as metric
spaces with current structures. We will prove in Proposition~\ref{prop-current-spaces} that a metric space is a nonzero integer rectifiable current space iff it
is a completely settled weighted oriented countably $\mathcal{H}^m$-rectifiable metric space. Note that the notion of a completely settled space does not appear
in Ambrosio-Kirchheim's work and is introduced here to allow us to understand
current spaces in an intrinsic way.
Integral current spaces will have an added condition that their boundaries
are integer rectifiable metric spaces as well.
\subsection{Reviewing Ambrosio-Kirchheim's Currents on Metric Spaces} \label{Subsect-Def-Current-Space-2}
In this subsection we review all definitions and theorems of Ambrosio-Kirchheim
and Federer-Fleming necessary to define current structures on metric spaces
\cite{AK}\cite{FF}.
For readers familiar with the Federer-Fleming theory of currents
one may recall that an $m$ dimensional current, $T$,
acts on smooth $m$ forms (e.g. $\omega=f d\pi_1\wedge\cdots \wedge d\pi_m$).
An integer rectifiable current is defined by integration over a rectifiable set
in a precise way with integer weight
and the notion of the boundary of $T$ is defined as in
Stokes theorem: $\partial T(\omega)=T(d\omega)$.
This approach extends naturally
to smooth manifolds but not to metric spaces which do not have differential
forms.
In the place of differential forms, Ambrosio-Kirchheim use
$m+1$ tuples, $\omega\in \mathcal{D}^m(Z)$,
\begin{equation}
\omega=f\pi=\left(f,\pi_1 ..., \pi_m\right) \in \mathcal{D}^m(Z)
\end{equation}
where
$f: Z \to \mathbb{R}$ is a bounded Lipschitz function and
$\pi_i: Z \to \mathbb{R}$ are Lipschitz. They credit this approach to DeGiorgi \cite{DeGiorgi}.
In \cite{AK} Definitions 2.1, 2.2, 2.6 and 3.1, an $m$ dimensional
current $T\in \curr_m(Z)$ is defined
as a multilinear functional on $\mathcal{D}^m(Z)$ such that
$T\left(f,\pi_1,..., \pi_m\right) $ satisfies a variety of functional properties similar to $T\left(\omega\right)$ where
$\omega= f d\pi_1 \wedge \dots \wedge d\pi_m$ in the
smooth setting as follows:
\begin{defn}[Ambrosio-Kirchheim]
An
$m$ dimensional {\bf \em current}, $T$, on a complete metric space, $Z$,
is a real valued
{\em multilinear functional} on $\mathcal{D}^m(Z)$, with the
following required properties:
i) {\bf Locality}:
\begin{equation} \label{def-locality}
T(f, \pi_1,..., \pi_m)=0 \textrm{ if }\exists i\in \{1,...m\} \textrm{ s.t. }\pi_i
\textrm{ is constant on a nbd of } \{f\neq0\}.
\end{equation}
ii) {\bf Continuity}:
$$
T \textrm{ is continuous with respect to the pointwise convergence
of the }\pi_i \textrm{ such that } \operatorname{Lip}(\pi_i)\le 1.
$$
iii) {\bf Finite mass}: there exists a
finite Borel measure $\mu$ on $Z$ such that
\begin{equation} \label{def-AK-current-iii}
|T(f,\pi_1,..., \pi_m)| \le \prod_{i=1}^m \operatorname{Lip}(\pi_i) \int_Z |f| \,d\mu \qquad \forall (f,\pi_1,..., \pi_m)\in \mathcal{D}^m(Z).
\end{equation}
The space of $m$ dimensional currents on $Z$, is denoted, $\curr_m(Z)$.
\end{defn}
\begin{example} \label{basic-current}
Given an $L^1$ function $h: A \subset \mathbb{R}^m \to \mathbb{Z}$, one can define an
$m$ dimensional current $\Lbrack h \Rbrack $
as follows
\begin{equation} \label{def-current-from-function}
\Lbrack h \Rbrack \left(f, \pi\right) := \int_{A \subset \mathbb{R}^m} h f \det\left(\nabla \pi\right) \, d\mathcal{L}^m =\int_{A \subset \mathbb{R}^m} hf \, d\pi_1 \wedge \dots \wedge d\pi_m.
\end{equation}
Given a Borel measurable set, $A\subset \mathbb{R}^m$, the
current $\Lbrack {\bf \rm{1}}_A \Rbrack$ is defined by the indicator function ${\bf \rm{1}}_A:\mathbb{R}^m \to \mathbb{R}$. Ambrosio-Kirchheim prove $\Lbrack h \Rbrack \in \curr_m(Z)$ \cite{AK}.
\end{example}
\begin{rmrk}
Stronger versions of locality and continuity, as well as product and
chain rules are proven in \cite{AK}[Theorem 3.5]. In particular, they prove
\begin{equation}
T(f, \pi_{\sigma(1)},..., \pi_{\sigma(m)})= \operatorname{sgn}(\sigma) T(f, \pi_1,..., \pi_m)
\end{equation}
for any permutation, $\sigma$, of $\{1,2,..., m\}$.
\end{rmrk}
The following definition will allow us to define the most
important currents explicitly:
\begin{defn}[Ambrosio-Kirchheim]
Given a Lipschitz map $\varphi:Z\to Z'$, the {\em push
forward} of a current $T\in \curr_m(Z)$
to a current $\varphi_\# T \in \curr_m(Z')$ is given in \cite{AK}[Defn 2.4] by
\begin{equation} \label{def-push-forward}
\varphi_\#T(f,\pi_1,..., \pi_m):=T(f\circ \varphi, \pi_1\circ\varphi,..., \pi_m\circ\varphi)
\end{equation}
exactly as in Federer-Flemming when everything is smooth.
\end{defn}
\begin{example}\label{basic-current-pushed}
If one has a bi-Lipschitz map, $\varphi:\mathbb{R}^m \to Z$, and
a Lebesgue function $h\in L^1(A,\mathbb{Z})$ where $A\subset \mathbb{R}^m$,
then $\varphi_\# \Lbrack h \Rbrack \in \curr_m(Z)$ is an example of an
$m$ dimensional current in $Z$. Note that
\begin{equation}
\varphi_\# \Lbrack h \Rbrack (f,\pi_1,..., \pi_m)=
\int_{A \subset \mathbb{R}^m} (h\circ \varphi )(f\circ\varphi) \,
d(\pi_1\circ \varphi) \wedge \dots \wedge d(\pi_m\circ\varphi)
\end{equation}
where $d(\pi_i\circ\varphi)$ is well defined almost everywhere
by Rademacher's Theorem. All currents of importance in this
paper are built from currents of this form.
\end{example}
The following are Definition 2.3
and Definition 2.5 in \cite{AK}:
\begin{defn}[Ambrosio-Kirchheim]\label{rmrk-def-boundary}
The {\bf boundary} of $T\in \curr_{m+1}(Z)$ is defined
\begin{equation} \label{def-boundary}
\partial T(f, \pi_1, ... , \pi_m):= T(1, f, \pi_1,..., \pi_m)
\end{equation}
since in the smooth setting
\begin{equation}
\partial T(f d\pi_1 \wedge \cdots \wedge d\pi_m)
=T(1df\wedge d\pi_1\wedge \cdots\wedge d\pi_m).
\end{equation}
Note that $\varphi_\#(\partial T)=\partial(\varphi_\#T)$ and $\partial \partial T=0$.
\end{defn}
\begin{defn}[Ambrosio-Kirchheim]
The {\bf restriction} $T\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \omega\in \curr_m(Z)$
of a current $T\in M_{m+k}(Z)$ by a $k+1$ tuple
$\omega=(g,\tau_1,..., \tau_k)\in \mathcal{D}^k(Z)$:
\begin{equation}
(T\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\:\omega)(f,\pi_1,..., \pi_m):=T(f\cdot g, \tau_1,..., \tau_k, \pi_1,..., \pi_m).
\end{equation}
\end{defn}
The following definition of the mass of a current is technical \cite{AK}[Defn 2.6].
A simpler
formula for mass will be given in Lemma~\ref{lemma-weight} when we restrict
ourselves to integer rectifiable currents.
\begin{defn}[Ambrosio-Kirchheim] \label{defn-mass}
The {\bf mass measure} $\|T\| $
of a current $T\in \curr_m(Z)$, is the smallest Borel measure $\mu$ such that
(\ref{def-AK-current-iii}) holds for all $m+1$ tuples, $(f,\pi)$.
The {\bf mass} of $T$ is defined
\begin{equation} \label{def-mass-from-current}
M\left(T\right) = || T || \left(Z\right) = \int_Z \, d\| T\|.
\end{equation}
\end{defn}
In particular
\begin{equation} \label{eqn-mass}
\Big| T(f,\pi_1,..., \pi_m) \Big| \le {\mathbf M}(T) |f|_\infty \operatorname{Lip}(\pi_1) \cdots \operatorname{Lip}(\pi_m).
\end{equation}
Note that the currents in $\curr_m(Z)$
defined by Ambrosio-Kirchheim have finite mass by definition. Urs Lang
develops a variant of Ambrosio-Kirchheim theory that does not rely
on the finite mass condition in \cite{Lang-local-currents}.
Note the integral current, $\Lbrack h\Rbrack \in \curr_m(\mathbb{R}^m)$,
in Example~\ref{basic-current} has mass measure,
\begin{equation}
||\Lbrack h \Rbrack ||= |h| d\mathcal{L}^m
\end{equation}
and mass
\begin{equation}
{\mathbf M}\left(\Lbrack h \Rbrack \right) =\int_A |h| d\mathcal{L}^m.
\end{equation}
\begin{rmrk}
In (2.4) \cite{AK}, Ambrosio-Kirchheim show that
\begin{equation} \label{mass-push}
||\varphi_\#T|| \le [\operatorname{Lip}(\varphi)]^m \varphi_\# ||T||,
\end{equation}
so that when $\varphi$ is an isometry $||\varphi_\#T||=\varphi_\#||T||$
and ${\mathbf M}(T)={\mathbf M}\left(\varphi_\#T\right)$.
\end{rmrk}
Computing the mass of the push
forward current in Example~\ref{basic-current-pushed}
is a little more complicated and will be done in the next section.
\subsection{Parametrized Integer Rectifiable Currents}\label{subsect-param}\label{Subsect-Def-Current-Space-3}
Ambrosio and Kirchheim define integer rectifiable currents, $\intrectcurr_m\left(Z\right)$, on an arbitrary complete
metric space $Z$ \cite{AK}[Defn 4.2]. Rather than giving their definition, we will
use their characterization of integer rectifiable currents given in \cite{AK}[Thm 4.5]:
{\em A current $T\in \curr_m(Z)$ is an integer rectifiable current iff it has a parametrization
of the following form:}
\begin{defn}[Ambrosio-Kirchheim] \label{def-param-rep}
A {\bf parametrization}
$\left(\{\varphi_i\}, \{\theta_i\}\right)$ of an integer rectifiable current $T\in \intrectcurr_m\left(Z\right)$ with $m\ge 1$ is a countable collection of
bi-Lipschitz maps $\varphi_i:A_i \to Z$ with $A_i\subset\mathbb{R}^m$ precompact
Borel measurable and with pairwise disjoint images and
weight functions $\theta_i\in L^1\left(A_i,\mathbb{N}\right)$ such that
\begin{equation}\label{param-representation}
T = \sum_{i=1}^\infty \varphi_{i\#} \Lbrack \theta_i \Rbrack \quad\text{and}\quad {\mathbf M}\left(T\right) = \sum_{i=1}^\infty {\mathbf M}\left(\varphi_{i\#}\Lbrack \theta_i \Rbrack\right).
\end{equation}
The mass measure is
\begin{equation}
||T|| = \sum_{i=1}^\infty ||\varphi_{i\#}\Lbrack \theta_i \Rbrack ||.
\end{equation}
\end{defn}
Note that the current in Example~\ref{basic-current-pushed} is an integer
rectifiable current.
\begin{ex} \label{basic-mani}
If one has an oriented Riemannian manifold, $M^m$, of finite volume
and a bi-Lipschitz map $\varphi:M^m\to Z$, then $T=\varphi_\#\Lbrack{\bf \rm{1}}_M\Rbrack$
is an integer rectifiable current of dimension $m$ in $Z$. If $\varphi$
is an isometry, and $Z=M$
then ${\mathbf M}(T)=\operatorname{Vol}(M^m)$. Note further that
$||T||$ is concentrated on $\varphi(M)$ which is a set of Hausdorff dimension $m$.
\end{ex}
In \cite{AK}[Theorem 4.6] Ambrosio-Kirchheim define a canonical set associated with any
integer rectifiable current:
\begin{defn}[Ambrosio-Kirchheim] \label{defn-set}
The {\bf canonical set} of a current, $T$,
is the collection of points in $Z$ with positive lower density:
\begin{equation} \label{def-set-current}
\rm{set}\left(T\right)= \{p \in Z: \Theta_{*m}\left( \|T\|, p\right) >0\},
\end{equation}
where the definition of lower density is given in (\ref{eqn-lower-density}).
\end{defn}
\begin{rmrk}\label{good-set}
In \cite{AK}[Thm 4.6], Ambrosio-Kirchheim prove
given a current $T \in \intrectcurr_m\left(Z\right)$ on a
complete metric space $Z$ with a parametrization
$\left(\{\varphi_i\}, \theta_i\right)$ of $T$, we have
\begin{equation} \label{eqn-lem-weight-1}
\mathcal{H}^m\left(\rm{set}\left(T\right) \Lambda \bigcup_{i=1}^\infty \varphi_i\left(A_i\right)\right)=0,
\end{equation}
where $\Lambda$ is the symmetric difference,
\begin{equation}
A \Lambda B= \left(A \setminus B\right) \cup \left(B \setminus A\right).
\end{equation}
In particular
the canonical set, $\rm{set}\left(T\right)$,
endowed with the restricted metric, $d_Z$,
is a countably $\mathcal{H}^m$ rectifiable metric
space, $\left(\rm{set}\left(T\right), d_Z\right)$.
\end{rmrk}
\begin{example} \label{basic-mani-sing}
Note that the current in Example~\ref{basic-mani}, has
\begin{equation} \label{basic-mani-sing-1}
\rm{set}\left(\varphi_\#\Lbrack {\bf \rm{1}}_M \Rbrack\right)=\varphi(M).
\end{equation}
when $M$ is a smooth oriented Riemannian manifold.
If $M$ has a conical singularity, then (\ref{basic-mani-sing-1})
holds as well. However if $M$ has a cusp singularity at a point $p$
then
\begin{equation} \label{basic-mani-sing-1}
\rm{set}\left(\varphi_\#\Lbrack {\bf \rm{1}}_M \Rbrack\right)=\varphi(M\setminus\{p\}).
\end{equation}
\end{example}
Recall that the support of a current (c.f. \cite{AK} Definition 2.8) is
\begin{equation} \label{eqn-def-support}
\operatorname{spt} (T) := \operatorname{spt} ||T|| = \{p \in Z: \|T\|(B_p(r)) >0\,\, \forall r>0\}.
\end{equation}
Ambrosio-Kirchheim show the closure of $\rm{set}(T)$ is $\operatorname{spt}(T)$.
\begin{rmrk} \label{trouble-with-support}
Note that there are integer rectifiable currents $T^m$ on $\mathbb{R}^n$
such that the support is all of $\mathbb{R}^n$. For example, take a countable dense collection of points $p_j \in \mathbb{R}^3$,
then $X=\bigcup_{j\in \mathbb{N}} \partial B_{p_j}\left(1/2^j\right)$ is the set of the current $T\in \intcurr_m\left(\mathbb{R}^3\right)$ defined
by integration over $X$ and yet the support is $\mathbb{R}^3$.
\end{rmrk}
\begin{rmrk} \label{def-param-rep-2}
Given a parametrization of an integer rectifiable current $T$ one may refine this parametrization by choosing Borel measurable subsets
$A_i'$ of the $A_i$ such that
$\varphi_i: A'_i \to set\left(T\right)$. The new collection of maps $\{\varphi_i: A_i'\to Z\}$ is also a parametrization of $T$
and we will call it a settled parametrization. Unless stated otherwise, all our parametrizations will be settled. We may also choose precompact
$A_i'\subset A_i$ such that $\varphi_i(A_i')\cap\varphi_j(A_j')=\emptyset$.
We will call such a parametrization a preferred settled parametrization.
\end{rmrk}
Recall the definition of orientation in Definition~\ref{def-orientation} and the definition of
multiplicity in Definition~\ref{def-weight}.
The next lemma allows one to define the orientation and multiplicity of an integer rectifiable current [Definition~\ref{def-orient-mult-T}].
\begin{lem} \label{lemma-param-equiv}
Given two currents $T,T' \in \intrectcurr_m\left(Z\right)$ on a complete metric space $Z$ and respective parametrizations
$\left(\{\varphi_i\}, \theta_i\right)$, $\left(\{\varphi'_i\}, \theta'_i\right)$ we have $T=T'$ iff the following hold:
i) The symmetric difference satisfies,
\begin{equation}
\mathcal{H}^m
\left(\bigcup_{i=1}^\infty \varphi_i\left(A_i\right) \Lambda \bigcup_{i=1}^\infty \varphi_i'\left(A'_i\right) \right)=0.
\end{equation}
ii) The union of the atlases $\{\varphi_i\}$ and $\{\varphi_i'\}$ is an oriented atlas of
\begin{equation}
X=\bigcup_{i=1}^\infty \varphi_i\left(A_i\right) \cup \bigcup_{i=1}^\infty \varphi_i'\left(A'_i\right).
\end{equation}
iii) The sums:
\begin{equation} \label{eqn-theta-prop1}
\sum_{i=1}^\infty \theta_i\circ\varphi_i^{-1}{\bf \rm{1}}_{\varphi_i\left(A_i\right)}
=
\sum_{i=1}^\infty \theta'_i\circ{\varphi'_i}^{-1}{\bf \rm{1}}_{\varphi'_i\left(A'_i\right)}
\qquad
\mathcal{H}^m a.e. \textrm{ on } Z.
\end{equation}
\end{lem}
\begin{defn} \label{def-orient-mult-T}
Given $T$, the sum in (\ref{eqn-theta-prop1}) will be called the {\bf multiplicity} function, $\theta_T$. This function is an $\mathcal{H}^m$ measurable
function from $Z$ to $\mathbb{N} \cup \{0\}$. The uniquely defined equivalence class of oriented
atlases of $\rm{set}\left(T\right)$
will be called the orientation of $T$.
\end{defn}
A similar result is in \cite{AK}[Thm 9.1]
with a less Riemannian approach to the notion of orientation. The $\theta$ in
their theorem is our $\theta_T$.
\begin{proof}
We begin by relating some equations and then prove the theorem.
Note that by restricting to $A_{i,j}:= \varphi_i\left(A_i\right) \cap {\varphi'}_j\left({A'}_j\right)$,
we can focus on one term in the parametrization at a time:
\begin{equation}
T \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j} = \sum_{k=1}^\infty \varphi_{k\#} \Lbrack \theta_k \Rbrack \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j} =
\varphi_{i\#} \Lbrack \theta_i \Rbrack \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j} =
\varphi_{i\#} \Lbrack \theta_i {\bf \rm{1}}_{\varphi_i^{-1}\left(A_{i,j}\right)} \Rbrack .
\end{equation}
Thus $T \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j}=T'\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A'_{i,j}$ iff
\begin{equation} \label{eqn41}
\varphi_{i\#} \Lbrack \theta_i {\bf \rm{1}}_{\varphi_i^{-1}\left(A_{i,j}\right)} \Rbrack
= {\varphi'_{j\#}} \Lbrack \theta'_j {\bf \rm{1}}_{{\varphi'}_j^{-1}\left(A_{i,j}\right)} \Rbrack
\,\,\,\textrm{ iff } \,\,\,
\Lbrack \theta'_j {\bf \rm{1}}_{{\varphi'}_j^{-1}\left(A_{i,j}\right)} \Rbrack={ \varphi'_{j\#}}^{-1}\varphi_{i\#} \Lbrack \theta_i {\bf \rm{1}}_{\varphi_i^{-1}\left(A_{i,j}\right)} \Rbrack.
\end{equation}
This is true iff for any Lipschitz function $f$ defined on $A'_j$ we have
\begin{equation}
\int_{\varphi_j^{'-1}\left(A_{i,j}\right)} \theta'_j \cdot f \, d\mathcal{L}^m = \int_{\varphi_i^{-1}\left(A_{i,j}\right)} \theta_i \cdot (f
\circ {\varphi'_j}^{-1} \circ \varphi_i )\, \det\left(\nabla\left({\varphi'_j}^{-1}\circ \varphi_i\right) \right) d\mathcal{L}^m.
\end{equation}
By the change of variables formula,
this is true iff
\begin{equation}
\int_{\varphi_j^{'-1}\left(A_{i,j}\right)} \theta'_j \cdot f \, d\mathcal{L}^m =
\int_{\varphi_j^{'-1}\left(A_{i,j}\right)} (\theta_i \circ \varphi_i^{-1}\circ \varphi'_j )
\cdot f \operatorname{sgn}\det\left(\nabla(\varphi_i^{-1}\circ\varphi'_j) \right)d\mathcal{L}^m
\end{equation}
because the change of variables formula involves the absolute value of the determinant.
This is true iff the following two equations hold:
\begin{equation} \label{eqn-new-iii}
\theta'_j=\theta_i\circ \varphi_i^{-1}\circ \varphi'_j \qquad \mathcal{L}^m \textrm{ a.e. on }{\varphi'_j}^{-1}\left(A_{i,j}\right)
\end{equation}
and
\begin{equation} \label{eqn-new-ii}
\operatorname{sgn}\det(\nabla(\varphi_i^{-1}\circ\varphi'_j))
=1 \qquad \mathcal{L}^m \textrm{ a.e. on }{\varphi_j'}^{-1}\left(A_{i,j}\right).
\end{equation}
Setting
\begin{equation}
Y:= \bigcup_{i=1}^\infty \varphi_i(A_i) \textrm{ and }
Y':= \bigcup_{j=1}^\infty \varphi'_j(A'_j),
\end{equation}
we have $X= Y \cup Y'$ and $\bigcup_{i,j=1}^\infty A_{i,j} = Y \cap Y'$.
Furthermore by Remark~\ref{good-set}, we have
\begin{equation} \label{eqn-set-Lambda-i}
(i) \qquad \textrm{ iff }\qquad
\mathcal{H}^m\left(Y\Lambda Y'\right)
\qquad\textrm{ iff }\qquad
\mathcal{H}^m\left(set\left(T\right) \Lambda set\left(T'\right)\right)=0.
\end{equation}
We may now prove the theorem.
If $T=T'$, then $\rm{set}\left(T\right)=set\left(T'\right)$ and we have (i). Furthermore
$T\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j}=T'\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j}$ for all $i,j$ which implies (\ref{eqn-new-ii}) which implies (ii).
We also have (\ref{eqn-new-iii}), which implies
\begin{equation} \label{eqn-theta-on-X}
\sum_{i=1}^\infty \theta_i\circ\varphi_i^{-1}{\bf \rm{1}}_{\varphi_i\left(A_i\right)}
=
\sum_{i=1}^\infty \theta'_i\circ\varphi_i^{-1}{\bf \rm{1}}_{\varphi'_i\left(A'_i\right)}
\end{equation}
holds $\mathcal{H}^m$ almost everywhere on $\bigcup_{i,j=1}^\infty A_{i,j} = Y\cap Y'$.
Since we already have (i) then (\ref{eqn-set-Lambda-i}) implies
(\ref{eqn-theta-on-X}) holds
$\mathcal{H}^m$ almost everywhere on $ Y\cup Y'=X$ and we get (iii).
Conversely if (i), (ii), (iii) hold for a pair of parametrizations, then (ii) implies (\ref{eqn-new-ii}) and (iii) implies
(\ref{eqn-new-iii}). Thus, by (\ref{eqn41}) we have $T\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j}= T'\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: A_{i,j}$ for all $i,j$.
Summing over $i$ and $j$ we have $T\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: X = T'\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: X$.
By (i) and (\ref{eqn-set-Lambda-i}),
we have
\begin{equation}
T= T \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \bigcup_{i=1}^\infty \varphi_i\left(A_i\right) =T \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: Y = T'\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: Y' = T'\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \bigcup_{j=1}^\infty \varphi_j\left(A'_j\right)=T'.
\end{equation}
\end{proof}
In Proposition~\ref{prop-current-spaces} we will prove that
{\em if $T\in\intrectcurr_m(Z)$ is an integer rectifiable current, then
$(\rm{set}(T), d_Z, [\{\varphi_i\}], \theta_T)$ as defined in Definition~\ref{def-orient-mult-T}
is a completely settled weighted oriented
countably $\mathcal{H}^m$ rectifiable metric space} as in
Definitions~\ref{def-weight} and~\ref{def-settled}.
To prove this we must show $\rm{set}(T)$ is completely settled.
Thus we must better understand the relationship between the
mass measure of $T$, $||T||$, which is used to define the canonical
set and the weight $\theta_T \mathcal{H}^m$ which is used to defined
settled. Both measures must have positive density at the same locations.
\begin{rmrk}
In the proof of \cite{AK}[Theorem 4.6], Ambrosio-Kirchheim note that
\begin{equation} \label{AK4.6-measure}
||T||=\Theta_{*m}(||T||, \cdot) \mathcal{H}^m\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \rm{set}(T).
\end{equation}
\end{rmrk}
\begin{example}\label{weight-match}
Suppose $T\in \intcurr_m(M^m)$ in a smooth oriented Riemannian manifold
of finite volume
is defined $T=\Lbrack {\bf \rm{1}}_M\Rbrack$. Then $\theta_T=1$
while $||T||$ is the Lebesgue measure on $M$.
Since the Hausdorff
and Lebesgue measures agree on a smooth Riemannian manifold,
we have $\Theta_{*m}(||T||,p)=1$ as well.
The Hausdorff and Lebesgue measures also agree on
manifolds that have point singularities
as in Example~\ref{basic-mani-sing}, so that $\rm{set}(T)$ is completely
settled with respect to $\theta_T d\mathcal{H}^m$ in both cases given in that example
as well. In that case we again have $\theta_T=1$ everywhere, but
$\Theta_{*m}(||T||,p)=\Theta_{*m}(\theta_T\mathcal{H}^m,p)<1$ at
conical singularities and $0$ at cusp points.
\end{example}
In general, however, the lower density of $T$ need not agree with the
weight, $\theta_T$.
To find a formula
relating the multiplicity $\theta_T$
to the lower density of $||T||$ we need a notion called the area
factor of a normed space $V$ (c.f. \cite{AK}(9.11)):
\begin{equation}\label{def-area-factor}
\lambda_V:=\frac{2^m}{\omega_m} \sup \,\,\left\{\frac{\mathcal{H}^m(B_0(1))}
{\mathcal{H}^m(R)} \right\},
\end{equation}
where the supremum is taken over all parallelepipeds $R\subset V$ which
contain the unit ball $B_0(1)$.
\begin{rmrk}\label{rmrk-lambda}
In \cite{AK}[Lemma 9.2], Ambrosio-Kirchheim prove that
\begin{equation} \label{eqn-lem-weight-lambda-pre}
\lambda_V\in [m^{-m/2}, 2^m/\omega_m]
\end{equation}
and
observe that
$\lambda_V=1$ whenever $B_0(1)$ is a solid ellipsoid. This
will occur when $V$ is the tangent space on a Riemannian
manifold because the norm is an inner product. It is also possible that $\lambda_V=1$
when $V$ does not have an inner product norm (c.f. \cite{AK} Remark 9.3).
\end{rmrk}
The following lemma consolidates a few results in \cite{AK} and \cite{Kirchheim}:
\begin{lem}\label{lemma-weight}
Given an integer rectifiable current $T \in \intrectcurr_m(Z)$,
in a complete metric space $Z$ there is a function
\begin{equation} \label{eqn-lem-weight-lambda}
\lambda:\rm{set}(T) \to [m^{-m/2}, 2^m/\omega_m]
\end{equation}
satisfying
\begin{equation} \label{eqn-lem-weight-new-key}
\Theta_{*m}(||T||,x)=\theta_T(x)\lambda(x),
\end{equation}
for $\mathcal{H}^m$ almost every $x\in \rm{set} (T)$
such that
\begin{equation} \label{eqn-lem-weight-2}
||T||=\theta_T \lambda \mathcal{H}^m \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \rm{set}(T).
\end{equation}
In particular $\rm{set}(T)$ with the restricted metric from $Z$
is a completely settled
weighted oriented countably $\mathcal{H}^m$ rectifiable metric space with respect to
the weight function $\theta_T$ defined in Definition~\ref{def-orient-mult-T}.
When $T=\varphi_\#\Lbrack {\bf \rm{1}}_A\Rbrack$, with a bi-Lipschitz function, $\varphi$,
then for $x\in \varphi(A)$ we have
$\lambda(x)=\lambda_{V_x}$ where $V_x$ is $\mathbb{R}^m$
with the norm defined by the metric differential
$md\varphi_{\varphi^{-1}(x)}$.
\end{lem}
\begin{proof}
On the top of page 58 in \cite{AK}, Ambrosio-Kirchheim observe that
for $\mathcal{H}^m$ almost every $x\in S=\rm{set}(T)$, one can define
a approximate tangent space $\operatorname{Tan}^m(S,x)$ which
is $\mathbb{R}^m$ with a norm.
Taking
$\lambda(x)=\lambda_{\operatorname{Tan}^m(S,x)}$ and applying \cite{AK}(9.10), one
sees they have proven
(\ref{eqn-lem-weight-2}).
We then deduce (\ref{eqn-lem-weight-new-key}) using the fact
that $\Theta_{*m}(\mathcal{H}^m \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \rm{set} (T),x)=1$ almost everywhere
\cite{Kirchheim}[Theorem 9].
The bounds on $\lambda$ in (\ref{eqn-lem-weight-lambda})
come from (\ref{eqn-lem-weight-lambda-pre}) and they allow us
to conclude that the lower density of $\theta_T \mathcal{H}^m$
and the lower density of $||T||$ are positive at the same collection
of points.
Examining the proof of \cite{AK}, Theorem 9.1, one sees that
$V_x=\operatorname{Tan}^m(S,x)$ in this setting.
\end{proof}
In this section we introduce the notion of an integer rectifiable current structure on a metric space and define
integer rectifiable current spaces. We then prove Proposition~\ref{prop-current-spaces} that integer rectifiable current
spaces are completely settled weighted oriented $\mathcal{H}^m$ rectifiable metric spaces using the lemmas
from Subsection~\ref{Subsect-Def-Current-Space-2}.
\begin{defn} \label{def-current-structure} \label{defn-current-space}
An $m$-dimensional {\bf integer rectifiable
current structure} on a metric space $\left(X,d\right)$ is an integer rectifiable current $T\in\intrectcurr_m\left(\bar{X}\right)$ on the completion, $\bar{X}$,
of $X$ such that $\rm{set}\left(T\right)=X$. We call such a space an {\bf integer rectifiable current space} and denote it $\left(X,d,T\right)$.
Given an integer rectifiable current space $M=\left(X,d,T\right)$ , we let
$\rm{set}\left(M\right)$ and $X_M$ denote $X$, $d_M=d$ and $\Lbrack M \Rbrack =T $.
\end{defn}
\begin{rmrk} \label{rmrk-separable}
By \cite{AK} Defn 4.2, any metric space with an $m$-dimensional current structure must be countably $\mathcal{H}^m$-rectifiable
because the set of an $m$ dimensional integer rectifiable current is countably $\mathcal{H}^m$ rectifiable.
By \cite{AK} Thm 4.5, there is a countably collection of
bi-Lipscitz charts with compact domains
which map onto a dense subset
of the metric space (because we only include points of
positive density). In particular, the space is separable.
\end{rmrk}
\begin{rmrk}
We do not use the support, $spt(T)$, in this definition as it
is not necessarily countably $\mathcal{H}^m$ rectifiable
and may have a higher dimension as described in Remark~\ref{trouble-with-support}. See Example~\ref{example-dense-support}.
\end{rmrk}
\begin{rmrk} \label{Lip-mani-structure}
Recall that in Remark~\ref{Lip-mani-charts} we said that any $m$ dimensional oriented
connected Lipschitz or Riemannian manifold, $M$, is endowed with a standard atlas of charts
with a fixed orientation. We will also view these spaces as having
multiplicity or weight $1$. If $M$ has finite volume and we've chosen an
orientation, then we can
define an integer rectifiable current structure, $T=\Lbrack M \Rbrack \in \intrectcurr_m\left(M\right)$,
parametrized by a finite disjoint selection of charts with weight $1$. It is easy to verify that
$\rm{set} \left(T\right)=M$.
\end{rmrk}
\begin{lem} \label{lemma-isom-to-a-set}\label{lemma-isom-to-set}
Suppose $\left(X,d,T\right)$ is an integer rectifiable current space and $Z$ is a complete metric space.
If ${\phi}: X \to Z$ is an isometric embedding then the induced map on the completion,
$\bar{\phi}:\bar{X} \to Z$, is also an isometric embedding. Furthermore
the pushforward
$
\bar{\phi}_\#T
$
is an integer rectifiable current on $Z$ and
\begin{equation}
\phi: X \to \rm{set}\left(\bar{\phi}_\# T\right)
\end{equation}
is an isometry.
\end{lem}
\begin{proof}
Follows from the fact that $\rm{set}\left(\bar{\phi}_\# T\right)= \bar{\phi}\left( \rm{set}\left(T\right) \right)$ \cite{AK}.
\end{proof}
Conversely, if $T$ is an integer rectifiable current in $Z$,
then $\left(\rm{set}\left(T\right), d_Z, T\right)$ is an an $m$ dimensional integer rectifiable current space.
\begin{prop} \label{prop-current-spaces}
There is a one-to-one correspondence between completely settled weighted oriented countably $\mathcal{H}^m$ rectifiable metric
spaces, $\left(X, d, [\{\phi\}], \theta\right)$, and integer rectifiable current spaces $\left(X,d,T\right)$ as follows:
Given $\left(X,d,T\right)$, we define a weight $\theta=\theta_T$
and orientation, $[\{\varphi_i\}]$ as in Definition~\ref{def-orient-mult-T},
so that
\begin{equation} \label{eqn-theta-prop2}
\theta:= \theta_T=\sum_{i=1}^\infty \theta_i\circ\varphi_i^{-1}{\bf \rm{1}}_{\varphi_i\left(A_i\right)},
\end{equation}
and the corresponding space is $(X,d,[\{\varphi_i\}],\theta )$.
Given $\left(X, d, [\{\varphi\}], \theta\right)$, we define a
unique induced current structure $T\in\intrectcurr_m\left(\bar{X}\right)$ given by
\begin{equation} \label{eqn-T-prop1}
T\left(f,\pi\right)= \sum \varphi_{i\#}\Lbrack \theta\circ\varphi_i\Rbrack \left(f,\pi\right) = \sum \int_{A_i} \theta\circ\varphi_i f \circ \varphi_i \det\left(\nabla\left(\pi\circ\varphi_i\right)\right)\, d\mathcal{L}^m,
\end{equation}
and the corresponding space is then $(X,d,T)$ because $\rm{set}(T)=X$.
\end{prop}
\begin{proof}
Given $\left(X, d, [\{\varphi_i\}], \theta\right)$ we first define a current on the completion $\bar{X}$ using a preferred oriented atlas
as in (\ref{eqn-T-prop1}). This is well defined because
\begin{equation}
\sum_{i=1}^\infty M \left(\varphi_{i\#} \Lbrack \theta \circ \varphi_i \Rbrack\right) \le
C_m \sum_{i=1}^\infty \int_{\varphi_i\left(A_i\right)}\theta \, d\mathcal{H}^m<\infty
\end{equation}
where $C_m$ is a constant that may be computed using Lemma~\ref{lemma-weight}.
The sum is then finite by Definition~\ref{def-weight}.
So we have a current with a parametrization $\left(\{\varphi_i\}, \{\theta_i\}\right)$ where $\theta_i := \theta \circ \varphi_i$. The weight
function $\theta_T$ of the current $T$ defined below Lemma~\ref{lemma-param-equiv}
agrees with the weight function $\theta$ on $X$ because for almost every $x\in X$
there is a chart such that $x \in \varphi_i\left(A_i\right)$, and
\begin{equation} \label{eqn-theta-T-theta}
\theta_T\left(x\right)=\theta_i\circ \varphi_i^{-1}\left(x\right)=\theta\left(x\right).
\end{equation}
Furthermore $\rm{set}\left(T\right)= \{p \in \bar{X}: \Theta_{*m}\left( \|T\|, p\right) >0\}$,
so by Lemma~\ref{lemma-weight} we have
\begin{equation}
\rm{set}\left(T\right)= \left\{p \in \bar{X}: \Theta_{*m}\left(\theta \, d\mathcal{H}^m \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \bigcup_{i=1}^\infty \varphi_i\left(A_i\right), p\right)>0\right\}
\end{equation}
which is $X$ because $X$ is completely settled. Since $X$ is a countably $\mathcal{H}^m$ rectifiable space, we know
$T \in \intrectcurr_m\left(\bar{X}\right)$. Thus we have an integer rectifiable current space $\left(X,d,T\right)$.
Conversely we start with $\left(X,d,T\right)$. Applying Lemma~\ref{lemma-param-equiv}, we have a unique well defined orientation and
weight function $\theta_T$. Thus $\left(\rm{set}\left(T\right), d, [\{\varphi_i\}], \theta_T\right)$ is an oriented weighted countably $\mathcal{H}^m$ rectifiable
metric space. Since $\rm{set}\left(T\right)=X$ in the definition of a current space, we have shown $\left(X, d, [\{\varphi_i\}], \theta_T\right)$
is an oriented weighted countably
$\mathcal{H}^m$ rectifiable metric space. As in the above paragraph,
we see that $\rm{set}\left(T\right)$ is a completely
settled subset of $\bar{X}$. So $X$ is completely settled.
Note that since the $\{\varphi_i\}$ from the preferred atlas are the $\{\varphi_i\}$ of the parametrization and the
weights agree in (\ref{eqn-theta-T-theta}), this pair of maps is a correspondence.
\end{proof}
We may now define the mass and relate it to the weighted volume:
\begin{defn} \label{defn-space-mass}
The {\bf mass} of an integer rectifiable current space $\left(X,d,T\right)$ is defined to be the
mass, ${\mathbf M}\left(T\right)$, of the current structure, $T$.
\end{defn}
Note that the mass is always finite by (iii) in the definition of a current.
\begin{lem} \label{lem-push-mass}
If $\varphi: X\to Y$ is a $1$-Lipschitz map, then ${\mathbf M}(\varphi_\#(T))
\le {\mathbf M}(T)$.
Thus if $\varphi: X\to Y$ is an isometric embedding, then ${\mathbf M}(T)={\mathbf M}(\varphi_\#(T))$.
\end{lem}
Recall Definition~\ref{def-weight} of the weighted volume, $Vol\left(X, \theta\right)$.
We have the following corollary of Lemma~\ref{lemma-weight} and Proposition~\ref{prop-current-spaces}:
\begin{lem}\label{lemma-mass-current-space}
The mass of an integer rectifiable current space $\left(X,d,T\right)$ with multiplicity or weight, $\theta_T$, satisfies
\begin{equation}
{\mathbf M}(T)=\int_X \theta_T(x)\lambda(x) d\mathcal{H}^m(x).
\end{equation}
In particular,
\begin{equation} \label{eqn-lem-weight-3}
M\left(T\right) \in \left[ m^{-m/2}Vol(X,\theta), \frac{2^m}{\omega_m} Vol(X,\theta)\right] ,
\end{equation}
where $Vol(X,\theta)$ is the weighted volume defined in Definition~\ref{def-weight}.
\end{lem}
Note that on a Riemannian manifold with multiplicity one, the mass and the weighted volume agree and are
both equal to the volume of the manifold. On reversible Finsler spaces,
$\lambda(x)$ depends on the norm of the tangent space at $x$.
\subsection{Integral Current Spaces} \label{Subsect-Def-Current-Space-4}
In this subsection, we define the boundaries of integer rectfiable current spaces and the notion of
an integral current space. We begin with Ambrosio-Kirchheim's
extension of Federer-Fleming's notion of an
integral current \cite{AK}[Defn 3.4 and 4.2]:
\begin{defn}[Ambrosio-Kirchheim]
An
{\bf integral current}
is an
integer rectifiable current, $ T\in\intrectcurr_m(Z)$, such that $\partial T$
defined as
\begin{equation}
\partial T \left(f, \pi_1,..., \pi_{m-1}\right) := T \left(1, f, \pi_1,..., \pi_{m-1}\right)
\end{equation}
satisfies the requirements to be
a current. One need only verify that $\partial T$ has finite mass as the
other conditions always hold.
We use the standard notation, $\intcurr_m\left(Z\right)$, to denote
the space of $m$ dimensional integral currents on $Z$.
\end{defn}
\begin{rmrk}
By the boundary rectifiability theorem of Ambrosio-Kirchheim
\cite{AK}[Theorem 8.6], $\partial T$ is then an integer rectifiable current
itself. And in fact it is an integral current whose boundary is $0$.
\end{rmrk}
Thus we can make the following new definition:
\begin{defn} \label{def-integral-space} \label{defn-integral-current-space}
An $m$ dimensional {\bf integral current space} is an
integer rectifiable current space,
$\left(X,d,T\right)$, whose current structure,
$T$, is an integral current (that is
$\partial T$ is an integer
rectifiable current in $\bar{X}$). The boundary of $\left(X,d,T\right)$ is then the
integral current space:
\begin{equation}
\partial \left(X,d_X,T\right) := \left(\rm{set}\left(\partial T\right), d_{\bar{X}}, \partial T\right).
\end{equation}
If $\partial T=0$ then we say $\left(X,d,T\right)$ is an integral current without boundary or with zero boundary.
\end{defn}
Note that $\rm{set}\left(\partial T\right)$ is not necessarily a subset of $\rm{set}\left(T\right)=X$ but it is always a subset
of $\bar{X}$. As in Definition~\ref{defn-current-space}, given an integer rectifiable current
space $M=\left(X,d,T\right)$ we will use
$\rm{set}\left(M\right)$ or $X_M$ to denote $X$, $d_M=d$ and $\Lbrack M \Rbrack =T $.
\begin{rmrk}
On an oriented Riemannian manifold with boundary,
$M$, the boundary $\partial M$ defined as a current space
agrees with the definition of $\partial M$ in Riemannian geometry.
In that setting an atlas of $M$ can be
restricted to provide an atlas for $\partial M$. It is not always possible to do this
on integer rectifiable current spaces.
In fact the boundaries of charts need not even have finite mass for an individual chart.
If a chart $\varphi: K \subset \mathbb{R}^m \to Z$ with $K$ compact, then
$\partial \varphi_\# \Lbrack {\bf \rm{1}}_K\Rbrack$ is an integral current iff $K$ has
finite perimeter.
\end{rmrk}
\begin{rmrk} \label{rmrk-biLip-matching}
Suppose $M$ and $N$ are connected $m$-dimensional oriented
Lipschitz manifolds
with the standard current structures, $\Lbrack M \Rbrack$ and $\Lbrack N \Rbrack$
as in Remark~\ref{Lip-mani-charts}
and $\psi: M\to N$ a bi-Lipschitz homeomorphism.
Then one can do a computation mapping charts on $M$ to
charts on $N$ and applying Lemma~\ref{lemma-param-equiv}, to see that
\begin{equation} \label{rmrk-biLip-matching-1}
\psi_\#\Lbrack M\Rbrack = \pm\Lbrack N\Rbrack.
\end{equation}
That is, the bi-Lipschitz
homeomorphism is either a current preserving or a current reversing map.
When $M$ and $N$ are isometric, then the isometry is also current preserving or
current reversing.
When $M$ and $N$ are integral current spaces, they may have multiplicity, so
that a bi-Lipschitz homeomorphism or isometry from $\rm{set}\left(M\right)$ to $\rm{set}\left(N\right)$ does not
in general push $\Lbrack M \Rbrack$ to $\Lbrack N \Rbrack$.
Even with multiplicity $1$, the fact that orientations are defined using disjoint charts
can lead to different signs on different charts so that (\ref{rmrk-biLip-matching-1}) fails.
\end{rmrk}
As in Federer, Ambrosio-Kirchheim define the total mass and we do as well:
\begin{defn} \label{def-total-mass}
The {\bf total mass} of an integral current with boundary, $T$, is
\begin{equation} \label{total-mass}
\mathbf N\left(T\right)={\mathbf M}\left(T\right) +{\mathbf M}\left(\partial T\right).
\end{equation}
Naturally we can extend this concept to current spaces: $\mathbf N\left(X,d,T\right)=\mathbf N\left(T\right)$.
\end{defn}
Recall that by Remark~\ref{rmrk-separable}, an integral current space is separable and
has a collection of disjoint biLipshitz charts whose image is dense and the boundary of
the integral current space has the same property. An integral current space need not be
precompact or bounded. An integral current space is not necessarily a geodesic space.
\section{{\bf The Intrinsic Flat Distance Between Current Spaces}}\label{sect-flat-distance}
Let $\mathcal{M}^m$ be the space of $m$ dimensional integral current spaces as
defined in Definition~\ref{defn-integral-current-space}.
Recall they have the form $M=\left(X_M,d_M,T_M\right)$ where
$T_M \in \intcurr_m\left(\bar{X}_M\right)$ and $\rm{set}(T_M)=X_M$. Note $\mathcal{M}^m$
also includes the zero current denoted ${\bf{0}}$.
Definition~\ref{def-flat1} in the introduction naturally applies to any $M,N \in \mathcal{M}^m$
so that:
\begin{equation}\label{eqn-local-defn}
d_{{\mathcal F}}\left(M,N\right):=\inf\{{\mathbf M}\left(U\right)+{\mathbf M}\left(V\right)\}
\end{equation}
where the infimum is taken over all complete metric spaces,
$\left(Z,d\right)$, and all integral currents,
$U\in\intcurr_m\left(Z\right), V\in\intcurr_{m+1}\left(Z\right)$,
such that there exists isometric embeddings
\begin{equation}
\varphi : \left(\bar{X}_M, d_{\bar{X}_M}\right)\to \left(Z,d\right) \textrm{ and }\psi: \left(\bar{X}_N,d_{\bar{X}_N}\right)\to \left(Z,d\right)
\end{equation}
with
\begin{equation} \label{eqn-Federer-Flat-2}
\varphi_\# T_M- \psi_\# T_N=U+\partial V.
\end{equation}
Here we consider the ${\bf{0}}$ space to isometrically embed into any $Z$
with $\varphi_\#0=0\in \intcurr_m\left(Z\right)$.
Note that, by the definition, $d_{{\mathcal F}}$ is clearly symmetric.
In Subsection~\ref{subsect-triangle} we prove
that $d_{\mathcal{F}}$ satisfies the triangle inequality on $\mathcal{M}^m$
[Theorem~\ref{triangle}].
As a consequence, the distance between integral
current spaces is always finite and is easy to estimate [Remark~\ref{finite}].
In Subsection~\ref{subsect-flat-dist-1}, we review the compactness theorems of
Gromov and of Ambrosio-Kirchheim,
and present a compactness theorem for
intrinsic flat convergence [Theorem~\ref{GH-to-flat}], which
follows immediately from theirs.
In Subsection~\ref{subsect-flat-dist-2}, we prove
Theorem~\ref{inf-dist-attained} that the infimum in the definition of
the intrinsic flat distance is attained
between precompact integral current spaces.
That is, there exists a common metric space, $Z$,
and integral currents, $U,V\in \intcurr_m(Z)$, achieving the infimum in (\ref{eqn-local-defn}).
In Subsection~\ref{subsect-flat-dist-3} we prove
that $d_{\mathcal{F}}$ is a distance on $\mathcal{M}_0^m$.
That is, we prove that when two precompact integral current
spaces are a distance zero apart, there is a current preserving isometry between them [Theorem~\ref{zero-mani}].
Thus $d_{{\mathcal F}}$ is a distance on $\mathcal{M}_0^m$ where
\begin{equation}
\mathcal{M}_0^m =\{ M \in \mathcal{M}^m: \,\, X_M \textrm{ is precompact}\}.
\end{equation}
\begin{rmrk} \label{rmrk-distinct-flat}
Note that the flat distance $d^Z_F$ given above Definition~\ref{def-flat1}
has an infimum that is taken over all
$U\in\intcurr_m\left(Z\right), V\in\intcurr_{m+1}\left(Z\right)$ where the supports of $U$ and $V$
may be noncompact or even unbounded as long as they have finite mass.
Thus we can have unbounded
limits [Example~\ref{ex-unbounded}] and bounded noncompact limits
[Example~\ref{ex-many-tips}].
\end{rmrk}
\subsection{The Triangle Inequality} \label{subsect-triangle}
In this section we prove the triangle inequality for the intrinsic flat distance
between integral current spaces:
\begin{thm} \label{triangle}
For all $M_1, M_2, N \in \mathcal{M}^m$, we have
\begin{equation}\label{eqn-triangle-here}
d_{{\mathcal F}}\left(M_1,M_2\right)\le d_{{\mathcal F}}\left(M_1,N\right)+d_{{\mathcal F}}\left(N, M_2\right).
\end{equation}
\end{thm}
In the proof of this theorem, we do not assume the infimum in (\ref{eqn-local-defn})
is finite. Naturally the theorem is immediately true if the right hand side
of (\ref{eqn-triangle-here}) is infinite. It is a consequence of the theorem that
when the right hand side is finite, the left hand side is finite as well.
Applying the theorem with $N_1={\bf 0}$, we may then conclude the
distance is finite and estimate it using the masses of $M_1$ and $M_2$:
\begin{rmrk}\label{finite}
Taking $U=M$ and $V=0$ in (\ref{eqn-local-defn}), we see that
$
d_{{\mathcal F}}\left(M,0\right) \le {\mathbf M}\left(M\right),
$
so the intrinsic flat distance between any pair of integral current spaces of finite mass
is finite
\begin{equation}
d_{{\mathcal F}}\left(M_1,M_2\right) \le d_{{\mathcal F}}\left(M_1,0\right)+d_{{\mathcal F}}\left(0,M_2\right) \le {\mathbf M}\left(M_1\right)+{\mathbf M}\left(M_2\right).
\end{equation}
In particular, when $M_i$ are Riemannian manifolds, then ${\mathbf M}\left(M_i\right)=\operatorname{Vol}\left(M_i\right)$ and
we have
\begin{equation}
d_{{\mathcal F}}\left(M_1,M_2\right) \le \operatorname{Vol}\left(M_1\right)+\operatorname{Vol}\left(M_2\right).
\end{equation}
\end{rmrk}
To prove Theorem~\ref{triangle} we apply the following well-known
gluing lemma (c.f. \cite{BBI}):
\begin{lem} \label{gluing-lemma}
Given three
metric spaces $\left(Z_1,d_1\right)$, $(Z_2,d_2)$ and $\left(X, d_X\right)$ and two isometric embeddings, $\varphi_i: X \to Z_i$,
we can glue $Z_1$ to $Z_2$ along the isometric images of $X$ to create a space
$Z=Z_1 \sqcup _{X} Z_2$ where $d_Z\left(x,x'\right)= d_i\left(x,x'\right)$ when $x,x'\in Z_i$ and
\begin{equation}
d_Z\left(z,z'\right)=\inf_{x\in X} \left(d_1\left(z,\varphi_1\left(x\right)\right) + d_2\left(\varphi_2\left(x\right), z'\right) \right)
\end{equation}
whenever $z\in Z_1, z'\in Z_2$. There exist natural isometric embeddings $f_i: Z_i \to Z$
such that $f_1\circ\varphi_1=f_2 \circ\varphi_2$ is an isometric embedding of $X$ into $Z$.
\end{lem}
We now prove Theorem~\ref{triangle}:
\begin{proof}
Let $M_i=(X_i, d_i, T_i)$ and $N=(X,d,T)$, and let $Z_1, Z_2$ be
metric spaces and let $\psi_i:\bar{X}_i \to Z_i$ and $\varphi_i:\bar{X} \to Z_i$
be isometric embeddings. Let $U_i \in \intcurr_m(Z_i)$ and $V_i\in \intcurr_{m+1}(Z_i)$
such that
\begin{equation} \label{triangle-1}
\varphi_{i\#} T- \psi_{i\#} T_i=U_i+\partial V_i.
\end{equation}
Applying Lemma~\ref{gluing-lemma}, we create a metric space $Z$ with
isometric embeddings $f_i: Z_i \to Z$ such that
$f_1 \circ \varphi_1=f_2 \circ \varphi_2$ is an isometric embedding of $X$ into $Z$.
Pushing forward the current structures to $Z$, we have
$f_{1\#} \varphi_{1\#} T =f_{2\#}\varphi_{2\#}T$, so
\begin{eqnarray}
f_{1\#}\psi_{1\#} T_1 - f_{2\#}\psi_{2\#}T_2 &= &
f_{1\#}\psi_{1\#} T_1 - f_{1\#} \varphi_{1\#} T + f_{2\#}\varphi_{2\#}T-f_{2\#}\psi_{2\#}T_2 \\
&=& f_{1\#}(\psi_{1\#} T_1 - \varphi_{1\#} T) + f_{2\#}(\varphi_{2\#}T-\psi_{2\#}T_2 )\\
&=& f_{1\#} (-U_1 - \partial V_1) + f_{2\#}( U_2 + \partial V_2)\\
&=& -f_{1\#} U_1 - \partial f_{1\#}V_1 + f_{2\#}U_2 + \partial f_{2\#}V_2\\
&=& f_{2\#} U_2 -f_{1\#}U_1 + \partial (f_{2\#} V_2 -f_{1\#} V_1).
\end{eqnarray}
So by (\ref{eqn-local-defn}) applied to the isometric embeddings
$f_i\circ\psi_i:\bar{X}_i\to Z$, we have
\begin{equation}
d_{{\mathcal F}}\left(M_1,M_2\right)
\le {\mathbf M}(f_{2\#} U_2 -f_{1\#}U_1) + {\mathbf M} (f_{2\#} V_2 -f_{1\#} V_1).
\end{equation}
Applying the fact that mass is a norm and Lemma~\ref{lem-push-mass}
we have,
\begin{eqnarray}
d_{{\mathcal F}}\left(M_1,M_2\right)
&\le& {\mathbf M}(f_{2\#} U_2) +{\mathbf M}(f_{1\#}U_1) + {\mathbf M} (f_{2\#} V_2) +{\mathbf M}(f_{1\#} V_1) \\
&=&{\mathbf M}(U_2) +{\mathbf M}(U_1) + {\mathbf M} ( V_2) + {\mathbf M} (V_1).
\end{eqnarray}
Taking an infimum over all $U_i$ and $V_i$ satisfying (\ref{triangle-1}), we see
that
\begin{equation}
d_{{\mathcal F}}(M_1,M_2) \le d^{Z_1}_F(\varphi_{1\#} T,\psi_{1\#} T_1) +d^{Z_2}_F(\varphi_{2\#} T,\psi_{2\#} T_2).
\end{equation}
Taking an infimum over all metric spaces $Z_1, Z_2$ and all isometric embeddings
$\psi_i:\bar{X}_i \to Z_i$ and $\varphi_i:\bar{X} \to Z_i$ we obtain the triangle inequality.
\end{proof}
\subsection{A Brief Review of Existing Compactness Theorems} \label{subsect-flat-dist-1}
Gromov defined the following distance between metric spaces in \cite{Gromov-metric}:
\begin{defn}[Gromov]\label{defn-GH}
Recall that the Gromov-Hausdorff distance between two metric spaces $\left(X, d_X\right)$ and $\left(Y, d_Y\right)$
is defined as
\begin{equation} \label{eqn-GH-def}
d_{GH}\left(X,Y\right) := \inf \, d^Z_H\left(\varphi\left(X\right), \psi\left(Y\right)\right)
\end{equation}
where $Z$ is a complete metric space, and $\varphi: X \to Z$ and $\psi:Y\to Z$ are
isometric embeddings and where the Hausdorff distance in $Z$ is defined as
\begin{equation}
d_{H}^Z\left(A,B\right) = \inf\{ \epsilon>0: A \subset T_\epsilon\left(B\right) \textrm{ and } B \subset T_\epsilon\left(A\right)\}.
\end{equation}
\end{defn}
Gromov proved that this is indeed a distance on compact metric spaces: $d_{GH}\left(X,Y\right)=0$
iff there is an isometry between $X$ and $Y$. There are many equivalent definitions
of this distance, we choose to state this version because it inspired our definition
of the intrinsic flat distance. Gromov also introduced the following notions:
\begin{defn}[Gromov]\label{defn-equibounded}
A collection of metric spaces is said to be equibounded or
uniformly bounded if there is a uniform upper bound
on the diameter of the spaces.
\end{defn}
\begin{rmrk}\label{rmrk-NXr}
We will write $N\left(X,r\right)$ to denote the maximal
number of disjoint balls of radius $r$ in a space $X$. Note that $X$ can always be covered
by $N\left(X,r\right)$ balls of radius $2r$.
\end{rmrk}
\begin{defn}[Gromov] \label{defn-equicompact}
A collection of spaces is said to
be equicompact or uniformly compact if they
have a common upper bound $N\left(r\right)$
such that $N\left(X,r\right) \le N\left(r\right)$ for all spaces $X$ in the collection.
\end{defn}
Note that Ilmanen's Example depicted in Figure~\ref{fig-hairy-sphere} is not
equicompact, as the number of balls centered on the tips approaches infinity
[Example~\ref{ex-hairy-sphere}].
Gromov's Compactness Theorem states that sequences of equibounded and equicompact
metric spaces have a Gromov-Hausdorff converging subsequence \cite{Gromov-French}.
In fact,
Gromov proves a stronger version of this statement in
a subsequent work, \cite{Gromov-poly}{p 65}, which
we state here so that we may apply it:
\begin{thm}[Gromov's Compactness Theorem] \label{Thm-Gromov}
If a sequence of compact metric spaces, $X_j$, is equibounded and equicompact, then
there is a pair of compact metric spaces, $Y \subset Z$, and a subsequence $X_{j_i}$ which
isometrically embed into $Z$: $\varphi_{j_i}: X_{j_i} \to Z$ such that
\begin{equation}
\lim_{i\to \infty} d_H^Z\left(\varphi_{j_i}\left(X_{j_i}\right), Y\right) =0.
\end{equation}
So $\left(Y,d_Z\right)$ is the Gromov-Hausdorff limit of the $X_{i_j}$.
\end{thm}
Gromov's proof of the stronger statement involves a construction of a metric
on the disjoint union of the sequence of spaces. This method of proving the Gromov compactness theorem relies on
the fact that infimum in (\ref{defn-GH}) can be estimated arbitrarily well by taking $Z$ to be a disjoint union of
the spaces and choosing a clever metric on $Z$.
The reason we have stated this stronger version of Gromov's Compactness Theorem
is because it can be applied in combination with Ambrosio-Kirchheim's compactness
theorem to prove our first compactness theorem for integral current spaces [Theorem~\ref{GH-to-flat}].
Recall the notion of total mass [Definition~\ref{def-total-mass}].
Ambrosio Kirchheim's Compactness Theorem, which extends Federer-Fleming's Flat Norm
Compactness Theorem, is stated in terms of weak convergence of
currents. See Definition 3.6 in \cite{AK} which extends Federer-Fleming's notion of weak convergence:
\begin{defn}[Weak Convergence] \label{def-weak}
A sequence of integral currents $T_j \in \intcurr_m\left(Z\right)$ is said to converge weakly to
a current $T$ iff the pointwise limits satisfy
\begin{equation}
\lim_{j\to \infty} T_j\left(f, \pi_1,..., \pi_m\right) = T\left(f, \pi_1,..., \pi_m\right)
\end{equation}
for all bounded Lipschitz $f$ and Lipschitz $\pi_i$.
\end{defn}
\begin{rmrk} \label{rmrk-unif-to-weak}
If we suppose one has a sequence of isometric embeddings, $\varphi_i: X \to Z$,
which converge uniformly to $\varphi: X\to Z$, and $T\in \intcurr_m(X)$,
then $\varphi_{i\#}T$ converges to $\varphi_\#T$. This can be seen by
applying properties (ii) and (iii) in the definition of a current
as follows:
\begin{eqnarray*}
\lim_{i\to\infty} \varphi_{i\#}T(f, \pi_1,..., \pi_m)
&=&\lim_{i\to\infty} T(f\circ \varphi_i, \pi_1\circ\varphi_i,..., \pi_m\circ\varphi_i)\\
&=& T(f\circ \varphi, \pi_1\circ\varphi,..., \pi_m\circ\varphi)=\varphi_\#T(f, \pi_1,..., \pi_m).
\end{eqnarray*}
\end{rmrk}
\begin{rmrk}
If $T_j\in \intcurr_m(Z)$ has
${\mathbf M}(T_j) \to 0$, then by (\ref{eqn-mass}),
\begin{equation}
\Big| T_j(f,\pi_1,..., \pi_m) \Big| \le {\mathbf M}(T_j) |f|_\infty \operatorname{Lip}(\pi_1) \cdots \operatorname{Lip}(\pi_m)\to 0,
\end{equation}
so $T_j$ converges weakly to $0$.
\end{rmrk}
\begin{rmrk}\label{rmrk-flat-implies-weak}
Note that flat convergence implies weak convergence because $T_j \stackrel {\mathcal{F}}{\longrightarrow} T$
implies there exists $U_j,V_j$ with ${\mathbf M}(U_j)+{\mathbf M}(V_j)\to 0$
such that $T_j-T=U_j +\partial V_j$. This implies that $U_j$ and
$V_j$ must converge weakly to $0$ and $\partial V_j$ must as well.
So $T_j -T\rightharpoonup 0$ and $T_j \rightharpoonup T$.
\end{rmrk}
\begin{rmrk} \label{rmrk-lower-mass}
Immediately below the definition of weak convergence \cite{AK} Defn 3.6,
Ambrosio-Kirchheim prove
the lower semicontinuity of mass. In particular,
if $T_j$ converges weakly to $T$, then $\liminf_{j\to\infty} {\mathbf M}(T_j) \ge {\mathbf M}(T)$.
\end{rmrk}
\begin{rmrk} It should be noted here that weak convergence as defined in Federer \cite{Federer}
is tested only with differential forms of compact support while weak convergence in
Ambrosio-Kirchheim does not require the test tuples to have compact support. Sequences
of unit spheres in Euclidean space whose centers diverge to infinity converge weakly to
$0$ in the sense of Federer but not in the sense of Ambrosio-Kirchheim.
\end{rmrk}
\begin{thm}[Ambrosio-Kirchheim Compactness]\label{AK-compact}
Given any complete metric space
$Z$, a compact set $K \subset Z$ and any sequence of integral currents $T_j \in \intcurr_m \left(Z\right)$
with a uniform upper bound on their total mass $\mathbf N\left(T_j\right)={\mathbf M}\left(T_j\right) +{\mathbf M}\left(\partial T_j\right) \le M_0$,
such that $\rm{set}\left(T_j\right) \subset K$, there exists a subsequence, $T_{j_i}$, and a limit current $T \in \intcurr_m\left(Z\right)$
such that $T_{j_i}$ converges weakly to $T$.
\end{thm}
The key point of this theorem is that the limit current is an integral current and has a rectifiable
set with finite mass and rectifiable boundary with bounded mass.
In order to apply Ambrosio-Kirchheim's result we need a result of the second author from
\cite{Wenger-flat}[Theorem 1.4] which generalizes a theorem of Federer-Fleming relating the
weak and flat norms. As in Federer-Fleming one needs a uniform bound on total mass
to have the relationship. To simplify the statement of \cite{Wenger-flat}[Theorem 1.4],
we restrict the setting to Banach spaces although his result is far more general:
\begin{thm}[Wenger Flat=Weak Convergence] \label{weak=flat}
Let $E$ be a Banach space and $m \ge 1$.
If we assume a sequence of integral currents,
$T_j \in \intcurr_m\left(E\right)$, has a uniform upper
bound on total mass ${\mathbf M}\left(T_j\right)+{\mathbf M}\left(\partial T_j\right)$, then $T_j$ converges weakly to
$T\in \intcurr_m\left(E\right)$ iff $T_j$ converges to $T$ in the flat sense.
\end{thm}
For our purposes, it suffices to have a Banach space, because we may apply Kuratowski's embedding theorem
to embed any complete metric space into a Banach space:
\begin{thm}[Kuratowski Embedding Theorem] \label{Kuratowski}
Let $Z$ be a complete metric space, and $\ell^\infty\left(Z\right)$ be the space of
bounded real valued functions on $Z$ endowed with the sup norm. Then the
map $\iota: Z \to \ell^\infty\left(Z\right)$ defined by fixing a basepoint $z_0\in Z$
and letting $\iota\left(z\right)= d_Z\left(z_0,\cdot\right)-d_Z\left(z, \cdot\right)$ is an isometric embedding.
\end{thm}
\begin{rmrk} \label{rmrk-Kuratowski}
By the Kuratowski embedding theorem,
the infimum in (\ref{eqn-local-defn})
can be taken over Banach spaces, $Z$.
\end{rmrk}
Combining Kuratowski's Embedding Theorem with
Gromov and Ambrosio-Kirchheim's Compactness Theorems we immediately obtain:
\begin{thm} \label{GH-to-flat}
Given a sequence of $m$ dimensional integral current spaces $M_j=\left(X_j, d_j, T_j\right)$ such that $X_j$ are equibounded and
equicompact and such that $\mathbf N\left(T_j\right)$ is uniformly bounded above, then a subsequence
converges in the
Gromov-Hausdorff sense $\left(X_{j_i}, d_{j_i}\right) \to \left(Y,d_Y\right)$ and in the
intrinsic flat sense
$\left(X_{j_i}, d_{j_i}, T_{j_i}\right) \to \left(X,d,T\right)$
where either $\left(X,d,T\right)$ is an $m$ dimensional integral current space
with $X \subset Y$
or it is the ${\bf 0}$ current space.
\end{thm}
Note that $X$ might be a strict subset of $Y$ as seen in Example~\ref{example-not-length} depicted
in Figure~\ref{figure-not-length}.
\begin{proof}
By Gromov's Compactness Theorem, there exists a compact space $Z$ and isometric embeddings $\varphi_j:X_j \to Z$
such that a subsequence of the $\varphi_j\left(X_j\right)$, still denoted $\varphi_j(X_j)$,
converges in the Hausdorff sense to a closed subset, $Y' \subset Z$.
We then apply Kuratowski's Theorem
to define isometric embeddings
$\varphi'_j=\iota\circ \varphi_j: X_j \to \ell^\infty\left(Z\right)$. Note that $K=\iota\left(Z\right)\subset \ell^\infty\left(Z\right)$ is compact and
\begin{equation}
\operatorname{spt} \varphi'_{j\#}\left(T_j\right) \subset Cl\left(\varphi'_j\left(X_j\right)\right) \subset \iota\left(Z\right) = K.
\end{equation}
Let $Y=\iota(Y')$ with the restricted metric.
We now apply the Ambrosio-Kirchheim Compactness Theorem to see that there exists a further
subsequence $\varphi'_{j_i \#} T_{j_i}$ converging weakly to an integral current
$S \in \intcurr_m\left(\ell^\infty\left(Z\right)\right)$. We claim $\operatorname{spt} S \subset Y$.
If not then there exists $x \in \operatorname{spt} S\setminus Y$, and there exists $r>0$
such that $B(x,r)\cap Y=0$. By definition of support,
$||S||(B(x,r/2)) >0$. By weak convergence, there is an $i$
sufficiently large that $||S_{j_i}||(B(x,r))>0$. In particular $x \in T_{r/2}(\operatorname{spt} S_{j_i})$.
Taking $i\to\infty$, we see that $x\in T_r(Y)$ because $Y$ is the Hausdorff
limit of the $\operatorname{spt} S_{j_i}$.
Since $E=\ell^\infty\left(Z\right)$ is a Banach space and there
is a uniform upper bound on the total mass, we apply Wenger's Flat=Weak Convergence
Theorem to see that
\begin{equation}
d^E_F\left(\varphi'_{j_i \#} T_{j_i}, S\right) \to 0.
\end{equation}
We now define our limit current space $\left(X,d,T\right)$ by taking $X=\rm{set}(S)$, $d=d_E$
and $T=S$. The identity map isometrically embeds $X$ into $E$ and takes $T$ to $S$.
Since $\rm{set}(S) \subset \operatorname{spt}(S) \subset Y$, we are done.
\end{proof}
We have the following immediate corollary of Theorem~\ref{GH-to-flat}:
\begin{cor} \label{lower-dim}
Given a sequence of precompact
$m$ dimensional integral current spaces, $M_j=\left(X_j, d_j, T_j\right)$,
with a uniform upper bound on their total mass such that $X_j$ converge in the
Gromov-Hausdorff sense to a compact limit space,$Y$,
of lower Hausdorff dimension, $dim_\mathcal{H}(Y)<m$, then $M_j$
converges in the intrinsic flat sense
to the $\bf{0}$ current space because the zero current is the only $m$ dimensional
integral current whose canonical set has Hausdorff dimension less than $m$.
\end{cor}
\begin{rmrk} \label{vol-to-zero}
Note that by Remark~\ref{finite} any collapsing sequence of Riemannian manifolds, $M_j^m$ such that
$\operatorname{Vol}\left(M_j\right) \to 0$, converges in the intrinsic flat sense to the
${\bf{0}}$ integral current
space. Thus even when the Gromov-Hausdorff limit is higher
dimensional as in Example~\ref{ex-jungle-gym}
the intrinsic flat limit may collapse to the ${\bf 0}$ current space.
\end{rmrk}
\subsection{The Infimum is Achieved} \label{subsect-flat-dist-2}
In this subsection we prove the infimum in the definition of the intrinsic flat
distance (\ref{eqn-local-defn}) is achieved for precompact integral current
spaces.
\begin{thm}\label{inf-dist-attained}\label{achieved}
Given a pair of precompact integral current spaces, $M=(X,d,T)$
and $M'=(X',d',T')$, there exists a compact metric space, $Z$,
integral
currents $U\in\intcurr_m\left(Z\right)$ and $V\in\intcurr_{m+1}\left(Z\right)$,
and isometric embeddings
$\varphi : \bar{X}\to Z$ and $\varphi':\bar{X}' \to Z$
with
\begin{equation} \label{eqn-Federer-Flat-3}
\varphi_\# T- \varphi'_\# T'=U+\partial V
\end{equation}
such that
\begin{equation}\label{eqn-local-defn-2}
d_{{\mathcal F}}\left(M,M'\right)={\mathbf M}\left(U\right)+{\mathbf M}\left(V\right).
\end{equation}
In fact, we can take $Z=\operatorname{spt}\left(U\right)\cup \operatorname{spt}(V)$.
\end{thm}
This theorem also holds for $M'=\bf{0}$, where we just take $T'=0$ and do not concern
ourselves with embedding $X'$ into $Z$.
In our proof of this theorem, we use the notion of an injective metric space and
Isbell's theorem regarding the existence of an injective envelope of a metric space
\cite{Isbell}:
\begin{defn}
A metric space $W$ is said to be injective iff it has the following property:
given any pair of metric spaces, $Y\subset Z$, and
any $1$ Lipschitz function, $f: Y\subset Z \to W$, we can extend $f$ to
a $1$ Lipschitz function $\bar{f}: Z \to W$.
\end{defn}
\begin{thm} [Isbell]
Given any metric space $X$, there is a smallest injective space,
which contains $X$, called the injective envelope.
Furthermore, when $X$ is compact, its injective envelope
is compact as well.
\end{thm}
We now prove Theorem~\ref{inf-dist-attained}.
\begin{proof}
Let $Z_n$ and $U_n\in\intcurr_m\left(Z_n\right)$ and $V_n\in\intcurr_{m+1}\left(Z_n\right)$
approach the infimum in the definition of the flat distance
between current spaces (\ref{eqn-local-defn}). That is there exists
isometric embeddings $\varphi_n: \bar{X}\to Z_n$ and $\varphi'_n: \bar{X}'\to Z_n$
such that
\begin{equation}
\varphi_{n\#}T - \varphi'_{n\#}T' = U_n +\partial V_n
\end{equation}
where
\begin{equation}
{\mathbf M}\left(U_n\right)+{\mathbf M}\left(V_n\right) \leq d_{{\mathcal F}}\left(M,M'\right) + \frac{1}{n}.
\end{equation}
We would like to apply Ambrosio-Kirchheim's Compactness Theorem, so we
need to find a common compact metric space, $Z$, and push $U_n$ and $V_n$
into this common space and then take their limits to find $U$ and $V$. We will
build $Z$ in a few stages using Gromov's Compactness Theorem and Isbell's
Theorem. The $Z_n$ we have right now need not be equicompact or
equibounded.
We first claim that $\varphi_n, \varphi_n'$ and $Z_n$ may be chosen so that
\begin{equation}
\operatorname{diam}(Z_n) \le 3 \operatorname{diam}(\varphi_n(\bar{X}))+ 3\operatorname{diam}(\varphi'_n(\bar{X}'))
=3 \operatorname{diam}(X) + 3 \operatorname{diam}(X').
\end{equation}
If not, then there exists $p_n\in \varphi_n(\bar{X})$ and $p_n'\in \varphi'_n(\bar{X}')$
such that the closed balls
\begin{equation}
\bar{B}(p_n,2\operatorname{diam}(X)) \cap \bar{B}(p_n',2\operatorname{diam}(X')) =\emptyset.
\end{equation}
Taking $A_n=Z_n\setminus (\bar{B}(p_n,2\operatorname{diam}(X)) \cup \bar{B}(p'_n,2\operatorname{diam}(X')))$,
we would then define
$
Z_n':=Z_n / A_n
$
with the quotient metric
\begin{equation}
d_{Z_n'}\left([z_1],[z_2]\right) := \inf \left\{ d_{Z_n}(x_1,a_1)+d_{Z_n}(a_2,x_2): \,\,
x_i \in [z_i], a_i \in A_n\right\}.
\end{equation}
Then $Z_n'$ has the required bound on diameter and we need only construct
the embeddings.
Let $p: Z_n \to Z_n/A$ be the projection. Then $p$ is an isometric
embedding when restricted to
$\varphi_n(X)\subset \bar{B}(p_n,\operatorname{diam}(X))$ or to
$\varphi_n(X')\subset \bar{B}(p'_n,\operatorname{diam}(X'))$. Thus
$p\circ\varphi_n:\bar{X} \to Z_n/A$ and
$p\circ\varphi'_n:\bar{X}' \to Z_n/A$ are isometric embeddings.
Furthermore $p$ is $1$-Lipschitz on $Z_n$, so
\begin{equation}
p_\#\varphi_{n\#}T -p_\# \varphi'_{n\#}T' = p_\#U_n +\partial p_\#V_n
\end{equation}
and, by Lemma~\ref{lem-push-mass},
\begin{equation}
{\mathbf M}\left(p_\#U_n\right)+{\mathbf M}\left(p_\#V_n\right) \leq
{\mathbf M}\left(U_n\right)+{\mathbf M}\left(V_n\right)
\end{equation}
So our first claim is proven.
Now let $Y_n:=\varphi_n(\bar{X}) \cup \varphi'_n(\bar{X}')\subset Z_n$ with the restricted
metric from $Z_n$. By our first claim, the diameters of the $Y_n$ are uniformly bounded.
In fact the $Y_n$ are equicompact because the number of disjoint
balls of radius $r$ may easily be estimated:
\begin{equation}
N(Y_n,r) \le N(\varphi_n(X),r) + N(\varphi_n'(X'),r)=N(X,r) + N(X',r).
\end{equation}
Thus, by Gromov's Compactness Theorem, there exists a compact metric
space, $Z'$, and isometric embeddings $\psi_n: Y_n \to Z'$.
Recall that $U_n\in\intcurr_m\left(Z_n\right)$ and $V_n\in\intcurr_{m+1}\left(Z_n\right)$,
so we need to extend $\psi_n$ to $Z_n$ in order to push forward
these currents into the common compact metric space, $Z$, and
take their limits.
By Isbell's Theorem, we may take $Z$ to be the injective envelope of $Z'$.
Since $Z$ is injective, we can extend the $1$-Lipschitz
maps, $\psi_n$, to $1$-Lipschitz maps, $\bar{\psi}_n: Z_n \to Z$.
So now we have a common compact metric space, $Z$,
isometric embeddings $\bar{\psi}_n\circ \varphi_n: \bar{X} \to Z$
and $\bar{\psi}_n\circ \varphi'_n:\bar{X}'\to Z$,
such that
\begin{equation} \label{four-currents}
\bar{\psi}_{n\#} \varphi_{n\#}T - \bar{\psi}_{n\#}\varphi'_{n\#}T' =
\bar{\psi}_{n\#}U_n +\partial \bar{\psi}_{n\#}V_n
\end{equation}
where
\begin{equation}
{\mathbf M}\left(\bar{\psi}_{n\#}U_n\right)+{\mathbf M}\left(\bar{\psi}_{n\#}V_n\right) \leq
d_{{\mathcal F}}\left(M,M'\right) + \frac{1}{n}.
\end{equation}
By Arzela-Ascoli's Theorem, after taking a subsequence, the isometric embeddings
$\bar{\psi}\circ\varphi_n: X\to Z$ and $\bar{\psi}\circ\varphi'_n: X'\to Z$
converge uniformly to isometric embeddings
$\varphi: X \to Z$ and $\varphi': X'\to Z$. As in Remark~\ref{rmrk-unif-to-weak},
we then have weak convergence:
\begin{equation}
\bar{\psi}_{n\#} \varphi_{n\#}T \rightharpoonup \varphi_\#T \textrm{ and }
\bar{\psi}_{n\#}\varphi'_{n\#}T' \rightharpoonup \varphi'_\# T'.
\end{equation}
By Ambrosio-Kirchheim's Compactness Theorem, after possibly taking
a further subsequence, there exists $U\in\intcurr_m\left(Z\right),
V\in\intcurr_{m+1}\left(Z\right)$ such that
\begin{equation}
\bar{\psi}_\#U_n \rightharpoonup U \textrm{ and } \bar{\psi}_\#V_n \rightharpoonup V.
\end{equation}
In particular, $\varphi_\#T - \varphi'_\# T'= U- \partial V$.
By the lower semicontinuity of mass (c.f. Remark~\ref{rmrk-lower-mass}),
\begin{equation}
{\mathbf M}(U)+{\mathbf M}(V) \leq
d_{{\mathcal F}}\left(M,M'\right) + \frac{1}{n} \qquad \forall n \in \mathbb{N}
\end{equation}
and we are done.
\end{proof}
\subsection{Current Preserving Isometries} \label{subsect-flat-dist-3
We can now prove that the intrinsic flat distance is a distance
on the space of precompact oriented Riemannian manifolds with boundary
and, more generally, on precompact integral current spaces in $\mathcal{M}_0^m$.
\begin{defn}
Given $M,N \in \mathcal{M}^m$, an isometry $f: X_M \to X_N$ is called
a current preserving isometry between $M$ and $N$, if its
extension $\bar{f}: \bar{X}_M \to \bar{X}_N$ pushes forward the
current structure on $M$ to the current structure on $N$:
$
\bar{f}_\# T_M= T_N
$
\end{defn}
When $M$ and $N$ are oriented Riemannian manifolds
or other Lipschitz manifolds with the standard
current structures as described in Remark~\ref{Lip-mani-charts}
then orientation preserving isometries are exactly current preserving
isometries. {See Remark~\ref{rmrk-biLip-matching}.}
\begin{thm} \label{zero-mani}
If $M,N$ are precompact integral current spaces such that $d_{{\mathcal F}}\left(M,N\right)=0$
then there is a current preserving isometry from $M$ to $N$.
Thus $d_{\mathcal{F}}$ is a distance on $\mathcal{M}^m_0$.
\end{thm}
It should be noted that a pair of precompact metric
spaces, $X,Y$ such that $d_{GH}(X,Y)=0$ need not be isometric (e.g.
the Gromov-Hausdorff distance between a Riemannian manifold, and
the same manifold with one point removed is $0$). However, if $X$ and $Y$
are compact, then Gromov proved
$d_{GH}(X,Y)=0$ implies they are isometric \cite{Gromov-metric}.
While we do not require that our spaces be complete, the definition
of an integral current space requires that the spaces be completely
settled [Defn~\ref{def-settled}] so that $X=\rm{set} (T)$
[Defn~\ref{defn-integral-current-space}]. This is as essential to
the proof of Theorem~\ref{zero-mani} as the compactness is
essential in Gromov's theorem.
Precompactness on the other hand, is not a necessary condition.
Theorem~\ref{zero-mani} can be extended to noncompact
integral current spaces applying Theorem 6.1 in the second author's
compactness paper \cite{Wenger-compactness}.
\begin{proof}
By Theorem~\ref{inf-dist-attained} and the fact that an integral current
has zero mass iff it is $0$,
we know there exists a compact
space $Z$ and isometric embeddings,
$\varphi : \left(\bar{X}_M, d_{\bar{X}}\right) \to \left(Z,d\right)$ and $\psi: \left(\bar{X}_N , d_{\bar{X}_N}\right)\to \left(Z,d\right)$,
with
\begin{equation} \label{eqn-zero-mani-1}
{\varphi}_\# T_M- {\psi}_\# T_N=0 \in \intcurr_m\left(Z\right).
\end{equation}
Thus
\begin{equation}
\rm{set}\left({\varphi}_\# T_M\right) =\rm{set} \left({\psi}_\# T_N\right).
\end{equation}
By Lemma~\ref{lemma-isom-to-set}, we know $\varphi: X_M \to \rm{set}\left({\varphi}_\# T_M\right)$
and $\psi: X_N \to \rm{set}\left({\psi}_\# T_N\right)$ are isometries.
We define our isometry $f: X_M \to X_N$ to be $f=\psi^{-1}\circ \varphi$.
Then $\bar{f}: \bar{X}_M \to \bar{X}_N$, pushes $T_M\in \intcurr_m\left(\bar{X}_M\right)$
to $\bar{f}_\# T_M \in \intcurr_m\left(\bar{X}_N\right)$,
so that with (\ref{eqn-zero-mani-1}) we have,
\begin{equation}
{\psi}_\# \bar{f}_\# T_M= {\varphi}_\# T_M = {\psi}_\# T_N.
\end{equation}
Since ${\psi}_\# \,\left(f_\#T_M - T_N\right) =0 \in \intcurr_m\left(Z\right)$ and ${\psi}$
is an isometry, we have
$f_\#T_M - T_N=0 \in \intcurr_m\left(\bar{X}_N\right)$.
\end{proof}
The following is an immediate consequence of Theorem~\ref{zero-mani}:
\begin{cor}
If $M^m$ and $N^m$ are precompact oriented Riemannian manifolds with finite volume,
then $d_{\mathcal{F}}(M^m,N^m)=0$ iff there is an orientation preserving isometry,
$\psi: M^m \to N^m$. Thus $d_{\mathcal{F}}$ is a distance on the space of precompact
oriented Riemannian manifolds with finite volume.
\end{cor}
\begin{rmrk}
Initially we were hoping to prove that if the intrinsic flat distance between
two Riemannian manifolds
is zero then the manifolds are isometric. This is false
unless the manifold has an orientation reversing isometry
as we prove in Theorem~\ref{zero-mani}. We thought we
might use a $\mathbb{Z}_2$ notion of integral currents to avoid the issue of
orientation. However, at the time there was no such theory, so we
settled on this version of the theorem with this notion of intrinsic
flat distance. Very recently Ambrosio-Katz \cite{Ambrosio-Katz}
and Ambrosio-Wenger \cite{Ambrosio-Wenger}
completed work covering this theory and one expects this will
lead to interesting new ideas.
Alternatively one could try to use
the even more recent theory of DePauw-Hardt \cite{DH-chains}.
\end{rmrk}
\section{{\bf Sequences of Integral Current Spaces}}
In this section we describe the properties of sequences of integral
current spaces which converge in the intrinsic flat sense.
In Subsection~\ref{subsectseq} we take a
Cauchy or converging sequence of precompact integral current
spaces and construct a common metric space, $Z$, into which the entire sequence
embeds [Theorem~\ref{cauchy} and Theorem~\ref{converge}]. Note that
$Z$ need not be compact even when the spaces are. Relevant examples
are given and an open question appears in Remark~\ref{rmrk-cauchy}.
In Subsection~\ref{subsect-properties} we remark on the properties
of converging sequences of integral current spaces.
We prove the lower semicontinuity
of mass [Theorem~\ref{mass-drops}] which is a direct consequence of
Ambrosio-Kirchheim \cite{AK}. We remark on the continuity of filling
volume which follows from work of the second author \cite{Wenger-flat}.
In Subsection~\ref{subsect-SorWen1} we state consequences of
the authors' first paper \cite{SorWen1} concerning limits of sequences
of Riemannian manifolds with contractibility conditions as in work
of Greene-Petersen \cite{Greene-Petersen}. We discuss how to
avoid the kind of cancellation depicted in Example~\ref{example-cancels}
depicted in Figure~\ref{figure-cancels} using Gromov's filling volume \cite{Gromov-filling}.
In Subsection~\ref{subsect-Ricci} we discuss noncollapsing sequences
of manifolds with nonnegative Ricci or positive scalar curvature particularly
Theorem~\ref{thm-ricci} and Conjecture~\ref{conj-scalar}
which appear in our first paper \cite{SorWen1}.
In Subsection~\ref{subsect-Wenger-compactness} we state the second
author's compactness theorem [Theorem~\ref{thm-Wenger-compactness}]
which is proven in \cite{Wenger-compactness}. We then prove
Theorem~\ref{Cauchy-to-Converge} which provides a common metric
space $Z$ for a Cauchy sequence bounded as in the compactness theorem.
In particular, any Cauchy sequence of integral current spaces with a uniform
upper bound on diameter and total mass converges to an integral current
space.
\subsection{Embeddings into a Common Metric Space}\label{subsectseq}\label{subsect-flat-dist-4}
In this subsection we prove Theorems~\ref{cauchy},~\ref{converge}
and~\ref{convergeto0} which describe how
Cauchy and converging
sequences of integral current spaces, $M_i$, can be isometrically embedded into a common
separable complete metric space $Z$ as a flat Cauchy or converging sequence.
These theorems are essential to understanding sequences of manifolds which
do not have Gromov-Hausdorff limits.
We will also apply them
to prove Theorem~\ref{Cauchy-to-Converge}.
\begin{thm}\label{cauchy}
Given an intrinsic flat Cauchy sequence of
integral current spaces, $M_j=\left(X_j, d_j, T_j\right) \in \mathcal{M}^m$,
there exists a separable complete
metric space
$Z$, and a sequence
of isometric embeddings $\varphi_j: X_j \to Z$
such that $\varphi_{j\#}T_j$ is a flat Cauchy
sequence of integral currents in $Z$.
\end{thm}
The classic example of a Cauchy sequence of integral currents converging to
Gabriel's horn shows that a uniform upper bound on mass is required to
have a limit space which is an integral current space [Example~\ref{ex-Gabriel's-horn}].
So the Cauchy sequence in this theorem need not converge without an
additional assumption on total mass.
In Example~\ref{ex-unbounded} we see that even with the uniform bound on
total mass, the sequence may have a limit which is unbounded.
In Example~\ref{ex-many-tips} depicted in Figure~\ref{fig-many-tips} we
see that even with a uniform bound on total mass and diameter, the limit
space need not be precompact. See also Remark~\ref{rmrk-cauchy} and Theorem~\ref{Cauchy-to-Converge}.
If we assume that the Cauchy sequence of integral current spaces converges
to a given integral current space, than we can embed the entire sequence
including the limit into a common metric space $Z$:
\begin{thm}\label{converge}\label{converges}
If a sequence of
integral current spaces,
$M_{j}=\left(X_j, d_j, T_j\right)$, converges
to an
integral current space,
$M_0=\left(X_0,d_0,T_0\right)$
in the intrinsic flat sense , then
there is a separable
complete metric space, $Z$, and isometric embeddings $\varphi_j: X_j \to Z$ such that
$\varphi_{j\#}T_j$ flat converges to $\varphi_{0\#} T_0$ in $Z$
and thus converge weakly as well.
\end{thm}
Note that one cannot construct a compact $Z$ as Gromov did in \cite{Gromov-groups}
even when one knows the sequence converges in the intrinsic flat sense to a compact
limit space and that the sequence has a uniform bound on total mass. The sequence of hairy spheres
in Example~\ref{ex-hairy-sphere} converges to a sphere in the flat norm but cannot be isometrically
embedded into a common compact space because the sequence is not equicompact.
The special case of Theorem~\ref{converges} where $M_j$ converges to the
$\bf{0}$ space can have prescribed pointed isometries:
\begin{thm}\label{convergeto0}
If a sequence of
integral current spaces
$M_{j}=\left(X_j, d_j, T_j\right)$ converges
in the intrinsic flat sense to the
zero integral current space, $\bf{0}$, then
we may choose points $p_j\in X_j$ and a
separable
complete metric space, $Z$, and isometric embeddings
$\varphi_j: \bar{X}_j \to Z$ such that
$\varphi_j(p_j)=z_0\in Z$ and
$\varphi_{j\#}T_j$ flat converges to $0$ in $Z$ and thus converges weakly as well.
\end{thm}
We prove this theorem first since it is the simplest.
\begin{proof}
By the definition of the flat distance, we know there exists a complete
metric space $Z_j$ and $U_j\in \intcurr_m(Z_j)$ and $V_j\in \intcurr_{m+1}(Z_j)$
and an isometry $\varphi_j: X_j\to Z_j$ such that
$\varphi_{j\#}T_j=U_j +\partial V_j$ and
\begin{equation}
d_\mathcal{F}(M_j, {\bf{0}}) \le{\mathbf M}(U_j)+{\mathbf M}(V_j) \to 0.
\end{equation}
We may choose $Z_j=\operatorname{spt} U_j \cup \operatorname{spt} V_j$, so it is separable.
We then create a common complete separable metric space $Z$ by gluing
all the $Z_j$ together at the common point $\varphi_j(p_j)$:
\begin{equation}
Z=Z_1 \sqcup Z_2 \sqcup\cdots
\end{equation}
where $d_Z(z_1,z_2)= d_{Z_i}(z_1,z_2)$ when there exists an $i$ with $z_1,z_2\in Z_i$
and
\begin{equation}
d_Z(z_i,z_j)=d_{Z_i}(z_i,\varphi_i(p_i))+d_{Z_j}(z_j,\varphi_j(p_j)).
\end{equation}
We then identify all the $\varphi_i(p_i)=\varphi_j(p_j)\in Z$ so that this is a metric.
Since mass is preserved under isometric embeddings, we have
$d^Z_F(\varphi_{j\#}T_j,0) \le {\mathbf M}(U_j)+{\mathbf M}(V_j) \to 0$.
\end{proof}
To prove Theorems~\ref{cauchy} and~\ref{converge}, we need to glue together
our spaces $Z$ in a much more complicated way. So we first prove the
following two lemmas and then prove the theorems.
We close the section with Remark~\ref{rmrk-cauchy}
which discusses a related open problem.
Recall the well known gluing lemma [Lemma~\ref{gluing-lemma}]
that we applied to prove the triangle inequality in Subsection~\ref{subsect-triangle}.
One may apply this gluing of metric spaces countably many times, to glue together countably
many distinct metric spaces:
\begin{lem} \label{lemma-tree-graph} \label{lemma-shrub}
We are given
a connected tree with countable vertices $\{V_i : i \in A \subset \mathbb{N}\}$ and edges
$\{E_{i,j} : (i,j)\in B \}$ where $B\subset \{(i,j): \, i<j, i,j \in A\}$,
and a corresponding countable collection of metric spaces
$\{X_i: i \in A\}$ and $\{ Z_{i,j} : \, \left(i,j\right) \in B\}$
and isometric embeddings
\begin{equation}
\varphi_{i, (i,j)}: X_i \to Z_{i,j} \textrm{ and }
\varphi_{j, (i,j)}: X_j \to Z_{i,j} \qquad \forall (i,j)\in B.
\end{equation}
Then there is a unique metric space $Z$ defined by gluing the $Z_{i,j}$ along the isometric
images of the $X_i$.
In particular there exists isometric embeddings $f_{i,j}: Z_{i,j} \to Z$ for all
$(i,j)\in B$ such that for all $(i,j),(j,k) \in B$ we have
isometric embeddings
\begin{equation} \label{eqn-f-i-j}
f_{i,j} \circ \varphi_{j,(i,j)} = f_{j,k} \circ \varphi_{j, (j,k)}: \, X_{j} \to Z.
\end{equation}
If $Z_{i,j}$ are separable then so is $Z$.
\end{lem}
\begin{proof}
Let $Z$ be the disjoint union of the $Z_{i,j}$. We define a quasimetric on $Z$ and then identify the
images of the $X_i$ so that the quasimetric becomes a metric. Let $z,z'\in Z$, so
each lies in one of the $Z_{i,j}$ and thus has a corresponding edge $E\left(z\right), E\left(z'\right)\in \{E_{i,j}: (i,j)\in B\}$.
If $E\left(z\right)=E\left(z'\right)$ then they lie in the same $Z_{i,j}$ and we
let $d_Z\left(z,z'\right):=d_{Z_{i,j}}\left(z,z'\right)$ which we denote as $d_{i,j}$ to avoid excessive subscripts below.
If $E(z)\neq E(z')$, then because the graph is a connected tree there is a unique sequence of distinct edges
$\{E_{i_0,i_1}, E_{i_1,i_2}, ..., E_{i_{n},i_{n+1}}\}$ where $E\left(z\right)=E_{i_0,i_1}$ and $E\left(z'\right)=E_{i_n,i_{n+1}}$.
We define
\begin{equation*}
\begin{split}
d_Z\left(z,z'\right) & = \inf \Bigg\{ \,\, d_{i_0,i_1}\left(z, \varphi_{i_1, (i_0,i_1)}\left(y_1\right)\right) \\
& \qquad \qquad
+ \sum_{j=1}^{n-1} d_{i_j,i_{j+1}}\left(\varphi_{i_j, (i_j,i_{j+1})}\left(y_j\right),
\varphi_{i_{j+1}, (i_j, i_{j+1}) }\left(y_{j+1}\right) \right)\\
& \qquad \qquad
+ \, d_{i_n, i_{n+1}}\left(\varphi_{i_n, (i_n,i_{n+1})}\left(y_n\right), z' \right)
\, :\,\left(y_1,..., y_n\right)\in X_{i_1}\times \cdots \times X_{i_n} \Bigg\}.
\end{split}
\end{equation*}
One may then easily verify the triangle inequality $d_Z\left(a,b\right)+ d_Z\left(b,c\right) \ge d_Z\left(a,c\right)$ by breaking into cases
regarding the location of $E\left(b\right)$ relative to $E\left(a\right)$ and $E\left(c\right)$.
Finally we identify points $z$ and $z'$ such that $d_Z\left(z,z'\right)=0$.
\end{proof}
We can now prove Theorem~\ref{cauchy}:
\begin{proof}
Recall that we have a Cauchy sequence of current spaces, so for all $\epsilon>0$, there exists $N_\epsilon \in \mathbb{N}$
such that
\begin{equation}
r_{i,j}= d_{{\mathcal F}}\left(M_i,M_j\right) < \epsilon \qquad \forall i,j \ge N_\epsilon.
\end{equation}
By the definition of the intrinsic flat distance between $M_i$ and $M_j$ in
(\ref{eqn-local-defn}),
there
exist metric spaces $Z_{i,j}$ and isometric embeddings
$\varphi_{i, (i,j)}: \bar{X}_i\to Z_{i,j}$ and $\varphi_{j,(i,j)}: \bar{X}_j\to Z_{i,j}$
and integral currents
$U_{i,j} \in \intcurr_m\left(Z_{i,j}\right)$ and $V_{i,j} \in \intcurr_{m+1}\left(Z_{i,j}\right)$
with
\begin{equation} \label{cauchy-1}
\varphi_{i, (i,j)\#}T_i-\varphi_{j, (i,j)\#}T_j=
U_{i,j} + \partial V_{i,j} \in \intcurr_m\left(Z_{i,j}\right)
\end{equation}
such that
\begin{equation} \label{cauchy-exact-1}
r_{i,j}:= d_{{\mathcal F}}\left(M_i,M_j\right) =
d_F^{Z_{i,j}}\left(\varphi_{i,(i,j)\#}T_i, \varphi_{j,(i,j)\#}T_j\right)\le {\mathbf M}\left(U_{i,j}\right)+{\mathbf M}\left(V_{i,j}\right) \le 3r_{i,j}/2.
\end{equation}
We choose $Z_{i,j}=\operatorname{spt} U_j \cup \operatorname{spt} V_j$ and so it is separable.
Since the sequence is Cauchy, we know
there exists a subsequence $j_k \in \Bbb{N}$ such that
$j_1=1$ and when $k\ge 2$ we have $r_{j_k,i}\le 1/2^k$
$\forall i \ge j_k$. In particular $r_{j_k,j_{k+1}}\le 1/2^k$ when $k\ge 2$.
We call this special subsequence, a {\em geometric subsequence}.
We now apply Lemma~\ref{lemma-tree-graph} to the graph whose
vertices are $\{V_i : i \in A = \mathbb{N} \}$ and edges
$\{E_{i,j} : \left(i,j\right) \in B \subset \mathbb{N}\times \mathbb{N} \}$ where
\begin{equation}
B=\{\left(j_k,j_{k+1}\right): k \in \mathbb{N}\} \cup \{ \left(j_k,i\right) : i=j_k, ..., j_{k+1} -1\}.
\end{equation}
Intuitively this is a tree whose trunk is the geometric subsequence and whose branches consist
of single edges attached to the nearest vertex on the trunk.
As a result we have a complete metric space $Z$ and isometric embeddings
$f_{i,j}:Z_{i,j} \to Z$ such that
\begin{equation} \label{cauchy2}
f_{i,j} \circ \varphi_{j,(i,j)} = f_{j,i'} \circ \varphi_{j, (j,i')}: \, X_{j} \to Z
\end{equation}
are isometric embeddings for all $(i,j), (j,i') \in B$. In particular
each current space $M_j$ has been mapped to a unique current in $Z$:
\begin{equation} \label{cauchy-4}
T'_j:= f_{i,j\#} \varphi_{j,(i,j)\#} T_j = f_{j,i'\#} \varphi_{j, (j,i')\#} T_j
\in \intcurr_m\left(Z\right)
\end{equation}
So $f_{i,j} \circ \varphi_{j,(i,j)}$
is a current preserving isometry from
$M_j=\left(X_j, d_j, T_j\right)$
to $\left(\rm{set} (T'_j), d_Z, T'_j\right)$.
Applying (\ref{cauchy-1}), we have for any $\left(i,j\right)\in B$:
\begin{equation}
T'_i-T'_j = f_{i,j \#}\varphi_{i, (i,j)\#}T_i-f_{i,j\#}\varphi_{j, (i,j)\#}T_j= f_{i,j\#}U_{i,j} + \partial f_{i,j\#}V_{i,j} \in \intcurr_m\left(Z\right).
\end{equation}
Since mass is conserved under isometries (c.f. Lemma~\ref{lem-push-mass}) we have
\begin{equation} \label{add-error-1}
d^Z_F\left(T'_i, T'_j\right) \le {\mathbf M}\left(f_{i,j\#}U_{i,j}\right) + {\mathbf M}\left( f_{i,j\#}V_{i,j}\right) = {\mathbf M}\left(U_{i,j}\right) + {\mathbf M} \left(V_{i,j}\right) = 3r_{i,j}/2.
\end{equation}
In particular by our choice of $B$ in (\ref{cauchy2}), we have for the geometric subsequence:
\begin{equation}
d^Z_F\left(T'_{j_k}, T'_{j_{k+1}}\right) \le 3/2^k \,\,\forall k\ge 2.
\end{equation}
For $i,i' \ge j_2$ we have
$k,k' \ge 2$ respectively such that $\left(i,j_k\right), \left(i', j_{k'}\right) \in B$ such that
\begin{equation}
d^Z_F\left(T'_{j_k}, T'_{i}\right) \le 3/2^k \textrm{ and }
d^Z_F\left(T'_{j_{k'}}, T'_{i'}\right) \le 3/2^{k'}.
\end{equation}
So we have
\begin{eqnarray}
d^Z_F\left(T'_i, T'_{i'}\right) &\le& d^Z_F\left(T'_{j_k}, T'_{i}\right) + \sum_{h=k}^{k'-1} d^Z_F\left(T'_{j_h}, T'_{j_{h+1}}\right)+ d^Z_F\left(T'_{j_{k'}}, T'_{i'}\right) \\
&\le& 3/2^k + \left(3/2^k + 3/2^{k+1} + \cdots +3/2^{k'}\right) < 9/2^k.
\end{eqnarray}
and thus our sequence of integral current spaces has been
mapped into a Cauchy sequence of integral currents.
\end{proof}
We now prove Theorem~\ref{converge}. Since we have
already proven Theorem~\ref{convergeto0}, we will assume
we have a nonzero limit in this proof
\begin{proof}
As in the proof of Theorem~\ref{cauchy}, we take a geometrically converging
subsequence of the converging sequence of current spaces.
This time we apply Lemma~\ref{lemma-tree-graph} to the tree whose
vertices are $\{V_i : i \in A = 0 \cup \mathbb{N}\}$ and edges $\{E_{i,j} : \left(i,j\right) \in B \subset \mathbb{N}\times \mathbb{N} \}$ where
\begin{equation}
B=\{\left(j_k,0\right): k \in \mathbb{N}\} \cup \{ \left(j_k,i\right) : i=j_k, ..., j_{k+1} -1\}.
\end{equation}
so that all the terms in the geometric subsequence will be directly attached to the limit, and everything else
will be attached to the subsequence as before. As in (\ref{cauchy-4}) we obtain unique currents
$T_j' \in \intcurr_m\left(Z\right)$ such that $\left(\rm{set} (T'_j), d_Z, T'_j\right)$ has a current preserving isometry with $\left(X_j, d_j, T_j\right)$. This time our currents flat converge, because for any $i \in [j_k, j_{k+1}-1]$ we have
\begin{equation}
d^Z_F\left(T'_i, T'_{0}\right) \le d^Z_F\left(T'_{j_k}, T'_{0}\right) + d^Z_F\left(T'_i, T'_{j_{k}}\right) \le 3/2^k + 3/2^k.
\end{equation}
Weak convergence then follows by Remark~\ref{rmrk-flat-implies-weak}.
\end{proof}
\begin{rmrk} \label{rmrk-cauchy}
We do not know if the sequence $\varphi_{j\#}T_j$ in Theorem~\ref{cauchy}
when given a uniform bound on total mass
converges in the flat sense to an integral current in $Z$. Without a uniform
bound on total mass it is possible there is no limit integral current space [Example~\ref{ex-Gabriel's-horn}].
It is an open question
whether flat Cauchy sequences with uniform upper bounds on total
mass have flat converging subsequences which converge to an integral current
in the sense of Ambrosio-Kirchheim. In Federer-Fleming, one needs to add
a diameter bound because integral currents in Federer-Fleming have compact support.
In Ambrosio-Kirchheim compactness is never assumed so an unbounded limit
like the one in Example~\ref{ex-unbounded} is not a counter example here.
In Theorem~\ref{Cauchy-to-Converge} we prove
that adding a uniform bound
on diameter as well as the bound on total mass,
we can find a common metric space $Z$
where where $\varphi_{j\#}T_j$ do converge. The metric space $Z$ in that theorem
may not be the metric space constructed in Theorem~\ref{cauchy}.
To prove that theorem we need Theorem~\ref{converge} as well as
the second author's compactness theorem, Theorem~\ref{thm-Wenger-compactness}.
It would be of interest to eliminate the bound on diameter or find a counter example.
\end{rmrk}
\subsection{Properties of Intrinsic Flat Convergence} \label{subsect-properties}
As a consequence of Theorems~\ref{converge} and~\ref{convergeto0}
and Kuratowski's Embedding Theorem,
we may now observe that sequences of integral current spaces that
converge in the intrinsic flat sense have all the same properties
Ambrosio-Kirchheim have proven for
sequences of integral currents that converge weakly in a Banach space.
Most importantly, one has the lower semicontinuity of mass. Applying work
of the second author in \cite{Wenger-flat} [Theorem 1.4], one also observes that one has
continuity of the filling volume. Here we only give the details on lower
semicontinuity of mass and leave it to the reader to extend the ideas to
other properties of integral currents.
\begin{thm}\label{mass-drops}
If a sequence of
integral current spaces
$M_{j}=\left(X_j, d_j, T_j\right)$ converges in the intrinsic flat sense
to
$M_0=\left(X_0,d_0,T_0\right)$
then $\partial M_j$ converges to $\partial M_0$ in the intrinsic flat sense.
\begin{equation}
\liminf_{j\to\infty} {\mathbf M}\left(M_j\right) \ge {\mathbf M} \left(M_0\right)
\textrm{ and }
\liminf_{j\to\infty} {\mathbf M}\left(\partial M_j\right) \ge {\mathbf M} \left(\partial M_0\right).
\end{equation}
\end{thm}
In Example~\ref{example-cancels} depicted in Figure~\ref{figure-cancels}
we see that the mass of the limit
space may be $0$ despite a uniform lower bound on the mass of the sequence.
\begin{proof}
First we isometrically embed the converging sequence into a common metric
space, $Z$, applying Theorem~\ref{converge} and Theorem~\ref{convergeto0}:
$\varphi_j: \bar{X}_j \to Z$ such that $\varphi_{j\#}T_j$ converges in the
flat sense in $Z$ to $\varphi_{0\#}T_0$. Note that
$$
d^Z_F(\partial \varphi_{j\#} T_j,\partial \varphi_{0\#} T_0) \le
d^Z_F(\varphi_{j\#} T_j,\varphi_{0\#} T_0) \to 0.
$$
By the definition of $\partial M=(\rm{set} (\partial T), d, \partial T)$
and the fact that $\partial \varphi_{j\#} T=\varphi_{j\#}\partial T$,
we have
\begin{equation}
d_{\mathcal{F}}(\partial M_j, \partial M_0) \le d^Z_F(\varphi_{j\#}\partial T_j,
\varphi_{0\#}T_0)\to 0.
\end{equation}
Immediately below the definition of weak convergence of currents in a metric
space $Z$ in \cite{AK}[Defn 3.6], Ambrosio
Kirchheim remark that the mapping $T \mapsto ||T||(A)$ is lower semicontinuous
with respect to weak convergence for any open set $A \subset Z$.
Since $\varphi_{j\#} T_j$ converge weakly to $\varphi_{0\#}T_0$, we may take
$A=Z$ and apply Lemma~\ref{lem-push-mass}, to see that
\begin{equation}
\liminf_{j\to\infty} {\mathbf M}(M_j)=\liminf_{j\to\infty}{\mathbf M}(\varphi_{j\#}T_j)
\ge {\mathbf M}(\varphi_{0\#} T_0) = {\mathbf M}(M_0).
\end{equation}
The same may be done to the boundaries to conclude
that
$$
\liminf_{j\to\infty} {\mathbf M}\left(\partial M_j\right) \ge {\mathbf M} \left(\partial M_0\right).
$$
\end{proof}
\begin{rmrk}
Note that there are also local versions of the lower semicontinuity of mass
which can be seen by taking $A$ in the proof above to be a ball $B_{\varphi_0(x_0)}(r)$.
These local versions require an application of Ambrosio-Kirchheim's
Slicing Theorem \cite{AK} Thm 5.6,
which implies that $\varphi_{j\#}T_j \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: B_{\varphi_0(x_0)}(r)$ is
an integral current for almost all values of $r$.
The reader is referred to \cite{SorWen1} where local versions of lower
semicontinuity of mass and continuity of filling volume are applied.
\end{rmrk}
\subsection{Cancellation and Intrinsic Flat Convergence} \label{subsect-SorWen1}
When a sequence of integral currents converges to the ${0}$ current
due to the effect
of two sheets of opposing orientation coming together, this is referred to as
cancellation. In Example~\ref{example-cancels} depicted in Figure~\ref{figure-cancels},
we see that the same effect can occur causing a sequence of Riemannian
manifolds to converge in the intrinsic flat sense to the ${\bf 0}$ current space. Naturally
it is of great importance to avoid this situation.
In \cite{SorWen1}, the authors proved a few theorems providing conditions
that prevent cancellation of certain weakly converging sequences of integral currents.
These theorems immediately
apply to prevent the cancellation of certain sequences of Riemannian manifolds
although they do not extend to arbitrary integral current spaces. The reader is
referred to \cite{SorWen1} for the most general statements of these results.
In this section we give some of the intuition that led to these results,
then review Greene-Petersen's compactness theorem and finally
review a result of \cite{SorWen1}, Theorem~\ref{thm-contractibility-no-cancellation},
which states that under the conditions of Greene-Petersen's theorem,
there is no cancellation and, in fact,
the intrinsic flat and Gromov-Hausdorff limits agree.
\begin{rmrk}
The initial observation that lead to the results in \cite{SorWen1} was that the
sequence in Example~\ref{example-cancels} depicted in
Figure~\ref{figure-cancels} has increasing topological type.
The only way to bring two sheets together with an intrinsic distance on a smooth
Riemannian manifold, was to create many small tubes between the two sheets,
and all these tubes lead to increasing local topology.
\end{rmrk}
\begin{rmrk} \label{rmrk-avoid-cancellation}
The second observation was that, in order to avoid cancellation, one needed
to locally bound the filling volume of spheres away from $0$. More precisely
the filling volumes of distance spheres of radius $r$ had to be bounded below by
$Cr^m$, so that the filling volumes in the limit would have the same bound.
Since the volume of a ball is larger than the filling volume of the sphere,
we could then prove the limit points had positive density.
\end{rmrk}
Note that
if a sequence of Riemannian manifolds converges to a Riemannian
manifold with a cusp singularity as in Example~\ref{ex-to-cusp}
depicted in Figure~\ref{fig-to-cusp}, the cusp point disappears
in the limit because it does not have positive density [Example~\ref{example-settled},
Example~\ref{basic-mani-sing}]. To avoid cancellation,
we need to prevent points from disappearing.
\begin{figure}[h]
\centering
\includegraphics[width=4.5in]{GHFcusp1.jpg}
\caption{The intrinsic flat limit does not include the tip of the cusp.}
\label{fig-to-cusp}
\end{figure}
In Gromov's
initial paper defining filling volume, he proved the filling volume could be
bounded from below by the filling radius and the filling radius could be bounded
from below by applying contractibility estimates \cite{Gromov-filling}.
Greene-Petersen applied Gromov's technique to estimate the filling
volumes of balls and consequently prove the following compactness
theorem \cite{Greene-Petersen}. They needed a uniform estimate
on contractibility to prove their theorem:
\begin{defn}
On a Riemannian manifold, $M^m$,
a geometric contractibility function, $\rho:(0,r_0] \to (0,\infty)$, is a function
such that $\lim_{r\to 0}\rho(r)=0$ and such that
any ball $B_p(r)\subset M^m$ is contractible in
$B_p\left(\rho\left(r\right)\right)\subset M^m$.
\end{defn}
\begin{thm}[Greene-Petersen] \label{thm-Greene-Petersen} If a
sequence of Riemannian manifolds $M^m_j$
without boundary have a uniform geometric contractibility function,
$\rho:(0,r_0] \to (0,\infty)$
then one can construct uniform lower bound $\nu_{\rho,m}: (0,D] \to (0,\infty)$ such that
\begin{equation}
\operatorname{Vol}\left(B_p\left(r\right) \right) \ge {\operatorname{Fillvol}}(\partial B_p(r)) \ge \nu_{\rho,m}(r)
\end{equation}
for all balls $B_p(r)$ in all the manifolds.
If, in addition, there is a uniform upper bound on volume $\operatorname{Vol}(M^m_j)\le V$,
then a subsequence $M_j^m \stackrel { \textrm{GH}}{\longrightarrow} Y$.
\end{thm}
Immediately below the statement of this theorem, Greene-Petersen mention
that if $\rho$ is linear, $\rho(r)=\lambda r$, then there exists a constant
$C_m>0$ such that $\nu_{\rho,m}(r)\ge C_m r^m$. This is exactly the bound
needed to avoid cancellation.
If the geometric contractibility function $\rho$ is not linear then one can have a sequence
of Riemannian manifolds which converge to a Riemannian
manifold with a cusp singularity as in Example~\ref{ex-to-cusp}
depicted in Figure~\ref{fig-to-cusp}.
The lack of a uniform linear geometric contractibility function for that sequence of
is depicted in Figure~\ref{fig-GHFcusp2}.
\begin{figure}[h]
\centering
\includegraphics[width=4.5in]{GHFcusp2.jpg}
\caption{The first ball contracts in a ball of twice its radius, the second in a ball of
3 times its radius, the next in a ball of five times its radius...}
\label{fig-GHFcusp2}
\end{figure}
Cones have linear geometric contractibility functions (as seen in
Figure~\ref{fig-GHFcusp3}). Riemannian manifolds with conical
singularities viewed as integral current spaces include their
conical singularities [Example~\ref{example-settled},
Example~\ref{basic-mani-sing}].
\begin{figure}[h]
\centering
\includegraphics[width=4.5in]{GHFcusp3.jpg}
\caption{The contractibility function is $\rho(r)=2r$ here.}
\label{fig-GHFcusp3}
\end{figure}
In \cite{SorWen1}, we dealt with a far more general class of integral current
spaces than Riemannian manifolds. We began by applying Gromov's
compactness theorem to isometrically embed the sequence into a
common metric space where
we used a notion of integral filling volume (c.f. \cite{Wenger-isoper}),
which is well defined for integral currents without boundary.
We did not use Greene-Petersen's
smoothing arguments applying Ambrosio-Kirchheim's Slicing Theorem instead.
We needed to adapt everything to integral filling
volumes, so we applied a new
Lipschitz extension theorem akin to that of Lang-Schlichenmaier \cite{Lang-Schlichenmaier}.
This lead to the following local theorem we could apply to avoid cancellation.
The following is a simplified restatement of \cite{SorWen1} Theorem 4.1:
\begin{thm} \cite{SorWen1}\label{thm-contractible}\label{contractibility-no-cancellation}
If $M^m$ is an oriented Lipschitz manifold of finite volume
with integral current structure, $T$, and if there is a ball,
$B_x(r)\subset M^m$, that has $\partial T \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: B_x(r)=0$ and if
$B_x(r)$ has a uniform linear geometric contractibility function,
$\rho:[0,2r]\to [0,\infty)$, with $\rho(r)=\lambda r$, then
\begin{equation} \label{sorwen1.estimate}
||T||(B_x(s))\ge \textrm{Fillvol}_\infty(\partial (T\:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: \bar{B}_x(r)) \ge C_\lambda s^m \,\,
a.e. \,s\in [0, r/(2^{m+6}\lambda^{m+1})].
\end{equation}
\end{thm}
\begin{example}
Note that the condition here that $\partial T \:\mbox{\rule{0.1ex}{1.2ex}\rule{1.1ex}{0.1ex}}\: B_x(r)=0$ is necessary.
If $M^m$ were a thin flat strip $[0,1]\times[0,\epsilon]$
, all balls in $M^m$ would have $\rho(r)=r$,
but the volumes of the balls would be less than $2r\epsilon$.
\end{example}
This theorem combined with the ideas described in
Remark~\ref{rmrk-avoid-cancellation} leads to the the following theorem
demonstrating that the limits occurring in Greene-Petersen's compactness
theorem have no cancellation:
\begin{thm} \cite{SorWen1}\label{thm-contractibility-no-cancellation}
If a sequence of connected oriented
Lipschitz manifolds
without boundary, $M^m_j=(X_j,d_j,T_j)$ has
a uniform linear geometric contractibility function,
$\rho:[0,r_0]\to [0,\infty)$, with $\rho(r)=\lambda r$,
and a uniform upper bound on volume, then a subsequence converges in
both the intrinsic flat sense and the Gromov-Hausdorff sense to
the same space $M^m=(X,d,T)$. In particular, $M^m$ is a countably
$\mathcal{H}^m$ rectifiable metric space.
\end{thm}
A more general version of Theorem~\ref{thm-contractibility-no-cancellation}
which allows for boundaries, is stated as Corollary 1.6 in our paper \cite{SorWen1}.
\begin{rmrk}\label{rmrk-Schul-Wenger}
If the contractibility function is not linear, Schul-Wenger
have shown the limit space need not be countably $\mathcal{H}^m$ rectifiable
in the Appendix of
\cite{SorWen1}. Note that Ferry-Okun have shown that without a uniform
upper bound on volume, these sequences can converge to an infinite
dimensional space \cite{Ferry-Okun}.
\end{rmrk}
\subsection{Ricci and Scalar Curvature}\label{subsect-Ricci}
Gromov proved that a sequence of manifolds, $M^m_j$, with nonnegative Ricci curvature
and a uniform upper bound on diameter, have a subsequence which converges in
the Gromov-Hausdorff sense to a compact geodesic space, $Y$ \cite{Gromov-metric}.
Cheeger-Colding
proved that in the noncollapsed setting, where the volumes are uniformly bounded
below, the manifolds converge in the metric measure sense to $Y$ with the
Hausdorff measure, $\mathcal{H}^m$. In particular, if $p_j \in M_j$ converge to
$y\in Y$ then $\operatorname{Vol}(B_{p_j}(r))$ converges to $\mathcal{H}^m(B_y(r))$.
Furthermore $Y$ is countably
$\mathcal{H}^m$ rectifiable with Euclidean tangent cones almost everywhere.
Points with Euclidean tangent cones are called {\em regular points} and, at such
points, the density of the Hausdorff measure is $1$. In fact
$\lim_{r\to 0} \mathcal{H}^m(B_y(r))/r^m=\omega_m$. \cite{ChCo-PartI}.
Such sequences
do not have uniform geometric contractibility functions as seen by Perelman's example
in \cite{Perelman-example}. In fact Menguy proved the limit space could
have infinite topological type \cite{Menguy-inf-top-type}. Nevertheless, in \cite{SorWen1},
the authors proved that the Gromov-Hausdorff and intrinsic flat distances agree in
this setting:
\begin{thm} \label{thm-ricci} \cite{SorWen1}
If a noncollapsing
sequence of oriented Riemannian manifolds
without boundary, $M^m_j=(X_j,d_j,T_j)$ has
nonnegative Ricci curvature and a uniform upper bound on diameter,
then a subsequence converges in
both the intrinsic flat sense and the Gromov-Hausdorff sense to
the same space $M^m=(X,d,T)$.
\end{thm}
This theorem can be viewed as an example of a noncancellation theorem.
The proof is based on Theorem~\ref{thm-contractible} and the fact
that Perelman proved that balls of large volume in a manifold with
nonnegative Ricci curvature are contractible \cite{Perelman-max-vol}.
We also applied the work of Cheeger-Colding \cite{ChCo-PartI}, which
states that in this setting the volumes of balls converge and that almost
every point in the Gromov-Hausdorff limit is a regular point. Regular
points have Euclidean tangent cones and
$\lim_{r\to 0} \mathcal{H}^m(B_y(r))/r^m=\omega_m$.
\begin{rmrk}
It would be interesting if one could prove this theorem directly without
resorting to the powerful theory of Cheeger-Colding. That would give new
insight perhaps allowing one to extend this result to situations with weaker
conditions on the curvature.
\end{rmrk}
In \cite{SorWen1} we presented an example of a sequence of three
dimensional Riemannian manifolds with positive scalar curvature that
converge in the intrinsic flat sense to the $0$ integral current space.
Example~\ref{example-cancels} depicted in Figure~\ref{figure-cancels}
is a 2 dimensional version of this example. The example with positive
scalar curvature is constructed by connecting a pair of standard three
dimensional spheres by an increasingly dense collection of tunnels.
Each tunnel is constructed using Schoen-Yau or Gromov-Lawson's
method \cite{MR535700} \cite{MR569070}.
This
sequence has increasingly negative Ricci and sectional curvatures
within the tunnels but the scalar curvature remains positive.
Note that each tunnel has a minimal two sphere inside. It is natural
in the study of general relativity, to require that a manifold have positive
scalar curvature and no interior minimal surfaces. The boundary is
allowed to consist of minimal surfaces.
The following conjecture is based upon discussions with Ilmanen:
\begin{conjecture} \label{conj-scalar}
A converging sequence of three dimensional
Riemannian manifolds with positive
scalar curvature, a uniform lower bound on volume,
and no interior minimal surfaces converges without
cancellation to a nonzero integral current space.
\end{conjecture}
A solution to this
conjecture would have applications in general relativity and is essential to
solving Ilmanen's 2004 proposal that a new weak form of convergence needs
to be developed to better understand manifolds with positive scalar curvature.
\vspace{.1cm}
\subsection{Wenger's Compactness Theorem}\label{subsect-Wenger-compactness}
In \cite{Wenger-compactness}, the second author has proven the key
compactness theorem for the intrinsic flat distance:
\begin{thm}\label{thm-Wenger-compactness} {\em{\cite{Wenger-compactness} [Theorem 1.2]}}
Let $m, C, D > 0$ and let $\bar{X}_j$ be a sequence of complete metric spaces.
Given $T_j \in \intcurr_m(\bar{X}_j)$ with uniform bounds on total mass and diameter:
\begin{equation} \label{eqn-compact-1}
{\mathbf M}(T_j) +{\mathbf M}(\partial T_j) \le C
\end{equation}
and
\begin{equation} \label{eqn-compact-2}
\operatorname{diam}(\operatorname{spt}(T_j) ) \le D
\end{equation}
then
there exists a subsequence $T_{j_i}$, a complete metric space $Z$, an integral current $T \in \intcurr_m(Z)$
and isometric embeddings $\varphi_{j_i}: \bar{X}_{j_i} \to Z$ such that
\begin{equation} \label{eqn-compact-3}
d^Z_F\left(\varphi_{j_i\#}T_{j_i},T\right) \to 0.
\end{equation}
In particular, if $M_n=(X_n, d_n,T_n)$ is a sequence of integral current spaces satisfying (\ref{eqn-compact-1})
and (\ref{eqn-compact-2}), then a subsequence converges in the intrinsic flat sense to an integral
current space of the same dimension. The limit space is in fact $M=(\rm{set}(T), d_Z, T)$.
\end{thm}
In particular, sequences of oriented Riemannian manifolds with boundary
with a uniform upper bound on volume, on the volume of the boundary and
on diameter have a subsequence which converges in the intrinsic flat sense
to an integral current space. Note that even when the sequence of manifolds
is compact, the limit space need not be precompact as seen in Example~\ref{ex-many-tips}
depicted in Figure~\ref{fig-many-tips}.
We now apply this compactness theorem combined
with techniques from the proof of
Theorem~\ref{converge} to prove Theorem~\ref{Cauchy-to-Converge}.
We do not apply this compactness theorem anywhere else in this paper.
Contrast this with Theorem~\ref{cauchy} and see Remark~\ref{rmrk-cauchy}.
\begin{thm} \label{Cauchy-to-Converge}
Given an intrinisic flat Cauchy sequence of
integral current spaces,
$M^m_j=\left(X_j, d_j, T_j\right)$
with a uniform bound on total mass, $\mathbf N(M_j) \le V_0$, and a uniform bound
on diameter, $\operatorname{diam}(M_j) \le D$, there exists a complete metric space
$Z$, and a sequence
of isometric embeddings $\varphi_j: X_j \to Z$
such that $\varphi_{j\#}T_j$ is a flat Cauchy
sequence of integer rectifiable currents in $Z$
which converges in the flat sense to an integral current $T\in \intcurr_m(Z)$.
Thus $M^m_j$ converges in the intrinsic flat sense to an integral current
space $(\rm{set}(T), d_Z, T)$.
\end{thm}
\begin{proof}
First there is a subsequence
$(X_{j_i}, d_{j_i}, T_{j_i})$ which converges in the intrinsic flat sense
to an integral current space $(X,d,T)$, by Wenger's compactness theorem.
Since $(X_j,d_j,T_j)$ is
Cauchy, it also converges to $(X,d,T)$. Then Theorem~\ref{converge}
then yields the claim.
\end{proof}
\section{{\bf Lipschitz Maps and Convergence}} \label{Sect-reln}
We review Lipschitz convergence and
prove that when sequences of manifolds
converge in the Lipschitz sense, then they converge in the intrinsic flat
sense. As a consequence, sequences of manifolds which converge in
the $C^{k,\alpha}$ sense or the $C^\infty$ sense, also converge in the
intrinsic flat sense. Lemmas in this section will also be useful when proving
the examples in the final section of the paper.
\subsection{Lipschitz Maps} \label{subsect-lip-maps}
The purpose of this subsection is to list some basic properties of the intrinsic
flat norm of an integral current space.
Some of the lemmas will be used later on for the construction
of examples in the Appendix.
Others will be used to relate the Lipschitz convergence to intrinisic flat convergence
[Theorem~\ref{thm-lip-to-flat}].
Recall that a metric space $X$ is called injective if for every metric space $Y$, every subset $A\subset Y$ and every Lipschitz map $\varphi: A\to X$ there exists a Lipschitz extension $\bar{\varphi}: Y\to X$ of $\varphi$ with the same Lipschitz constant. It is not difficult to check that given a set $Z$, the Banach space $l^\infty\left(Z\right)$ of bounded functions, endowed with the supremum norm, is injective
(c.f. \cite{BenLinText} p 12-13).
Given a complete metric space $X$ and $T\in\intcurr_m\left(X\right)$ we define
\begin{equation}
\mathcal F_X\left(T\right):= \inf\left\{{\mathbf M}\left(U\right) + {\mathbf M}\left(V\right): U\in\intcurr_m\left(X\right), V\in\intcurr_{m+1}\left(X\right), T = U+\partial V\right\}
\end{equation}
whereas
\begin{equation}
\mathcal F\left(T\right):= \inf\left\{\mathcal F_Z\left(\varphi_{\#}T\right): \text{ $Z$ metric space, $\varphi: X\hookrightarrow Z$ isometric embedding}\right\}.
\end{equation}
\begin{lem}\label{lemma:injective-spaces-flatnorm} \label{lem-wen-1.1}
Given $X$ an injective metric space and $T\in\intcurr_m\left(X\right)$ we have $\mathcal F\left(T\right) = \mathcal F_X\left(T\right).$
\end{lem}
\begin{proof}
Let $Z$ be a metric space and $\varphi: X\hookrightarrow Z$ an isometric embedding. Since $X$ is injective there exists a $1$-Lipschitz extension $\psi: Z\to X$ of $\varphi^{-1}: \varphi\left(X\right) \to X$. Let $U\in\intcurr_m\left(Z\right)$ and $V\in\intcurr_{m+1}\left(Z\right)$ with $\varphi_{\#}T=U+\partial V$ and observe that $U':= \psi_{\#}U$ and $V':= \psi_{\#}V$ satisfy $T = U' + V'$ and
\begin{equation}
{\mathbf M}\left(U'\right)+{\mathbf M}\left(V'\right) \leq {\mathbf M}\left(U\right) + {\mathbf M}\left(V\right).
\end{equation}
Since $U$ and $V$ were arbitrary, it follows that $\mathcal F_X\left(T\right)\leq\mathcal F\left(T\right)$.
\end{proof}
\begin{lem}\label{lemma-Lipschitz-fillvol} \label{lem-wen-1.2}
Let $X$ and $Y$ be complete metric spaces and let $\varphi: X \to Y$ be a $\lambda$-bi-Lipschitz map
Then for each
$T\in\intcurr_m\left(X\right)$ we have
$$\mathcal F\left(T\right)\leq\lambda^{m+1}\mathcal F_Y\left(\varphi_\# T\right).$$
\end{lem}
\begin{proof}
Let $\iota: X\to l^\infty\left(X\right)$ be the Kuratowski embedding and let $\bar{\varphi}: Y\to l^\infty\left(X\right)$ be a $\lambda$-Lipschitz extension of $\iota\circ\varphi^{-1}$. Given $U\in\intcurr_m\left(Y\right)$ and $V\in\intcurr_{m+1}\left(Y\right)$ with $\varphi_\#T= U+\partial V$ then $\iota_\#T=\bar{\varphi}_{\#}U + \partial\left(\bar{\varphi}_{\#}V\right)$ and thus
\begin{equation}
\mathcal F\left(T\right)= \mathcal F_{l^\infty\left(X\right)}\left(\iota_\# T\right) \le {\mathbf M}\left(\bar{\varphi}_\#U\right) + {\mathbf M}\left(\bar{\varphi}_\#V\right)\leq \lambda^{m}{\mathbf M}\left(U\right) + \lambda^{m+1}{\mathbf M}\left(V\right).
\end{equation}
Minimizing over all $U$ and $V$ selected as above completes the proof.
\end{proof}
\begin{lem} \label{Euclidean-fillvol} \label{lem-wen-1.3}
Let $X$ be a complete metric space and $\varphi: X\to\mathbb{R}^N$ a $\lambda$-lipschitz map
where $\lambda \ge 1$. For $T\in\intcurr_m\left(X\right)$ we have
\begin{equation}
\mathcal F\left(T\right)\geq \left(\sqrt{N}\lambda\right)^{-\left(m+1\right)}\mathcal F_{\mathbb{R}^N}\left(\varphi_\#T\right).
\end{equation}
\end{lem}
We illustrate the use of the lemma by a simple example: Let $M$ be an $m$-dimensional oriented submanifold of $\mathbb{R}^N$ of finite volume and finite boundary volume. Endow $M$ with the length metric and call the so defined metric space $X$. Clearly, the inclusion $\varphi: X\to\mathbb{R}^N$ is $1$-Lipschitz. Let $T$ be the integral current in $X$ induced by integration over $M$. The above lemma thus implies
\begin{equation}
\mathcal F\left(T\right)\geq N^{-\frac{m+1}{2}}\mathcal F_{\mathbb{R}^N}\left(\Lbrack M\Rbrack\right)
\end{equation}
where $\Lbrack M\Rbrack$ is the current in $\mathbb{R}^N$ induced by integration over $M$.
\begin{proof}
\CS{Let $A=\iota\left(X\right) \subset l^\infty\left(X\right)$ where $\iota: X\to l^\infty\left(X\right)$ denotes the Kuratowski embedding.
Then $\varphi\circ \iota^{-1}: A \to \mathbb{R}^N$ is a $\lambda$-Lipschitz map.}
By McShane's extension theorem there exists a $\sqrt{N}\lambda$-Lipschitz extension $\psi: l^\infty\left(X\right)\to\mathbb{R}^N$ of $\varphi\circ\iota^{-1}: A \to \mathbb{R}^N$ \cite{McShane-34}.
Thus, if $U\in\intcurr_m\left(l^\infty\left(X\right)\right)$ and $V\in\intcurr_{m+1}\left(l^\infty\left(X\right)\right)$ are such that $\iota_\# T = U + \partial V$ then
\begin{equation}
\varphi_\#T= \psi_\# \iota_\# T= \psi_\# U + \psi_\# \left(\partial V\right)=\psi_\#U + \partial\left(\psi_\#V\right)
\end{equation}
and
\begin{equation}
\mathcal F_{\mathbb{R}^N}\left(\varphi_\#T\right)\leq {\mathbf M}\left(\psi_\#U\right) + {\mathbf M}\left(\psi_\#V\right) \leq \left(\sqrt{N} \lambda\right)^{m+1}[{\mathbf M}\left(U\right) + {\mathbf M}\left(V\right)].
\end{equation}
We now obtain the claim by minimizing over all $U$ and $V$ and using \lemref{lemma:injective-spaces-flatnorm}.
\end{proof}
\CS{ In the following lemma we bound the intrinsic flat distance between an integral
current space and its image under a bi-Lipschitz map. Recall the total mass
$\mathbf N\left(T\right)={\mathbf M}\left(T\right)+{\mathbf M}\left(\partial T\right)$ [Definition~\ref{def-total-mass}].}
\begin{lem} \label{Lip-In}
Let $X$ and $Y$ be complete metric spaces and let $\varphi: X \to Y$ be a $\lambda$-bi-Lipschitz map for some $\lambda>1$.
Then for $T\in\intcurr_m\left(X\right)$ viewed as an integral current space
$T=\left(\rm{set}\left(T\right), d_X,T\right)$ and $\varphi_\#T=\left(\rm{set}\left(\varphi_\#T\right), d_Y, \varphi_\# T\right)$
we have
\begin{equation}
d_{\mathcal{F}}\left(T, \varphi_\#T\right) \le
k_{\lambda,m}
\max\{\operatorname{diam}\left(\operatorname{spt} T\right), \operatorname{diam}\left(\varphi\left(\operatorname{spt} T\right)\right)\}\,\mathbf N\left(T\right)
\end{equation}
where $k_{\lambda,m}:=\frac{1}{2}\left(m+1\right)\lambda^{m-1}\left(\lambda - 1\right)$.
\end{lem}
\begin{proof}
Let $C_0:= \operatorname{spt} T$, $C_1:= \varphi\left(C_0\right)$, and denote by $d_0$ and $d_1$ the metric on $C_0$ and $C_1$, respectively. Let $D:= \max\{\operatorname{diam} C_0, \operatorname{diam} C_1\}$. Let $d_Z$ be the metric on $Z:= C_0\sqcup C_1$ which extends $d_0$ on $C_0$ and $d_1$ on $C_1$ and which satisfies
\begin{equation}
d_Z\left(x, x'\right)= \inf\{d_0\left(x,\bar{x}\right)+d_1\left(\varphi\left(\bar{x}\right), x'\right): \bar{x}\in C_0\} + \lambda'D,
\end{equation}
whenever $x\in C_0$ and $x'\in C_1$ and where $\lambda':= \frac{1}{2}\lambda^{-1}\left(\lambda - 1\right)$.
It is not difficult to verify that $d_Z$ is in fact a metric.
Let $\varphi_i: C_i\to l^\infty\left(Z\right)$
be the composition of the inclusion map with the Kuratowski embedding.
Note that these are isometric embeddings. Define a map
$\psi: [0,1]\times C_0\to l^\infty\left(Z\right)$ using linear interpolation:
\begin{equation} \label{psi-def-1}
\psi\left(t,x\right):= \left(1-t\right)\varphi_0\left(x\right)+t\varphi_1\left(\varphi\left(x\right)\right).
\end{equation}
It is then clear that
\begin{equation} \label{eqn-lip1}
\operatorname{Lip}\left(\psi\left(\cdot,x\right)\right)= \lambda' D \quad \forall x\in C_0
\quad \textrm{and} \quad \operatorname{Lip}\left(\psi\left(t,\cdot\right)\right)\leq\lambda \quad \forall t\in[0,1].
\end{equation}
We now apply the linear interpolation to define two currents,
\begin{equation}
\begin{split}
U & :=\psi_\#\left([0,1]\times \partial T\right)\in \intcurr_m\left(l^\infty\left(Z\right)\right)
\textrm{ and } \\
V & := \psi_{\#}\left([0,1]\times T\right) \in \intcurr_{m+1}\left(l^\infty\left(Z\right)\right),
\end{split}
\end{equation}
where the product of currents is defined as in \cite{Wenger-isoper} Defn 2.8.
By Theorem 2.9 in \cite{Wenger-isoper},
\begin{equation}
\partial\left([0,1]\times T\right) = [1]\times T - [0]\times T - [0,1]\times \partial T.
\end{equation}
So if we push forward by $\psi$ applying (\ref{psi-def-1})
we get
\begin{eqnarray*}
\partial V &=& \psi_\# \left([1]\times T\right) -\psi_\#\left( [0]\times T\right) -\psi_\#\left( [0,1]\times \partial T\right)\\
&=&\varphi_{1\#}\varphi_{0\#}T - \varphi_{0\#}T - U.
\end{eqnarray*}
Since $\varphi_0$ is an isometric embedding we have
\begin{equation}
d_{\mathcal{F}}\left(\varphi_\# T, T\right) \le d_F^Z\left(\varphi_{0\#}\varphi_\# T, \varphi_{0\#}T\right) \le
{\mathbf M}\left(U\right) + {\mathbf M}\left(V\right).
\end{equation}
By Proposition 2.10 in \cite{Wenger-isoper}, we have
\begin{equation} \label{mass-est}
{\mathbf M}\left(U\right) + {\mathbf M}\left(V\right) \le m\lambda^{m-1}\lambda' D\, {\mathbf M}\left(\partial T\right) + \left(m+1\right)\lambda^m\lambda' D\, {\mathbf M}\left(T\right).
\end{equation}
Thus we obtain the lemma.
\end{proof}
\subsection{Lipschitz and Smooth Convergence}\label{subsect-lip-conv}
Over the years various notions
of smooth convergence and compactness theorems have been proven. We recommend
Petersen's textbook \cite{Petersen-text} for a survey of these various notions of
convergence progressing from $C^{1,\alpha}$ to $C^\infty$ convergence. All these notions
involve maps $f_j:M_j\to M_\infty$ and the push forward of the metric tensors $g_j$
from $M_j$ to positive definite tensors $f_{j*}g_j$ on $M$ and then studying the
appropriate convergence of these tensors to $g$.
A weaker notion than these
notions is Gromov's Lipschitz convergence introduced in 1979
which does not require one to examine
the metric tensors but rather just the distances on the spaces \cite{Gromov-metric}[Defn 1.1 and Defn 1.3].
In this section we will briefly review Lipschitz convergence and prove that
whenever a sequence of manifolds converges in the Lipschitz sense then it
converges in the intrinsic flat sense [Theorem~\ref{thm-lip-to-flat}]. As a consequence,
$C^{1,\alpha}$ convergence and all other smooth forms of convergence are
stronger than intrinsic flat convergence as well. That is, any sequence of manifolds
converging in the smooth sense to a manifold, converges in the intrinsic flat sense
as well.
\begin{defn}[Gromov] \label{defn-lip-conv}
The Lipschitz distance between two metric spaces $X,Y$, is defined as
\begin{equation}
d_L\left(X,Y\right)= \inf \{ \,|\log \textrm{dil} \left(f\right)| + |\log \textrm{dil}\left( f^{-1} \right)|: \textrm{ bi-Lipschitz }f:X\to Y \, \}
\end{equation}
where
\begin{equation}
\textrm{dil}\left(f\right)=\sup \left\{ \frac {d\left(f\left(x\right),f\left(y\right)\right)}{d\left(x,y\right)}: \,\, x,y\in X \textrm{ s.t. } x\neq y\, \right\}.
\end{equation}
When there is no bi-Lipschitz map from $X$ to $Y$ one says $d_L\left(X,Y\right)=\infty$.
\end{defn}
Note that if a sequence of orientable Riemannian manifolds $M_j$ converges in the Lipschitz
sense to a metric space $M$, then $M$ is bi-Lipschitz to an
orientable Riemannian manifold.
In particular $M$ is an orientable Lipschitz manifold and by Remarks~\ref{rmrk-biLip-matching}
and~\ref{Lip-mani-structure}, it has a natural structure as an integral current space
determined completely by choosing an orientation on the space.
\begin{thm} \label{thm-lip-to-flat}
If $M_j$ are orientable Lipschitz manifolds converging in the Lipschitz sense
to an oriented Lipschitz manifold $M$, then after matching orientations of
the $M_j$ to the limit manifold, $M$, the oriented Lipschitz manifolds $\Lbrack M_j\Rbrack$
converge in the intrinsic flat sense to $\Lbrack M \Rbrack$.
In fact, whenever $M$ and $N$ are Lipschitz manifolds with matching orientations,
\begin{equation} \label{mani-flat-lip-1}
d_{\mathcal{F}}\left(M, N\right) < k_{\lambda,m} \max \{\operatorname{diam}\left(M\right), \operatorname{diam}\left(N\right) \} \, ( \operatorname{Vol}\left(M\right) +\operatorname{Vol}(\partial M))
\end{equation}
where $k_{\lambda,m}:=\frac{1}{2}\left(m+1\right)\lambda^{m-1}\left(\lambda - 1\right)$
and where
$\lambda=e^{d_L\left(M,N\right)}$.
\end{thm}
Gromov has proved that Lipschitz convergence implies
Gromov-Hausdorff convergence \cite{Gromov-metric}[Prop 3.7]. So that in
this setting the Gromov-Hausdorff limits and intrinsic flat limits agree.
Gromov's proof applies to any sequence of metric spaces. We cannot
extend our theorem to arbitrary integral current spaces because, in general,
one cannot just reverse orientations to match the orientations between
a pair of bi-Lipschitz homeomorphic integral current spaces.
\begin{proof}
Recall Remarks~\ref{rmrk-biLip-matching} and~\ref{Lip-mani-structure},
that when $\psi: M^m\to N^m$ is a bi-Lipschitz homeomorphism
between connected oriented Lipschitz
manifolds then $\psi_\#\Lbrack M\Rbrack = \pm\Lbrack N\Rbrack$.
Once the orientations have been fixed to match, the sign becomes positive.
Lemma~\ref{Lip-In} implies that
\begin{equation} \label{mani-flat-lip-1}
d_{\mathcal{F}}\left(M, N\right) \le \frac{1}{2}\left(m+1\right)\lambda^{m-1}\left(\lambda - 1\right) \max\{\operatorname{diam}\left(M\right), \operatorname{diam}\left(N\right)\}\,(\operatorname{Vol}\left(M\right)+ \operatorname{Vol}(\partial M))
\end{equation}
where $\lambda>1$ is the bi-Lipschitz constant for $\psi$.
Note further that
\begin{equation}
\log \lambda \le |\log dil\left(\psi\right)|+|\log dil\left(\psi^{-1}\right)|\le 2\log \lambda.
\end{equation}
Taking the infimum of this sum over all $\psi$
and applying (\ref{mani-flat-lip-1}), we see that
\begin{equation} \label{mani-flat-lip-2}
d_{\mathcal{F}}\left(M, N\right) \le k_{\lambda.m}
\max\{\operatorname{diam}\left(M\right), \operatorname{diam}\left(N\right)\}\,(\operatorname{Vol}\left(M\right)+ \operatorname{Vol}(\partial M))
\end{equation}
where $\lambda=e^{d_L\left(M,N\right)}$.
Now whenever a sequence of
Lipschitz manifolds, $M_j$, converges in the Lipschitz sense
to a Lipschitz manifold, $M$, then
\begin{equation}
\lambda_j= e^{d_L\left(M_j,M\right)} \to 1
\textrm{ and }\operatorname{diam}\left(M_j\right) \to \operatorname{diam}\left(M\right).
\end{equation}
Thus $d_{\mathcal{F}}\left(M_j, M\right)$ is less than or equal to
\begin{equation} \label{mani-flat-lip-3}
k_{\lambda_j,m} \max\left\{\operatorname{diam}\left(M_j\right), \operatorname{diam}\left(M\right)\right\}\,\left( \operatorname{Vol}\left(M\right)_{\,}^{\,}+\operatorname{Vol}(\partial M)\right)
\end{equation}
which converges to $0$ as $j\to \infty$.
\end{proof}
|
\section{Introduction}
Let $q$ be a positive integer.
The $q$-state Potts partition function of a graph~$G=(V,E)$,
with uniform interactions of strength~$\gamma\geq-1$ along the edges, is
defined as
\begin{equation}\label{eq:PottsGph}
Z_\mathrm{Potts}(G;q,\gamma) =
\sum_{\sigma:V\rightarrow [q]}
\prod_{e=\{u,v\}\in E}
\big(1+\gamma\,\delta(\sigma(u) ,\sigma(v))\big),
\end{equation}
where $[q]=\{1,\ldots,q\}$ is a set of $q$~spins or colours,
and $\delta(s,s')$ is~$1$ if $s=s'$, and 0 otherwise.
The partition function is a sum over ``configurations''~$\sigma$
which assign spins to vertices in all possible ways.
Later, we shall widen this definition to allow the interaction strength $\gamma$
to be a function of the edge~$e$, and allow $G$ to be a hypergraph (a generalisation
of graph in which edges contain an arbitrary number of vertices), but the
above restricted case is sufficient for this overview. Mostly we shall concentrate in this
paper on the ferromagnetic situation, characterised by $\gamma>0$. In the
ferromagnetic Potts model, configurations $\sigma$ with many adjacent like spins
make a greater contribution to the partition function $Z_\mathrm{Potts}(G;q,\gamma)$ than
those with few.
The statistical mechanical model just described was introduced
by Potts~\cite{Potts} and generalises the classical Ising model from
two to $q$~spins.
Definition (\ref{eq:PottsGph}) applies only when $q$ is a positive integer.
However, it transpires that, regarding $q$ as an indeterminate, (\ref{eq:PottsGph})
defines a polynomial in~$q$, and in this way we can make sense of the Potts partition
function for non-integer~$q$, even though the underlying physical model has
no meaning.
An equivalent, but more concrete way of approaching the partition function
when $q$ is non-integer
is via the Tutte polynomial, which in its ``random cluster'' formulation is
defined as follows:
\begin{equation}\label{eq:TutteGph}
\ZTutte(G;q,\gamma)=\sum_{F\subseteq E}
q^{\kappa(V,F)}\gamma^{|F|},
\end{equation}
where $\kappa(V,F)$ denotes the number of connected components in
the graph $(V,F)$.
The notation is as before, except that now $q$ is an arbitrary real number.
Again, for simplicity, we are assuming that $G$ is a usual graph (not
a hypergraph) and that the edge weight $\gamma$ is uniform over edges.
The multivariate definition can be guessed at and will in any case appear
later in the paper. For readers who are familiar with the classical
$(x,y)$-parameterisation
of the Tutte polynomial, the transformation
between that and the one here is given by $\gamma=y-1$ and $q=(x-1)(y-1)$.
Although (\ref{eq:PottsGph}) and (\ref{eq:TutteGph}) are formally very
different, they define the same polynomial in~$q$:
see Observation~\ref{obs:FK}. We continue the discussion now in terms
of the Tutte polynomial~(\ref{eq:TutteGph}), remembering all along that
we include as a special case the Potts partition function, and as an even
more special case that of the Ising model. We denote by $\textsc{Tutte}(q,\gamma)$
the computational task of computing $\ZTutte(G;q,\gamma)$ given a graph $G$ as
problem instance. Then each pair $(q,\gamma)$ defines a separate computational
problem, and we can study the computational complexity of this problem as $q$ and~$\gamma$
vary. It is important to note that $q$ and $\gamma$ do not form part of the problem instance,
which consists simply of the graph~$G$.
For the purposes of this discussion, we may assume that $q$ and $\gamma$ are rational,
in order to avoid representation issues, but in the main body of the paper we
work in the wider class of ``efficiently approximable'' real numbers.
In a seminal paper, Jaeger, Vertigan and Welsh~\cite{JVW90} examined
the problem of computing $\ZTutte(G;q,\gamma)$ exactly. In the exact setting,
they completely classified the complexity of $\textsc{Tutte}(q,\gamma)$
for all $q,\gamma$ (in fact for all complex $q,\gamma$).
It transpires that $\textsc{Tutte}(q,\gamma)$ is $\mathrm{\#P}$-hard (i.e., as hard as determining the
number of satisfying assignments to a CNF Boolean formula), except when $q=1$,
or when $(q,\gamma)$ is one of a finite number of ``special points''; in these cases
$\textsc{Tutte}(q,\gamma)$ is polynomial-time computable.
In light of Jaeger et al.'s strong negative result, attention turned to the question
of whether $\ZTutte(G;q,\gamma)$ could be approximated with arbitrarily small
specified relative error. In the context of computing partition functions,
the appropriate notion of efficient approximate computation is the ``Fully polynomial
randomised approximation scheme'' or FPRAS,
which is rigorously defined in~\S\ref{sec:FPRAS}. An early positive
result was provided by Jerrum and Sinclair~\cite{JS93}, who presented an FPRAS
for the case $q=2$ and $\gamma>0$, that is to say, for the ferromagnetic Ising model.
Sadly, no further generally applicable positive results have appeared since then,
though FPRAS's have been proposed for restricted classes of graphs, e.g., dense
or degree-bounded~\cite{dense}.
Greater progress has been made in the negative direction.
Goldberg and Jerrum~\cite{tuttepaper}
showed, under the reasonable complexity-theoretic assumption $\mathrm{RP}\not=\mathrm{NP}$,
that no FPRAS exists for $\textsc{Tutte}(q,\gamma)$ for a wide range of values for the parameters
$(q,\gamma)$. Stated informally, $\mathrm{RP}\not=\mathrm{NP}$ is the assumption that
there are problems in NP that cannot be decided by a polynomial-time randomised
algorithm. Intuitively, it is only slightly stronger than the more usual
$\mathrm{P}\not=\mathrm{NP}$ assumption.
As an indicative example of what is known,
the intractability result of~\cite{tuttepaper} covers the entire half-plane
$\gamma<-2$
except for the tractable case $q=1$ and
the case $q=2$ where the problem is equivalent to approximately
counting perfect matchings. Similar results apply when $q/\gamma<-2$.
The restriction to planar graphs was treated in a follow-up
paper~\cite{planartutte}. However none of the existing intractability results apply
to the region $q>0$ and $\gamma>0$ that concerns us here, and which
is perhaps the one of greatest physical interest.
Our goal here is to present the first evidence that $\textsc{Tutte}(q,\gamma)$
is computationally hard in the region $q>2$ and $\gamma>0$, i.e., the
region corresponding to the ferromagnetic Potts model with $q>2$ states.
We achieve this, but under a stronger complexity-theoretic assumption than
$\mathrm{RP}\not=\mathrm{NP}$. To explain this assumption, a digression into
computational complexity is required.
The complexity class $\mathrm{\#RH}\Pi_1$
of counting problems was introduced by
Dyer, Goldberg, Greenhill and Jerrum~\cite{APred} as a means
to classify a wide class of approximate counting problems that were
previously of indeterminate computational complexity. The problems in $\mathrm{\#RH}\Pi_1$
are those that
can be expressed in terms of counting the number of models of a logical formula
from a certain syntactically restricted class.
(Although the authors were not aware of it at the time, this
syntactically restricted
class had already been studied under the title ``restricted Krom SNP''~\cite{Dalmau05}.
Yet another terminological variation is to say that problems in $\mathrm{\#RH}\Pi_1$
enumerate solutions to a linear Datalog program.)
The complexity class $\mathrm{\#RH}\Pi_1$ has a completeness class (with respect to
approximation-preserving ``AP-reductions'')
which includes a wide and ever-increasing
range of natural counting problems, including:
independent sets in a bipartite graph,
downsets in a partial order,
configurations in the Widom-Rowlinson model (all~\cite{APred}),
the partition function of the
ferromagnetic Ising model with mixed external field (i.e., not
consistently favouring one or other spin)~\cite{ising},
and stable matchings~\cite{stablematchings}.
Either all of these problems admit an FPRAS (i.e., are efficiently approximable),
or none do. No FPRAS is known for any of them at the time of writing, despite
much effort having been expended on finding one.
All the problems in the completeness class mentioned above are inter-reducible
via AP-reductions, so
any
of them could be said to exemplify the completeness class.
However, mainly for historical reasons,
the particular problem $\BIS$, of counting independent sets
in a bipartite graph, tends to be taken as the exemplar of the class,
much in the same way that $\textsc{Sat}$ has
a privileged status in the theory on NP-completeness.
Our main result is
\begin{theorem}
\label{thm:main}
Suppose that
$q>2$ and $\gamma>0$ are efficiently approximable. Then
$\BIS\leq_\mathrm{AP} \textsc{Tutte}(q,\gamma)$.
\end{theorem}
Here, $\leq_\mathrm{AP}$ is the symbol for ``is AP-reducible to'', and ``efficiently
approximable'' is a concept defined in \S\ref{sec:FPRAS};
suffice it to say for now that
the rational numbers are trivially efficiently approximable.
One limitation of our result is that the
resulting inapproximability of the problem $\textsc{Tutte}(q,\gamma)$
is conditional on there being no FPRAS for $\BIS$ (and the rest of the completeness
class), rather than on the weaker assumption
$\mathrm{NP}\not=\mathrm{RP}$.
In fact, we conjecture that $\BIS$ does not admit an
FPRAS. The basis for our conjecture is empirical ---
namely
that the collection of known $\BIS$-equivalent problems is growing
and that the problem itself has survived its first decade
despite considerable efforts to find an FPRAS.
For example,
Ge and \v Stefankovi\v c~\cite{BISpoly}
recently proposed an interesting new MCMC algorithm for sampling indepedent sets
in bipartite graphs. Unfortunately, however, the relevant Markov chain mixes slowly~\cite{slow}
so even this interesting new idea does not give an FPRAS.
Despite the fact that our results are limited by a strong complexity-theoretic assumption,
we feel there are counterbalancing strengths that fully
justify this investigation. One is the range and intrinsic interest of
the problem under consideration. Whether in the guise of the Potts partition
function, or
of the Tutte plane, the computational complexity of $\textsc{Tutte}(q,\gamma)$
has received considerable attention since it was first studied by Jaeger
et al.~\cite{JVW90}: see, for example,
\cite{dense, bipartite, planar, knots}.
So it seems worth striving for a complexity classification
even under
a strong assumption such
as the one we are making.
The situation is similar to working
with the Unique Games Conjecture in the area of approximation algorithms
for optimisation problems, or employing the class PPAD in analysing
the complexity of Nash equilibria.
Futhermore, Theorem~\ref{thm:main}
has a wide range of applicability,
covering as it does the whole region $q>2$, $\gamma>0$, which, in the classical
parameterisation of the Tutte polynomial, equates to the entire upper quadrant
of the Tutte plane above the positive branch of the hyperbola
$H_2=\{(x,y):(x-1)(y-1)=2\}$. Note that the
$\BIS$-hard
region extends right to the tractable hyperbola~$H_2$.
Another potential strength of the work
is that the reduction introduces a novel technique
that may have wider applicability. The idea is conceptually simple and can
be sketched informally here. In the first step, we reduce $\BIS$ to a
hypergraph version of the Tutte polynomial. (The conventional Tutte polynomial may be
recovered as the specialisation to 2-uniform hypergraphs.) This step, if
not routine, is at least standard in its techniques. After this, we
show how to
simulate each hyperedge containing $t$~vertices by a graph gadget
with $t$ distinguished vertices or terminals.
At this point we exploit the
first order phase transition that is a feature of the so-called
random cluster model when $q>2$. The configurations
of the random cluster model on a graph~$G$ are spanning subgraphs of~$G$,
which are weighted
according to the numbers of edges and connected components they contain.
As formulated in (\ref{eq:TutteGph}), the Tutte polynomial
is the partition function of this model.
The gadget is designed and carefully tuned so that is has
two coexisting ``phases'': one in which the random cluster
configurations (spanning subgraphs) have a large connected (or
``giant'') component, and one in which they don't. We show that it is
possible to arrange for the $t$~terminals to be (with high probability)
in a single component
in one phase and in $t$~distinct components in the other.
This provides us with a bistable gadget that simulates a potentially
large hyperedge using many 2-vertex edges.
Note that AP-reductions often exploit phase transitions, playing
one class of configurations off against
another. See the examples in \cite{APred} and~\cite{KelkPhD}.
What is new here is that,
as far as we are aware, this is the first
time anyone managed to derive stronger complexity results using
the complex phase transitions that arise
in actual models studied in statistical physics. Unfortunately, the delicate
nature of the gadgets needed to exploit this kind of phase transition does
lead to significant technical complexity in our analysis.
Non-trivial phase transitions have been used
in the past to rule out certain natural Markov chain Monte Carlo (MCMC)
approaches to approximate counting problems. Indeed, the
first-order
phase
transition in the random cluster model was already exploited
by Gore and Jerrum~\cite{GJ99} to demonstrate that the Swendsen-Wang algorithm
is not always effective for the ferromagnetic Potts model.
(See also Borgs et al.~\cite{torpid} for a more thorough working out
of this idea.)
In those applications
the aim was to rule out a certain algorithmic approach, namely MCMC, whereas here
our goal is to show inherent intractability.
Finally, note that Theorem~\ref{thm:main} establishes $\BIS$-hardness of $\textsc{Tutte}(q,\gamma)$ but not $\BIS$-equivalence.
It would be very interesting to know whether there is an AP-reduction from $\textsc{Tutte}(q,\gamma)$ to $\BIS$.
Note that the complexity of approximate counting is complicated. Bordewich~\cite{bordewich} has shown that
if any problem
in~\#P
fails to have an FPRAS, then there is an infinite approximation hierarchy within~\#P.
The rest of the paper is laid out as follows. Sections~\ref{sec:TuttePotts}
and~\ref{sec:rc} introduce the models we shall be working with.
Section~\ref{sec:natural} is the technical heart of the paper.
It analyses the random cluster model on some
families of graphs, building up to the constuction of the hyperedge simulation
gadget. Sections \ref{sec:FPRAS} and~\ref{sec:APred}
cover the computational framework
we work in, while \S\ref{sec:results}--\ref{shift} present
and analyse the reductions themselves.
The final \S\ref{sec:3uniform} turns the hypergraph intractability
result on its head by focusing on the $3$-uniform case. By reduction
to the graphic case, we obtain an FPRAS for the partition function.
\section{The Tutte polynomial of a hypergraph and the Potts partition function}
\label{sec:TuttePotts}
Let $H=(\calV,\calE)$ be a hypergraph
with vertex set $\calV$ and hyperedge (multi)set~$\calE$.
Following the usual convention
for the Tutte polynomial~\cite{Sokal05},
a hypergraph is allowed to have parallel edges. This is why $\calE$
is a multiset. Also, the vertices in a particular hyperedge need not
be distinct (so each hyperedge can be viewed as a multiset).
The multivariate Tutte polynomial of $H$ is
defined as follows
$$
\ZTutte(H;q,\boldgamma)=\sum_{\mathcal{F}\subseteq\calE}
q^{\kappa(\calV,\mathcal{F})}
\prod_{f\in\mathcal{F}}
\gamma_f,
$$
where
$q$ and $\boldgamma=\{\gamma_f\}_{f\in\calE}$ are
commuting indeterminates
and
$\kappa(\calV,\mathcal{F})$ denotes the number
of connected components in the subhypergraph $(\calV,\mathcal{F})$.
(Two vertices $u$, $v$, are in the same component of $(V,\mathcal{F})$ if $u=v$, or there
is a sequence $f_1,\ldots,f_\ell\in\mathcal{F}$ of hyperedges with
$u\in f_1$, $v\in f_\ell$ and $f_i\cap f_{i+1}\not=\emptyset$ for
$1\leq i<\ell$.)
This partition function was studied (under a different name) by Grimmett~\cite{Grimmett}.
An undirected graph $G$ can be viewed as a $2$-uniform hypergraph
(a hypergraph in
which every hyperedge has size~$2$).
In this case, $\ZTutte(G;q,\boldgamma)$ coincides with the usual definition of the multivariate Tutte polynomial~\cite{Sokal05}.
Let $q$ be a positive integer. The $q$-state Potts partition function of $H$ is
defined as follows:
$$
Z_\mathrm{Potts}(H;q,\boldgamma) =
\sum_{\sigma:\calV\rightarrow [q]}
\prod_{f\in\calE}
\big(1+\gamma_f\delta(\{\sigma(v) \mid v\in f\})\big),$$
where $[q]=\{1,\ldots,q\}$ is a set of $q$~spins or colours,
and $\delta(S)$ is~$1$ if its argument is a singleton and 0 otherwise.
The partition function is a sum ranging over objects, in this case
assignments of spins to vertices, which are often referred to as ``configurations''.
The following observation is due to Fortuin and Kastelyn.
\begin{observation}
\label{obs:FK}
If $q$ is a positive integer then
$Z_\mathrm{Potts}(H;q,\boldgamma) = \ZTutte(H;q,\boldgamma)$.
\end{observation}
\begin{proof}
A proof can
be found in~\cite[Proposition 3.1]{CSS} or in~\cite{Grimmett}, but we include it here
for completeness.
The argument is a straightforward generalisation
of the standard proof for the graph case. See \cite[Theorem 2.3]{Sokal05}.
Consider
$$\sum_{\sigma: \calV\rightarrow [q]}
\sum_{\mathcal{F}\subseteq\calE}
\prod_{f\in \mathcal{F}}
\big(\gamma_f\delta(\{\sigma(v) \mid v\in f)\big).$$
Now ``integrate out'' (i.e., explicitly sum over) $\mathcal{F}$ to see that this quantity is equal to
$Z_\mathrm{Potts}(H;q,\boldgamma)$ and
integrate out $\sigma$ to see that this quantity is equal to
$\ZTutte(H;q,\boldgamma)$.
\end{proof}
The \emph{ferromagnetic} Potts model corresponds to the Potts model in the
special case in which the edge weights $\gamma_f$ are non-negative. In this case, a mono-chromatic
edge contributes more weight than an edge with multiple spins.
For a subset $\mathcal{F}\subseteq \calE$ of the hyperedges of a graph,
we use $\gamma({\mathcal{F}})$ to denote $\prod_{f\in\mathcal{F}}
\gamma_f$.
\section{The Random Cluster Model}
\label{sec:rc}
Consider a graph $G=(V,E)$.
Every edge $e\inE$ is associated with a
quantity $p(e)\in[0,1]$.
Then for a set of edges $A\subseteqE$ define
\begin{equation*}
\widetilde{P}(G;A,q,p) = q^{\kappa(V,A)}
\prod_{e\in A} p(e)
\prod_{e \in E\setminus A} (1-p(e)).
\end{equation*}
Let
\begin{equation*}
Z_\mathrm{rc}(G;q,p) = \sum_{A\subseteq E} \widetilde{P}(G;A,q,p)
= \ZTutte(G;q,\boldgamma)\prod_{e\in E}(1-p(e)),
\end{equation*}
where $\gamma_e = p(e)/(1-p(e))$.
Then the probability of edge-set~$A$ in the random cluster model is given by
\begin{equation*}
P(G;A,q,p) = \frac{\widetilde{P}(G;A,q,p)}
{Z_\mathrm{rc}(G;q,p)}.
\end{equation*}
The \emph{random cluster model} refers to the distribution
$\mathrm{RC}(G;q,p)$, in which
a subset $A$ of edges is chosen
with probability
$P(G;A,q,p)$.
The difference between $\ZTutte$ and $Z_\mathrm{rc}$ is simply one of parameterisation.
Nevertheless, the change of parameter is useful, as it allows us to employ
probabilistic terminology and exploit existing results from the random graph
literature.
A central ingredient which helps us to understand the random cluster model is to
compare it to the (multivariate) Erd\H os-R\'enyi model of a random graph.
Consider a graph $G=(V,E)$.
Every edge $e\inE$ is associated with a
quantity $p'(e)\in[0,1]$.
In
the Erd\H os-R\'enyi model $\mathrm{ER}(G;p')$, a subset $A$
of edges
is chosen with probability
$$\prod_{e \in A} p'(e)
\prod_{e \in E \setminusA} (1-p'(e)).$$
Thus, to choose a configuration $A$ in the $\mathrm{ER}(G;p')$ model, each edge $e\in E$ is included
in~$A$ independently with probability $p'(e)$.
\subsection{Stochastic domination results}
We will use several simple
stochastic domination results,
which are close in spirit, and indeed in their proofs, to results of
Holley~\cite{Holley}; see, in particular, Theorem~(6) of that article
and its proof.
If $A^+$ and $A^-$ are disjoint subsets of $E$,
let $\mathrm{RC}(G;q,p;A^+,A^-)$
be the random cluster model
conditioned on the fact that the chosen subset $A$
contains every edge in $A^+$ and no edges in
$A^-$.
(To avoid trivialities, we assume that no edge~$e\inA^+$
has $p(e)=0$ and that no edge $e\in A^-$ has $p(e)=1$.)
Similarly, let
$\mathrm{ER}(G;p';A^+,A^-)$
be the
Erd\H os-R\'enyi model
with this conditioning.
We give conditioned versions of the stochastic domination results.
\begin{lemma}
Consider a graph $G=(V,E)$
in which each edge $e\inE$ is associated with a
quantity $p(e)\in[0,1]$.
Let $A^+$ and $A^-$ be disjoint subsets of $E$
such that $p(e)>0$ for $e\inA^+$
and $p(e)<1$ for $e\inA^-$.
Suppose $q\geq 1$. Then
the random cluster model
$\mathrm{RC}(G;q,p;A^+,A^-)$
is stochastically dominated by the Erd\H os-R\'enyi model
$\mathrm{ER}(G;p;A^+,A^-)$
in the sense that we can select a pair
$(A,A')$
such that $A$ is drawn from
$\mathrm{RC}(G;q,p;A^+,A^-)$
and $A'$ is drawn from
$\mathrm{ER}(G;p;A^+,A^-)$
and $A \subseteq A'$.
\label{lem:coupleone}
\end{lemma}
\begin{proof}
Let $\mathcal{M}$ be the ``heat-bath on an edge'' Markov chain
for $\mathrm{RC}(G;q,p;A^+,A^-)$
and let
$\mathcal{M}'$ be the ``heat-bath on an edge'' Markov chain
for $\mathrm{ER}(G;p;A^+,A^-)$.
Start $\mathcal{M}$ in state $A_0$
and $\mathcal{M}'$ in state $A'_0$
where $A_0 \subseteq A'_0$
and $A_0$ and $A'_0$ contain every edge in $A^+$
and no edges in $A^-$. We can then couple the evolution
of the chains from the $i$'th pair of
states $(A_i,A'_i)$ as follows, guaranteeing
that $A_{i+1} \subseteq A'_{i+1}$. Choose the
same edge~$e\inE \setminus A^+\cup A^-$ in both chains.
In $\mathcal{M}$ the probability of
putting $e$ in $A_{i+1}$ is
either
$p(e)$
or
$p(e)/(p(e)+ q(1-p(e)))$,
each of which is
at most
$p(e)$,
which is the probability of putting it in
$A_{i+1}'$.
\end{proof}
\begin{lemma}
Consider a graph $G=(V,E)$
in which each edge $e\inE$ is associated with a
quantity $p(e)\in[0,1]$.
For each edge $e \inE$, let $p'(e)=p(e)/q$.
Let $A^+$ and $A^-$ be disjoint subsets of $E$
such that $p(e)>0$ for $e\inA^+$
and $p(e)<1$ for $e\inA^-$.
Suppose $q\geq 1$. Then
the Erd\H os-R\'enyi model
$\mathrm{ER}(G;p';A^+,A^-)$
is stochastically dominated by
the random cluster model
$\mathrm{RC}(G;q,p;A^+,A^-)$
in the sense that we can select a pair
$(A,A')$
such that $A$ is drawn from
$\mathrm{ER}(G;p';A^+,A^-)$
and $A'$ is drawn from
$\mathrm{RC}(G;q,p;A^+,A^-)$
and $A \subseteq A'$.
\label{lem:coupletwo}
\end{lemma}
\begin{proof}
Let $\mathcal{M}$ be the ``heat-bath on an edge'' Markov chain
for
$\mathrm{ER}(G;p';A^+,A^-)$
and let
$\mathcal{M}'$ be the ``heat-bath on an edge'' Markov chain
for
$\mathrm{RC}(G;q,p;A^+,A^-)$
Start $\mathcal{M}$ in state $A_0$
and $\mathcal{M}'$ in state $A'_0$
where $A_0 \subseteq A'_0$
and $A_0$ and $A'_0$ contain every edge in $A^+$
and no edges in $A^-$. We can then couple the evolution
of the chains from the $i$'th pair of
states $(A_i,A'_i)$ as follows, guaranteeing
that $A_{i+1} \subseteq A'_{i+1}$. Choose the
same edge~$e\inE \setminus A^+\cup A^-$ in both chains.
In $\mathcal{M}'$ the probability of
putting $e$ in $A'_{i+1}$ is
either
$p(e)$
or
$p(e)/(p(e)+ q(1-p(e)))$,
each of which is
at least
$p'(e)$,
which is the probability of putting it in
$A_{i+1}$.
\end{proof}
\begin{lemma}
\label{lem:coupleRC}
Consider a graph $G=(V,E)$
in which each edge $e\inE$ is associated with a
quantity $p(e)\in[0,1]$.
Suppose that, for each edge $e \inE$, $p'(e)\geqp(e)$.
Let $A^+$ and $A^-$ be disjoint subsets of $E$
such that $p(e)>0$ for $e\inA^+$
and $p(e)<1$ for $e\inA^-$.
Suppose $q\geq 1$. Then
the random cluster model
$\mathrm{RC}(G;q,p;A^+,A^-)$
is stochastically dominated by
the random cluster model
$\mathrm{RC}(G;q,p';A^+,A^-)$
in the sense that we can select a pair
$(A,A')$
such that $A$ is drawn from
the distribution
$\mathrm{RC}(G;q,p;A^+,A^-)$
and $A'$ is drawn from
$\mathrm{RC}(G;q,p';A^+,A^-)$
and $A \subseteq A'$.
\end{lemma}
\begin{proof}
Let $\mathcal{M}$ be the ``heat-bath on an edge'' Markov chain
for $\mathrm{RC}(G;q,p;A^+,A^-)$
and let
$\mathcal{M}'$ be the ``heat-bath on an edge'' Markov chain
for $\mathrm{RC}(G;q,p';A^+,A^-)$
Start $\mathcal{M}$ in state $A_0$
and $\mathcal{M}'$ in state $A'_0$
where $A_0 \subseteq A'_0$
and $A_0$ and $A'_0$ contain every edge in $A^+$
and no edges in $A^-$. We can then couple the evolution
of the chains from the $i$'th pair of
states $(A_i,A'_i)$ as follows, guaranteeing
that $A_{i+1} \subseteq A'_{i+1}$. Choose the
same edge~$e\inE \setminus A^+\cup A^-$ in both chains.
If the endpoints of the edge are in the same component of $A_i-e$
then they are in the same component of $A'_i-e$.
The probability of putting $e$ in $A_{i+1}$ is
$p(e)$
and the probability of putting $e$ in $A'_{i+1}$ is
$p'(e)$,
which is at least as big.
Otherwise, the probability of putting $e$ in $A_{i+1}$ is
$p(e)/(p(e)+ q(1-p(e)))$.
This is at most
$p'(e)/(p'(e)+ q(1-p'(e)))$
and the probability of putting $e$ in
$A'_{i+1}$ is at least
$p'(e)/(p'(e)+ q(1-p'(e)))$.
\end{proof}
\subsection{The Fundamental Lemma of Bollob\'as, Grimmett and Jansen}
Another way to compare the random cluster model to the
Erd\H os-R\'enyi model is to use the following lemma, which is
a multivariate version of Bollob\'as, Grimmett and Jansen's
``Fundamental Lemma'', \cite[Lemma~3.1]{BGJ}.
To do the comparison, the lemma should be applied with $r=1/q$
so that the distribution $\mathrm{RC}(\induced{G}{V_1} ; r q,p)$ in the statement of the lemma below
is identical to the distribution $\mathrm{ER}(\induced{G}{V_1} ;p)$.
Let $0\leq r \leq 1$ be fixed.
Given a subset $A$ of edges chosen from
$\mathrm{RC}(G;q,p)$, colour each component of
$(V,A)$ either \emph{red}, with probability~$r$,
or \emph{green}, with probability~$1-r$: different components are coloured independently of one another. The union of the red components is the \emph{red subgraph}
and the union of the green components is the \emph{green subgraph}.
Let $R$ be the set of vertices in the red subgraph.
Given a subset $U$ of vertices, let $\induced{G}{U}$
be the subgraph of~$G$ induced by~$U$.
\begin{lemma}
\label{lem:FL} (Bollob\'as, Grimmett and Jansen~\cite[Lemma~3.1]{BGJ}.)
Let $V_1 \subseteq V$.
Conditioned on $R=V_1$, the red subgraph is distributed
according to $\mathrm{RC}(\induced{G}{V_1}; r q,p)$
and the green subgraph is distributed according
to $\mathrm{RC}(\induced{G}{V\setminus {V_1}};(1-r)q,p)$. Conditioned on $R=V_1$, the red subgraph and green subgraph are independent of each other.
\end{lemma}
\begin{proof}
Let $V_2 = V\setminus V_1$.
Let $E_1 = E \cap V_1^{(2)}$
and $E_2 = E \cap V_2^{(2)}$,
where the notation $V_1^{(2)}$ denotes the set of unordered pairs
of vertices from~$V_1$.
Now consider $A_1 \subseteq E_1$
and $A_2 \subseteq E_2$.
The (unconditional)
probability that the red subgraph is $(V_1,A_1)$
and the green subgraph is $(V_2,A_2)$
is
$$
\frac{\widetilde{P}(G;A_1\cupA_2,q,p)\,
r^{\kappa(V_1,A_1)}
{(1-r)}^{\kappa(V_2,A_2)}}
{Z_\mathrm{rc}(G;q,p)}
$$
so the conditional probability, conditioned on $R=V_1$
is
$$
\frac{\widetilde{P}(G;A_1\cupA_2,q,p)\,
r^{\kappa(V_1,A_1)}
{(1-r)}^{\kappa(V_2,A_2)}}
{
\sum_{A'_1\subseteq E_1,A'_2\subseteq E_2}
{\widetilde{P}(G;A'_1\cupA'_2,q,p)\,
r^{\kappa(V_1,A'_1)}
{(1-r)}^{\kappa(V_2,A'_2)}}
}$$
which is
$$ P(\induced{G}{V_1}; A_1,r q,p)
P(\induced{G}{V\setminus {V_1}};A_2,(1-r)q,p).
$$
\end{proof}
\section{The random cluster model on some natural graphs}
\label{sec:natural}
In this section we consider the random cluster model on a clique, and also on
a pair of connected cliques. The latter is used as a gadget in our constructions.
First, however, we need a technical lemma,
\subsection{A technical lemma}
The following lemma is not about the random cluster model, but we will use
it in our analysis of the random cluster model on a pair of connected cliques.
\begin{lemma}\label{lem:yellowblue}
Suppose we have a partition of the set $[\nu]$ into $s$~blocks
of size at most $\nu_{\max}$.
Randomly colour a subset of the elements of $[\nu]$
yellow according to a Bernoulli process with success probability $\sprob$.
Independently, randomly colour a subset of the elements of $[\nu]$ blue
according to a Bernoulli process with success probability $\sprob$ (so an element
can be coloured yellow, blue, both or neither).
We say that a block of the partition is {\em bicoloured} if it
contains both yellow and blue elements. Then
$$
\Pr(\textup{no block is bicoloured})\leq
[(1-\sprob)^{\nu/s}(2-(1-\sprob)^{\nu/s})]^s
$$
(which is increasing with $s$ and decreasing with $\nu$) and
$$
\Pr(\textup{some block is bicoloured})
\leq
\nu[1-(1-\sprob)^{\nu_{\max}}]^2
$$
(which is increasing in both $\nu$ and $\nu_{\max}$).
\end{lemma}
\begin{proof}
Let the block sizes be $\nu_1,\ldots,\nu_s$, so that
$\max_j\nu_j \leq \nu_{\max}$.
Observe that
$$
\Pr(\text{$j$th block contains a yellow element})=1-(1-\sprob)^{\nu_j}
=\sprob_j\text{ (say)},
$$
and that the same bound applies, of course, to blue elements. Thus,
using the convention that the index $j$ always ranges over $1\leq j\leq s$,
$$\Pr(\text{some block is bicoloured})\leq
\sum_j\sprob_j^2\leq\sum_j[1-(1-\sprob)^{\nu_{\max}}]^2,$$
which gives the second bound. Clearly, this bound is monotonically
increasing in $\nu$ and~$\nu_{\max}$ as claimed.
Now for the first bound.
\begin{align}
\Pr(\text{no block is bicoloured})&=\prod_j(1-\sprob_j^2)
=\prod_j(1-\sprob_j)\prod_j(1+\sprob_j)\notag\\
&=\prod_j(1-\sprob)^{\nu_j}\prod_j(1+\sprob_j)\notag\\
&=(1-\sprob)^\nu\prod_j(1+\sprob_j).\label{eq;opt2}
\end{align}
To get an upper bound we are interested in evaluating the supremum of
\begin{equation}\label{eq:opt1}
\prod_j(1+\sprob_j)=\prod_j[2-(1-\sprob)^{\nu_j}]
\end{equation}
over the domain defined by the linear inequalities
$0\leq\nu_j\leq\nu$ and $\sum_j\nu_j=\nu$.
We consider this as an optimisation problem over~$\mathbb{R}^s$ even though
the $\nu_j$ are all integers; of course, this will if anything only
increase the supremum. We are considering a continuous function
over a closed, bounded set, so the supremum is achieved at some
point; we claim that this (unique) point is
$\nu_1=\cdots=\nu_s=\nu/s$. For if not, then at least one pair,
say $\nu_1$ and $\nu_2$ would be unequal. But then it is easily checked
that replacing $\nu_1$ and $\nu_2$ by their average would increase
$[2-(1-\sprob)^{\nu_1}][2-(1-\sprob)^{\nu_2}]$, and hence increase
the right hand side of (\ref{eq:opt1}): a contradiction.
Substituting $\nu_j=\nu/s$ into equations (\ref{eq:opt1})
and then~(\ref{eq;opt2}), we obtain
\begin{align}
\Pr(\text{no block is bicoloured})
&\leq (1-\sprob)^\nu[2-(1-\sprob)^{\nu/s}]^s\label{eq:ub1}\\
&=[(1-\sprob)^{\nu/s}(2-(1-\sprob)^{\nu/s})]^s,\label{eq:ub2}
\end{align}
as desired.
In only remains to verify the monotonicity claims about (\ref{eq:ub2}).
Let $u(\nu)=(1-\sprob)^{\nu/s}\in(0,1]$. Regarding $s>0$ as fixed,
$u(\nu)$ decreases monotonically with $\nu$.
Also, $[u(2-u)]^s$ increases monotonically as a function of $u$ in the
range $(0,1)$. Thus, expression (\ref{eq:ub2}) decreases monotonically
with $\nu$.
Now regard $\nu$ as fixed and make the change of variable $s=ax$,
where $a=\nu\ln((1-\sprob)^{-1})$, and note that $x>0$.
Then expression~(\ref{eq:ub1})
becomes $[e^{-1}(2-e^{-1/x})^x]^a$. Thus, it is enough to show that
$f(x)=x\ln(2-e^{-1/x})$ increases monotonically with~$x$.
Now
$$f''(x)=-2x^{-3}e^{-1/x}(2-e^{-1/x})^{-2}<0,$$
and
$$f'(x)=\ln(2-e^{-1/x})-x^{-1}e^{-1/x}(2-e^{-1/x})^{-1}\to0\quad
\text{as $x\to\infty$}.
$$
These two facts imply $f'(x)>0$ for $x>0$.
\end{proof}
\subsection{The random cluster model on a clique}
\label{sec:clique}
Bollob\'as, Grimmett and Jansen \cite{BGJ} studied the random cluster model on the complete $N$-vertex graph~$K_N$.
More detailed analyses have since been performed, for example
by Luczak and \L uczak \cite{LuczakLuczak}, but the approach of the earlier paper is
easier to adapt to our needs.
For fixed~$q$ and a fixed constant~$\lambda$,
they studied the distribution $\mathrm{RC}(K_N,q,p)$ where $p$ is the constant function which
assigns every edge $e$ of $K_N$ the value $p(e)=\lambda/N$.
They show that there is a critical value~$\lambda_c$, depending on~$q$,
so that, if $\lambda>\lambda_c$ then, as $N\rightarrow \infty$, with high probability
a configuration~$A$ drawn from $\mathrm{RC}(K_N,q,p)$
will have a large component (of size linear in~$N$) and otherwise, with high probability
the largest component will be much smaller (of size logarithmic in~$N$).
For $q>2$, the critical value $\lambda_c$ is defined as follows.
$$\lambda_c = 2\left(\frac{q-1}{q-2}\right)\ln(q-1).$$
It is important for our analysis that $\lambda_c < q$ (see \cite[p.16]{BGJ}, or
by calculus).
We define $\delta = (q-\lambda_c)/2>0$
and $\lambda = \lambda_c + \delta$.
Let $\theta=(q-2)/(q-1)$.
We will use the following lemma, which follows from
Theorem~2.2 and Equation~(5) of~\cite{BGJ}.
\begin{lemma}
\label{lem:clique}
Fix $q>2$ and define $\lambda$ and $\theta$ as above. Let $p$ be the constant function which assigns
every edge $e$ of $K_N$ the value $p(e)=\lambda/N$.
Let $A$ be drawn from $\mathrm{RC}(K_N,q,p)$. The probability that $A$ has a connected component of size
at least $\theta N$ tends to~$1$ as $N\rightarrow \infty$.
\end{lemma}
\subsection{The random cluster model on a pair of connected cliques}
\label{sec:gadget}
Let $\varGamma$
be the
complete graph with vertex set $\gadgetvertices = K\cup T$.
Let $\gadgetedges$ denote the edge set of~$\varGamma$ and let
$N = |K|$ and $\terms = |T|$.
Let $K^{(2)}$ denote the set of unordered pairs of distinct elements in~$K$
and define $T^{(2)}$ similarly.
Let $\critprob$ be a value in $[0,1]$.
Define $p$ as follows.
$$p(e) = \begin{cases}
\critprob, &\text{if $e\in K^{(2)}$,}\\
N^{-3/4}, &\text{if $e\in K\times T$, and}\\
1, &\text{if $e\in T^{(2)}$.}
\end{cases}
$$
Ultimately, we will use the graph~$\varGamma$
(or, more precisely, $\varGamma$ with the edges $T^{(2)}$ deleted) as a
gadget to simulate the contribution of a hyperedge on the set~$T$ to the multivariate
Tutte polynomial of~$\varGamma$.
Thus, we refer to vertices in~$T$ as ``terminals'' of the graph~$\varGamma$.
For a subset $A\subseteq \gadgetedges$, let $Y(A)$ be the number of connected components in the graph
$(\gadgetvertices,A \setminus T^{(2)})$
that contain
terminals.
A remark about the gadget~$\varGamma$ and its eventual use.
When we come to use the gadget, the edges in $T^{(2)}$ will not be
present. It is for this reason that we are interested in the structure of
connected components in $\varGamma$ in the absence of these edges,
and specifically the random variable $Y(A)$. However, it turns out that
the key properties of the gadget are easier to verify if we work with
a random cluster distribution associated with $\varGamma$, exactly
as given above, with the edges $T^{(2)}$ present. Informally, the appropriate
``boundary condition'' for the gadget is the one in which the terminals are
joined with probability~1.
The following two lemmas establish some useful properties of the gadget.
The second shows that distribution of $Y(A)$ is concentrated at the extremes
of its range, i.e., $Y=1$ or $Y=\terms$. This concentration property holds for a wide
range of values for the edge probability~$\critprob$. The first lemma, which is easier, shows that we
can tune~$\critprob$ so that the balance of probability between those extremes is
the desired value~$\gamma$. Later, in Lemmas \ref{lem:computeZ} and~\ref{lem:computerho},
we shall show that a sufficiently
close approximation to this~$\critprob$ can be efficiently computed.
\begin{lemma}\label{lem:excludedmiddleone}
Fix $q>2$ and let $\lambda = \lambda_c + (q-\lambda_c)/2$.
Fix a weight $\gamma>0$ and let $N_0$ be a sufficiently
large quantity depending on $q$ and $\gamma$.
Suppose a number of terminals $\terms>1$ is given
and fix
$N\geq\max\{\terms^{16}, N_0\}$.
Then there is a parameter~$\critprob$
satisfying $N^{-3} \leq \critprob \leq
\lambda/N\leq \tfrac14$
such that, if $A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$ then
\begin{equation}
\Pr(Y(A)=1)=\gamma\Pr(Y(A)=\terms)\label{eq:balance}.
\end{equation}
\end{lemma}
\begin{proof}
Define $\theta$ (depending on~$q$) as in Section~\ref{sec:clique}.
Let $\psi(\critprob) = \Pr(Y(A)=\terms)/\Pr(Y(A)=1)$.
We will use stochastic domination to show
\begin{itemize}
\item $\psi(\critprob)$ is monotonically decreasing as a function of $\critprob$,
\item $\psi(N^{-3})> 1/\gamma$, and
\item $\psi(\lambda/N)< 1/\gamma$.
\end{itemize}
Since $\psi(\critprob)$ is a rational function in~$\critprob$ and the denominator is never zero, we conclude that
$\psi(\critprob)$ is continuous in~$\critprob$ and there is a value $\critprob\in(N^{-3},\lambda/N)$
such that $\psi(\critprob)=1/\gamma$. This gives~(\ref{eq:balance}).
Note that the lower bound $N^{-3}$ for~$\critprob$ is very
crude, but this is all that we will need.
First, to show that $\psi(\critprob)$ is monotonically decreasing as a function of $\critprob$,
note from
Lemma~\ref{lem:coupleRC}
that $\Pr(Y(A)=\terms)$ is monotonically decreasing in~$\critprob$ and
$\Pr(Y(A)=1)$ is monotonically increasing in~$\critprob$.
Next, to show that $\psi(N^{-3})> 1/\gamma$, we will assume $\critprob=N^{-3}$
(and that $N$ is sufficiently large)
and we will show
$\Pr(Y(A)=\terms)> 1/(1+\gamma)$, which suffices.
By Lemma~\ref{lem:coupleone},
$\Pr(Y(A)=\terms)\geq \Pr(Y(\overline\es)=\terms)$ where
$\overline\es$ is drawn from $\mathrm{ER}(\varGamma;p)$.
Now the probability that
$\overline\es \cap K^{(2)}=\emptyset$ is at least $1-\binom{N}{2}
N^{-3}$. In this case,
the probability that a particular
pair of terminals is connected in~$\overline\es \setminus T^{(2)}$
is at most $N\times N^{-3/4}\times N^{-3/4}=N^{-1/2}$,
and the probability that there exists a connected pair is at most
$\terms^2N^{-1/2}\leq N^{-3/8}$.
So $\Pr(Y(\overline\es)=\terms) \geq1
- \binom{N}{2}
N^{-3}
-N^{-3/8}>1/(1+\gamma)$.
To finish the proof of~(\ref{eq:balance}),
we will show that $\psi(\lambda/N)< 1/\gamma$. To do this, we will assume $\critprob=\lambda/N$
(and that $N$ is sufficiently large)
and we will show $\Pr(Y(A)=1)> \gamma/(1+\gamma)$, which suffices.
We will again use stochastic domination to compare the random cluster model to the
Erd\H os-R\'enyi model, but this time we need some conditioning.
First, we will show that the probability that the graph
$(K,A\cap K^{(2)})$ has a connected component of size at least $\theta N$
tends to~$1$ as $N \rightarrow\infty$.
To do this, let $A^*$ be drawn from the distribution $\mathrm{RC}(\Gamma,q,\hatp)$
where
$$\hatp(e) = \left\{
\begin{array}{cc}
\critprob, &\mbox{if $e\in K^{(2)}$,}\\
0, &\mbox{otherwise.}
\end{array}\right.
$$
Since $\hatp(e)\leq p(e)$, Lemma~\ref{lem:coupleRC}
guarantees that the probability in question
is at least the probability that the graph
$(K,A^*)$ has a connected component of size at least $\theta N$.
By Lemma~\ref{lem:clique}, this probability tends to~$1$ as $N\rightarrow \infty$.
Next, we will consider
the generation of configuration~$A$ from the distribution
$\mathrm{RC}(\varGamma;q,p)$ as follows. First, we will select a set $A^+\subseteq K^{(2)}$ from
the appropriate induced distribution.
Letting
$A^-=K^{(2)}\setminus A^+$, we will select $A$ from the distribution
$\mathrm{RC}(\varGamma;q,p;A^+,A^-)$.
We will finish by showing that, as long as $(K,A^+)$ has a connected component of size
at least $\theta N$,
$\Pr(Y(A)=1 \mid A^+\subseteq A, A\cap A^-=\emptyset)$
is greater than $\gamma/(1+\gamma)$.
To do this, let $p'(e) = p(e)/q$ and let $\underline\es$ be a random variable drawn from
$\mathrm{ER}(\varGamma;p';A^+,A^-)$.
By Lemma~\ref{lem:coupletwo},
$\Pr(Y(A)=1 \mid A^+\subseteq A, A\cap A^-=\emptyset)\geq \Pr(Y(\underline\es)=1)$.
Now in $\underline\es$,
the probability that a particular terminal is not
connected to the large component is at most
$(1-q^{-1}N^{-3/4})^{\theta N}
\leq\exp(-\theta q^{-1}N^{1/4})$,
and the probability that there exists a terminal that is not connected to this component is at most
$t\exp(-\theta q^{-1}N^{1/4})$.
So $\Pr(Y(\underline\es)=1)>\gamma/(1+\gamma)$.
\end{proof}
\begin{lemma}\label{lem:excludedmiddletwo}
Fix $q>2$ and
let $\lambda = \lambda_c + (q-\lambda_c)/2$.
Fix a weight $\gamma>0$ and let $N_0$ be a sufficiently
large quantity depending on $q$ and $\gamma$.
Suppose a number of terminals $\terms>1$ and a tolerance $0<\eta<1$
are given and fix
$N\geq\max\{\terms^{16},\eta^{-1/8},N_0\}$.
For every value of~$\critprob$
in the range $[N^{-3},
\lambda/N
]$,
if $A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$ then
\begin{equation}\label{eq:dichotomy}
\Pr(1<Y(A)<\terms)<\eta.
\end{equation}
\end{lemma}
\begin{proof}
Let $C_1,C_2,\ldots$ be the connected
components of $A \cap K^{(2)}$, ordered in non-increasing size.
We are going to be relying on the phase transition of the
random cluster model.
The main fact that we will use is that $|C_1|$ is likely to
either be very small (around order $\log(N)$)
or very large (a constant fraction of $N$).
Thus, it is unlikely that $|C_1|$ is close to $N^{1/8}$.
We will not need much detail about the phase transition.
We will show
\begin{equation}
\Pr\big((|C_1|\leq N^{1/8})\wedge(Y(A)<\terms)\big)\leq \eta/2,
\label{eq:first}
\end{equation}
and
\begin{equation}
\label{eq:second}
\Pr\big((|C_1|> N^{1/8})\wedge(Y(A)>1)\big) \leq \eta/2.
\end{equation}
(Actually, the event mentioned in (\ref{eq:second})
holds with all but exponentially small
probability.) The required inequality~(\ref{eq:dichotomy})
follows directly from (\ref{eq:first}) and~(\ref{eq:second}).
Inequality~(\ref{eq:first}) is easier. To generate
a configuration~$A$ from
$\mathrm{RC}(\varGamma;q,p)$ we first select a set $A^+\subseteq K^{(2)}$ from
the appropriate induced distribution. Then,
letting
$A^-=K^{(2)}\setminus A^+$, we select $A$ from the distribution
$\mathrm{RC}(\varGamma;q,p;A^+,A^-)$.
The size of $C_1$ is entirely determined by~$A^+$.
If $A^+$ has a component of size greater than $N^{1/8}$
then
$$\Pr((|C_1|\leq N^{1/8})\wedge(Y(A)<\terms) \mid
A^+\subseteq A, A\cap A^-=\emptyset
)
=0.$$
Otherwise, this probability is equal to
$ \Pr(Y(A) < \terms \mid
A^+\subseteq A, A\cap A^-=\emptyset
)$.
By Lemma~\ref{lem:coupleone}, this is at most
$\Pr(Y(\overline\es) < \terms)$
where $\overline\es$ is generated from
$\mathrm{ER}(\varGamma;p;A^+,A^-)$.
So now consider $\overline\es$.
Fix attention on two terminals,
and colour the vertices in $K$ that are adjacent to the first terminal yellow,
and those adjacent to the second terminal blue. We are in the situation
of Lemma~\ref{lem:yellowblue}
with $\nu=N$, $\nu_{\max}\leqN^{1/8}$
and $\sprob=N^{-3/4}$.
Thus the probability that there exists a bicoloured block
(i.e., that the terminals are connected via some connected component~$C_j$)
is at most
$N[1-(1-N^{-3/4})^{N^{1/8}}]^2
\leq N[1-1+N^{-5/8}]^2=N^{-1/4}$.
Thus the probability that there exists a pair of connected terminals is
at most $\binom{\terms}{2}N^{-1/4}\leqN^{-1/8}/2<\eta/2$.
Note that the event ``there exists a pair of connected terminals'' is
the same as the event $Y(\overline\es)<\terms$.
At a high level, the path to establishing inequality~(\ref{eq:second})
is as follows. We define events $\mathcal{L}$ (``large component'')
and $\mathcal{W}$ (``weighty components'') and establish
\begin{align}
\Pr\big((|C_1|>N^{1/8})\wedge \neg\mathcal{L}\big)&<\eta/6,\label{eq:bd1}\\
\Pr(\mathcal{L}\wedge\neg\mathcal{W})&<\eta/6,\quad\text{and}\label{eq:bd2}\\
\Pr\big(\mathcal{W}\wedge(Y(A)>1)\big)&<\eta/6.\label{eq:bd3}
\end{align}
Then the required inequality~(\ref{eq:second})
follows by elementary algebra of sets (events).
We will say that an event holds ``with high probability''
(abbreviated whp)
if the probability that it holds is $o(1/Q(n))$ for
any fixed polynomial~$Q$.
First, we prove inequality~(\ref{eq:bd1}).
Let $S_1,S_2,\ldots$ be the connected components of $A$
and let $\widehat C_1,\widehat C_2,\ldots$ be the sets
$S_1 \cap K^{(2)},S_2 \cap K^{(2)},\ldots$, ordered in non-increasing size.
(Thus, $\{\widehat C_j\}$
is a coarsening of $\{C_j\}$.) Event $\mathcal{L}$ is
$|\widehat C_1|>N^{15/16}$.
Let $\mathcal{E}_1$ be the event
$(|C_1|>N^{1/8})\wedge \neg\mathcal{L}$.
We are interested showing that the
event $\mathcal{E}_1$ is unlikely.
Construct the red subgraph and green subgraph
of $(\gadgetvertices,A)$
as in Lemma~\ref{lem:FL}
with $r=1/q$.
Let $\mathcal{E}_2$ be the event
$$(|R\cap K|\leq N/q+ N^{15/16} )\wedge (C_1\subseteq R) \wedge (|C_1|>N^{1/8})
.$$
We will show
$$\Pr(\mathcal{E}_2)\geq\Pr(\mathcal{E}_1)/q^2$$
and
$$\Pr(\mathcal{E}_2) \leq \eta/(6 q^2),$$
which proves~(\ref{eq:bd1}).
The first of these follows from $\Pr(\mathcal{E}_2 \mid \mathcal{E}_1)\geq q^{-2}$,
which follows since
$\Pr(C_1\subseteq R \mid \mathcal{E}_1) = q^{-1}$ (by the definition of the red subgraph)
and
$\Pr\bigl(|R \cap K|\leq N/q+ N^{15/16}\bigm|
\mathcal{E}_1 \wedge (C_1\subseteq R)
\bigr)\geq q^{-1}$
(also by the definition of the red subgraph --- with probability $q^{-1}$, red is the rarest colour among
vertices in $K\setminus \widehat C_i$ where $\widehat C_i$ is the element of $\{\widehat C_j\}$ containing $C_1$).
Now the probability of event $\mathcal{E}_2$
is at most
$$
\Pr\big( (C_1\subseteq R) \wedge (|C_1|>N^{1/8}) \bigm|
|R\cap K|\leq N/q+ N^{15/16}\big)$$
which is at most
the conditional probability that
$A[R\capK]$ has a component of size at least
$N^{1/8}$,
conditioned on $|R\cap K|\leq N/q+ N^{15/16}$.
But, by Lemma~\ref{lem:FL},
$A[R\capK]$ is
an Erd\H os-R\'enyi
random graph with
$|R\capK|\leq(q^{-1}+N^{-1/16})N$ vertices and
subcritical edge probability $\critprob \leq \lambda/N<|R\capK|^{-1}$.
So the probability that it has a
component of
size at least $N^{1/8}$ is at most
$\epsilon(N)$, where $\epsilon(\cdot)$ is smaller than
any inverse polynomial~\cite[Proof of Theorem~5.4]{JLR}.
Thus, we have shown $\Pr(\mathcal{E}_2) \leq \eta/(6 q^2)$
(provided $N$ is sufficiently large), so we have proved~(\ref{eq:bd1}).
Next we prove inequality~(\ref{eq:bd2}).
The event $\mathcal{W}$ is that $|C_1|+\cdots+|C_s|\geq
N^{15/16}$, for some $s\leq2N^{5/16}$.
The component $\widehat C_1$
is in general composed of a number of components from~$\{C_j\}$.
However, the number of constituent components cannot be larger than
the number $|\edgesubset\cap(K\timesT)|$ of edges joining
$K$ to $T$. This number can be upper-bounded using a
bounding configuration~$\overline\es$ drawn from $\mathrm{ER}(\varGamma;p)$.
We see from a Chernoff bound that whp the number
$|\overline\es\cap(K\timesT)|$ of edges joining
$K$ to $T$ is
less than $2\termsN^{1/4}\leq2N^{5/16}$,
i.e., twice the expected number,
so whp,
at most $2N^{5/16}$
components from~$\{C_j\}$ contribute to $\widehat C_1$.
Thus $\Pr(\mathcal{W}\mid\mathcal{L})=1-\epsilon(N)$
and (\ref{eq:bd2}) follows.
Finally we prove inequality~(\ref{eq:bd3}).
As before, generate~$A$ by selecting a
set $A^+\subseteq K^{(2)}$ from
the appropriate induced distribution,
letting
$A^-=K^{(2)}\setminus A^+$, and selecting $A$ from the distribution
$\mathrm{RC}(\varGamma;q,p;A^+,A^-)$.
We will condition on the event $\mathcal{W}$ (which is entirely determined by $A^+$).
As above,
$$\Pr\big(Y(A)=1\bigm| A^+\subseteq A, A\cap A^-=\emptyset\big)
\geq \Pr(Y(\underline\es)=1),$$
where
$p'(e) = p(e)/q$ and
$\underline\es$ is a random variable drawn from
$\mathrm{ER}(\varGamma;p';A^+,A^-)$.
Consider two terminals $i,j$ in $T$.
We can use Lemma~\ref{lem:yellowblue} to find an upper bound for the probability that
$i$ and $j$ are {\it not\/} connected
in $\underline\es$
via one of the components $C_1,\ldots,
C_s$. In $\underline\es$, edges from $T$ to $K$
are selected independently with probability $q^{-1}N^{-3/4}$.
If we colour vertices in $K$ adjacent to $i$ (respectively, $j$)
yellow (respectively, blue) then we are in the situation
of Lemma~\ref{lem:yellowblue}, with $\sprob=q^{-1}N^{-3/4}$.
From the remarks about monotonicity in $s$ and~$\nu$, we may assume
for an upper bound that $\nu=N^{15/16}$
and $s=2N^{5/16}$.
The probability we want to bound is that
of not having a bicoloured component.
Observe
$$(1-\sprob)^{\nu/s}=(1-q^{-1}N^{-3/4})^{\frac12N^{5/8}}
\leq 1-
\tfrac14
q^{-1}N^{-1/8},$$
for $N$ sufficiently large,
and since the function $u\mapsto u(2-u)$ is increasing in the range $(0,1)$,
$$(1-\sprob)^{\nu/s}(2-(1-\sprob)^{\nu/s})
\leq (1-
\tfrac14
q^{-1}N^{-1/8})(1+
\tfrac14
q^{-1}N^{-1/8}).$$
Applying Lemma~~\ref{lem:yellowblue},
$$\Pr(\text{$i\not\sim j$ in $\underline\es$} )
=(1-
\tfrac1{16}
q^{-2}N^{-1/4})^{2N^{5/16}}
<\exp(-
\tfrac18
q^{-2}N^{1/16}).$$
So whp $i\sim j$. It follows that whp all vertices in
$T$ are connected to each other, and
so $\Pr(Y(\underline\es)=1 )\geq1-\epsilon(N)$.
This deals with~(\ref{eq:bd3}) and completes the proof.
\end{proof}
\subsection{The random cluster model on a clique connected to an independent set}
\label{sec:variant}
Construct~$\varGamma$ as in Section~\ref{sec:gadget}.
Let $\vargadget=(\gadgetvertices,\gadgetedges \setminusT^{(2)})$ be
the graph derived from~$\varGamma$ by deleting edges within~$T$.
The vertices in~$T$ are called the ``terminals'' of $\vargadget$.
Let $\boldgamma=\{\gamma_\graphedge\}$ be the set of edge weights
defined by $\gamma_\graphedge = p(\graphedge)/(1-p(\graphedge))$.
For an edge subset $A'\subseteq\gadgetedges\setminusT^{(2)}$,
let $\kappa'(\gadgetvertices,A')$ denote the number of connected components
that do not contain terminals
in the graph $(\gadgetvertices,A')$.
Let $\mathcal{A}^k$ denote the set of edge subsets $A'\subseteq\gadgetedges
\setminusT^{(2)}$
for which the terminals of $(\gadgetvertices,A')$
are contained in exactly $k$~connected components.
Let $\mathcal{A} = \bigcup_{k\in[t]} \mathcal{A}^k$ (this is the set of all edge subsets of $\vargadget$).
Let $Z^k$ be $q^{-k}$ times the contribution to $\ZTutte(\vargadget;q,\boldgamma)$ from
edge subsets $A'\in \mathcal{A}^k$.
Formally, $Z^k = \sum_{A' \in \mathcal{A}^k} q^{\kappa'(V,A')}\gamma({A'})$.
Let $Z = \sum_{k=1}^t Z^k$.
We will use the following lemma to apply
Lemmas~\ref{lem:excludedmiddleone}
and~\ref{lem:excludedmiddletwo}
from Section~\ref{sec:gadget}
in our reductions.
\begin{lemma}
\label{lem:wiring}
$Z^k/Z=\Pr(Y(A)=k)$, where
$A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$.
\end{lemma}
\begin{proof}
From the definitions of the random cluster model,
$$\Pr(Y(A)=k) =
\frac{\sum_{A'\in \mathcal{A}^k} \widetilde{P}(\varGamma;A'\cup T^{(2)},q,p)}
{
\sum_{A'\in \mathcal{A}} \widetilde{P}(\varGamma;A'\cup T^{(2)},q,p)}.$$
Plugging in the definition of $\widetilde{P}(\varGamma;A'\cup T^{(2)},q,p) $,
this is
$$
\frac{\sum_{A'\in \mathcal{A}^k} \gamma({A'})q^{\kappa'(\gadgetvertices,A')+1} \prod_{e\in \gadgetedges \setminus T^{(2)}}(1-p(e))}
{\sum_{A'\in \mathcal{A}} \gamma({A'})q^{\kappa'(\gadgetvertices,A')+1}
\prod_{e\in \gadgetedges \setminus T^{(2)}}(1-p(e))},$$
which is what we require, once we cancel a factor of
$$q \prod_{e\in \gadgetedges \setminus T^{(2)}}(1-p(e))$$ from the numerator and denominator.
\end{proof}
\section{Computational problems, fully polynomial randomised approximation schemes and efficiently approximable real numbers}
\label{sec:FPRAS}
Fix real numbers $q>2$ and $\gamma>0$ and consider the
following computational problem, which is parameterised by~$q$ and~$\gamma$.
\begin{description}
\item[Problem] $\textsc{Tutte}(q,\gamma)$.
\item[Instance] Graph $G=(V,E)$.
\item[Output] $\ZTutte(G;q,\boldgamma)$,
where
$\boldgamma$ is the constant function with $\boldgamma_\graphedge = \gamma$ for every $\graphedge\inE$.
\end{description}
We will have much more to say about computational approximations of real numbers below. For the moment, it may help
the reader to think of~$q$ and~$\gamma$ as being rational.
Jaeger, Vertigan and Welsh~\cite{JVW90}
have shown that $\textsc{Tutte}(q,\gamma)$ is \#P-hard
for every fixed $q>2$ and $\gamma>0$.
Thus, it is unlikely that there is a polynomial-time algorithm for exactly
solving this problem. (If there were such an algorithm, this would entail
$\mathrm P=\mathrm{\#P}$, and of course $\mathrm P=\mathrm{NP}$.)
We are interested in the complexity of \emph{approximately} solving $\textsc{Tutte}(q,\gamma)$.
We start by defining the relevant concepts.
A \emph{randomised approximation scheme\/} is an algorithm for
approximately computing the value of a function~$f:\Sigma^*\rightarrow
\mathbb{R}$.
The
approximation scheme has a parameter~$\varepsilon>0$ which specifies
the error tolerance.
A \emph{randomised approximation scheme\/} for~$f$ is a
randomised algorithm that takes as input an instance $ x\in
\Sigma^{\ast }$ (e.g., for the problem $\textsc{Tutte}(q,\gamma)$, the
input would be an encoding of a graph~$G$) and a rational error
tolerance $\varepsilon >0$, and outputs a rational number $z$
(a random variable of the ``coin tosses'' made by the algorithm)
such that, for every instance~$x$,
\begin{equation}
\label{eq:3:FPRASerrorprob}
\Pr \big[e^{-\epsilon} f(x)\leq z \leq e^\epsilon f(x)\big]\geq \frac{3}{4}\, .
\end{equation}
The randomised approximation scheme is said to be a
\emph{fully polynomial randomised approximation scheme},
or \emph{FPRAS},
if it runs in time bounded by a polynomial
in $ |x| $ and $ \epsilon^{-1} $.
Note that the quantity $3/4$ in
Equation~(\ref{eq:3:FPRASerrorprob})
could be changed to any value in the open
interval $(\frac12,1)$ without changing the set of problems
that have randomised approximation schemes~\cite[Lemma~6.1]{JVV86}.
We say that a real number~$z$ is \emph{efficiently approximable} if there is an FPRAS
for the problem which maps any input to the output~$z$.
Approximations to real numbers are useful.
For example, if $\hat q$ and $\hat \gamma$ are
approximations to~$q$ and~$\gamma$ satisfying
$$e^{-\tfrac{\varepsilon}{n+m}} q \leq \hat q \leq
e^{\tfrac{\varepsilon}{n+m}} q$$
and
$$e^{-\tfrac{\varepsilon}{n+m}} \gamma \leq \hat \gamma \leq
e^{\tfrac{\varepsilon}{n+m}}
\gamma$$
and $\hat\boldgamma_\graphedge = \hat\gamma$ for every $\graphedge\inE$
then
\begin{equation}
\label{approxparams}
e^{-\varepsilon }
\ZTutte(G;q,\boldgamma) \leq
\ZTutte(G;\hat q,\hat \boldgamma) \leq
e^{\varepsilon }
\ZTutte(G; q, \boldgamma).
\end{equation}
Thus, to approximate $\ZTutte(G;q,\boldgamma)$,
it suffices to first compute rational approximations $\hat{q}$ and $\hat \gamma$, and then approximate
$\ZTutte(G;\hat q,\hat \boldgamma)$.
When the parameters are efficiently approximable
reals, it is possible
to approximate quantities associated with the gadgets~$\varGamma$
and~$\vargadget$
that we studied in Section~\ref{sec:gadget} and~\ref{sec:variant}.
\begin{lemma}
\label{lem:computeZ}
Suppose $q>2$ is an efficiently computable real.
Consider the gadget $\vargadget$ from Section~\ref{sec:variant} with
parameters~$\terms$, $N$ and $\critprob$
where
$\critprob\in[0,1]$ is a rational number and $N^{1/4}$ is an integer.
There is an FPRAS for computing $Z^1$ and $Z^\terms$, given
inputs~$\terms$, $N$ and $\critprob$.
\end{lemma}
\begin{proof}
In this proof only, the notation $\vargadget_{N,\terms}$ is used to make explicit the
size of the gadget~$\vargadget$.
Let $E(\vargadget_{N,\terms})$ be the set of edges of $\vargadget_{N,\terms}$.
Let $\mathcal{A}^{k,\ell}$ denote the set of edge subsets $A\subseteq E(\vargadget_{N,\terms})$
with $k$~connected components containing vertices in $T$, and $\ell$
other connected components.
Let $w(\terms,N,k,\ell) = \sum_{A\in \mathcal{A}^{k,\ell}} \gamma(A)$.
Thus
$Z^k=\sum_{\ell=0}^N w(\terms,N,k,\ell)q^\ell$. We exhibit recurrence
relations for $w(\terms,N,j,k)$, from which it follows that
these
can be computed in
polynomial time by dynamic programming.
The rest of this proof is straightforward, but provides the details.
To reduce the number of boundary cases
we need to consider, it is convenient to allow one or other of
$k$ and $\ell$ to take on the value $-1$.
Of course, we stipulate
$$
w(\terms,N,-1,\ell)=w(\terms,N,k,-1)=0, \quad
\text{for all $\terms, N,k,\ell\geq0$}.
$$
Another easy-to-verify boundary case is
$$
w(\terms,0,k,\ell)=\begin{cases}1,&\text{when $k=\terms$ and $\ell=0$};\\
0,&\text{otherwise},
\end{cases}
$$
which is valid for all $\terms,k,\ell\geq0$.
Also $w(\terms,N,k,\ell)=0$ if exactly one of $\terms$ and $k$ is $0$
and $w(0,N,0,0)$ is~$1$ is $N=0$ and~$0$ otherwise.
Finally, $w(0,0,k,\ell) = 0$ if $k+\ell>0$.
The general recurrence, covering all situations other than the boundary cases
already mentioned can now be given.
\begin{align}
w(\terms,N,k,\ell)
&=\sum_{\textstyle{{1\leq i\leq\terms\atop1\leq j\leqN}}}
\binom{\terms}{i}\binom{N-1}{j-1}w(i,j,1,0)w(\terms-i,N-j,k-1,\ell)
\label{eq:rec1}\\
&\qquad\null+\sum_{1\leq j\leqN}\binom{N-1}{j-1}
w(0,j,0,1)w(\terms,N-j,k,\ell-1).
\label{eq:rec2}
\end{align}
(Although valid for all $k,\ell\geq0$, the recurrence becomes trivial
when $k+\ell=1$, a point we must return to at the end.)
The explanation is as follows. We partition the sum defining $w(\terms,N,k,\ell)$
according to the connected component~$C$ in $A$ containing some distinguished
vertex $v$ in~$K$.
(Note that $N>0$, since one of the boundary cases covers $N=0$.)
Let $i=|C\cap T|$ be the number of vertices of~$C$ that lie in~$T$, and $j=|C\cap K|$
the number that lie in~$K$. Summation (\ref{eq:rec1}) deals with the situation $i>0$ and
(\ref{eq:rec2}) with the situation $i=0$. The binomial coefficients in~(\ref{eq:rec1})
count the number of
connected components $C\ni v$ that contain the distinguished vertex, and
have the correct intersections $|C\capT|=i$ and $|C\capK|=j$
with $T$ and~$K$; the factor $w(i,j,1,0)$ counts the weight of {\it connected\/}
subgraphs of $\vargadget[C]$; and $w(\terms-i,N-j,k-1,\ell)$
the weight of subgraphs of $\vargadget[K\cupT\setminus C]$ that
have the correct number of connected components (i.e., $k-1$ with vertices in~$T$,
and $\ell$ without). The analysis of summation~(\ref{eq:rec2}) is entirely similar.
Assuming $k+\ell>1$, the recurrence is well founded, in the sense that all occurrences
of $w(\terms',N',\cdot,\cdot)$ on the right hand side (apart from those that
get multiplied by zero) have
$\terms'+N'<\terms+N$.
When $k+\ell=1$, i.e., $(k,\ell)\in\{(0,1),(1,0)\}$ the recurrence (\ref{eq:rec1},\ref{eq:rec2})
becomes the trivial $w(\terms,N,k,\ell)=w(\terms,N,k,\ell)$.
Note that $(k,\ell)=(0,1)$ entails $\terms=0$
(otherwise $w(\terms, N,k,\ell)=0$)
and $(k,\ell)=(1,0)$ entails $\terms>0$,
so only one of the possibilities $(k,\ell)\in\{(0,1),(1,0)\}$ occurs for a given pair
$(\terms, N)$.
In order to make progress
in this situation, we apply a preprocessing step:
$$
w(\terms, N,1,0)=\prod_{e\in E(\vargadget_{N,\terms})}(1+\gamma_e)
-\sum_{\textstyle{0\leq k\leq \terms,0\leq\ell\leqN\atop k+\ell>1}}
w(\terms,N,k,\ell).
$$
Exactly the same formal expression applies to $w(\terms, N,0,1)$. The identity
merely expresses complementation: the weight of subgraphs with $k+\ell=1$ equals
the total weight of subgraphs, less the weight of subgraphs with $k+\ell>1$.
With this modification, the recurrence becomes well founded.
Since only a polynomial number of distinct tuples $(\terms,N,k,\ell)$ arise,
the recurrence can be solved by dynamic programming in polynomial time.
Note that the computation of $\prod_{e\in E(\vargadget_{N,\terms})}(1+\gamma_e)$
in the preprocessing step can be done exactly since $\critprob$ is a rational
and $N^{1/4}$ is an integer.
Once the $w(\terms,N,k,\ell)$ values are computed, we wish to estimate
$$Z^k=\sum_{\ell=0}^N w(\terms,N,k,\ell)q^\ell.$$
Let $\varepsilon$ be the desired accuracy in the approximation-preserving
reduction. Compute a rational $\hat q$
in the range
$e^{-\varepsilon/N} q \leq \hat q \leq e^{\varepsilon/N} q$ and return
$Z^k=\sum_{\ell=0}^N w(\terms,N,k,\ell){\hat q}^\ell$.
\end{proof}
It will also be necessary for us to approximate the critical edge
probability~$\critprob$, so that, with this approximation, the graph~$\varGamma$
approximately satisfies Equation~(\ref{eq:balance}) in Lemma~\ref{lem:excludedmiddleone}.
The following lemma shows that this is
possible.
\begin{lemma}
\label{lem:computerho}
Suppose $q>2$ is an efficiently computable real. Fix $\gamma>0$ and
let $\lambda = \lambda_c + (q-\lambda_c)/2$.
Suppose that $\chi>0$ is rational. Consider the gadget $\varGamma$ from Section~\ref{sec:gadget} with parameters~$\terms$, $N$, and $\critprob$. There is
a randomised algorithm
whose running time is at most a polynomial in $\chi^{-1}$, $N$ and $\terms$ which takes input $N$ and $\terms$ (where it is assumed that $N^{1/4}$ is an integer and that $N\geq\max\{\terms^{16}, N_0\}$ for the constant $N_0$
from Lemma~\ref{lem:excludedmiddleone})
and, with probability at least~$3/4$, computes
a rational ${\critprob}$ in the range $[N^{-3},\lambda/N]$ such that, if $A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$, then \begin{equation} \label{eq:compbalance} e^{-\chi} \gamma \leq \frac{\Pr(Y(A)=1)} {\Pr(Y(A)=\terms)} \leq e^{\chi} \gamma.
\end{equation}
\end{lemma}
\begin{proof}
First, we establish a useful preliminary fact.
If $0<\critprob\leq 1/4$, $0<\delta\leq 1$
and
$\frac{1}{ 1+\delta} \critprob \leq \hat \critprob \leq (1+\delta) \critprob$
then
\begin{equation}
\label{eq:trivia}
e^{-2\delta} \frac{\critprob}{1-\critprob}
\leq \frac{\hat\critprob}{1-\hat\critprob}
\leq e^{2\delta} \frac{\critprob}{1-\critprob}.
\end{equation}
To see this, note that $e^{\delta} \critprob \leq 3/4$
so $1-e^{\delta} \critprob\geq \critprob$. Then
letting $x = e^{\delta}-1$,
$$1-\critprob
\leq
1 -\critprob + x(1-e^{\delta}\critprob) - x \critprob
= (1+x)(1-(1+x)\critprob)
= e^{\delta}(1-e^{\delta} \critprob),$$
which, together with $1+\delta \leq e^{\delta}$, gives the right-most inequality in~(\ref{eq:trivia}).
Similarly,
$\critprob e^{-\delta} \leq 1/4$ so $1-e^{-\delta} \critprob \geq
\critprob $ so
letting $x=1-e^{-\delta}$,
$$1-\critprob
\geq
1 -\critprob + x \critprob - x(1-e^{-\delta}\critprob)
= (1-x)(1-(1-x)\critprob)
= e^{-\delta}(1-e^{-\delta} \critprob),$$
which, together with $e^{-\delta} \leq \frac{1}{1+\delta}$, gives the left-most inequality in~(\ref{eq:trivia}).
Now consider the gadget $\varGamma$ with parameters~$\terms$,
$N$, and $\critprob$.
Let $n$ and $m$ be the number of vertices, and edges, respectively, of
the variant $\vargadget$.
We will be interested in three quantities which depend
upon~$\critprob$: $Z^1$, $Z^\terms$ and
$\zeta \doteq \frac{\Pr(Y(A)=1)}
{\Pr(Y(A)=\terms)}$.
For this proof only, we will denote these as $Z^1_\critprob$,
$Z^\terms_\critprob$ and $\zeta_\critprob$ to make the dependence on
$\critprob$ clear.
By Lemma~\ref{lem:wiring}, $\zeta_\critprob = Z^1_\critprob/Z^\terms_\critprob$.
Lemma~\ref{lem:computeZ} provides an FPRAS for computing $\zeta_\critprob$ for
rational parameters~$\critprob$.
We will
set $\delta = \chi/(16(n+m))$ and we will
use the FPRAS with accuracy parameter~$\chi/4$ to estimate
$\zeta_\critprob$ for every rational
$$\critprob = \frac{1}{N^3} {(1+\delta)}^\mu$$
where $\mu$ is a non-negative integer
with $ \frac{1}{N^3} {(1+\delta)}^\mu \leq \frac{\lambda}{N}$.
Note that the number of integers~$\mu$ in this range is $O(\delta^{-1}
\log N)$.
We will first ``power up'' the success probability of the FPRAS using standard techniques~\cite{jvv} so
that the probability that any of these $O(\delta^{-1}
\log N)$ calls to the FPRAS fails is at most~$3/4$.
If we find a rational $\hat \critprob$ for which our
estimate $\hat \zeta_{\hat \critprob}$ satisfies
$e^{-\chi/2} \gamma \leq \hat \zeta_{\hat \critprob} \leq e^{\chi/2}
\gamma$, we will return this value~$\hat\critprob$.
Clearly, this will be an acceptable answer, since
$e^{-\chi/4} \zeta_{\hat\critprob} \leq
\hat \zeta_{\hat \critprob} \leq
e^{\chi/4} \zeta_{\hat\critprob}$.
To finish the proof, we just need to argue that we will try a
$\hat\critprob$ which we accept.
Lemma~\ref{lem:excludedmiddleone}
guarantees that there is a value $\critprob^*\in
[N^{-3},\lambda N^{-1}]$ such that
$\zeta_{\critprob^*} = \gamma$.
In our computations, we will compute $\zeta_{\hat \critprob}$ for some
$\hat \critprob$ in the range
$ \frac{1}{1+\delta} \critprob^* \leq \hat \critprob \leq (1+\delta) \critprob^*$.
Recall that every edge~$e$ of
the gadget has edge-weight $\gamma_e=p(e)/(1-p(e))$,
and that $p(e)=\critprob$ for edges within the part of the gadget
corresponding to the clique~$K$. The weights of the other edges do not
depend on $\critprob$.
Thus, by~(\ref{eq:trivia})
and by analogy to~(\ref{approxparams}),
$e^{-\chi/4} \zeta_{\critprob^*} \leq \zeta_{\hat\critprob} \leq
e^{\chi/4} \zeta_{\critprob^*}$. Thus, $\hat \critprob$ will be accepted.
\end{proof}
\section{Approximation-preserving reductions and \BIS}
\label{sec:APred}
Our main tool for understanding the relative difficulty of
approximation counting problems is \emph{approximation-preserving reductions}.
We use
Dyer, Goldberg, Greenhill and Jerrum's notion of
approximation-preserving reduction~\cite{APred}.
Suppose that $f$ and $g$ are functions from
$\Sigma^{\ast }$ to~$\mathbb{R}$. An ``approximation-preserving
reduction'' from~$f$ to~$g$ gives a way to turn an FPRAS for~$g$
into an FPRAS for~$f$. Here is the definition. An {\it approximation-preserving reduction\/}
from $f$ to~$g$ is a randomised algorithm~$\mathcal{A}$ for
computing~$f$ using an oracle for~$g$. The algorithm~$\mathcal{A}$ takes
as input a pair $(x,\varepsilon)\in\Sigma^*\times(0,1)$, and
satisfies the following three conditions: (i)~every oracle call made
by~$\mathcal{A}$ is of the form $(w,\delta)$, where
$w\in\Sigma^*$ is an instance of~$g$, and $0<\delta<1$ is an
error bound satisfying $\delta^{-1}\leq\mathop{\mathrm{poly}}(|x|,
\varepsilon^{-1})$; (ii) the algorithm~$\mathcal{A}$ meets the
specification for being a randomised approximation scheme for~$f$
(as described above) whenever the oracle meets the specification for
being a randomised approximation scheme for~$g$; and (iii)~the
run-time of~$\mathcal{A}$ is polynomial in $|x|$ and
$\varepsilon^{-1}$.
If an approximation-preserving reduction from $f$ to~$g$
exists we write $f\leq_\mathrm{AP} g$, and say that {\it $f$ is AP-reducible
to~$g$}.
Note that if $f\leq_\mathrm{AP} g$ and $g$ has an FPRAS then $f$ has an FPRAS\null.
(The definition of AP-reduction was chosen to make this true).
If $f\leq_\mathrm{AP} g$ and $g\leq_\mathrm{AP} f$ then we say that
{\it $f$ and $g$ are AP-interreducible}, and write $f\equiv_\mathrm{AP} g$.
The definitions allow us to construct approximation-preserving
reductions between problems~$f$ and~$g$ with real parameters without
insisting that the parameters themselves be efficiently approximable. Nevertheless, some
of our results restrict attention to efficiently approximable parameters.
According to the definition, approximation-preserving reductions may use randomisation. Nevertheless,
the reductions that we present in this paper are deterministic,
except for where they make use of an FPRAS to approximate a real parameter.
A word of warning about terminology:
Subsequent to~\cite{APred}, the notation $\leq_\mathrm{AP}$ has been
used
to denote a different type of approximation-preserving reduction which applies to
optimisation problems.
We will not study optimisation problems in this paper, so hopefully this will
not cause confusion.
Dyer et al.~\cite{APred} studied counting problems in \#P and
identified three classes of counting problems that are interreducible
under approximation-preserving reductions. The first class, containing the
problems that admit an FPRAS, are trivially AP-interreducible since
all the work can be embedded into the reduction (which declines to
use the oracle). The second class is the set of problems that are
AP-interreducible with \textsc{\#Sat}, the problem of counting
satisfying assignments to a Boolean formula in CNF\null.
Zuckerman~\cite{zuckerman}
has shown that \textsc{\#Sat}{} cannot have an FPRAS unless
$\mathrm{RP}=\mathrm{NP}$. The same is obviously true of any problem
to which \textsc{\#Sat}{} is AP-reducible.
The third class appears to be of intermediate complexity.
It contains all of the counting problems
expressible in a certain logically-defined complexity class. Typical
complete problems include counting the downsets in a partially ordered
set~\cite{APred},
computing the partition function of the ferromagnetic Ising model with
varying interaction energies and local external magnetic
fields~\cite{ising}
and counting the independent sets in a bipartite graph,
which is defined as follows.
\begin{description}
\item[Problem] $\BIS$.
\item[Instance] A bipartite graph $B$.
\item[Output] The number of independent sets in $B$.
\end{description}
We showed in \cite{APred} that \BIS\ is complete for the
logically-defined
complexity class $\mathrm{\#RH}\Pi_1$ with respect to approximation-preserving
reductions.
We presume that there is no FPRAS for \BIS, but this is not
known.
\section{Our results}
\label{sec:results}
The next three sections of this paper give an approximation-preserving
reduction from \BIS\ to the problem $\textsc{Tutte}(q,\gamma)$.
We start by defining the problem of computing the Tutte polynomial of
a uniform hypergraph~$H$
with fixed positive edge weights.
For fixed positive real numbers~$q$ and $\gamma$
the problem $\textsc{UniformHyperTutte}(q,\gamma)$ is defined as follows.
\begin{description}
\item[Problem] $\textsc{UniformHyperTutte}(q,\gamma)$.
\item[Instance] A uniform hypergraph $H=(\calV,\calE)$.
\item[Output] $\ZTutte(H;q,\gamma)$,
where $\boldgamma$ is the constant function with $\boldgamma_f = \gamma$ for every $f\in\calE$.
\end{description}
In Section~\ref{bistohypertutte}
we show that, for every $q>1$, there is an approximation-preserving
reduction from \BIS\ to $\textsc{UniformHyperTutte}(q,q-1)$.
$\textsc{UniformHyperTutte}(q,\gamma)$ is the problem of computing the Tutte polynomial of
a uniform hypergraph.
The most difficult part of the paper is reducing this approximation problem
to the problem of approximately computing the (multivariate) Tutte polynomial of an undirected
graph (a $2$-uniform hypergraph).
Section~\ref{sec:tographs}
shows that for any positive real numbers $q>2$ and $\gamma>0$,
there is an approximation-preserving reduction from $\textsc{UniformHyperTutte}(q,\gamma)$
to the following problem.
\begin{description}
\item[Problem] $\TwoWeightFerroTutte(q)$.
\item[Instance] Graph $G=(V,E)$
with an edge-weight function $\boldgamma': E \rightarrow
\{\gamma',\gamma''\}$
where $\gamma'$ and $\gamma''$ are rationals in the interval $[|V|^{-3},1]$.
\item[Output] $\ZTutte(G;q,\boldgamma')$.
\end{description}
Finally, Section~\ref{shift} gives an approximation-preserving reduction
from
the problem
$\TwoWeightFerroTutte(q)$ to
the problem $\textsc{Tutte}(q,\gamma)$ for any
$q>2$ and $\gamma>0$.
\section{Approximately computing the Tutte polynomial of a uniform hypergraph}
\label{bistohypertutte}
We first
define a parameterised version of $\BIS$, and
also a restricted version of this in which vertices on the right-hand side
are required to have the same degree.
\begin{description}
\item[Problem] $\BIS(\mu)$.
\item[Instance] Bipartite graph $B$.
\item[Output] $Z_{\textsc{IS}}(B;\mu)=\sum_I\mu^{|I|}$,
where the sum is over all independent sets $I$ in $B$.
\end{description}
\begin{description}
\item[Problem] $\#\textsc{SemiRegularBIS}(\mu)$.
\item[Instance] Bipartite graph $B=(U,V,E)$ in which every vertex in $V$ has
the same degree.
\item[Output] $Z_{\textsc{IS}}(B;\mu)=\sum_I\mu^{|I|}$,
where the sum is over all independent sets $I$ in $B$.
\end{description}
\begin{lemma}
\label{lem:bis}
Suppose that
$\mu>0$ is efficiently approximable. Then $\BIS \leq_\mathrm{AP} \#\textsc{SemiRegularBIS}(\mu)$.
\end{lemma}
\begin{proof}
In~\cite[Lemma 15]{APred}, we gave an AP-reduction from
$\BIS$ to \textsc{\#BipartiteMaxIS}, the problem of counting \emph{maximum}
independent sets in a bipartite graph.
Here, we give two AP-reductions --- first, a reduction from \textsc{\#BipartiteMaxIS}\ to
$\BIS(\mu)$,
and then a reduction from
the intermediate problem
$\BIS(\mu)$ to $\#\textsc{SemiRegularBIS}(\mu)$.
Let $B$ be an instance of $\textsc{\#BipartiteMaxIS}$ with $n$ vertices and $m$ edges
and let $\varepsilon$ be the desired accuracy of the
approximation-preserving reduction.
Let $\xi$ be the size of a maximum independent set of~$B$
and let $Y$ be the number of maximum independent sets.
To do the construction, we first need a rough estimate of $\mu$,
so compute a rational value $\tilde{\mu}$ in the range
$\tfrac34 \mu \leq \tilde{\mu} \leq \tfrac54 \mu$.
Let
$s$ be an integer satisfying
$$ s- 1 \leq \left\lceil \frac{n+3}{\lg(1+2\tilde \mu/3)}\right\rceil \leq s.$$
Note that $s \geq \frac{n+3}{\lg(1+\mu/2)}$.
Let $B'$ be the graph with
vertex set $\{(u,i)\mid u\in V(B), i\in [s]\}$
and edge set $\{((u,i),(v,j)) \mid (u,v)\in E(B)\}$.
Next, we need a more accurate estimate of~$\mu$.
Compute a rational number $\hat \mu$
in the range
$e^{-\epsilon/(60 ns)} \mu \leq \hat \mu \leq e^{\epsilon/(60
ns)}\mu$.
Now since there are $n s$ vertices in $B'$,
$$e^{-\epsilon/60} Z_{\textsc{IS}}(B';\mu) \leq Z_{\textsc{IS}}(B';\hat \mu) \leq e^{\epsilon/60}
Z_{\textsc{IS}}(B';\mu).$$ Thus, by using our oracle for $\BIS(\mu)$ with accuracy parameter $\epsilon/60$,
we can compute a value $Z$ satisfying
\begin{equation}
\label{APredeq}
e^{-\epsilon/30} Z_{\textsc{IS}}(B';\hat \mu) \leq Z \leq e^{\epsilon/30}
Z_{\textsc{IS}}(B';\hat \mu).
\end{equation}
Now note that every independent set of~$B'$ points out an independent
set of~$B$. The vertex~$u$ of~$B$ is in the independent set of~$B$ if
there is at least one vertex~$(u,i)$ that is in the independent set
of~$B'$.
A size-$k$ independent set of~$B$ thus makes a contribution of
${({(1+\hat\mu)}^s-1)}^k$ to $Z_{\textsc{IS}}(B';\hat \mu)$.
Since the number of independent sets of~$B$ is at most~$2^n$,
we have
\begin{equation}
\label{APredanother}
Y \leq \frac{Z_{\textsc{IS}}(B';\hat \mu)}{{({(1+\hat\mu)}^s-1)}^\xi} \leq
Y + \frac{2^n}{({(1+\hat\mu)}^s-1)}
\leq Y + \frac14,
\end{equation}
where the final inequality follows from the definition of~$s$ since
$\hat \mu \geq \mu/2$.
The following simple procedure now gives a sufficiently accurate
estimate to~$Y$. Take the value~$Z$ from Equation~(\ref{APredeq}),
divide it by ${({(1+\hat\mu)}^s-1)}^\xi$ and round down to the nearest
integer.
The fact that the accuracy is sufficient follows from~(\ref{APredeq})
and~(\ref{APredanother}). See~\cite[Theorem~3]{APred}
Finally, we present an AP-reduction from $\BIS(\mu)$ to $\#\textsc{SemiRegularBIS}(\mu)$.
Let $B=(U,V,E)$ be an $n$-vertex instance of $\BIS(\mu)$ in which the
maximum degree of a vertex in $V$ is $d>1$.
We'll construct $B'=(U',V',E')$ -- an instance of $\#\textsc{SemiRegularBIS}(\mu)$
in which every vertex in $V'$ has degree~$d$.
Let $\varepsilon$ be the desired accuracy of the
approximation-preserving reduction.
As above, compute a rational value $\tilde{\mu}$
such that
$\tfrac45 \tilde \mu \leq \mu \leq \tfrac43 \tilde \mu$.
Let $\mu^-$ denote the computed lower bound
$\tfrac45 \tilde \mu$ and let $\mu^+$ denote the computed upper bound
$\tfrac43 \tilde \mu$.
Let $D(x) = {(1+x)}^{d-1}$, $U(x) = x D(x)$,
$L(s,x) = (1+x)^s$, and $Y(s,x) = L(s,x) + D(x)-1$.
We start by computing an integer~$s$ such that
\begin{equation}
\label{eq:s}
\frac{1}{L(s,\mu^-)}
\max(D(\mu^+)-1,U(\mu^+))
\leq \frac{\varepsilon}{6 dn}.
\end{equation}
Note that $s=O( n \epsilon^{-1})$
so we can efficiently compute such a value~$s$ by starting with $s=1$,
and increasing $s$ one-by-one until we find a value for which~(\ref{eq:s}) holds.
Now
let $\Psi$ be a complete bipartite graph
with vertex sets $\{z_1,\ldots,z_{d}\}$ and $\{y_1,\ldots,y_s\}$.
For each vertex $v\in V$ of degree $\delta$,
take $d-\delta$ new copies of $\Psi$, and attach $v$ to vertex~$z_1$
of each copy.
Let $g < d n $ be the number of copies of $\Psi$ that get included in $B'$.
Now
$ Z_{\textsc{IS}}(\Psi;\mu) = L(s,\mu) + (1+\mu)D(\mu)-1$
and the total contribution to $Z_{\textsc{IS}}(\Psi;\mu)$ from independent
sets including the vertex~$z_1$ is $U(\mu)$.
Let $Y = Z_{\textsc{IS}}(\Psi;\mu) - U(\mu) = Y(s,\mu)$.
Clearly, $Y$ is the total contribution to $Z_{\textsc{IS}}(\Psi;\mu)$ from independent
sets not including the vertex~$z_1$
Then
$$Z_{\textsc{IS}}(B;\mu) Y^g \leq Z_{\textsc{IS}}(B';\mu) \leq Z_{\textsc{IS}}(B;\mu)
Z_{\textsc{IS}}(\Psi;\mu)^g$$
so
$$Z_{\textsc{IS}}(B;\mu) \leq \frac{Z_{\textsc{IS}}(B';\mu)}{Y^g} \leq Z_{\textsc{IS}}(B;\mu)
{\left(\frac{Z_{\textsc{IS}}(\Psi;\mu)}{Y}\right)}^g$$
Now
\begin{itemize}
\item $Z_{\textsc{IS}}(B';\mu)$ may be estimated with accuracy
parameter~$\epsilon/3$ using the
oracle.
\item $Y^g$ may be estimated directly (to the same accuracy)
by computing a value~$\hat \mu$
satisfying
$e^{-\epsilon/(6 d n s)} \mu \leq \hat \mu \leq e^{\epsilon/(6 d n s)} \mu $
which ensures that
$$e^{-\epsilon/(6 d n )} L(s,\mu) \leq L(s,\hat \mu) \leq e^{\epsilon/(6 d n ) }L(s,\mu) $$
And noting from (\ref{eq:s}) that
$$L(s,\mu) \leq Y = L(s,\mu)\left(1+\frac{D(\mu)-1}{L(s,\mu)}\right) \leq L(s,\mu) e^{\epsilon/(6 dn)}.$$
\item Finally, (\ref{eq:s}) gives
$${\left(\frac{Z_{\textsc{IS}}(\Psi;\mu)}{Y }\right)}^g = {\left(1+ \frac{U(\mu)}{Y(s,\mu)}\right)}^g \leq e^{\epsilon/3},$$
which finishes the reduction.
\end{itemize}
\end{proof}
We are interested in the situation
$\mu>1$ so that we have $q=\mu+1>2$ in
the parameters of $\textsc{UniformHyperTutte}(q,\gamma)$ in
Lemma~\ref{lem:one} below.
\begin{lemma}
\label{lem:one}
Suppose that
$\mu>0$ is efficiently approximable. Then
$$\#\textsc{SemiRegularBIS}(\mu) \leq_\mathrm{AP}\textsc{UniformHyperTutte}(\mu+1,\mu).$$
\end{lemma}
\begin{proof}
Let $B=(U,V,E)$ be an instance of $\#\textsc{SemiRegularBIS}(\mu)$.
Let $H=(\calV,\calE)$ be an instance of $\textsc{UniformHyperTutte}(\mu+1,\mu)$
constructed as follows.
Let~$s$ be a new vertex that is not in $U\cup V$ and let $\calV=U\cup\{s\}$.
For $v\in V$, let $\Gamma(v)=\{u\in U:(u,v)\in E\}$
and let
$F_v=\Gamma(v) \cup\{s\}$.
Let
$\calE=\bigcup_{v\in V} F_v$.
We will show below that
\begin{equation}
\label{eq:temp}
Z_{\textsc{IS}}(B,\mu)=(\mu+1)^{-1}\ZTutte(H;\mu+1,\mu).
\end{equation}
Now let
$\varepsilon$ be the desired accuracy in the approximation-preserving
reduction. To complete the reduction,
we first compute a value~$x$
in the range
$$e^{-\epsilon/2}
(\mu+1)^{-1} \leq x \leq
e^{\epsilon/2}
(\mu+1)^{-1}.
$$
This is easy to do since $\mu$ is efficiently approximable
and
$$\frac{e^{-\epsilon/2}}{\mu+1}
\leq \frac{1}{e^{\epsilon/2}\mu+1}\mbox{ and }
\frac{1}{e^{-\epsilon/2}\mu+1} \leq
\frac{e^{\epsilon/2}}{\mu+1}.$$
Then we use the oracle to estimate
$\ZTutte(H;\mu+1,\mu)$ with accuracy parameter~$\varepsilon/2$.
We finish the proof by establishing~(\ref{eq:temp}).
Let $S\subseteq V$. The contribution to $Z_{\textsc{IS}}(B,\mu)$
from independent sets $I$ with $I\cap V=S$ is
$\mu^{|S|}(\mu+1)^{|U|-|\Gamma(S)|}$, where $\Gamma(S)=\bigcup_{v\in S}\Gamma(v)$.
That is,
$$
Z_{\textsc{IS}}(B,\mu)=\sum_{S\subseteq V}\mu^{|S|}(\mu+1)^{|U|-|\Gamma(S)|}.
$$
On the other hand, the contribution to $\ZTutte(H;\mu+1,\mu)$
from the hyperedge set $\calE=\{F_v:v\in S\}$ is
$$\mu^{|S|}(\mu+1)^{\kappa(\calV,\calE)}=
\mu^{|S|}(\mu+1)^{|U|-|\Gamma(S)|+1},$$
since the vertices in $\Gamma(S)$ together
with $s$ form one connected component, and all other vertices are isolated.
Thus
$$\ZTutte(H;\mu+1,\mu)=
\sum_{S\subseteq V}\mu^{|S|}(\mu+1)^{|U|-|\Gamma(S)|+1}.$$
\end{proof}
\section{Approximately computing the multivariate Tutte polynomial of a graph}
\label{sec:tographs}
\begin{lemma}
Suppose that
$q>2$ and $\gamma>0$ are efficiently approximable. Then
$\textsc{UniformHyperTutte}(q,\gamma) \leq_\mathrm{AP} \TwoWeightFerroTutte(q)$.
\label{lem:last}
\end{lemma}
\begin{proof}
Start with a $\terms$-uniform hypergraph
$H=(\calV,\calE)$.
$H$ is an instance of
the problem
$\textsc{UniformHyperTutte}(q,\gamma)$.
For convenience, let $\calV=\{v_1,\ldots, v_n\}$
and $\calE=\{f_1,\ldots, f_m\}$.
The basic idea is to simulate each hyperedge with a copy of the
gadget $\varGamma$ from Section~\ref{sec:gadget}
(actually with the variant $\vargadget$ from Section~\ref{sec:variant}).
We start by setting the parameters. Let $\varepsilon$ be the desired accuracy in the approximation-preserving
reduction (see the definition of an AP-reduction in Section~\ref{sec:APred}).
Let $\chi = \varepsilon/(4m)$
The tolerance $\eta$, which is a parameter to Lemma~\ref{lem:excludedmiddletwo},
may be chosen to be any value such that
$$
\left(1+ \frac{\eta(1+e^{\chi}\gamma)}{1-\eta}\right)
\leq e^\chi.$$
Note that this inequality is achieved for $\eta=O(\varepsilon/m)$.
The reduction will construct an instance~$\widehat G=(V,E)$ of
the target problem
$\TwoWeightFerroTutte(q)$
along with an appropriate edge-weight function $\boldgamma': E \rightarrow
\{\gamma',\gamma''\}$. We will show that using
an oracle to approximate the solution to this instance with accuracy
parameter $\delta=\varepsilon/2$ enables us to get within
$\exp(\pm\varepsilon)$ of
$\ZTutte(H;q,\gamma)$.
Let $\lambda = \lambda_c + (q-\lambda_c)/2$ and let
$N_0$ be the quantity from
Lemmas~\ref{lem:excludedmiddleone} and~\ref{lem:excludedmiddletwo}.
Let $N$ be the smallest integer greater than
$\max\{\terms^{16},\eta^{-1/8},N_0\}$
for which $N^{1/4}$ is an integer.
Using the algorithm from Lemma~\ref{lem:computerho}
(suitably powered up so that its failure probability is at most~$3/8$)
we can compute a rational ${\critprob}$ in the
range $[N^{-3},\lambda/N]$
such that, if
$A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$, then
Equation~(\ref{eq:compbalance}) holds.
Note from Lemma~\ref{lem:excludedmiddletwo} that
Equation~(\ref{eq:dichotomy}) holds.
The construction is as follows. For every $j\in[m]$, let $K_j$ be a set of $N$ vertices
and let $V_j = K_j \cup f_j$.
Let $E_j = K_j^{(2)} \cup K_j \times f_j$ and let $G_j$ be the graph $(V_j,E_j)$.
Note that $G_j$ is a copy of the graph $\vargadget$ from Section~\ref{sec:variant} --- $f_j$
is the set of terminals.
Now we construct the graph $\widehat G = (V,E)$
where $V = \calV \cup \bigcup_{j\in[m]} K_j$
and $E = \bigcup_{j\in[m]} E_j$.
Define
$$ \hatp(e) = \begin{cases}
\critprob, &\text{if $e\in K_j^{(2)}$ for some~$j\in[m]$, and}\\
N^{-3/4}, &\text{if $e\in K_j\times f_j$ for some~$j\in[m]$.}
\end{cases}
$$
Let $\gamma_e = \hatp(e)/(1-\hatp(e))$, and let $\boldgamma'=\{\gamma_e\}_{e\inE}$.
Note that $\gamma' = \critprob/(1-\critprob)$
and $\gamma'' =
N^{-3/4}/(1-N^{-3/4})$. These are both rational values in
the interval $[|V|^{-3},1]$, as required, since $\lambda/N \leq 1/2$
by Lemma~~\ref{lem:excludedmiddleone}.
Denote by $\Pi_\edge$ the set of all partitions of $f_\edge$.
Given $\pi_\edge\in\Pi_\edge$,
let $\mathcal{A}_\edge^{\pi_\edge}$
be the collection of all edge subsets $A_\edge \subseteqE_\edge$
that induce the partition $\pi_\edge$ (into connected components) on $f_\edge$.
For $A_\edge \subseteq E_\edge$, let $Y(A_\edge)$ denote
the number of
connected components that contain
terminals (vertices in $f_\edge$) in the graph
$(V_\edge,A_\edge)$.
Let $\kappa'(V_\edge,A_\edge) = \kappa(V_\edge,A_\edge) -
Y(A_\edge)$ be the number of remaining connected components (that do not contain terminals).
For partitions $\pi$ and $\pi'$, $\pi\vee\pi'$ denotes the
finest partition that is a common coarsening of $\pi$ and
$\pi'$.
If two elements are together in $\pi$ or in $\pi'$ then they are together in the coursening.
The ``coarsest'' partition has one block and the ``finest'' partition consists of singleton blocks.
Technically, $\pi_\edge\in \Pi_\edge$ is a partition of $f_\edge$
but we consider it as a partition of the entire vertex set $\calV$
by extending it with singleton blocks.
Now $$\ZTutte(\widehat G;q,\boldgamma')=\sum_{A \subseteq E}
q^{\kappa(V,A)}
\prod_{e\inE}
\gamma_e.$$
Let $A_\edge = A \cap E_\edge$.
Let $\gamma({A_\edge}) = \prod_{e\in A_\edge} \gamma_e$.
Then
\begin{align*}\ZTutte(\widehat G;q,\boldgamma')&=
\sum_{\textstyle{\pi_\edge \in\Pi_\edge \atop \forall \edge \in[m]}}\>
\sum_{\textstyle{A_\edge\in \mathcal{A}_j^{\pi_j}
\atop\forall \edge\in[m]}}\>
q^{\kappa(V,
\bigcup_{\edge\in[m]} A_\edge)}
\prod_{\edge\in[m]}
\gamma(A_\edge)\\
&=
\sum_{\textstyle{\pi_\edge \in\Pi_\edge \atop \forall \edge \in[m]}}\>
q^{\kappa(\pi_1\vee\cdots\vee\pi_m)}
\sum_{\textstyle{A_\edge\in \mathcal{A}_j^{\pi_j} \atop\forall \edge
\in[m]}}\>
\prod_{\edge\in[m]}
\gamma(A_\edge)
q^{\kappa'(V_\edge,
A_\edge)},
\end{align*}
where $\kappa(\pi_1\vee\cdots\vee\pi_m)$ denotes the number of blocks
in the
partition $\pi_1\vee\cdots\vee\pi_m$.
Now let
$$
Z_\edge(\pi_\edge)=\sum_{A_\edge\in\mathcal{A}_\edge^{\pi_\edge}}\gamma(A_\edge)\,q^{\kappa'(V_\edge
,A_\edge)}
$$
and pull out the contribution of each~$\edge$ to get
\begin{equation}\label{eq:ZPottsbasic}
\ZTutte(\widehat G;q,\boldgamma')
=
\sum_{\textstyle{\pi_\edge \in\Pi_\edge \atop \forall \edge \in[m]}}\>
q^{\kappa(\pi_1\vee\cdots\vee\pi_m)}
\prod_{\edge=1}^m Z_\edge(\pi_\edge).
\end{equation}
At this point, we make some connections to Section~\ref{sec:variant}.
Let $\Pi_\edge^k$ denote the set of all partitions of
$f_\edge$ into $k$ blocks and let
$Z_\edge^k = \sum_{\pi_\edge\in\Pi_\edge^k} Z_\edge(\pi_\edge)$.
From Lemma~\ref{lem:wiring}
we deduce that
\begin{equation}
\label{eq:usewiring}
\frac{Z_\edge^k}{Z_\edge^{k'}} =
\frac{\Pr(Y(A)=k)}{\Pr(Y(A)=k')},
\end{equation}
where $A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$
for the function~$p$ from Section~\ref{sec:gadget}.
Denote by $\bot_\edge$ the finest partition (with $\terms$ blocks) of
$f_\edge$
and by $\top_\edge$ the coarsest partition (with one block)
on the same set.
By Lemma~\ref{lem:computerho} Equation~(\ref{eq:compbalance})
and Equation~(\ref{eq:usewiring})
we have
\begin{equation}
\label{chione}
e^{-\chi} \gamma \leq
\frac
{Z_\edge(\top_\edge)}
{Z_\edge(\bot_\edge)} \leq e^{\chi} \gamma.
\end{equation}
Use the FPRAS from Lemma~\ref{lem:computeZ}
(again, suitably powered up so that its failure probability is at most~$3/8$)
to compute a value~$\lastc$
which satisfies
\begin{equation}
\label{chitwo}
e^{-\chi} Z_\edge(\bot_\edge) \leq \lastc \leq e^{\chi} Z_\edge(\bot_\edge).
\end{equation}
Now, define $b_\edge(0)=\bot_\edge$ and $b_\edge(1)=\top_\edge$.
Then, by restricting
the sum in (\ref{eq:ZPottsbasic}),
to terms satisfying $\pi_\edge\in\{\bot_\edge,\top_\edge\}$ for all $\edge$,
we get
$$\ZTutte(\widehat G;q,\boldgamma')\geq
\sum_{z\in\{0,1\}^m}
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))}.$$
Letting $\|z\|$ denote the Hamming weight of $z$
and using Equation~(\ref{chione}),
the right-hand side is at least
$$ {Z(\bot_\edge)}^m e^{-\chi m}\sum_{z\in\{0,1\}^m}\notag
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))} {\gamma}^{\|z\|}.
$$
Using Equation~(\ref{chitwo}), we get
\begin{align}
\ZTutte(\widehat G;q,\boldgamma')&\geq \label{eq:ineq1}
\lastc^m e^{-2\chi m}\sum_{z\in\{0,1\}^m}
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))} {\gamma}^{\|z\|}\\
&=\lastc^m e^{-2\chi m}\,\ZTutte(H;q,\gamma).\notag
\end{align}
To finish, we just have to show that
this lower-bound for $\ZTutte(\widehat G;q,\boldgamma')$ is actually a good
estimate
because
the terms that we threw away
in inequality~(\ref{eq:ineq1}) don't amount to much.
For $\edge\in[m]$, let
$\Pi_\edge^1=\{\top_\edge\}$ and let
$\Pi_\edge^0=\Pi_\edge\setminus\{\top_\edge\}$.
Then, starting from (\ref{eq:ZPottsbasic}),
$$\ZTutte(\widehat G;q,\boldgamma')
= \sum_{z\in\{0,1\}^m}
\sum_{\textstyle{\pi_\edge \in\Pi_\edge^{z_\edge} \atop \forall \edge \in[m]}}\>
q^{\kappa(\pi_1\vee\cdots\vee\pi_m)}
\prod_{\edge=1}^m Z_\edge(\pi_\edge).$$
Now note the partition $\pi_\edge \in\Pi_\edge^{z_\edge} $ is refined by
$b_\edge(z_\edge)$. Thus,
$b_1(z_1)\vee\cdots\vee b_m(z_m)$ has
at least as many connected
components as $\pi_1\vee\cdots\vee\pi_m$.
So we get
\begin{align*}\ZTutte(\widehat G;q,\boldgamma')
&\leq \sum_{z\in\{0,1\}^m}
\sum_{\textstyle{\pi_\edge \in\Pi_\edge^{z_\edge} \atop \forall \edge \in[m]}}\>
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))}
\prod_{\edge=1}^m Z_\edge(\pi_\edge)\\
&=\sum_{z\in\{0,1\}^m}\>
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))}
\sum_{\textstyle{\pi_\edge \in\Pi_\edge^{z_\edge} \atop \forall \edge \in[m]}}\>
\prod_{\edge=1}^m Z_\edge(\pi_\edge)\notag\\
&=\sum_{z\in\{0,1\}^m}\>
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))}
\prod_{\edge=1}^m Z_\edge(\Pi_\edge^{z_\edge}),\notag\\
\end{align*}
where $Z_\edge(\Pi_\edge^{z_\edge}) =
\sum_{\pi_\edge \in\Pi_\edge^{z_\edge}}
Z_\edge(\pi_\edge)$.
Now, $Z_\edge(\Pi_\edge^1)=Z_\edge(\top_\edge) \leq e^{2\chi} \gamma \lastc$.
Also, $Z_\edge(\Pi_\edge^0) = \sum_{k=2}^\terms Z_\edge^k $.
By Equation~(\ref{eq:usewiring}), this is
$$Z_j^\terms\left(1 + \frac{\sum_{k=2}^{\terms-1} \Pr(Y(A)=k)}
{\Pr(Y(A) = \terms)}\right) ,$$
where $A$ is drawn from $\mathrm{RC}(\varGamma;q,p)$
for the function~$p$ from Section~\ref{sec:gadget}.
Since $Z_j^\terms\leq e^{\chi}\lastc$, we have
$$Z_\edge(\Pi_\edge^0)
\leq e^{\chi} \lastc \left(1 + \frac{\sum_{k=2}^{\terms-1}
\Pr(Y(A)=k)}{\Pr(Y(A)=\terms)}\right).$$
Now by Lemma~\ref{lem:excludedmiddletwo}
Equation~(\ref{eq:dichotomy}),
$\sum_{k=2}^{\terms-1}
\Pr(Y(A)=k)\leq \eta$.
Thus, $\Pr(Y(A)=1)+\Pr(Y(A)=\terms)\geq 1-\eta$ and
by Equation~(\ref{eq:compbalance}),
$$\Pr(Y(A)=\terms)\geq \frac{1-\eta}{1+e^{\chi}\gamma}.$$
Thus,
$$
Z_\edge(\Pi_\edge^0)
\leq e^{\chi}\lastc
\left(1+ \frac{\eta(1+e^{\chi}\gamma)}{1-\eta}\right)
\leq e^{2 \chi} \lastc
.$$
Thus,
$$\ZTutte(\widehat G;q,\boldgamma') \leq
e^{2 \chi m} \lastc^m
\sum_{z\in\{0,1\}^m}\>
q^{\kappa(b_1(z_1)\vee\cdots\vee b_m(z_m))}
\gamma^{\|z\|} = e^{2 \chi m} \lastc^m \ZTutte(H;q,\gamma)
.$$
Thus, if we knew a quantity $\psi$ in the
range
$$e^{-\epsilon/2} \leq \frac{\psi}{\ZTutte(\widehat G;q,\boldgamma')}
\leq e^{\epsilon/2},$$
we would have
$$e^{-\epsilon} \leq \frac{\psi \lastc^{-m}}{\ZTutte(H;q,\gamma) }
\leq e^{\epsilon},$$
so we have completed the AP-reduction.
\end{proof}
\section{Approximately computing the Tutte polynomial of a graph}
\label{shift}
In this section, we complete the approximation-preserving reduction
from~\BIS\ to $\textsc{Tutte}(q,\gamma)$
by giving an approximation-preserving reduction from
the problem
$\TwoWeightFerroTutte(q)$ to
the problem
$\textsc{Tutte}(q,\gamma)$.
For this, we first need to define what it means to ``implement'' an
edge weight. This description is mainly taken from~\cite[Section~1.6]{planartutte}.
Fix $q>2$.
Let
$W$ be a set of edge weights.
For example, $W$ might contain the edge weight
$\gamma$ (a parameter of the problem $\textsc{Tutte}(q,\gamma)$) or
it might contain a selection of edge weights that we have already
implemented using~$\gamma$.
Let $\gamma^*$ be a weight (which may not be in $W$) which we want to ``implement''.
Suppose that
there is a graph~$\Upsilon$,
with distinguished vertices $s$ and~$t$
and an edge-weight function $\hat\boldgamma: E(\Upsilon) \rightarrow
W$
such that
\begin{equation}
\label{eq:implement}
\gamma^* = \frac{q Z_{st}(\Upsilon)}{Z_{s|t}(\Upsilon)},
\end{equation}
where $Z_{st}(\Upsilon)$ denotes the contribution to
$\ZTutte(\Upsilon;q,\hat\boldgamma)$ arising from edge-sets $A$ in which $s$ and $t$ are
in the same component.
That is,
$Z_{st}(\Upsilon) = \sum_{A} \hat\boldgamma(A) q^{\kappa(V,A)}$,
where the sum is over subsets $A\subseteq E(\Upsilon)$ in which
$s$ and $t$ are in the same component.
Similarly, $Z_{s|t}$ denotes the contribution to
$\ZTutte(\Upsilon;q,\hat\boldgamma)$ arising from edge-sets $A$ in which $s$ and $t$ are in different components.
In this case, we say that $\Upsilon$ and $\hat\boldgamma$ implement
$\gamma^*$
(or even that $W$ implements $\gamma^*$).
The purpose of ``implementing'' edge weights is this.
Let $G$ be a graph with edge-weight function $\boldgamma$.
Let $f$ be some edge of $G$ with edge weight $\gamma_f=\gamma^*$.
Suppose that $W$ implements $\gamma^*$.
Let $\Upsilon$ be a graph with distinguished vertices $s$ and $t$
with a weight function $\hat\boldgamma$ satisfying (\ref{eq:implement}).
Construct the weighted graph $\widetilde G$
by replacing edge $f$ with a copy of $\Upsilon$ (identify $s$ with either endpoint of $f$
(it doesn't matter which one) and identify $t$ with the other endpoint of $f$ and remove edge $f$).
Define the weight function $\tilde\boldgamma$ as follows.
$$\tilde\gamma_e = \begin{cases}
\hat\gamma_e, &\text{if $e\in E(\Upsilon)$, and}\\
\gamma_e, & \text{otherwise.}
\end{cases}$$
Then the definition of the multivariate Tutte polynomial gives
\begin{equation}
\label{eq:shift}
\ZTutte(\widetilde G;q,\tilde\boldgamma) = \frac{Z_{s|t}(\Upsilon)}{q^2} \ZTutte(G;q,\boldgamma).\end{equation}
So, as long as $Z_{s|t}(\Upsilon)$ is easy to evaluate
and $q$ is efficiently approximable,
approximating the multivariate Tutte polynomial of $\widetilde G$ with weight function $\tilde\boldgamma$ is
essentially the same as approximating the multivariate Tutte polynomial of $G$ with weight function~$\boldgamma$.
Two especially useful implementations are series and parallel compositions.
These are explained in detail in \cite[Section 2.3]{JacksonSokal}.
So we will be brief here.
Parallel composition is the case in which $\Upsilon$ consists of two parallel edges $e_1$ and $e_2$
with endpoints $s$ and $t$ and $\hat\gamma_{e_1}=\gamma_1$ and $\hat\gamma_{e_2}=\gamma_2$.
It is easily checked from Equation~(\ref{eq:implement})
that $\gamma^* = (1+\gamma_1)(1+\gamma_2)-1$. Also, the extra factor in Equation~(\ref{eq:shift}) cancels,
so in this case $\ZTutte(\widetilde G;q,\tilde\boldgamma) = \ZTutte(G;q,\boldgamma)$.
Series composition is the case in which $\Upsilon$ is a length-2 path from $s$ to $t$ consisting of edges $e_1$ and $e_2$
with $\hat\gamma_{e_1}=\gamma_1$ and $\hat\gamma_{e_2}=\gamma_2$.
It is easily checked from Equation~(\ref{eq:implement})
that $w^* = \gamma_1\gamma_2/(q+\gamma_1+\gamma_2)$.
Also, the extra factor in Equation~(\ref{eq:shift}) is $q+\gamma_1+\gamma_2$,
so in this case $\ZTutte(\widetilde G;q,\tilde\boldgamma) = (q+\gamma_1+\gamma_2) \ZTutte(G;q,\boldgamma)$.
It is helpful to note that
$\gamma^*$ satisfies
$$\left(1+\frac{q}{\gamma^*}\right) = \left(1+\frac{q}{\gamma_1}\right) \left(1+\frac{q}{\gamma_2}\right).$$
We are now ready to prove this lemma.
\begin{lemma}
\label{lem:shiftaround}
Suppose that
$q>2$ and $\gamma>0$ are efficiently approximable. Then
$\TwoWeightFerroTutte(q) \leq_\mathrm{AP} \textsc{Tutte}(q,\gamma)$.
\end{lemma}
\begin{proof}
Let $G=(V,E)$ be an instance of
$\TwoWeightFerroTutte(q)$
with edge-weight function $\boldgamma': E \rightarrow
\{\gamma',\gamma''\}$
where $\gamma'$ and $\gamma''$ are rationals in the interval $[|V|^{-3},1]$.
We will assume without loss of generality that $|E|$ is sufficiently
large with respect to the fixed parameters~$q$ and~$\gamma$.
let $\epsilon$ be the desired accuracy in the approximation-preserving reduction.
Let
$\chi = \epsilon/(4(|V|+|E|^2))$
Let $\hat q$ be a rational in the range
$e^{-\chi} q \leq \hat q \leq e^{\chi} q$ and
let $\hat\gamma$ be a rational in the range
$e^{-\chi} \gamma \leq \hat \gamma \leq e^{\chi} \gamma$.
Since $q$ and $\gamma$ are efficiently approximable, the amount of time
that it takes to compute $\hat q$ and $\hat \gamma$
is at most a polynomial in $|V|$, $|E|$ and $\varepsilon^{-1}$.
The idea of the proof is to show how to use
series and parallel compositions from
the set $W=\{\hat \gamma\}$ to
implement edge-weights ${\gamma'}^*$
and ${\gamma''}^*$
satisfying
\begin{equation}
\label{eq:finalfinal}
e^{- \chi} {\gamma'} \leq {\gamma'}^* \leq
e^{ \chi} \gamma'
\end{equation}
and
$$e^{- \chi} {\gamma''} \leq {\gamma''}^* \leq
e^{ \chi} \gamma''.$$
Letting $\boldgamma^*$ be the edge-weight function derived from~$\boldgamma'$ by
replacing $\gamma'$ with ${\gamma'}^*$ and ${\gamma''}$ with ${\gamma''}^*$,
Equation~(\ref{approxparams}) gives
$$
e^{-\varepsilon /4}
\ZTutte(G;q,\boldgamma') \leq
\ZTutte(G;\hat q,\boldgamma^*) \leq
e^{\varepsilon /4}
\ZTutte(G; q, \boldgamma').
$$
Let $\hat\boldgamma$ be the edge-weight function which assigns every edge weight $\hat\gamma$.
We can think of our implementations as constructing a graph $\widehat{G}$
such that
$\ZTutte(G;\hat q,\boldgamma^*)$ is equal to
the product of $\ZTutte(\widehat{G};\hat q,\hat \boldgamma)$
and an easily-computed function of~$\hat q$ and~$\hat \gamma$.\footnote{This easily-computed
function arises from the extra factor $Z_{st}(\Upsilon)$ in Equation~(\ref{eq:implement}). It is easy to compute
because our implementations use only series and parallel composition}
We will ensure that each implementation uses at most $|E|$ edges,
so the total number of edges in $\widehat{G}$ is at most $|E|^2$.
To finish, we note (from (\ref{approxparams})) that
$$e^{-\varepsilon/4 }
\ZTutte(\widehat{G};q,\boldgamma) \leq
\ZTutte(\widehat{G};\hat q,\hat \boldgamma) \leq
e^{\varepsilon /4}
\ZTutte(\widehat{G}; q, \boldgamma),$$
where $\boldgamma$ is the constant edge-weight function which assigns every edge weight $\gamma$.
We finish the approximation of
$ \ZTutte(\widehat{G};\hat q,\hat \boldgamma)$
by using the oracle to approximate
$\ZTutte(\widehat{G};q,\boldgamma) $ using accuracy parameter $\delta = \epsilon/2$.
It remains to show how to do the implementations.
Taking
$$\pi = \frac{\chi}{2|V|^3} \leq \frac{\gamma' \chi}{2} \leq \gamma'(1-e^{-\chi}),$$
We show how to use $W=\{\hat \gamma\}$
to implement an edge-weight
${\gamma'}^*$
which satisfies ${\gamma'} - \pi \leq {\gamma'}^* \leq {\gamma'}$.
This ensures that Equation~(\ref{eq:finalfinal}) holds.
The implementation of ${\gamma''}^*$ is similar.
Our implementation is taken from Section~2.1 of our
paper~\cite{planartutte}. First, we can
implement a weight $\gamma_1\leq\tfrac14$
by taking a series composition of $k$ edges of weight~$\hat\gamma$
for sufficiently large~$k$. It suffices to take
$$k = \left\lceil
\frac
{\log(1+4\hat q)}
{\log(1+\hat q/\hat \gamma)}
\right\rceil.$$
Then implement a weight $\gamma_j$ by taking a series composition of
$j$
copies of $\gamma_1$.
The following (recursive) definitions are
from~\cite[Section~2.1]{planartutte} for integers~$j\geq 1$.
$$d_j = \left\lfloor
\frac{\log((1+{\gamma'})
\prod_{\ell=1}^{j-1}{(1+\gamma_\ell)}^{-d_\ell}
)
}
{\log(1+\gamma_j)}
\right\rfloor, \mbox{ and }
m = \left\lceil
\frac
{\log( \hat q (1+{\gamma'}) /\pi+1)}{\log(
\hat q/\gamma_1+1)}\right\rceil.
$$
Then the implementation combines, in parallel, $d_j$ edges with
edge-weight $\gamma_j$, for all $j\in[m]$.
The calculation in~\cite[Section~2.1]{planartutte}
shows that the implemented value~${\gamma'}^*$
satisfies ${\gamma'} - \pi \leq {\gamma'}^* \leq {\gamma'}$, as required.
Now to finish we need to show that $d_1 + \cdots + d_m \leq |E|$, which we
used above.
First, note that the fixed parameters~$q$ and~$\gamma$
give fixed upper and lower bounds on
$\hat q$ and $\hat \gamma$.
Using the upper bound $\gamma'\leq 1$, we see that
the value $m$ is
at most
logarithmic in $\pi^{-1}$ which is at most logarithmic in $|V|$,
$|E|$, and
$\varepsilon^{-1}$.
The calculation in~\cite[Section~2.1]{planartutte} shows that the same
is true of $d_1,\ldots,d_m$. In fact, there is a fixed upper bound
for~$d_j$ depending only on~$\hat q$ and~$\hat \gamma$. The proof of this fact uses the
fact that $0< \gamma_j \leq \gamma_1 \leq \tfrac14$.
Since we assumed, without loss of generality,
that $|E|$ is sufficiently large with respect to the fixed parameters~$q$ and~$\gamma$,
we conclude that $d_1 + \cdots + d_m \leq |E|$, as required.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main}]
Theorem~\ref{thm:main} follows from Lemmas \ref{lem:bis}, \ref{lem:one},
\ref{lem:last} and~\ref{lem:shiftaround}.
\end{proof}
\section{$3$-Uniform Hypergraphs}
\label{sec:3uniform}
Lemmas~\ref{lem:bis}
and \ref{lem:one} have the following corollary.
\begin{corollary}
\label{cor:uhhard}
Suppose that
$q>0$ is efficiently approximable. Then
$\BIS \leq_\mathrm{AP}\textsc{UniformHyperTutte}(q,q-1).$
\end{corollary}
Thus, assuming that there is no FPRAS for \BIS, we can conclude that there is no
FPRAS for computing the Tutte polynomial
of a uniform hypergraph when the edge-weights are set to $\gamma=q-1$.
The \emph{Ising model} corresponds to the $q=2$ case of the Potts model.
Thus, we conclude that
there is no FPRAS for computing the partition function of
the Ising model on a uniform hypergraph in which every edge has weight~$1$.
We conclude this paper with a contrasting positive result for $3$-uniform hypergraphs.
Consider the following problem.
\begin{description}
\item[Problem] $\textsc{$3$-UniformHyperTutte}(q,\gamma)$
\item[Instance] A $3$-uniform hypergraph $H=(\calV,\calE)$.
\item[Output] $\ZTutte(H;q,\gamma)$,
where $\boldgamma$ is the constant function with $\boldgamma_f = \gamma$ for every $f\in\calE$.
\end{description}
\begin{lemma}
Suppose that
$\gamma>0$ is efficiently approximable.
There is an FPRAS for $\textsc{$3$-UniformHyperTutte}(2,\gamma)$.
\end{lemma}
\begin{proof}
Jerrum and
Sinclair~\cite{JS93}
have given an FPRAS for $\textsc{Tutte}(2,\gamma')$ for every $\gamma'>0$.
We will give a reduction from~$\textsc{$3$-UniformHyperTutte}(2,\gamma)$
to $\textsc{Tutte}(2,\gamma')$ where $\gamma' = {(1+\gamma)}^{1/2}-1$.
Let $H=(\calV,\calE)$ be a
$3$-uniform hypergraph, an instance of
the target problem
$\textsc{$3$-UniformHyperTutte}(2,\gamma)$.
Let $\boldgamma$ be the constant function with $\boldgamma_f = \gamma$ for every $f\in\calE$.
Let $y=\gamma+1$.
Now, by Observation~\ref{obs:FK},
$$\ZTutte(H;2,\boldgamma) = Z_\mathrm{Potts}(H;2,\boldgamma) =
\sum_{\sigma:\calV\rightarrow \{0,1\}}
y^{\mathrm{mono}(\sigma)},$$
where $\mathrm{mono}(\sigma)$ denotes the number of hyperedges $f\in\calE$
that are monochromatic in configuration~$\sigma$.
Construct a (multi-)graph~$G$ with vertex set~$\calV$
and edge set
$$E = \bigcup_{(u,v,w)\in \calE}
\{(u,v),(v,w),(u,w)\}.$$
Let $\boldgamma'$ be the constant function with $\boldgamma'_\edge = \gamma'$ for every $\edge\inE$.
Let $y'=\gamma'+1 = y^{1/2}$.
Now if a hyperedge $f\in \calE$ is monochromatic in~$\sigma$, it contributes
${y'}^3$ to the corresponding term in $Z_\mathrm{Potts}(G;2,\boldgamma')$.
Otherwise, it contributes $y'$ to the term.
Thus, $$\ZTutte(G;2,\boldgamma)=
Z_\mathrm{Potts}(G;2,\boldgamma') = {y'}^{|\calE|}Z_\mathrm{Potts}(H;2,\boldgamma),$$
which completes the proof.
\end{proof}
\bibliographystyle{plain}
|
\section*{Introduction}
\noindent The famous Monty Hall Problem as quoted by Marilyn vos Savant in her ``Ask Marilyn'' column in \emph{Parade} magazine (p.~16, 9 September 1990), as a problem posed to her by a correspondent Mr. Craig Whitaker, was the following:
\begin{quote}
Suppose you're on a game show, and you're given the choice of three doors: Behind one door is a car; behind the others, goats. You pick a door, say No.~1, and the host, who knows what's behind the doors, opens another door, say No. 3, which has a goat. He then says to you, ``Do you want to pick door No.~2?'' Is it to your advantage to switch your choice?
\end{quote}
The literature on this problem is enormous, and at the end of this paper I simply cite the references which for me have been most valuable: the 2008 paper by Jeff Rosenthal and the 2009 book by Jason Rosenhouse. The latter has a huge reference list and discusses the pre-history and the post-history of vos Savant's problem, as well as many variants. Aside from these two references, the wikipedia discussion pages (on English and on Dutch wikipedia) are a fabulous though every-changing resource, and everything that I write here was learnt from those pages.
Despite making homage here to both cited authors for their wonderful work, I emphasize that I strongly disagree with both Rosenhouse (``the canonical problem'') and Rosenthal (``the original problem'') on what the essential Monty Hall problem is. I am more angry with certain other authors, who will remain nameless but for the sake of argument I'll just call Morgan et al, for unilaterally declaring their Monty Hall problem to be the only possible sensible problem, for arrogantly calling everyone who solved different problems stupid, and for getting an incorrect theorem published in the peer reviewed literature.
But: deciding unilaterally (Rosenhouse, 2009) that a certain formulation is \emph{canonical} is asking for a schism and an excommunication. And calling a version \emph{original} (Rosenthal, 2008) without historical research is asking for a contradiction (in view of the pre-history of the problem, the notion is not well defined). Monty Hall is part of folk-culture, culture is alive, the Monty Hall problem is not owned by a particular kind of mathematician who looks at such a problem from a particular point of view, adding for them ``natural'' extra assumptions which merely have the role of allowing their solution to work. Thus any canonical Monty Hall problem is actually an example of \emph{solution driven science} --- you have learnt a clever trick and want to show that it solves lots of problems.
I will distinguish three different Monty Hall problems. One of them is simply to answer the question literally posed by Marilyn vos Savant, ``would you switch?''. The other two are popular mathematizations, particularly among experts or teachers of elementary probability theory: one asks for the unconditional probability that switching gets the car, the other asks for a conditional probability, given the choices made so far:
\begin{itemize}
\item[Q-0:] Marilyn vos Savant's question ``should you switch?''
\item[Q-1:] A mathematician's question ``what is the unconditional probability that switching gives the car?''
\item[Q-2:] A mathematician's question ``what is the conditional probability that switching gives the car?''
\end{itemize}
The free internet (and freely editable) encyclopedia Wikipedia is the scene of a furious debate as to which mathematization Q-1 or Q-2 is the right starting point for answering the original verbal question Q-0; and after that, what supplementary conditions are obviously implicitly being made (the assumptions which ensure that the question has a nice answer). My own humble opinion is ``neither, though the unconditional approach comes closer; and no supplementary conditions should be made''.
My analysis of the mathematical problems yields the good answer ``$2/3$'' for both unconditional and conditional probabilities, under minimal assumptions, and almost without computation or algebraic manipulation. I use Boris Tsirelson's proposal on the wikipedia talk pages to use symmetry to deduce the conditional probability from the unconditional one. (Boris graciously gave me permission to cite him here, but this should not be interpreted to mean that anything written here also has his approval). I finally use a game-theoretic point of view to answer the original question posed by Marilyn vos Savant: if you were the player, should you switch doors? I make one supplementary assumption: that you know game theory, and therefore in advance have randomized your choice of door.
Let the three doors be numbered in advance $1$, $2$, and $3$. Introduce four random variables taking values in the set of door-numbers $\{1,2,3\}$:
\begin{itemize}
\item[]$C$: the quiz-team hides the Car behind door $C$,
\item[]$P$: the Player chooses door $P$,
\item[]$Q$: the Quizmaster opens door $Q$,
\item[]$S$: he asks the player if she'd like to Switch to door $S$.
\end{itemize}
Because of the standard story of the Monty Hall show, we certainly have:
\begin{itemize}
\item[] $Q \ne P$, the quizmaster opens a door different to the player's first choice,
\item[] $Q \ne C$, opening that door reveals a goat,
\item[] $S \ne P$, the player is invited to switch to another door,
\item[] $S \ne Q$, no player wants to go home with a goat.
\end{itemize}
It does not matter for mathematical analysis whether probabilities are subjective (Bayesian) or objective (frequentist); nor does it matter whose probabilities they are supposed to be, at what stage of the game. Some writers think of the player's initial choice as fixed. For them, $P$ is degenerate. I simply write down some assumptions and deduce consequences of them.
\section{Solution to problem 1:\\ unconditional chance that switching wins}
By the rules of the game and the definition of $S$, if $P\ne C$ then $S =C$, and vice-versa. A ``switcher'' would win the car if and only if a ``stayer'' would lose it. Therefore:
\begin{quotation}
\noindent \emph{If ${\mathrm{Prob}}(P=C)=1/3$ then ${\mathrm{Prob}}(S=C)=2/3$, since the two events are complementary.}
\end{quotation}
\section{Solution to problem 2:\\ conditional chance that switching wins}
Take the door chosen by the player as fixed, $P\equiv x$, say. We are to compute ${\mathrm{Prob}}(S=C|Q=y)$ for a further specific value $y\ne x$. Let $y'$ denote the remaining door-number, besides $x$ and $y$. \emph{Assume that all doors are equally likely to hide the car and assume that the quizmaster chooses completely at random when he has a choice}: $C$ is uniform and the distribution of $Q$ given $C$ is uniform on the available possibilities.
From the first assumption
$${\mathrm{Prob}}(P=C)~=~{\mathrm{Prob}}(C=x)=1/3$$
and therefore by the solution to Problem~1, ${\mathrm{Prob}}(S=C)=2/3$.
This unconditional probability is the weighted average of the two conditional probabilities ${\mathrm{Prob}}(S=C|Q=y)$ and ${\mathrm{Prob}}(S=C|Q=y')$, weighted by the unconditional probabilities ${\mathrm{Prob}}(Q=y)$ and ${\mathrm{Prob}}(Q=y')$. The distribution of $C$ is unchanged on exchanging $y$ and $y'$. Regarding the set of conditional laws of $Q$ given each value of $C$,
$${\mathrm{Prob}}(Q=y|C=x)~=~{\mathrm{Prob}}(Q=y'|C=x)~=~1/2,$$
$${\mathrm{Prob}}(Q=y'|C=y)~=~{\mathrm{Prob}}(Q=y|C=y')~=~1.$$
This set of conditional laws is symmetric on interchange of $y$ and $y'$. Finally, regarding the set of conditional laws of $S$ given $C$ and $Q$, $S$ is simply defined as the unique door different from $Q$ and $x$. This definition too is symmetric under exchange of $y$ and $y'$.
\begin{quotation}
\noindent \emph{If $P$ is fixed, $C$ is uniform, and $Q$ is symmetric, then that switching gives the car is independent of the quizmaster's choice:}
\end{quotation}
$${\mathrm{Prob}}(S=C|Q=y)~=~{\mathrm{Prob}}(S=C|Q=y')~=~{\mathrm{Prob}}(S=C)~=~2/3.$$
Instead of starting with fixed $P$, one can \emph{without loss of generality pretend} that is random and uniform, and add the uniformity of $C$ given $P$ and that of $Q$ given $C$ and $P$. The joint law of $C,P,Q,S$ is now invariant under renumberings of the three doors. Hence ${\mathrm{Prob}}(S=C|P=x,Q=y)$ is the same for all $x\ne y$ and hence equal to the unconditional ${\mathrm{Prob}}(S=C)=2/3$.
Some readers may prefer a direct calculation. Using Bayes' theorem in the form ``posterior odds equal prior odds times likelihoods'' is a particularly efficient way to do this. The probabilities and conditional probabilities below are all conditional on $P=1$, or if your prefer with $P\equiv 1$.
We have uniform prior odds $${\mathrm{Prob}}(C=1):{\mathrm{Prob}}(C=2):{\mathrm{Prob}}(C=3)~=~1:1:1.$$
The likelihood for $C$, the location of the car, given data $Q=3$, is (proportional to) the discrete density function of $Q$ given $C$ (and $P$)
$${\mathrm{Prob}}(Q=3|C=1):{\mathrm{Prob}}(Q=3|C=2):{\mathrm{Prob}}(Q=3|C=3)~=~\frac 1 2:1:0.$$
The posterior odds are therefore proportional to the likelihood. It follows that the posterior probabilities are
$${\mathrm{Prob}}(Q=3|C=1)=\frac 13,~ {\mathrm{Prob}}(Q=3|C=2)=\frac 2 3,~ {\mathrm{Prob}}(Q=3|C=3)=0.$$
\section{Answer to Marylin vos Savant: \\
would I switch doors?}
\noindent Yes. Recall, \emph{I only know that Monty Hall always opens a door revealing a goat}. I didn't know what strategy the quiz-team and quizmaster were going to use for their choices of the distribution of $C$ and the distribution of $Q$ given $P$ and $C$, so naturally I had picked my door uniformly at random. My strategy of choosing $C$ uniformly at random guarantees that ${\mathrm{Prob}}(C=P)=1/3$ and hence that ${\mathrm{Prob}}(S=C)=2/3$.
In fact, I know elementary game theory as well as elementary probability theory, so it was easy for me to find out that this combined strategy, which I'll call ``symmetrize and switch'', is my minimax strategy.
On the one hand, ``symmetrize and switch'' guarantees me a $2/3$ (unconditional) chance of winning the car, whatever strategy used by the quizmaster and his team.
On the other hand, if the quizmaster and his team use their ``symmetric'' strategy ``hide the car uniformly at random and toss a fair coin to open a door if there is choice'', then I cannot win the car with a \emph{better} probability than $2/3$.
The fact that my ``symmetrize and switch'' strategy gives me ``at least'' $2/3$, while the quizmaster's ``symmetry'' strategy prevents me from doing better, proves that these are our respective minimax strategies, and $2/3$ is the game-theoretic value of this two-party zero-sum game.
There is not much point for me in worrying about my conditional probability of winning given my specific initial choice and the specific door opened by the quizmaster, say doors $1$ and $3$ respectively. I don't know this conditional probability anyway, since I don't know the strategy used by quiz-team and the quizmaster. (Even though I know probability theory and game theory, they maybe don't). However, it is maybe comforting to learn, by easy calculation, that if the car is hidden uniformly at random, then the conditional probability cannot be \emph{smaller} than $1/2$. So in that case at least, it certainly never \emph{hurts} to switch door.
\section{Discussion}
Above I tried to give short clear mathematical solutions to three mathematical problems. Two of them were problems of elementary probability theory, the third is a problem of elementary game theory. As such, it involves not much more than elementary probability theory and the beautiful minimax theorem of John von Neumann (1928). That a finite two-party zero-sum game has a saddle-point, or in other words, that the two parties in such a game have matching minimax strategies (if randomization is allowed), is not obvious. It seems to me that probabilists ought to know more about game theory, since every ordinary non-mathematician who hears about the problem starts to wonder whether the quiz-master is trying to cheat the player, leading to an infinite regress: if I know that he knows that I know that....
It am told that the literature of mathematical economics and of game theory is full of Monty Hall examples, but no-one can give me a nice reference to a nice game-theoretic solution of the problem. Probably game-theorists like to keep their clever ideas to themselves, so as to make money from playing the game. Only losers write books explaining how the reader could make money from game theory.
It would certainly be interesting to investigate more complex game-theoretic versions of the problem. If we take Monty Hall as a separate player to the TV station, and note that TV ratings are probably helped if nice players win while annoying players lose, we leave elementary game theory and must learn the theory of Nash equilibria.
Then there is a sociological or historical question: who ``owns'' the Monty Hall problem? I think the answer is obvious: no-one. A beautiful mathematical paradox, once launched into the real world, lives it own life, it evolves, it is re-evaluated by generation after generation. This point of view actually makes me believe that Question 0: \emph{would you switch} is the right question, and no further information should be given beyond the fact that you know that the quizmaster knows where the car is hidden, and always opens a door exhibiting a goat. Question 0 is a question you can ask a non-mathematician at a party, and if they have not heard of the problem before, they'll give the wrong answer (or rather, one of the two wrong answers: \emph{no} because nothing is changed, or \emph{it doesn't matter} because it's now 50--50). My mother, who was one of Turing's computers at Bletchley Park during the war, but who had had almost no schooling and in particular never learnt any mathematics, is the only person I know who immediately said: \emph{switch}, by immediate intuitive consideration of the 100-door variant of the problem. The problem is a \emph{paradox} since you can next immediately convince anyone (except lawyers, as was shown by an experiment in Nijmegen), that their initial answer is wrong.
The mathematizations Questions 1 and 2 are not (in my humble opinion!) \emph{the} Monty Hall problem; they are questions which probabilists might ask, anxious to show off Bayes' theorem or whatever. Some people intuitively try to answer Question 0 via Questions 1 and 2; that is natural, I do admit. And sometimes people become very confused when they realize that the answer to Question 2 can only be given its pretty answer ``2/3'' under further conditions. It is interesting how in the pedagogical mathematical literature, the further conditions are as it were held under your nose, e.g. by saying ``three \emph{identical} doors'', or replacing Marilyn's "say, door 1'' by the more emphatic ``door 1''.
It seems to me that adding into the question explicitly the remarks that the three doors are equally likely to hide the car, and that when the quizmaster has a choice he secretly tosses a fair coin to decide, convert this beautiful paradox into a probability puzzle with little appeal any more to non experts.
It also converts the problem into one version of the three prisoner's paradox. The three prisoners problem is isomorphic to the conditional probabilistic three doors problem. I always found it a bit silly and not very interesting, but possibly that problem too should be approached from a sophisticated game theoretic point of view.
By the way, Marilyn vos Savant's original question is semantically ambiguous, though this might not be noticed by a non-native English speaker. Are the mentioned door numbers, huge painted numbers on the front of the doors \emph{a priori}, or are we just for convenience \emph{naming} the doors by the choices of the actors in our game \emph{a posteriori}. Marilyn stated in a later column in \emph{Parade} that she had originally been thinking of the latter. However, her own offered solutions are not consistent with a single unambiguous formulation. Probably she did not find the difference very interesting.
This little paper containing nothing new, and only almost trivial mathematics, is a plea for future generations to preserve the life of The True Monty Hall paradox, and not let themselves be misled by probability purists who say ``you \emph{must} compute a conditional probability''.
\section*{References} ~
\vskip -0.5cm
\raggedright
\frenchspacing
\parskip 0.2 cm
\leftskip 0.5 cm
\parindent -0.5 cm
Rosenhouse, Jason (2009), \emph{The Monty Hall Problem}, Oxford University Press.
Rosenthal, Jeffrey S. (2008), Monty Hall, Monty Fall, Monty Crawl, \emph{Math Horizons}, September 2008, 5--7. Reprint: \texttt{http://probability.ca/jeff/writing/montyfall.pdf}
\end{document}
\section{Introduction}
\noindent In the above abstract to this paper, I reproduced The Monty Hall Problem, as it was defined by Marilyn vos Savant in her ``Ask Marilyn'' column in \emph{Parade} magazine (p.~16, 9 September 1990). Marilyn's solution to the problem posed to her by a correspondent Craig Whitaker sparked a controversy which brought the Monty Hall Problem to the attention of the whole world. Though MHP probably originated in a pair of short letters to the editor in \emph{The American Statistician} by biostatistician Steve Selvin (1975a,b), from 1990 on it was public property, and has sparked research and controversy in mathematical economics and game theory, quantum information theory, philosophy, psychology, ethology, and other fields, as well as having become a fixed point in the teaching of elementary statistics and probability.
This has resulted in an enormous literature on MHP. Here I would like to draw attention to the splendid book by Jason Rosenhouse (2009), which has a huge reference list and which discusses the pre-history and the post-history of vos Savant's problem as well as many variants. My other favourite is Rosenthal (2008), one of the few papers where a genuine attempt is made to argue to the layman why MHP has to be solved with conditional probability. Aside from these two references, the English language wikipedia page on the Monty Hall Problem, as well as its discussion page, is a rich though every-changing resource. Much that I write here was learnt from the many editors of those pages, both allies and enemies in the never ending edit-wars which plague it.
The battle among wikipedia editors could be described as a battle between intuitionists versus formalists, or to use other words, between simplists versus conditionalists. The main question which is endlessly discussed is whether simple arguments for switching, which typically show that the \emph{unconditional} probability that the switching gets the car is 2/3, may be considered rigorous and complete solutions of MHP. The opposing view is that vos Savant's question is only properly answered by study of ``the'' \emph{conditional} probability that the switching gets the car, \emph{given} the initial choice of door by the player and door opened by the host. This more sophisticated approach requires making more assumptions, and that leads to the question whether those supplementary conditions are implicitly implied by vos Savant's words. What particularly interests me, however, is that the conditionalists take on a dogmatic stance: their point of view is put forward as a moral imperative. This leads to an impasse, and the clouds of dust thrown up by what seems a religious war conceal what seem to me to be much more interesting, though more subtle, questions.
My personal opinion on the wikipedia-MHP-wars is that they are fights about the wrong question. Craig Whitaker, through the voice of Marilyn vos Savant, asks for an action, not a probability. I think that game theory gives a more suitable framework in which to represent our ignorance of the mechanics of the set-up (where the car is hidden) and of the mechanics of the host's choice, than subjectivist probability.
Therefore, though Rosenhouse's book is a wonderful resource, I strongly disagree with him, as well as with many other authors, on what the essential Monty Hall problem is (and that is the main point of this paper). Deciding unilaterally (Rosenhouse, 2009) that a certain formulation is \emph{canonical} is asking for schism. Calling a specific version \emph{original} (Rosenthal, 2008) is asking for contradiction. Rosenthal states without any argument at all that additional assumptions are implicitly contained in vos Savant's formulation. Selvin (1975a) did state all those assumptions explicitly but strangely enough did not use all of them. His second paper Sevin (1975b) gave a new solution using all his original assumptions but the author does not seem to notice the difference. At the same time, he quotes with approval a simplist solution of Monty Hall himself, who sees randomness in the choice of the player rather than in the actions of the team who prepare the show in advance, and the quiz-master himself. Vos Savant did not use the full set of assumptions which others find implicit in her question. Her relatively simple explanation of why one should switch seems to satisfy everyone except for the writers of elementary texts in statistics and probability. I have the impression that words like original, canonical, standard, complete are all used to hide the paucity of argument of the writer who needs to make that extra assumption in order to be able to apply the tool which they are particularly fond of, conditional probability.
One of the most widely cited but possibly least well read papers in MHP studies is Morgan et al.~(1991a), published together with a discussion by Seymann (1991) and a rejoinder Morgan et al.~(1991b). Morgan et al.~(1991a) firmly rebuke vos Savant for not solving Whitaker's problem as they consider should be done, namely by conditional probability. They use only the assumption that all doors are initially equally likely to hide the car; this assumption is hidden within their calculations. The paper was written during the peak of public interest and heated emotions about MHP which arose from vos Savant's column. It actually contains an unfortunate error which was only noticed 20 years later by wikipedia editors Hogbin and Nijdam (2010): if the player puts a non-informative and hence symmetric Bayesian prior on the host's bias in opening a door when he has a choice, it will be equally likely (for the player) that the host will open either door when he has the opportunity. Morgan et al.~(2010) acknowledge the error and also reproduce part of Craig Whitaker's original letter to Marilyn vos Savant whose wording is even more ambiguous than vos Savant's.
Rosenhouse (2009), Rosenthal (2005, 2008), Morgan et al.~(1991a,b, 2010), and Selvin (1975b) (but not Selvin, 1975a) solve MHP using elementary \emph{conditional} probability. In order to do so they are obliged to add mathematical assumptions to vos Savant's words, without which the conditional probability they are after is not determined. Actually, and I think tellingly, almost no author gives any argument at all why we \emph{must} solve vos Savant's question by computing a conditional probability that the other door hides the car, conditional on which door was first chosen by the player and which opened by the host.
For whatever reasons, it has become conventional in the elementary statistics literature, where MHP features as an exercise in the chapter on Bayes' theorem in discrete probability, to take it for granted that the car is initially hidden ``at random'', and the host's choice, if he is forced to make one, is ``at random'' too. Morgan et al.~(1991a) are notable in only making the first assumption. Many writers also have the player's initial choice ``at random'' too. ``At random'' is a code phrase for what I would prefer to call \emph{completely} at random. The student is apparently supposed to make these assumptions by default, though sometimes they are listed without motivation as if they are always the right assumptions to make.
In my opinion, this approach to MHP is an example of \emph{solution driven science}, and hence an example of bad practise in mathematical modelling. Taking for granted that unspecified probability distributions must be uniform or normal, depending on context, is the cause of such disasters as the miscarriage of justice concerning the Dutch nurse Lucia de Berk, or the doping case of the German skater Claudia Pechstein. Of course, MHP does indeed provide a nice exercise in conditional probability, provided one is willing to fill in gaps without which conditional probability does not help you answer the question whether you should stay or switch. Morgan et al. (1991a)'s original contribution is to notice the minimal condition under which conditional probability does give an unequivocal solution.
Precisely because of all these issues, MHP presents a beautiful playground for learning the art of mathematical modelling. For me, MHP is the problem of how to build a bridge from vos Savant's words to a mathematical problem, solve that problem, \emph{and} translate the solution back to the real world. \emph{If} we use probability as a tool in this enterprise, we are going to have to motivate probabilistic assumptions. We must also \emph{interpret} probabilistic \emph{conclusions}. Like it or not, the interpretation of probability plays a role, going both directions.
Real world problems are often brought to a statistician because the person with the question, for some reason or other, thinks the statistician must be able to help them. The client has often already left out some complicating factors, or made some simplifications, which he thinks that the statistician doesn't need to know. The first job of the consulting statistician is to find out what the real question is with which the client is struggling, which may often be very different from the imaginary statistical problem that the client thinks he has. The first job of the statistical consultant is to undo the pre-processing of the question which his client has done for him.
In mathematical model building we must be careful to distinguish the parts of the problem statement which are truly determined by the background real world problem, and parts which represent hidden assumptions of the client who thinks he needs to enlist the statistician's aid and therefore has already pre-processed the original question so as to fit in the client's picture of what a statistician can do. The result of a statistical consultation might often be that the original question posed by the client is reformulated, and the client goes home, happier, with a valuable answer to a more meaningful question than the one he brought to the statistician. Maybe this is the real message which the Monty Hall Problem should be telling us? What if vos Savant's opening words had been ``\emph{Suppose you're \emph{going to be} on a game show tonight. If you make it to the last round, you'll be given the choice of three doors...}''?
\section{The mathematical facts}
In this section, I present some elementary mathematical facts, firstly from probability theory, secondly from game theory. The results are formulated within a mathematical framework which does not make any assumptions restricting the scope of the present discussion. Modelling all the various door choices as random variables does not exclude the case that they are fixed. It also leaves the question completely open how we think of probability: in a frequentist or in a Bayesian sense. I impose only the ``structural'' conditions on the sequence of choices, or moves, which are universally agreed to be implied by vos Savant's story.
\subsection{What probability theory tells us}
I distinguish four consecutive actions:
\begin{enumerate}
\item Host: hiding the car before the show behind one of three doors, \textsf{Car}
\item Player: choosing a door, \textsf{P1}
\item Host: revealing a goat by opening a different door, \textsf{Goat}
\item Player: switching or staying, final choice door \textsf{P2}
\end{enumerate}
The doors are conventionally labelled ``1'', ``2'' and ``3'', and we can represent the door numbers specified by the four actions with random variables \textsf{Car, P1, Goat, P2}. Since two doors hide goats and one hides a car and the host knows the location of the car, he can and will open a door different to that chosen by the player and revealing a goat. I allow both the location of the car \textsf{Car} and the initial choice of the player \textsf{P1} to be random, and assume them to be statistically independent. From different modelling points of view, we might want to take either of these two variables to be fixed; the independence assumption is then of course harmless. Given the location of the car and the door chosen by the player, the host opens a different door \textsf{Goat} revealing a goat, according to a probability distribution over the two doors available to him when he has a choice (which includes the case that he follows some fixed selection rule). Then the player makes his choice \textsf{P2}, deterministically or according to a probability distribution if he likes, but in either case only depending on his initial choice and the door opened by the host. Finally we can see whether he goes home with a car or a goat by opening his door and revealing his \textsf{Prize}.
The probabilistic structure of the four actions together with the final result \textsf{Prize} (whether the player goes home with a car or a goat) can be represented in the graphical model or Bayes net shown in Figure 1. This diagram (drawn using the \textsf{gRain} package in the statistical package R) was inspired by Burns and Wieth (2004) who performed psychological experiments to test their hypothesis that people fail MHP because of their inability to internalise the \emph{collider principle}: conditional on a consequence, formerly independent causes become correlated. In this case, the initially statistically independent initial moves \textsf{Car} and \textsf{P1} of host and player are conditionally \emph{dependent} of one another given the door \textsf{Goat} opened by the host.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{MHP_dag.pdf}\\[0.2cm]
\caption{{A Graphical Model (Bayes Net) for MHP}}
\end{center}
\end{figure}
I now write down three simple propositions, each making in turn a stronger mathematical assumption, and each in turn giving a better reason for switching.
\begin{proposition} If the player's initial choice has probability $1/3$ to hit the car, then always switching gives the player the car with (unconditional) probability $2/3$ \emph{(Monty Hall, as reported by Selvin, 1975b)}.
\end{proposition}
\begin{proposition} If initially all doors are equally likely to hide the car, then, given the door initially chosen and the door opened, switching gives the player the car with conditional probability at least $1/2$ \emph{(Morgan et al., 1991a)}.
Consequently, not only does ``always switching'' give the player the car with unconditional probability $2/3$, but no other strategy gives a higher success chance.
\end{proposition}
\begin{proposition} If initially all doors are equally likely to hide the car and if the host is equally likely to open either door when he has a choice, then, given the door initially chosen and the door opened, switching gives the player the car with conditional probability $2/3$, whatever door was initially chosen and which door was opened \emph{(Morgan et al., 1991a,b).}
\end{proposition}
\begin{proof} ~
\\
\noindent \underline{Prop.~1}: This implication is trivial once we observe that a ``switcher'' wins if and only if a ``stayer'' loses.
\\
\\
\noindent \underline{Prop.~2}: We use Bayes' theorem in the form
$$\textit{posterior odds equals prior odds times likelihood ratio.}$$
The initial odds that the car is behind doors 1, 2 and 3 are 1:1:1. The posterior odds are therefore proportional to the probabilities that the host opens Door 3 given the player chose Door 1 and the car is behind Door 1, 2 and 3 respectively. These probabilities are $q$, $1$ and $0$ respectively, where
$$q~=~{\mathrm{Prob}}(\,\textrm{Host opens Door 3} \,|\, \textrm{Player chose Door 1, car is behind same}\,).$$
The posterior odds for switching versus staying are therefore $1:q$, so that staying does not have an advantage over switching, whatever $q$ might be.
Since all doors are initially equally likely to hide the car, the door chosen by the player hides the car with probability 1/3. The unconditional probability that switching gives the car is therefore 2/3. By the law of total probability, this can be expressed as the sum over all six conditions (door chosen by player, door opened by host), of the probability of that condition times the conditional probability that switching gives the car, under the condition. Each of these conditional probabilities is at least 1/2. The strategy of always switching can't be beaten, since the success probability of any other strategy is obtained from the success probability of always switching by replacing one or more of the conditional probabilities of getting the car by switching by probabilities which are never larger.
\\
\\
\noindent \underline{Prop.~3}: If all doors are equally likely to hide the car then by independence of the initial choice of the player and the location of the car, the probability that the initial choice is correct is $1/3$. Hence the unconditional probability that switching gives the car is $2/3$. If the player's initial choice is uniform and the two probability distributions involved in the host's choices are uniform, the problem is symmetric with respect to the numbering of the three doors. Hence the conditional probabilities we are after in Proposition 3 are all the same, hence by the law of total probability equal to the unconditional probability that switching gives the car, 2/3.
\end{proof}
Proposition 3 also follows from the inspection of the posterior odds computed in the proof of Proposition 2. On taking $q=1/2$, the posterior odds in favour of switching are 2:1 (Morgan et al., 1991a).
In the literature, Proposition 3 is usually proven by explicit computation or tabulation, i.e., by going back to first principles to compute the conditional probability in question. For instance, Morgan et al.~(1991a) also give this direct computation, attributing it to Mosteller's (1965) solution of the Prisoner's dilemma paradox. It is often offered as an example of Bayes' theorem, but really is just an illustration of conditional probability via its definition. On the other hand, Bayes' theorem in its odds form (which I used to prove Morgan et al.'s Proposition 2) is a genuine \emph{theorem}, and offers to my mind a much more satisying route for those who like to see a computation and at the same time learn an important concept and a powerful tool. To my mathematical mind the most elegant proof of Proposition 3 is the argument by symmetry starting from Proposition 1: the conditional probability is the same as the unconditional since all the conditional probabilities must be the same. I learnt this proof from Boris Tsirelson on wikipedia discussion pages, but it is also to be found in Morgan et al.~(1991b).
This proof also supplies one reason why the literature is so confused as to what constitutes a solution to MHP. One could apply symmetry at the outset, to argue that we only want an unconditional probability. There is no point in conditioning on anything which we can see in advance is irrelevant to the question at hand.
The pages of wikipedia, as well as a number of papers in the literature, are the scene of a furious controversy mainly as to whether Proposition 1 and a proof thereof, or Proposition 3 and a proof thereof, is a ``complete and correct solution to MHP''. These two solutions can be called the simple or popular or unconditional solution, and the full or complete or conditional solution respectively. The situation is further complicated by the fact that many supporters of the popular solution do accept all the symmetry (uniformity) conditions of Proposition 3, for a variety of reasons. I will come back to this in the next main section, but first consider a rather different kind of result which can be obtained within exactly the same general framework as before.
\subsection{What game theory tells us}
Let us think of the four actions of the previous subsection as two pairs of moves in a two stage game between the host and the player in which the player wants to get the car, the host wants to keep it. Von Neumann's minimax theorem tells us that there exist a pair of minimax strategies for player and host, and a value of the game, say $p$, having the following defining characteristics. The minimax strategy of the player (minimizes his maximum chance of losing) guarantees him at least probability $p$ of winning, whatever the strategy of the host; the minimax stategy of the host (minimizes his maximum probability of losing) guarantees him at most probability $p$ of losing. If both player and host play their minimax strategy then indeed the player will win with probability $p$.
\begin{proposition}The player's strategy ``initial choice uniformly at random, thereafter switch'' and the host's strategy ``hide the car uniformly at random, open a door uniformly at random when there is a choice'' form the minimax solution of the finite two-person zero-sum game in which the player tries to maximize his probability of getting the car, the host tries to minimize it. The value of the game is $2/3$.
\end{proposition}
\begin{proof}
We must verify two claims. The first is that whatever strategy is used by the host, the player's minimax strategy guarantees the player a success chance of at least 2/3. The second is that whatever strategy is used by the player, the host's minimax strategy prevents the player from achieving a success chance greater than 2/3.
For the first claim notice that if the player chooses a door uniformly at random and thereafter switches, he'll get the car with probability exactly 2/3; that follows from Proposition 1.
For the second, suppose the host hides the car uniformly at random and thereafter opens a door uniformly at random when he has a choice. Making the initial choice of door in any way, and thereafter switching, gives the player success chance 2/3, and by Proposition 2 (or 3, if you prefer) there is no way to improve this.
\end{proof}
Note that I did not use von Neumann's theorem in any way to get this result: I simply made use of the probabilistic results of the previous subsection.
MHP was brought to the attention of the mathematical economics and game theory community by a paper of Nalebuff (1987), which contains a number of game-theoretic or economics choice puzzles. He considered MHP as an amusing problem with which to while away a few minutes during a boring seminar. After describing the problem he very briefly reproduced the short solution corresponding to Proposition 1. He enigmatically drops the names of Neumann-Morgenstern and Bayes as he ponders why most people in real life took the wrong decision, but he does not waste any more time on MHP.
Variants of the MHP in which the host does not always open a door, or where he might be trying to help you, or might be trying to cheat you, lend themselves very well to game theoretic study, see wikipedia or Rosenhouse (2009) for references.
For present purposes, the important point which I think is brought out by a game theoretic approach is that the player does have two decision moments. The player has control over his initial choice. Vos Savant describes the situation at the moment the player must make up his mind whether to switch or stay, and most, but not all, people will instinctively feel that this is the only important decision moment. But the player earlier had a chance to choose any door he liked. Perhaps he would have been wise to think about what he would do if he did make it to the last round of the show, before setting off to the TV studio. There is no way he can ask the advice of a friendly mathematician as he stands on the stage under the dazzling studio lights while the audience is shouting out conflicting advice.
Van Damme (1995; in Dutch) goes a little deeper into the question of why real human players did not behave rationally on the Monty Hall show; this is one of the main questions studied in the psychology, philosophy, artificial intelligence and animal behaviour literature on MHP. Since ``rational expectations'' play a fundamental role in modern economic theory, the actual facts of the real world MHP, where players almost never switched doors, bodes ill for the application of economic theory to real world economics. The usual rationale for human rational expectations in economics is that humans learn from mistakes. However, the same person did not get to play several times times in the final round of the Monty Hall show, and apparently no-one kept a tally of what had happened to previous contestants, so learning simply did not take place. Nobody thought there would be a point in learning! Instead, the players used their brains, came to the conclusion that there was no advantage in switching, and mostly stuck to their original choice. At this point they do make a rational choice: there would be a much larger emotional loss to their ego on switching and losing, than on staying and losing. Sticking to your door demonstrates moral fortitude. Switching is feckless and deserves punishment.
Interestingly, pigeons (specifically, the Rock Pigeon, \emph{Columba Livia}, the pigeon which tourists feed in city squares all over the world) are very good at learning how to win the Monty Hall game, see Herbranson and Schroeder (2010). They do not burden their little brains thinking about what to do but just go ahead and choose. There is a lot of variation in their initial decisions whether to switch or stay, and observing the results gives them a chance to learn from the past without thinking at all. Only a very small brain is needed to learn the optimal strategy. And these birds are evolutionarily speaking very succesful indeed.
\section{Which assumptions?}
Propositions 1, 2 and 3 tell us in different ways that switching is a good thing. Notice that the mathematical conditions made are successively stronger and the conclusion drawn is successively stronger too. As the conditions get stronger, the scope of application of the result gets narrower: there are more assumptions to be justified. From a mathematical point of view none of these results are stronger than any of the others: they are all \emph{strictly different}.
The literature on MHP focusses on variants of Proposition 1, and of Proposition 3. These correspond to what are called the popular or simple or unconditional solution, and the full or conditional solution to MHP. The full solution is mainly to be found in introductory probability or statistics texts, whereas the simple solution is popular just about everywhere else. The intermediate ``Proposition 2'' is only occasionally mentioned in the literature. The full list of conditions in Proposition 3 is often called, at least in the kind of texts just mentioned, the standard or canonical or original MHP. I will just refer to them as the \emph{conventional supplementary assumptions}.
Regarding the word ``original'', it is a historical fact that Selvin (1975a) gave MHP its name, did this in a statistics journal, and wrote down the conventional full list of uniformity assumptions. He proceeded to compute the \emph{unconditional} probability that switching gives the car by enumeration of \emph{nine} equally likely cases, for which he took both the player's initial choice and the location of the car as uniform random, and of course independent of one another. In his second note, Selvin (1975b), he computed the \emph{conditional} probability using now his full list of assumptions concerning the host's behaviour, and fixing the initial choice of the player, but without noting any conceptual, let alone technical, difference at all with his earlier solution. Of course, the number ``2/3'' is the same. In the same note he quotes with approval from a letter from Monty Hall himself who gave the argument of Proposition 1: switching gives the car with probability 2/3 because the initial choice is right with probability 1/3. We know Monty will open a door revealing a goat. Conditioning on an event of probability one does not change the probability that the initial choice was right.
Thus Selvin set the seeds for subsequent confusion. Let me call his approach the \emph{practical-minded approach}:
\begin{quote} The \emph{right} answer to MHP is ``2/3''. There are many ways to get to this answer, and I am not too much concerned how you get there. As long as you get the right answer 2/3, we're happy. After all, the whole point of MHP is that the initial instinct of everyone hearing the problem is to say ``since the host could anyway open a door showing a goat, his action doesn't tell me anything. There are still two doors closed so they still are equally likely to hide the car. So the probability that switching would give the car is ``1/2'', so I am not going to switch, thank you.
\end{quote}
Selvin's two papers together gave MHP a firm and more or less standard position in the elementary statistics literature. There is a conventional complete specification of the problem. This enables us to write down a finite sample space and allocate a probability to every single outcome. Usually the player's initial choice is taken, in the light of the other assumptions without loss of generality, as fixed. All randomness is in the actions of the host, or \emph{in our lack of any knowledge} about them. This corresponds to whether the writer has a frequentist or a subjectivist slant, often not explicitly stated, but implicit in verbal hints. The question is not primarily ``should you switch or stay?'', but ``what is \emph{the} probability, or \emph{your} probability, that switching will give the car?'' Typically, as in Selvin's second, conditional, approach, the player's initial choice is already fixed in the problem statement, the host's two actions are already seen as completely random, whether because we are told they are, objectively, or because we are completely ignorant of how they are made, subjectively. The problem typically features in the chapter which introduces conditional probability and Bayes' theorem in the discrete setting. Thus the problem is posed by a maths teacher who wants the student to learn conditional probability. The problem is further reformulated as ``what is the \emph{conditional} probability that switching will give the car''.
In such a context not much attention is being paid to the meaning of probability. After all, right now we are just busy getting accustomed to its calculus. Most of the examples figure playing cards, dice and balls in urns, and so the probability spaces are usually completely specified (all outcomes can be enumerated) and mostly they are symmetric, all elementary probabilities are equal. The student is either supposed to ``know'', or he is told explicitly, that the car is initially equally likely to be behind any of the three doors. The host is supposed to choose at random (shorthand for uniformly, or completely, at random) when he has a choice. Since these facts are given or supposed to be guessed, the initial choice of the player is irrelevant, and we are indeed always told that the player has already picked Door 1.
Well, if MHP is merely an exercise in conditional probability where the mathematical model is specified in advance by the teacher, then it is clear how we are to proceed. But I prefer to take a step back and to \emph{imagine you are on a game show}. How could we ``know'' these probabilities? This is especially important when one has the task of ``explaining'' MHP to non-mathematicians and to non-statisticians.
This is where philosophy, or if you prefer, metaphysics, raises its head. How can one ``know'' a probability; what does it mean to ``know'' a probability?
I am not going to answer those questions. But I am going to compare a conventional frequentist view -- probability is out there in the real world -- to a conventional Bayesian view -- probability is in the information which we possess. I hope to do this neutrally, without taking a dogmatic stance myself. It is a fact that many amateur users of probability are instinctive subjectivists, not so many are instinctive frequentists. Let us try to work out where either instinct would take us. An important thing to realise is: Bayesian probabilities in, Bayesian probabilities out; frequentist in, frequentist out. I will also emphasize the difference which comes from seeing randomness in the \emph{host's moves} or in the \emph{player's moves}, and the difference which comes from seeing the question as asking for an \emph{action}, or more passively for a \emph{probability}.
For a subjectivist (Bayesian) the MHP is very simple indeed. We only know what we have been told by vos Savant (Whitaker). The wording ``\emph{say}, Door 1'' and ``\emph{say}, Door 3'' (my italics) emphasize that we know nothing about the behaviour of the host, whether in hiding cars or in opening doors. Knowing nothing, the situation \emph{for us} is indistinguishable from the situation in which we had been told in advance that car hiding and door opening was actually done using perfect fair randomizers (unbiased dice throws or coin tosses). Probability is a representation of our uncertain knowledge about the single instance under consideration. Probability calculus is the unique internally consistent way to manipulate uncertain knowledge. To start off with, since we know nothing, we may as well choose our door initially according to our personal lucky number, so we picked Door 1. Having seen the host open Door 3, we would now be prepared to bet at odds 2:1 that the car is behind Door 2. The new situation is indistinguishable for us from a betting situation with fair odds 2:1 based on a perfect fair randomizer (by which I simply refer to the kind of situation in which subjectivists and objectivists tend to agree on the probabilities, even if they think they mean something quite different).
Does the Bayesian (a subjectivist) need Bayes' theorem in order to come to his conclusion? I think the answer is \emph{no}. For a subjectivist the door numbers are irrelevant. The problem is unchanged by renumbering of the doors. His beliefs about whether the car would be behind the other door in any of the six situations (door chosen, door opened) would be the same. He has no need to actively compute the conditional probability in order to confirm what he already knows. He could use Proposition 3 but is only interested in Proposition 1. The symmetry argument of my proof of Proposition 3 is the mathematical expression of his prior knowledge that he may ignore the door numbers and just compute an unconditional probability. Do you notice the symmetry in advance and take advantage of it, or just compute away and notice it afterwards? It doesn't matter. The answer is $2/3$ and it is a conditional and unconditional probability at the same time.
What is important to realise is that the probability computed by a subjectivist is also a subjective probability. Starting from probabilities which express prior personal expectations, the probability we have derived expresses how our prior personal expectations as to the location of the car should logically be modified on seeing the host open Door 3 and reveal a goat in response to our choice of Door 1. These probabilities say nothing about what we would expect to see if the game was repeated many times. We might well expect to see a systematic bias in the location of the car or the choice of the host. Our uniform prior distributions express the fact that our prior beliefs or prior information about such biases are invariant under permutations of the door numbers.
For a frequentist, MHP is harder -- unless the problem has already been mathematized, and the frequentist has been told that the car is hidden completely at random and the host chooses completely at random (when he has a choice) too. Personally, I don't find this a very realistic scenario. I can think of one semi-realistic scenario, and that is the scenario proposed by Morgan et al.~(1991a). Suppose we have inside information that every week, the car is hidden uniformly at random, in order to make its location totally unpredictable to all players. However Monty's choice of door to open, when he has a choice, is something which goes on in his head at the spur of the moment. In this situation we may as well let our initial choice of door be determined by our lucky number, e.g., Door 1. Proposition 2 tells us that not only is always switching a wise strategy, it tells us that we cannot do better. No need to worry our heads about \emph{what is} the conditional probability. It is never against switching.
There is just one solution which does not require any prior knowledge at all; instead it requires prior action. Taking our cue from the game theoretic solution, we realize that the player has two opportunites to act, not one. We allow ourselves the latitude to reformulate vos Savant's words as ``You are \emph{going to be} on a game show...''. We advise vos Savant, or her correspondent Craig Whitaker, to take fate into his or her own hands. Before the show, pick your lucky number (1, 2 or 3) by a toss of a fair die. When you make it to the final round, choose that door and thereafter switch. By Proposition 1 you'll come home with the car with probability 2/3, and by Proposition 4 that's the best you can hope for.
Both frequentists and subjectivists will agree that you win the car with probability 2/3 in this way. They will likely disagree about whether the conditional probability that you win, given door chosen and door opened, is also 2/3. I think the frequentist will say that he does not know it since he doesn't know anything about the two host actions, while the subjectivist will say that he does know the conditional probability (and it's 2/3) for the very same reason. So what?
\section{Conclusion}
The Monty Hall Problem offers much more to the student than a mindless exercise in conditional probability. It also offers a challenging exercise in mathematical modelling. I notice three important lessons. (1) The more you assume, the more you can conclude, but the more limited are your conclusions. The honest answer is not one mathematical solution but a range of solutions. (2) Whether you are a subjectivist or a frequentist affects the ease with which you might make probabilistic assumptions but simultaneously affects the meaning of the conclusions. (3). Think out of the box. Vos Savant asks for an \emph{action}, not for a \emph{probability}. The player has \emph{two} decision moments during the show, not one.
\section*{References} ~
\vskip -0.5cm
\raggedright
\frenchspacing
\parskip 0.2 cm
\leftskip 0.5 cm
\parindent -0.5 cm
Burns, B. D. and Wieth, M. (2004), The collider principle in causal reasoning: why the Monty Hall dilemma is so hard,
{\it J. Experimental Psychology: General\/} {\bf 133\/} 434--449. {\tt http://www.psych.usyd.edu.au/staff/bburns/Burns\_Wieth04\_man.pdf}
van Damme, E. E. C. (1995), Rationaliteit. {\it Econ. Stat. Berichten} 15--11--1995, 1019.
Herbranson, W. T. and Schroeder, J. (2010), Are birds smarter than mathematicians? Pigeons (Columba livia) perform optimally on a version of the Monty Hall Dilemma. {\it J. Comp. Psychol.\/} {\bf 124\/} 1--13. {\tt http://people.whitman.edu/$\sim$herbrawt/HS\_JCP\_2010.pdf}
Hogbin, M. and Nijdam, W. (2010), Letter to the editor. {\it Am. Statist.\/} {\bf 64} 193.
Morgan, J. P., Chaganty, N. R., Dahiya, R. C., and Doviak, M. J. (1991a), Let's make a deal: The player's dilemma. {\it Am. Statist.\/} {\bf 45\/} 284--287.
Morgan, J. P., Chaganty, N. R., Dahiya, R. C., and Doviak, M. J. (1991b), Rejoinder to Seymann's comment on ``Let's make a deal: the player's dilemma''. {\it Am. Statist.\/} {\bf 45\/} 289.
Morgan, J. P., Chaganty, N. R., Dahiya, R. C., and Doviak, M. J. (2010), Response to Hogbin and Nijdam's letter, {\it Am. Statist.\/} {\bf 64} 193--194.
Mosteller, F. (1965), \emph{Fifty Challenging Problems in Probability with Solutions}. Dover, New York.
Nalebuff, B. (1987), Puzzles: Choose a curtain, duel-ity, two point conversions, and more, {\it J. Econ. Perspectives\/} {\bf 1} (2) 157--163.
Rosenhouse, J. (2009), \emph{The Monty Hall Problem}, Oxford University Press.
Rosenthal, J. S. (2005), {\it Struck by Lightning: The Curious World of Probabilities}. Harper Collins, Canada.
Rosenthal, J. S. (2008), Monty Hall, Monty Fall, Monty Crawl, \emph{Math Horizons\/} {\bf 16\/} 5--7. \texttt{http://probability.ca/jeff/writing/montyfall.pdf}
Selvin, S. (1975a), A problem in probability (letter to the editor). {\it Am. Statist.} {\bf 29} 67.
Selvin, S. (1975b), On the Monty Hall problem (letter to the editor). {\it Am. Statist.\/} {\bf 29\/} 134.
Seyman, R. G. (1991), Comment on ``Let's make a deal: the player's dilemma''. {\it Am. Statist.\/} {\bf 64} 287--288.
\end{document}
|
\section{Introduction}
\label{sectionIntro}
When a system is governed by an autonomous nonlinear algebraic
partial differential equation (PDE),
it frequently admits some permanent profile structures
such as fronts, pulses, kinks, etc
\cite{vS2003},
and usually these profiles are mathematically some
single-valued solutions
of the travelling wave reduction $(x,t) \to x-ct$
of the PDE to an ordinary differential equation (ODE).
The physical motivation of the present work
is to find such solutions in closed form.
Since this is a difficult mathematical problem,
we restrict here to a simple case (a third order nonlinear ODE)
and solve it completely.
The method we used here is a refinement of Eremenko's method
used in \cite{Eremenko1982} as well as \cite{EremenkoKS}
and \cite{ELN} which is based on the local singularity analysis
of the meromorphic solutions of the given differential equations
as well as the zero distribution and growth rate of the meromorphic
solutions by using Nevanlinna theory.
This is a very powerful method.
For example, it was used by Eremenko \cite{EremenkoKS}
to characterize all meromorphic traveling wave solutions of the
Kuramoto-Sivashinsky (KS) equations.
In fact, Eremenko showed that all the meromorphic traveling wave solutions
of the KS equations belong to the class $W$ (like Weierstrass),
which consists of
elliptic functions and their successive degeneracies,
i.e.:
elliptic functions,
rational functions of one exponential $\exp(kz), k\in\mathbb{C}$
and rational functions of $z$.
In general, even if we know that the solutions belong to the class $W$,
it is still difficult to find their explicit form.
To overcome this problem,
we shall apply the subequation method introduced in \cite{MC2003}
and developed in \cite{CM2009}.
In order to emphasize the method,
we will choose a test equation according to the following criteria:
\begin{enumerate}
\item
to have a small differential order $n$,
\item
to have only nonrational Fuchs indices,
apart from the ever present $-1$ index,
\item
to be of the form
$u^{(n)}=P(u^{(n-1)},\dots,u',u)$,
with $P$ a polynomial of its arguments,
\item
to have movable poles of order one,
\item
to be \textit{complete} in the classical sense \cite{PaiBSMF}
(see details in \cite[p.~122]{Cargese1996Conte})
i.e.~to include all admissible nondominant terms,
\end{enumerate}
The requirement for nonrational Fuchs indices sets $n \ge 3$.
Let us take the complete autonomous
third order polynomial ODE with simple poles,
\begin{eqnarray}
& &
d_0 u''' + d_1 u u'' + d_2 {u'}^2 + d_3 u^2 u' + d_4 u^4
\nonumber \\ & &
+c_1 u'' +c_2 u u' +c_3 u^3
+c_4 u'+c_5 u^2
+c_6 u
+c_7=0.
\label{eqODE3Complete}
\end{eqnarray}
This equation is indeed complete in the sense that it
includes all polynomial terms having a singularity degree at most
equal to four,
as seen from the generating function
\begin{eqnarray}
& &
\frac{1}{(1-t u)(1 -t^2 u')(1-t^3 u'')}
\nonumber\\ & &
=1 + u t
+(u^2+u') t^2 + (u^3 + u u' + u'') t^3
\nonumber\\ & & \phantom{1234567890}
+(u^4 + u^2 u' +{u'}^2 + u u'') t^4+O(t^5).
\end{eqnarray}
Let us choose one particular set of dominant terms
(the ones with coefficients $d_j$, which have quadruple poles)
so as to enforce from the beginning
the condition that the Fuchs indices be nonrational.
After setting $c_3=0$ without loss of generality,
our test equation will be normalized as
\begin{eqnarray}
& &
c_0 u''' + 6 u^4
+c_1 u'' +c_2 u u'
+c_4 u'+c_5 u^2
+c_6 u
+c_7=0,
\label{eqODE3}
\end{eqnarray}
Let $u$ be a meromorphic solution of the ODE (\ref{eqODE3}).
We first check that if $u$ has a movable
pole at $z=z_0$,
then $u$ has only three distinct Laurent series expansions at $z_0$.
Note that if $z_0$ is a pole of $u$, it must be a simple pole.
Therefore, in a neighbourhood of $z=z_0$,
the Laurent series of the meromorphic solution $u$ is of the form
\begin{eqnarray}
& &
u(z)= u_{-1}(z-z_0)^{-1} +u_0 +u_1 (z-z_0)+\cdots,\ u_{-1}\not=0.
\label{eqLaurent}
\label{eqODE3Laurent}
\end{eqnarray}
Denote $a$ any one
of the cubic roots of $c_0$.
Substituting the above Laurent series into the ODE (\ref{eqODE3})
and balancing the leading terms,
we obtain $u_{-1}=a$,
and $u_0=(-2 c_1 a + c_2 a^2)/(24 c_0)$.
We are going to prove that there are at most
three distinct Laurent series expansions at $z_0$.
If one linearizes the ODE (\ref{eqODE3})
around the movable singularity $z=z_0$ \cite[p.~114]{Cargese1996Conte},
the resulting linear ODE has the Fuchsian type at $z_0$,
and its three Fuchs indices $r$ are defined by
\begin{eqnarray}
& &
(r+1) (r^2-7r+18)=0.
\end{eqnarray}
Hence, the Fuchs indices
are equal to $r=-1,(7 \pm \sqrt{-23})/2$.
Because of the absence of any positive integer
in the set of values of $r$,
all other cofficients $u_i$ are uniquely determined
\cite[p.~90]{Cargese1996Conte}
by the leading coefficient $u_{-1}$.
Hence, there are at most three meromorphic functions
with poles at $z=z_0$ satisfying the ODE (\ref{eqODE3}).
We shall study the third order nonlinear
differential equation (\ref{eqODE3})
and show that all meromorphic solutions of this differential
equation belong to the class $W$.
More specifically, our results are the following.
\begin{thm}
If the ODE (\ref{eqODE3}) has a particular meromorphic solution $u$,
then $u$ belongs to the class $W$.
Moreover, a necessary and sufficient condition for the ODE (\ref{eqODE3})
to admit a particular meromorphic solution is
to belong to the following list,
\begin{eqnarray}
& &
S_{3a}:\
c_1,c_6=\mbox{\rm arbitrary},\
c_2=0,\ c_5=0,\ c_7=0,\ c_4=\frac{c_1^2}{12 c_0};\
\\
& &
S_{3b}:\
c_5,c_6=\hbox{\rm arbitrary},\
c_1=0,\ c_2=0,\ c_4=0,\ c_7=\frac{c_5^2}{128};\
\\
& &
S_{2A}:\
c_1,c_4=\hbox{\rm arbitrary},\
c_2=0,\
c_5=\frac{c_1^2 - 12 a^3 c_4}{4 a^4},\
c_6=-\frac{c_1 (c_1^2 + 36 a^3 c_4)}{144 a^6},\
\nonumber \\ & & \phantom{12345}
c_7=\frac{(12 a^3 c_4 -c_1^2)(36 a^3 c_4-11 c_1^2)}{1536 a^8};\
\\
& &
S_{2B}:\
c_1,c_2=\hbox{\rm arbitrary},\
c_4= \frac{44 c_1^2 + 8 a c_1 c_2 - a^2 c_2^2}{144 a^3},\
\nonumber \\ & & \phantom{12345}
c_5=\frac{-32 c_1^2 -24 a c_1 c_2 - 7 a^2 c_2^2}{48 a^4},\
c_6=-\frac{(c_1+a c_2)(12 c_1^2 + 6 a c_1 c_2 + a^2 c_2^2)}{144 a^6},\
\nonumber \\ & & \phantom{12345}
c_7=-\frac{(4 c_1 + 3 a c_2)(48 c_1^2 +20 a c_1 c_2 + a^2 c_2^2)}{55296 a^7};
\\
& &
S_{1}:\
c_1,c_2,c_4,c_5=\hbox{\rm arbitrary},\
\nonumber \\ & & \phantom{1234}
1152 a^6 c_6=-56 c_1^3 + 60 a c_1^2 c_2 -18 a^2 c_1 c_2^2
+a^3 c_2^3 +288 a^3 c_1 c_4
\nonumber \\ & & \phantom{1234567890123}
-144 a^4 c_2 c_4-96 a^4 c_1 c_5 +48 a^5 c_2 c_5,
\nonumber \\ & & \phantom{1234}
2^{13} 3^2 a^8 c_7=-176 c_1^4 +128 a c_1^3 c_2 +24 a^2 c_1^2 c_2^2
-32 a^3 c_1 c_2^3
\nonumber \\ & & \phantom{1234567890123}
+5 a^4 c_2^4 +2688 a^3 c_1^2 c_4
-1536 a^4 c_1 c_2 c_4 +96 a^5 c_2^2 c_4
-6912 a^6 c_4^2
\nonumber \\ & & \phantom{1234567890123}
+128 a^4 c_1^2 c_5 -512 a^5 c_1 c_2 c_5
+ 224 a^6 c_2^2 c_5
\nonumber \\ & & \phantom{1234567890123}
+ 4608 a^7 c_4 c_5 + 2304 a^8 c_5^2.
\end{eqnarray}
\end{thm}
\bigskip
We shall apply Eremenko's method \cite{EremenkoKS}
to prove the first part of Theorem 1.
Here, we shall assume the readers are familiar with the standard terminology
and results of Nevanlinna theory.
The standard reference of this theory are \cite{Hayman64}
and \cite{Laine-book,Nevanlinna-book}
(see also \cite{EremenkoKS} for a quick introduction).
Our argument is slightly different from that of Eremenko
and it makes use of the following version of Clunie's Lemma
(\cite[Lemma 2.4.2]{Laine-book}, see also \cite{YZ}).
\begin{lemma} Let $f$ be a transcendental meromorphic solution of
$$f^nP(z,f)=Q(z,f),$$
where $P(z,f)$ and $Q(z,f)$ are polynomials in $f$
and its derivatives with meromorphic coefficients
$\{a_\lambda|\lambda \in I\}$ such that
$m(r,a_{\lambda})=S(r,f)$ for all $\lambda \in I$.
If the total degree of $Q(z,f)$ as a polynomial in $f$
and its derivatives is less than or equal to $n$, then
$$m(r,P(r,f))=S(r,f).$$
\label{lemma2}
\end{lemma}
Now let $u$ be a function meromorphic in the complex plane
which satisfies the above ODE (\ref{eqODE3}).
If $u$ is rational, then we are done.
So suppose $u$ is a transcendental meromorphic solution
of equation (\ref{eqODE3}), then we have
\begin{eqnarray}
& &
-6 u^4=c_0 u'''
+c_1 u'' +c_2 u u'
+c_4 u'+c_5 u^2
+c_6 u
+c_7,
\end{eqnarray}
Take $f=u, P=u, n=3$ and apply Clunie's lemma (Lemma 1)
to the above equation,
we conclude that $m(r,u)=S(r,u)$ and hence $(1-o(1))T(r,u)=N(r,u)$.
We claim that $u$ must have infinitely many poles.
Assume it is not the case, then $N(r,u)= O(\log r)$.
Therefore, $T(r,u)=O(\log r)$ which is impossible as $u$ is transcendental.
Secondly, we prove that if $u$ is a transcendental
meromorphic solution,
then $u$ is a periodic function.
Recall that there are at most three meromorphic functions
with poles at $z=z_0$ satisfying the ODE (\ref{eqODE3}).
Now let $z_j, j=1,2,3,\cdots$ be the poles of $u(z)$,
then the functions $w_j(z) = u(z+z_j-z_0)$ are meromorphic
solutions of the ODE (\ref{eqODE3}) with a pole at $z_0$.
Thus some of them must be equal.
Consequently, $u$ is a periodic function.
Without loss generality, we may assume that $u$ has a period of $2\pi i$.
Let
$D=\{z:0\le {\rm Im} z <2\pi\}$. If $u$ has more than three poles in $D$,
then by the previous argument,
we can conclude that $u$ is periodic in $D$
and hence it is indeed an elliptic function and we are done.
Now suppose $u$ has at most three poles in $D$.
Since $u$ is a periodic function with period $2\pi i$, we have
$N(r,u)= O(r),$ as $r\to\infty$.
It follows from $(1-o(1))T(r,u)=N(r,u)$ that $T(r,u)=O(r)$.
By Nevanlinna's First Fundamental Theorem, we know that
for any $a\in\mathbb{C}$, $N(r,1/(u -a)) =O(r)$ as $r\to\infty$.
By the periodicity of $u$, we conclude that $u$ take each $a$ finitely many
times in $D$. Hence, the function $R(z)=u(\ln z)$ is a
single-valued analytic function in the punctured plane
$\mathbb{C}\backslash \{0\}$ and takes each $a \in \mathbb{C}$ finitely many times.
It follows that $0$ is a removable singularity of $R$ and $R$
can then be extended to a meromorphic function on $\mathbb{C}$. Hence $R$ is a rational function
as it takes each complex number finitely many times.
Therefore, $u(z)=R(e^z)$ belongs to the class $W$ and this completes the proof of
the first part of Theorem 1.
\medskip
\noindent
{\bf Remark.}
{}From the above proof, we notice that if $u$ is an elliptic solution,
then $u$ has at most three (simple) poles in each fundamental polygon
$\Omega$.
Recall that the residue of $u$ at any pole must be one of
$a,\omega a,\omega^2 a$ where $\omega$ is the cubic root of unity.
Since the sum of the residues of all the poles in any fundamental
polygon $\Omega$ is zero,
$u$ must have three distint simple poles in $\Omega$
and hence we have three distinct Laurent series at $z_0$.
\noindent
{\bf Remark.}
If $u(z)=R(e^{kz})$ where $R$ is some rational function,
then $R$ has at most three (simple) poles in $\mathbb{C}\backslash \{0\}$.
We are going to show that $R$ cannot have a pole at $0$.
Suppose we write
$u(z)=R(Z)= r_0/Z^n + \sum_{i=1}^{3} r_i/(Z-Z_i) + q(Z)$,
where $q$ is a polynomial in $Z= e^{kz}$.
Substituting $u(z)=R(Z)$ into ODE (\ref{eqODE3}) and letting $Z$ tend to infinity,
we can conclude that $q$ equals to some constant $C$.
Now letting $Z$ tend to $0$, we can deduce that $r_0=0$.
Hence, $u(z)= \sum_{i=1}^{3} r_i/(e^{kz} -Z_i) + C$,
where $Z_i, C \in \mathbb{C}$.
Finally, if $u$ is rational,
then $u$ will have at most three (simple) poles in $\mathbb{C}$.
Similarly, we can show that $u$ must be of the form $\sum_{i=1}^{3}r_i/(z-z_i) + C$,
where $r_i, C \in \mathbb{C}$.
\section{Explicit solutions in the class $W$}
\label{section2}
Let us determine the constraints on the coefficients
$c_j$ of (\ref{eqODE3}) for meromorphic solutions to exist,
and let us determine all these meromorphic solutions
in closed form.
According to section \ref{sectionIntro},
these solutions are necessarily elliptic or degenerate of elliptic
(i.e.~rational in one exponential $e^{k z}, k \in \mathbb{C}$
or rational in $z$),
i.e.~they belong to the class $W$.
If the meromorphic solution is elliptic,
by a classical theorem,
the sum of the residues of the three Laurent series
for $u$, Eq.~(\ref{eqLaurent}),
must vanish,
and similarly for any rational function of $u,u',u''$.
These necessary conditions \cite{Hone2005}
are first established in section
\ref{sectionResidueConditions}.
If the solution is elliptic,
one knows the elliptic orders of $u$ and $u'$,
they are respectively equal to
three (three simple poles)
and six (three double poles).
Therefore,
by a classical theorem of Briot and Bouquet
\cite[p.~277]{BriotBouquet},
\cite[part II, chap.~IX p.~329]{HalphenTraite},
\cite[p.~424]{Hille}
the elliptic solution obeys a first order algebraic equation
whose degree in $u'$ is the order of $u $ (three)
and degree in $u $ is the order of $u'$ (six),
\begin{eqnarray}
& &
F(u,u') \equiv
\sum_{k=0}^{m} \sum_{j=0}^{2m-2k} a_{j,k} u^j {u'}^k=0,\
a_{0,m}\not=0,\
\label{eqsubeqODEOrderOnePP}
\end{eqnarray}
with $m=3$.
The complex constants $a_{j,k}$, with $a_{0,m}\not=0$,
are then determined by the algorithm presented in
\cite{MC2003},
i.e.~by requiring each of the three Laurent series
(\ref{eqLaurent}) to obey (\ref{eqsubeqODEOrderOnePP}).
The search for all third degree subequations
(\ref{eqsubeqODEOrderOnePP})
obeyed by the three Laurent series
is performed in section \ref{sectionsubeqdeg3}.
As to those solutions of (\ref{eqODE3}) which are
degenerate of elliptic,
they also obey a first order equation (\ref{eqsubeqODEOrderOnePP}),
whose degree $m$ is at most three.
Because of the singularity structure of (\ref{eqODE3})
(three \textit{distinct} Laurent series),
any $m$-th degree subequation, $1 \le m \le 3$,
must have $m$ \textit{distinct} Laurent series.
The search for all second or first degree subequations
(\ref{eqsubeqODEOrderOnePP})
is performed in sections (\ref{sectionsubeqdeg2})
and (\ref{sectionsubeqdeg1}).
Let us first establish all these first order subequations.
Their general solution may be either singlevalued
(and hence in class $W$) or multivalued.
The explicit integration of the singlevalued subset
will provide as a final output
all the meromorphic solutions of (\ref{eqODE3})
in closed form.
\subsection{Residue conditions}
\label{sectionResidueConditions}
If (\ref{eqODE3}) admits an elliptic solution,
it is necessary that,
for any rational function of $u$ and its derivatives,
the sum of the residues inside a period parallelogram be zero,
\begin{eqnarray}
& &
\forall k \in \mathbb{N}\
\forall n \in \mathbb{N}\:
\mathop{\rm res}\nolimits \sum_{i=1}^{3} \left(u^{(k)}\right)^n=0.
\label{eqresidueconditions}
\end{eqnarray}
The first conditions are
\begin{eqnarray}
& &
\left\lbrace
\begin{array}{ll}
\displaystyle{
k=0,n=2:\ c_2=0,
}\\ \displaystyle{
k=0,n=3:\ c_4 =\frac{c_1^3}{12 a^3},
}\\ \displaystyle{
k=0,n=5:\ c_1 c_5 =0,
}\\ \displaystyle{
k=0,n=7:\ c_1 c_7=0,
}\\ \displaystyle{
k=1,n=4:\ (c_6 (c_5^2-128 c_7)=0 \hbox{ if } c_1=0),\
(c_7 (c_1^3+36 a_0^2 c_6)=0 \hbox{ if } c_1\not=0).
}
\end{array}
\right.
\end{eqnarray}
When the computation is limited to $k \le 4,n \le 10$,
this defines three and only three distinct
sets of fixed coefficients for a possible elliptic solution,
\begin{eqnarray}
& &
c_2=0,\ c_1=0,\ c_4=0,\ c_6 \not=0,\ c_7=\frac{c_5^2}{128},\
\label{eqrell1K0}
\label{eqrella}
\\
& &
c_2=0,\ c_1=0,\ c_4=0,\ c_6=0,\
\label{eqrell1K1}
\label{eqrellb}
\\
& &
c_2=0,\ c_1\not=0,\ c_4=\frac{c_1^2}{12 a^3},\ c_5=0,\ c_7=0
\label{eqrell2K3}
\label{eqrellc}
\end{eqnarray}
\subsection{Subequations of degree three}
\label{sectionsubeqdeg3}
Denoting $\omega_k,k=1,2,3$, cubic roots of unity,
each such subequation has the necessary form
\begin{eqnarray}
& &
F(u,u') \equiv
-(\omega_1 a u' + u^2)(\omega_2 a u' + u^2)(\omega_3 a u' + u^2)
\nonumber \\ & & \phantom{12345678}
+b_1 {u'}^2 u + b_2 u' u^3 + b_3 u^5
+b_4 {u'}^2 + b_5 u' u^2 + b_6 u^4
\nonumber \\ & & \phantom{12345678}
+ b_7 u' u + b_8 u^3
+ b_9 u' + b_b u^2
+ b_a u
+ b_c
+ b_0=0,
\label{eqsubeqODEOrderOnePP3Repeat}
\end{eqnarray}
with all $\omega_j$ distinct
and the additional condition to be irreducible.
The first order third degree subequation is precisely defined as
\begin{eqnarray}
& &
F(u,u') \equiv
-a^3 {u'}^3 - u^6
+b_1 {u'}^2 u + b_2 u' u^3 + b_3 u^5
+b_4 {u'}^2 + b_5 u' u^2 + b_6 u^4
\nonumber \\ & & \phantom{12345678}
+ b_7 u' u + b_8 u^3
+ b_9 u' + b_b u^2
+ b_a u
+ b_c
+ b_0=0.
\label{eqsubeqODEOrderOnePP3}
\end{eqnarray}
The algorithm \cite{MC2003} to compute the coefficients $b_k$
is to substitute $u$ by one of
the Laurent series (\ref{eqLaurent}),
which makes
the right hand side of (\ref{eqsubeqODEOrderOnePP3}) become a Laurent series
\begin{eqnarray}
& &
F(u,u') \equiv
\sum_{j=0}^{+\infty} F_j (z-z_0)^{j-6},
\label{eqLaurentF}
\end{eqnarray}
then to solve the infinite set of equations
\begin{eqnarray}
& &
\forall a\ \forall j:\ F_j=0.
\label{eqSystemFj}
\end{eqnarray}
The practical resolution is as follows.
First, the 21 equations $F_j=0, j=0,...,6$
define a linear system for the $b_k$,
which admits a unique solution
and generates six nonlinear constraints among the six $c_k$.
By considering slightly more equations in (\ref{eqSystemFj})
(in the present case, going to $j=8$ is enough),
the set of nonlinear constraints among the $c_k$'s
admits exactly two solutions,
and all the remaining equations $F_j=0$ identically vanish.
These two solutions are
\begin{eqnarray}
& &
\left\lbrace
\begin{array}{ll}
\displaystyle{
S_{3a}:\
c_1,c_6=\hbox{arbitrary},\
c_2=0,\ c_5=0,\ c_7=0,\ c_4=\frac{c_1^2}{12 a_0},\
\label{eqgenusonecase1}
}\\ \displaystyle{
S_{3b}:\
c_5,c_6=\hbox{arbitrary},\
c_1=0,\ c_2=0,\ c_4=0,\ c_7=\frac{c_5^2}{128},
\label{eqgenusonecase2}
}
\end{array}
\right.
\end{eqnarray}
and they are identical to the two residue conditions
(\ref{eqrell2K3}) and (\ref{eqrell1K0}).
The corresponding subequations have genus one
\begin{eqnarray}
& &
(a u' + 4 k_1 u)^2 (a u' - 2 k_1 u) + (u^3 + 20 k_1^3 + k_6)^2=0,\
c_1=12 a^2 k_1,\
c_6=4 k_6,\
\label{eqsubeq3a}
\\
& &
(a u')^3 + (u^3 -3 k_5^2 u + k_6)^2=0,\
c_5=-16 k_5^2,\
c_6=4 k_6.
\label{eqsubeq3b}
\end{eqnarray}
The method to integrate them \cite[\S 249 p.~393]{BriotBouquet}
is to build a birational transformation to the canonical equation
of Weierstrass
\begin{eqnarray}
& &
{\wp'}^2=4 (\wp-e_1)(\wp-e_2)(\wp-e_3)=4 \wp^3 - g_2 \wp - g_3.
\end{eqnarray}
To do that, it proves convenient to introduce one of the roots $e_0$
of the cubic polynomial of $u(x)$ appearing as a square
in (\ref{eqsubeq3a}) and (\ref{eqsubeq3b}),
i.e.~to redefine $k_6$ by the respective relations
\begin{eqnarray}
& &
e_0^3 + 20 k_1^3 + k_6=0
\hbox{ and }
e_0^3 - 3 k_5^2 e_0 + k_6=0.
\end{eqnarray}
The subequation (\ref{eqsubeq3b})
is one of the five first order binomial equations
of Briot and Bouquet \cite[p.~122]{Cargese1996Conte},
its general solution is classical
\begin{eqnarray}
& &
\frac{1}{u-e_0}=\frac{\wp'(z-z_0,g_2,g_3)-A}{N_1},\
g_2=0,\
g_3=\frac{ (e_0^2-k_5^2)^2 (e_0^2-4 k_5^2)}{243 a^6},\
\nonumber \\ & &
N_1=\frac{2 (e_0^2-k_5^2)^2}{3 a^3},\
A=\frac{e_0 (e_0^2-k_5^2)}{3 a^3}.
\label{eqBinomial3}
\end{eqnarray}
The subequation (\ref{eqsubeq3a})
has been integrated by Briot and Bouquet \cite[\S 250 p.~395]{BriotBouquet}
by introducing a function $w$ defined by
\begin{eqnarray}
& &
a u' + 4 k_1 u = \frac{u^3 - e_0^3}{u-e_0} w,
\end{eqnarray}
then by establishing the birational tranformation
\begin{eqnarray}
& &
w=\frac{a u' + 4 k_1 u}{u^2 + e_0 u + e_0^2},\
u=\frac{-3 a w w' - e_0 w^3 + 6 k_1 w^2 + 2 e_0}{2 (w^3 +1)},
\end{eqnarray}
finally by integrating the ODE for $w$,
\begin{eqnarray}
& &
w=\frac{2 k_1}{e_0}+\frac{A}{\wp - B},\
g_2=\frac{4 k_1 (k_1^3 - e_0^3)}{3 a^4},\
g_3=\frac{e_0^6 - 20 e_0^3 k_1^3 - 8 k_1^6}{17 a^6},\
\nonumber \\ & &
g_2^3-27 g_3^2=-\frac{(8 k_1^3 + e_0^3)^3 e_0^3}{27 a^{12}},
A=-\frac{e_0^3 + 8 k_1^3}{3 a^2},\
B=-\frac{k_1^2}{ a^2}.
\end{eqnarray}
More generally,
birational transformations from $(u,u')$ to $(\wp,\wp')$
are obtained with an algorithm due to Poincar\'e,
implemented for instance by the command \verb+Weierstrassform+
of the computer algebra package \verb+algcurves+ \cite{MapleAlgcurves}.
\subsection{Subequations of degree two}
\label{sectionsubeqdeg2}
Let us
define the second degree subequation as
\begin{eqnarray}
& &
F(u,u') \equiv
a^2 {u'}^2 - a u^2 u' + u^4
+b_4 u' u + b_3 u^3
+b_5 u' + b_2 u^2
+ b_1 u
+b_0=0,
\label{eqsubeqODEOrderTwoPP1}
\end{eqnarray}
with the additional condition to be irreducible.
Computations similar to those mentioned in section
\ref{sectionsubeqdeg3}
provide two solutions,
\begin{eqnarray}
& &
\left\lbrace
\begin{array}{ll}
\displaystyle{
S_{2A}:\
c_1,c_4=\hbox{arbitrary},\
c_2=0,\
c_5=\frac{c_1^2 - 12 a^3 c_4}{4 a^4},\
}\\ \displaystyle{\phantom{12345}
c_6=-\frac{c_1 (c_1^2 + 36 a^3 c_4)}{144 a^6},\
}\\ \displaystyle{\phantom{12345}
c_7=\frac{(12 a^3 c_4 -c_1^2)(36 a^3 c_4-11 c_1^2)}{1536 a^8},\
}\\ \displaystyle{\phantom{12345}
u=v-\frac{k_1}{2},\
c_1=-3 a^2 k_1,\
c_4=2 a b^2 +\frac{3}{4} a k_1^2,\
}\\ \displaystyle{\phantom{12345}
\left(a v'-\frac{v^2-b^2}{2}\right)^2+\frac{3}{4} (v+b) (v-b) (v-k_1)^2=0,\
b \not=0,
}
\end{array}
\right.
\label{eqsubeq2sol6}
\end{eqnarray}
and
\begin{eqnarray}
& &
\left\lbrace
\begin{array}{ll}
\displaystyle{
S_{2B}:\
c_1,c_2=\hbox{arbitrary},\
c_4= \frac{44 c_1^2 + 8 a c_1 c_2 - a^2 c_2^2}{144 a^3},\
}\\ \displaystyle{\phantom{12345}
c_5=\frac{-32 c_1^2 -24 a c_1 c_2 - 7 a^2 c_2^2}{48 a^4},\
}\\ \displaystyle{\phantom{12345}
c_6=-\frac{(c_1+a c_2)(12 c_1^2 + 6 a c_1 c_2 + a^2 c_2^2)}{144 a^6},\
}\\ \displaystyle{\phantom{12345}
c_7=-\frac{(4 c_1 + 3 a c_2)(48 c_1^2 +20 a c_1 c_2 + a^2 c_2^2)}{55296 a^7},
}\\ \displaystyle{\phantom{12345}
u=v+\frac{b}{4} + \frac{c_1}{12 a^2},\ c_2=-2 \frac{c_1}{a} + 6 a b,\
}\\ \displaystyle{\phantom{12345}
\left(a v'-\frac{v^2-b^2}{2}\right)^2+\frac{3}{4} (v+b)^3 (v-b)=0,\
b \not=0.
}
\end{array}
\right.
\label{eqsubeq2sol2}
\end{eqnarray}
For $k_1^2 \not= b^2$,
the point transformation
\begin{eqnarray}
& &
v=k_1+\frac{1}{w},\
w=-\frac{1}{k_1+b}-\frac{1}{k_1-b}
+N \left(\lambda-\frac{1}{\lambda}\right),
N^2=- \frac{b^2}{(k_1^2-b^2)^2},
\end{eqnarray}
maps the ODE (\ref{eqsubeq2sol6}) to the Riccati ODE
\begin{eqnarray}
& &
a N \lambda' - M \lambda-\frac{b^2}{4 (k_1^2-b^2)} (\lambda^2+1)=0,\
M^2=\frac{3 b^2}{4(k_1^2-b^2)},
\end{eqnarray}
whose general solution is a M\"obius function of one exponential so that
$v$ is a rational function of one exponential.
For $k_1^2=b^2$, i.e.~for instance for $k_1=-b$,
the ODE (\ref{eqsubeq2sol2}) integrates as
\begin{eqnarray}
& &
v=-b+\frac{2 b}{w},\
w=1 + 3 \left(1 + e^{b (z-z_0)/(2 a)} \right)^2.
\end{eqnarray}
\subsection{Subequations of degree one}
\label{sectionsubeqdeg1}
These first degree subequations
\begin{eqnarray}
& &
F(u,u') \equiv
a u' + u^2 + b_1 u + b_0=0,
\label{eqsubeqODEOrderOnePP1}
\end{eqnarray}
are determined by requiring their vanishing
when $u$ is the Laurent series (\ref{eqODE3Laurent}).
This results in
\begin{eqnarray}
& &
\left\lbrace
\begin{array}{ll}
\displaystyle{
S_{1}:\
c_1,c_2,c_4,c_5=\hbox{arbitrary},\
}\\ \displaystyle{\phantom{1234}
b_1=\frac{2 c_1 - a c_2}{12 a^2},
}\\ \displaystyle{\phantom{1234}
b_0=\frac
{44 c_1^2 - 32 a c_1 c_2 + 5 a^2 c_2^2 -144 a^3 c_4 +144 a^4 c_5}
{1152 a^4},
}\\ \displaystyle{\phantom{1234}
1152 a^6 c_6=-56 c_1^3 + 60 a c_1^2 c_2 -18 a^2 c_1 c_2^2
+a^3 c_2^3 +288 a^3 c_1 c_4
}\\ \displaystyle{\phantom{1234567890123}
-144 a^4 c_2 c_4-96 a^4 c_1 c_5 +48 a^5 c_2 c_5,
}\\ \displaystyle{\phantom{1234}
2^{13} 3^2 a^8 c_7=-176 c_1^4 +128 a c_1^3 c_2 +24 a^2 c_1^2 c_2^2 -32 a^3 c_1 c_2^3
}\\ \displaystyle{\phantom{1234567890123}
+5 a^4 c_2^4 +2688 a^3 c_1^2 c_4
-1536 a^4 c_1 c_2 c_4 +96 a^5 c_2^2 c_4
}\\ \displaystyle{\phantom{1234567890123}
-6912 a^6 c_4^2 +128 a^4 c_1^2 c_5 -512 a^5 c_1 c_2 c_5 + 224 a^6 c_2^2 c_5
}\\ \displaystyle{\phantom{1234567890123}
+ 4608 a^7 c_4 c_5 + 2304 a^8 c_5^2.
}
\end{array}
\right.
\label{eqSolutionOneExp}
\end{eqnarray}
The solution of this Riccati equation is
either a rational function of one exponential or a rational function,
\begin{eqnarray}
& &
u=
\left\lbrace
\begin{array}{ll}
\displaystyle{
-\frac{b_1}{2} + a \frac{k}{2} \coth \frac{k}{2} (z-z_0),\
k^2=\frac{b_1^2 - 4 b_0}{2 a^2}\not=0,
}\\ \displaystyle{
-\frac{b_1}{2} + \frac{a}{z-z_0},\ b_1^2 - 4 b_0=0.
}
\end{array}
\right.
\label{eqRational}
\end{eqnarray}
\medskip
\noindent
{\bf Acknowledgement.}
The authors would like to thank the referee for the valuable suggestions.
|
\section{Introduction}
A set $P$ in $\ensuremath{\mathbb{R}}^d$ is said to be a \notion{$d$-dimensional polyhedron} if $P$ is intersection of finitely many closed half-spaces and has non-empty interior. Thus, a natural description of $P$ is given in terms of its $\mathcal{H}$-{\em representation}, i.e. a collection of non-strict affine inequalities. As was suggested in \cite{MR1976602}, \cite{MR2166533} and \cite{Henk06PolRep}, one could try to use more general polynomial representations of $P$, called $\mathcal{P}$-{\em representations}, for possibly developing new effective methods for solving combinatorial optimization problems. That is, one may look for representations of the form
\begin{equation*}
P = \{p_1 \ge 0, \ldots, p_n \ge 0 \} := \setcond{x \in \ensuremath{\mathbb{R}}^d}{p_1(x) \ge 0,\ldots,p_n(x) \ge 0},
\end{equation*}
where $p_1,\ldots,p_n$ are real polynomials in $d$ variables. Naturally, one tries to find $\mathcal{P}$-representations of $P$ with a small number $n$ of polynomials.
Possible choices of $n$, also for the more general case when $P$ is a semi-algebraic set, can be derived using results from real algebraic geometry, see \RAGbook, \cite{MR1393194}. It has been recently established that every $d$-dimensional polyhedron has a $\mathcal{P}$-representation with $d$ polynomials, see the end of this section for
references and further details.
In this note we are interested in small $\mathcal{P}$-representations of polyhedra consisting of polynomials of bounded degree.
We present a result in the $2$-dimensional case, namely, we determine asymptotically the minimal number of polynomials of degree at most $k$ which are
needed to describe a $2$-dimensional polyhedron with $m$ edges, for all degrees $k\le m/\log_2 m$. This improves earlier work by \cite{HenkMatzke}, see below.
Let us call a $2$-dimensional polyhedron $P$ a \notion{polygon} (thus, we also allow unbounded polygons). Besides polynomial representations of $P$ we shall also investigate polynomial representations of the interior $\mathop{\mathrm{int}}\nolimits P$ of $P$ in terms of strict polynomial inequalities. Our main result is the following:
\begin{theorem} \label{main:cor}
Let $m, k \in \ensuremath{\mathbb{N}}$. Let $N(m,k)$ (resp. $\overline{N}(m,k)$) denote the minimal $n \in \ensuremath{\mathbb{N}}$ such that for every polygon $P$
with $m$ edges there exist real polynomials $p_1,\ldots,p_n$ satisfying the conditions conditions (A) and (B) (resp.
($\overline{A}$) and (B)), where
\begin{enumerate}
\item[(A)] \label{repr:cond:open} $\mathop{\mathrm{int}}\nolimits P=\rep{p_1}{p_n}{>}$;
\item[($\overline{A}$)] \label{repr:cond:closed} $P=\rep{p_1}{p_n}{\ge}$;
\item[(B)] \label{prod:cond:1:open} every $p_i$ has degree at most $k.$
\end{enumerate}
Then for all $k$ satisfying $$k \le m / \log_2 m$$ one has
$$m/k \le N(m,k)\le (1+ \varepsilon_m) m/k,$$ where $\varepsilon_m \rightarrow 0$ as $m \rightarrow +\infty$.\\
The same statement holds with $\overline{N}(m,k)$ instead of $N(m,k)$.
\end{theorem}
The upper bound in Theorem \ref{main:cor} will be established by using polynomials $p_i$ each of which being a product of affine
functions. In fact, for such polynomials we are able to determine their minimal number $n$ asymptotically without
any restriction on $k$:
\begin{theorem} \label{main:thm}
Let $m, k \in \ensuremath{\mathbb{N}}$. Let $N(m,k)$ (resp. $\overline{N}(m,k)$) denote the minimal $n \in \ensuremath{\mathbb{N}}$ such that for every polygon $P$ with $m$ edges there exist real
polynomials $p_1,\ldots,p_n$ satisfying the conditions conditions (A) and (C) (resp.
($\overline{A}$) and (C)), where
\begin{enumerate}
\item[(A)] \label{repr:cond:open} $\mathop{\mathrm{int}}\nolimits P=\rep{p_1}{p_n}{>}$;
\item[($\overline{A}$)] \label{repr:cond:closed} $P=\rep{p_1}{p_n}{\ge}$;
\item[(C)] \label{prod:cond:1:open} every $p_i$ is a product of at most $k$ affine functions each non-negative on
$P$ and vanishing on some edge of $P$.
\end{enumerate}
Then, with $s(m,k) := \max \{m/k, \log_2 m \}$,
$$s(m,k) \le N(m,k)\le \overline{N}(m,k) \le (1+ \varepsilon_m) s(m,k),$$
where $\varepsilon_m \rightarrow 0$ as $m \rightarrow +\infty.$
\end{theorem}
Prior to Theorem \ref{main:thm}, an upper bound $\overline{N}(m,k)\le m/k+\log_2 k$ was observed in \cite{HenkMatzke}
(see also \cite[Satz\,2.4]{Henk06PolRep}), which, together with the lower bound in Theorem \ref{main:thm}, determines
$\overline{N}(m,k)$ already up to a factor of $2$.
Theorem \ref{main:cor} will follow from Theorem \ref{main:thm}. The proof of Theorem~\ref{main:thm} consists of a geometric and a combinatorial part. In the geometric part, given in Section~\ref{geometric arguments}, we show that essentially it suffices to
consider unbounded polygons whose two unbounded edges are not parallel and express the minimal $n$ from Theorem~\ref{main:thm} for such polygons in purely combinatorial terms. In the combinatorial part, given in Section~\ref{combinatorial arguments}, we estimate this minimal $n$ by employing results on Gray codes (cf. \cite{MR1491049} and \cite{Knuth05}).
By Theorem~\ref{main:cor}, $N(m,k) \sim \overline{N}(m,k) \sim m/k$ for $k \le m/\log_2 m$ as $m \rightarrow \infty.$ Let us give a few remarks for other choices of $k.$ We only discuss $N(m,k)$ here, since the comments on $\overline{N}(m,k)$ are analogous. Using results from real algebraic geometry, one can show that there exists a function $k_0(m)$ such that $N(m,k) = 2$ for every $k \ge k_0(m)$. This follows directly from the Theorem of Br\"ocker and Scheiderer \RAGbook[\S\,6.5,\,\S\,10.4] and \cite[Theorem~5.1]{PrestelBooklet}, which are formulated in the framework of real closed fields. One should point out that the existence of $k_0(m)$ is a non-constructive result. In fact, on the one hand it is known that every $m$-gon $P$ can be described by two polynomials $p_1, p_2$, see \cite{Bernig98}. On the other hand, known formulas for $p_1, p_2$ depend on the metric structure of the $m$-gon $P$ so that the degrees of $p_1, p_2$ are not bounded from above for any fixed $m.$ Already for the case $m=5$ it is not clear how to (efficiently) construct polynomials $p_1, p_2$ representing an arbitrary given pentagon $P$ and having degrees bounded from above by an absolute constant, for further details see \cite{AveHenkDCG}. Finally, it would be interesting to estimate $N(m,k)$ for $m/\log_2 m \le k \le k_0(m)$, where $N(m,k)$ changes from about $\log_2 m$ to $2.$
As for $\mathcal{P}$-representations of polyhedra of arbitrary dimensions, we point out that recently it was shown \cite{AveBroe10}
that every $d$-dimensional polyhedron can be represented by $d$ non-strict polynomial inequalities. This provides a positive answer to a conjecture posed in \cite{MR2166533}. We also refer to
\cite{vomHofe}, \cite{Bernig98}, \cite{MR2166533}, \cite{AveHenkDCG}, \cite{AveHenkRepSimplePolytopes} for preliminary results,
\cite{AveSomePropSemiAlg}, \cite{AveRepElemSemiAlg} for related results, and \cite{Henk06PolRep} for a survey on that topic. For the case $d\ge 3$ currently no results on the interplay between the degrees and the number of polynomials in $\mathcal{P}$-representations of $d$-dimensional polyhedra seem available.
\section{Geometric arguments} \label{geometric arguments}
Let $P$ be a polygon in $\ensuremath{\mathbb{R}}^2.$ By $E$ we denote the set of all edges of $P$. For every edge $I \in E$,
we fix a real polynomial $p_I(x)$ of degree one, vanishing on $I$ and non-negative on $P$.
The following proposition is easily derived by standard arguments from real algebraic geometry.
For completeness we give a proof (see also \cite[Prop. 2.1]{MR1976602} and \cite{AveSomePropSemiAlg}).
\begin{proposition} \label{08.07.22,14:01}
Let $P$ be a polygon and let $p_1,\ldots,p_n$ be real polynomials such that $P=\{p_1\ge 0,\dots,p_n\ge 0\}$ or $\mathop{\mathrm{int}}\nolimits P = \rep{p_1}{p_n}{>}$. Then for every
$I \in E$ the polynomial $p_I$ is a factor of odd multiplicity of some $p_i$, $1\le i\le n$.
\end{proposition}
\begin{proof}[{\bf Proof}]
Fix an edge $I\in E$ and let $\prod_{i=1}^n p_i=p_I^{e}\prod_{j=1}^m q_j$ with $e\in\ensuremath{\mathbb{N}}\cup\{0\}$ and irreducible polynomials
$q_j$ not divisible by $p_I$, $1\le j\le m$. Since every $q_j$ has only finitely many zeros in $\{p_I=0\}$ there exists an
$x\in I$ with $q_j(x)\not=0$ for all $j=1,\dots,m$, and hence, by continuity, there is an neighboorhood $\mathcal{N}$
of $x$ on which for all $j=1,\dots,m$ the sign of $q_j$ is constant.
If $p_I$ would have even multiplicity in every $p_i$ then the sign of every $p_i$ would be constant on
$\mathcal{N}\setminus\{p_I=0\}$ so that, in both cases $P=\{p_1\ge 0,\dots,p_n\ge 0\}$ and $\mathop{\mathrm{int}}\nolimits P = \rep{p_1}{p_n}{>}$,
either $\mathcal{N}\setminus\{p_I=0\}\subseteq P$ or $(\mathcal{N}\setminus\{p_I=0\})\cap \mathop{\mathrm{int}}\nolimits P=\emptyset$.
But this is impossible since $x$ is a boundary point of $P$ and
hence $\mathcal N \cap\{p_I<0\}\not\subseteq P$ and $\mathcal N \cap\{p_I>0\}\cap \mathop{\mathrm{int}}\nolimits P\not=\emptyset$.
\end{proof}
Recall that a real polynomial is called {\em squarefree} if every irreducible factor of it occurs with multiplicity one.
\begin{lemma}\label{lemma:squarefree}
Let $m,k\in\ensuremath{\mathbb{N}}$. The quantity $N(m,k)$ (resp. $\overline{N}(m,k)$) from Theorem \ref{main:thm} is attained at
squarefree polynomials $p_1,\dots,p_n$ satisfying $(A)$ and $(C)$ (resp. $\overline{A}$ and $(C)$).
\end{lemma}
\begin{proof}[{\bf Proof}]
Given a polygon $P$ with $m$ edges and polynomials $p_1,\ldots,p_n$ satisfying the conditions ($A$) and ($C$), we can clearly
replace factors of $p_1,\ldots,p_n$ with odd multiplicities by the same factors with multiplicity one while maintaining the conditions (A) and (C).
Furthermore, factors
with even multiplicities can just be dropped out. In fact, if for some edge $I$, $p_I$ is a factor of $p_1$ with
even multiplicity, say with multiplicity two, then, in view of Proposition~\ref{08.07.22,14:01}, $p_I$ is a factor of odd multiplicity of some $p_i$, $1<i\le n$, say $p_2$, so that $\{p_1>0,p_2>0\}=\{p_1 / p_I^{2}>0,p_2>0\}$.
Similarly, given polynomials $p_1,\ldots,p_n$ satisfying the conditions ($\overline{A}$) and ($C$), odd multiplicities
can be replaced by multiplicity one, and factors with even multiplicities can be dropped out, since, with the notation above,
$P\subseteq \{p_1/p^2_I\ge 0, p_2\ge0\}\subseteq\{p_1\ge 0,p_2\ge 0\}$.
Performing the above reductions we obtain polynomials $p_1,\dots,p_n$ each of which being a product of at most $k$ different
$p_I$, $I\in E$.
\end{proof}
For every set $F \subseteq E$ of edges of the polygon $P$ we put
$$p_F(x) := \prod_{I \in F} p_I(x).$$
For $x \in \ensuremath{\mathbb{R}}^2 \setminus P$ we define the set $E_{<}(x) := \setcond{I \in E}{p_I(x) <0}.$ Geometrically, $E_{<}(x)$ is the set of edges of $P$ illuminated from $x.$ The sets $E_{\le }(x)$ and $E_{=}(x)$ are defined analogously.
\begin{lemma} \label{illum:crit}
Let $P$ be a polygon and let $F_1,\ldots,F_n \subseteq E$. Then the following statements hold.
\begin{enumerate}[I.]
\item \label{08.07.22,13:59} If $P = \rep{p_{F_1}}{p_{F_n}}{\ge}$ then $\mathop{\mathrm{int}}\nolimits P = \rep{p_{F_1}}{p_{F_n}}{>}.$
\item \label{08.07.22,14:00} If $\mathop{\mathrm{int}}\nolimits P = \rep{p_{F_1}}{p_{F_n}}{>}$ then $E = F_1 \cup \ldots \cup F_n.$
\item \label{illum:crit:nonneg} $P=\rep{p_{F_1}}{p_{F_n}}{\ge}$ if and only if for every $x \in \ensuremath{\mathbb{R}}^2 \setminus P$ there exists $i \in \{1,\ldots,n\}$ such that $\mathop{\mathrm{card}}\nolimits (F_i \cap E_{<}(x))$ is odd and $F_i \cap E_{=}(x) = \emptyset.$
\item \label{illum:crit:pos} $\mathop{\mathrm{int}}\nolimits P = \rep{p_{F_1}}{p_{F_n}}{>}$ if and only if for every $x \in \ensuremath{\mathbb{R}}^2 \setminus P$ there exists $i \in \{1,\ldots,n\}$ such that $\mathop{\mathrm{card}}\nolimits (F_i \cap E_{<}(x) )$ is odd.
\end{enumerate}
\end{lemma}
\begin{proof}[\bf{Proof}]
The proof is straightforward. Part~\ref{08.07.22,14:00} follows directly from Proposition~\ref{08.07.22,14:01}. Let us show only
\ref{illum:crit:pos}. We start with the ``only if'' part.
Let $x \in \ensuremath{\mathbb{R}}^2 \setminus P.$ If $p_I(x) \ne 0$ for every $I \in E$ then, by assumption,
there exists $i \in \{1,\ldots,n\}$ such that $p_{F_i}(x) <0$ and consequently $\mathop{\mathrm{card}}\nolimits(F_i \cap E_{<}(x) )$ is odd.
If $p_I(x)=0$ for some $I \in E$ then we can choose an $y \in \ensuremath{\mathbb{R}}^d \setminus P$
with $E_{<}(y)=E_{<}(x)$ and $p_J(y) \ne 0$ for every $J \in E$. Thus, by the previous case, there exists
$i \in \{1,\ldots,n\}$ such that $\mathop{\mathrm{card}}\nolimits(F_i \cap E_{<}(x) )$ is odd. Next we show the ``if'' part. Since obviously
$\mathop{\mathrm{int}}\nolimits P\subseteq \{p_{F_1}>0,\dots,p_{F_n}>0\}$, we have to show that for every $x\in\ensuremath{\mathbb{R}}^2\setminus\mathop{\mathrm{int}}\nolimits P$ there exists
$i\in\{1,\dots,n\}$ with $p_{F_i}(x)\le 0$. If $x\notin P$ this is clear by assumption. If
$x\in P\setminus\mathop{\mathrm{int}}\nolimits P$ we have $x\in I$ for some $I\in E$, and we can choose an $y\in \ensuremath{\mathbb{R}}^2\setminus P$ with
$E_{<}(y)=\{I\}$. Then, by assumption, there exists $i\in\{1,\dots,n\}$ with $I\in F_i$, so $p_{F_i}(x)=0$.
\end{proof}
We remark that the converse of Lemma~\ref{illum:crit}.\ref{08.07.22,13:59} is not true in general. For example, the interior of the quadrant $P:=\setcond{x \in \ensuremath{\mathbb{R}}^2}{x_1 \ge 0, x_2 \ge 0}$ is equal to $$\setcond{x \in \ensuremath{\mathbb{R}}^2}{x_1 > 0, \ x_2 > 0}= \setcond{x \in \ensuremath{\mathbb{R}}^2}{x_1 > 0, \ x_1 x_2 >0},$$ and relaxing the inequalities on the right hand side gives
$\setcondbegin{x \in \ensuremath{\mathbb{R}}^2}{x_1 \ge 0,} \setcondend{x_1 x_2 \ge 0}$, the union of $P$ with the $x_1$-axis.
By $n(P,k)$ (resp. $\overline{n}(P,k)$) we denote the minimal $n \in \ensuremath{\mathbb{N}}$ such that there exist $F_1,\ldots,F_n \subseteq E$, each $F_i$ of cardinality at most $k$, such that $\mathop{\mathrm{int}}\nolimits P=\rep{p_{F_1}}{p_{F_n}}{>}$ (resp. $P=\rep{p_{F_1}}{p_{F_n}}{\ge}$).
Clearly, in view of Lemma~\ref{lemma:squarefree}, the quantity $N(m,k)$ (resp. $\overline{N}(m,k)$) in Theorem~\ref{main:thm} is the maximum of $n(P,k)$ (resp. $\overline{n}(P,k)$) over all polygons $P$ with $m$ edges.
We can now state our two main propositions in this section.
\begin{proposition} \label{08.07.22,15:12} Let $P$ be an unbounded polygon with $m$ edges such that the unbounded edges of $P$ are not parallel. Then the quantities $n(P,k)$ and $\overline{n}(P,k)$ depend only on $m$ and $k.$ Furthermore, $n(m,k):=n(P,k)$ and $\overline{n}(m,k):=\overline{n}(P,k)$ are determined as follows:
\begin{enumerate}[I.]
\item \label{09.02.19,17:05} $n(m,k)$ is the minimal $n \in \ensuremath{\mathbb{N}}$ such that there exist sets $S_1,\ldots, S_n \subseteq \{1,\ldots,m\}$ satisfying the following conditions:
\begin{enumerate}
\item[(I)] \label{n_1 odd cov} for every $a, b \in \ensuremath{\mathbb{N}}$ with $1 \le a \le b \le m$ there exists $i \in \{1,\ldots,n\}$ such that $\mathop{\mathrm{card}}\nolimits (S_i \cap \{a,\ldots,b\})$ is odd;
\item[(K)] \label{n_1 card bound} $\mathop{\mathrm{card}}\nolimits S_i \le k$ for all $i \in \{1,\ldots,n\}$.
\end{enumerate}
\item \label{08.07.31,13:46} $\overline{n}(m,k)$ is the minimal $n \in \ensuremath{\mathbb{N}}$ such that there exist sets $S_1,\ldots,S_n \subseteq \{1,\ldots,m\}$ satisfying the following conditions:
\begin{enumerate}
\item[(J)] \label{n_2 odd cov} for every $a, b \in \ensuremath{\mathbb{N}}$ with $1 \le a \le b \le m$ there exists $i \in \{1,\ldots,n\}$ such that $\mathop{\mathrm{card}}\nolimits (S_i \cap \{a,\ldots,b\})$ is odd and $\{a-1,b+1 \} \cap S_i = \emptyset$;
\item[(K)] \label{n_2 card bound} $\mathop{\mathrm{card}}\nolimits S_i \le k$ for all $i \in \{1,\ldots,n\}$.
\end{enumerate}
\end{enumerate}
\end{proposition}
\begin{proof}[{\bf Proof}]
Let $I_1, I_2,\dots, I_m$ denote the edges of $P$ in consecutive order. Thus, $I_1$ and $I_m$ are the non-parallel unbounded
edges. By identifying edge sets $F_1,\dots, F_n\subseteq E=\{I_1,\dots,I_m\}$ with subsets
$S_1,\dots,S_n\subseteq\{1,\dots,m\}$ in the obvious way, part \ref{09.02.19,17:05} follows from Lemma~\ref{illum:crit}.\ref{illum:crit:pos} and
the equality
\[
\setcond{E_<(x)}{x \in \ensuremath{\mathbb{R}}^2 \setminus P} = \setcond{\{I_a,\ldots,I_b\}}{1 \le a \le b \le m},
\]
and part \ref{08.07.31,13:46} follows from Lemma~\ref{illum:crit}.\ref{illum:crit:nonneg} and the equality
\begin{multline*}
\setcond{\big(E_<(x), E_=(x)\big)}{x \in \ensuremath{\mathbb{R}}^2 \setminus P} = \\
\setcond{\big(\{I_a,\ldots,I_b\},F\big)}{1 \le a \le b \le m,\, F\subseteq\{I_{a-1},I_{b+1}\}\cap E}.
\end{multline*}
\end{proof}
\begin{proposition} \label{08.11.12,12:28} For $m, k \in \ensuremath{\mathbb{N}}$ let $N(m,k)$ and $\overline{N}(m,k)$ be defined as in Theorem~\ref{main:thm}, and $n(m,k)$ and $\overline{n}(m,k)$ as in Proposition~\ref{08.07.22,15:12}. Then one has
\begin{align*}
n(m,k) & \le N(m,k) \le 1+n(m-1,k), \\
\overline{n}(m,k) & \le \overline{N}(m,k) \le 1+\overline{n}(m-1,k).
\end{align*}
\end{proposition}
\begin{proof}[{\bf Proof}]
The lower bounds for $N(m,k)$ and $\overline{N}(m,k)$ are clear by Proposition \ref{08.07.22,15:12} and the remark
preceding it. As for the upper bounds consider an arbitrary polygon $P$ with $m+1$ consecutive edges $I_0,\ldots,I_m$.
Notice that
\begin{equation*}
\setcond{E_<(x)}{x \in \ensuremath{\mathbb{R}}^2 \setminus P, \ I_0 \not\in E_<(x) } = \setcond{\{I_a,\ldots,I_b\}}{1 \le a \le b \le m}.
\end{equation*}
Consider sets $S_1,\ldots,S_n \subseteq \{1,\ldots,m\}$ satisfying {\itshape(I)} and {\itshape(K)}. As in the proof of
Proposition~\ref{08.07.22,15:12} we identify each set $S_i$ with the corresponding edge set $F_i \subseteq \{I_1,\ldots,I_m\}$.
In view of Lemma~\ref{illum:crit}.\ref{illum:crit:pos} we have $\mathop{\mathrm{int}}\nolimits P = \{p_{I_0} > 0, p_{F_1}>0,\ldots,p_{F_n}>0\}$.
Hence $N(m+1,k) \le 1+ n(m,k).$ The upper bound for $\overline{N}(m,k)$ is proved analogously and employs Lemma~\ref{illum:crit}.\ref{illum:crit:nonneg}.
\end{proof}
\section{Combinatorial arguments and conclusion} \label{combinatorial arguments}
In view of the results from Section~\ref{geometric arguments}, the proofs of Theorem~\ref{main:cor} and \ref{main:thm}
require the determination of the asymptotic behavior of the functions $n(m,k)$ and $\overline{n}(m,k)$
(defined in Proposition \ref{08.07.22,15:12}).
We will show the following.
\begin{theorem}~\label{comb_prop}
Let $m, k \in \ensuremath{\mathbb{N}}$, $k \le m.$ Then
\begin{align*}
\max\{m/k, \log_2 (m\!+\!1)\} \le n(m,k) \le \overline{n}(m,k)\le
\max \left\{\frac{m}{k} + 3 \log_2 \frac{m}{k} +C,\,\ceil{\log_2(m\!+\!1)}\right\},
\end{align*}
where $C>0$ is an absolute constant.
\end{theorem}
In particular, for every $\varepsilon>0$ and all $(1+\varepsilon) m / \log_2 m \le k \le m$ with $m$ sufficiently large, we have the exact
values $\overline{n}(m,k)=n(m,k) = \ceil{\log_2(m+1)}$.
The proof of Theorem~\ref{comb_prop} employs results on Gray codes and is based on the following
reformulations for $n(m,k)$ and $\overline{n}(m,k)$.
For subsets $S_1,\dots,S_n$ of $\{1,\dots,m\}$ let $M$ denote their incidence matrix, i.e., the $n\times m$-matrix whose
$(i,j)$-entry is $1$ if $j\in S_i$ and $0$ otherwise. Let $c_j$ be the $j$-th column of $M$, and let $M'$ denote the binary
$n\times(m+1)$ matrix whose $j$-th column is $c'_j$ is $c_1+\dots+c_j \modulo{2}$, $j=0,\dots,m$. In particular, $c'_0$ is the zero column. Clearly, $c_j=c'_{j-1}+c'_j \modulo{2}$ for all $j=1,\dots,m$.
Now it is easy to see that the conditions $(I)$, $(J)$ and $(K)$ (in Proposition \ref{08.07.22,15:12}) on the
sets $S_1,\dots,S_n$ are equivalent to the following conditions $(I')$, $(J')$ and $(K')$, respectively, on the matrix $M'$:
\begin{enumerate}
\item[$(I')$] \label{gray cond 1} the $m+1$ columns of $M'$ are pairwise distinct;
\item[$(J')$] \label{gray cond 2} for every $a, b \in \ensuremath{\mathbb{N}}$ with $0 \le a < b \le m$ there exists
$i \in \{1,\ldots,n\}$ such that $M'_{i,a-1}=M'_{i,a}\neq M'_{i,b}=M'_{i,b+1}$
(with $M'_{i,-1}:=M'_{i,0}$ and $M'_{i,m+1}:=M'_{i,m}$);
\item[$(K')$] \label{gray cond 3} every row of $M'$ has at most $k$ bit changes.
\end{enumerate}
Here a {\em bit change in row $i$} of $M'$ is a column index $j\in\{1,\dots,m\}$ for which $M'_{i,j-1}\not=M'_{i,j}$.
Thus, $n(m,k)$ (resp. $\overline{n}(m,k)$) is the minimal $n\in\ensuremath{\mathbb{N}}$ such that there exits a binary $n\times(m+1)$-matrix $M'$ starting with the zero column and satisfying $(I')$ and $(K')$ (resp. $(J')$ and $(K')$).
Next we cite some facts about Gray codes (cf. \cite{MR1491049} and \cite{Knuth05}). Recall that an \emph{$n$-bit Gray code} is a binary
$n\times 2^n$-matrix $G$ with distinct columns $g_j$, $j=0,\dots 2^n-1$, such that $g_j$ and $g_{j+1}$ differ in exactly
one coordinate, $j=0,\dots,2^n-1$, where $g_{2^n}:=g_0$.
Without loss of generality we shall always assume that $g_0$ is the zero column.
A {\em bit run} of length $\ell$ of $G$ is a $1\times\ell$-submatrix $(G_{i,j+1},\dots,G_{i,j+\ell})$
of $G$ with $G_{i,j}\not=G_{i,j+1}=G_{i,j+2}=\dots=G_{i,j+\ell}\not=G_{i,j+\ell+1}$, where the columns are indexed modulo $2^n$.
The upper bound in Theorem \ref{comb_prop} essentialy follows from the following result due to Goddyn and Gvozdjak \cite{GoddynGvozdjak03}, see also
\cite[Section 7.2.1.1]{Knuth05}.
\begin{theorem} \label{gray code results} For every $n \in \ensuremath{\mathbb{N}}$ there exists an $n$-bit Gray code for which every bit run has length at least $n-3\log_2 n$. \eop
\end{theorem}
In fact, as shown in \cite{GoddynGvozdjak03}, Theorem~\ref{gray code results} holds with the slightly stronger bound $\floor{n-2.001\log_2 n}$.
We will also need the following simple observation.
\begin{lemma}\label{J'lemma}
Every $n$-bit Gray code $G$ satisfies condition (J') (with $M'=G$).
\end{lemma}
\begin{proof}[{\bf Proof}]
Fix $a,b\in\mathbb{N}$ with $0<a<b<m$. Consider the row indices $j,k\in\{1,\dots,n\}$ for which $M'_{j,a-1}\not=M'_{j,a}$ and
$M'_{k,b}\not=M'_{k,b+1}$. Then, since $M'$ is a Gray code, $M'_{i,a-1}=M'_{i,a}$ for all $i\not=j$, and $M'_{i,b}=M'_{i,b+1}$
for all $i\not=k$. In particular, the submatrix of $M'$ corresponding to the row indices $j,k$ and column
indices $a-1,a,b,b+1$ contains two equal columns (in both cases $j\not=k$ and $j=k$). Since the four columns of
$M'$ corresponding to the column indices $a-1,a,b,b+1$ are pairwise different, there exists an $i\not=j,k$ for which
$M'_{i,a-1}=M'_{i,a}\neq M'_{i,b}=M'_{i,b+1}$.
A similar argument works if $a=0$ or $b=m$.
\end{proof}
\begin{proof}[\bf{Proof of Theorem~\ref{comb_prop}}] We shall use the matrices $M$ and $M'$ introduced above.
The lower bound on $n(m,k)$ is clear in view of the conditions on $M$ and $M'$. Indeed, with $a=b$ in
condition $(I)$ we see that every element from $\{1,\dots,m\}$ is covered by at least one of the sets $S_1,\dots,S_n$,
which in view of $(K)$ yields $n\ge m/k$. (Alternatively, we can count the total number of bit changes in $M'$.)
Further, we have $n\ge\log_2(m+1)$ since the $m+1$ columns of $M'$ are distinct.
The bound $n(m,k) \le \overline{n}(m,k)$ is trivial.
It remains to show the upper bound on $\overline{n}(m,k)$. Fix the minimal $n \in \ensuremath{\mathbb{N}}$ satisfying $m\le 2^n-1$ and $m \le k(n-3\log_2n)$
and consider an $n$-bit Gray code $G$ as in Theorem~\ref{gray code results}. Then the
$n\times(m+1)$-matrix $M'$ consisting of the first $m+1$ columns of $G$ starts with the zero column and clearly satisfies
condition $(K')$. By Lemma \ref{J'lemma}, $M'$ also satisfies condition $(J')$. Thus, $n(m,k)\le n$, from which the desired upper
bound follows by elementary calculations. Indeed, if $n>\ceil{\log_2(m+1)}$ then, by minimality of $n$,
$m/k>((n-1)-3\log_2(n-1))\ge (n-1)/4$ (the last inequality for $n\ge 8$) and thus
\[
n-1<m/k+3\log_2(n-1)<m/k+3\log_2(4m/k)=m/k+3\log_2(m/k)+6.
\]
\end{proof}
\begin{proof}[\bf{Proof of Theorem~\ref{main:thm}}]
The inequality $N(m,k)\le\overline{N}(m,k)$ is clear in view of Proposition \ref{illum:crit}.\ref{08.07.22,13:59}.
The remaining inequalities follow from Proposition~\ref{08.11.12,12:28} and Theorem~\ref{comb_prop}.
\end{proof}
\begin{proof}[\bf{Proof of Theorem~\ref{main:cor}}] The lower bounds on $N(m,k)$ and $\overline{N}(m,k)$ follow from
Proposition~\ref{08.07.22,14:01}. The upper bounds follow directly from Theorem~\ref{main:thm}.
\end{proof}
\section*{Acknowledgement}
We thank Ludwig Br\"ocker, Martin Henk, Bettina Matzke and Claus Scheiderer for helpful discussions.
\small
\bibliographystyle{amsalpha}
|
\section{Introduction}
\label{sec:intro}
In many application domains, software systems need to perpetually and
rapidly evolve to cope with new user and technology requirements.
Being able to modify existing systems or redesign new systems to
rapidly take in account new functionalities or preferences has led to
the proposition of several software engineering approaches such as the
Agile software development methodology~\cite{agile}. One of the key
principles of Agile software development is to build software through
an incremental and iterative process. Each iteration adds a new
feature and produces a fully working system by going through the whole
the software lifecycle, {\em i.e.,} the analyze, develop and test
phases. Another particularity of Agile development is that the testing
activity is not just confined to the classical test phase but rather
integrated throughout the entire lifecycle, meaning that the software
is continuously tested throughout its development, from its
specifications to the final running system, in order to augment the
overall software system quality.
Another software engineering approach that contributes to facilitating
the rapid development of software systems is the use of
component-based software architectures. In this context, the overall
structure of the application is first described with an architecture
description language (ADL)~\cite{Medvidovic2000}. Such description
highlights the needed components and their assembly, which facilitates
the understanding and analysis of the application's properties, such
as behavioral or quality of service properties. If the specifications
are coherent, the application is eventually instantiated, deployed and
executed to be tested.
Although Agile software development and component-based software
engineering (CBSE) may appear quite different approaches, some
works~\cite{Stojanovic-etal-03,xpagile} have identified that both
approaches could benefit to each other, CBSE bringing for example the
capability of building large software and enhancing reusability, and
Agile development offering more flexible development processes for
shorter time-to-market products. Nevertheless we believe that there is
still a bridge between these two approaches, one reason being the lack
of component frameworks that allow incremental and iterative
development processes, as well as throughout-lifecycle testing.
To address this problematic, we have developed a model-based
framework, named CALICO, that enables architects to design and test
component-based systems in an iterative and uniformed
process~\cite{waignier-MODELS-08, waignier-QoSA-09}.
CALICO allows software architects to specify their architectures as
models, and to analyze them with respect to application and platform
constraints. Our approach enables the testing of the system throughout
the system lifecycle. More concretely, CALICO analyses architecture
models and creates contracts by composing contractual application
properties, {\em e.g.}, behavioral, dataflow, QoS properties. This
composition allows compatible and incompatible interaction to be
identified, as well as partially compatible interactions, which
require runtime checking~\cite{waignier-QoSA-08}. When runtime
checking is needed, CALICO automatically instruments the application
to reify runtime information to complete the resolution of the
partially compatible interaction contract and thus detects if the
given interaction may lead to an error. By using this framework in
iterative software design processes, architects get design feedback,
{\em i.e.,} information on identified interaction errors, and can then
modify the models accordingly. Each modification performed on the
model is propagated to the running system since CALICO ensures the
synchronization between the model and the runtime system, both of which
thus coexist during the whole application development. Furthermore,
the solution offered by CALICO is generic regarding underlying
platforms, allowing component platforms to benefit from all the
analyses integrated into CALICO.
The rest of this paper is organized as follows.
Section~\ref{sec:presentation} gives an overview of the CALICO
iterative and incremental development process.
Section~\ref{sec:example} illustrates with a concrete scenario the
CALICO approach. Finally, Section~\ref{sec:conclusion} provides some
information about the current status of our framework implementation.
\section{CALICO Overview}
\label{sec:presentation}
CALICO is composed of two levels: a model level and a platform level
as shown in Figure~\ref{fig:calico}. The model level is independent
of any component-based or service-oriented platform. It contains the
CALICO Architecture Description metamodels that enable an architect to
describe the structure and the properties, \textit{i.e.}, structural,
behavioral, dataflow and QoS properties, of an application. It is also
possible to specify some contextual adaptation rules, independently of
any platform, in order to allow the debugging of autonomic systems. The
platform level holds the running system on a target platform.
The iterative and incremental development process of CALICO,
illustrated in Figure~\ref{fig:calico}, is as follows :
{\bf (1) Design :} The architect specifies the design of the desired
application using the CALICO metamodels. The {\tt system structure}
metamodel enables architects to describe the structure of their
architecture, independently of any component platform. CALICO provides
also four contract metamodels to allow architects to specify
structural, behavioral, dataflow and QoS properties for each
component.
{\bf (2) Static analysis :} The interaction analysis tool checks the
coherence of the system architecture. For each partially compatible
interaction, a test to be performed at execution time is automatically
inserted into the CALICO {\tt debug} metamodel. For each incompatible
interaction, the architect is notified of the problem and he/she may
thus provide some modification of the application design. As long as
some incompatible interactions remain, the next steps of the
development process can not be reached. Once all of the problems are
fixed, the architect specifies the runtime platform on which the
application is to be executed and CALICO verifies that the
specifications do not go against the platform constraints in order to
make sure that the application can be indeed deployed on that specific
platform.
{\bf (3) Code generation :} If a component or service does not
already exist, then the {\tt generation} tool generates code skeleton such
that only business code needs to be provided by the developers.
{\bf (4) Instrumentalisation :} This step makes the link between the
static analysis and the dynamic checks of the application at runtime.
The {\tt instrumentation} tool takes the {\tt debug} model as input
and automatically instruments the application code to enable the
capture of the needed runtime information to complete the resolution
of the partially compatible interaction. This instrumentalisation
relies on an aspect-oriented approach and is independent of the
underlying platform.
{\bf (5) Instantiation :} The {\tt loader} instantiates the
application on the target platform as described by the architect's
structural model. Concretely, the running system is created
incrementally by calling the appropriate sequence of system
construction operations, such as creating/removing components and
connectors.
{\bf (6) Reification :} As the testers run the application in different
execution contexts, the instrumented application automatically reifies
any context changes and monitored information.
{\bf (7) Dynamic debugging :} During the debugging phase, the {\tt
debug} tool analyzes the information reified by the running system
and triggers when needed the tests contained in the debugging model.
The architect is notified each time an error is detected, allowing
him/her to correct the application design. Other debugging action
rather than the notification action maybe chosen, such as logging the
information into a file, or executing a reconfiguration script that
will automatically modify the design and trigger the step 2 of the
process. This latter case may be useful to tune/test adaptation
policies for autonomic system.
{\bf (8) Evolution of the design :} The architect can modify the design with
respect to the debugging information if problems have occurred. They
can also adapt at anytime the design of the application to address new
user or application requirements. After any modification, the
development cycle iterates again starting at step 2.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\textwidth]{calicobenevol}
\caption{Overview of CALICO iterative development process}
\label{fig:calico}
\end{figure}
\section{Illustrative Example}
\label{sec:example}
To illustrate the agile development process offered by CALICO, we use an
example of architecture in the context of the French Personal Health
Record system (PHR)~\cite{dmp}. PHR is the French personal health
record system that is intended to provide health-care
professionals with the information needed for their patients care.
Figure~\ref{fig:dmp} represents a possible architecture of the PHR system.
All medical information, (such as biological analyses, X-rays,
medications, \textit{etc.}), will be stored in distributed {\tt databases}
and will be made accessible through an on-line interface {\tt Client}.
In order to build a robust PHR application, architects need to be able
to express several application properties. A first requirement of
this system architecture is related to authentication issues since not
everybody should have access to anybody's health records. The
architecture of this system must thus provide some authentication
mechanism. The {\tt Authentication} architecture element logs a
health-care professional in and returns a session ticket through the
functionality {\tt getTicket} that is offered by {\tt SessionServer}.
For security reason, the functionality {\tt getTicket} can be used
only by the element {\tt Authentication} to avoid that an
unauthenticated user get a session ticket. Finally the session ticket
must be validated by the {\tt SessionServer} before retrieving any
medical data from the database.
Another requirement is the high reliability of the system. Such
system has to be able to handle very heterogeneous medical
information, going from light-weight text records to gigabytes of
echographies. Furthermore, the devices used to display this
information are also heterogeneous. They range from desktop computers
with high-quality large-screen monitors and gigabyte network
connexions to simple PDAs with small screens and low-bandwidth GPRS
network connexion. Handling such data in a reliable way is critical
because the system must be able to determine if a given data can be
displayed appropriately with no loss of information, as well as in
which time-frame, depending on the available resources and amount of
information to display. For example, a dataflow constraint may express
that medical data received on a terminal of type PDA Nokia N800 should
be less than 10 megaoctets. Another constraint can also specify that
only text or jpg documents can be read on that terminal.
Overall the constraints may evolve in time or just not reflect exactly
one given execution context. There is thus a need to iterate the whole
process to check if the declared application constraints can be all
checked, statically or dynamically.
\begin{figure}[!ht]
\centering
\includegraphics[width=1.0\textwidth]{archidmpbenevol.eps}
\caption{Structure of the PHR application}
\label{fig:dmp}
\end{figure}
{\bf (1)} The architect specifies the architecture and the properties
of the PHR application. For example, the first requirement mentioned
above can be specified using the structural contract metamodel.
{\bf (2)} The overall coherence of the constraints is statically
verified. A partially compatible interaction is detected between the
{\tt MedicalServer} and the {\tt PDS} since data sent by the server
could be greater than 10 megaoctets or in a different format than txt
or jpg. Accordingly CALICO adds some rules to validate in the {\tt
debug} model. These rules specify that the size and the data type
must be captured at runtime.
{\bf (3)} Code skeletons are generated and developers can provide the
business code of each components.
{\bf (4)} Following the information contained in the {\tt debug}
model, the application is automatically modified to capture the size
and the type of the medical data that enters the PDA.
{\bf (5)} The application is deployed on the target application.
{\bf (6)} At this step different execution contexts are tested. One
may consists in the use of the PHR application by a druggist, who
typically uses the PHR only to consult text documents. Another test
scenario considers a radiologist. During the test scenario execution,
monitored informations are reified.
{\bf (7)} The {\tt debug} tool resumes the interaction compatibility
checks that were partially compatible. In the case of the druggist, no
error is detected, whereas for the radiologist, the analyse indicates
that the data are too large for the PDA.
{\bf(8)} The architect can accordingly modify the application design
by inserting a new component {\tt DataConverter} between the PDA and
the {\tt GlobalSearch} component in order to reduce the size of a too
large radiography.
The whole process is then iterated again. If no error is detected
statically, the new component {\tt DataConverter} is automatically
integrated into the already deployed application, and new test
scenarios may be executed.
\section{Conclusion and Current Implementation Status}
\label{sec:conclusion}
CALICO is a model-based framework that enables the design and debug of
systems in an iterative and incremental way, bridging a little more
the gap between component-based software engineering and agile
development approaches. Our framework is generic and highly
extensible. All metamodels for specifying the structure, the
application properties and the adaptation rules are independent of any
underlying platform. This enables architects to perform various
architecture analyses on their applications even if the underlying
component or service framework does not provide any verification
tools.
The current implementation of CALICO is developed in Java. All CALICO
metamodels are implemented with the Eclipse Modeling Framework (EMF).
A graphical editor, implemented with the Graphical Modeling Framework
(GMF), enables the architect to edit the model during the whole
development cycle.
We have integrated several existing tools to verify the coherence of
the component interactions in term of structural, behavioral, dataflow
and quality of service properties. Structural constraints are
expressed in OCL~\cite{ocl}, using the EMF-OCL library. Behavioral
specifications are based on existing process algebra, such as
CSP~\cite{csp}, FSP~\cite{fsp}, SFSP. The current implementation uses
the Fractal behavioral protocol checker~\cite{fractalbpc} to verify
that a given component composition does not introduce a deadlock. We
have developed a dataflow analysis based on the algorithm of constant
propagation in partial program validation~\cite{Gary}. The QoS
metamodel has been inspired by the QML~\cite{qml} and WSLA~\cite{wsla}
approaches. The associated analysis is based on prediction of quality
property in a workflow of WEB services~\cite{10.1109/ICEBE.2007.18}.
Furthermore, application instrumentation has been implemented with
Spoon~\cite{spoon}. The sensor framework Wildcat~\cite{wildcat} has
been integrated in CALICO. Our current implementation supports four
component platforms (Fractal~\cite{FractalJava},
OpenCCM~\cite{opencom}, OpenCOM~\cite{BRICLET:2004:HAL-00003287:1} and
FraSCAti~\cite{frascati}) and one service-oriented platform(Web
services~\cite{soa}). CALICO has been carefully designed to allow new
extensions in terms of support for new platforms, new QoS sensors and
new kinds of debugging actions.
We have performed benchmarks on our implementation and showed that CALICO is usable to
design reliable large systems up-to 10000 components, which is the maximum
load of most runtime platforms.
CALICO is still being developed, to support more extensions. The
current implementation is freely available at
http://calico.gforge.inria.fr.
\section*{References}
|
\section{Introduction}
\label{sec:introduction}
Unveiling the mass distribution, dynamical structure and dark matter
content of early-type galaxies is of great interest both as a subject
in its own right, considering their importance in the local Universe,
where a large fraction of the total stellar mass is contained within
E/S0 systems \citep*{Fukugita1998}, and in order to provide stringent
tests for the galaxy formation and evolution models.
It is not surprising, therefore, that nearby early-type galaxies have
been the object of intense study during the last decades, by taking
advantage of the diverse available observational tracers. These
include stellar kinematics (see e.g. \citealt*{Saglia1992},
\citealt{Franx1994}, \citealt{Rix1997},
\citealt{Loewenstein-White1999}, \citealt{Gerhard2001},
\citealt{Borriello2003}, \citealt{Cappellari2007},
\citealt{Thomas2007b}, \citealt{Weijmans2009}), globular clusters and
planetary nebulae kinematics \citep[e.g.][]{Cote2003, Romanowsky2003,
deLorenzi2008}, the occasional $\ion{H}{I}$ disk or ring
(e.g.\ \citealt*{Franx1994}; \citealt{Weijmans2008}) and hot X-ray
emission \citep[e.g.][]{Matsushita1998, Fukazawa2006, Humphrey2006,
Humphrey-Buote2010}. The general picture emerging from many of these
studies is that the total mass density profile of ellipticals can be
well described by a power-law form close to $\rho_{\mathrm{tot}}
\propto 1/r^{2}$, often referred to as the isothermal
profile. Moreover, while the inner regions of early-type galaxies are
clearly dominated by the stellar component, the dark matter component
is usually found to play already a non negligible role, with fractions
of approximately $10$ to $40$ per cent of the total mass within an
effective radius. Studies based on stellar population and dynamical
models \citep[e.g.][]{Padmanabhan2004} indicate that the dark matter
fraction increases with the mass of the galaxy, a trend that is more
conspicuous in the case of slow-rotator ellipticals
\citep{Tortora2009}.
The analysis of early-type galaxies beyond the local Universe,
i.e. beyond redshift $z \approx 0.1$, holds great promise in view of
understanding the structural evolution of these objects, but it also
presents several difficulties which hinder the application of
traditional techniques. Stellar dynamics studies, in particular, are
limited by the degeneracy between the galaxy mass profile and the
anisotropy of the stellar velocity dispersion tensor. Taking into
account higher order velocity moments can provide a solution
\citep[see][]{Gerhard1993, vanderMarel-Franx1993, Lokas-Mamon2003},
but unfortunately carrying out such measurement for distant systems is
not viable with the current instruments on 8-10 meter class
telescopes.
However, galaxies at $z \gtrsim 0.1$ have a far greater chance of
acting as strong gravitational lenses \citep*{Turner1984}, thus
providing a very helpful additional diagnostic tool. This is
particularly valuable since it allows an accurate and robust
determination of the total mass enclosed, in projection, within the
region delimited by the Einstein radius
\citep{Kochanek1991}. Unfortunately, the diagnostic power of strong
lensing to constrain internal mass distribution of the deflector is
limited, chiefly by the mass-sheet and mass-slope degeneracies
\citep{Falco1985, Wucknitz2002}, although the latter can be partially
overcome by studying spatially extended lensed sources
\citep[e.g.][]{Warren-Dye2003, Suyu2010}. A very effective way to
overcome these difficulties and robustly recover various structural
properties of the galaxy is to combine the gravitational lensing
analysis with the complementary constraints provided by stellar
dynamics (see \citealt{Koopmans-Treu2002},
\citeyear{Koopmans-Treu2003}, \citealt{Treu-Koopmans2002b},
\citeyear{Treu-Koopmans2002a}, \citeyear{Treu-Koopmans2004},
\citealt{Barnabe-Koopmans2007}, hereafter BK07, and, e.g.,
\citealt{Rusin-Kochanek2005}, \citealt{Jiang-Kochanek2007},
\citealt{vandeVen2008}, \citealt{Trott2010}, \citealt{Grillo2010} for
further applications of this approach).
Up to very recently, the availability of only a handful of lens
galaxies suitable for the joint analysis represented a major
limitation. This concern has been dispelled by the Sloan Lens ACS
Survey, SLACS (\citealt{Bolton2006}, \citeyear{Bolton2008a},
\citeyear{Bolton2008b}, \citealt{Koopmans2006},
\citeyear{Koopmans2009}, \citealt{Treu2006}, \citeyear{Treu2009},
\citealt{Gavazzi2007}, \citeyear{Gavazzi2008}, \citealt{Auger2009}),
which has led to the discovery of a large and homogeneous sample of
almost a hundred strong gravitational lenses, mostly early-type
galaxies, in the redshift range of $z \approx 0.05 - 0.5$. For a
subset of about~$30$ SLACS systems, the data set is complemented by
two-dimensional kinematic maps of the lens obtained from spectroscopic
observations carried out either with the Very Large Telescope (VLT)
instrument VIMOS or with the Low Resolution Imager and Spectrograph
(LRIS, see \citealt{Oke1995}) mounted on the Keck-I telescope. This
has provided further motivation to expand the combined analysis
technique into a more general and self-consistent method which makes
full use of the available data sets (i.e.\ surface brightness
distribution of both the lensed source and the lens galaxy, and
two-dimensional kinematic maps of the latter), and is coherently
embedded in the framework of Bayesian statistics (BK07). The current
implementation of the method, the {\textsc{cauldron}} code --- based on the
assumptions of axial symmetry and two-integral stellar DF for the lens
galaxy --- has been used to conduct an in-depth study of a SLACS
subsample of six systems representative of the survey in terms of
redshifts and velocity dispersions (see \citealt{Czoske2008} and
\citealt{Barnabe2009}, hereafter B09). As shown in those works, the
more sophisticated approach makes it possible to extract much more
information out of the data set, allowing to recover, in addition to
the slope of the total density profile, several other important
properties of the lens galaxies, including the flattening of the
density distribution, lower limits for the dark matter fraction at
different radii and insights on the dynamical structure (angular
momentum, anisotropy, contribution of rotation and random
motions). These quantities are all of relevance to the formation
history of these galaxies.
In this paper we carry out a detailed combined lensing and dynamics
analysis of the SLACS system SDSS\,J0728$+$3835, employing the
{\textsc{cauldron}} algorithm. The lens is an early-type galaxy at $z =
0.206$, with an aperture averaged velocity dispersion $\sigma = 214
\pm 11$ km s$^{-1}$\ measured from SDSS spectroscopy and a half-light radius
$R_{\mathrm{e}} = 1.78 \arcsec$ in the $I$~band. The background source is
located at $z = 0.688$ and the Einstein radius is $R_{\mathrm{Einst}} = 1.25
\arcsec$. With respect to the systems considered in \citet{Czoske2008}
and B09, all followed-up with VLT VIMOS integral-field unit, the main
difference in the observables lies in the kinematic data set: the
{SDSS\,J0728} velocity moments maps are obtained from LRIS Keck long-slit
spectroscopic observations, using three slits parallel to the major
axis, and offset along the minor axis, in order to mimic
integral-field capabilities. Remarkably, this kinematic data set
extends significantly farther than those of the previously examined
systems, reaching up to 1.7~$R_{\mathrm{e}}$, thus providing us important
constraints beyond the inner regions of the galaxy. Moreover, for
the first time, we use the stellar masses determined from stellar
population analysis \citep{Auger2009} to set the normalization of the
luminous mass distribution of the lens galaxy, enabling us to
disentangle the luminous and dark matter contributions and to compare
different choices of the initial mass function (IMF).
The paper is organized as follows: after introducing the data set in
Section~\ref{sec:observations}, we present and discuss the results of
the combined analysis in Section~\ref{sec:analysis} and draw
conclusions in Section~\ref{sec:conclusions}.
Throughout this paper we adopt a concordance $\Lambda$CDM model
described by $\Omega_{\mathrm{M}}=0.3$, $\Omega_{\Lambda} = 0.7$ and
$H_{0} = 100\,h\,\mathrm{km\,s^{-1}\,Mpc^{-1}}$ with $h=0.7$, unless
stated otherwise.
\section{Observations}
\label{sec:observations}
\subsection{High-resolution imaging data}
The lensing analysis requires deep high-resolution imaging data and
this is provided by the SLACS survey's {\it Hubble Space Telescope}
({\it HST}) imaging. In particular, SLACS has obtained one {\it HST}
orbit ($\approx$ 2200s) of data in the F814W filter. The data are
processed as described in \citet{Bolton2008a}; to briefly summarize,
the four individual exposures are background subtracted, cosmic
ray-cleaned, registered, resampled to an output grid with square
pixels that are 0\farcs05 on a side, and stacked with an additional
cosmic ray-rejection step. Synthetic point-spread function (PSF)
images created with TinyTim are likewise resampled and combined to
create a composite model PSF for the output image.
The light distribution of the galaxy is then fit with a B-spline model
\citep[e.g.,][]{Bolton2008a} and this model, convolved with the seeing
present during the spectroscopic observations, is used as the input
surface brightness distribution for the dynamics modelling. The
residual image (the B-spline model subtracted from the data image)
contains the flux from the lensed background source and is used as
lensing constraint on the mass model of the lens galaxy (see BK07 for
details). The galaxy-subtracted image is shown in the bottom
right-hand panel of Fig.~\ref{fig:J0728_LEN}.
\subsection{Pseudo-IFU Spectroscopy}
\label{ssec:IFU}
Our previous joint lensing and dynamics studies used VLT-VIMOS
integral field unit (IFU) spectroscopy to generate resolved stellar
kinematic maps. In this study we employ `pseudo-IFU' spectroscopy,
which consists of dithering a traditional longslit in the spectral
direction over the target galaxy to provide spatially-resolved
kinematic data perpendicular to the slit direction in addition to the
spatial information that is obtained along each slit (see Figure
\ref{fig:pseudoIFU}). This technique yields significantly fewer
spaxels per exposure than proper IFU spectroscopy but the throughput
from the longslit is a factor of a few larger than the IFU throughput
and we are therefore able to efficiently produce kinematic maps,
despite the less efficient way of sampling the kinematic
field. Additionally, we are more sensitive to the low-surface
brightness outer regions of the galaxy and can therefore extend our
kinematic maps to larger radii than is possible with the IFU data.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth,clip]{SDSSJ0728_pseudoifu.eps}
\end{center}
\caption{{\it HST} F814W image of {SDSS\,J0728} with the slit locations
overplotted as solid red lines. The most narrow slit (0\farcs7) is
centered on the light distribution and follows the semi-major
axis. The other slits are 1\arcsec and 1\farcs5 wide and are
offset so as to be adjacent to the previous slit. The dotted
red lines delineate the 7 apertures with width 0\farcs43 that were
extracted from each longslit observation.}
\label{fig:pseudoIFU}
\end{figure}
The spectroscopic data for SDSS J0728 were obtained with LRIS on Keck
I during the nights of 22 and 23 December 2006. The observing
conditions were clear with $\approx 0\farcs8$ seeing, and we used the
460 dichroic to split the beam to the blue and red sides of the
spectrograph. Here we only use data from the red side, which employed
the 900/5500 grating with a dispersion scale of 0.85\,\AA
pixel$^{-1}$. Three slit positions aligned along the semi-major axis
of the galaxy were observed, including: two 900s exposures with a
0\farcs7 slit positioned on the center of the galaxy; four 1200s
exposures with a 1\arcsec slit offset 0\farcs85 from the center of the
galaxy; and 4 1800s exposures with a 1\farcs5 slit offset 2\farcs1
from the center of the galaxy. The data were reduced using custom
Python scripts \citep[for details see][]{Suyu2010} and one dimensional
spectra were extracted from apertures of width 2 pixels ($\approx
0\farcs43$) at seven points along each slit as indicated in Figure
\ref{fig:pseudoIFU}, yielding 21 spectra in total (e.g., Figure
\ref{fig:keckSpectra}).
\begin{figure*}
\begin{center}
\includegraphics[width=0.3\textwidth,clip]{SDSSJ0728_central.eps}
\includegraphics[width=0.3\textwidth,clip]{SDSSJ0728_offset_10.eps}
\includegraphics[width=0.3\textwidth,clip]{SDSSJ0728_offset_15.eps}
\end{center}
\caption{Extracted one-dimensional spectra for the central aperture
of the 0\farcs7 slit (left), 1\arcsec slit (center), and 1\farcs5
slit (right). The red line is the best-fit model, the green line
is the polynomial continuum model, and the lower panel shows the
residuals. Greyed-out regions indicate parts of the spectra that
were not included in the fit (two strong night-sky features, the
Mgb line of {SDSS\,J0728}, and the \ion{O}{II} features from the
background source).}
\label{fig:keckSpectra}
\end{figure*}
The stellar velocity dispersion and velocity offset was computed for
each spectrum as in \citet{Suyu2010} using the rest-frame wavelength
range 4200\,\AA\ to 5450\,\AA\ and errors were determined from Markov
chain Monte Carlo simulations. The resulting velocity and velocity
dispersion profiles for each slit position are shown in Figure
\ref{fig:velocitydata}. We compare our data with the velocity
dispersion derived from the SDSS spectrum by combining the 14 spectra
from the inner two slits and determining a composite velocity
dispersion. This is found to be 236 km s$^{-1}$, rather larger than the value
of 214 km s$^{-1}$\ found for the SDSS spectrum. We have re-analysed the SDSS
spectrum and find $\sigma = 212$ km s$^{-1}$, but if we mask the \ion{O}{II}
lines (as is done in the analysis of the Keck spectra) we find $\sigma
= 225$ km s$^{-1}$. This is still somewhat lower than, although consistent
within the errors with, the central velocity dispersion derived from
the composite Keck spectrum. However, we might expect the SDSS value
to be lower due to poor seeing pushing more flux from large radii
(where the velocity dispersion is lower) into the SDSS fibre aperture.
\begin{figure*}
\begin{center}
\includegraphics[width=0.45\textwidth,clip]{SDSSJ0728_velocity_1d.eps}
\includegraphics[width=0.45\textwidth,clip]{SDSSJ0728_veldisp_1d.eps}
\end{center}
\caption{Line-of-sight velocity (left) and velocity dispersion
(right) profiles for the lens {SDSS\,J0728}. The black circles are for
the central slit, the blue squares are for the slit offset
0\farcs85 from the center, and the red diamonds are for the widest
slit offset 2\farcs1 from the galaxy center.}
\label{fig:velocitydata}
\end{figure*}
\section{Analysis and results}
\label{sec:analysis}
\subsection{Joint gravitational lensing and stellar dynamics analysis}
\label{ssec:code}
Here we briefly recall the general features of the {\textsc{cauldron}} code,
the tool employed to carry out the combined self-consistent
gravitational lensing and stellar dynamics analysis for the galaxy in
exam, {SDSS\,J0728}. The reader is referred to BK07 for a detailed
description of the algorithm.
We characterize the lens galaxy by means of its total density
distribution $\rho(\vec{\eta})$, where $\vec{\eta}$ is a set of parameters
describing the density profile. Via the Poisson equation, we calculate
the total gravitational potential $\Phi$ and we use it simultaneously
for both the gravitational lensing and the stellar dynamics modelling
of the data set, which typically includes the surface brightness and
velocity moments maps of the lens galaxy and the surface brightness
map of the lensed image. Both these modelling problems can be
formalized as a set of regularized linear equations, for which
standard solving techniques are readily available. Thus --- given a
combined data set --- for each choice of the parameters $\vec{\eta}$ we
can calculate the surface brightness distribution of the unlensed
source, and the weights of the elementary stellar dynamics building
blocks. In order to determine the ``best'' (in an Occam's razor sense)
density model given the data, this linear optimization scheme has been
embedded within the framework of Bayesian statistics. This allows to
objectively quantify (and therefore rank) the plausibility of each
model by means of the evidence merit function \citep[see
e.g.][]{MacKay1999, MacKay2003}. In this way, by maximizing the
evidence, one recovers the set of non-linear parameters $\vec{\eta}$
corresponding to the best density model, i.e.\ the model which
maximizes the joint posterior probability density function (PDF),
hence called maximum \emph{a posteriori} (MAP) model.
The method as described is extremely flexible and can in principle
support any density profile, by adopting for example a completely
general pixelized density distribution. However, its current practical
implementation, the {\textsc{cauldron}} algorithm, is more restricted in order
to make it computationally efficient and assumes axial symmetry
(i.e.\ a density distribution of the form $\rho(R,z)$) and a
two-integral stellar DF $f = f(E, L_{z})$, where $E$ and $L_{z}$ denote
the two classical integrals of motion, i.e., respectively, energy and
angular momentum along the rotation axis. As shown in BK07, under
these assumptions it is possible to take advantage of a fast Monte
Carlo numerical implementation of the two-integral Schwarzschild
method described by \citet{Cretton1999} and
\citet{Verolme-deZeeuw2002}, which allows a dynamical model to be
built in a matter of seconds and, therefore, makes it possible to
explore large regions of the parameter space. The distinguishing
feature of this method is that the building blocks employed for the
construction of the dynamical model are constituted by two-integral
components (TICs) rather than stellar orbits as in the classical
Schwarzschild method (\citealt{Schwarzschild1979}; see
e.g.\ \citealt{Thomas2007b} and \citealt{vandenBosch2008} for modern
implementations). A TIC can be visualized as an elementary toroidal
system, completely specified by a particular choice of $E$ and
$L_{z}$. TICs have simple $1/R$ radial density distributions and
analytic unprojected velocity moments, and by superposing them one can
build $f(E, L_{z})$ models for arbitrary spheroidal potentials
\citep[cf.][]{Cretton1999}: all these characteristics contribute to
make TICs particularly valuable and inexpensive building blocks when
compared to orbits.
Notwithstanding these restrictions, \citet{Barnabe2008} have shown
that {\textsc{cauldron}} works robustly even when applied to simulated systems
which depart significantly from the method's assumptions (including
the assumption of axial symmetry), reliably recovering several
important global properties of such systems, in particular the slope
of the total mass density profile, which is determined within less
than 10~per cent of the correct value. When the system displays
rotation in the kinematical maps, as is the case for {SDSS\,J0728}, the
main dynamical quantities (such as the global anisotropy parameter
$\delta$, the angular momentum and the ordered to random motions
ratio, see Sect.~\ref{ssec:globdyn}) are recovered with an accuracy
of~10 to~25 per cent.
\subsection{The galaxy model}
\label{ssec:model}
Stellar dynamics \citep[e.g.][]{Gerhard2001}, strong and weak
gravitational lensing \citep[e.g.][]{Koopmans2009, Gavazzi2007} and
X-ray studies \citep[see e.g.][and references therein]{Humphrey-Buote2010}
all concur in indicating that the total mass profile of elliptical
galaxies is remarkably well described by a single power-law model over
a large radial range.
As our model for the total mass density profile of the analyzed lens
galaxy we adopt, therefore, an axially symmetric power-law
distribution
\begin{equation}
\label{eq:rho}
\rho(m) = \frac{\rho_{0}}{m^{\gamma'}}
\end{equation}
with $\rho_{0}$ being a density scale, $0 < \gamma' < 3$ the
logarithmic slope of the density profile, and $m$ the elliptical
radius, i.e.\
\begin{equation}
\label{eq:m}
m^2 = \frac{R^2}{a_0^2} + \frac{z^2}{c_0^2}
= \frac{R^2}{a_0^2} + \frac{z^2}{a_0^2 q^2} ,
\end{equation}
where $c_0$ and $a_0$ are length-scales and the axial ratio $q\equiv
c_0/a_0$. The corresponding (inner) gravitational potential
$\Phi(R,z)$ associated with the distribution $\rho(m)$ can be obtained
straightforwardly by means of the classical \citet{Chandrasekhar1969}
formula, which entails the computation of a single integral.
In case the assumption of a power-law density profile is an
oversimplified description of the true mass distribution of the
galaxy, this is expected to have visibly disrupting effects on the
reconstructed observables, in particular for the lensing ones. As
illustrated in the \citet{Barnabe2008} tests, these can include a
strongly irregular reconstructed source, with sharp transitions in
intensity between close pixels (despite the usual adoption of a
regularization term), and the presence of recognizable arc-like
features in the lens image residuals. We emphasize, however, that such
clear discrepancies have never surfaced in previous analyses of the
SLACS systems (\citealt{Czoske2008}, B09), which supports instead the
effectiveness of the simple power-law model.
With this choice, the galaxy model is therefore characterized by three
non-linear physical parameters, i.e. the slope $\gamma'$, the axial
ratio $q$ and the dimensionless lens strength $\alpha_{0}$, which is
directly related to the normalization of the three dimensional
potential (see Appendix~B of BK07). To these, we must add the four
``geometrical'' parameters defining the configuration of the system in
the sky, i.e.\ the position angle $\vartheta_{\mathrm{PA}}$, the inclination $i$ and the
coordinates of the lens galaxy centre. The latter, as well as the
angle $\vartheta_{\mathrm{PA}}$, are typically strongly constrained by the lens image
brightness distribution and can be accurately determined by means of
fast preliminary explorations and, therefore, kept fixed afterwards in
order to decrease the number of free non-linear parameters during the
more computationally expensive optimization and error analysis
runs. If necessary, external shear can be also accounted for, by
introducing shear strength and shear angle as additional parameters.
Finally, we have three so-called hyperparameters which control the
level of the regularization in the reconstructed quantities: one for
the surface brightness distribution of the unlensed source, and two
for the TIC weights map. Their values are set by optimizing the
Bayesian posterior probability.
\subsection{Uncertainties}
\label{ssec:errors}
In order to correctly assess the model uncertainties within the
framework of Bayesian statistics, one needs to evaluate the posterior
probability distribution of the parameters, i.e., by denoting the data
set as $\vec{d}$ and the considered model or hypothesis as
$\mathcal{H}$,
\begin{equation}
\label{eq:posterior}
\mathcal{P}(\vec{\eta}) \equiv \mathrm{Pr}(\vec{\eta} \, | \, \vec{d},\mathcal{H})
\propto \mathcal{L}(\vec{\eta}) \times p(\vec{\eta}) \, ,
\end{equation}
where $\mathcal{L} (\vec{\eta}) \equiv \mathrm{Pr}(\vec{d} \, | \, \vec{\eta},
\mathcal{H})$ is the likelihood and $p (\vec{\eta}) \equiv \mathrm{Pr} (\vec{\eta}
\, | \, \mathcal{H})$ is the prior. The individual parameter
confidence intervals can be obtained by marginalizing the posterior
over each of them.
Sampling a multidimensional distribution such as $\mathcal{P}$ is in
general a challenging and computationally expensive task. An effective
technique to tackle this problem is the nested sampling Monte Carlo
method introduced by \citet{Skilling2004} which, in calculating the
evidence, produces posterior inferences as valuable by-products. For
our error analysis, we make use of the \textsc{MultiNest} algorithm
developed by \citet{Feroz-Hobson2008} and \citet*{Feroz2009}, which
provides an efficient and robust implementation of the nested sampling
method, and has been shown to yield reliable posterior inferences even
in presence of multi-modal and degenerate multivariate distributions.
The model parameters that we consider are the ones introduced in the
previous Section, i.e. the inclination~$i$, the lens strength~$\alpha_{0}$,
the slope~$\gamma'$ and the axial ratio~$q$ (the additional parameters
which can be estimated by means of preliminary runs, such as the lens
center and the position angle, however, are kept fixed here in order
to reduce the computational load), to which we must add the three
hyper-parameters. We formalize our ignorance by adopting flat priors
(or flat in logarithm for the hyper-parameters), constructed around
the MAP value of each parameter, and taken wide enough to include the
bulk of posterior probability distribution. Finally, we conduct a
thorough exploration of this 7-dimensional surface by launching
\textsc{MultiNest} with 400 live points\footnote{The live points
(sometimes also called active points) are the initial samples drawn
from the full prior distribution $p (\vec{\eta})$, from which the
nested sampling exploration of the posterior is started. Our choice
of~400 live points for the relatively well-behaved 7-dimensional
distribution at hand is very conservative. In fact, as shown by the
test problems examined in \citet{Feroz2009}, 1000 or 2000 active
points are sufficient for the application of \textsc{MultiNest} even
to, respectively, highly dimensional problems (with up to 30
parameters) or pathologically multimodal distributions (e.g. the
egg-box likelihood presented in their Section~6.1). Moreover, if one
is more interested in determining the marginalized posterior
distribution of the parameters rather than in accurately calculating
the value of the total evidence, a much smaller number of live
points (of order~50) is shown to be already enough to obtain a
reliable estimate.}, from which the individual marginalized
posterior probability distributions are obtained. In the following, we
quote the 99 per cent confidence interval calculated from these
distributions as our error.
\begin{table}
\centering
\begin{minipage}{0.80\hsize}
\caption{Recovered model parameters for lens galaxy {SDSS\,J0728}.}
\begin{center}
\begin{tabular}{ c @{\hspace{4em}} c c c c}
\hline
\noalign{\smallskip}
parameter & $\eta_{\textsc{map}}$ & $\eta_{-}$ & $\eta_{\mathrm{max}}$ & $\eta_{+}$ \\
\noalign{\smallskip}
\hline
$ \gamma' $ & 2.082 & 2.055 & 2.077 & 2.119 \\
\noalign{\smallskip}
$ \alpha_{0} $ & 0.325 & 0.317 & 0.323 & 0.331 \\
\noalign{\smallskip}
$ q $ & 0.602 & 0.574 & 0.603 & 0.681 \\
\noalign{\smallskip}
$ i $ & 68.1 & 67.8 & 68.6 & 73.6 \\
\hline
\end{tabular}
\end{center}
\label{tab:eta}
\textit{Note.} The listed parameters are: the logarithmic slope
$\gamma'$; the dimensionless lens strength $\alpha_{0}$; the axial ratio
$q$ and the inclination $i$ (in degrees) with respect to the
line-of-sight. The second column presents the MAP parameters,
i.e.\ the parameters that maximize the joint posterior
distribution. Columns~3 to~5 encapsulate a description of the
one-dimensional marginalized posterior distributions, plotted in
Fig.~\ref{fig:errors}. The parameter $\eta_{\mathrm{max}}$
corresponding to the maximum of that distribution is listed in
column~4, while $\eta_{-}$ and $\eta_{+}$ are, respectively, the
lower and upper limits of the 99 per cent confidence interval.
\end{minipage}
\end{table}
\begin{figure*}
\centering
\resizebox{0.97\hsize}{!}{\includegraphics[angle=-90]
{J0728_LENcomp.ps}}
\caption{Best model lens image reconstruction for the system
{SDSS\,J0728}. \emph{Top row:} reconstructed source model; $1 \sigma$
uncertainty on the source pixels; significance of the
reconstructed source. \emph{Bottom row:} \textit{HST}/ACS data
showing the lens image after subtraction of the lens galaxy; lens
image reconstruction; residuals. In the panels, North is up and
East is to the left.}
\label{fig:J0728_LEN}
\end{figure*}
\begin{figure*}
\resizebox{0.97\hsize}{!}{\includegraphics[angle=-90]
{J0728_DYNcomp.ps}}
\caption{Best dynamical model for the galaxy {SDSS\,J0728}. \emph{Top
row:} observed surface brightness distribution, projected
line-of-sight velocity and line-of-sight velocity
dispersion. \emph{Middle row:} corresponding reconstructed
quantities for the best model. \emph{Bottom row:} residuals. In
the panels, North is up and East is to the left.}
\label{fig:J0728_DYN}
\end{figure*}
\subsection{Results for the best model reconstruction}
\label{ssec:bestmodel}
The {\textsc{cauldron}} code has been applied to the combined data set
described in Section~\ref{sec:observations}. In order to avoid
possible spurious effects in the lensing reconstruction, we have
masked out the central regions of the lensed image map, where the
residuals of the galaxy subtraction are still appreciable. We use the
best-fitting B-spline model of the lens galaxy as our data set for its
surface brightness distribution, to avoid ``contamination'' from the
background galaxy lensed images, which is particularly bright and
could bias the reconstructed model. A similar approach had been
followed in the analysis of lens system {SDSS\,J2321} \citep{Czoske2008}.
The recovered non-linear parameters for the best reconstructed model,
i.e. the maximum \emph{a posteriori} model, are presented in
Table~\ref{tab:eta}. The uncertainties on the individual parameters
are quantified by marginalizing, over each of them, the joint
posterior distribution (see Fig.~\ref{fig:errors}). The parameter
values corresponding to the maximum of the one-dimensional
marginalized posterior and the limits of the 99 per cent confidence
interval are also listed in Table~\ref{tab:eta}.
We find for {SDSS\,J0728} a logarithmic slope $\gamma' =
2.077^{+0.042}_{-0.022}$ (errors indicate the 99\% confidence level),
very close to the so-called isothermal (i.e.\ $\rho \sim 1/r^{2}$)
profile which appears to be a characterizing feature of early-type
galaxies, and in general agreement with previous combined lensing and
dynamics studies of the SLACS sample (\citealt{Koopmans2009},
B09). For this specific system --- by using the 3-arcsec aperture
averaged SDSS velocity dispersion measure as the only kinematic
constraint --- \citet{Koopmans2009} determine a slightly different
slope $\gamma' = 1.85 \pm 0.10$ (68\% CL), which, however, we find
here to be too shallow to correctly reproduce the kinematic maps. One
reason for this discrepancy is that the velocity dispersion derived
from the SDSS is lower (typically of $\sim 20$ km s$^{-1}$) than the value
obtained from Keck spectroscopy, as discussed in Sect.~\ref{ssec:IFU}.
Moreover, \citet{Koopmans2009} adopt a simpler dynamical model, based
on solving spherical Jeans equations, which might have difficulties in
describing a system characterized by a significant flattening like
{SDSS\,J0728}.
The recovered lens strength is $\alpha_{0} = 0.323^{+0.008}_{-0.006}$. The
axial ratio of the total density distribution is found to be $q =
0.602^{+0.079}_{-0.028}$, slightly flatter than the intrinsic axial
ratio $q_{\star} = 0.688$ of the luminous distribution, calculated by
deprojecting the observed isophotal axial ratio $q_{\mathrm{\star, 2D}}$ by making use
of the best model value $i = 68.6^{+5.0}_{-0.8}$ obtained for the
inclination. Moreover, the position angle $\vartheta_{\mathrm{PA}} = 67\fdg5$ is found to
be extremely close to the value inferred from the light distribution,
indicating alignment between the dark and luminous mass components in
the inner regions of the galaxy.
The marginalized posterior probability distributions of these
parameters (with the exception of the position angle, which, as
previously described, is kept fixed after the preliminary run), as
well as those of the regularization hyper-parameters, obtained as
described in Section~\ref{ssec:errors}, are shown in
Fig.~\ref{fig:errors}. We note that these represent the statistical
errors on the considered power-law model, and do not take into account
the potential systematic uncertainties due to issues in the generation
of the data sets (e.g.\ the procedure for galaxy subtraction, as
pointed out by \citealt{Marshall2007}) or to incorrect model
assumptions (as in the case of, e.g., non axially symmetric density
distribution or flattening that varies with radius). In the latter
case, a quantitative estimate of the upper limits of the systematic
errors can be obtained by looking at the findings of the
\citet{Barnabe2008} `crash-test' where the {\textsc{cauldron}} code is applied
to a non-symmetric simulated galaxy. As mentioned in
Sect.~\ref{ssec:code}, the error on the logarithmic slope $\gamma'$ is
less than 10 per cent even in this quite extreme case. Since real
early-type galaxies are unlikely to depart from axisymmetry as
drastically as this simulated system, we expect systematic
uncertainties to remain within a few per cent level.
\begin{figure*}
\centering
\resizebox{0.95\hsize}{!}{\includegraphics[angle=-90]
{J0728_NSerr.ps}}
\caption{Marginalized posterior probability distributions of the
power-law model parameters (inclination, lens strength,
logarithmic slope, and axial ratio) and hyper-parameters, obtained
from the nested sampling exploration of the posterior. The
uncertainties on the parameters quoted in the text are calculated
by considering the interval around the peak which contains 99 per
cent of the probability.}
\label{fig:errors}
\end{figure*}
There is no evidence that any significant external shear effect needs
to be included in order to model this system: when the external shear
angle and strength are allowed to vary, the MAP model value of the
latter is found to be negligibly small. This is consistent with the
absence of massive structures near the lens galaxies found in
\citet{Treu2009}. We also explored the possibility of introducing a
core radius in the density profile, without finding any improvement to
justify the inclusion of this additional parameter.
In Figure~\ref{fig:J0728_LEN} and~\ref{fig:J0728_DYN} we present ---
respectively for lensing and dynamics --- the {SDSS\,J0728} data set, the
reconstructed observables obtained for the MAP model, and the
corresponding residuals. This system displays an unusually structured
lensed image, which we find to be produced by the presence of multiple
components in the source plane. The reconstructed background source is
shown in the top-left panel of Fig.~\ref{fig:J0728_LEN}; the
top-middle and top-right panels show, respectively, the standard
errors and the significance of the reconstructed source.
The surface brightness map is reconstructed very accurately. Since the
adopted data set is a B-spline model and therefore noiseless, the
small residuals (at the 1 per cent level) are dominated by the
discreteness effects of the TIC superposition, which determines the
concentric ripples, while the structure aligned with the galaxy minor
axis is caused by the toroidal shape of the building blocks. These
undesired effects have been kept under control by increasing both the
number of TIC employed in the dynamical model (of almost a factor of
four, from $N_{E} = 10 \times N_{L_{z}} = 10$ elements adopted in the
previous studies conducted with {\textsc{cauldron}} to $20 \times 9$ elements)
and the number of particles populating each TIC. An additional and
more compelling motivation for this improvement of the TIC library,
which further justifies the increased computational burden, is that it
proves to be important for the reconstruction of the kinematic
observables (particularly the velocity dispersion map) at larger
distances from the galaxy major axis, i.e. that part of our data set
which best allows us to probe the system under study beyond the
effective radius.
\subsection{Global dynamical quantities}
\label{ssec:globdyn}
The recovered weighted stellar DF for the best combined model is
presented in Fig.~\ref{fig:DF} as a map of the relative TIC weights
over the integral space $(E, L_{z})$ (the grid in the radial
coordinate $R_{\rm c}$ is related to a grid in energy as explained in
BK07). This representation encodes, in a very compact way, much of the
information on the dynamical structure of the galaxy that can be
obtained under the assumptions of the adopted two-integral
model. However, it is often useful to distill such information into
quantities that allow for a more straightforward physical
interpretation.
The global properties of the stellar velocity dispersion tensor are
customarily encapsulated in the three anisotropy parameters
\begin{equation}
\label{eq:AP}
\beta \equiv 1 - \frac{\Pi_{zz}}{\Pi_{RR}}, \quad
\gamma \equiv 1 - \frac{\Pi_{\varphi\varphi}}{\Pi_{RR}}
\quad \textrm{and} \quad
\delta \equiv 1 - \frac{2 \Pi_{zz}}{\Pi_{RR} + \Pi_{\varphi\varphi}},
\end{equation}
where we indicate with
\begin{equation}
\label{eq:AP:PI}
\Pi_{kk} = \int \rho \sigma^{2}_{k}\, \mathrm{d}^{3} \vec{x}
\end{equation}
the total unordered kinetic energy in the coordinate direction $k$ and
$\sigma_{k}(\vec{x})$ is the velocity dispersion along the direction
$k$ \citep[see][]{Cappellari2007,BT08}.
For {SDSS\,J0728}, we compute the integral of Eq.~(\ref{eq:AP:PI}) within a
cylinder of radius and height equal to $R_{\mathrm{e}}$, i.e. inside a region
which is very well constrained by the data, finding a mild anisotropy
$\delta = 0.08\pm0.02$, which falls within the typical
range of values for early-type galaxies both in the local Universe
\citep[][]{Cappellari2007,Thomas2009} and up to redshift $z \sim 0.35$
(see B09). Since we make use of a two-integral DF dynamical model, the
velocity dispersion tensor is isotropic in the meridional plane
(i.e. $\sigma^{2}_{R} (\vec{x}) = \sigma^{2}_{z} (\vec{x})$ for each
$\vec{x}$) and therefore $\beta \equiv 0$ by construction, and
$\gamma$ is univocally linked to $\delta$ by the relation $\gamma = 2
\delta / (\delta - 1)$. For the analyzed system, we have $\gamma =
-0.18\pm0.04$.
The importance of rotation with respect to random motions is among the
most defining aspects of the dynamical structure of a stellar
system. In order to explore how this property varies with the position
in the meridional plane, we calculate the local ratio
$\langle v_{\varphi} \rangle/\bar{\sigma}$ between the mean rotation velocity around the
$z$-axis and the mean velocity dispersion $\bar{\sigma}^{2} \equiv
(\sigma_{R}^{2} + \sigma_{\varphi}^{2} + \sigma_{z}^{2})/3$ and we
plot it in Fig.~\ref{fig:VoS} up to one effective radius. The inner
regions --- within approximately $1\arcsec$ --- are dominated by
random motions, while rotation becomes more important at large radii,
a trend somewhat reminiscent of what is found, in B09, for the fast
rotator {SDSS\,J0959} (although it should be remembered that the
$\langle v_{\varphi} \rangle/\bar{\sigma}$ map of the latter only extends up to
$R_{\mathrm{e}}/2$). Not surprisingly, therefore, the presence of large-scale
ordered motions is reflected also in the quite high value $j_{z} =
0.28^{+0.05}_{-0.01}$ obtained for the intrinsic rotation parameter,
which is a dimensionless proxy for the galaxy angular
momentum (refer to B09 for the definition).
\subsection{Dark and luminous mass distribution}
\label{ssec:mass}
The spherically averaged profile of the galaxy total mass
corresponding to the best reconstructed model (solid black curve in
Fig.~\ref{fig:massprof}) can be calculated straightforwardly from
Eq.~(\ref{eq:rho}), while the radial profile of the luminous component
is obtained from the recovered stellar DF. However, since within our
method stars are treated as tracers of the total potential, the
normalization of the luminous profile is not fixed, and must be
constrained by means of an independent determination of the stellar
mass-to-light ratio or by introducing additional assumptions.
One particularly informative assumption, known as the ``maximum
bulge'' approach, consists in maximally rescaling the luminous mass
distribution without exceeding the total mass profile at any
radius. This provides a consistent and robust way to assess a lower
limit for the dark matter fraction in the analyzed system (under the
hypothesis that the stellar mass-to-light ratio does not change too
drastically with the position in the galaxy). By applying this
procedure to {SDSS\,J0728}, we determine a value $\log(M_{\star}/M_{\sun})
= 11.50$ for the total stellar mass. With this normalization of the
luminous mass profile, one finds a dark matter fraction of 16 per cent
of the total mass within the (three-dimensional) spherical radius $r =
R_{\mathrm{e}}/2$, which rises to 28 per cent at $r = R_{\mathrm{e}}$ --- corresponding
to about $6$ kpc --- and up to almost 40 per cent at $r \sim 10$ kpc,
which is approximately the outer limit of the region over which we
have direct information from the stellar kinematic maps. This result
is consistent with the findings of purely dynamical studies of the
inner regions of early-type galaxies in the local Universe
\citep[e.g.][]{Gerhard2001,Cappellari2006,Thomas2007b}. It is also in
good agreement with the conclusions of our previous analysis of six
SLACS lens systems (where the same maximum bulge prescription was
adopted), for which, however, the kinematic data set does not extend
beyond $R_{\mathrm{e}}$, with the exception of galaxy {SDSS\,J0959} (see B09). The
corresponding upper limit for the stellar mass-to-light ratio in the
$B$ band of {SDSS\,J0728} is $M_{\star}/L_{B} = 3.14$, which is in the
lower end of the typical values reported for slow-rotating elliptical
galaxies in the local Universe \citep{Kronawitter2000, Gerhard2001,
Trujillo2004}. This is not too surprising, since the
$M_{\star}/L_{B}$ is expected to increase by a significant amount (a
factor of $1.4$ according to \citeauthor{Treu2002}
\citeyear{Treu2002}) between $z = 0.2$ and $z = 0$, simply due to
passive evolution of stellar populations. Moreover, the evidence of
rotation at large radii (cf.\ \S~\ref{ssec:globdyn}) might indicate
the presence of a disk component, typically characterized by a lower
mass-to-light ratio.
\begin{figure}
\centering
\resizebox{0.85\hsize}{!}{\includegraphics[angle=-90]{J0728_DFrec.ps}}
\caption{Reconstruction of the weighted two-integral DF of the
system {SDSS\,J0728} obtained from the MAP model. The value of each
pixel in the two-integral space represents the relative
contribution of the corresponding TIC to the stellar component of
the modelled system.}
\label{fig:DF}
\end{figure}
\begin{figure}
\centering
\resizebox{0.85\hsize}{!}{\includegraphics[angle=-90]{J0728_VoS.ps}}
\caption{Map of the local $\langle v_{\varphi} \rangle/\bar{\sigma}$ ratio between the
mean rotation velocity around the $z$-axis and the mean velocity
dispersion, plotted up to $R_{\mathrm{e}}$ in the positive quadrant of
the meridional plane}.
\label{fig:VoS}
\end{figure}
It is interesting to compare this maximum bulge upper limit with the
stellar masses determined from stellar population analysis. By
applying a novel Bayesian stellar population analysis code to
multi-band imaging data, \citet{Auger2009} determine for galaxy
{SDSS\,J0728} --- without including any priors from lensing --- a value
$\log(M_{\star}/M_{\sun}) = 11.44 \pm 0.12$ for a \citet{Chabrier2003}
IMF and a value $\log(M_{\star}/M_{\sun}) = 11.69 \pm 0.12$ for a
\citet{Salpeter1955} IMF (quoted errors are 1-sigma). We note that,
since the thermally pulsing asymptotic giant branch stars do not
dominate the luminosity of old stellar populations, the
\citet{Bruzual-Charlot2003} models used by \citet{Auger2009} should
not be biased by ignoring them. In fact, the stellar masses for SLACS
galaxies are found to be consistent with the masses determined from
\citet{Maraston2005} models \citep[see][]{Treu2010}.
By using these stellar mass values to rescale the (spherically
averaged) luminous profile, one obtains the red curves shown in
Fig.~\ref{fig:massprof}. It can be clearly seen that the Salpeter IMF
produces a stellar mass distribution which (unphysically) overshoots
the total mass profile up to and beyond the effective radius. The
Chabrier IMF, on the other hand, yields a physically acceptable
luminous profile, which, interestingly, is also very close to the one
determined through the maximum bulge assumption.
\citet{Treu2010} argue -- based on the mass determinations from
lensing, dynamics and stellar populations synthesis models for 56
SLACS systems -- that early-type galaxies cannot have both universal
IMF and dark matter profiles. In fact, if a universal Navarro, Frenk
\& White (NFW, \citealt*{Navarro1996}, \citeyear{Navarro1997}) halo is
assumed, the IMF shows a trend with velocity dispersion: a
Chabrier-like normalization is more appropriate for less massive
systems with $\sigma$ of the order of $200$ km s$^{-1}$, while more
massive galaxies are better described with a Salpeter-like IMF. The
results of our detailed analysis of {SDSS\,J0728} are consistent with this
general picture. With a SDSS velocity dispersion $\sigma = 214 \pm 11$
km s$^{-1}$, this system definitely belongs to the lower mass end of
the SLACS sample, and its halo profile is consistent with NFW (as
discussed in the next Section). Clearly, since this conclusion is
based on only one system, a full analysis of the sample still needs to
be done to further strengthen this tentative trend.
\begin{figure*}
\begin{center}
\subfigure{\label{fig:mass-chab}\includegraphics[angle=-90,width=0.48\textwidth]{Chab_profM.ps}}\hfill
\subfigure{\label{fig:mass-salp}\includegraphics[angle=-90,width=0.48\textwidth]{Salp_profM.ps}}
\caption{Spherically averaged mass distribution for lens galaxy
{SDSS\,J0728}. The solid black line shows the total mass profile
obtained from the best reconstructed model, with the statistical
uncertainty represented as a grey band. The solid red line shows
the luminous mass profile obtained from the recovered stellar DF
and rescaled using the stellar masses obtained from the
\citet{Auger2009} stellar population analysis with Chabrier and
Salpeter IMFs (left and right panel, respectively); the upper and
lower error bars quoted in that paper set the limits for the
red-shaded regions. The vertical lines provide an indication of
the region probed by the data set, showing the three-dimensional
radius $r$ which equals the Einstein radius (dashed line), the
effective radius (dotted line) and the outermost boundary of the
kinematic maps (dash-dotted line).}
\end{center}
\label{fig:massprof}
\end{figure*}
\subsection{Is the isothermal profile consistent with a NFW halo?}
The combined analysis of {SDSS\,J0728} has provided us with the total mass
density profile of this galaxy within its inner regions. As discussed
in the previous Section, and visualized in Fig.~\ref{fig:massprof},
the luminous component alone cannot account for this mass profile over
the entire radial range, even if its contribution is maximized (unless
the stellar mass-to-light ratio changes with radius in a very
fine-tuned manner). One, therefore, needs to invoke an additional mass
component characterized by the specific profile that complements the
luminous distribution. Interestingly, we find that the mass
distribution of this dark component can be consistent with a NFW halo
profile.
To show this, we attempt to describe the (spherically-averaged) total
mass distribution of the galaxy as the sum of the luminous component,
calculated as before from the recovered stellar DF, and a dark
component modelled as a standard NFW halo, i.e.
\begin{equation}
\label{eq:rho_NFW}
\rho_{\mathrm{NFW}}(r) = \frac{\rho_{\mathrm{crit}} \delta_{c}}
{(r/r_{\mathrm{b}})\,(1 + r/r_{\mathrm{b}})^{2}} \, ,
\end{equation}
where $\rho_{\mathrm{crit}}$ is the critical density, the
characteristic overdensity of the halo $\delta_{c}$ is a dimensionless
parameter connected to the halo concentration $c$ by the relation
\begin{equation}
\label{eq:deltac}
\delta_{c} = \frac{200}{3} \frac{c^{3}}
{\left[ \ln (1+c) - c/(1+c) \right]}
\end{equation}
and $r_{\mathrm{b}}$ denotes the break radius. We adopt $r_{\mathrm{b}} = 25$ kpc, based on
the \citet{Gavazzi2007} weak lensing analysis of the SLACS sample and
the consideration that {SDSS\,J0728} is slightly less massive than the
average SLACS system (for which $\langle \sigma_{\mathrm{SDSS}}
\rangle \simeq 250$ km s$^{-1}$).
We then determine the (non-negative) normalization coefficients for
the dark and luminous distribution which allow to best reproduce (in a
least-square sense) the total mass profile of the galaxy. Remarkably
--- despite the fact that the only two free parameters here are the
rescaling factors --- the superposition of these two simple components
proves to be enough to reproduce the total profile with great accuracy
over the whole radial range covered by the observations, as shown in
Fig.~\ref{fig:NFW}. Moreover, the combined profile remains consistent
with the one predicted by a nearly isothermal density distribution
even beyond $10$ kpc and up to the break radius.
The luminous profile obtained in this way almost coincides with the
one determined by means of the maximum bulge approach, with
$\log(M_{\star}/M_{\sun}) = 11.50 $, while the normalization for the
dark halo profile translates into a concentration parameter $c \simeq
11$, from which one infers a virial radius $r_{200} = c r_{\mathrm{b}} \simeq
280$ kpc and a halo mass $M_{200} \simeq 3.1 \times 10^{12}
M_{\sun}$. This is a mildly high value for the concentration when
compared with the range $c \sim 3 - 10$ obtained from numerical
simulations of relaxed dark matter haloes of corresponding mass
\citep*[see in particular][]{Maccio2008}. However, concentrations
higher than the theoretical predictions are found in dynamical studies
of slow-rotating early-type galaxies \citep[see e.g.][and references
therein]{Romanowsky2010}. We note that, in our case, lower
concentrations are obtained by setting a larger value for the break
radius, e.g. $c \sim 10$ for $r_{\mathrm{b}} = 30$ kpc, while the fit to the
total mass profile becomes only slightly worse, with discrepancy of a
few percent.
\section{Conclusions}
\label{sec:conclusions}
We have carried out a detailed study of the mass profile and dynamical
structure of the inner regions of the early-type lens galaxy {SDSS\,J0728},
located at a redshift $z = 0.21$, using a composite data set
consisting of \emph{HST}/ACS high-resolution imaging and
two-dimensional kinematic maps constructed from long-slit
spectroscopic observations obtained with the Keck instrument LRIS (the
slit --- aligned with the galaxy major axis --- has been positioned at
three different heights along the minor axis, allowing to mimic
integral-field spectroscopy). We have modelled the system by making
use of the {\textsc{cauldron}} code for combined gravitational lensing and
stellar dynamics analysis, which operates under the assumptions of
axial symmetry and two-integral stellar DF. With respect to sample
studies of the SLACS lens galaxies such as \citet{Koopmans2009}, the
approach used here presents a number of improvements: it employs a
self-consistent framework where the same total potential is used for
both lensing and dynamics; it allows one to construct genuine
axisymmetric dynamical models (albeit restricted to two integrals of
motions); it extracts much more information from the data set, making
use --- in addition to the lensed image --- of the lens galaxy surface
brightness and velocity moments maps rather than being limited to a
single measure of SDSS velocity dispersion. This permits a much
greater level of detail to be recovered when modelling the system.
{SDSS\,J0728} is the first galaxy for which it has been possible to conduct
this kind of in-depth combined analysis by taking advantage of a
kinematic data set that extends well past the effective radius: the
outermost pixels of the velocity moments maps probe a region up to a
distance of about $3$ arcsec from the center, corresponding to $\sim
1.7$ $R_{\mathrm{e}}$. For comparison, in the sample of six SLACS lens galaxies
examined in B09, the outermost boundary $R_{\mathrm{kin}}$ of the kinematic maps
(obtained with VLT VIMOS integral-field spectroscopy) is in the range
$0.30 - 0.85$ $R_{\mathrm{e}}$, with the single exception of system {SDSS\,J0959}
for which $R_{\mathrm{kin}}$ exceeds the effective radius of about $15$ per cent.
\begin{figure}
\centering
\resizebox{1.00\hsize}{!}{\includegraphics[angle=-90]
{J0728_combprof.ps}}
\caption{Best-fit decomposition of the total mass profile of
{SDSS\,J0728} (black curve) in the sum of a NFW dark matter halo with
break radius $r_{\mathrm{b}} = 25$ kpc (blue curve) and the luminous profile
obtained from the recovered stellar DF (green curve). The dotted
red line which very closely matches the black curve over the
whole radial interval is the profile obtained by adding up the
luminous and dark components. The vertical dash-dotted line
indicates the outermost boundary of the kinematic maps.}
\label{fig:NFW}
\end{figure}
The main results of the analysis are summarized and discussed below:
\begin{enumerate}
\item The total mass density profile of the galaxy inner regions, up
to a radius of $\sim$ 1.7 $R_{\mathrm{e}}$, is found to be satisfactorily
described by a simple axisymmetric distribution $\rho \propto
1/m^{\gamma'}$, with a logarithmic slope $\gamma' =
2.08^{+0.04}_{-0.02}$ (99~\% CL). This is very close to the
isothermal profile (i.e.\ $\rho \propto 1/m^{2}$), which seems to be
an ubiquitous characteristic of early-type galaxies at least up to
redshift $z \sim 1$, as revealed by a number of dynamics, X-ray and
combined lensing and dynamics studies (see
e.g. \citealt{Kronawitter2000}, \citealt{Gerhard2001},
\citealt{Humphrey-Buote2010}, \citealt{Koopmans2009} and references
therein). Moreover, weak lensing studies show that the total
density profile remains consistent with being isothermal also well
beyond the galaxy inner regions, up to radial distances of about 100
effective radii \citep{Gavazzi2007}.
The physical mechanisms which cause the total density distribution
to be shaped into this particular structure, with little but non
negligible intrinsic scattering ($\lesssim 10$ per cent in $\gamma'$,
see B09 and \citealt{Koopmans2009}), are not well understood. The
apparent ``conspiracy'' between the luminous and dark components to
generate a nearly isothermal combined profile appears even more
surprising when considering that numerical simulations including gas
dynamics find dark matter density profiles which depend both on the
amount of baryons and on the details of the assembly processes
\citep[see e.g.][]{Tissera2009,Abadi2009}.
Interestingly, despite the predicted complications, for {SDSS\,J0728} we
find that it is possible to reproduce very accurately the
(spherically averaged) total mass distribution by combining two very
intuitive and simple building blocks: (1)~the luminous mass profile
obtained from the stellar DF, almost maximally rescaled and (2)~a
NFW dark matter halo (with a break radius $r_{\mathrm{b}} = 25$ kpc and a
concentration parameter $c \sim 11$).
\item We find the total density distribution to be quite flattened
within the probed region, with an axial ratio $q =
0.60^{+0.08}_{-0.03}$, which is flatter than the axial ratio
$q_{\star}$ of the luminous distribution (obtained by using the best
model recovered inclination to deproject the two-dimensional
isophotal axial ratio). This characteristic differentiates {SDSS\,J0728}
from the six lens galaxies studied in B09, for which $q/q_{\star}
\gtrsim 1$.
\item The system is characterized by a very mild anisotropy $\delta =
0.08\pm0.02$. On examining the dynamical structure of the galaxy by
means of the $\langle v_{\varphi} \rangle/\bar{\sigma}$ map, one notices that the
contribution of ordered motions becomes more important outside the
inner regions, which determines the moderately high value $j_{z} =
0.28$ for the dimensionless angular momentum parameter. This result
is obtained by integrating within a cylindrical region of radius and
height equal to $R_{\mathrm{e}}$. If the integration is limited to $R_{\mathrm{e}}/2$,
in order to allow a direct comparison with results of B09, one finds
a lower $j_{z} = 0.18$, fully consistent with the typical values
obtained for the galaxies in that sample (with the exception of the
clearly fast-rotating {SDSS\,J0959}).
\item Under the assumptions of maximum bulge and position-independent
stellar mass-to-light ratio, we determine for the dark matter
fraction a lower limit of 28 per cent within the spherical radius $r
= R_{\mathrm{e}}$. Within $r \sim 10$ kpc, i.e.\ the approximate extension of
the area directly probed by the kinematic data, the contribution of
the dark matter to the total cumulative mass is about 40 per cent,
almost matching the luminous component. This is in agreement with
the findings of dynamical studies of nearby ellipticals
\citep{Gerhard2001,Cappellari2006,Thomas2007} as well as with the
combined lensing and dynamics analysis of six SLACS galaxies at $z =
0.08 - 0.33$ (\citealt{Czoske2008}, B09). Interestingly, numerical
simulations of early-type galaxy formation from cosmological initial
conditions also predict a dark matter fraction of about $20 - 40$
per cent within this radius \citep{Naab2007}.
In alternative to the previous approach, we have also rescaled the
luminous profile by using the stellar masses calculated from the
\citet{Auger2009} stellar population analysis. The obtained luminous
mass distribution is too high in the case of a Salpeter IMF,
exceeding the total mass in places, while for a Chabrier IMF it
remains lower than $M_{\mathrm{tot}}(r)$ and, moreover, close to the
profile predicted under the maximum bulge hypothesis. This suggests,
in agreement with the conclusions of \citet{Treu2010}, that a
Chabrier functional form might be more suited to describe the IMF
for less massive early-type galaxies such as {SDSS\,J0728}.
In order to further test if this description is correct, we plan to
extend this study to a wider sample of SLACS lens galaxies, covering
a broad range of velocity dispersions $\sigma \approx 200 - 350$ km
s$^{-1}$, for which two-dimensional kinematic data sets are
available.
\end{enumerate}
In conclusion, despite being located at a redshift greater than $0.2$,
the system {SDSS\,J0728} shows structural characteristics --- namely nearly
isothermal total density profile, dark matter fraction, anisotropy
parameter $\delta$, local ratio of ordered to random motions ---
broadly consistent with what is observed in the nearby Universe for
early-type galaxies of comparable luminosity and velocity dispersion
(e.g.\ \citealt{Gerhard2001}, \citealt{Thomas2007},
\citealt{Cappellari2007}). The upper limit for the B-band stellar
mass-to-light ratio, obtained from the maximum bulge assumption, is
also in line with the values determined for local ellipticals
\citep[e.g.][]{Kronawitter2000}, once the ageing of the stellar
populations is taken into account. This study, therefore, provides an
indication that the density profile as well as the global dynamical
structure of the inner regions of massive ellipticals did not undergo
any dramatic change or significant evolution across the last two
billion years.
\section*{Acknowledgments}
M.B. is grateful to Phil Marshall and Simona Vegetti for useful
discussion and to Farhan Feroz for his help with
\textsc{MultiNest}. M.B. acknowledges support from the Department of
Energy contract DE-AC02-76SF00515. T.T. acknowledges support from the
NSF through CAREER award NSF-0642621, and from the Packard Foundation
through a Packard Fellowship. L.K. is supported through an NWO-VIDI
program subsidy (project number 639.042.505). Some of the data
presented herein were obtained at the W.M. Keck Observatory, which is
operated as a scientific partnership among the California Institute of
Technology, the University of California and the National Aeronautics
and Space Administration. The Observatory was made possible by the
generous financial support of the W.M.~Keck Foundation. The authors
wish to recognize and acknowledge the very significant cultural role
and reverence that the summit of Mauna Kea has always had within the
indigenous Hawaiian community. We are most fortunate to have the
opportunity to conduct observations from this mountain. This paper is
also based on observations made with the NASA/ESA Hubble Space
Telescope, obtained from the data archive at the Space Telescope
Institute. STScI is operated by the association of Universities for
Research in Astronomy, Inc. under the NASA contract NAS 5-26555. This
work was supported by NASA through HST grants GO-10886 and 11202.
|
\section{Introduction}
\indent
The Riccati equation
\begin{equation}\label{ricceq}
\frac{dy}{dt}=b_0(t)+b_1(t)y+b_2(t)y^2,
\end{equation}
is the simplest non-linear differential equation \cite{CRL07,CarRamGra} and it appears
in many different fields of Mathematics and Physics \cite{CL09,CLR10,CMN,Ch09,MH05,RM08,SHC07,DS08,PW}. It is essentially the only first-order ordinary differential
equation on the real line admitting a
non-linear superposition principle \cite{LS,PW} and in spite of its apparent simplicity,
its general solution cannot be described by means
of quadratures except in some very particular cases \cite{AS64,CarRam,Ib08,Ib09II,Kamke,Ko06,Zh98,Zh99,Mu60,Na99,Na00,Pr80,Ra61,Ra62,RU68,RDM05,Stre}.
In this paper we review the geometric approach to Riccati equations
according to the results of the works \cite{CRL07,CarRam} with the aim of proving that
integrability conditions of Riccati equations can be understood in a very
general way from the point of view of the theory of Lie systems
\cite{CR02,Gu93,Ib96,Ib99,Ib89,LS, Ve93,PW}. Furthermore, we
recover various known results as particular cases of our approach and the method here derived can be applied to any other Lie system, e.g. \cite{CL09,CLR10}.
Each Lie system is associated with a Lie algebra of vector fields, the so-called Vessiot--Guldberg Lie algebra \cite{CLR10,GGL08,Gu93,Ib96,Ve93}. This Lie algebra can be used to classify those Lie systems that can be integrated by quadratures \cite{CarRamGra}. For instance, it is a known fact that Lie systems related to solvables Vessiot--Guldberg Lie algebras, e.g. affine homogeneous systems or linear homogeneous systems, can be integrated by quadratures \cite{CarRamGra,Ib09II,Ib09}. Nevertheless, the general solution of Lie systems related to non-solvables Lie algebras, e.g. Riccati equations, cannot be completely determined and it frequently relies on the knowledge of certain special functions \cite{Ib09II}, the solution of other equations \cite{Na99,Na00,DS08}, etc.
The method developed here allows us to determine integrable cases of Lie systems related to non-solvables Vessiot--Gulberg Lie algebras. Such a procedure is detailed for Riccati equations, which are associated with a non-solvable Vessiot--Guldberg Lie algebra isomorphic to $\mathfrak{sl}(2,\mathbb{R})$, but it can also be applied to other Lie systems related to Vessiot--Guldberg Lie algebras isomorphic to this one, for example, the Lie systems connected to Milne--Pinney equations \cite{CL09Milne}, Ermakov systems \cite{CLR08e}, harmonic oscillators \cite{CLR10}, etc.
We finally analyse the linearisation of Riccati equation by means of our new approach and recover a characterisation previously proved by Ibragimov \cite{Ib08}. Furthermore, we detail a, as far as we know, new result about the properties of linearisation of Riccati equations.
The paper is organised as follows. For the sake of completeness,
we report some known facts on the integrability of Riccati equations in Sec. 2 and we review the geometric interpretation of
the general Riccati equation as a $t$-dependent vector field on
the one-point compactification of the real line in Sec. 3. As a consequence of the latter, Riccati equations can be studied through equations on $SL(2,\mathbb{R})$.
Sec. 4 is devoted to reporting some known results on the action of the group of curves in $SL(2,\mathbb{R})$
on the set of Riccati equations and how this action can be seen in terms of
transformations
of the corresponding equations on $SL(2,\mathbb{R})$, see \cite{CarRam05b}. In Sec. 5 we build up a Lie system describing the transformation process of Riccati
equations through the action of curves of $SL(2,\mathbb{R})$ and then, in Sec. 6, we analyse
the general characteristics of our approach into integrability conditions and
how the transformation processes described by the previous Lie system can be used to give a unified approach to the results of \cite{CRL07,CarRamGra}. In Sec. 7 we develop
a particular case of the procedures of Sec. 6 in order to recover some results found in the literature \cite{AS64,Ko06,Ra61,Ra62,RU68,RDM05}. Sec. 8 is devoted to analyzing the theory of integrability through reduction from our new viewpoint. Finally, in
Sec. 9 we describe how our Lie system for studying
integrability conditions enables us to explain when certain linear fractional transformations allow us to linearise Riccati equations. As a particular instance we obtain a result given in \cite{Ib08,RDM05}.
\section{Integrability of Riccati equations}\label{IntRicEqu}
\indent
In order to provide a first insight into the study of integrability conditions for Riccati equations, and for any Lie system in general, we review in this Section some known results about the integrability of Riccati equations.
As a first particular example, Riccati equations (\ref{ricceq}) are integrable by quadratures when
$b_2=0$. Indeed, in such a case these equations reduce to an inhomogeneous linear equation
and two quadratures allow us to find the
general solution.
Additionally, under the change of variable
$w=-1/y$ the Riccati equation
(\ref{ricceq}) reads
$$
\frac{dw}{dt}=b_0(t)\, w^2-b_1(t)\,w+b_2(t)
$$
and if we suppose $b_0=0$ in Eq. (\ref{ricceq}),
then the mentioned change of variable transforms the given equation into an integrable
linear one.
Another very well-known property on integrability of Riccati equations is
that given a particular solution $y_1(t)$ of Eq. (\ref{ricceq}),
then the change of variable $y=y_1(t)+z$ leads to a new Riccati equation for which the coefficient of the
term independent of $z$
is zero, i.e.
\begin{equation}
\frac {dz}{dt}=[2\, b_2(t)\, y_1(t)+ b_1(t)] z+ b_2(t)\,z^2, \label{Bereq}
\end{equation}
and, as we pointed out before, it can be reduced to an inhomogeneous linear equation with the change
$z=-1/u$. Therefore, given one particular solution, the general
solution can be found by means of two quadratures.
If not only one but two particular solutions,
$y_1(t)$ and $y_2(t)$, of Eq. (\ref{ricceq}) are known, the general solution can be found by means of only one quadrature. In fact, the change of variable
$z=(y-y_1(t))/(y-y_2(t))$ transforms the original equation into a homogeneous first-order linear differential equation in the new variable $z$
and therefore the general solution can immediately be found.
Finally, giving three particular solutions, $y_1(1),y_2(t),y_3(t)$, the general solution can be
written, without making use of any quadrature, in the following way
$$
y(t)=\frac{y_1(t)(y_3(t)-y_2(t))-ky_2(t)(y_1(t)-y_3(t))}{(y_3(t)-y_2(t))-k(y_1(t)-y_3(t))}.
$$
This is a non-linear superposition rule studied in
\cite{CMN} from a group theoretical perspective.
The simplest case of Eq. (\ref{ricceq}), when it is an autonomous equation ($b_0$, $b_1$ and $b_2$ constants), has been fully studied (see
e.g. \cite{CarRamdos} and references therein) and it is integrable by
quadratures. This result can be considered as a consequence of the existence of a constant
(maybe complex) solution enabling us to reduce the Riccati equation into an
inhomogeneous linear one. Moreover, the separable Riccati equations of the form
\begin{equation*}
\frac{dy}{dt}=\varphi(t)(c_0+c_1\,y+c_2\,y^2),
\end{equation*}
with $\varphi(t)$ a non-vanishing function on a certain open interval $I\subset \mathbb{R}$ and $c_0$, $c_1$, $c_2$ real numbers, are integrable because a new time function $\tau=\tau(t)$
such that $d\tau/dt=\varphi(t)$ reduces the above equation into an autonomous one. Furthermore, the above Riccati equations are also integrable as they accept, in similarity to the autonomous case, a constant (maybe complex) solution.
\section{Geometric approach to Riccati equations}
\noindent
Let us report in this Section some known results about the geometrical approach to the Riccati equation \cite{CRL07}. Such a point of view is used in next Sections to investigate integrability conditions for these equations and, in general, for any Lie system.
From the geometric viewpoint, the Riccati equation
(\ref{ricceq})
can be considered
as a differential equation determining the integral curves for the
$t$-dependent vector field \cite{Car96}
\begin{equation}
X(t,y)=\left[b_0(t)+b_1(t)y+b_2(t)y^2\right]\frac{\partial}{\partial y}\ .\label{vfRic}
\end{equation}
This $t$-dependent vector field is a linear combination with $t$-dependent coefficients $b_0(t)$, $b_1(t)$ and $b_2(t)$
of the three vector fields
\begin{equation}
L_0 =\frac{\partial}{\partial y}\,, \quad
L_1 =y\,\frac{\partial}{\partial y}\, , \quad
L_2 = y^2\,\frac{\partial}{\partial y}\,, \label{sl2gen}
\end{equation}
with defining relations
\begin{equation}\label{conmutL}
[L_0,L_1] = L_0\,, \quad
[L_0,L_2] = 2L_1\,, \quad
[L_1,L_2] = L_2 \,,
\end{equation}
and therefore spanning a three-dimensional Lie algebra of vector fields $V$. Consequently, Riccati equations are Lie systems \cite{LS} and the Lie algebra $V$, the so-called Vessiot--Guldberg Lie algebra \cite{Gu93,Ve93}, is isomorphic to $\mathfrak{sl}(2,\mathbb{R})$ being here considered as made up by traceless $2\times 2$ matrices. A particular basis for $\mathfrak{sl}(2,\mathbb{R})$ is given by
\begin{equation}
M_0=\left(\begin{matrix}
0&-1\\0&0
\end{matrix}\right)\,,
M_1=\frac{1}{2}\left(\begin{matrix}
-1&0\\0&1
\end{matrix}\right)\,,
M_2=\left(\begin{matrix}
0&0
\\1&0
\end{matrix}\right)\ .
\label{base_matrices}
\end{equation}
Moreover, it can be checked that the linear map $\rho:\mathfrak{sl}(2,\mathbb{R})\rightarrow V$ obeying $\rho(M_j)=L_j$, with $j=0,1,2$, is
a Lie algebra isomorphism.
Note that $L_2$ is not a complete vector field on $\mathbb{R}$. However we can do the one-point compactification $\overline{\mathbb{R}}=\mathbb{R}\cup \{\infty\}$ of $\mathbb{R}$ and then $L_0$, $L_1$ and $L_2$ are complete vector fields on $\overline{\mathbb{R}}$. Consequently, these vector fields are fundamental vector fields
corresponding to the action $\Phi:(A,y)\in SL(2,{\mathbb{R}})\times \overline{\mathbb{R}}\mapsto \Phi(A,y)\in\overline{\mathbb{R}}$ given by
\begin{equation}\label{Action}
\Phi(A,y)=\left\{
\begin{aligned}
&{\frac{\alpha\, y+\beta}{\gamma\, y+\delta}}\quad &y&\neq-{\frac{\delta}{\gamma}},\,\,\,y\neq\infty,\\
&\frac{\alpha}{\gamma} \quad &y&=\infty,\\
&\infty \quad&y&=-\frac{\delta}{\gamma}, \\
\end{aligned}
\right.\quad {\rm with}\quad A=\left(\begin{array}{cc}\alpha& \beta\\\gamma&\delta\end{array}\right)\in SL(2,\mathbb{R}).
\end{equation}
Denote by $X^{\tt R}_j$ and $ X^{\tt L}_j$, $j=1,2,3$, the right- and left-invariant vector fields on $SL(2,{\mathbb{R}})$ such that $X^{\tt
R}_j(I)=X^{\tt
L}_j(I)=M_j$. Moreover, these vector fields satisfy that $X^{\tt R}_j(A)=M_j\cdot A$ and $X^{\tt L}_j(A)=A\cdot M_j$, with ``$\cdot$'' the usual matrix multiplication.
A remarkable property is that if $A(t)$ is
the integral curve for
the $t$-dependent
vector field
$$X(t)=-\sum_{j=0}^2 b_j(t)\, X^{\tt R}_j\,,
$$
starting from the
neutral element in $SL(2,\mathbb{R})$, i.e. $A(0)=I$, then $A(t)$ satisfies the equation
\begin{equation}\label{eLA}
\dot{A}(t)A^{-1}(t)=-\sum_{j=0}^2b_j(t)M_j\equiv {\rm a}(t),
\end{equation}
and the solution of Riccati equation (\ref{ricceq}) with initial
condition $y(0)=y_0$ is given by $y(t)=\Phi(A(t),y_0)$ \cite{CR02}.
Note that the r.h.s. in Eq. (\ref{eLA}) is a curve in
$T_ISL(2,\mathbb{R})$ that can be identified to a curve in the Lie algebra
$\mathfrak{sl}(2,\mathbb{R})$ of left-invariant vector fields on $SL(2,\mathbb{R})$
through the usual isomorphism: we relate each left-invariant vector field $X^{\tt L}$ to
the element $X^{\tt L}(I)\in T_ISL(2,\mathbb{R})$. From now on, we do not distinguish explicitly
elements in $T_ISL(2,\mathbb{R})$ and its corresponding ones in $\mathfrak{sl}(2,\mathbb{R})$.
In summary, the general solution of Riccati equations (\ref{ricceq}) can be
obtained through solutions of an equation like (\ref{eLA}) starting from $I$. Consequently, we have reduced the problem of finding the general solution of Riccati equations to determining the solution of Eq.
(\ref{eLA}) beginning at the neutral element of $SL(2,\mathbb{R})$. Note that, in a similar way, this procedure can be applied to any Lie system \cite{CL09}.
\section{Transformation laws of Riccati equations}\label{TL}
\noindent
In this Section we briefly describe an important property of Lie systems, in the
particular case of Riccati equations, which plays a very relevant r\^ole for
establishing, as indicated in \cite{CarRam}, integrability
criteria: {\it The group $\mathcal{G}$ of curves in a Lie group $G$ associated with
a Lie system, here $SL(2, {\mathbb{R}})$, acts on the set of these Lie systems, here
Riccati equations}.
More explicitly, fixed a basis of vector fields on $\overline{\mathbb{R}}$, for instance $\{L_j\,|\,j=0,1,2\}$, which spans a Vessiot--Guldberg Lie algebra of vector fields isomorphic to $\mathfrak{sl}(2,\mathbb{R})$, each Riccati equation (\ref{ricceq}) can be considered as a
curve $(b_0(t),b_1(t),b_2(t))$ in $\mathbb{R}^3$. The point now is that
each element of the group of smooth curves in $SL(2, \mathbb{R})$, i.e. $\bar A\in \mathcal{G}\equiv{\rm Map}(\mathbb{R},\,SL(2,\mathbb{R}))$,
transforms every curve $y(t)$ in $\overline{\mathbb{R}}$
into a new curve $y'(t)$ in $\overline{\mathbb{R}}$ given by $y'(t)=\Phi(\bar A(t),y(t))$. Moreover, the $t$-dependent change of variables $y'(t)=\Phi(\bar A(t),y(t))$ transforms the Riccati equation (\ref{ricceq})
into a new
Riccati equation with new $t$-dependent coefficients, $b'_0,b'_1, b'_2$ given by
\begin{equation}\label{trans}
\left\{\begin{aligned}
b'_2&={\delta}^2\,b_2-\delta\gamma\,b_1+{\gamma}^2\,b_0+\gamma {\dot{\delta}}-\delta \dot{\gamma}\ ,\\
b'_1&=-2\,\beta\delta\,b_2+(\alpha\delta+\beta\gamma)\,b_1-2\,\alpha\gamma\,b_0
+\delta \dot{\alpha}-\alpha \dot{\delta}+\beta \dot{\gamma}-\gamma \dot{\beta}\ , \\
b'_0&={\beta}^2\,b_2-\alpha\beta\,b_1+{\alpha}^2\,b_0+\alpha\dot{\beta}-\beta\dot{\alpha},
\end{aligned}\right.
\end{equation}
with
$$
\bar{A}(t)=\left(
\begin{matrix}
\alpha(t)&\beta(t)\\
\gamma(t)&\delta(t)
\end{matrix}\right).
$$
The above transformation defines an affine action (see e.g. \cite{LM87} for the general definition of
this concept) of the group
$\mathcal{G}$ on the set of
Riccati equations, see \cite{CarRam}.
The group $\mathcal{G}$ also acts on the set of equations of the form (\ref{eLA}) on $SL(2,\mathbb{R})$. In order to show this, note first that $\mathcal{G}$ acts on the left on the
set of curves in $SL(2,\mathbb{R})$ by left translations, i.e. given two curves $A(t)$ and $\bar A(t)$ in $SL(2,\mathbb{R})$, the curve $\bar A(t)$ transforms the curve $A(t)$ into a new one $A'(t)=\bar A(t) A(t)$. Moreover, if
$A(t)$ is a solution of Eq. (\ref{eLA}), then the new curve $A'(t)$ satisfies a new equation like (\ref{eLA}) but with a different right
hand side ${\rm a}'(t)$. Differentiating the relation $A'(t)=\bar A(t) A(t)$ in terms of time and taking into account the form of (\ref{eLA}), we get that the relation between the curves ${\rm a}(t)$ and ${\rm a}'(t)$ in $\mathfrak{sl}(2,\mathbb{R})$ is
\begin{equation}
{\rm a}'(t)=\bar A(t){\rm a}(t)\bar A^{-1}(t)+\dot{\bar{A}}(t)\bar A^{-1}(t)
=-\sum_{j=0}^2b'_j(t)M_j\, \label{newricc}
\end{equation}
and such a relation implies the expressions (\ref{trans}). Conversely, if $A'(t)=\bar A(t) A(t)$ is the solution for the equation corresponding to the
curve ${\rm a}'(t)$ given by the transformation rule (\ref{newricc}), then $A(t)$ is the solution of Eq. (\ref{eLA}).
To sum up, we have shown that it is possible to associate each Riccati equation with an equation on
the Lie group $SL(2,\mathbb{R})$ and to define an infinite-dimensional group of
transformations acting on the set of Riccati equations. Additionally, this
process
can be easily derived in a similar way for any Lie system. In such a case, we
must consider an equation on a Lie group $G$ associated with the corresponding Lie
system
and the group $\mathcal{G}$ of curves in $G$ acting on the set of curves in $G$
in the
form $A'(t)=L_{\bar A(t)}A(t)$ instead of $A'(t)=\bar A(t)A(t)$. This action
induces
other action of $\mathcal{G}$ on the set of equations of the form (\ref{eLA})
but on the Lie group $G$.
More explicitly, a curve $\bar A(t)\in \mathcal{G}$ transforms an equation on
$G$ of the form (\ref{eLA})
determined by a curve ${\rm a}(t)\subset T_IG$ into a new one determined by the
new curve
${\rm a}'(t)\subset T_IG$ given by
\begin{equation}\label{TransConecc}
{\rm a}'(t)={\rm Ad}_{\bar A(t)}{\rm a}(t)+R_{\bar A^{-1}(t)*\bar A(t)}\dot{\bar{A}}(t).
\end{equation}
\section{Lie structure of an equation of transformation of Lie systems}
\indent
Our aim in this Section is to construct a Lie system describing the curves in $SL(2,\mathbb{R})$
relating two Riccati equations associated with a pair of equations in
$SL(2,\mathbb{R})$ characterised by two
curves ${\rm a}(t), {\rm
a}'(t)\subset \mathfrak{sl}(2,{\mathbb{R}})$. By means of this Lie system we are
going to explain in next Sections the developments of \cite{CRL07, CarRamGra} and other works from a unified viewpoint.
Let us multiply Eq. (\ref{newricc}) on the right by $\bar A(t)$ to get
\begin{equation}\label{MatrixRicc}
\dot{\bar{A}}(t)={\rm a}'(t)\bar A(t)-\bar A(t){\rm a}(t)\,.
\end{equation}
If we consider Eq. (\ref{MatrixRicc}) as a first-order differential equation in the
coefficients of the
curve $\bar A(t)$ in $SL(2,\mathbb{R})$, with
$$
\bar A(t)=\left(
\begin{matrix}
\alpha(t) &\beta(t)\\
\gamma(t)& \delta(t)
\end{matrix}\right)\,,\quad \alpha(t)\delta(t)-\beta(t)\gamma(t)=1,
$$
then system (\ref{MatrixRicc}) reads
\begin{equation}\label{FS}
\left(\begin{matrix}
\dot\alpha\\
\dot\beta\\
\dot\gamma\\
\dot\delta
\end{matrix}\right)
=
\left(\begin{matrix}
\frac{b'_1-b_1}{2}&b_2 &b'_0&0\\
-b_0& \frac{b'_1+b_1}{2}&0 &b'_0\\
-b'_2&0 &-\frac{b'_1+b_1}{2}& b_2\\
0&-b_2' &-b_0& -\frac{b'_1-b_1}{2}
\end{matrix}\right)
\left(\begin{matrix}
\alpha\\
\beta\\
\gamma\\
\delta
\end{matrix}\right).
\end{equation}
In order to determine the solutions $x(t)=(\alpha(t),\beta(t),
\gamma(t),\delta(t))$ of the above system relating two
different Riccati equations, we should check that actually the matrices $\bar A(t)$,
whose elements are the corresponding components of $x(t)$, are related to
matrices in $SL(2,\mathbb{R})$, i.e.
we have to verify that at any time
$\alpha\delta-\beta\gamma=1$. Nevertheless, we can drop such a restriction because it can be automatically implemented by a restraint on the initial conditions for
the solutions and hence we can deal with the variables $\alpha,\beta,\gamma, \delta$ in the
system (\ref{FS}) as being independent. Consider now the vector fields
{\small
\begin{equation*}
\begin{array}{ll}
N_0=-\alpha\dfrac{\partial}{\partial\beta}-\gamma\dfrac{\partial}{\partial\delta}, &N'_0=\gamma\dfrac{\partial}{\partial\alpha}+\delta\dfrac{\partial}{\partial\beta},\cr
N_1=\frac
12\left(\beta\dfrac{\partial}{\partial\beta}+\delta\dfrac{\partial}{\partial\delta}-\alpha\dfrac{\partial}{\partial\alpha}-\gamma\dfrac{\partial}{\partial\gamma}\right),
&N'_1=\frac 12\left(\alpha\dfrac{\partial}{\partial\alpha}+\beta\dfrac{\partial}{\partial\beta}-\gamma\dfrac{\partial}{\partial\gamma}-\delta\dfrac{\partial}{\partial\delta}\right),\cr
N_2=\beta\dfrac{\partial}{\partial\alpha}+\delta\dfrac{\partial}{\partial\gamma},& N'_2=-\alpha\dfrac{\partial}{\partial\gamma}-\beta\dfrac{\partial}{\partial\delta},\nonumber
\end{array}
\end{equation*}}
satisfying the non-null commutation relations
\begin{eqnarray*}
&&\left[ N_0,N_1\right]=N_0, \qquad [N_0,N_2]=2 N_1, \qquad [N_1,N_2]=N_2,\cr
&&[N'_0,N'_1]=N'_0, \qquad [N'_0, N'_2]=2
N'_1,\qquad [N'_1, N'_2]=N'_2\,.\nonumber
\end{eqnarray*}
Note that as $[N_i,N'_j]=0$, for $i,j=0,1,2$, the linear system of differential equation (\ref{FS}) is a Lie system on $\mathbb{R}^4$ associated with a Lie algebra of vector fields isomorphic to $\mathfrak{g}\equiv\mathfrak{sl}(2,\mathbb{R})\oplus\mathfrak{sl}(2,\mathbb{R})$. This Lie
algebra decomposes into a direct sum of two Lie algebras of vector fields isomorphic to
$\mathfrak{sl}(2,\mathbb{R})$: the first one is spanned by $\{N_0,N_1,N_2\}$ and the second
one by $\{N'_0,N'_1,N'_2\}$.
If we denote $x\equiv\left(\alpha,\beta,\gamma,\delta\right)\in \mathbb{R}^4$, the
system (\ref{FS}) is a differential equations on $\mathbb{R}^4$
\begin{equation*}
\frac{dx}{dt}=N(t,x),
\end{equation*}
with $N$ being the $t$-dependent vector field
\begin{equation*}
N(t,x)=\sum_{j=0}^2\left(b_\alpha(t)N_\alpha(x)+b'_\alpha(t)N'_\alpha(x)\right).
\end{equation*}
The vector fields $\{N_0,N_1,N_2,N'_0,N'_1,N'_2\}$ span a regular involutive distribution $\mathcal{D}$ with rank
three in almost any point of $\mathbb{R}^4$ and thus there exists, at least locally, a
first-integral. We can check that the function
$$I:x\equiv (\alpha,\beta,\gamma,\delta)\in\mathbb{R}^4\longrightarrow I(x)\equiv\det x\equiv\alpha\delta-\beta\gamma \in \mathbb{R}$$
is a first-integral for the vector fields in the distribution $\mathcal{D}$. Moreover, such a first-integral is related to the
determinant of the matrix $\bar A$ with coefficients given by the components
of $x=(\alpha,\beta,\gamma,\delta)$. Therefore, if we have a solution of the system (\ref{FS})
with an initial condition $\det x(0)=\alpha(0)\delta(0)-\beta(0)\gamma(0)=1$,
then $ \det x(t)=1$ at any time $t$ and the solution can be understood
as a curve in $SL(2,\mathbb{R})$.
In summary, we have proved that:
\begin{theorem}\label{THLS} The curves in $SL(2,\mathbb{R})$ transforming equation (\ref{eLA}) into a new equation of the same form but characterised by a new curve ${\rm a}'(t)=-\sum_{j=0}^2b'_j(t)M_j\,$ are described through the solutions of the Lie system
\begin{equation}\label{Sys}
\frac{dx}{dt}=N(t,x)\equiv\sum_{j=0}^2\left(b_j(t)N_j(x)+b'_j(t)N'_j(x)\right)\,
\end{equation}
such that $\det x(0)=1$. Furthermore, the above Lie system is related to a non-solvable Vessiot--Guldberg Lie algebra isomorphic to $\mathfrak{sl}(2,\mathbb{R})\oplus\mathfrak{sl}(2,\mathbb{R})$.
\end{theorem}
\begin{corollary} \label{CorCur} Given two Riccati equations associated with
curves ${\rm a}'(t)$ and ${\rm a}(t)$ in
$\mathfrak{sl}(2,\mathbb{R})$ there always exists a curve $\bar A(t)$ in $SL(2,\mathbb{R})$
transforming the Riccati equation related to ${\rm a}(t)$ into the one associated with ${\rm a}'(t)$. If
furthermore $\bar A(0)=I$, this curve is uniquely defined.
\end{corollary}
\begin{proof}
Given a matrix $A(0)\in SL(2,\mathbb{R})$ and an element $x(0)$ related to it, according to the theorem of existence and uniqueness of solutions for differential equations, the system (\ref{Sys}), with the chosen ${\rm a}'(t)$ and ${\rm a}(t)$, admits a solution $x(t)$ with initial condition $x(0)$. As ${\rm det}(x(t))={\rm det}(x(0))=1$, such a solution, considered as a matrix $\bar
A(t)$, belongs to
$SL(2,\mathbb{R})$ and therefore there exists a solution $x(t)$ for the system (\ref{Sys}) with initial condition $x(0)$ related to $\bar A(0)$. This proves the first statement of our corollary.
If the curve $\bar A(t)$ connecting two curves in $\mathfrak{sl}(2,\mathbb{R})$
satisfies $\bar A(0)=I$, it is the curve $x(t)$ in $\mathbb{R}^4$
being the solution
of system (\ref{Sys}) with initial condition $x(0)=(1,0,0,1)$, which is
uniquely determined because of the theorem of existence and uniqueness of
solutions of systems of first-order differential equations.
\end{proof}
Even if we know that given two equations on the Lie group $SL(2,\mathbb{R})$ there
always exists a transformation relating both, in order to obtain such a curve
we need to solve the Lie system (\ref{Sys}). Unfortunately, such a Lie system is associated with a non-solvable Lie algebra and it is
not easy in general to find its solutions, i.e. it is not integrable by
quadratures and therefore such a curve cannot be easily found
in the general case.
Nevertheless, we will explain many known properties and obtain new integrability conditions for Riccati equations by means of Theorem \ref{THLS}. Furthermore, the procedure to obtain the Lie system (\ref{Sys}) can be generalised to deal with any Lie system related to a Lie group $G$ with Lie algebra $\mathfrak{g}$. In this general case, relation (\ref{TransConecc}) implies that
\begin{equation*}
\dot{\bar{A}}(t)=R_{\bar{A}(t)*I}{\rm a}'(t)-L_{\bar{A}(t)*I}{\rm a}(t).
\end{equation*}
As $X^{\tt R}(t,\bar A)=R_{\bar{A}*I}{\rm a}'(t)$ is a $t$-dependent
right-invariant vector field
on $G$ and $X^{\tt L}(t,\bar A)=-L_{\bar{A}*I}{\rm a}(t)$ a left-invariant one,
the above system
is the equation determining the integral curves of a time-dependent vector
field with values in
the linear space spanned by right- and left- invariant vector fields on
$G$. Note that the family
of left-invariant (right-invariant) vector fields on $G$ spans a Lie algebra
isomorphic to
$\mathfrak{g}$ and, as right- and left-invariant vector fields commute among
them, the set
of vector fields spanned by both families is a Lie algebra of vector fields
isomorphic to
$\mathfrak{g}\oplus\mathfrak{g}$. In this way, we get that the above system,
relating two Lie
systems associated with curves ${\rm a}(t)$ and ${\rm a}'(t)$ in
$\mathfrak{g}$,
is a Lie system related to a Vessiot--Guldberg Lie algebra isomorphic to $\mathfrak{g}\oplus\mathfrak{g}$.
\section{Lie systems and integrability conditions}
\indent
In this section some integrability conditions are analysed
from the perspective of the theory
of Lie systems with $SL(2,\mathbb{R})$ as associated Lie group, with the aim of giving a unified approach to the reduction and
transformations procedures described in \cite{CRL07, CarRamGra}. More explicitly, these
methods are related to conditions for the existence of a curve in a previously
chosen family of curves in $SL(2,\mathbb{R})$ connecting a curve ${\rm a}(t)\subset\mathfrak{sl}(2,\mathbb{R})$ with a curve ${\rm a}'(t)$
in a
solvable Lie subalgebra of $\mathfrak{sl}(2,\mathbb{R})$. It is also shown that
this viewpoint
enables us to explain many of the previous results scattered
in the literature about this topic and to prove other new properties.
As it was shown in Sec. \ref{TL}, if the curve $\bar A(t)\subset SL(2,\mathbb{R})$ transforms the equation on this Lie group defined by the curve ${\rm a}(t)$ into
another one characterised by ${\rm a}'(t)$ and $A'(t)$
is a solution for the equation similar to (\ref{eLA}) for the primed system, i.e. characterised by ${\rm a}'(t)$, then
$A(t)=\bar A^{-1}(t)A'(t)$ is a solution for the equation in $SL(2,\mathbb{R})$ characterised by ${\rm a}(t)$. Moreover, if ${\rm a}'(t)$ lies in a solvable Lie subalgebra of $\mathfrak{sl}(2,\mathbb{R})$,
we can obtain $A'(t)$ in many ways, e.g. by quadratures or by other methods
as those used in \cite{CarRamGra}. Then, once $A'(t)$ is obtained, the knowledge of the curve $\bar A(t)$ transforming the
curve ${\rm a}(t)$ into ${\rm a}'(t)$ provides the curve $A(t)$.
Therefore if we begin with a curve
${\rm a}'(t)$ in a solvable Lie subalgebra of $\mathfrak{sl}(2,\mathbb{R})$ and consider the solutions for the system (\ref{Sys}) in a subset of $SL(2,\mathbb{R})$, we can
relate the curve ${\rm a}'(t)$, and therefore its Riccati equation, to other possible curves ${\rm a}(t)$, finding in this way
a family of Riccati equations that can be exactly solved. Note that, if we do not consider solutions of the system (\ref{Sys}) in a subset of $SL(2,\mathbb{R})$, it is generally difficult to check whether a particular Riccati equation belongs to the family of integrable Riccati equations so obtained.
Suppose we impose some restrictions on the family of curves solutions of the system (\ref{Sys}), for instance $\beta=\gamma=0$. Consequently, the system may not have solutions compatible with such restrictions, i.e. it may be impossible to connect the curves ${\rm a}(t)$ and ${\rm a}'(t)$ by a curve in $SL(2,\mathbb{R})$ satisfying the assumed restrictions. This gives rise to some
compatibility conditions for the existence of these special solutions, some of them algebraic and
other differential
ones, between the $t$-dependent coefficients of ${\rm a}'(t)$ and ${\rm
a}(t)$. It will be shown later on that
such restrictions correspond to integrability conditions previously proposed in the literature.
Therefore, there are two ingredients to take into account:
\begin{enumerate}
\item {\it The equations on the Lie group characterised by curves ${\rm a}'(t)$ for which
we can obtain an explicit solution}.
We always suppose that ${\rm a}'(t)$ is related to a solvable Lie subalgebra of
$\mathfrak{sl}(2,\mathbb{R})$ and we leave open other possible restrictions for further
study.
\item {\it The conditions imposed on the solutions of system} (\ref{FS}). We follow two principal approaches in next Sections
where the
solutions of this system are related to curves in certain one-parameter or two-parameter subsets of $SL(2,\mathbb{R})$.
\end{enumerate}
Consider the next example of our theory: suppose we try to connect any ${\rm a}(t)$ with a final curve of the form ${\rm a}'(t)=-D(t)(c_0{\rm a}_0+c_1{\rm a}_1+c_2{\rm a}_2)$,
where $c_0,c_1$ and $c_2$ are real numbers. In this way, the
system (\ref{FS}) describing the curve $\bar A(t)\subset SL(2,\mathbb{R})$ connecting these curves is
\begin{equation}\label{Lie2}
\frac{dx}{dt}=\sum_{j=0}^2\left(b_j(t)N_j(x)+D(t)
c_j N'_j(x)\right)=N(t,x).
\end{equation}
Now, as the vector field
\begin{equation*}
N'=\sum_{j=0}^2c_j N'_j,
\end{equation*}
is such that
\begin{equation*}
\left[N_j,N'\right]=0,\quad\quad j=0,1,2,
\end{equation*}
the Lie system (\ref{Lie2}) is related to a non-solvable Lie algebra of vector fields
isomorphic to $\mathfrak{sl}(2,\mathbb{R})\oplus \mathbb{R}$. Hence, it is not integrable by
quadratures and the solution cannot be easily found in the general case. Nevertheless, note that system (\ref{Lie2}) always has a solution.
In this way, we can consider some particular cases of Lie system (\ref{Lie2}) for which the
resulting system of differential equations can be easily integrated. As a first
instance, take $x$ related to a one-parameter family of
elements of $SL(2,\mathbb{R})$. Such a restriction implies that system (\ref{Lie2}) has not
always a solution because sometimes it is not possible to connect ${\rm
a}(t)$ and ${\rm a}'(t)$ by means of the chosen
family of curves. This fact induces differential and/or algebraic restrictions on the
initial $t$-dependent functions $b_j$, with $j=0,1,2$, that
describe some known integrability conditions and may be some new ones developing the ideas of \cite{CRL07}. From this viewpoint we can obtain new
integrability conditions that can be used, for instance, to obtain exact solutions.
Otherwise, if we choose a two-parameter set for the restriction, we find in
some cases that we need a particular solution of the initial Riccati equation
to obtain the reduction of the given Riccati equation into an integrable one. This is the point of view shown in \cite{CarRamGra} where integrability conditions were related to reduction methods.
\section{Description of known integrability conditions}\label{DIC}
\indent
Let us first remark that Lie systems on $G$ of the form (\ref{eLA}) and determined by a constant curve,
${\rm a}=-\sum_{j=0}^2 c_j M_j$,
are integrable and consequently the same happens for
curves of the form ${\rm a}(t)=-D(t)\left(\sum_{j=0}^2 c_j M_j\right)$, where $D$ is any non-vanishing function, because a time-reparametrisation reduces the problem to
the previous one.
Our aim in this Section is to determine
the curves $\bar A(t)$ in $SL(2,\mathbb{R})$ relating two
equations on $SL(2,\mathbb{R})$ characterised by the curves ${\rm a}(t)$ and ${\rm a}'(t)=-D(t)(c_0M_0+c_1M_1+c_2M_2)$ with $D(t)$ a non-vanishing function and $c_0$, $c_1$ and $c_2$ real constants such that $c_0c_2\neq 0$. As the final
equation is integrable, the transformation
establishing the relation to such a final integrable equation allows us to find
by quadratures the solution of the initial equation and, therefore, the solution for its associated Riccati equation. In order to get such a
transformation, we look for curves $\bar A(t)$ in $SL(2,\mathbb{R})$ satisfying certain conditions in order to get an integrable equation (\ref{Lie2}).
Nevertheless, under the assumed restrictions, we may obtain a
system of differential equations admitting no solution. As an
application,
we show that many known results can be
recovered and explained in this way.
We have already showed that the Riccati equations (\ref{ricceq}) with either $b_0\equiv 0$ or $b_2\equiv 0$ are reducible to linear
differential equations and therefore they are always
integrable. Hence, they are not interesting in our study and we focus our attention on reducing a Riccati equation (\ref{ricceq}), with $b_0b_2\ne 0$ in an open
interval in $t$, into an integrable one by means of the action of a curve in $SL(2,\mathbb{R})$. With this aim,
we consider the family of curves in $SL(2,\mathbb{R})$ with $\beta=0$ and $\gamma=0$, i.e. we take curves of the form
$$\bar A(t)=\left(\begin{matrix}\alpha(t)&0\\0&\delta(t)\end{matrix}\right)\in SL(2,\mathbb{R})\,,\quad\alpha(t)\delta(t)=1.$$
We already pointed out that a curve $\bar A(t)$ in $SL(2,\mathbb{R})$ induces a $t$-dependent change of variables in $\bar{\mathbb{R}}$ given by $y'(t)=\Phi(\bar A(t),y(t))$. In view of (\ref{Action}) and as $\alpha\delta=1$, we get that, in our case, such a change of variables is given by
\begin{equation}\label{yprime}
y'=\alpha^2(t)y=G(t)y\,,\quad G(t)\equiv \frac{\alpha(t)}{
\delta(t)}>0.
\end{equation}
In view of the relations (\ref{trans}), the initial Riccati equation is transformed by means of the curve $\bar A(t)$ into the new Riccati equation with $t$-dependent coefficients
$$b'_2=\delta^2\,b_2\,,\qquad b'_1=\alpha\,\delta\,b_1+\dot \alpha\,\delta-\alpha\,\dot
\delta\,,\qquad b'_0=\alpha^2\, b_0.$$
Furthermore, the functions $\alpha$ and $\delta $ are solutions of
system (\ref{FS}), which in this case reads
\begin{equation}\label{RLFS}
\left(\begin{matrix}
\dot\alpha\\
0\\
0\\
\dot\delta
\end{matrix}\right)
=\left(\begin{matrix}
\frac{b'_1-b_1}{2}&b_2 &b'_0&0\\
-b_0& \frac{b'_1+b_1}{2}&0 &b'_0\\
-b'_2&0 &-\frac{b'_1+b_1}{2}& b_2\\
0&-b_2' &-b_0& -\frac{b'_1-b_1}{2}
\end{matrix}\right)\left(
\begin{matrix}
\alpha\\
0\\
0\\
\delta
\end{matrix}\right).
\end{equation}
The existence of particular solutions for the above system related to elements of $SL(2,\mathbb{R})$ and satisfying the required conditions determines integrability conditions for Riccati equations by the described method. Thus, let us analyse the existence of such solutions to get these integrability conditions.
From some of the relations of the system (\ref{RLFS}), we get that
$$-b_0\,\alpha+b'_0\, \delta=0\,,\qquad -b_2'\,\alpha+b_2\,\delta=0.$$
As $\alpha(t)\delta(t)= 1$, the above relations imply that $b_0b_2= b'_0b'_2$ and
\begin{equation*}
\alpha^2=\frac{b_0'}{b_0}=\frac{b_2}{b_2'}\equiv G>0\,.
\end{equation*}
Hence, the transformation formulas (\ref{trans}) reduce to
\begin{equation}
b'_2=\alpha^{-2}\,b_2\,,\qquad b'_1=b_1+2\frac{\dot \alpha}\alpha
\,,\qquad b'_0=\alpha^2 b_0\,.\label{transfb}
\end{equation}
Then, in order to exist a $t$-dependent function $D$ and two real constants
$c_0$ and $c_2$, with $c_0c_2\neq 0$, such that
$b'_2=Dc_2$ and $b'_0=Dc_0$, the function $D$ must be given by
\begin{equation*}
D^2c_0c_2=b_0b_2\Longrightarrow D=\pm\sqrt{\frac{b_0b_2}{c_0c_2}}\,,
\end{equation*}
where we have used that $b'_0b'_2=b_0b_2$. On the other hand, as $b'_0/b_0=\alpha^2>0$, we have to fix the sign $\kappa$ of the function $D$ in order to satisfy this relation, i.e. ${\rm sg}(c_0D)={\rm sg}(b_0)$. Therefore,
$$
\kappa={\rm sg}(D)={\rm sg}(b_0/c_0).
$$
Also, as $b_0b_2=b'_0b'_2$, we get that ${\rm sg}(b_0b_2)={\rm sg}(c_0c_2D^2)={\rm sg}(c_0c_2)$. Furthermore, in view of the relations (\ref{transfb}), $\alpha$ is determined, up to a sign, by
\begin{equation}\label{otroalfa}
\alpha=\sqrt{\frac{Dc_0}{b_0}}=\left(\frac{c_0}{c_2}\,\frac{b_2}{b_0} \right)^{1/4}\,,
\end{equation}
and therefore the change of variables (\ref{yprime}) reads:
\begin{equation}\label{Chang}
y'=\frac{D(t)c_0}{b_0(t)}y\,.
\end{equation}
Finally, as a consequence of (\ref{transfb}), in order for $b'_1$ to be the product $b'_1=c_1\, D$, we see that
\begin{equation}\label{eq10}
b_1+2\, \frac{\dot \alpha}{\alpha}=\kappa c_1 \sqrt{\frac{b_0b_2}{c_0c_2}}\,.
\end{equation}
Using (\ref{otroalfa}) we get
$$4\,\frac{\dot
\alpha}{\alpha}=\frac 1{\alpha^4}\, \frac {d\alpha^4}{dt}=\frac{b_0}{b_2}\,\frac
d{dt}\left(\frac{b_2}{b_0}\right)=\frac{b_0}{b_2}\,\, \frac{\dot b_2b_0-\dot
b_0 b_2}{b_0^2}=\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0},
$$
and replacing $2\dot\alpha/\alpha$ in (\ref{eq10}) for the value obtained above, we see that the
required integral condition is
\begin{equation*}
\sqrt{\frac{c_0c_2}{b_0b_2}}\left[b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0}\right)\right]=\kappa c_1\,.
\end{equation*}
Conversely, it can be verified that if the above integrability condition holds and $D^2c_0c_2=b_0b_2$, then the change of variables (\ref{Chang}) transforms the Riccati equation (\ref{ricceq}) into $dy'/dt=D(t)(c_0+c_1y'+c_2y'^2)$, with $c_0c_2\neq 0$.
In summary:
\begin{theorem}\label{TU} The necessary and sufficient condition
for the existence of a transformation
\begin{equation*}
y'=G(t)y,\quad G(t)>0,
\end{equation*}
relating the Riccati equation
\begin{equation*}
\frac{dy}{dt}=b_0(t)+b_1(t)y+b_2(t)y^2\,, \qquad b_0b_2\ne 0,
\end{equation*}
to an integrable one given by
\begin{equation}
\frac{dy'}{dt}=D(t)(c_0+c_1y'+c_2y'^2)\,,\quad c_0c_2\neq 0\label{eqDcs}
\end{equation}
where $c_0, c_1, c_2$ are real numbers and $D(t)$ is a non-vanishing function, are
\begin{equation}
D^2c_0c_2=b_0b_2,\qquad \left(b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0}\right)\right)\sqrt{\frac{c_0c_2}{b_0b_2}}=\kappa c_1,\label{DinTh2}
\end{equation}
where $\kappa={\rm sg}(D)=sg(b_0/c_0)$. The transformation is then uniquely defined by
\begin{equation*}
y'=\sqrt{\frac{b_2(t)c_0}{b_0(t)c_2}}\,y\,.
\end{equation*}
\end{theorem}
As a consequence of Theorem \ref{TU}, given a Riccati equation
\begin{equation*}
\frac{dy}{dt}=b_0(t)+b_1(t)y+b_2(t)y^2\,, \qquad b_0(t)b_2(t)\ne 0,
\end{equation*}
if there are real constants $c_0,c_1$ and $c_2$, with $c_0c_2\neq 0$, such that
\begin{equation*}
\sqrt{\frac{c_0c_2}{b_0b_2}}\left(b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0}\right)\right)=\kappa c_1,
\end{equation*}
there exists a $t$-dependent linear change of variables transforming the given equation into
an integrable Riccati equation of the form
\begin{equation}
\frac{dy'}{dt}=D(t)(c_0+c_1y'+c_2y'^2), \qquad c_0c_2\neq 0,\label{ricc1dim}
\end{equation}
and the function $D$ is given by (\ref{DinTh2}) with the sign determined by $\kappa$.
From the previous results, it can be derived the following corollary.
\begin{corollary}\label{CTU}
A Riccati equation (\ref{ricceq}) with $b_0b_2\ne 0$ can be transformed into a Riccati equation of the form (\ref{ricc1dim}) by a $t$-dependent change of variables $y'=G(t)y$, with $G(t)>0$, if and only if
\begin{equation}
\frac{1}{\sqrt{|b_0b_2|}}\left(b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot
b_0}{b_0}\right)\right)=K,
\label{resCor2}
\end{equation}
for a certain real constant $K$. In such a case, the Riccati equation (\ref{ricceq}) is integrable by quadratures.
\end{corollary}
According to Theorem \ref{TU}, if we start with the integrable Riccati
Eq. (\ref{ricc1dim}), we can
obtain the set of all Riccati equations that can be reached from it by means of a
transformation of the form (\ref{yprime}).
\begin{corollary}\label{C2TU} Given an integrable Riccati equation
\begin{equation*}
\frac{dy}{dt}=D(t)(c_0+c_1y+c_2y^2),\qquad c_0c_2\neq 0,
\end{equation*}
with $D(t)$ a non-vanishing function, the set of Riccati equations which can be obtained with a
transformation $y'=G(t)y$, with $G(t)>0$, are
those of the form:
\begin{equation*}
\frac{dy'}{dt}=b_0(t)+\left( \frac{\dot b_0(t)}{b_0(t)}-\frac{\dot D(t)}{D(t)}+c_1D(t)\right) y'+\frac{D^2(t)c_0c_2}{b_0(t)}y'^2\,,
\end{equation*}
with
$$G=\frac{Dc_0}{\sqrt{b_0}}\,.$$
\end{corollary}
Therefore starting with an integrable equation we can generate a family of
solvable Riccati equations whose coefficients are parametrised by a non-vanishing function $b_0$. Moreover, the integrability condition
to check whether a Riccati equation belongs to this family can be easily verified.
These results can now be used for a better understanding of
some integrability conditions found in the literature.
\medskip
$\bullet$ {\it The case of Allen and Stein}:
\medskip
The results of the paper by Allan and Stein \cite{AS64} can be recovered through our general approach. In that work, a Riccati equation (\ref{ricceq}) with $b_0b_2>0$ and $b_0$, $b_2$ differentiable functions satisfying the condition
\begin{equation}\label{ALintegra}
\frac{b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0}\right)}{\sqrt{b_0b_2}}=C,
\end{equation}
where $C$ is a real constant, was transformed into the integrable one
\begin{equation}\label{FREAS64}
\frac{dy'}{dt}=\sqrt{b_0(t)b_2(t)}\left(1+Cy'+y'^2\right),
\end{equation}
through the $t$-dependent linear transformation
\begin{equation*}
y'=\sqrt{\frac{b_2(t)}{b_0(t)}}y\,.
\end{equation*}
If integrability condition (\ref{ALintegra}) is satisfied by a Riccati equation, such an equation also holds the assumptions of the Corollary \ref{CTU} and, therefore, the integrability condition given in Theorem \ref{TU} with
\begin{equation*}
c_0=1=c_2,\quad c_1=C,\quad D=\sqrt{b_0b_2}.
\end{equation*}
Consequently, the corresponding transformation given by Theorem \ref{TU} reads
\begin{equation*}
y'=\sqrt{\frac{b_2(t)}{b_0(t)}}y\,,
\end{equation*}
showing that the transformation in \cite{AS64} is a particular case of our results. This is not an unexpected result because Theorem \ref{TU} shows that if such a time-dependent change of variables is used to transform a Riccati equation (\ref{ricceq}) into one of the form (\ref{eqDcs}), this change of variables must be of the form (\ref{Chang}) and the initial Riccati equation must hold the integrability conditions (\ref{DinTh2}).
\medskip
$\bullet$ {\it The case of Rao and Ukidave}:
\medskip
Rao and Ukidave stated in their work \cite{RU68} that the Riccati equation (\ref{ricceq}), with $b_0b_2>0$, can
be transformed into
an integrable Riccati equation of the form
\begin{equation*}
\frac{dy'}{dt}=\sqrt{cb_0b_2}\left(1-ky'+\frac{1}{c}{y'}^2\right),
\end{equation*}
through a $t$-dependent linear transformation
\begin{equation*}
y'=\frac{1}{v(t)}y,
\end{equation*}
if there exist real constants $c$ and $k$ such that following integrability condition holds
\begin{equation}\label{CondRU1}
b_2=\frac{b_0}{cv^2},
\end{equation}
with $v$ being a solution of the differential
equation
\begin{equation}\label{CondRU2}
\frac{dv}{dt}=b_1(t)v+kb_0(t)\,.
\end{equation}
Note that, in view of (\ref{CondRU1}), necessarily $c>0$ and if the integrability conditions (\ref{CondRU1}) and (\ref{CondRU2}) hold with constants $c$ and $k$ and a negative solution $v(t)$, the same conditions hold for the constants $c$ and $-k$ and a positive solution $-v(t)$. Consequently, we can restrict ourselves to studying the integrability conditions (\ref{CondRU1}) and (\ref{CondRU2}) for positive solutions $v(t)>0$. In such a case, the previous method uses a $t$-dependent linear change of coordinates of the form (\ref{yprime}) and the final Riccati equation are of the type described in our work (\ref{eqDcs}), therefore
the integrability conditions derived by Rao and Ukidave must be a particular instance of the integrable cases provided by
Theorem \ref{TU}.
Using the value of $v(t)$ in terms of the constant $c$ and the functions $b_0$ and $b_2$ obtained from formula (\ref{CondRU1}) and Eq. (\ref{CondRU2}), we get that
\begin{equation*}
\frac{1}{\sqrt{|b_0b_2|}}\left(b_1+\frac 12\left(\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0}\right)\right)=-k\,{\rm sg}(b_0)\sqrt{c}.
\end{equation*}
Hence, the Riccati equations obeying conditions (\ref{CondRU1}) and (\ref{CondRU2}) satisfy the integrability conditions of Corollary \ref{CTU}. Moreover, if we choose
\begin{equation*}
D^2=cb_0b_2,\quad c_0=1,\quad c_1=-k,\quad c_2=c^{-1}\,,
\end{equation*}
then $D=\sqrt{c b_0b_2}$ and the only possible transformation (\ref{yprime}) given by Theorem \ref{TU} reads
\begin{equation*}
y'=\alpha^2(t)y=\sqrt{\frac{cb_2(t)}{b_0(t)}}y,
\end{equation*}
and then
\begin{equation*}
\frac{1}{v}=\sqrt{\frac{cb_2}{b_0}}.
\end{equation*}
In this way, we recover one of the results derived by Rao and Ukidave in \cite{RU68}.
\medskip
$\bullet$ {\it The case of Kovalevskaya}:
\medskip
Kovalevskaya showed in the paper \cite{Ko06} that the Riccati equation
\begin{equation*}
\frac{dy}{dt}=F(t)+\left(L+\frac{\dot F(t)}{F(t)}\right)y-\frac{K}{F(t)}y^2,
\end{equation*}
where $K$ and $L$ are real constant, can be integrated through quadratures. It can be verified that the above family of Riccati equations
holds the assumption of Corollary \ref{CTU}. Indeed, taking $c_0=1$, $c_2=-K$, $c_1=L$ we get that $\kappa=1$,
\begin{equation*}
\sqrt{\frac{c_0c_2}{b_0b_2}}\left(b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot
b_0}{b_0}\right)\right)=L=c_1,
\end{equation*}
and $D=\sqrt{{b_2b_2}/{c_2c_0}}=1$. Therefore, Theorem \ref{TU} shows that the above family of Riccati equations can be integrated.
Moreover, taking the above values of the constants $c_0$, $c_1$, $c_2$ and the function $b_0(t)=F(t)$, Corollary \ref{C2TU} reproduces the family of Riccati equations analysed by Kovalevskaya.
\medskip
$\bullet$ {\it The case of Hong-Xiang}:
\medskip
As a final example we can consider the Riccati equation
\begin{equation*}
\frac{dy}{dt}=- y^2-\left(2 b G(t)-\frac{\dot G(t)}{G(t)}\right) y-c G^2(t),
\end{equation*}
used to analyse a certain integrable linear differential equation in \cite{Ro07} which was also analysed by Hong-Xiang \cite{HX82}. The above Riccati equation
satisfies the integrability condition (\ref{resCor2}) and hence it can be integrated. Indeed, we have
$$
\left\{
\begin{aligned}
b_0(t)&=-c G^2(t),\\
b_1(t)&=-\left(2bG(t)-\frac{\dot G(t)}{G(t)}\right),\\
b_2(t)&=-1,
\end{aligned}\right.
$$
and therefore we get that
$$
\frac{b_1+\frac{1}{2}\left(\frac{\dot b_2}{b_2}-\frac{\dot b_0}{b_0}\right)}{\sqrt{|b_0b_2|}}={\rm const.}
$$
In summary, many integrability conditions shown in the literature are equivalent to or
particular instances of those given in our more general statements.
\section{Integrability and reduction}
\indent
In this Section we develop a procedure that is similar
to the one derived throughout the previous Sections but we here consider solutions of system
(\ref{FS}) in two-parameter subsets of $SL(2,\mathbb{R})$. In this case, we recover
some known integrability conditions, e.g. a certain kind of integrability used in \cite{CarRamGra}. More specifically, we try to relate a Riccati equation
(\ref{ricceq}) to an integrable one associated, as a Lie system, with a curve of the form
${\rm a}'(t)=-D(t)(c_0{\rm a}_0+c_1{\rm a}_1+c_2{\rm a}_2)$, with $c_2\neq 0$ and a non-vanishing function $D=D(t)$. Furthermore, we consider solutions of system (\ref{Sys}) with $\gamma=0$ and $\alpha>0$ related to elements of $SL(2,\mathbb{R})$, i.e. we analyse transformations
$$y'=\frac{\alpha(t)}{\delta(t)}y+\frac{\beta(t)}{\delta(t)}=\alpha^2(t)\,y+
\alpha(t)\beta(t)\,.$$
In this case, using the expression in coordinates (\ref{FS}) of system (\ref{Sys}), we get that
\begin{equation}\label{PC}
\left(\begin{matrix}
\dot\alpha\\
\dot\beta\\
0\\
\dot\delta
\end{matrix}\right)=\left(
\begin{matrix}
\frac{b'_1-b_1}{2}&b_2 &b'_0&0\\
-b_0& \frac{b'_1+b_1}{2}&0 &b'_0\\
-b'_2&0 &-\frac{b'_1+b_1}{2}& b_2\\
0&-b_2' &-b_0& -\frac{b'_1-b_1}{2}
\end{matrix}\right)\left(
\begin{matrix}
\alpha\\
\beta\\
0\\
\delta
\end{matrix}\right)\,,
\end{equation}
where $b'_j=D\,c_j$ and $c_j\in\mathbb{R}$ for $j=0,1,2$. As we suppose $b'_2\neq 0$, the third equation of the above system implies
\begin{equation*}
\frac{\alpha}{\delta}=\frac{b_2}{b_2'}.
\end{equation*}
As $\alpha\delta=1$ in order to obtain a solution of (\ref{Sys}) related to an element of $SL(2,\mathbb{R})$ and $b_2'=D c_2$,
we get
\begin{equation}\label{Drelation}
\alpha^2=\frac{b_2}{D c_2}.
\end{equation}
Hence, $\alpha$ is determined, up to a sign, by the values of $b_2(t), D$ and $c_2$. In this way, if we take $\alpha$ to be positive, the first differential equation of system (\ref{PC}) gives us
the value of $\beta$ in terms of the related initial and final Riccati equation, i.e.
$$
\beta=\frac{1}{b_2}\left(\dot \alpha-\frac{b'_1-b_1}{2}\alpha\right).
$$
Taking into account the relation (\ref{Drelation}), the above expression is equivalent to the differential equation
\begin{equation*}
\frac{dD}{dt}=\left(b_1(t)+\frac{\dot b_2(t)}{b_2(t)}\right)D-c_1D^2-2b_2(t)D\beta \left(\frac{c_2D}{b_2(t)}\right)^{1/2},
\end{equation*}
and, as $\alpha\delta= 1$, we can define $M=\beta/\alpha$ and rewrite the above expression as follows
\begin{equation*}
\frac{dD}{dt}=\left(b_1(t)+\frac{\dot b_2(t)}{b_2(t)}\right)D-c_1D^2-2b_2(t)MD.
\end{equation*}
Considering the differential equation in $\dot \beta$ in terms of $M$, we get the equation
\begin{equation*}
\frac{dM}{dt}=-b_0(t)+\frac{c_0c_2}{b_2(t)}D^2+b_1(t) M-b_2(t) M^2\,.
\end{equation*}
Finally, as $\delta\alpha=1$ is a first-integral of system (\ref{Sys}), if the system for the variables $M$ and $D$ and all the obtained conditions are satisfied, the value $\delta=\alpha^{-1}$ satisfies its corresponding differential equation of the system (\ref{PC}). To sum up, we have obtained the following result.
\begin{theorem}\label{FT2} Given a Riccati equation (\ref{ricceq}),
there exists a transformation
\begin{equation*}
y'=G(t)y+H(t)\,,\qquad G(t)>0\,,
\end{equation*}
relating it to the integrable equation
\begin{equation}\label{fequation}
\frac{dy'}{dt}=D(t)(c_0+c_1y'+c_2y'^2),
\end{equation}
with $c_2\neq 0$ and $D$ a non-vanishing function, if and only if there exist functions $D$ and $M$ satisfying the following system
\begin{eqnarray*}
\left\{\begin{aligned}
\frac{dD}{dt}&=\left(b_1(t)+\frac{\dot b_2(t)}{b_2(t)}\right)D-c_1D^2-2b_2(t)MD,\\
\frac{dM}{dt}&=-b_0(t)+\frac{c_0c_2}{b_2(t)}D^2+b_1(t) M-b_2(t) M^2.
\end{aligned}\right.
\end{eqnarray*}
The transformation is then given by
\begin{equation}\label{ChangeT3}
y'=\frac{b_2(t)}{D(t)c_2}(y+M(t))\,.
\end{equation}
\end{theorem}
Consider $c_0=0$ in Eq. (\ref{fequation}). Thus, the system determining the curve
in $SL(2,\mathbb{R})$ performing the transformation of Theorem \ref{FT2} is
\begin{equation}
\left\{
\begin{array}{rcl}
\dfrac{dD}{dt}&=&\left(b_1(t)+\dfrac{\dot b_2(t)}{b_2(t)}\right)D-c_1D^2(t)-2b_2(t)MD,\\
\dfrac{dM}{dt}&=&-b_0(t)+b_1(t) M-b_2(t) M^2.
\end{array}\label{RedSep}\right.
\end{equation}
On one hand, this system does not involve any integrability condition because, as a consequence of the Theorem of existence and uniqueness of solutions,
there always exists a solution for every initial condition. On the other hand, such solutions can be as difficult to be found as the general solution of the initial
Riccati equation. Hence, in order to find a particular solution, we need to look for some
simplifications. For instance, we can consider the case in which $M=b_1/b_2$. In this case, the first differential equation of the above system does not depend on $M$ and reads
$$
\frac{dD}{dt}=\left(-b_1(t)+\frac{\dot b_2(t)}{b_2(t)}\right)D-c_1D^2
$$
and it is integrable by quadratures. Its solution reads
\begin{equation*}
D(t)=\frac{\exp\left(\int_0^t A(t')dt'\right)}{C+c_1\int^t_0\exp\left(\int_0^{t''} A(t')dt'\right)dt''}\,,\qquad A(t)=\left(-b_1(t)+\frac{\dot b_2(t)}{b_2(t)}\right).
\end{equation*}
Meanwhile, the condition for $M=b_1/b_2$ to be a solution of the second equation in
(\ref{RedSep}) is
\begin{equation*}
\frac{d}{dt}\left(\frac{b_1}{b_2}\right)=-b_0\,,
\end{equation*}
giving rise to an integrability condition. This summarises one of the integrability conditions considered in \cite{Ra62}.
Next, we recover from this new viewpoint the well-known result that the knowledge of a particular solution of the Riccati equation allows us to solve the system (\ref{RedSep}). In fact, under the change of variables $M\mapsto -y$, system (\ref{RedSep}) becomes
\begin{eqnarray}\label{eq8}
\left\{\begin{aligned}
\frac{dD}{dt}&=\left(b_1(t)+\frac{\dot b_2(t)}{b_2(t)}\right)D-c_1D^2+2b_2(t)yD,\\
\dfrac{dy}{dt}&=b_0(t)+b_1(t) y+b_2(t) y^2.
\end{aligned}\right.
\end{eqnarray}
Note that each particular solution of the above system is the form $(D_p(t),y_p(t))$, with $y_p(t)$ a particular solution of the Riccati equation (\ref{ricceq}). Therefore, given such a particular solution $y_p(t)$, the function $D_p=D_p(t)$, corresponding to the particular solution $(D_p(t),y_p(t))$ of system (\ref{eq8}), holds the integrable equation
\begin{equation}\label{PS}
\frac{dD_p}{dt}=\left(b_1(t)+\frac{\dot b_2(t)}{b_2(t)}+2b_2(t)y_p(t)\right)D_p-c_1D_p^2.
\end{equation}
Hence, the knowledge of a particular solution $y_p(t)$ of the Riccati equation (\ref{ricceq}) enables us to get a particular solution $(D_p(t),y_p(t))$ of system (\ref{eq8}) and, taking into account the change of variables $y\mapsto -M$, a particular solution $(D_p(t),M_p(t))=(D_p(t),-y_p(t))$ of system (\ref{RedSep}). Finally, the functions $M_p(t)$ and $D(t)$ determines a change of variables (\ref{ChangeT3}) given by Theorem \ref{FT2} transforming the initial Riccati equation (\ref{ricceq}) into another one related, as a Lie system, to a solvable Lie algebra of vector fields. In this way, we describe a reduction process similar to that one pointed out in \cite{CarRamGra}. Nevertheless, we here directly obtain a reduction to a Riccati equation related, as a Lie system, to a one dimensional Lie subalgebra of $\mathfrak{sl}(2,\mathbb{R})$ through one of its particular
solutions.
There exists many ways to impose conditions on the coefficients of the second
equation of (\ref{eq8}) for being able to obtain one of its particular
solutions easily. Now, we give some particular examples of this.
If there exists a real constant $c$ such that for the time-dependent functions $b_0$, $b_1$ and $b_2$ we have that $b_0+b_1 c+b_2 c^2=0$, then $c$ is a particular solution. This resumes some cases found in
\cite{CarRamGra, Stre}. For instance:
\begin{enumerate}
\item $b_0+b_1+b_2=0$ means that $c=1$ is a particular solution.
\item $c_1^2b_0+c_1c_2b_1+c_2^2b_2=0$ means that $c=c_2/c_1$ is a particular solution.
\end{enumerate}
In these particular instances, we can find $D$ through the first differential
equation of (\ref{eq8}).
As a first application of this last case we can integrate the Riccati equation
\begin{equation}\label{Hovy}
\frac{dy}{dt}=-\frac{n}{t}+\left(1+\frac{n}{t}\right)y-y^2.
\end{equation}
related to Hovy's equation \cite{Ro07}. This Riccati equation admits the particular constant solution $y_p(t)=1$. Using such a particular solution in Eq. (\ref{PS}) and fixing, for instance, $c_1=0$, we can obtain a particular solution for Eq. (\ref{PS}), e.g. $D_p(t)=t^ne^{-t}$. Therefore, we get that $(t^ne^{-t},1)$ is a solution of the system (\ref{eq8}) related to Eq. (\ref{Hovy}) and $(t^ne^{-t},-1)$ is a solution of the system (\ref{RedSep}). In this way, Theorem \ref{FT2} states that the transformation (\ref{ChangeT3}), determined by the $D_p(t)=t^ne^{-t}$ and $M_p(t)=-1$, of the form
\begin{equation}\label{rel}
y'=-t^{-n}e^tc_2^{-1}(y-1),
\end{equation}
relates the solutions of Eq. (\ref{Hovy}) to the integrable one
$$
\frac{dy'}{dt}=e^{-t}t^n(c_0+c_2y'^2).
$$
If we fix $c_0=1$ and $c_2=1$ the solution for the above equation is
\begin{equation*}
y'(t)=-\frac{1}{-K+\Gamma(1+n,t)},
\end{equation*}
where $K$ is an integration constant and $\Gamma(a,b)$ is the incomplete Euler's Gamma function
\begin{equation*}
\Gamma(a,t)=\int^\infty_t t'^{a-1}e^{-t'}dt'.
\end{equation*}
In view of the change of variables (\ref{rel}), the solutions $y(t)$ of the
Riccati equation (\ref{Hovy})
and $y'(t)$ are related by means of the expression
$y'(t)=-t^{-n}e^tc_2^{-1}(y(t)-1)$. Therefore, if we
substitute the general solution $y'(t)$ in this expression we can derive the
general solution
for the Riccati equation (\ref{Hovy}), that is,
\begin{equation*}
y(t)=1-\frac{e^{-t}t^n}{\Gamma(n+1,t)+K}.
\end{equation*}
Another approach that can be summarised by Theorem \ref{FT2} is the factorisation
method developed in \cite{Ro07} to explain an integrability process for second-order
differential equations. In that work, it was analysed the differential equation:
\begin{equation}\label{SOE}
\frac{d^2y}{dt^2}+2P(t)\frac{dy}{dt}+\left(\frac{dP}{dt}+P^2(t)-\frac{d\phi}
{dt}-\phi^2(t)\right)y=0\,.
\end{equation}
We know that invariance under dilations leads to consider an adapted
variable $z$, such that $y=e^z$. Under this change of variables the equation obtained for $\psi=\dot z$ is the
Riccati equation
\begin{equation}\label{FRE}
\frac{d\psi}{dt}=-\psi^2-2P(t)\psi-\left(\frac{dP}{dt}(t)+P^2(t)-\frac{d\phi}
{dt}(t)-\phi^2(t)\right).
\end{equation}
This equation was integrated through a factorisation method in
\cite{Ro07}. Nevertheless, we can also integrate
this equation if we take into account that $\psi_p(t)=\phi(t)-P(t)$ is
a particular solution of the above
differential equation and then applying the same procedure as for Eq.
(\ref{Hovy}). Indeed,
as $\psi_p(t)$ is a particular solution for the Riccati equation (\ref{FRE}),
we can obtain a
particular solution $D_p=D_p(t)$ for Eq. (\ref{PS}) and by means of
the functions
$M_p(t)=-\psi_p(t)$ and $D_p(t)$ we can obtain the solution of the Riccati
equation (\ref{FRE}).
Finally, inverting the change of variables used to relate Eq.
(\ref{SOE}) to (\ref{FRE})
we obtain the solution for Eq. (\ref{SOE}).
\section{Linearisation of Riccati equations}
\indent
One can also study the problem of the linearisation of Riccati equations
through the linear fractional
transformations (\ref{yprime}). This set
of time-dependent transformations is general enough to include many of the
time-dependent or time-independent changes of variables already used to study Riccati equations, e.g. it
allows us to recover the results of
\cite{RDM05}. As a main result, we state in this Section some integrability
conditions to be able to
transform a $t$-dependent
Riccati equation into a linear one by means of a diffeomorphism on
$\overline{\mathbb{R}}$ associated with
certain linear fractional transformations.
As a first insight in the linearisation process, note that Corollary \ref{CorCur} shows that
there always exists a curve in $SL(2,\mathbb{R})$, and then a
$t$-dependent
linear fractional
transformation on ${\overline{\mathbb{R}}}$, transforming a given Riccati
equation into any other one.
In particular, if we fix $b_2'=0$ in the final Riccati equation, we obtain that there is
a $t$-dependent linear fractional change of variables transforming any
Riccati equation (\ref{ricceq})
into a linear one.
Nevertheless, as the Lie system (\ref{Sys}) describing such a transformation is
not related to a
solvable Lie algebra of vector fields, it is not easy to find such a transformation in the general case.
Let us try to relate a Riccati equation (\ref{ricceq}) to a linear differential
equation by means of
a linear fractional transformation (\ref{Action}) determined by a vector
$(\alpha,\beta,\gamma,\delta)\in \mathbb{R}^4$
with $\alpha\delta-\beta\gamma=1$. In this case, the existence of solutions of
the system (\ref{FS}) performing
such a transformation is an easy task and we can look for integrability
conditions to get the corresponding
change of variables. Note that as $(\alpha,\beta,\gamma,\delta)$ is a constant,
we have
$\dot\alpha=\dot\beta=\dot\gamma=\dot\delta=0$ and, in view of (\ref{FS}), the
diffeomorphism on
$\overline{\mathbb{R}}$ performing the transformation is related to a vector in the kernel of the matrix
\begin{equation}\label{EM}
B=\left(\begin{matrix}
\frac{b'_1-b_1}{2}&b_2 &b'_0&0\\
-b_0& \frac{b'_1+b_1}{2}&0 &b'_0\\
0&0 &-\frac{b'_1+b_1}{2}& b_2\\
0&0 &-b_0& -\frac{b'_1-b_1}{2}
\end{matrix}\right),
\end{equation}
where we assume $b_0b_2\neq 0$ in an open interval in the variable $t$. We
leave out the study of the case
$b_0b_2=0$ in an open interval because, as it was shown in Sec.
\ref{IntRicEqu}, this case is
known to be integrable.
The necessary and sufficient condition for a non-trivial $\ker B$ is $\det B=
0$ and, therefore, a
short calculation shows that $\dim\,\, {\rm ker}\, B>0$ if and only if
$(-b_1^2+b_1'^2(t)+4 b_0 b_2)^2=0.$ Thus, $b_1'=\pm \sqrt{b_1^2-4 b_0 b_2}$ and
$b_1'$ is fixed, but a sign,
by the values of $b_0$, $b_1$ and $b_2$. Let us study the kernel of the matrix
$B$ in the positive
and negative cases for $b'_1$.
$\bullet$ Positive case:
The kernel of the matrix (\ref{EM}) is given by the vectors
\begin{equation*}
\left(\delta\frac{b_0'}{b_0}+\beta\frac{b_1+\sqrt{b_1^2-4 b_0 b_2}}{2 b_0}
,\beta,-\delta\frac{-b_1+\sqrt{b_1^2-4 b_0b_2}}{2 b_0},\delta\right), \qquad \delta,\beta\in\mathbb{R}.
\end{equation*}
Recall that we only consider the constant elements of $\ker B$, therefore there
should be two real
constants $K_1$ and $K_2$ such that
\begin{eqnarray*}
K_1=\delta\frac{b_0'}{b_0}+\beta\frac{b_1+\sqrt{b_1^2-4 b_0 b_2}}{2 b_0},\qquad K_2=\frac{-b_1+\sqrt{b_1^2-4 b_0b_2}}{2 b_0}.
\end{eqnarray*}
Moreover, in order to relate these vectors to elements in $SL(2,\mathbb{R})$ we
have to
impose that $\det (K_1,\beta,-\delta K_2,\delta)=\delta(K_1 +\beta K_2)=1$.
The second condition imposes a restriction on the coefficients of the
initial Riccati equation to be linearisable by a constant linear fractional
transformation (\ref{Action}).
Then, if this condition is satisfied we can fix $\beta,\gamma, K_1$ and $b_0'$
to satisfy the other conditions.
Thus, the only linearisation condition is the condition on $K_2$.
$\bullet$ Negative case: In this case, $\ker\,B$ reads
\begin{equation*}
\left(\delta \frac{b_0'}{b_0}+\beta\frac{b_1-\sqrt{b_1^2-4 b_0 b_2}}{2 b_0}
,\beta,-\delta\frac{-b_1-\sqrt{b_1^2-4 b_0b_2}}{2 b_0},\delta\right),\qquad \delta,\beta\in\mathbb{R},
\end{equation*}
and now the new conditions reduce to the existence of two real constants $K_1$ and $K_2$ such that
\begin{eqnarray*}
K_1=\delta \frac{b_0'}{b_0}+\beta\frac{b_1-\sqrt{b_1^2-4 b_0 b_2}}{2
b_0},\qquad K_2=
\frac{-b_1-\sqrt{b_1^2-4 b_0b_2}}{2 b_0},
\end{eqnarray*}
with $\delta(K_1+\beta K_2)=1$. If the condition in $K_2$ is satisfied we can
proceed as in
the positive case to obtain the transformation performing the linearisation of
the
initial Riccati equation.
In summary:
\begin{theorem}
The necessary and sufficient condition for the existence of a diffeomorphism on
$\bar{\mathbb{R}}$ of linear
fractional type associated with a transformation
of $SL(2,\mathbb{R})$ transforming the Riccati equation (\ref{ricceq}) into a
linear differential equation
is the existence of a real constant $K$ such that
\begin{equation}\label{IntCond}
K=\frac{-b_1\pm\sqrt{b_1^2-4 b_0b_2}}{2 b_0}.
\end{equation}
\end{theorem}
As a Riccati equation (\ref{ricceq}) holds condition (\ref{IntCond}) if and
only if $K$ is a constant
particular solution, we get the following corollary:
\begin{corollary}
A Riccati equation can be linearised by means of a diffeomorphism on
$\overline{\mathbb{R}}$ of the form
(\ref{Action}) if and only if it admits a constant particular solution.
\end{corollary}
Ibragimov showed that a Riccati equation (\ref{ricceq}) is linearisable by
means of a change
of variables $z=z(y)$ if and only if the Riccati equation admits a constant
solution \cite{Ib08}.
Additionally, we have proved that in such a case, the change of variables
can be described by
means of a transformation of the type (\ref{Action}).
Now, it can be checked that the example given in \cite{Ib08,RDM05} satisfies
the above integrability
condition. In this work, the differential equation
\begin{equation*}
\frac{dy}{dt}=P(t)+Q(t)y+k(Q(t)-kP(t))y^2,
\end{equation*}
was studied. The only interesting case is that with $k\neq 0$ because the other
ones are linear.
In this latter case, $b_0(t)=P(t)$, $b_1(t)=Q(t)$ and $b_2(t)=k(Q(t)-kP(t))$. Hence,
\begin{eqnarray*}
\begin{aligned}
\frac{-b_1-\sqrt{b_1^2-4 b_0b_2}}{2 b_0}=-k,
\end{aligned}
\end{eqnarray*}
and the integrability condition (\ref{IntCond}) holds. Now we may fix $K_1=0$
and we look for a
solution for the condition $\det(K_1,\beta,-\delta K_2,\delta)=1$ reading
$k\delta\beta=-1$.
As $k\neq 0$, we can take $\beta=-1/k$ to get from the above condition that
$\delta=1$.
Thus the transformation is that one associated with the vector $(0,-1/k,k,1)$,
i.e.
the linear fractional transformation
\begin{equation*}
y'=\frac{-1/k}{ky+1}
\end{equation*}
that is the same found in \cite{RDM05}.
In this way we only have to obtain $b_0'$ from the condition
$$
K_1=0=\delta\frac{b_0'}{b_0}+\beta\frac{b_1+\sqrt{b_1^2-4 b_0 b_2}}{2 b_0},
$$ to get the final linear differential equation, that is,
\begin{equation*}
\frac{dy'}{dt}=\frac{Q(t)}{P(t)}+(Q(t)-2P(t)k)y',
\end{equation*}
as it appears in \cite{CRL07}.
\section{Conclusions and outlook}
\indent
It has been shown that previous works about the integrability of the Riccati
equation can be explained from the unifying viewpoint of Lie systems. The
transformations
used in the study of the integrability condition for these equations have been
understood as
induced by curves in $SL(2,\mathbb{R})$.
We have investigated a Lie system characterising the $t$-dependent fractional
transformations
relating different Riccati equations associated, as Lie systems,
with curves in $\mathfrak{sl}(2,\mathbb{R})$. We have used this differential equation
and considered some simple instances. These simplifications have been
used to analyse known integrability conditions and provide new ones.
We have also shown that the system (\ref{FS}) is a good way to describe
linear fractional time--dependent transformations and found necessary and sufficient
conditions for the linearisability or simplification of a Riccati equation
through time-independent
and time-dependent transformations obtained from curves in $SL(2,\mathbb{R})$.
There are many ways of simplifying (\ref{Sys}) and some of them have been
developed here. Other ways can be used to obtain new
integrability conditions.
Finally, the theory used here can be extended to any other Lie system to
provide new or
recover known integrability conditions. This fact is to be developed in forthcoming works.
\section*{Acknowledgements}
\indent
Partial financial support by research projects MTM2009-11154, MTM2009-08166-E
and E24/1 (DGA)
are acknowledged. JdL also acknowledges
a F.P.U. grant from Ministerio de Educaci\'on y Ciencia.
|
\section{Introduction}
\begin{wrapfigure}{r}{0.45\textwidth}
\vbox{\hrule height0pt
\vskip -2pc
\hbox{\includegraphics[width=0.4\textwidth]{fig1.eps}}}
\caption{The solar axions produced via the Primakoff process in the
solar core are, then, converted into X-rays via the reverse process
in the magnet.}
\label{fig:principle}
\end{wrapfigure}
The axion is a Nambu--Goldstone boson of a spontaneously broken U(1)
symmetry, the Peccei--Quinn symmetry, which was introduced to solve
the strong CP problem in quantum chromodynamics (QCD)
\cite{axion-bible1}.
If its mass $m_a$ is at around a few electronvolts, the sun can be
a powerful source of axions and the so-called `axion helioscope'
technique may enable us to detect such axions
\cite{sikivie1983,bibber1989}.
The principle of the axion helioscope is illustrated in
Fig.~\ref{fig:principle}.
Axions are expected to be produced in the solar core through the
Primakoff process.
The average energy of the solar axions is 4.2\,keV and their
differential flux expected at the Earth is approximated by
\cite{bahcall2004,raffelt2005}
\begin{eqnarray}
{\rm d}\Phi_a/{\rm d}E&=&6.020\times10^{10}[\mathrm{cm^{-2}s^{-1}keV^{-1}}]
\nonumber\\
&&{}\times\left(g_{a\gamma}\over10^{-10}\mathrm{GeV}^{-1}\right)^2
\left( {E\over1\rm\,keV} \right)^{2.481}
\exp \left( -{E\over1.205\rm\,keV} \right),
\label{eq:aflux}
\end{eqnarray}
where $E$ is the energy of the axions and $g_{a\gamma}$ is the axion-photon
coupling constant.
Then, they would be converted into X-ray photons through the inverse
process in a strong magnetic field at a laboratory.
The conversion rate is given by
\begin{equation}
P_{a\to\gamma} = {g_{a\gamma}^2\over4}
\exp\left[-\int_0^L\!\!\mathrm{d}z\,\Gamma \right]
\times
\left| \int_0^L\!\!{\rm d}z\,B_\bot\exp
\left[i \int_0^z\!\!\mathrm{d}z^\prime
\left(q - {i\Gamma\over2}\right) \right]\right|^2,
\label{eq:prob}
\end{equation}
where $z$ and $z^\prime$ are the coordinate along the incident solar
axion, $B_\bot$ is the strength of the transverse magnetic field,
$L$ is the length of the field along $z$-axis, $\Gamma$ is the X-ray
absorption coefficient of the filling medium,
$q=(m_\gamma^2-m_a^2)/2E$ is the momentum transfer by the virtual
photon, and $m_\gamma$ is the effective mass of the photon in medium.
In light gas, such as hydrogen or helium, $m_\gamma$ is well
approximated by
\begin{equation}
m_\gamma=\sqrt{4\pi\alpha N_e/m_e},
\end{equation}
where $\alpha$ is the fine structure constant, $m_e$ is the electron
mass, and $N_e$ is the number density of electrons.
In vacuum, the axion helioscope is not sensitive to massive axions due
to a loss of coherence by non-zero $q$.
However, coherence can be restored if one can adjust $m_\gamma$ to $m_a$.
We adopted cold \iso{4}{He} gas as a dispersion-matching medium.
The gas was kept at almost the same temperature as the magnet,
$T\lesssim6\rm\,K$, which is just above the critical temperature of
\iso{4}{He}, $T_c=5.2\rm\,K$.
It is worth noting that axions as heavy as a few electronvolts can be
reached with helium gas of only about one atmosphere and \iso{4}{He}
will not liquify at any pressure at this temperature.
In 1997, the first phase measurement \cite{sumico1997} was performed
without the gas container.
In this measurement, the sensitive mass region was limited to
$m_a<0.03\rm\,eV$ since the conversion region was vacuum.
In 2000, the second phase measurement \cite{sumico2000} was performed
to search for sub-electronvolt axions.
This experiment, together with the first phase measurement
\cite{sumico1997}, yielded an upper limit of
$g_{a\gamma}<\hbox{6.0--10.5}\times10^{-10}\rm GeV^{-1}$ (95\% CL) for
$m_a<0.27\rm\,eV$.
Here, we will present the result of the third phase measurement in
which the mass region around 1\,eV was scanned using the upgraded
apparatus to withstand higher pressure gas.
\section{Experimental apparatus}
\begin{wrapfigure}{r}{0.45\textwidth}
\vbox{\hrule height 0pt
\vskip -2pc
\hbox{\includegraphics[width=0.45\textwidth]{fig_sumico_grey.eps}}%
\hrule height 0pt}
\caption{The schematic view of the axion helioscope called
the Sumico V detector.}
\label{fig:sumico}
\end{wrapfigure}
The axion helioscope consists of a superconducting magnet, X-ray
detectors, a gas container, and an altazimuth mounting to track the sun.
The schematic figure is shown in Fig.~\ref{fig:sumico}.
The superconducting magnet \cite{sato1997} consists of two 2.3-m long
race-track shaped coils running parallel with a 20-mm wide gap between
them.
The magnetic field in the gap is 4\,T perpendicular to the helioscope
axis.
The coils are kept at 5--6\,K during operation.
The magnet is made cryogen-free by two Gifford-McMahon refrigerators
cooling directly by conduction, and is equipped with a persistent
current switch.
The container to hold dispersion-matching gas is inserted in the gap
of the magnet.
Its body is made of four stainless-steel square pipes welded side by
side to each other, and is wrapped with 5N high purity aluminium sheet
to achieve high uniformity of temperature.
The uniformity of the temperature guarantees the homogeneous density
along the length of the container.
The detector side of the container is ended with an X-ray window which
is transparent to X-ray above 2\,keV and can withstand up to 0.3\,MPa.
The container is fixed to the magnet at this side through a
temperature-stablilized thermal linkage.
The solar end at the opposite side is blind-ended and is suspended by
three Kevlar cords, so that thermal flow through this end is highly
suppressed.
\hyphenation{ho-ri-ba-stec}
To have automatic sequential pressure settings, a gas handling system
is built with HORIBASTEC Piezo valves and a YOKOGAWA precision
pressure gauge.
For emergency exhaust of the gas in case of rapid temperature increase
due to a magnet quenching, a cryogenic rupture disk, which is designed
to break at 0.248 MPa, is also introduced into the gas handling system
to avoid destruction of the X-ray window by the over pressure.
Sixteen PIN photodiodes, Hamamatsu Photonics S3590-06-SPL, are used as
the X-ray detectors \cite{naniwaPIN}.
In the present measurement, however, twelve of them are used for the
analysis because four went defective through thermal stresses.
The chip size of a photodiode is $11\times11\times0.5\rm\,mm^3$, and
the effective area is larger than $9\times9\,\mathrm{mm^2}$.
It has an inactive surface layer of $0.35\,\mu\mathrm{m}$
\cite{akimotoPIN}.
The output from each photodiode is fed to a charge sensitive
preamplifier and waveforms of the preamplifier outputs are digitized
using FADCs.
We applied off-line pulse shaping to the recorded waveforms as
described in Ref.~\cite{sumico2000}.
Each photodiode was calibrated by 5.9-keV Mn X-rays from a \iso{55}{Fe}
source installed in front of them.
The source is manipulated from the outside and is completely retracted
behind the shield during the observations.
The entire axion detector is constructed in a vacuum vessel and the
vessel is mounted on a computer-controlled altazimuth mount.
Its trackable altitude ranges from $-28^\circ$ to $+28^\circ$ and its
trackable azimuthal range is almost $360^\circ$.
However, in the present measurement, the azimuthal range without a
human intervention is restricted to about 60$^\circ$ because a cable
handling system for its unmanned operation is not completed yet.
This range corresponds to an exposure time of about a quarter of a
day in observing the sun.
During the rest three quarters of a day, background was measured.
\section{Measurement and Analysis}
From December 21 2007 through April 21 2008, a new measurement was
performed for 34 gas-density settings with about three days of running
time per setting.
The scanned mass range was 0.84--1 eV.
Since we had not completed the gas relief system, the highest density
was determined so that the gas pressure would not exceed the breakage
pressure of the rupture disk even during a magnet quenching.
Energy spectra for the solar observation and the background are
obtained for each density settings based on the measured direction of
the helioscope.
Event reduction process was applied in the same way as described in
Ref.~\cite{sumico2000}.
We searched for expected axion signals which scale with $g_{a\gamma}^4$ for
various $m_a$ in these spectra by applying a series of least $\chi^2$
fittings assuming various $m_a$ values.
Data from the 34 different gas density settings were combined by using
the summed $\chi^2$ of the 34.
The energy region of 4--20\,keV was used for fitting.
As a result, no significant excess was seen for any $m_a$, and thus
an upper limit on $g_{a\gamma}$ at 95\% confidence level was given.
Fig.~\ref{fig:exclusion} shows the limit plotted as a function of
$m_a$.
The previous limits from the first \cite{sumico1997} and the second
\cite{sumico2000} phase measurements and some other bounds are
also plotted in the same figure.
\begin{figure}
\includegraphics[scale=0.85]{fig_limit.eps}
\hfill
\includegraphics[scale=0.85]{fig_limit_closer.eps}
\caption{
The left figure show the exclusion plot for $g_{a\gamma}$ as a function
of $m_a$.
The shaded area corresponds to the preferred axion models
\cite{GUT_axion}.
Two slopes in the area correspond to two typical axion models.
Other lines show the upper limits and preliminary ones are in
dashed lines.
The new limit and the previous ones\cite{sumico1997,sumico2000}
are plotted in blue lines.
Limits by other axion helioscope experiments: Lazarus
{\it et al.}~\cite{Lazarus} and CAST~\cite{CAST} are shown in red lines.
Limits by axion Bragg scattering experiments \cite{Pascos}:
SOLAX~\cite{solax1999}, COSME~\cite{cosme2002},
DAMA~\cite{DAMA2001}, and CDMS~\cite{cdms2008} are shown in green
lines.
The black solid line shows the solar limit inferred from the solar
age consideration.
The right figure show the magnified view of the new limit.}
\label{fig:exclusion}
\end{figure}
\section{Conclusion}
The axion mass around 1\,eV has been scanned with an axion helioscope
with cold helium gas as the dispersion-matching medium in the
$4{\rm\,T}\times2.3\rm\,m$ magnetic field, but no evidence for solar
axions was seen.
A new limit on $g_{a\gamma}$ shown in Fig. \ref{fig:exclusion} was set for
$0.84<m_a<1.00\rm\,eV$.
This is the first result to search for the axion in the $g_{a\gamma}$-$m_a$
parameter region of the realistic axion models \cite{GUT_axion} with a
magnetic helioscope. Full description of the present result is
published in Ref.~\cite{sumico2007}.
\acknowledgments The authors thank the former director general of KEK,
Professor H. Sugawara, for his support in the beginning of the
helioscope experiment.
This research was partially supported by the Japanese Ministry of
Education, Science, Sports and Culture, Grant-in-Aid for COE Research
and Grant-in-Aid for Scientic Research (B), and also by the
Matsuo Foundation.
|
\section{Introduction}
This paper describes a geometrical study of the system of nonlinear partial differential equations
\begin{equation}\label{nWaveSystem}
\frac{\partial A_{j\,k}}{\partial x_i}=A_{j\,i}A_{i\,k},\ \ \ \ \text{for all}\ \ (i,\,j,\,k)\in \text{perm}_3(n)
\end{equation}
where $\text{perm}_3(n)$ denotes the set of all permutations of 3 letters out of $n$. Note that the convention of summation over repeated indices is not in force in equation (\ref{nWaveSystem}) and will not be used in this paper. Any summation will make explicit use of the summation symbol.
As we will explain in more detail in the next section, system (\ref{nWaveSystem}) plays a important role in aspects of submanifold geometry and integrable systems theory [\ref{Darboux10}, \ref{ZakharovManakov85}, \ref{Zakharov98}, \ref{Tsa}, \ref{JoshiKitaevTreharne08}] as well as arising as model equations in a variety of physical phenomena including nonlinear optics and plasma physics [\ref{ZakharovManakov73}, \ref{ZakharovManakov85}]. It also plays a role in mathematical physics, for instance,
[\ref{MangazeevSergeev01}, \ref{Sergeev07}].
A particularly important special case of (\ref{nWaveSystem}) is when $n=3$, known as the $2+1$-{\it dimensional three-wave resonant interaction system}
({\sf 3WRI}), which serves as a model equation for nonlinear wave interaction.
The main objectives of this paper are firstly, to introduce a new geometric setting for system (\ref{nWaveSystem}), namely, {\it $n$-hyperbolic manifolds}, and then show how this leads to new solutions and new perspectives on this integrable system. Secondly, to describe some of the consequences arising from these results for the above mentioned areas of application especially to {\it systems of hydrodynamic type}, which have a rich mathematical structure. The central object that draws all these many issues together is a class of overdetermined systems of linear hyperbolic partial differential equations, the {\it Darboux-Manakov-Zakharov systems} which we now discuss.
\section{Darboux-Manakov-Zakharov Systems}
\label{intro}
To set the results to be presented in context, we provide background on the crucial tool that is used in this paper, namely, Darboux-Manakov-Zakharov systems and explain the links that they have with various nonlinear phenomena.
The explicit modern study of overdetermined linear systems of partial differential equations of the form
\begin{equation}\label{KTsystem1}
\frac{{\partial}^2 u}{{\partial} x_i{\partial} x_j}-\Gamma_{j\,i}(x)\frac{{\partial} u}{{\partial} x_i}-
\Gamma_{ij}(x)\frac{{\partial} u}{{\partial} x_j}+C_{ij}(x)u=0,\
\ 1\leq i< j\leq n,
\end{equation}
for a real-valued function $u(x_1,\ldots,x_n)$, where\footnote{Warning: We take this opportunity to remind the reader that the convention of summation over repeated indices {\it will not} be used in this paper and is not in force in equation (\ref{KTsystem1}).} coefficients
$\Gamma_{ij},\Gamma_{j\,i}, C_{ij}=C_{j\,i}$ are smooth functions of the $x_i$
began in the work of Manakov \& Zakharov [\ref{ZakharovManakov85}, \ref{Zakharov98}] and Kamran \& Tenenblat [\ref{KT96}, \ref{KT98}, \ref{KT2000}]. The former studies were motivated by nonlinear optics and founded on the extensive classical works of 19th century geometers on the problem of triply orthogonal coordinate systems, particularly those of Darboux [\ref{Darboux1896}, \ref{Darboux10}]. The studies of Kamran and Tenenblat arose in their search for a suitable venue for generalising the classical Laplace Transformation to higher dimensions following work of Chern's [\ref{Chern44}] on the theory of Cartan submanifolds. The work of all these authors turns out to be extremely interesting and is connected with a variety of phenomena in submanifold geometry and integrable systems theory.
However a challenge is that for $n>2$, system (\ref{KTsystem1}) is not involutive in general and the applications mentioned above very much rely on its involutivity. This state of affairs presents both a problem and an opportunity since the complicated integrability conditions that arise have been the object of study in mathematical physics for about three decades, at least in the case $n=3$, and much longer in differential geometry [\ref{Darboux1896}]. In the 3-dimensional case, they are equivalent to the well known 2+1-dimensional 3-wave resonant interaction ({\sf 3WRI}) equations and form part of the Lam\'e equations for triply orthogonal coordinate systems [\ref{ZakharovManakov85}, \ref{Zakharov98}, \ref{Darboux10}].
Until now essentially two approaches have been taken, in the recent literature, to the construction of solutions of the 2+1-dimensional {\sf 3WRI} system. Firstly, the Darboux-Manakov-Zakharov linear problem [\ref{ZakharovManakov85}] or else the Ablowitz-Haberman linear problem [\ref{AblowitzHaberman75}] have been used to obtain some solutions via the inverse scattering transform (IST) [\ref{Cornille79}, \ref{Kaup80}, \ref{Kaup81}, \ref{Zakharov98}]. In relation to this we mention also the works [\ref{Kaup81a}, \ref{LeviPilloniSantini81}] which construct B\"acklund transformations for (among other equations) the 2+1-dimensional {\sf 3WRI} system. Important solutions of the 2+1-dimensional {\sf 3WRI} system obtained in this way include the lump solutions [\ref{Kaup81}]. However, it should be mentioned that by comparison to other integrable systems, the IST solution of the {\sf 3WRI} system is very complicated and not so staightforward to apply.
Secondly, Lie symmetry methods have been extensively applied to obtain explicit solutions; references here include [\ref{Kitaev90}, \ref{LeoMartinaSoliani86}, \ref{MartinaWinternitz89}]. In the last reference a detailed study of the Lie point symmetry group of the 2+1-dimensional {\sf 3WRI} system is shown to be infinite-dimensional with Lie algebra of Kac-Moody-Virosoro type. Numerous explicit group-invariant solutions are constructed via symmetry reduction using the one- and two-dimensional Lie subgroups of the full symmetry group of the system. In particular, as first pointed out by Kitaev [\ref{Kitaev90}], there are solutions expressible in terms of the Painlev\'e transcendents. This fact has been used in more recent work [\ref{JoshiKitaevTreharne08}] to study the linearization of Painlev\'e equations.
Of course, another source of solutions of the {\sf 3WRI} system and extensive information on linear system (\ref{KTsystem1}) comes to us from the nineteenth century geometers. Many of these results are summarised in Darboux's famous treatise [\ref{Darboux10}]. Additionally, formulas relevant to the modern treatment of the Lam\'e equations via the inverse scattering transform are discussed in detail in the classical work [\ref{Darboux1896}, Chapter 12], without however, explicitly anticipating IST itself. A linear PDE system of the form (\ref{KTsystem1}) {\it which is involutive} in the sense of Cartan will be called a {\it linear Darboux-Manakov-Zakharov system} which we abreviate to {\it linear {\sf DMZ} system} or simply {\it {\sf DMZ} system} whenever there is little danger of confusion.
This paper takes a radically different approach to the problem of constructing solutions of the 2+1-dimensional {\sf 3WRI} equations by providing a geometric characterisation and construction of {\it generalised Darboux-Manakov-Zakharov systems}
\begin{equation}\label{GDMZ}
\frac{\partial^2 u}{\partial x_i\partial x_j}=f_{ij}\Big(x_k,u,\frac{\partial u}{\partial x_l}\Big),\ \ \
1\leq i<j\leq n,\ \ k,l\in\{1,\ldots,n\},
\end{equation}
which we sometimes denote {\it {\sf GDMZ} systems}, and then showing how to specialize to linear {\sf DMZ} systems. In consequence new tools are created for the study of linear {\sf DMZ} systems with remarkably many applications arising from the central role they play in the theory of integrable equations, submanifold geometry and mathematical physics. For instance, we construct, in case $n=3$, linear {\sf DMZ} systems which therefore provide new explicit solutions for the 2+1-dimensional {\sf 3WRI} equations. Note that the restriction to the case $n=3$ is not essential so that all the results can be easily extended to $n>3$. It is noteworthy that the problem of constructing linear {\sf DMZ} systems was first proposed by Darboux [\ref{Darboux1896}, Ch. 12]. There Darboux derives numerous results subject to the proviso of explicit knowledge of a linear {\sf DMZ} system; he states\footnote{pp 278-279}
\begin{center}
{\it ``Supposons que l'on ait obtenu par un moyen quelconque un syst\'eme\\ d'\'equations aux d\'eriv\'ees partielles} { [\,(\,\ref{KTsystem1}\,)\,]} {\it pour lequel les conditions d'int\'egrabilit\'e soient satisfaites et dont on ait d\'etermin\'e l'int\'egrale $u$."}
\end{center}
Darboux's motivation for this problem arose from the fact that linear {\sf DMZ} systems are deeply linked to the problem of constructing $n$-orthogonal coordinate systems in an $n$-dimensional flat Riemannian manifold, say $\mathbb{E}^n$, $n$-dimensional Euclidean space. To see this, let $x^1,\ldots,x^n$ be coordinates on $\mathbb{E}^n$. The problem of $n$-orthogonal coordinate systems is to find functions $u^i(x^1,\ldots,x^n)$, $i=1,\ldots,n$ such that the map $\boldsymbol{u}:x\mapsto u(x)$ is a local diffeomorphism and satisfies the PDE system
$$
\sum_{k=1}^n\frac{\partial u^i}{\partial x^k}\frac{\partial u^j}{\partial x^k}=0,\ \ \ \ \forall\ \ \ \ i,j\in\{1,2,\ldots,n\},\ \ \ i\neq j.
$$
The problem for $n=2$ is easy but much more difficult for $n\geq 3$ and it was intensively studied by numerous important 19th century geometers, leading to the so called {\it Lam\'e equations}. We will briefly outline the construction of these equations as they will play a role in subsequent sections of this paper. Let us denote the inverse of $\boldsymbol{u}$ by
$
\boldsymbol{x}(u^1,\ldots,u^n)
$
and define
$$
h_i^2=\sum_{k=1}^n\left(\frac{\partial x^i}{\partial u^k}\right)^2.
$$
The metric in $\mathbb{E}^n$ expressed in the $(u^i)$-coordinate system has the form
$$
\sum_{i=1}^n\,h_i^2\,(du^i)^2.
$$
The Christoffel symbols for this metric are
$$
\begin{aligned}
&\Gamma^i_{ml}=0,\ \ \ \ \ i\neq l\neq m,\cr
&\Gamma^i_{il}=\frac{1}{h_i}\frac{\partial h_i}{\partial u^l},\cr
&\Gamma^i_{ll}=-\frac{h_l}{h_i^2}\frac{\partial h_l}{\partial u^i},\ \ \ \ \ \ i\neq l.
\end{aligned}
$$
The vanishing of the Riemann curvature tensor leads to the PDE system
\begin{equation}\label{LameEquations}
\begin{aligned}
&\frac{\partial^2 h_i}{\partial u^l\partial u^m}-
\frac{1}{h_l}\frac{\partial h_l}{\partial u^m}\frac{\partial h_i}{\partial u^l}-
\frac{1}{h_m}\frac{\partial h_m}{\partial u^m}\frac{\partial h_i}{\partial u^l}=0,\cr
&\frac{\partial}{\partial u^l}\left(\frac{1}{h_l}\frac{\partial h_i}{\partial u^l}\right)
+\frac{\partial}{\partial u^i}\left(\frac{1}{h_i}\frac{\partial h_l}{\partial u^i}\right)
+\sum_{k\neq i,l}^n\,\frac{1}{h_k^2}\frac{\partial h_i}{\partial u^k}\frac{\partial h_l}{\partial u^k}=0
\end{aligned}
\end{equation}
One can show that system (\ref{LameEquations}) is involutive in the sense of the Cartan-K\"ahler theorem [\ref{BC3G},\ref{IveyLandsberg03}] and the solution depends upon $n(n-1)/2$ functions, each of 2 variables.
We will now make the connection between the $2+1$-dimensional {\sf 3WRI} equations, the Lam\'e equations and {\sf DMZ} system (\ref{KTsystem1}) more precise by formulating the following result that is essentially contained in [\ref{Darboux1896}, Chapter 12].
\begin{thm}[Darboux]\label{geometricBackgroundThm}
Suppose the linear PDE system
\begin{equation}\label{3WRI_LP}
\frac{\partial^2u}{\partial x_i\partial x_j}-
\Gamma_{j\,i}(x)\frac{\partial u}{\partial x_i}-\Gamma_{ij}(x)\frac{\partial u}{\partial x_j}=0,\ \ \ \ 1\leq i<j\leq 3
\end{equation}
for functions $u:\mathbb{R}^3\to\mathbb{R}$, where $\Gamma_{ij},\Gamma_{j\,i}$ are functions of $x=(x_1,x_2,x_3)$ is involutive. Then
\begin{enumerate}
\item[{\rm 1.}] There are real-valued functions $h_i(x)$, $1\leq i\leq 3$, such that
$$
\Gamma_{ij}=\frac{1}{h_j}\frac{\partial h_j}{\partial x_i}
$$
\item[{\rm 2.}] If $u$ is a solution of (\ref{3WRI_LP}) then the functions
\begin{equation}\label{3WRI_def}
\psi_i=\frac{1}{h_i}\frac{\partial u}{\partial x_i},\ \ \ \
\ A_{ij}=\frac{1}{h_i}\frac{\partial h_j}{\partial x_i}
\end{equation}
satisfy the 2+1-dimensional 3-wave resonant interaction system and its linear problem
\begin{equation}\label{3WRI}
\frac{\partial A_{j\,k}}{\partial x_i}=A_{j\,i}A_{i\,k}\ \ \ \ \
\frac{\partial \psi_j}{\partial x_i}=A_{j\,i}\psi_i,
\end{equation}
respectively
\item[{\rm 3.}] The functions $h_1,h_2,h_3$ satisfy the nonlinear system
\begin{equation}\label{halfLame}
\frac{\partial^2 h_i}{\partial x_j\partial x_k}-
\frac{1}{h_j}\frac{\partial h_j}{\partial x_k}\frac{\partial h_i}{\partial x_j}-
\frac{1}{h_k}\frac{\partial h_k}{\partial x_k}\frac{\partial h_i}{\partial x_j}=0,
\ \ \ (i,j,k)\in\ \text{\rm perm}\,(1,2,3).
\end{equation}
\end{enumerate}
\end{thm}
\proof{ The integrability conditions for linear system (\ref{3WRI_LP}) are
\begin{equation}\label{KT_0Compatibility}
\frac{\partial\Gamma_{ij}}{\partial x_k}-\Gamma_{ik}\Gamma_{kj}-\Gamma_{ki}\Gamma_{ij}+\Gamma_{ij}\Gamma_{kj}=
\frac{\partial \Gamma_{ij}}{\partial x_k}-\frac{\partial \Gamma_{kj}}{\partial x_i}=0,\ \ (i,j,k)\in\ \text{\rm perm}\,(1,2,3)
\end{equation}
From the second set of equations in (\ref{KT_0Compatibility}) we deduce that there are real valued functions functions $h_j$ on $\mathbb{R}^3$ such that
$$
\Gamma_{ij}=\frac{1}{h_j}\frac{\partial h_j}{\partial x_i}
$$
Next, let $u$ be any solution of (\ref{3WRI_LP}) and define functions $\psi_i$ and $A_{ij}$ as in (\ref{3WRI_def}). Then we have
$$
\begin{aligned}
\frac{{\partial} \psi_j}{{\partial} x_i}-A_{ji}\psi_i=&
\frac{{\partial}}{{\partial} x_i}\left(\frac{1}{h_j}\frac{{\partial} u}{{\partial} x_j}\right) -\frac{1}{h_j}\frac{{\partial} h_i}{{\partial} x_j}\frac{1}{h_i}
\frac{{\partial} u}{{\partial} x_i}\cr
=&\frac{1}{h_j}\left( \frac{{\partial}^2 u}{{\partial} x_i{\partial} x_j}
-\frac{1}{h_i}\frac{{\partial} h_i}{{\partial} x_j}\frac{{\partial} u}{{\partial} x_i}
-\frac{1}{h_j}\frac{{\partial} h_j }{{\partial} x_i}\frac{{\partial} u}{{\partial} x_j} \right)\cr
=&\frac{1}{h_j}\left(\frac{{\partial}^2 u}{{\partial} x_i{\partial} x_j}
-\Gamma_{j\,i}\frac{{\partial} u}{{\partial} x_i}-\Gamma_{ij}\frac{{\partial} u}{{\partial} x_j}\right)=0.
\end{aligned}
$$
It is then easily verified that the equations $(\ref{3WRI})_2$ have the 2+1-dimensional 3-wave resonant interaction equations
$(\ref{3WRI})_1$ as integrability conditions. This proves assertion 2 of the theorem. Finally, substituting the form of $A_{ij}$ from (\ref{3WRI_def}) into $(\ref{3WRI})_1$ gives (\ref{halfLame}).}\hfill \qed
\vskip 5 pt
Notice that equations (\ref{halfLame}) form part of the Lam\'e equations for triply orthogonal coordinate
systems, namely, $(\ref{LameEquations})_2$. When the $h_i$ satisfy only (\ref{halfLame}) rather than system (\ref{LameEquations}), they determine a triply orthogonal coordinate system in a Riemannian manifold $M$ which is not flat. However, it can be shown that the metric of $M$
$$
g=\sum_{i=1}^n h_i^2\,dx_i^2
$$
in the coordinates $x_i$ that arises from any solution of (\ref{halfLame}) has diagonal Ricci tensor. Moreover, the coefficients $\Gamma_{ij},\, \Gamma_{ji}$ in {\sf DMZ} system (\ref{3WRI_LP}) agree with the Levi-Civita connection coefficients $\Gamma^i_{ij},\, \Gamma^j_{ij}$, for $g$, respectively.
Zakharov [\ref{Zakharov98}] calls such Riemannian manifolds {\it spaces of diagonal curvature} and compares them with systems of hydrodynamic type which are integrable in a sense to be described but which do not possess a Hamiltonian structure. These are the so called {\it semi-Hamiltonian systems} about which we shall say more below.
Note also that equations (\ref{3WRI})
contain the 2+1-dimensional 3-wave resonant interaction equations, $(\ref{3WRI})_1$, together with it's standard linear problem $(\ref{3WRI})_2$. In other words, system $(\ref{3WRI})_1$ expresses the integrability conditions of $(\ref{3WRI})_2$, in the linear case (\ref{3WRI_LP}).
From this theorem we see that the linear system (\ref{3WRI_LP}) is the key that unifies the 2+1-dimensional {\sf 3WRI} system and certain triply orthogonal coordinate systems {\it provided (\ref{3WRI_LP}) is involutive}. An important goal of this paper is to derive feasible means of constructing linear {\sf DMZ} systems (\ref{3WRI_LP}) without being required to directly solve the nonlinear integrability conditions (\ref{KT_0Compatibility}) for the coefficients $\Gamma_{ij},\Gamma_{j\,i}$ in (\ref{3WRI_LP}). In fact, we examine the {\it geometric properties} of generalised {\sf DMZ} systems (\ref{GDMZabstract}) in order to characterise and construct them, thereby constructing new solutions for the 2+1-dimensional {\sf 3WRI} system.
As a concrete example of a {\sf DMZ} system, consider the triply orthogonal system defined by {\it oblate spheroidal coordinates}
$$
u=\cosh\,x\,\,\cos\,y\,\,\cos\,z,\ \ \ v=\cosh\,x\,\,\cos\,y\,\,\sin\,z,\ \ \
w=\sinh\,x\,\,\sin\,y.
$$
The corresponding flat metric is
$$
g=(\cosh^2x-\cos^2y)\,(dx^2+dy^2)+\cosh^2x\ \cos^2y\ dz^2
$$
and the relevant linear system corresponding to (\ref{3WRI_LP}) is
$$
\frac{{\partial}^2 u }{{\partial} x_i{\partial} x_j}-\Gamma^i_{ij}\frac{{\partial} u}{{\partial} x_i}
-\Gamma^j_{ij}\frac{{\partial} u}{{\partial} x_j}=0.
$$
Computing Christoffel coefficients delivers the linear system
\begin{equation}\label{oblateSpheroidal}
\begin{aligned}
&\frac{{\partial}^2 u }{{\partial} x{\partial} y}-\left(\frac{\cos\,y\ \sin\,y}{\cosh^2x-\cos^2y}\right)\frac{{\partial} u}{{\partial} x}
-\left(\frac{\cosh\,x\ \sinh\,x}{\cosh^2x-\cos^2y}\right)
\frac{{\partial} u}{{\partial} y}=0,\cr
&\frac{{\partial}^2 u }{{\partial} x{\partial} z}-\tanh\,x\frac{{\partial} u}{{\partial} z}=0,
\ \ \frac{{\partial}^2 u }{{\partial} y{\partial} z}+\tan\,y\frac{{\partial} u}{{\partial} z}=0.
\end{aligned}
\end{equation}
System (\ref{oblateSpheroidal}) is involutive in the sense of the Cartan-K\"ahler theorem, as can be verified; it is therefore a {\sf DMZ} system. One way to see this is to restrict the contact system on jet space $J^2(\mathbb{R}^3,\mathbb{R})$
to the 10-dimensional submanifold $\mathcal{R}\subset J^2(\mathbb{R}^3,\mathbb{R})$ defined by (\ref{oblateSpheroidal}). This defines a rank 4 linear Pfaffian system $\theta^i$, $0\leq i\leq 3$ on $\mathcal{R}$ with independence forms $\omega^1=dx,\omega^2=dy,\omega^3=dz$. These 7 differential 1-forms can be extended to a coframe on $\mathcal{R}$ by $\pi_1,\pi_2,\pi_3$ satisfying the structure equations
\begin{equation}\label{mainStructure}
\begin{aligned}
&d\theta^0\equiv 0,\cr
&d\theta^1\equiv \pi_1\wedge\omega^1,\cr
&d\theta^2\equiv \pi_2\wedge\omega^2,\cr
&d\theta^3\equiv \pi_3\wedge\omega^3,
\end{aligned}\mod\{\theta^0,\theta^1,\theta^2,\theta^3\}.
\end{equation}
This verifies that there is no essential torsion, that is, no integrability conditions, and that the tableau [\ref{BC3G},\,\ref{IveyLandsberg03}] satisfies Cartan's involutivity test. Accordingly, by Theorem \ref{geometricBackgroundThm}, the functions $A_{ij}$ defined by
(\ref{3WRI_def}) satisfy the 2+1-dimensional {\sf 3WRI} system. However, there are many more {\sf DMZ} systems (\ref{3WRI_LP}) than those provided by triply orthogonal coordinate systems and in this paper we will show how to construct many such systems. In consequence of Theorem 1.1, we construct orthogonal coordinate systems in spaces of diagonal curvature.
An integrable system of partial differential equations closely related to the {\sf 3WRI} system is the 2+1-dimensional {\it modified three-wave resonant interaction system}, which we denote {\sf m3WRI}, discovered by V. Mangazeev and S. Sergeev [\ref{MangazeevSergeev01}, \ref{Sergeev07}]. Like the {\sf 3WRI} system, the {\sf m3WRI} system has a Hamiltonian structure and plays a role in discrete differential geometry and quantum field theory [\ref{Sergeev07}]; it has the form
\begin{equation}\label{m3WRI}
\frac{{\partial} \Gamma_{jk}}{{\partial} x_i}=(\Gamma_{ij}-\Gamma_{ik})(\Gamma_{jk}-\Gamma_{ji}),\ \ \ (i,j,k)\in\text{perm}\{1, 2, 3\}.
\end{equation}
As shown by Mangazeev and Sergeev, it arises as the integrability condition of a {\sf DMZ} system of the form
\begin{equation}\label{MangazeevSergeev_LP}
\left(\P {x_i}\P {x_j}-\Gamma_{ij}\P {x_j}-\Gamma_{ji}\P {x_i}+
\Gamma_{ij}\Gamma_{ji}\right)\phi=0,\ \ \ 1\leq i<j\leq 3.
\end{equation}
We demonstrate in this paper how our general construction gives rise to new solutions of the {\sf m3WRI} system by constructing {\sf DMZ} systems of the form (\ref{MangazeevSergeev_LP}).
The problem of orthogonal coordinates and the 2+1-dimensional {\sf 3WRI} system are, in turn, linked to certain nonlinear partial differential equations of mathematical physics that have attracted a good deal of interest in recent years. To explain this let $\boldsymbol{u}=(u^1,u^2,\ldots,u^n)$ be a vector valued function of space variable $x$ and time $t$ and consider PDE of the form
\begin{equation}\label{quasilinear}
\frac{\partial u^i}{\partial t}-\sum_{k=1}^n\,v^i_k(\boldsymbol{u})\frac{\partial u^k}{\partial x}=0.
\end{equation}
Dubrovin and Novikov [\ref{DN}] showed that
if there is a flat Levi-Civita connection $\nabla$ and function $h$ on phase space such that
$$
v^i_j=\nabla^i\nabla_j\,h,
$$
then (\ref{quasilinear}) has a Hamiltonian structure. Subsequently, S. Tsarev showed that among non-Hamiltonian systems (\ref{quasilinear}) there is a significant subset of them which possess remarkable integrability properties.
To describe these we restrict ourselves to systems (\ref{quasilinear}) that are expressible in distinct Riemann invariants; these are the {\it strongly hyperbolic} systems. For strongly hyperbolic systems there is a change of dependent variables such that
$$
(v^j_k)=\text{diag}(w^1(\bar{u})\ \ w^2(\bar{u})\ \ \ldots\ \ w^n(\bar{u}))
$$
and for all $i,j$, $w^i(\bar{u})\neq w^j(\bar{u})$.
Attention focusses on those strongly hyperbolic systems that satisfy the additional constraint
\begin{equation}\label{diagonalCondition}
\partial_k\left(\frac{\partial_j w^i}{w^j-w^i}\right)=\partial_j\left(\frac{\partial_k w^i}{w^k-w^i}\right).
\end{equation}
Even though strongly hyperbolic systems (\ref{quasilinear}) whose ``diagonal velocities" satisfy (\ref{diagonalCondition}) are not, in general Hamiltonian, they nevertheless have significant integrability properties; for this reason they are called {\it semi-Hamiltonian}. For such systems Tsarev [\ref{Tsa}] has established a number of notable properties (see section 5 of this paper). The main point to note at this stage is that any {\sf DMZ} system (\ref{3WRI_LP}) gives rise to a semi-Hamiltonian PDE system in (1+1)-dimensions (\ref{quasilinear}). We demonstrate in section 5 how the geometric properties of {\sf DMZ} systems derived in sections 2 and 3, lead to new examples of such systems.
Finally we mention a link between the {\sf DMZ} systems (\ref{KTsystem1}) and a class of submanifolds whose study was initiated by Cartan [\ref{Cartan19}] and then subsequently developed by Chern [\ref{Chern44}] and much more recently by Kamran and Tenenblat in their generalisation of Laplace transformations to dimensions higher the two [\ref{KT96}, \ref{KT98}, \ref{KT2000}]. Below we schematically summarise the central role that {\sf DMZ} systems (\ref{KTsystem1}) play in the above mentioned fields of submanifold geometry and integrable systems.
\vskip 5 pt
\begin{equation}
\begindc{\commdiag}[40]
\obj(0,0){DMZ}
\obj(1,0){E}
\obj(-1,0){W}
\obj(0,-1){S}
\obj(0,1){N}
\mor{DMZ}{N}{}
\mor{DMZ}{S}{}
\mor{DMZ}{E}{}
\mor{DMZ}{W}{}
\enddc
\end{equation}
where
\begin{enumerate}
\item \underline{DMZ} denotes the Darboux-Manakov-Zakharov linear systems (\ref{KTsystem1}) in the case $n=3$; [\ref{ZakharovManakov85}, \ref{Zakharov98}].
\item \underline{N} denotes Lam\'e Equations for Triply Orthogonal Coordinate Systems; [\ref{Zakharov98}]
\item \underline{S} denotes Semi-Hamiltonian Systems of Hydrodynamic Type; [\ref{Tsa}], [\ref{Se}]
\item \underline{E} denotes the 2+1-dimensional Three-Wave Resonant Interaction Systems; [\ref{Kaup80},\ref{Sergeev07}]
\item \underline{W} denotes Cartan Sub-manifolds and Higher-Dimensional Laplace Transformations; [\ref{KT96}]
\end{enumerate}
\vskip 5 pt
We will have opportunity, in sections 4 and 5 of this paper, to discuss and illustrate the links \underline{S} and \underline{E} in more detail, using the theory developed in sections 3 and 4. For the present, it suffices to emphasise the importance of finding a means of characterising and constructing {\sf GDMZ} systems (\ref{GDMZ}) since it will enable us to significantly affect a number of problems in geometry, integrable systems and mathematical physics. We turn to this problem next.
\section{Intrinsic geometry of {\sf GDMZ} partial differential equations}
\label{sec:1}
In this section we derive a method for geometrically characterising involutive overdetermined systems of partial differential equations in 1 dependent variable $u$ and $n$ independent variables $x_1,\ldots,x_n$ of the form
\begin{equation}\label{mainEquations}
\frac{\partial^2 u}{\partial x_i\partial x_j}=f_{ij}\Big(x_k,u,\frac{\partial u}{\partial x_l}\Big),\ \ \ 1\leq i<j\leq n,\ \ k,l\in\{1,\ldots,n\}.
\end{equation}
In particular, we will be interested in {\it explicitly} constructing {\it linear} systems in the class (\ref{mainEquations}).
\begin{defn}
In case system (\ref{mainEquations}) is involutive, then we shall refer to it as a {\it generalised Darboux-Manakov-Zakharov} ({\sf GDMZ}) PDE system. If the $f_{ij}$ in (\ref{mainEquations}) are such that it is a {\rm linear} involutive PDE system then we shall refer to it as a {\it linear Darboux-Manakov-Zakharov} system or simply as a {\it Darboux-Manakov-Zakharov} ({\sf DMZ}) system.
\end{defn}
For $n=2$ system (\ref{mainEquations}) comprises just one equation and is automatically involutive. For $n>2$ the system is overdetermined; it's involutive if and only if integrability conditions are satisfied. The contribution of this section is to derive a geometric characterisation of systems (\ref{mainEquations}) for which the integrability conditions are satisfied ``automatically". That is, a geometric characterisation of generalised Darboux-Manakov-Zakharov systems. A significant by-product of this result in the linear case, is the construction of explicit solutions of the 3-wave resonant interaction equations and the modified 3-wave resonant interaction equations. This will be discussed in section 4. Another is the construction of new semi-Hamiltonian systems of hydrodynamic type, explained in section 5.
The abovementioned geometric characterisation of {\sf GDMZ} systems relies on the existence of certain structured manifolds and special coordinate systems on them. For any $n\geq 2$, let $M_{3n+1}$ be a smooth manifold of dimension $3n+1$ equiped with a rank $2n$ distribution $H$ which decomposes as a direct sum of $n$ rank 2 distributions $H_i$
\begin{equation}\label{nHyp}
H=\bigoplus_{i=1}^nH_i
\end{equation}
satisfying the structure equations
\begin{equation}\label{genHypStr}
[H_i,H_j]\equiv 0,\ [H_i,H_i]\equiv Z_i\mod H,\ \forall\,i\ \text{and}\ j\neq i
\end{equation}
where $Z_1\wedge\cdots\wedge Z_n\neq 0$. Establishing that such distributions exist in plentiful supply, can be explicitly constructed and have interesting properties and applications are aims of this paper.
\begin{defn}
Let $n\geq 2$ be an integer. A pair $(M_{3n+1},H)$ consisting of a manifold $M$ of dimension $3n+1$ and a distribution $H$ of the form (\ref{nHyp}) satisfying structure equations (\ref{genHypStr}) and derived type
$$
\mathfrak{d}(H)=[[2n,0],[3n,n],[3n+1,3n+1]]
$$
will be called an $n$-{\it hyperbolic manifold}, while $H$ will be called an $n$-{\it hyperbolic structure} or $n$-{\it hyperbolic distribution}.\hfill $\clubsuit$
\end{defn}
Note that the $j^{\,\text{th}}$, ordered, 2-component element of the list of lists $\mathfrak{d}(H)$ records the dimension $\dim\,H^{(j)}$ of the $j^{\,\text{th}}$ derived distribution $H^{(j)}$ of $H$ and the dimension $\dim\,\ch\,H^{(j)}$ of its Cauchy distribution
$$
[\dim\,H^{(j)},\dim\,\ch\,H^{(j)}].
$$
Recall that for any smooth distribution $\mathcal{V}\subseteq TM$ on a smooth manifold $M$, the {\it Cauchy} or {\it characteristic} distribution, $\ch\mathcal{V}$ is defined by
$$
\ch\mathcal{V}=\{X\in\mathcal{V}~|~[X,\mathcal{V}]\subseteq\mathcal{V}\}.
$$
The derived distribution $\mathcal{V}^{(1)}$ of $\mathcal{V}$ is defined by
$$
\mathcal{V}^{(1)}={\partial}\mathcal{V}=\mathcal{V}+[\mathcal{V},\mathcal{V}]
$$
and the higher derived distributions are defined by iteration: $\mathcal{V}^{(j)}={\partial}^j\mathcal{V}$. It is easy to prove, in the case $\mathcal{V}$ and $\ch\mathcal{V}$ have constant rank on $M$ that $\ch\mathcal{V}$ is Frobenius integrable. See [\ref{Vassiliou06a},\, \ref{Vassiliou06b}], for further details. Throughout this paper, constant rank assumptions are made for all vector field distributions and codistributions.
\begin{defn}
Let $n\geq 2$ be an integer. We say that an $n$-hyperbolic manifold $(M_{3n+1},H)$ is {\it Daboux integrable} if there is a coordinate system
\begin{equation}\label{adaptedCoords}
\text{\sf C}: x_1,x_2,\ldots,x_n,y_0,y_1,\ldots,y_n,u_1,\ldots,u_n
\end{equation}
on $M$, {\it adapted} to its $n$-hyperbolic structure $H$ in the sense that
\begin{enumerate}
\item[1).]
The Cauchy distribution ${\ch H^{(1)}}$ of the derived distribution $H^{(1)}$ satisfies
$$
\dim \ch H^{(1)}=n,\ \ \ch H^{(1)}\subset H
$$
and
$x_1,x_2,\ldots,x_n,y_0,y_1,\ldots,y_n$ are the invariants of $\ch H^{(1)}$, while $u_1,\ldots,u_n$ extend these to a coordinate system on $M$ from which it follows that
$$
\ch H^{(1)}=\{\P {u_1},\ldots,\P {u_n}\}
$$
\item[2).] For each $i$, there is a basis for $H$ of the form
\begin{equation}\label{adaptedBasis}
X^1_i=\P {x_i}+P_i,\ \ X^2_i=\P {u_i},\ \ 1\leq i\leq n
\end{equation}
where
$$
P_i=\sum_{a=0}^n\rho^a_i(x,y,u)\P {y_a}\mod\{\P {u_1},\ldots,\P {u_n}\}.
$$
and $\{X^1_i,\P {u_i}\}$ is a basis for $H_i$, for some real-valued functions $\rho^a_i$.
\end{enumerate}
The vector fields (\ref{adaptedBasis}) of a Darboux integrable $n$-hyperbolic manifold are said to form an {\it adapted basis} for $H$.\hfill $\clubsuit$
\end{defn}
That Darboux integrable $n$-hyperbolic manifolds exist and can be constructed will be established in section 3.
Our aim now is to show that a Darboux integrable $n$-hyperbolic manifold determines an immersion $\Phi$ of the $M$ into $J^2(\mathbb{R}^n,\mathbb{R})$ whose image is a {\sf GDMZ} system, (\ref{mainEquations}).
From now on we drop the superscript 1 on $X^1_i$ and rename this vector field $X_i$.
\begin{lem}\label{AdaptedCoords1}
Let $(M_{3n+1},H)$ be an Darboux integrable $n$-hyperbolic manifold with adapted coordinates (\ref{adaptedCoords}) and adapted basis $\{X_i,\P {u_i}\}$ for $H_i$. Then for all $i\neq j$ with $i,j\in \{1,2,\ldots,n\}$ we have
$$
[X_i,X_j]\equiv 0,\ [X_i,\P {u_j}]\equiv 0\mod\{\P {u_1}, \P {u_2},\ldots,\P {u_n}\}.
$$
\end{lem}
\vskip 5 pt
\proof{ By the hyperbolic structure equations (\ref{genHypStr}) we have
$$
[X_i,X_j]=\sum_{k=1}^nA^k_{ij}X_k+\sum_{k=1}^nB^k_{ij}\P {u_k}
$$
for some functions $A,B$. However, the assumed local form (\ref{adaptedBasis}) of the adapted basis for $H$ shows that $A^k_{ij}=0$ for all $i, j, k$. A similar argument establishes the second congruence.}\hfill\qed
\vskip 5 pt
\begin{lem}\label{AdaptedCoords2}
Let $(M,H)$ be a Darboux integrable $n$-hyperbolic manifold with adapted coordinates (\ref{adaptedCoords}) and adapted basis $\{X_i,\P {u_i}\}$ for $H_i$. Then
$$
\mathcal{A}=\Big\{\P {u_1},\P {u_2},\ldots,\P {u_n},[\P {u_1},X_1],[\P {u_2},X_2],\ldots,[\P {u_n},X_n]\Big\}\subset H^{(1)}
$$
is a rank $2n$, integrable distrubution.
\end{lem}
\vskip 5 pt
\proof{
Since $\dim H^{(1)}=3n$ and the $X_j$ pairwise commute modulo $\{\P {u_1}\ldots,\P {u_n}\}$ by the previous Lemma, it follows that $\dim\mathcal{A}=2n$. By the Jacobi identity and again by the previous Lemma, we deduce
for all $i\neq j$ that $[\P {u_j},[\P {u_i},X_i]]\equiv 0\mod \{\P {u_1}\ldots,\P {u_n}\}$. Similarly, we deduce that $[X_j,[\P {u_i},X_i]]\equiv 0\mod \{\P {u_1}\ldots,\P {u_n}\}$. Next set $\xi_i=[\P {u_i},X_i]$, apply the Jacobi identity to get $[\xi_i,[\P {u_j},X_j]]+[X_j,[\xi_i,\P {u_j}]]+[\P {u_j},[X_j,\xi_i]]=0$. From this and using our previous results we deduce that
$$
[[\P {u_i},X_i],[\P {u_j},X_j]]\equiv 0\mod\mathcal{A}.
$$
Hence $\mathcal{A}$ is rank $2n$ and integrable.}\hfill\qed
\begin{cor}\label{AdaptedCoords3}
Let $(M,H)$ be an Darboux integrable $n$-hyperbolic manifold. The invariants of the distribution $\mathcal{A}\subset H^{(1)}$ are spanned by $x_1,\ldots,x_n,p$ for some function $p$ on $M$.
\end{cor}
\begin{thm}\label{KTdistributions}
Let $(M,H)$ be a Darboux integrable $n$-hyperbolic manifold. Let functions $x_1,\ldots,x_n,p$ on $M$ span the invariants of $\mathcal{A}$ and for $1\leq i<j\leq n$, define functions $p_i,p_{ii},f_{ij}$ by
$$
p_i=X_ip,\ p_{ii}=X_ip_i,f_{ij}=X_jp_i.
$$
Then the map $\varphi$ given by
$$
\varphi(x,y,u)=(x_1,\ldots,x_n,p,p_1,\ldots,p_n,p_{11},\ldots,p_{nn})
$$
is a local diffeomorphism and defines an immersion
$$
\Phi : M\hookrightarrow J^2(\mathbb{R}^3,\mathbb{R})
$$
whose image is a {\sf GDMZ} system
$$
\frac{\partial^2 u}{\partial x_i\partial x_j}=f_{ij}\Big(x_k,u,\frac{\partial u}{\partial x_l}\Big),\ \ \
1\leq i<j\leq n,\ \ k,l\in\{1,\ldots,n\}.
$$
\end{thm}
\vskip 5 pt
\proof First we show that the functions
$$
x_1,\ldots,x_n,\, p,\, p_1,\ldots,p_n,\, p_{11},\ldots,p_{nn}
$$
as defined in the statement of the theorem form a local coordinate system on $M$. Begin by extending the independent invariants $x_1,\ldots,x_n,p$ of $\mathcal{A}$ by $y_1,\ldots,y_n,\allowbreak u_1,\ldots,u_n$, where the $y_l$ extend $x_1,x_2\ldots,x_n,p$ to be a complete set of invariants of $\ch H^{(1)}$. In these coordinates, the local form of each $X_i$ is
$$
X_i=\P {x_i}+\rho^0_i\P p+\sum_{l=1}^n\rho^l_i\P {y_l}\mod \P {u_j}
$$
and we therefore have
$$
[\P {u_i},X_i]=\frac{\partial \rho^0_i}{\partial u_i}\P p+
\sum_{l=1}^n\frac{\partial \rho^l_i}{\partial u_i}\P {y_l}\mod \{\P {u_j}\},\ \ 1\leq i\leq n.
$$
But the left-hand-side is an element of $\mathcal{A}$ and hence
$$
0=[\P {u_i},X_i]p=\frac{\partial \rho^0_i}{\partial u_i}
$$
and thus
$$
[\P {u_i},X_i]=
\sum_{l=1}^n\frac{\partial \rho^l_i}{\partial u_i}\P {y_l}\mod \{\P {u_j}\},\ \ 1\leq i\leq n.
$$
It follows that the matrix
$$
J=\left(\frac{\partial\rho^j_i}{\partial u_i}\right)_{i,j=1}^n
$$
is nonsingular. We deduce that
$$
H^{(1)}=\Big\{\P {x_1}+\rho^0_1\P p,\ldots,\P {x_n}+\rho^0_n\P p,\P {y_1}, \ldots, \P {y_n},\P {u_1}, \ldots, \P {u_n}\Big\}
$$
and that the quotient of $H^{(1)}$ by $\ch H^{(1)}$
$$
K:=H^{(1)}/\ch H^{(1)}=\Big\{\P {x_1}+\rho^0_1\P p,\ldots,\P {x_n}+\rho^0_n\P p,\P {y_1}, \ldots, \P {y_n}\Big\}
$$
is a rank $2n$ distribution defined on the $2n+1$-dimensional quotient manifold\newline
$M/\ch H^{(1)}$, satisfying $\dim K^{(1)}=2n+1$
and $\ch{K}=\{0\}$. It follows from Pfaff-Darboux theorem that new coordinates
$$
p_l=\rho^0_l
$$
can be taken in place of the $y_l$, expressing $K$ in the canonical form
$$
K=\Big\{\P {x_1}+p_1\P p,\ldots,\P {x_n}+p_n\P p,\P {p_1}, \ldots, \P {p_n}\Big\}.
$$
This means there are functions $\widetilde{\rho}^{\,l}_i$ of the independent coordinates on $M$ given by $x,p,p_l,u_j$ such that
$$
X_i=\P {x_i}+p_i\P p+\sum_{l=1}^n\widetilde{\rho}^{\,l}_i\P {p_l}\mod\{\P {u_j}\}.
$$
The functions $p_{ij}$ are defined by
\begin{equation}\label{derivatives2}
p_{ij}=X_jp_i=\widetilde{\rho}^{\,j}_i.
\end{equation}
and since the $X_i$ pairwise commute modulo $\ch\,H^{(1)}$, we have
\begin{equation}\label{lower_symmetric}
p_{ij}=\widetilde{\rho}^{\,j}_i=\widetilde{\rho}^{\,i}_j=p_{ji}.
\end{equation}
To see this we calculate
$$
p_{ij}-p_{j\,i}=X_ip_j-X_jp_i=[X_i,X_j]p=0
$$
since $[X_i,X_j]\in\{\P {u_1},\, \P {u_2}\ldots,\P {u_n}\}$, for all $i,j$, by Lemma 2.1.
Furthermore, we have
\begin{equation}\label{1stDerivatives}
\begin{aligned}
\frac{\partial p_k}{\partial u_l}&=\frac{\partial}{\partial u_l}\left(X_kp\right)\cr
&=\frac{\partial}{\partial u_l}\left(X_kp\right)-X_k\left(\frac{\partial p}{\partial u_l}\right),\ \text{by}\ \text{Corollary}\ \ref{AdaptedCoords3}\cr
&=[\P {u_l},X_k]p=0
\end{aligned}
\end{equation}
because $[\P {u_l},X_k]\in\mathcal{A}$ for all $k,l$ by Lemmas \ref{AdaptedCoords2} and \ref{AdaptedCoords3}. This means that
$$
[\P {u_l},X_k]=\frac{{\partial} p_k}{{\partial} u_l}\P p+\sum_{i=1}^n\frac{{\partial}\widetilde{\rho}^{\,i}_k}{{\partial} u_l}\P {p_i}=
\sum_{i=1}^n\frac{{\partial}\widetilde{\rho}^{\,i}_k}{{\partial} u_l}\P {p_i}.
$$
But for $l\neq k$, we have $[\P {u_l}, X_k]\in\{\P {u_1},\ldots,\P {u_n}\}$ and hence
\begin{equation}\label{nondiagonal}
\frac{\partial \widetilde{\rho}^{\,i}_k}{\partial u_l}=0,\ \ \ \ \forall\ \ i\ \text{and}\ l\neq k.
\end{equation}
Equations (\ref{derivatives2}), (\ref{lower_symmetric}) and (\ref{nondiagonal}) imply that the matrix $J$ has the form
$$
J=\text{diag}\,\left(\begin{matrix} \frac{\partial\rho^1_1}{\partial u_1} & \frac{\partial\rho^2_2}{\partial u_2} &
\ldots & \frac{\partial\rho^n_n}{\partial u_n}\end{matrix}\right)
$$
and since $J$ is non-singular we deduce that
\begin{equation}\label{indepDerivatives2}
\prod_{i=1}^n\frac{\partial\rho^i_i}{\partial u_i}\neq 0
\end{equation}
The condition (\ref{indepDerivatives2}) and our argument till now proves the claim that the functions defined by the theorem statement are functionally independent.
Next, we prove that the local diffeomorphism
$$
\varphi: (x,y,u)\mapsto (x_1,\ldots,x_n,p,p_1,\ldots,p_n,p_{11},\ldots,p_{nn})
$$
induces an immersion of $M$ into $J^2(\mathbb{R}^3,\mathbb{R})$ whose image is a {\sf GDMZ} system. We will prove that $\varphi$ identifies $H$ with the canonical distribution on
$J^2(\mathbb{R}^3,\mathbb{R})$ spanned by
$$
\Big\{D_1^f,\ldots,D_n^f,\P {p_{11}},\ldots,\P {p_{nn}}\Big\}
$$
where
$$
D_i^f=\P {x_i}+p_i\P p+\sum_{j=1}^np_{ij}\P {p_j},
$$
and where for all $i\neq j$
$$
p_{ij}=p_{j\,i}=f_{ij}.
$$
More precisely, we prove that there is a function $\boldsymbol{g} : M\to GL(2n)$ such that
\begin{equation}
\varphi_*\mathbb{X}=\boldsymbol{g}\;\mathfrak{D}
\end{equation}
where
\begin{equation}\label{equivalence}
\begin{aligned}
&\mathbb{X}=\left(\begin{matrix} X_1 & X_2 &\cdot&\cdot & X_n &\P {u_1} &\P {u_2}
&\cdot &\cdot &\P {u_n}\end{matrix}\right)^T,\cr
&\mathfrak{D}=\left(\begin{matrix} D^f_1 & D^f_2 &\cdot &\cdot & D^f_n &\P {p_{11}} &
\P {p_{22}} &\cdot &\cdot &\P {p_{nn}}\end{matrix}\right)^T.
\end{aligned}
\end{equation}
In fact we will show that $\boldsymbol{g}$ has the form
\renewcommand{\arraystretch}{1.3}
\begin{equation}\label{G_structure}
\boldsymbol{g}=\left(\begin{matrix} I_n & \mathbf{g}^1_n\cr
\mathbf{0}_n &\mathbf{g}^2_n\end{matrix}\right)
\end{equation}
where $I_n,\mathbf{0}_n$ are the $n\times n$ identity and zero matrices respectively; $\mathbf{g}^a_n$ are certain $n\times n$ matrices determined by the equivalence $\varphi$. Recalling that the components of local diffeomorphism $\varphi$
are $x_i,p,p_j,p_{kk}$, then from (\ref{equivalence}) and (\ref{G_structure}) we obtain the equations
\begin{equation}\label{equivalencePDE}
\begin{aligned}
&p_i=X_ip,\ \ \P {u_i}p=0,\ 1\leq i\leq n\cr
&p_{ii}=X_ip_i,\ \ f_{ij}=X_jp_i,\ \ \ 1\leq i<j\leq n,\cr
&\P {u_k}p_l=0,\ \ \ \ k,l\in\{1,2,\ldots,n\}.
\end{aligned}
\end{equation}
for the components of $\varphi$. The remaining equations define the entries of $\mathbf{g}^1_n$ and $\mathbf{g}^2_n$.
Equations $(\ref{equivalencePDE})_1$ and $(\ref{equivalencePDE})_2$ are satisfied by definition of $p,\,p_i$ and $p_{ii}$. Equations $(\ref{equivalencePDE})_3$ are satisfied by virtue of (\ref{1stDerivatives}). Hence, we have shown that $\varphi$ maps the distribution $H$ to the canonical distribution for the overdetermined system
$$
\frac{\partial^2 u}{\partial x_i\partial x_j}=f_{ij}\Big(x_k,u,\frac{\partial u}{\partial x_l},
\frac{\partial^2 u}{\partial x_l^2}\Big),\ \ \
1\leq i<j\leq n,\ \ k,l\in\{1,\ldots,n\}.
$$
Next we show that the $f_{ij}$ do not depend upon the second order partial derivatives
$$
\frac{\partial^2 u}{\partial x_l^2}.
$$
That is, we show that $f_{ij}$ don't, in fact, depend upon $p_{ll}$ for any $l$. For suppose for some $i,j$,\ $i\neq j$ and some $l$, that
$$
\frac{\partial f_{ij}}{\partial p_{ll}}\neq 0.
$$
We calculate
$$
\begin{aligned}
\frac{\partial f_{ij}}{\partial u_l}&=
\sum_{m=1}^n\left(\frac{\partial f_{ij}}{\partial x_m}\frac{\partial x_m}{\partial u_l}+
\frac{\partial f_{ij}}{\partial p}\frac{\partial p}{\partial u_l}+
\frac{\partial f_{ij}}{\partial p_m}\frac{\partial p_m}{\partial u_l}+
\frac{\partial f_{ij}}{\partial p_{mm}}\frac{\partial p_{mm}}{\partial u_l}\right)\cr
&=\frac{\partial f_{ij}}{\partial p_{ll}}\frac{\partial p_{ll}}{\partial u_l}\ \ \ \ \text{(no summation over $l$})\cr
&\neq 0\ \ \ \ \ \ \text{by equation}\ \ (\ref{indepDerivatives2})
\end{aligned}
$$
However, this contradicts equation (\ref{nondiagonal}). Hence we have shown that $\varphi$ maps $H$ to the canonical distribution for the PDE system
\begin{equation}\label{MyEquations}
\frac{\partial^2 u}{\partial x_i\partial x_j}=f_{ij}\Big(x_k,u,\frac{\partial u}{\partial x_l}\Big),\ \ \
1\leq i<j\leq n,\ \ k,l\in\{1,\ldots,n\}.
\end{equation}
as we wanted.
Finally, we prove that the $n$-hyperbolic structure $H$, and hence equation (\ref{MyEquations}), is involutive with respect to the independence form
$dx_1\wedge dx_2\wedge\cdots\wedge dx_n$. In particular, the integrability conditions for (\ref{MyEquations}) are satisfied. Set $\Theta=H^\perp$. We will show that $\Theta$ is an involutive linear Pfaffian system with respect to independence form $\omega^1\wedge \omega^2\wedge\cdots\wedge\omega^n$, where $\omega^j=dx_j$. Recall the structure equations for $H$ have the form
\begin{equation}\label{StructureEqs1}
\begin{aligned}
&[X_i, \P {u_i}]\equiv Z_i\mod H,\ \ \ 1\leq i\leq n,\ \ \ \ \text{and},\cr
&[X_i, X_j]\equiv 0,\ \ \ [X_i, \P {u_j}]\equiv 0\mod \ch H^{(1)},\ \ i,j\in\{1,2,\ldots,n\},\ i\neq j.
\end{aligned}
\end{equation}
where $Z_1\wedge Z_2\cdots\wedge Z_n\neq 0$. Consequently, we can write
$$
H^{(1)}=\{X_1,\ldots,X_n,\P {u_1},\ldots,\P {u_n}\}\oplus\{Z_1,\ldots,Z_n\}
$$
which we can extend by $U$ to make a frame on $M$. We let
\begin{equation}
\omega^1,\omega^2,\ldots,\omega^n,\pi^1,\pi^2,\ldots,\pi^n,\theta^0,\theta^1,\ldots,\theta^n
\end{equation}
be the coframe dual to
$$
X_1,\ldots,X_n,\P {u_1},\ldots,\P {u_n}, U, Z_1,\ldots,Z_n.
$$
In particular $\{\theta^0,\theta^1,\ldots,\theta^n\}=H^\perp=\Theta$. If $X,Y$ are vector fields on $M$ then use of the identity
$$
d\theta^p(X,Y)=X\left(\theta^p\rfloor Y\right)-X\left(\theta^p\rfloor Y\right)-\theta^p\,\rfloor\,[X,Y]
$$
and structure equations (\ref{StructureEqs1}) permits one to establish the structure equations
\begin{equation}\label{formsStructureEqs}
\begin{aligned}
&d\theta^0\equiv 0,\cr
&d\theta^1\equiv \pi^1\wedge\omega^1,\cr
&d\theta^2\equiv \pi^2\wedge\omega^2,\cr
&\ \ \ \ \ \ \vdots\cr
&d\theta^n\equiv \pi^n\wedge\omega^n
\end{aligned}\mod\Theta
\end{equation}
Equations (\ref{formsStructureEqs}) prove that $\Theta$ has no integrability conditions (zero essential torsion) and an involutive tableau with one nonzero Cartan character, $s_1=n$.
\hfill\qed
\begin{prop}[Linear {\sf DMZ} systems]
Let
\begin{equation}\label{linearDMZfull}
\frac{{\partial}^2 u}{{\partial} x_i{\partial} x_j }-\sum_{k=1}^n\Gamma^k_{ij}(x)\frac{{\partial} u}{{\partial} x_k}+C_{ij}(x)u=0,\ \ \ 1\leq i<j\leq n,
\end{equation}
be a system of linear partial differential equations for one dependent variable $u$ in independent variables $x=(x_1,\ldots x_n)$, $n\geq 2$, where $\Gamma^k_{ij},\ C_{ij}$ are real vauled smooth functions, symmetric in their lower indices. Then system (\ref{linearDMZfull}) is involutive if and only if
$$
\Gamma^k_{i j}=0,\ \ \ \ \text{for distinct}\ i, j, k,
$$
and for each $(i,j,k)\in\text{\rm perm}_3(n)$,
\begin{equation}\label{DMZfullCompatibility}
\begin{aligned}
&\frac{{\partial}\Gamma^j_{ij}}{{\partial} x_k}-\Gamma^k_{ik}\Gamma^j_{kj}-\Gamma^i_{ki}\Gamma^j_{ij}+\Gamma^j_{ij}\Gamma^j_{kj}+C_{ik}=0,\cr
&\frac{{\partial} \Gamma^j_{ij}}{{\partial} x_k}-\frac{{\partial} \Gamma^j_{kj}}{{\partial} x_i}=0,\cr
&\frac{{\partial} C_{ij}}{{\partial} x_k}-\frac{{\partial} C_{ik}}{{\partial} x_j}+
C_{kj}(\Gamma^k_{ik}-\Gamma^j_{ij})+C_{ij}\Gamma^i_{ik}-C_{ik}\Gamma^i_{ij}=0.
\end{aligned}
\end{equation}
If (\ref{linearDMZfull}) is involutive then its solutions depend upon $n$ functions, each of one variable.
\end{prop}
\vskip 3 pt
\proof This is a calculation using Cartan's theory of linear Pfaffian systems, [\ref{BC3G}, \ref{IveyLandsberg03}], applied to (\ref{linearDMZfull}). The left hand sides of (\ref{DMZfullCompatibility}) are components of the torsion tensor. The top reduced Cartan character is $s_1=n$. \hfill\qed
\vskip 5 pt
The main point here is that system (\ref{DMZfullCompatibility}) is precisely the {\sf nWRI} system in the case $C_{ik}=0$, as explained in Theorem \ref{geometricBackgroundThm}, and it is the modified {\sf nWRI} system in case $C_{ik}=\Gamma^i_{ik}\Gamma^k_{ik}$. Hence, if we can explicitly construct Darboux integrable $n$-hyperbolic manifolds whose associated {\sf GDMZ} systems, as guaranteed by Theorem \ref{KTdistributions}, are {\it linear} then we will have constructed
\begin{enumerate}
\item[{\bf i)}] Solutions of the multi-dimensional {\sf nWRI} system if $C_{ik}=0$
\item[{\bf ii)}] Solutions of the modified {\sf nWRI} system if $C_{ik}=\Gamma^i_{ik}\Gamma^k_{ik}$
\item[{\bf iii)}] Semi-Hamiltonian systems of strongly hyperbolic conservation laws and their commuting flows if $C_{ik}=0$
\end{enumerate}
In sections 3 and 4, we will lay out theory that solves each of these construction problems. In sections 4 and 5, we illustrate these constructions with explicit applications, which show that the calculations require only the Frobenius theorem as well as quadrature, entailing integration which is very often easy to carry out. This procedure provides solutions for the full integrability conditions (\ref{DMZfullCompatibility}) in general and, in particular, leads to a very wide class of solutions of the multi-dimensional $n$-wave resonant interaction systems as well as the construction of new semi-Hamiltonian systems of hydrodynamic type and their commuting flows.
\section{Constructing a class of {\sf GDMZ} systems from $n$-hyperbolic manifolds}
In the previous section we proved, in Theorem \ref{KTdistributions}, that to explicitly construct a {\sf GDMZ} system it is sufficient to have constructed an explicit Darboux integrable $n$-hyperbolic manifold. In this section a method will be described for constructing a large class of these.
We hasten to point out that Darboux integrable $n$-hyperbolic manifolds are far more numerous than those which we construct in this section. However, in this paper we are primarily interested in {\it linear} {\sf DMZ} systems and for this purpose the construction described below is adequate. Having said this, nonlinear {\sf GDMZ} systems are not excluded from our construction.
Consider cartesian products of jet spaces of the form
\begin{equation}\label{jetProds}
\mathbb{J}=J^{k_1}(\mathbb{R},\mathbb{R})\times J^{k_2}(\mathbb{R},\mathbb{R})\times\cdots\times J^{k_r}(\mathbb{R},\mathbb{R})
\end{equation}
We will study certain actions of finite Lie groups upon such products and show that one can thereby construct an important class of Darboux integrable $n$-hyperbolic manifolds. The main idea can be summarised as follows. Let $\mathbb{J}_1,\mathbb{J}_2$ each be manifolds of the form (\ref{jetProds}). It is natural to consider the distributions on $\mathbb{J}_1$ and $\mathbb{J}_2$ formed by the direct sum of the contact distributions $\mathcal{C}^{k_l}$ of each jet space $J^{k_l}(\mathbb{R},\mathbb{R})$ in a cartesian product $\mathbb{J}$. Let us call these {\it multi-contact distributions} and denote them by symbols $\mathcal{C}(\kappa_1)$ and $\mathcal{C}(\kappa_2)$ on $\mathbb{J}_1$ and $\mathbb{J}_2$, respectively, where
\begin{equation}\label{contact}
\mathcal{C}(\kappa_a):=\mathcal{C}^{k^a_1}\oplus\mathcal{C}^{k^a_2}\oplus\cdots\oplus\mathcal{C}^{k^a_{n_a}},\ a=1,2
\end{equation}
and where the sequences of positive integers
$$
\kappa_a=\langle k^a_1,k^a_2,\ldots,k^a_{n_a}\rangle
$$
completely specify the distributions under consideration. Suppose that free, regular, diagonal actions $\mu_1:G\times \mathbb{J}_1\to \mathbb{J}_1$ and $\mu_2: G\times \mathbb{J}_2\to \mathbb{J}_2$ of a finite Lie group $G$ are symmetries of $(\mathbb{J}_1,\mathcal{C}(\kappa_1))$ and $(\mathbb{J}_2,\mathcal{C}(\kappa_2))$, respectively\footnote{In this paper we ignore those symmetries of $\mathcal{C}(\kappa_a)$ that permute its factors}. Hence, $\mu_1$ and $\mu_2$ restrict to be finite-dimensional contact transformation groups on each factor in $\mathbb{J}_1$ and $\mathbb{J}_2$, respectively. We will confine our study to such actions and show that the quotient of the product $(\mathbb{J}_1\times \mathbb{J}_2,\mathcal{C}(\kappa_1)\oplus\mathcal{C}(\kappa_2))$ by the diagonal action of $G$ determined by $\mu_1$ and $\mu_2$ determines a Darboux integrable $n$-hyperbolic manifold. For more information on symmmetry reduction of exterior differential systems see [\ref{Itskov01}], [\ref{AndersonFels05}], [\ref{AFV09}]. Because all finite dimensional contact transformation groups acting on $J^k(\mathbb{R},\mathbb{R})$ have been classified, [\ref{DoubrovKomrakov}], it is possible to contemplate the problem of classifying all Darboux integrable $n$-hyperbolic manifolds that arise by this proceedure and the PDE systems (\ref{mainEquations}) that they intrinsically determine.
We now describe this construction in detail.
\subsection{The actions on $\mathbb{J}_1$ and $\mathbb{J}_2$}
Express the canonical bases for the multi-contact systems $\mathcal{C}^{(k_i^1)}$ on $\mathbb{J}_1$ by
$$
\Big\{X_i=\P {\alpha_i}+x^i_1\P {x^i}+x^i_2\P {x^i_1}+\cdots+x^i_{k^1_i}\P {x^i_{k^1_i-1}},\ \P {x^i_{k^1_i}}\Big\},\ \ 1\leq i\leq n_1
$$
and $\mathcal{C}^{(k_i^2)}$ on $\mathbb{J}_2$ by
$$
\Big\{Y_j=\P {\beta_j}+y^j_1\P {y^j}+y^j_2\P {x^j_1}+\cdots+y^j_{k^2_j}\P {y^j_{k^2_j-1}},\ \P {y^j_{k^2_j}}\Big\},\ \ 1\leq j\leq n_2.
$$
Using the Fels-Olver method of moving frames [\ref{FO}], we will construct Lie group actions $\mu_1$ on $\mathbb{J}_1$ and $\mu_2$ on $\mathbb{J}_2$ such that
\begin{enumerate}
\item[A1)] Each action $\mu_1,\mu_2$ is free and regular and $\mu_1$ preserves $\alpha_i$ and $\mu_2$ preserves $\beta_j$ for all $i,j$
\item[A2)] For each $i,\ 1\leq i\leq n_1$ and each $j$,\ $1\leq j\leq n_2$, there is exactly one invariant $u_i$ of $\mu_1$ such that
$$
\frac{\partial u_i}{\partial x^i_{k^1_i}}\neq 0
$$
and exactly one invariant $v_j$ such that
$$
\frac{\partial v_j}{\partial y^j_{k^2_j}}\neq 0.
$$
\item[A3)] All other invariants of $\mu_a$,\ $a=1,2$,\ and their moving frame components have order lower than
$\text{min}\,\big(k^a_1,k^a_2,\ldots,k^a_{n_a}\big)$.
\item[A4)] Symmetry group $G$ is chosen so that $\dim\mathbb{J}_1+\dim\mathbb{J}_2-\dim G=3n+1$, where $n=n_1+n_2$.
\end{enumerate}
\begin{defn}
We will call an action of a finite Lie group acting diagonally on a cartesian product $\mathbb{J}$ defined by (\ref{jetProds}) an {\it admissible action on} $\mathbb{J}$ if it is a symmetry of its multi-contact distribution and satisfies conditions A1, A2, A3 and A4.\hfill $\clubsuit$
\end{defn}
\begin{exmp}\label{exC22C2@A1}
Let $G$ be the 2-dimensional non-abelian Lie group. Let $\mathbb{J}_2=J^2(\mathbb{R},\mathbb{R})$ and
$\mathbb{J}_1=J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})$. The multi-contact systems on $\mathbb{J}_1$ and $\mathbb{J}_2$ are
$$
\begin{aligned}
\mathcal{C}\langle 2,2\rangle&=\{\P {\alpha_1}+x^1_1\P {x^1}+x^1_2\P {x^1_1},\P {x^1_2}\}\oplus
\{\P {\alpha_2}+x^2_1\P {x^2}+x^2_2\P {x^2_1},\P {x^2_2}\},\cr
\mathcal{C}\langle 2\rangle &=\{\P {\beta_1}+y^1_1\P {y^1}+y^1_2\P {y^1_1},\P {y^1_2}\},
\end{aligned}
$$
respectively. The action of $G$ on $\mathbb{J}_1$ obtained by prolongation to second order of the action on $J^0(\mathbb{R},\mathbb{R})\times J^0(\mathbb{R},\mathbb{R})$ given by
$$
\mu_1(g)(\alpha_1,x^1,\alpha_2,x^2)=(\alpha_1,ax^1+b,\alpha_2,ax^2+b)
$$
is admissible on $\mathbb{J}_1$, as may be checked. The action of $G$ on $\mathbb{J}_2$ obtained by the prolongation to second order of the action on $J^0(\mathbb{R},\mathbb{R})$ given by
$$
\mu_2(g)(\beta_1,y^2)=(\beta_1,ay^1+b)
$$
is admissible on $\mathbb{J}_2$. \hfill $\spadesuit$
\end{exmp}
The next step is to construct a {\it right moving frame}, [\ref{FO}], $\rho:\mathbb{J}_1\to G$ for the action $\mu_1$ and a left moving frame
$\lambda:\mathbb{J}_2\to G$, for the action $\mu_2$.
Let $\alpha_i,\varpi_1,\varpi_2,\ldots,\varpi_{s_1},u_i$, $1\leq i\leq n_1$ be the differential invariants of $\mu_1$ and let $\beta_j, \varsigma_1,\varsigma_2,\allowbreak \ldots,\varsigma_{s_2},v_j$, $1\leq j\leq n_2$ be the differential invariants of $\mu_2$.
Denote local coordinates on $G$ by $\boldsymbol{w}=w_1,\ldots,w_r$.
The free, regular action $\mu_1$ determines a principal $G$-bundle $\pi^1: \mathbb{J}_1\to \mathbb{J}_1/G:=\mathcal{S}_1$, with fibre diffeomorphic to $G$. Similarly, we have the principal $G$-bundle $\pi^2:\mathbb{J}_2\to \mathbb{J}_2/G:=\mathcal{S}_2$ arising from $\mu^2$. Local trivialisations of $\boldsymbol{\phi^1}$ and $\boldsymbol{\phi^2}$ are local diffeomorphisms
$$
\boldsymbol{\phi^1}:\mathbb{J}_1\to\mathcal{S}_1\times G
$$
and
$$
\boldsymbol{\phi^2}:\mathbb{J}_2\to\mathcal{S}_2\times G
$$
by
\begin{equation}\label{phi1}
\boldsymbol{\phi^1}(\boldsymbol{x})=(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u},\rho(\boldsymbol{x})),\ \ \ \ \forall\ \ \boldsymbol{x}\in\mathbb{J}_1
\end{equation}
and
\begin{equation}\label{phi2}
\boldsymbol{\phi^2}(\boldsymbol{y})=(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v},\lambda(\boldsymbol{y})),
\ \ \ \ \forall\ \ \boldsymbol{y}\in\mathbb{J}_2
\end{equation}
respectively, where,
$$
\begin{aligned}
&\boldsymbol{\alpha}=\alpha_1,\ldots,\alpha_{n_1},\ \ \boldsymbol{\beta}=\beta_1,\ldots,\beta_{n_2},\cr
&\boldsymbol{\varpi}=\varpi_1,\varpi_2,\ldots,\varpi_{s_1},\ \ \boldsymbol{\varsigma}=\varsigma_1,\varsigma_2,\ldots,\varsigma_{s_2},\cr
&\boldsymbol{u}=u_1,\ldots,u_{n_1},\ \ \boldsymbol{v}=v_1,\ldots,v_{n_2}.
\end{aligned}
$$
and $\boldsymbol{x}$, $\boldsymbol{y}$ denote the canonical multi-contact coordinates on $\mathbb{J}_1$ and
$\mathbb{J}_2$, respectively. We shall call the maps $\boldsymbol{\phi^1},\boldsymbol{\phi^2}$ the {\it right-} and {\it left augmented moving frames}, respectively.
\begin{defn}
Let $\Omega$ be a Pfaffian system on smooth manifold $M$ and $\mu : G\times M\to M$ a free, regular Lie group action on $M$ which preserves $\Omega$,
$$
\mu(g)^*\Omega\subseteq\Omega,\ \ \ \ \forall\ \ g\in G.
$$
Then the {\it quotient of} $\Omega$ {\it by} $G$ is the Pfaffian system $\Omega/G$ on $M/G$ where,
$$
\Omega/G=\{\omega\in \Lambda^1(M/G)~|~\pi^*\omega\in\ \Omega\},
$$
and where $\pi : M\to M/G$ is the natural projection from $M$ onto the quotient space of $M$ by $G$.\footnote{We assume here that the quotient of a Pfaffian system is also a Pfaffian system; see [\ref{Itskov01}], [\ref{AndersonFels05}] for more details.}\hfill $\clubsuit$
\end{defn}
Naturally, one can likewise define the quotient of a vector field distribution in the obvious way.
\begin{defn}
Let $\mathcal{D}$ be a vector field distribution on manifold $M$ and $\mu : G\times M\to M$ a free and regular Lie group action on $M$ which preserves $\mathcal{D}$,
$$
\mu(g)_*\mathcal{D}\subseteq\mathcal{D},\ \ \ \ \forall\ \ g\in G.
$$
Let $\pi : M\to M/G$ be the natural projection from $M$ onto the quotient space of $M$ by $G$. Then the {\it quotient of} $\mathcal{D}$ {\it by} $G$ is the distribution on $M/G$ defined by
$$
\mathcal{D}/G=\ker\,\left(\Omega/G\right).
$$
\end{defn}
Let us denote the images $\boldsymbol{\phi^a}(\mathbb{J}_a)$ by $\Sigma_a$. By construction, for each $a$, $\Sigma_a$ is locally diffeomorphic to $\mathcal{S}_a\times G$, where $\mathcal{S}_1$ has local coordinates
$(\boldsymbol{\alpha}, \boldsymbol{\varpi}, \boldsymbol{u})$
and $\mathcal{S}_2$ has local coordinates
$(\boldsymbol{\beta}, \boldsymbol{\varsigma}, \boldsymbol{v})$.
It is important to keep distinct the $G$ factors in each of $\Sigma_1$ and $\Sigma_2$. For $\Sigma_1$, we shall denote by $\mathfrak{R}_1$ and $\mathfrak{L}_1$ the infinitesimal right- and left-translations, respectively on its copy of $G$ which we shall sometimes denote $G_{\boldsymbol{a}}$; local coordinates on $G_{\boldsymbol{a}}$ will be denote by $\boldsymbol{a}=(a_1,\ldots,a_r)$. Similarly, infinitesimal right- and left-translations on the $\Sigma_2$ copy of $G$, $G_{\boldsymbol{b}}$, will be denoted by $\mathfrak{R}_2$ and $\mathfrak{L}_2$, respectively and local coordinates on $G_{\boldsymbol{b}}$ will be denote by $\boldsymbol{b}=(b_1,\ldots,b_r)$.
\begin{thm}\label{DMZdistributions1}
For $a=1,2$, let $\mathbb{J}_a$ be products of jet spaces of the form (\ref{jetProds}) with multi-contact distributions $\mathcal{C}(\kappa_a)$. Suppose admissible actions of a Lie group $G$, $\mu_a:G\times \mathbb{J}_1\times\mathbb{J}_2\to\mathbb{J}_1\times\mathbb{J}_2$ preserve $\mathcal{C}(\kappa_a)$ for each $a$ and let $\boldsymbol{\phi^a}$ be their right and left augmented moving frames, (\ref{phi1}), (\ref{phi2}). Then, there are functions $\vec{h}_j$,$\vec{k}_j$ on $\mathcal{S}_1$ and functions $\vec{m}_j$,$\vec{n}_j$ on $\mathcal{S}_2$ such that
$$
\begin{aligned}
&\boldsymbol{\phi^1}_*\mathcal{C}(\kappa_1)=\bigoplus_{j=1}^{n_1}\Bigg\{\P {\alpha_j}+
\vec{h}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u})\cdot\P {\boldsymbol{\varpi}}+
\vec{k}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u})\cdot\boldsymbol{L}_1,\ \P {u_j}\Bigg\},\cr
&\boldsymbol{\phi^2}_*\mathcal{C}(\kappa_2)=\bigoplus_{l=1}^{n_2}\Bigg\{\P {\beta_l}+
\vec{m}_l(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v})\cdot\P {\boldsymbol{\varsigma}}+
\vec{n}_l(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v})\cdot\boldsymbol{R}_2,\ \P {v_l}\Bigg\},
\end{aligned}
$$
where $\boldsymbol{L}_1$ is a basis for $\mathfrak{L}_1$ and $\boldsymbol{R}_2$ is a basis for $\mathfrak{R}_2$.
\end{thm}
\vskip 3 pt
\noindent{\it Proof.} By suitable prolongations we can arrange that group actions on $\mathbb{J}_a$ are free and regular. By the Olver-Fels method of moving frames [\ref{FO}] we can arrange that the free, regular actions on $\mathbb{J}_1$ and $\mathbb{J}_2$ are admissible. By construction the maps $\boldsymbol{\phi^1}$ and $\boldsymbol{\phi^2}$ are right- and left-equivariant, respectively. That is,
\begin{equation}
\big(\boldsymbol{\phi^1}\circ\mu_1(g)\big)(\boldsymbol{x})=(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u},
R_{g^{-1}}^1\boldsymbol{w}^1)=\mathbf{R}_{g^{-1}}^1\boldsymbol{\phi^1}(\boldsymbol{x})
\end{equation}
and
\begin{equation}
\big(\boldsymbol{\phi^2}\circ\mu_2(g)\big)(\boldsymbol{y})=(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v},
L_{g}^2\boldsymbol{w}^2)=\mathbf{L}_g^2\boldsymbol{\phi^2}(\boldsymbol{y})
\end{equation}
where $R_g^1,L_g^2$ are right- and left-translations on $G_1$ and $G_2$, respectively and $\mathbf{R}_g^1=\text{id}\times R_g^1$ and
$\mathbf{L}_g^2=\text{id}\times L_g^2$. We now show that the infinitesimal generators of the action on $\mathbb{J}_1$ pushforward under $\boldsymbol{\phi^1}$ to infinitesimal right-translations $\mathfrak{R}_1$ on $G_1$ while the infinitesimal generators of the action on $\mathbb{J}_2$ pushforward under $\boldsymbol{\phi^2}$ to infinitesimal left-translations $\mathfrak{L}_2$ on $G_2$.
Also, it will be convenient to let $\boldsymbol{R}_1$, $\boldsymbol{L}_1$ denotes bases for $\mathfrak{R}_1$ and $\mathfrak{L}_1$, respectively. Similarly $\boldsymbol{R}_2$, $\boldsymbol{L}_2$ denote bases for the Lie algebras $\mathfrak{R}_2$ and $\mathfrak{L}_2$, respectively.
Let $\mathbb{X}_a$ denote bases for the infinitesimal generators of the actions $\mu_a:G\times \mathbb{J}_a\to \mathbb{J}_a$ of $G$ on $\mathbb{J}_a$,\ $a=1,2$. We have
\begin{equation}
\big(\mu_a(g)\big)_*\mathbb{X}_a=\text{Ad}(g)\,\mathbb{X}_a,\ \ \ a=1,2.
\end{equation}
Hence, in case $a=1$,
$$
\begin{aligned}
(\mathbf{R}_{g^{-1}}^1)_*\boldsymbol{\phi^1}_*\mathbb{X}_1&=\boldsymbol{\phi^1}_*\mu_1(g)_*\mathbb{X}_1\cr
&=\boldsymbol{\phi^1}_*\left(\text{Ad}(g)\,\mathbb{X}_1\right)\cr
&=\text{Ad}(g)\,\boldsymbol{\phi^1}_*\mathbb{X}_1\cr
&=\left(\mathbf{R}_{g^{-1}}^1\right)_*\left(\mathbf{L}_g^1\right)_*\boldsymbol{\phi^1}_*\mathbb{X}_1,
\end{aligned}
$$
and therefore
$$
\left(\mathbf{L}_g^1\right)_*\boldsymbol{\phi^1}_*\mathbb{X}_1=\boldsymbol{\phi^1}_*\mathbb{X}_1.
$$
Thus $\boldsymbol{\phi^1}_*\mathbb{X}_1$ is a left-invariant $\mathbb{R}^r$-valued vector field on $G_1$. A similar argument shows that $\boldsymbol{\phi^2}_*\mathbb{X}_2$ is a right-invariant $\mathbb{R}^r$-valued vector field on $G_2$.
Now fix any total differential operator $\mathcal{D}_j\in \mathcal{C}(\kappa_1)$ and any infinitesimal generator $X$ of $\mu_1$.
Then since $X$ is an infinitesimal symmetry of $\mathcal{C}(\kappa_1)$, it follows that
$[X,\mathcal{D}_j]\in \mathcal{C}(\kappa_1)$. By hypothesis, an admissible group action fixes the ``independent variables" $\alpha_i$ and hence $X\alpha_i=0$ for all $i$ from which it easily follows that
$$
[X,\mathcal{D}_j]\in\big\{\P {x^1_{k_1}},\ \P {x^2_{k_2}},\ldots,\P {x^{n_1}_{k_{n_1}}}\big\}.
$$
Also, by property A3) for admissible action we deduce that
$$
\boldsymbol{\phi^1}_*(\P {x_{k^1_j}})=\P {u_j},\ \ \boldsymbol{\phi^2}_*(\P {y_{k^2_l}})=\P {v_l}.
$$
From these facts and because
$\P {\boldsymbol{\alpha}},\P {\boldsymbol{\varpi}},\P {\boldsymbol{u}},\boldsymbol{L}_1$ form a frame-field on $\boldsymbol{\phi^1}(\mathbb{J}_1)$, we deduce that
$$
\begin{aligned}
\boldsymbol{\phi^1}_*(\mathcal{D}_j)=
\P {\alpha_j}+\vec{h}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u},\boldsymbol{a})\cdot\P {\boldsymbol{\varpi}}+
\vec{k}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},&\boldsymbol{u},\boldsymbol{a})\cdot\boldsymbol{L}_1\cr
&\mod \{\P {u_1},\P {u_2},\ldots,\P {u_n}\}
\end{aligned}
$$
for some functions $\vec{h}_j, \vec{k}_j$ on $\boldsymbol{\phi^1}(\mathbb{J}_1)$. Since $\boldsymbol{\phi^1}_*(X)\in\mathfrak{R}_1$, it follows that
$\vec{h}_j,\vec{k}_j$ are independent of $\boldsymbol{a}$ and so depend only on the invariants of $\mu_1$. They therefore have the form
$$
\boldsymbol{\phi^1}_*(\mathcal{D}_j)=\P {\alpha_j}+
\vec{h}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u},)\cdot\P {\boldsymbol{\varpi}}+
\vec{k}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u},)\cdot\boldsymbol{L}_1\mod \{\P {u_1},\P {u_2},\ldots,\P {u_n}\}.
$$
Similarly, if $\mathcal{E}_l$ is a total differential operator in $\mathcal{C}(\kappa_2)$ then
$$
\boldsymbol{\phi^2}_*(\mathcal{E}_l)=\P {\beta_l}+
\vec{m}_l(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v})\cdot\P {\boldsymbol{\varsigma}}+
\vec{n}_l(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v})\cdot\boldsymbol{R}_2\mod \{\P {v_1},\P {v_2},\ldots,\P {v_n}\}
$$
for some functions $\vec{m}_l,\vec{n}_l$ on $\boldsymbol{\phi^2}(\mathbb{J}_2)$, independent of $\boldsymbol{b}$.
\hfill\hfill\qed
\vskip 5 pt
Next, let $\mathbf{id}_1 :G_{\boldsymbol{a}}\to G$ and $\mathbf{id}_2: G_{\boldsymbol{b}}\to G$ be identity maps. Let a distribution $H$ be defined on $\mathcal{S}_1\times\mathcal{S}_2\times G$ by
\begin{equation}\label{H}
\begin{aligned}
H=&\bigoplus_{j=1}^{n_1}\Bigg\{\P {\alpha_j}+
\vec{h}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u})\cdot\P {\boldsymbol{\varpi}}+
\vec{k}_j(\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u})\cdot\boldsymbol{L},\ \P {u_j}\Bigg\}\oplus\cr
&\bigoplus_{l=1}^{n_2}\Bigg\{\P {\beta_l}+
\vec{m}_l(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v})\cdot\P {\boldsymbol{\varsigma}}+
\vec{n}_l(\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v})\cdot\boldsymbol{R},\ \P {v_l}\Bigg\}=H_1\oplus H_2
\end{aligned}
\end{equation}
where $\boldsymbol{L}={\mathbf{id}_1}_*\boldsymbol{L}_1$ and $\boldsymbol{R}={\mathbf{id}_2}_*\boldsymbol{R}_2$ form bases for the infinitesimal left- and right-translations on $G$; for later use we denote their local expression by
$$
L_i=\sum_{j=1}^r\Lambda_i^j(w)\P {w_j},\ \ \ R_i=\sum_{j=1}^r\Pi_i^j(w)\P {w_j},
$$
where $\boldsymbol{w}=(w_1,\ldots,w_r)$ are local coordinates on $G$.
Next, we define an action on $\Sigma_1\times\Sigma_2=(\mathcal{S}_1\times G_{\boldsymbol{a}})\times (\mathcal{S}_2\times G_{\boldsymbol{b}})$ by
$$
{\mu}_D(g)(\mathcal{I}_1,R_{g^{-1}}\boldsymbol{a},\mathcal{I}_2,L_g\boldsymbol{b}),\ \ \ \forall\ g\in G.
$$
where $\mathcal{I}_1=\boldsymbol{\alpha},\boldsymbol{\varpi},\boldsymbol{u}$ and $\mathcal{I}_2=\boldsymbol{\beta},\boldsymbol{\varsigma},\boldsymbol{v}$. Action $\mu_D$ is a free, regular left-action of $G$ and we denote the quotient of $\Sigma_1\times\Sigma_2$ by the orbits of $\mu_D$ by $\Sigma_1\times_G\Sigma_2$. Observe that
$$
\pi(\mathcal{I}_1,\boldsymbol{a},\mathcal{I}_2,\boldsymbol{b})=
(\mathcal{I}_1,\mathcal{I}_2,\boldsymbol{a}\cdot\boldsymbol{b})
$$
is $\mu_D$-invariant, mapping $G$-orbits in $\Sigma_1\times\Sigma_2$ to points in $\Sigma_1\times_G\Sigma_2$. Hence
$$
\pi:\Sigma_1\times\Sigma_2\to\Sigma_1\times_G\Sigma_2
$$
is the quotient map.
We denote the final term in the derived flag of any distribution $\mathcal{E}$ by the symbol $\mathcal{E}^{(\infty)}$. It is the smallest integrable distrubution containing $\mathcal{E}$. By construction, the maximal integral submanifolds of $H_1^{(\infty)}$ are the leaves of the foliation in $\mathcal{S}_1\times\mathcal{S}_2\times G$ defined locally by $\mathcal{I}_2$=constant. Similarly, those of $H_2^{(\infty)}$ are locally defined by $\mathcal{I}_1$=constant.
Let $S_1$ denote a fixed integral submanifold of $H_2^{(\infty)}$ and $S_2$ a fixed integral submanifold of $H_1^{(\infty)}$ in $\Sigma_1\times_G\Sigma_2\simeq\mathcal{S}_1\times\mathcal{S}_2\times G$. Then as well as $\pi$, define projections $\pi_1,\pi_2$, and inclusions $\iota_1,\iota_2$ such that the following diagram commutes
$$
\begindc{\commdiag}[50]
\obj(1,2){$(\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2)$}
\obj(0,1){$(S_1, \Theta_1)$}
\obj(2,1){$(S_2, \Theta_2$)}
\obj(1,0){$(\Sigma_1\times_G\Sigma_2, \Theta)$}
\mor{$(\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2)$}{$(\Sigma_1\times_G\Sigma_2, \Theta)$}{$\pi$}
\mor{$(\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2)$}{$(S_1, \Theta_1)$}{$\pi_1$}[-1,0]
\mor{$(\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2)$}{$(S_2, \Theta_2$)}{$\pi_2$}
\mor{$(S_1, \Theta_1)$}{$(\Sigma_1\times_G\Sigma_2, \Theta)$}{$\iota_1$}[-1,0]
\mor{$(S_2, \Theta_2$)}{$(\Sigma_1\times_G\Sigma_2, \Theta)$}{$\iota_2$}[1,0]
\enddc
$$
Explicitly
$$
\begin{aligned}
&\pi_1(\mathcal{I}_1,\boldsymbol{a},\mathcal{I}_2,\boldsymbol{\beta})=(\mathcal{I}_1, \boldsymbol{a}),\ \
\pi_2(\mathcal{I}_1,\boldsymbol{a},\mathcal{I}_2,\boldsymbol{b})=(\mathcal{I}_2, \boldsymbol{b}),\cr
&\iota_1(\mathcal{I}_1,\boldsymbol{w})=(\mathcal{I}_1,\mathcal{I}_2,\boldsymbol{w}),\ \
\iota_2(\mathcal{I}_2,\boldsymbol{w})=(\mathcal{I}_1,\mathcal{I}_2,\boldsymbol{w}).
\end{aligned}
$$
We have also defined the following Pfaffian systems
$$
\Theta=H^\perp,\ \Theta_1=\iota_1^*\Theta,\ \ \Theta_2=\iota_2^*\Theta,\ \ \Omega_1=\pi_1^*\Theta_1,\ \ \Omega_2=\pi_2^*\Theta_2.
$$
\begin{thm}\label{DMZdistributions2}
Distribution $H$, locally defined by (\ref{H}), can be identified with the quotient of $(\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2)$ by $\mu_D$ giving rise to the differential system $(\Sigma_1\times_G\Sigma_2,\Theta)$. The pair
$$
\left(\Sigma_1\times_G \Sigma_2,H\right)
$$
is a Darboux integrable $n$-hyperbolic manifold.
\end{thm}
\vskip 3 pt
{\proof
Pfaffian system $\Theta$ has local expression
\begin{equation}\label{eqQuotient5}
\Big\{\Upsilon_1,\ \Upsilon_2,\ \theta^j\Big\}_{j=1}^r
\end{equation}
where
$$
\theta^j=dw_j-\sum_{l=1}^{r}\sum_{t=1}^{n_1}k^l_t(\mathcal{I}_1)\Lambda^j_l(w)d\alpha_t
-\sum_{l=1}^r\sum_{s=1}^{n_2}n^l_s(\mathcal{I}_2)\Pi^j_l(w)d\beta_s,\ \ 1\leq j\leq r.
$$
The symbol $\Upsilon_1$ denotes 1-forms on $S_1$, while $\Upsilon_2$ denotes 1-forms on $S_2$.
Denote by $e$ the identity in $G$ and let $(\sigma,\tau)$ be local coordinates around $(e,e)\in G\times G$. In terms of the composition function $\mathbf{c}$ on $G$, we have
\begin{equation}\label{eqQuotient6}
\Lambda^j_i(w)=\frac{\partial \mathbf{c}^j}{\partial {\sigma_i}}(e,w),\
\Pi^j_i(w)=\frac{\partial \mathbf{c}^j}{\partial {\tau_i}}(w,e),\ i,j=1,\ldots,r.
\end{equation}
Recalling the right-invariance of the $L_i$ and the left-invariance of the $R_i$, we have, for $\sigma,\tau\in G$
\begin{equation}\label{eqQuotient7}
\begin{aligned}
&(\rho_\tau)_*\sum_{j=1}^r\Lambda^j_i(\sigma)\;{\P {s_j}}_{|_\sigma}=\sum_{j=1}^r\Lambda^j_i(u\cdot v)\;
{\P {s_j}}_{|_{\sigma\cdot \tau}},\cr
&(\lambda_\sigma)_*\sum_{j=1}^r\Pi^j_i(\tau)\;{\P {s_j}}_{|_\tau}=\sum_{j=1}^r\Pi^j_i(\sigma\cdot \tau)\;{\P {s_j}}_{|_{\sigma\cdot \tau}},
\end{aligned}
\end{equation}
where $\sigma\cdot \tau$ denotes multiplication on $G$ and $\lambda_\sigma, \rho_\tau$ denote left- and right-translations on $G$, respectively. However, we also have
\begin{equation}\label{eqQuotient8}
\begin{aligned}
&(\rho_\tau)_*\sum_{j=1}^r\Lambda^j_i(\sigma)\;{\P {s_j}}_{|_\sigma}=\sum_{j,k=1}^r\Lambda^k_i(\sigma)\frac{\partial \mathbf{c}^j}{\partial \sigma_k}(\sigma,\tau)\;
{\P {s_j}}_{|_{\sigma\cdot \tau}},\cr
&(\lambda_\sigma)_*\sum_{j=1}^r\Pi^j_i(\tau)\;{\P {s_j}}_{|_\tau}=\sum_{j,k=1}^r\Pi^k_i(\tau)\frac{\partial \mathbf{c}^j}{\partial \sigma_k}(\sigma,\tau)\;
{\P {s_j}}_{|_{\sigma\cdot \tau}}.
\end{aligned}
\end{equation}
Recall the projection $\pi :\Sigma_1\times\Sigma_2\to M$ defined by
$$
\pi(\mathcal{I}_1,\boldsymbol{a};\,\mathcal{I}_2,\boldsymbol{b})=
\big(\mathcal{I}_1,\mathcal{I}_2,\mathbf{c}(\boldsymbol{a},\boldsymbol{b})\big).
$$
Upon making use of (\ref{eqQuotient5})-(\ref{eqQuotient8}) we compute that for each $j=1,\ldots,r$
$$
\begin{aligned}
\pi^*\theta^j=&d\mathbf{c}(\boldsymbol{a,b})-\sum_{m,l=1}^r\sum_{t=1}^{n_1}k^l_t(\mathcal{I}_1)\Lambda^m_l(\boldsymbol{a})
\frac{\partial {c}^j}{\partial a_m}\;d\alpha_t\cr
&\hskip 110 pt-\sum_{m,l=1}^r\sum_{s=1}^{n_2}n^l_s(\mathcal{I}_2)\Pi^m_l(\boldsymbol{b})\frac{\partial {c}^j}{\partial v_m}\;d\beta_s\cr
=&\sum_{m=1}^r\frac{\partial {c}^j}{\partial a_m}\Big(
da_m-\sum_{l=1}^r\sum_{t=1}^{n_1}k^l_t(\mathcal{I}_1)\Lambda^m_l(\boldsymbol{a})\;d\alpha_t\Big)\cr
&\hskip 50 pt+\sum_{m=1}^r\frac{\partial {c}^j}{\partial b_m}\Big(db_m-\sum_{l=1}^r\sum_{s=1}^{n_2}n^l_s(\mathcal{I}_2)\Pi^m_l(\boldsymbol{b})\;d\beta_s\Big).
\end{aligned}
$$
On the other hand
$$
\begin{aligned}
&\Omega_1=\pi_1^*\iota_1^*\Theta=\Big\{\Upsilon_1,\ da_j-\sum_{l=1}^r\sum_{t=1}^{n_1}k^l_t(\mathcal{I}_1)\Lambda^j_l(\boldsymbol{a})\;d\alpha_t\Big\}_{j=1}^r\cr
&\Omega_2=\pi_2^*\iota_2^*\Theta=\Big\{\Upsilon_2,\ db_j-\sum_{l=1}^r\sum_{s=1}^{n_2}n^l_s(\mathcal{I}_2)\Pi^j_i(\boldsymbol{b})\;d\beta_s\Big\}_{j=1}^r.
\end{aligned}
$$
We have therefore shown that
$$
\pi^*\Theta\subseteq\Omega_1\oplus\Omega_2.
$$
That is, $(\Sigma_1\times_G\Sigma_2,\Theta)$ is the quotient of the product structure $(\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2)$ by the action $\mu_D$ as we wanted.
Thus, we have the identification,
$$
(\Sigma_1\times_G\Sigma_2,\Theta)\approx (\Sigma_1\times\Sigma_2,\Omega_1\oplus\Omega_2 )/G.
$$
We now show that $(\Sigma_1\times_G\Sigma_2,H)$ (equivalently $(\Sigma_1\times_G\Sigma_2,\Theta)$) is a Darboux integrable $n$-hyperbolic manifold. We need to check that $\ch H^{(1)}=\{\P {u_1},\ldots,\P {u_{n_1}},\P {v_1},\ldots,\P {v_{n_2}}\}$ and that the derived type of $H$ is $[[2n,0],[3n,n],[3n+1, 3n+1]]$. Recall that
\begin{equation}
\begin{aligned}
&\ch\mathcal{C}(\kappa_1)^{(1)}=\Big\{\P {x^1_{k_1}}, \P {x^2_{k_2}},\ldots,\P {x^{s}_{k_{s}}}\Big\},\cr
&\ch\mathcal{C}(\kappa_2)^{(1)}=\Big\{\P {y^1_{l_{r_1}}}, \P {y^2_{l_{r_2}}},\ldots,\P {y^{r}_{l_{r}}}\Big\},
\end{aligned}
\end{equation}
where $s=n_1, r=n_2,k=k^1,l=k^2$. Hence,
$$
\ch H_1^{(1)}=\{\P {u_1},\ldots,\P {u_s}\},\ \ \ch H_2^{(1)}=\{\P {v_1},\ldots,\P {v_r}\}.
$$
From the local form of $H$, we see that
$$
\{\P {u_1},\ldots,\P {u_s}\}\oplus\{\P {v_1},\ldots,\P {v_r}\}\subseteq\ch H^{(1)}.
$$
Let $\chi\in\ch H^{(1)}$ and as $H_1\cap H_2=\{\ 0\ \}$, we have $\chi=\chi_1+\chi_2$ where, for $l=1,2,$ $\chi_l\in H_l$.
But this implies that $\chi_l\in\ch H^{(1)}_l$ and therefore $\chi_l\in(\text{\bf id}_l)_*\boldsymbol{\phi^l}_*\xi_l$ for some
$\xi_l\in\ch\mathcal{C}(\kappa_l)^{(1)}$. Hence $\chi\in\{\P {u_1},\ldots,\P {u_s}\}\oplus\{\P {v_1},\ldots,\P {v_r}\}$. That is,
$$
\ch H^{(1)}\subseteq\{\P {u_1},\ldots,\P {u_s}\}\oplus\{\P {v_1},\ldots,\P {v_r}\}.
$$
Hence, we've shown that
$\ch H^{(1)}=\{\P {u_1},\ldots,\P {u_s}\}\oplus\{\P {v_1},\ldots,\P {v_r}\}$.
Finally, we must check that $\dim H^{(1)}=3n$ and $\dim H^{(2)}=3n+1$. To establish the first equality, observe that $\dim\mathcal{C}(\kappa_a)^{(1)}=\dim\mathcal{C}(\kappa_a)+n_a$ and therefore $\dim H_a^{(1)}=\dim H_a+n_a$. Now $H^{(1)}_1\cap H^{(1)}_2=\{\ 0\ \}$. For suppose there is an nonzero $Y\in H^{(1)}_1\cap H^{(1)}_2$. This implies, in particular that $Y\in H^{(1)}$ and hence $[Y,H^{(1)}_2]\in H^{(1)}_2$. By the same token, $[Y,H^{(1)}_1]\in H^{(1)}_1$. This implies that
$Y\in\ch H^{(1)}_1\cap\ch H^{(1)}_2=\{\ 0\ \}$ and hence
$$
\begin{aligned}
\dim H^{(1)}=\dim H^{(1)}_1+\dim H^{(1)}_2&=\dim H_1+n_1+\dim H_2+n_2\cr
&=2n+n_1+n_2=3n.
\end{aligned}
$$
To see that $\dim H^{(2)}=3n+1$ we note that as $\dim M=3n+1$ and $\dim H^{(1)}=3n$, we have either $\dim H^{(2)}=3n$ or $\dim H^{(2)}=3n+1$. In the former case, $H^{(1)}$ is Frobenius integrable in which case there is a regular function $\eta$ on $M$ which is an invariant of $H^{(1)}$. In particular, $\eta$ is annihilated by each vector field in $H^{(1)}_1$ and each vector field in $H^{(1)}_2$. But the only invariants of $H^{(1)}_1$ are $\boldsymbol{\beta, \varsigma, v}$, while the only invariants of $H^{(1)}_2$ are $\boldsymbol{\alpha, \varpi, u}$. Hence, there are regular functions $\mathcal{F}_1,\mathcal{F}_2$ such that
$$
\mathcal{F}_1(\boldsymbol{\alpha,\varpi,u})=\eta=\mathcal{F}_2(\boldsymbol{\beta,\varsigma,v}).
$$
This implies that $\mathcal{F}_a$ are constant functions; a contradiction. We have now proved that $(\Sigma_1\times_G\Sigma_2,H)$ is a Darboux integrable $n$-hyperbolic manifold, as we wanted.}
\hfill\hfill\qed
\vskip 5 pt
\begin{exmp}
This is a continuation of {\it Example \ref{exC22C2@A1}}. To simplify notation we rewrite the multi-contact systems as
$$
\begin{aligned}
\mathcal{C}\langle 2,2\rangle &=\{\P x+x_1\P {x_0}+x_2\P {x_1},\P {x_2}\}\oplus
\{\P y+y_1\P {y_0}+y_2\P {y_1},\P {y_2}\}\cr
\mathcal{C}\langle 2\rangle &=\{\P z+z_1\P {z_0}+z_2\P {z_1},\P {z_2}\}
\end{aligned}
$$
Then $\mu_1(g):\mathbb{J}_1\to \mathbb{J}_1$ is given by
$$
\mu_1(a,b)(\boldsymbol{x},\boldsymbol{y})=(x,ax_0+b,ax_1,ax_2,y,ay_0+b,ay_1,ay_2)
$$
and $\mu_2(g):\mathbb{J}_2\to \mathbb{J}_2$ is given by
$$
\mu_2(a,b)(\boldsymbol{z})=(z,az_0+b,az_1,az_2).
$$
A complete list of $\mu_1$ invariants is
$$
x,y,\frac{x_0-y_0}{x_1},\frac{y_1}{x_1},\frac{x_2}{x_1},\frac{y_2}{y_1}
$$
whose right-moving frame is
$$
\rho : \mathbb{J}_1\to G
$$
given by
$$
\rho(\boldsymbol{x},\boldsymbol{y})=\left(\begin{matrix}\frac{1}{x_1}\ &\ -\frac{x_0}{x_1}\cr
0 & 1\end{matrix}\right)=\left(\begin{matrix}\ a\ &\ \ b\ \cr
0\ &\ 1\end{matrix}\right).
$$
A complete list of $\mu_2$ invariants is
$$
z, \frac{z_2}{z_1}
$$
whose left moving frame is
$$
\lambda(\boldsymbol{z})=\left(\begin{matrix}\frac{1}{z_1}\ &\ -\frac{z_0}{z_1}\cr
0 & 1\end{matrix}\right)^{-1}=\left(\begin{matrix} z_1\ &\ -z_0\cr
0\ &\ 1\end{matrix}\right).
$$
Consequently, the diffeomorphism $\boldsymbol{\phi^1}:\mathbb{J}_1\to\mathbb{R}^6\times G_{\boldsymbol{a}}$ for this example is
$$
\boldsymbol{\phi^1}(\boldsymbol{x},\boldsymbol{y})=\left(x,y,\frac{x_0-y_0}{x_1},\frac{y_1}{x_1},\frac{x_2}{x_1},\frac{y_2}{y_1},
\frac{1}{x_1}, -\frac{x_0}{x_1}\right)=(x,y,\varpi_1,\varpi_2,u_1,u_2,a_1,a_2)
$$
while diffeomorphism $\boldsymbol{\phi^2} : \mathbb{J}_2\to \mathbb{R}^2\times G$ is
$$
\boldsymbol{\phi^2}(\boldsymbol{z})=\left(z,\frac{z_2}{z_1},z_1,-z_0\right)=(z,v_1,b_1,b_2).
$$
The pushforward of $\mathcal{C}(2,2)$ by $\boldsymbol{\phi^1}$ yields
$$
\begin{aligned}
\boldsymbol{\phi^1}_*\mathcal{C}\langle 2,2\rangle=&\big\{\P x+(1-\varpi_1u_1)\P {\varpi_1}-\varpi_2u_1\P {\varpi_2}-u_1^2\P {u_1}-u_1a_1\P {a_1}-(1+u_1a_2)\P {a_2}, \P {u_1}\big\}\cr
&\hskip 140 pt\oplus\{\P y-\varpi_2\P {\varpi_1}+\varpi_2u_2\P {\varpi_2}-u_2^2\P {u_2},\P {u_2}\}\cr
=&\{\tilde{X}_1,\P {u_1}\}\oplus\{\tilde{X}_2,\P {u_2}\},
\end{aligned}
$$
and the pushforward of $\mathcal{C}\langle 2\rangle$ by $\boldsymbol{\phi^2}$ yields
$$
\begin{aligned}
\boldsymbol{\phi^2}_*\mathcal{C}\langle 2\rangle=&\big\{\P z-v_1^2\P {v_1}+vb_1\P {b_1}-b_1\P {b_2},\P {v_1}\big\}\cr
=&\{\tilde{Y}_1,\P {v_1}\}.
\end{aligned}
$$
Next, in accordance with Theorem \ref{DMZdistributions2}, we form the differential system
$(\Sigma_1\times_G \Sigma_2,H)$ with
$$
H=\{X_1,\P {u_1}\}\oplus\{X_2,\P {u_2}\}\oplus\{Y_1,\P {v_1}\}
$$
where
$$
\begin{aligned}
&X_1=\P x+(1-\varpi_1u_1)\P {\varpi_1}-\varpi_2u_1\P {\varpi_2}-u_1^2\P {u_1}-u_1w_1\P {w_1}-(1+u_1w_2)\P {w_2},\cr
&X_2=\P y-\varpi_2\P {\varpi_1}+\varpi_2u_2\P {\varpi_2},\cr
&Y_1=\P z-v_1^2\P {v_1}+vw_1\P {w_1}-w_1\P {w_2}.
\end{aligned}
$$
It can be checked that $(\Sigma_1\times_G \Sigma_2,H)$ is a Darboux integrable 3-hyperbolic manifold. Indeed, its derived type is
$$
[[6,0],[9,3],[10,10]]
$$
and
$$
\ch H^{(1)}=\{\P {u_1},\P {u_2},\P {v_1}\},
$$
proving that $H$ is a Darboux integrable 3-hyperbolic distribution, since it also has the correct local normal form. In accordance with Theorem \ref{DMZdistributions2}, $H$ is the quotient of
$\mathcal{C}\langle 2,2\rangle\oplus\mathcal{C}\langle 2\rangle$ by the diagonal action and the natural map
$$
\pi:\Sigma_1\times \Sigma_2\to \Sigma_1\times_G \Sigma_2
$$
is given by
$$
\pi(\boldsymbol{x},\boldsymbol{y},\boldsymbol{z})=(x,y,\varpi_1,\varpi_2,u_1,u_2,z,v_1,g_1,g_2)
$$
where $g_1,g_2$ satisfy
$$
\left(\begin{matrix}\, g_1\ &\ g_2\,\cr
0\ &\ 1\end{matrix}\right)=\left(\begin{matrix} a_1\ &\ a_2\cr
0 & 1\end{matrix}\right)
\left(\begin{matrix} b_1\ &\ b_2\cr
0 & 1\end{matrix}\right).
$$
We can now implement Theorem \ref{KTdistributions}. We find that
$$
\mathcal{A}=\{\P {u_1},\P {u_2},\P {v_1},\P {\varpi_2},\P {w_1},\varpi_1\P {\varpi_1}+w_2\P {w_2}\}
$$
whose invariants are $x,y,z,p=\varpi_1/w_2$. Computing coordinates as in Theorem \ref{KTdistributions} and relabelling $Y_1$ as $X_3$
$$
p_i=X_ip,p_{ij}=X_ip_j
$$
we find
\begin{equation}\label{Ex3.1defingEqs1}
p=\varpi_1/w_2,p_1=\varpi_1/w_2^2+1/w_2,p_2=-\varpi_2/w_2,p_3=\varpi_1w_1/w_2^2
\end{equation}
and
\begin{equation}\label{Ex3.1defingEqs2}
p_{12}=f_{12}=-\varpi_2/w_2^2,p_{13}=f_{13}=w_1(w_2+2\varpi_1)/w_2^3,p_{23}=f_{23}=-\varpi_2w_1/w_2^2
\end{equation}
Solving equation (\ref{Ex3.1defingEqs1}) for $\varpi_1,\varpi_2,w_1,w_2$ in terms of $p,p_1,p_2,p_3$ and substituting into equation (\ref{Ex3.1defingEqs2}) reveals the {\sf GDMZ} system in standard jet coordinates
\begin{equation}\label{Ex3.1PDE}
\boxed{
u_{xy}=\frac{2u+1}{u(u+1)}u_xu_z,\ u_{xz}=\frac{1}{u+1}u_xu_y,\ u_{yz}=\frac{1}{u}u_yu_z
}
\end{equation}
As predicted by Theorem \ref{KTdistributions}, system (\ref{Ex3.1PDE}) is semilinear and involutive.\hfill $\spadesuit$
\end{exmp}
By its construction, the {\sf GDMZ} system (\ref{Ex3.1PDE}) is {\it Darboux integrable}. Using the methods of [\ref{AFV09}], its general solution can be expressed in finite terms of three arbitrary functions, each of one variable. This fact will be used in the next section when solutions of a given {\sf DMZ} system are required.
Though the characterisation of {\sf GDMZ} systems provided by Theorem \ref{KTdistributions} is quite general our specific interest in this paper is in {\it linear} {\sf GDMZ} systems since these are the ones that are known to have interesting applications. To construct these, we use the earlier method of quotienting products of jet spaces by a group action but now choosing the $G$-action to be that of an {\it abelian} group. To ilustrate the significance of this, we now briefly indicate how to use it to construct solutions of {\sf 3WRI} and {\sf m3WRI} systems. A complete treatment is presented in [\ref{SergeevVassiliou09}]
\section{Solutions of {\sf 3WRI} systems from 3-hyperbolic manifolds. Role of gauge transformations}
\label{sec:3}
The purpose of this section is to give an indication of how the theory developed so far can be applied to construct solutions for the $n$-wave resonant interaction system and the modified $n$-wave resonant interaction system that we discussed briefly in the Introduction. For simplicity of exposition, the case $n=3$ will be discussed but the reader will quickly surmise its generalisation to arbitrary $n$.
We consider the following action of $\mathbb{R}^4$ on $J^4(\mathbb{R},\mathbb{R})$
$$
\begin{aligned}
\mu^1_g(z,q,q_1,q_2,q_3,q_4)\mapsto \Big(z,q-\frac{z^2}{2}t_1+\frac{z}{2}t_2-&\frac{t_3}{6}+\frac{z^3}{6}t_4,\cr
q_1-zt_1+\frac{1}{2}t_2+\frac{z^2}{2}t_4,q_2-&t_1+zt_4,q_3+t_4,q_4\Big),
\end{aligned}
$$
where we have deviated a little from the notation of section 3 by denoting the standard coordinates on $J^4(\mathbb{R},\mathbb{R})$ by $z,q,q_1,q_2,q_3,q_4$ and coordinates on $G=\mathbb{R}^4$ are $t_1,\ldots,t_4$.
Denote the standard coordinates on $J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})$ by $x,m,m_1,m_2,y,n,n_1,n_2$ and define the action of $\mathbb{R}^4$ by
$$
\begin{aligned}
\mu^2_g(x,m,m_1,m_2,y,n,n_1,n_2)=(x,m+t_2-xt_1,&m_1-t_1,m_2,\cr
&y,n+t_4-yt_3,n_1-t_3,n_2)
\end{aligned}
$$
The invariants of $\mu^1$ are $z,q_2$ while the $G$-equivariant moving frame is
$$
\rho(z,q,q_1,q_2,q_3,q_4)=(q_2-zq_3,2zq_2-z^2q_3-2q_1,6q-6zq_1+3z^2q_2-z^3q_3,-q_3).
$$
The invariants of $\mu_2$ are $x,m_2,y,n_2$ and the $G$-equivariant moving frame is
$$
\lambda(x,m,m_1,m_2,y,n,n_1,n_2)=(m_1,xm_1-m,n_1,yn_1-n).
$$
Forming the diffeomorphisms $\boldsymbol{\phi^1,\phi^2}$ are indicated we find that
$$
\boldsymbol{\phi^1}_*\mathcal{C}\langle 4\rangle=\Big\{\P z-u(z\P {a_1}+z^2\P {a_2}+z^3\P {a_3}+\P {a_4}),\ \P u\Big\}
$$
and
$$
\boldsymbol{\phi^2}_*\mathcal{C}\langle 2,2\rangle=\Big\{\P x+v_1(\P {b_1}+x\P {b_2}),\ \P y+v_2(\P {b_3}+y\P {b_4}),\ \P {v_1},\ \P {v_2}\}
$$
As proved in section 3, the differential system $(M,H)$ where
$$
H=\{X,\P {v_1}\}\oplus\{Y,\P {v_2}\}\oplus\{Z,\P u\}
$$
on
$$
M=\boldsymbol{\phi^1}\Big(J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})\Big)\times_G\boldsymbol{\phi^2}\Big(J^2(\mathbb{R},\mathbb{R})\Big)
$$
is a Darboux integrable 3-hyperbolic manifold, where
$$
\begin{aligned}
&X=\P x+v_1(\P {w_1}+x\P {w_2}),\ \ Y=\P y+v_2(\P {w_3}+y\P {w_4})\cr
&Z=\P z-u(z\P {w_1}+z^2\P {w_2}+z^3\P {w_3}+\P {w_4}).
\end{aligned}
$$
Implementing the procedure in Theorem \ref{KTdistributions} we compute the integrable distribution
$$
\mathcal{A}=\{\P {v_1}, \P {v_2}, \P u, \P {w_1}+x\P {w_2}, \P {w_3}+y\P {w_4}, z\P {w_1}+z^2\P {w_2}+z^3\P {w_3}+\P {w_4}\}.
$$
Its invariants are spanned by
$$
x,\ y,\ z,\ (yz^3-1)(xw_1-w_2)+z(z-x)(yw_3-w_4)
$$
According to Theorem \ref{KTdistributions} we can take
$$
u=(yz^3-1)(w_2-xw_1)+z(z-x)(w_4-yw_3)
$$
as the dependent variable after which differentiation by $x-,\, y-,\, z-\,$ total differential operators
$X, Y, Z$, respectively generates all the higher order jet variables
$$
\begin{aligned}
&u_x=-yw_1z^3-zw_4+zyw_3+w_1,\ \ u_y=z^3(w_2-xw_1)-z(z-x)(1+z)w_3,\cr
&u_z=3yz^2(w_2-xw_1)+2zw_4-2zyw_3-xw_4+xyw_3,\cr
\cr
&u_{xy}=z(w_3-z^3w_1),\ \ u_{xz}=-3yz^2w_1-w_4+yw_3,\cr
&u_{yz}=3z^2w_2-3z^2xw_1-2w_3z+xw_3,
\end{aligned}
$$
and leads to the linear {\sf DMZ} system
\begin{equation}\label{m3WRIsol}
\begin{aligned}
&u_{xy}+\frac{z^3}{1-z^3y}u_x-\frac{1}{x-z}u_y-\frac{z^3}{(x-z)(1-z^3y)}u=0,\cr
&u_{xz}+\frac{3yz^2}{1-z^3y}u_x-\frac{1+2yz^3}{2xyz^3+x-2z-yz^4}u_z-\cr
&\hskip 150 pt\frac{3yz^2(1+2yz^3)}{(1-z^3y)(2xyz^3+x-2z-yz^4)}u=0\cr
&u_{yz}-\frac{x-2z}{z(x-z)}u_y-\frac{z^3(2x-z)}{2xyz^3+x-2z-yz^4}u_z+\cr
&\hskip 150 pt \frac{(x-2z)(2x-z)z^2}{(x-z)(2xyz^3+x-2z-yz^4)}u=0.
\end{aligned}
\end{equation}
It can be check, in accordance with previous results that this overdetermined system of three equations in one unknown is a {\sf DMZ} system. In particular, its coefficients satisfy the integrability conditions (\ref{DMZfullCompatibility}) and is therefore involutive.
Note that system (\ref{m3WRIsol}) satisfies the constraint
\begin{equation}\label{SergeevConstraint}
C_{ij}=\Gamma_{ij}\Gamma_{ji},\ \ \text{for}\ \ \ (i,j)=(1,2),\ (1,3),\, (2,3),
\end{equation}
and therefore the coefficients $\Gamma_{ij}$ in (\ref{m3WRIsol}) constitute a solution of the {\it modified} {\sf 3WRI} system. From this example, it is evident that a great many {\sf DMZ} systems can be constructed by applying the results of sections 2 and 3.
\subsection{Gauge transformations} Even if the products of jet spaces and the diagonal action is fixed, there is still freedom in the choice of invariant $u$ that ultimately becomes the dependent variable. This is because we have the freedom to replace $u$ by an arbitrary function of $x,y,z$ times $u$: $u\mapsto \lambda(x,y,z)u$, while preserving the linearity of the resulting {\sf GDMZ} system. This ``gauge transformation" raises the important question as to how to construct those special {\sf DMZ} systems that lead to solutions of the {\sf 3WRI} system or the {\sf m3WRI} system.
To this end, recall that the general Darboux-Manakov-Zakharov linear problem is the {\sf DMZ} system
$$
\left(\P {x_i}\P {x_j}-\Gamma_{ij}\P {x_j}-\Gamma_{j\,i}\P {x_i}+
C_{ij}\right)\phi=0,\ \ 1\leq i<j\leq 3,
$$
where $\Gamma_{ij},\Gamma_{j\,i},C_{ij}$ are functions of $x_1,x_2,x_3$. The case $C_{ij}=0$ is the linear problem for the {\sf 3WRI} system while the case $C_{ij}=\Gamma_{ij}\Gamma_{j\,i}$ is the linear problem for the {\sf m3WRI} system. How can we construct {\sf DMZ} systems that fall into these special classes? To answer this question it is natural to invoke the study of gauge transformations and gauge equivalence of differential operators. Thus, we define the matrix-valued differential operator
\begin{equation}\label{involutiveOperator}
\mathcal{D}=\left(\begin{matrix}
\Q {x_1,x_2}-\Gamma_{21}\P {x_1}-\Gamma_{12}\P {x_2}+C_{12}\cr
\Q {x_1,x_3}-\Gamma_{31}\P {x_1}-\Gamma_{13}\P {x_3}+C_{13}\cr
\Q {x_2,x_3}-\Gamma_{32}\P {x_2}-\Gamma_{23}\P {x_3}+C_{23}
\end{matrix}\right)=
\left(\begin{matrix} L_{12}\cr
L_{13}\cr
L_{23}\end{matrix}\right)
\end{equation}
so that $\mathcal{D}\,u=0$ is a {\sf DMZ} system. We shall call such an operator $\mathcal{D}$ {\it involutive}. Each operator element $L_{ij}$
can be expressed as a composition of first order operators modulo a correction term
$$
L_{ij}=\big(\P {x_i}-\Gamma_{j\,i}\big)\circ\big(\P {x_j}-\Gamma_{ij}\big)+h_{ij}(\mathcal{D}),\ \ \ (i,j)\in\text{perm}_2\{1,2,3\}
$$
where
$$
h_{ij}(\mathcal{D})=\frac{{\partial} \Gamma_{ij}}{{\partial} x_j}-\Gamma_{ij}\Gamma_{ji}+C_{ij}.
$$
We define the infinite Lie pseudogroup of gauge transformations acting on operators $\mathcal{D}$ by
$$
\bar{\mathcal{D}}=\mathbf{T}_\lambda\mathcal{D}=e^{-\lambda}\mathcal{D}\circ e^\lambda=\left(\begin{matrix}
\Q {x_1,x_2}-\bar{\Gamma}_{21}\P {x_1}-\bar{\Gamma}_{12}\P {x_2}+\bar{C}_{12}\cr
\Q {x_1,x_3}-\bar{\Gamma}_{31}\P {x_1}-\bar{\Gamma}_{13}\P {x_3}+\bar{C}_{13}\cr
\Q {x_2,x_3}-\bar{\Gamma}_{32}\P {x_2}-\bar{\Gamma}_{23}\P {x_3}+\bar{C}_{23}
\end{matrix}\right)
$$
where $\lambda(x)$ is an arbitrary function of $x_i$ in the domain of the operator $\mathcal{D}$. The coefficients of the image operator are
\begin{equation}\label{gaugedCoefficients}
\bar{\Gamma}_{ij}=\Gamma_{ij}-\P {x_j}\lambda,\ \ \ \ \ \ \bar{C}_{(ij)}=L_{ij}\lambda.
\end{equation}
It is straightforward to prove that the functions $h_{ij}: \boldsymbol{\mathcal{Z}}\to C^{\infty}(\mathbb{R}^3)$ viewed as maps from the set of differential operators $\boldsymbol{\mathcal{Z}}$ of the form $\mathcal{D}$, such that $\mathcal{D}u=0$ is a {\sf DMZ} system, to the smooth real-valued functions $C^{\infty}(\mathbb{R}^3)$ on $\mathbb{R}^3$ are invariants of the gauge action
$$
h_{ij}(\mathbf{T}_\lambda\mathcal{D})=h_{ij}(\mathcal{D}), \ \ \forall\ \lambda\in C^{\infty}(\mathbb{R}^3),\ \ \text{and}\ \ \mathcal{D}\in\boldsymbol{\mathcal{Z}}.
$$
In fact functions $h_{ij}$ are {\it complete invariants} in the sense that operators $\mathcal{D}$ and $\mathcal{D}'$ are gauge equivalent, $\mathcal{D}'=\mathbf{T}_\lambda\mathcal{D}$ for some $\lambda$ if and only if the all their gauge invariants agree
$$
h_{ij}(\mathcal{D}')=h_{ij}(\mathcal{D}),\ \ \forall\ \ (i,j)\in\text{perm}_2\{1,2,3\}.
$$
\begin{prop}
If an operator $\mathcal{D}$ is involutive then so is each operator in its gauge orbit.
\end{prop}
\proof Gauge equivalence of operators is derived from the prolongation to second order of the local diffeomorphism
$$
\phi: x_1\mapsto x_1,\ x_2\mapsto x_2,\ x_3\mapsto x_3,\ u\mapsto \lambda(x_1,x_2,x_3)u,
$$
a contact equivalence of the differential equations that they define. Clearly $\phi$ preserves the independence form $dx_1\wedge dx_2\wedge dx_3$. The image of an involutive differential equation under a contact transformation which preserves the independence form is also involutive with the same independence form. \hfill\qed
\vskip 5 pt
A general involutive operator (\ref{involutiveOperator}) generates a solution of the {\sf 3WRI} equations once a solution of $\mathcal{D}u=0$ is known. Because the involutive operators arising from our construction are Darboux integrable, it follows that solutions of $\mathcal{D}u=0$ can be explicitly constructed.
\begin{prop}\label{General2_3wave}
Let $\mathcal{D}$ be any involutive differential operator (\ref{involutiveOperator}). Let $u$ be any nonzero solution of $\mathcal{D}u=0$. Then the coefficients of $\mathcal{D}_0=\mathbf{T}_u\mathcal{D}$ provide a solution of the {\sf 3WRI} system.
\end{prop}
\proof Formulas (\ref{gaugedCoefficients}) and the previous Proposition prove
that $\mathcal{D}_0$ is an involutive operator with $\bar{C}_{ij}=0$. Theorem \ref{geometricBackgroundThm} proves that the remaining nonzero coefficients of
$\mathcal{D}_0$ provide a solution of the {\sf 3WRI} system via the Lame potentials $h_1,h_2,h_3$.\hfill\qed
\vskip 5 pt
Next, we point out that an involutive operator (\ref{involutiveOperator}) satisfying $C_{ij}=0$ for all $i,\ j$, determines a solution of the {\it modified} {\sf 3WRI} system, that is a solution of (\ref{m3WRI}) once a certain quadrature can be performed.
\vskip 5 pt
\begin{prop}
Suppose $\mathcal{D}_0\in\mathcal{Z}$ is an involutive operator (\ref{involutiveOperator}) with $C_{ij}=0$ for all $(i,j)$. Then
\begin{enumerate}
\item The linear inhomogeneous PDE system
\begin{equation}\label{SergeevGauge}
\frac{{\partial}^2 \lambda}{{\partial} x_i{\partial} x_j}=\Gamma_{ij}\Gamma_{j\,i},\ \ \ (i,j)\in\{(1,2),\ (1,3),\ (2,3)\},
\end{equation}
for function $\lambda(x_1,x_2,x_3)$ is involutive, and
\item If $\lambda$ satisfies (\ref{SergeevGauge}) then the coefficients of $T_\lambda\mathcal{D}_0$ solve the modified {\sf 3WRI} system.
\end{enumerate}
\end{prop}
\proof In order for $\mathcal{D}_0$ to be transformed by some gauge to the linear problem $\mathcal{D}_1$ for the {\sf m3WRI} system we require
$$
\mathcal{D}_0\lambda=\left(\begin{matrix} (\Gamma_{21}-\lambda_{x_2})(\Gamma_{12}-\lambda_{x_1})\cr
(\Gamma_{31}-\lambda_{x_3})(\Gamma_{13}-\lambda_{x_1})\cr
(\Gamma_{32}-\lambda_{x_3})(\Gamma_{23}-\lambda_{x_2})\end{matrix}\right)
=\left(\begin{matrix} h_{21}(\mathbf{T}_\lambda\mathcal{D}_0)-\P {x_1}(\Gamma_{21}-\lambda_{x_2})\cr
h_{31}(\mathbf{T}_\lambda\mathcal{D}_0)-\P {x_1}(\Gamma_{31}-\lambda_{x_3})\cr
h_{32}(\mathbf{T}_\lambda\mathcal{D}_0)-\P {x_2}(\Gamma_{21}-\lambda_{x_3})\end{matrix}\right).
$$
Since the $h_{ij}$ are gauge-invariant this simplifies to equation (\ref{SergeevGauge})
whose integrability conditions are
\begin{equation}\label{mDMZtransIC}
\frac{{\partial}}{{\partial} x_k}\left(\Gamma_{ij}\Gamma_{ji}\right)=\frac{{\partial}}{{\partial} x_j}\left(\Gamma_{ik}\Gamma_{ki}\right),\ \ \ (i,j,k)\in\ \text{perm}(1,2,3).
\end{equation}
We recall that the $\Gamma_{ij}$ satisfy the integrability conditions (\ref{KT_0Compatibility}).
An calculation shows that the integrability conditions (\ref{mDMZtransIC}) are satisfied modulo (\ref{KT_0Compatibility}).\hfill\qed
\vskip 5 pt
Thus each solution of the {\sf 3WRI} system gives rise to solutions of the {\sf m3WRI} system.
\begin{exmp}
Proposition \ref{General2_3wave} asserts that every involutive operator determines another, $\mathcal{D}_0$, with $C_{ij}=0$. In turn $\mathcal{D}_0$ determines a solution of the 3-wave resonant interaction system as described in Theorem \ref{geometricBackgroundThm}. We shall now illustrate this construction beginning with the involutive system (\ref{m3WRIsol}) which was constructed by the theory of sections 2 and 3.
Performing the gauge transformation $\mathbf{T}_\lambda\mathcal{D}$, with
$\lambda=-\ln {z(x-z)}$ produces the operator
\begin{equation}\label{newDMZ}
\bar{\mathcal{D}}=\mathbf{T}_\lambda\mathcal{D}=\left(\begin{matrix}
\Q {x_1,x_2}-\bar{\Gamma}_{21}\P {x_1}-\bar{\Gamma}_{12}\P {x_2}\cr
\Q {x_1,x_3}-\bar{\Gamma}_{31}\P {x_1}-\bar{\Gamma}_{13}\P {x_3}\cr
\Q {x_2,x_3}-\bar{\Gamma}_{32}\P {x_2}-\bar{\Gamma}_{23}\P {x_3}
\end{matrix}\right)
\end{equation}
where
$$
\begin{aligned}
&\bar{\Gamma}_{21}=\frac{z^3}{yz^3-1},\cr
&\bar{\Gamma}_{31}=\frac{x+2xyz^3-yz^4-2z}{z(x-z)(yz^3-1)},\
\ \bar{\Gamma}_{13}=\frac{z(1-yz^3)}{(x+2xyz^3-yz^4-2z)(x-z)},\cr
&\phantom{\bar{\Gamma}_{31}=\frac{x+2xyz^3-yz^4-2z}{z(x-z)(yz^3-1)}}\ \ \ \ \bar{\Gamma}_{23}=\frac{z^3(2x-z)}{x+2xyz^3-yz^4-2z},
\end{aligned}
$$
with $\bar{\Gamma}_{12}=\bar{\Gamma}_{32}=0$. From these coefficients we obtain Lame potentials
$$
h_1=\frac{yz^3-1}{z(z-x)},\ \ h_2=1,\ \ h_3=\frac{-z^4y-2z+x+2xz^3y}{x-z},
$$
and the corresponding solution of the $2+1$\,-\,dimensional 3-wave resonant interaction equations, in accordance with Theorem \ref{geometricBackgroundThm} is easily computed to be
\begin{equation}\label{3WRIsol0}
\boxed{
\left(\begin{matrix}
0&A_{12}&A_{13}\cr
A_{21}&0&A_{23}\cr
A_{31}&A_{32}&0
\end{matrix}\right)
=\frac{1}{x-z}\left(\begin{matrix} 0&0 & z^2\cr
-z^2y^{-1} & 0 & z^3(2x-z)y^{-1}\cr
-z^{-2} &0 & 0\end{matrix}\right)}
\end{equation}
Because of Proposition \ref{General2_3wave}, we can generate new solutions of the 2+1-\newline dimensional {\sf 3WRI} system from any given solution arising from an appropriate operator $\mathcal{D}_0$, such as we have constructed in this example, by computing the image of $\mathcal{D}_0$ under a gauge transformation, $\mathbf{T}_u\mathcal{D}_0$, where $u$ is any nonconstant solution of $\mathcal{D}_0u=0$. Normally there is little hope of finding any such solutions but the situation is quite otherwise here because the operator $\mathcal{D}_0$ arises from a Darboux integrable $n$-hyperbolic distribution (with $n=3$ in this case)
$$
\begin{aligned}
H=\Big\{\P x+v_1(\P {w_1}+x\P {w_2}),\ \ \P {v_1}\Big\}\oplus&\Big\{\P y+v_2(\P {w_3}+y\P {w_4}),\P a,\ \ \P {v_2}\Big\}\oplus\cr
\Big\{\P z-u&(z\P {w_1}+z^2\P {w_2}+z^3\P {w_3}+\P {w_4}),\ \ \P u\Big\}\cr
&\hskip 45 pt =H_1\oplus H_2\oplus H_3.
\end{aligned}
$$
This distribution can be integrated using methods of [\ref{AFV09}]. By construction its {\it Vessiot group} [\ref{AFV09}] is $\mathbb{R}^4$. Furthermore $H_1,H_2$ and $H_3$ are each easily integrated and the fact that the Vessiot group is $\mathbb{R}^4$ means that the superposition formula is the usual linear superposition. From this we easily obtain the general integral submanifold of $H$ to be
$$
\begin{aligned}
\sigma(x,y,z)&=\cr
&\Big(-A_{xx},\ -B_{yy},\ -\frac{1}{6}z^3C_{zzzz}-2z^2C_{zzz}-6C_{zz}z-4C_{z},\cr
&\frac{1}{6}z^4C_{zzz}+\frac{4}{3}z^3C_{zz}+2z^2C_z-A_x,\cr
&\frac{7}{6}z^4C_{zz}+\frac{1}{6}z^5C_{zzz}+\frac{4}{3}z^3C_z+A-xA_x,\cr
&z^4C_z+\frac{1}{6}z^6C_{zzz}+z^5C_{zz}-B_y,\cr
&\frac{1}{6}z^3C_{zzz}+\frac{3}{2}z^2C_{zz}+3C_zz+C+B-yB_y\Big)\cr
&=(a,b,c,w_1,w_2,w_3,w_4).
\end{aligned}
$$
where $A(x),\ B(y),\ C(z)$ are arbitrary real-valued functions. We recall that
$$
u=\frac{(yz^3-1)(xw_1-w_2)+z(z-x)(yw_3-w_4)}{z(x-z)}
$$
generates the involutive operator (\ref{newDMZ}) and hence
\begin{equation}\label{DMZsol}
\begin{aligned}
\lambda=&\sigma^*u=\cr
&\Big(6xyz^5C_z-2z^4C_{zz}-10z^3C_z-6z^2C-6z^2B+2xyz^6C_{zz}+xz^3C_{zz}+6xz^2C_z-\cr
&yz^7C_{zz}-2yz^6C_z-6yz^3A+6xzC+6xzB+6A\Big)\Big(z(x-z)\Big)^{-1}
\end{aligned}
\end{equation}
is the general solution of $\bar{\mathcal{D}}\lambda=0$. It follows that the coefficients of $e^{-\lambda}\bar{\mathcal{D}}\circ e^\lambda$ determine a new solution of the 2+1-dimensional {\sf 3WRI} equations by Proposition \ref{General2_3wave}.
For instance, choosing $A(x)=B(y)=C(z)=1$ yields the particular solution
$$
\lambda=\frac{6-12z^2-6yz^3+12xz}{z(x-z)}
$$
of $\bar{\mathcal{D}}\lambda=0$. We compute
\begin{equation}
\widehat{\mathcal{D}}=\mathbf{T}_\lambda\bar{\mathcal{D}}=\left(\begin{matrix}
\Q {x_1,x_2}-\hat{\Gamma}_{21}\P {x_1}-\hat{\Gamma}_{12}\P {x_2}\cr
\Q {x_1,x_3}-\hat{\Gamma}_{31}\P {x_1}-\hat{\Gamma}_{13}\P {x_3}\cr
\Q {x_2,x_3}-\hat{\Gamma}_{32}\P {x_2}-\hat{\Gamma}_{23}\P {x_3}
\end{matrix}\right)
\end{equation}
where
$$
\begin{aligned}
&\hat{\Gamma}_{21}=\frac{2(z-x)z^4}{(yz^3-1)q},\ \ \hskip 62 pt\hat{\Gamma}_{12}=\frac{1-yz^3}{(z-x)q},\cr
&\hat{\Gamma}_{31}=\frac{2(z^4y-2yxz^3+2z-x)}{(yz^3-1)q},\
\ \hat{\Gamma}_{13}=\frac{(2yz^3+2z^2+1)(1-yz^3)}{(z^4y-2yxz^3+2z-x)q},\cr
&\hat{\Gamma}_{32}=\frac{(z^4y-2yxz^3+2z-x)}{(x-z)zq},\ \ \ \ \hat{\Gamma}_{23}=
\frac{(2z^2-4xz-3)(z-x)z^3}{(z^4y-2yxz^3+2z-x)q},
\end{aligned}
$$
and where
$$
q=2z^2+yz^3-2xz-1.
$$
These coefficients give rise to new Lam\'e potentials,
$$
h_1=\frac{yz^3-1}{q},\ \ h_2=\frac{(z-x)z}{q},\ \ h_3=\frac{2z-x-2yxz^3+z^4y}{q}.
$$
The corresponding explicit solution of the 3-wave resonant interaction equations is
\begin{equation}\label{3WRIsol1}
\boxed{
\left(\begin{matrix}
0&A_{12}&A_{13}\cr
A_{21}&0&A_{23}\cr
A_{31}&A_{32}&0
\end{matrix}\right)
=\frac{1}{q}\left(\begin{matrix} 0&-z & 2yz^3+2z^2+1\cr
\ \ \ 2z^3\ \ &\ \ \ 0 & \ \ \ z^2(2z^2-4xz-3)\cr
\ \ \ 2 &\ \ \ 1 & 0\end{matrix}\right)
}
\end{equation}
We have shown that commencing with a certain diagonal action of $\mathbb{R}^4$ on $J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})\times J^4(\mathbb{R},\mathbb{R})$ allows us to explicitly construct a solution of the {\sf m3WRI} equations and solution (\ref{3WRIsol0}) of the {\sf 3WRI} equations. Furthermore, because we are able to integrate differential system $H$ explicitly by methods of [\ref{AFV09}], we were able to construct, from solution (\ref{3WRIsol0}) of the {\sf 3WRI} system, the new solution (\ref{3WRIsol1}).
Evidently, the formula (\ref{DMZsol}) permits the construction of a solution of the {\sf 3WRI} system that depends on 3 functions, each of one variable, $A(x),\, B(y),\, C(z)$. For different choices of these functions, these solutions all belong to the same gauge class and are, to some extent, labeled by the gauge invariants $h_{ij}$ of the {\sf DMZ} linear system (\ref{m3WRIsol}). In this case, four of the six gauge invariants vanish $h_{12}=h_{21}=h_{13}=h_{31}=0$ while
$$
h_{32}=\frac{2(y^2z^6-2yz^3+1)}{(x-yz^4+2xyz^3-2z)^2},\ \ h_{23}=\frac{6z^2(z-x)^2}{(x-yz^4+2xyz^3-2z)^2}.
$$
The fact that $h_{32}$ and $h_{23}$ are nonzero means that we can perform Laplace transformations in the $(3,2)$ and $(2,3)$-directions [\ref{KT96}] to obtain new {\sf DMZ} systems and consequently solutions of the {\sf 3WRI} and {\sf m3WRI} systems that belong to a gauge class distinct from that of solution (\ref{3WRIsol1}).
\end{exmp}\hfill $\spadesuit$
\begin{rem}
Let $\mathbb{J}=J^{k_1}(\mathbb{R},\mathbb{R})\times J^{k_2}(\mathbb{R},\mathbb{R})\times J^{k_3}(\mathbb{R},\mathbb{R})$ and let
$$
\mu_D: \mathbb{R}^r\times\mathbb{J}\to\mathbb{J}
$$
be an admissible diagonal action of a Lie group $G$ as described in Theorems \ref{DMZdistributions1} and \ref{DMZdistributions2}, such that
\begin{equation}\label{dimensionConstraint}
\dim\mathbb{J}-\dim G=10.
\end{equation}
By choosing $k_i$ and $\dim G$ to be arbitrarily large, while maintaining the constraint (\ref{dimensionConstraint}) we can, by the means described above, obtain solutions of the {\sf 3WRI} system and the {\sf m3WRI} system, which depend on 3 smooth functions each of one variable and any number of their derivatives. An interesting question is whether or not these solutions are dense in the space of all solutions of the (say) {\sf 3WRI} system.
Despite their importance, up to now very \textcolor{red}{\bf few} explicit {\sf DMZ} systems were known in the literature. Significant known {\sf DMZ} systems are firstly those arising from known triply orthogonal coordinate systems and the ``1-periodic" systems, derived by Kamran \& Tenenblat [\ref{KT98}]. Another example is given in [\ref{KT96}, Example 3]; however, the candidate system may contain a typographical error since it doesn't appear to be involutive.
\end{rem}
\section{Semi-Hamiltonian systems and $n$-hyperbolic manifolds}
A second application of our geometric construction of {\sf GDMZ} systems is related to strongly hyperbolic PDE systems of {\it hydrodynamic type} expressible in Riemann invariants in the form
\begin{equation}\label{hydroType}
u^i_t=v^i(u^1,u^2,\ldots,u^n)u^i_x,\ \ 1\leq i\leq n.
\end{equation}
The term {\it strongly hyperbolic} here means that the functions $v^i(u)$ in (\ref{hydroType}) are pair-wise distinct.
Here it is shown how to use the theory developed in previous sections for linear {\sf DMZ} systems of the form
\begin{equation}\label{linearKT}
\frac{\partial^2 u}{\partial x_i\partial x_j}-\Gamma^i_{ij}\frac{\partial u}{\partial x_i}-
\Gamma^i_{ij}\frac{\partial u}{\partial x_j}=0,\ 1\leq i<j\leq n
\end{equation}
to construct new examples of PDE systems (\ref{hydroType}) which are {\it semi-Hamiltonian}. Such systems play a role in a range of applications in mathematics and physics and have been the subject of interesting developments; for instance see [\ref{Ferapontov97}, \ref{FeraMoroSokolov09}]. The geometric study of such systems arose in the work of Dubrovin \& Novikov [\ref{DN}], D. Serre [\ref{Se}] and Tsarev [\ref{Tsa}] and we will briefly review some of their results.
There is an interesting subclass of strongly hyperbolic systems of hydrodynamic type which we now define.
\vskip 5 pt
\begin{defn}
A strongly hyperbolic system (\ref{hydroType}) is said to be {\it semi-Hamiltonian} or {\it rich in conservation laws} if
\begin{equation}\label{def_semiHam}
\left(\frac{v^i_{,j}}{v^j-v^i}\right)_{,k}=\left(\frac{v^i_{,k}}{v^k-v^i}\right)_{,j}
\end{equation}
\end{defn}
\begin{exmp} The so called {\it chromotography system}
$$
u^i_t=u^i\prod_{l=1}^Nu^lu^i_x,\ \ 1\leq i\leq N,
$$
is Hamiltonian when $N=2$ but only semi-Hamiltonian for $N>2$.
\end{exmp}
Various important results have been proven about semi-Hamiltonian systems. For instance, D. Serre [\ref{Se}] proved that Lax's classical result [\ref{Lax64}] on the blow-up in finite time for generic $2\times 2$ systems extends to the class of
$n\times n$ semi-Hamiltonian systems. Another interesting result is due to Tsarev. To describe it we require a result of Darboux [\ref{Darboux10}]; see also [\ref{Tenenblat98}].
The following paraphrases results of Tsarev [\ref{Tsa}].
\begin{thm}[Tsarev,{[\ref{Tsa}]}]\label{TsarevSemiham}
Let
\begin{equation}\label{hydroType_thm}
u^i_t=v^i(u^1,u^2,\ldots,u^n)u^i_x,\ \ 1\leq i\leq n,
\end{equation}
be a strongly hyperbolic system of hydrodynamic type and consider the linear overdetermined PDE system
\begin{equation}\label{linearP_KT}
\frac{\partial^2 P}{\partial u_i\partial u_j}-\frac{v^i_{,j}}{v^j-v^i}\frac{\partial P}{\partial u_i}-
\frac{v^j_{,j}}{v^i-v^j}\frac{\partial P}{\partial u_j}=0,\ 1\leq i<j\leq n
\end{equation}
for a real-valued function $P(u^1,\ldots,u^n)$. Then (\ref{hydroType_thm}) is semi-Hamiltonian if and only if (\ref{linearP_KT}) is a {\sf DMZ} system.
\end{thm}
{\proof Suppose (\ref{linearP_KT}) is a {\sf DMZ} system. The integrability conditions
\begin{equation}\label{KT_semihamCompatibility}
\begin{aligned}
&\frac{{\partial}\Gamma^j_{ij}}{{\partial} u_k}-\Gamma^k_{ik}\Gamma^j_{kj}-\Gamma^i_{ki}\Gamma^j_{ij}+\Gamma^j_{ij}\Gamma^j_{kj}=0,\cr
&\frac{{\partial} \Gamma^j_{ij}}{{\partial} u_k}-\frac{{\partial} \Gamma^j_{kj}}{{\partial} u_i}=0,
\end{aligned}(i,j,k)\in\ \text{\rm perm}\,(1,2,3).
\end{equation}
are therefore satisfied, where
\begin{equation}\label{Darboux}
\Gamma^i_{ij}(u)=\frac{v^i_{\,,j}}{v^j-v^i},\ 1\leq i\neq j\leq n.
\end{equation}
Equations $(\ref{linearP_KT})_2$ are easily seen to entail the condition (\ref{def_semiHam}).
Conversely, if (\ref{hydroType_thm}) is semi-Hamiltonian, then a calculation verifies that (\ref{KT_semihamCompatibility}) will be satisfied given that $\Gamma^i_{ij}(u)$ also satisfy (\ref{Darboux}).\hfill\qed
}
\vskip 5 pt
In order to use the theory of this paper to construct new examples of semi-Hamiltonian hydrodynamic type systems, we are required from a given {\sf DMZ} system to solve the linear PDE system (\ref{Darboux}). Remarkably, the integrability condition arising from (\ref{Darboux}) is precisely $(\ref{KT_semihamCompatibility})_1$. This fact is attributed to Darboux [\ref{Darboux10}] in [\ref{Tenenblat98}], where a proof is given.
\begin{thm} [Darboux]\label{DarbouxTHM}
Let $\Gamma^i_{ij}(u^1,\ldots,u^n)$,\ $1\leq i\neq j\leq n$, be a collection of $n(n-1)$ smooth functions, symmetric in their lower indices and satisfying the linear {\sf DMZ} system integrability conditions with $C_{ij}=0$, equations (\ref{KT_0Compatibility}). That is,
\begin{equation}\label{KT_semihamCompatibilityDarboux}
\begin{aligned}
&\frac{{\partial}\Gamma^j_{ij}}{{\partial} u_k}-\Gamma^k_{ik}\Gamma^j_{kj}-\Gamma^i_{ki}\Gamma^j_{ij}+\Gamma^j_{ij}\Gamma^j_{kj}=0,
\end{aligned}(i,j,k)\in\ \text{\rm perm}\,(1,2,3).
\end{equation}
Then the linear first order PDE system
\begin{equation}\label{Darboux2}
\frac{1}{w^j-w^i}\frac{\partial w^i}{\partial u^j}=\Gamma^i_{ij}(u),\ 1\leq i\neq j\leq n,
\end{equation}
has smooth solutions depending on $n$ arbitrary functions each of one variable.
\end{thm}
\vskip 5 pt
Note that if we have by some means constructed a {\sf DMZ} system of the form
\begin{equation}\label{DMZconserved_density}
\frac{{\partial}^2 P}{{\partial} x_i{\partial} x_j}-\Gamma^i_{ji}(u)\frac{{\partial} P}{{\partial} x_i}-\Gamma^j_{ij}(u)\frac{{\partial} P}{{\partial} x_j}=0
\end{equation}
for a function $P(u^1,\ldots,u^n)$, then its coefficients will automatically satisfy (\ref{KT_semihamCompatibility}).
The significance of this remark arises from the following result of Tsarev [\ref{Tsa}].
\vskip 5 pt
\begin{thm}[\,Tsarev,\ {[\ref{Tsa}]}\,]\label{TsarevMain}
Let
\begin{equation}\label{hydroTypeTHM}
u^i_t=v^i(u^1,u^2,\ldots,u^n)u^i_x,\ \ 1\leq i\leq n
\end{equation}
be a semi-Hamiltonian system of hydroynamic type. Then
\begin{enumerate}
\item[1)] If $w^i(u)$ is a solution of
\begin{equation}\label{semihamFromDMZ}
\frac{1}{w^j-w^i}\frac{\partial w^i}{\partial u^j}=\Gamma^i_{ij}(u)=\frac{1}{v^j-v^i}\frac{\partial v^i}{\partial u^j},\ 1\leq i\neq j\leq n,
\end{equation}
then
$$
u^i_t=w^i(u)u^i_x
$$
is semi-Hamiltonian and defines a flow that commutes with (\ref{hydroTypeTHM}).
\item[2)] If $w^i(u)$ is a solution of (\ref{semihamFromDMZ}) then the vector field
$$
\sum_{i=1}^nw^i(u)u^i_x\P {u^i}
$$
is a first order generalised symmetry of (\ref{hydroTypeTHM}).
\end{enumerate}
\end{thm}
{\proof This follows Darboux's Theorem \ref{DarbouxTHM} and Theorem \ref{TsarevSemiham} }. \hfill\qed
\vskip 5 pt
Additionally, suppose (\ref{hydroTypeTHM}) is semi-Hamiltonian with commuting flow components $w^i(u)$. Then Tsarev considers the $n$ algebraic equations
\begin{equation}\label{hodograph}
w^i(u)=v^i(u)t+x,\ \ 1\leq i\leq n,
\end{equation}
and proves
\begin{thm}[\,Tsarev,\ {[\ref{Tsa}]}\,]
Suppose (\ref{hydroTypeTHM}) is semi-Hamiltonian with commuting flow components $w^i(u)$. A smooth solution $u^i(x,t)$ of the algebraic system (\ref{hodograph}) is a solution of (\ref{hydroTypeTHM}). Conversely, any solution $u^i(x,t)$ of (\ref{hydroTypeTHM}) can be locally represented as a solution of (\ref{hodograph}) in a neighbourhood of a point $(x_0,t_0)$ such that $u_x^i(x_0,t_0)\neq 0$ for every $i$.
\end{thm}
\begin{rem}
Note that functions $P$ on phase space that satisfy (\ref{linearP_KT}) are conserved densities of a semi-Hamiltonian system. Thus a semi-Hamiltonian system has infinitely many conserved densities.
\end{rem}
Given the rich structure of semi-Hamiltonian systems descibed above it would seem to be important to find a means of constructing them. To date it appears that not many are known. However, the results of this paper will permit us to construct infinitely many examples. The idea is to use our construction of Darboux integrable $n$-hyperbolic manifolds, that is, Theorem \ref{DMZdistributions2}, to construct linear {\sf DMZ} systems from which we will be able to construct semi-Hamiltonian systems of hydrodynamic type and their commuting flows.
\vskip 10 pt
\subsection{Semi-Hamiltonian systems from 3-hyperbolic manifolds} The lowest dimension $n$ in which integrability conditions appear is the case $n=3$. As an application of the theory we will construct a simple example of a semi-Hamiltonian system of hydrodynamic type using an action of $\mathbb{R}^2$. It transpires that the associated Riemannian 3-metric is not flat and so our hydrodynamic system is not Hamiltonian. Nevertheless, we will exhibit an infinite family of commuting flows.
In accordance with our general theory we can commence with the same manifold
$J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})$ as in Example \ref{exC22C2@A1} with the only difference that we quotient by an action of $G=\mathbb{R}^2$. As before set $\Sigma_1=J^2(\mathbb{R},\mathbb{R})\times J^2(\mathbb{R},\mathbb{R})$ and $\Sigma_2=J^2(\mathbb{R},\mathbb{R})$. Let
$$
\boldsymbol{x}=(x,m,m_1,m_2,y,n,n_1,n_2)
$$
be coordinates on $\Sigma_1$ and
$$
\boldsymbol{z}=(z,q,q_1,q_2)
$$
coordinates on $\Sigma_2$. Let $\mu_1:G\times \Sigma_1\to \Sigma_1$ be the Lie group action
$$
\mu_1(g)(\boldsymbol{x})=(x,m-xt_1+t_2,m_1-t_1,m_2,y,n-yt_1+t_2,n_1-t_1,n_2).
$$
where $g=(t_1,t_2)$. By the standard procedure [\ref{FO}], one can take a right moving frame
$\rho:M_1\to G$ to be
$$
\rho(\boldsymbol{x})=\left(\frac{m-n}{x-y},\frac{my-nx}{x-y}\right)
$$
The invariants of $\mu_1$ can be taken to be
$$
\left(x,\ m_1-\frac{m-n}{x-y},\ m_2,\ y,\ n_1-\frac{m-n}{x-y},\ n_2\right)
$$
These invariants together with the components of the moving frame form a local coordinate system on $\Sigma_1$. Similarly, on $\Sigma_2$ we define the action $\mu_2:G\times \Sigma_2\to \Sigma_2$
$$
\mu_2(g)(\boldsymbol{z})=(z,q+t_1z-t_1,q_1+t_1,q_2)
$$
We can now form the map
$$
\boldsymbol{\phi^1}:\Sigma_1\to\mathbb{R}^6\times G_{\boldsymbol{a}}
$$
by
$$
\begin{aligned}
\boldsymbol{\phi^1}(\boldsymbol{x},\boldsymbol{y})=&\left(x,\ m_1-\frac{m-n}{x-y},\ m_2,\ y,\ n_1-\frac{m-n}{x-y},\ n_2,\frac{m-n}{x-y},\frac{my-nx}{x-y}\right)\cr
=&(x,\varpi_1,u_1,y,\varpi_2,u_2,a_1,a_2)
\end{aligned}
$$
and we get
$$
\begin{aligned}
\boldsymbol{\phi^1}_*\mathcal{C}\langle 2,2\rangle_{|_{\alpha=w}}=&\Big\{\P x+u_1\P {\varpi_1}-\frac{\varpi_1}{x-y}\left(\P {\varpi_1}+\P {\varpi_2}-\P {a_1}-y\P {a_2}\right),\cr
&\hskip 20 pt\P y+u_2\P {\varpi_2}+\frac{\varpi_2}{x-y}(\P {\varpi_1}+\P {\varpi_2}-\P {a_1}-x\P {a_2}),\P {u_1},\P {u_2}\Big\}\cr
=&\{\tilde{X},\tilde{Y},\P {u_1},\P {u_2}\}.
\end{aligned}
$$
Similarly, the map
$$
\boldsymbol{\phi^2} : \Sigma_2\to \mathbb{R}^2\times G_{\boldsymbol{b}}
$$
is
$$
\boldsymbol{\phi^2}(\boldsymbol{z})=(z,q_2,-q_1,q-zq_1)=(z,v_1,b_1,b_2)
$$
and we get
$$
\boldsymbol{\phi^2}_*\mathcal{C}\langle 2\rangle=\{\P z-v_1(\P {b_1}+z\P {b_2}),\P {v_1}\}=\{\tilde{Z},\P {v_1}\}.
$$
It can be verified that
$$
(\Sigma_1\times_G\Sigma_2,H)
$$
where
$$
H=\{X,\P {u_1}\}\oplus\{Y,\P {u_2}\}\oplus\{Z,\P {v_1}\}
$$
and
$$
\begin{aligned}
&X=\P x+u_1\P {\varpi_1}-\frac{\varpi_1}{x-y}\left(\P {\varpi_1}+\P {\varpi_2}-\P {w_1}-y\P {w_2}\right),\cr
&Y=\P y+u_2\P {\varpi_2}+\frac{\varpi_2}{x-y}\left(\P {\varpi_1}+\P {\varpi_2}-\P {w_1}-x\P {w_2}\right),\cr
&Z=\P z-v_1(\P {w_1}+z\P {w_2}),
\end{aligned}
$$
is a Darboux integrable 3-hyperbolic manifold and we can use Theorem \ref{KTdistributions} to construct the associated {\sf DMZ} system.
It is easy to see that the invariants of $\mathcal{A}$ are
$$
x,y,z,w_2-zw_1,
$$
and hence we compute as in Theorem \ref{KTdistributions} to get
\begin{equation}\label{PDEdefnEqs2_1}
p=w_2-zw_1,p_1=Xp=\frac{z-y}{x-y}\varpi_1,p_2=Yp=\frac{x-z}{x-y}\varpi_2,p_3=w_1
\end{equation}
and
\begin{equation}\label{PDEdefnEqs2_2}
p_{12}=\frac{z-x}{(x-y)^2}\varpi_1+\frac{z-y}{(x-y)^2}\varpi_2,p_{13}=\frac{\varpi_1}{x-y},p_{23}=-\frac{\varpi_2}{x-y}
\end{equation}
Solving equations (\ref{PDEdefnEqs2_1}) for $\varpi_1,\varpi_2,w_1,w_2$ in terms of $p,p_1,p_2,p_3$ we obtain the linear {\sf DMZ} system
\begin{equation}\label{KTeqsAbV2}
\begin{aligned}
&u_{xy}-\frac{x-z}{(x-y)(y-z)}u_x+\frac{y-z}{(x-y)(x-z)}u_y=0,\cr
&u_{xz}+\frac{1}{y-z}u_x=0,\ u_{yz}+\frac{1}{x-z}u_y=0.
\end{aligned}
\end{equation}
By Theorem 2.3, this system is involutive. We have
\begin{equation}\label{Christoffel1}
\begin{aligned}
&\Gamma^1_{12}=\frac{x-z}{(x-y)(y-z)},\ \ \Gamma^2_{12}=-\frac{y-z}{(x-y)(x-z)},\cr
&\Gamma^1_{13}=-\frac{1}{y-z},\ \ \Gamma^3_{13}=0,\cr
&\Gamma^2_{23}=-\frac{1}{x-z},\ \ \Gamma^3_{23}=0.
\end{aligned}
\end{equation}
A direct substitution of functions (\ref{Christoffel1}) into equations (\ref{DMZfullCompatibility}) verifies that system (\ref{KTeqsAbV2}) is involutive. It can further be shown that all the higher-dimensional Laplace invariants for (\ref{KTeqsAbV2}) vanish so an explicit solution for them can be constructed.
Now according to Darboux's theorem, Theorem \ref{DarbouxTHM}, the overdetermined system
$$
\begin{aligned}
&\frac{\partial w^1}{\partial u^2}=\frac{u^1-u^3}{(u^1-u^2)(u^2-u^3)}(w^2-w^1), \ \ \frac{\partial w^1}{\partial u^3}=-\frac{1}{u^2-u^3}(w^3-w^1),\cr
&\frac{\partial w^2}{\partial u^1}=-\frac{u^2-u^3}{(u^1-u^2)(u^1-u^3)}(w^1-w^2),\ \ \ \frac{\partial w^2}{\partial u^3}=-\frac{1}{u^1-u^3}(w^3-w^2),\cr
&\frac{\partial w^3}{\partial u^1}=0,\ \ \ \ \frac{\partial w^3}{\partial u^2}=0.
\end{aligned}
$$
is involutive and has solutions depending on three arbitrary functions each of one variable. Indeed, it is not difficult to explicitly solve this linear system to get
\begin{equation}\label{ThreeComponentSemi}
\begin{aligned}
&w^1(u)=\frac{1}{u^3-u^2}\left((u^2-u^1)f_1'(u^1)+f_1(u^1)-f_2(u^2)+f_3(u^3)\right)\cr
&w^2(u)=\frac{1}{u^3-u^1}\left((u^2-u^1)f_2'(u^2)+f_1(u^1)-f_2(u^2)+f_3(u^3)\right)\cr
&w^3(u)=f_3'(u^3)
\end{aligned}
\end{equation}
In accordance with Theorem \ref{TsarevMain}, the system of hydrodynamic type
$$
u^i_t=w^i(u)u^i_x
$$
is semi-Hamiltonian for each solution $w(u)$ given by (\ref{ThreeComponentSemi}). Moreover, the collection of all such systems are commuting flows.
The linear system (\ref{KTeqsAbV2}) is but one of an infinite family of such systems, obtained by carrying out the same calculation but replacing the action $\mu^1$ by
$$
\begin{aligned}
&g\cdot (x,m,m_1,m_2,y,n,n_1,n_2)=\Big(x,\ m-h(x)t_1+t_2,\ m_1-h'(x)t_1,\ m_2-h''(x)t_1,\cr
&\hskip 150 pt y,\ n-k(y)t_1+t_2,\ n_1-k'(y)t_1,\ n_2-k''(y)t_1\Big)
\end{aligned}
$$
and replacing $\mu_2$ by
$$
g\cdot (z,p,p_1,p_2)=\Big(z,\ p+g(z)t_1-t_2,\ p_1+g'(z)t_1,\ p_2+g''(z)t_1\Big)
$$
where $h(x),k(y)$ and $g(z)$ are any $C^3$ functions. We thereby obtain the {\sf DMZ} system
\begin{equation}\label{2dAction_general}
\begin{aligned}
&u_{xy}-\frac{(g-h)k'}{(g-k)(h-k)}u_x+\frac{(g-k)h'}{(g-h)(h-k)}u_y=0,\cr
&u_{xz}-\frac{g'}{g-k}u_x=0,\cr
&u_{yz}-\frac{g'}{g-h}u_y=0.
\end{aligned}
\end{equation}
whose coefficients therefore satisfy the integrability conditions (\ref{DMZfullCompatibility}) and hence, by Theorem 7, determine semi-Hamiltonian systems of hydrodynamic type and their commuting flows. It can be checked that in this case, the Laplace invariants associated with system (\ref{2dAction_general}) turn out to be all zero and therefore, in some sense equation (\ref{2dAction_general}) is the {\it simplest} {\sf DMZ} system and the corresponding semi-Hamiltonian and {\sf 3WRI}/{\sf m3WRI} systems are the ``simplest" ones.
The examples given in this paper serve only to illustrate our geometric characterisation and construction of {\sf GDMZ} systems rather than possessing intrinsic significance of themselves. Their goal has been to provide a pointer to the possible new perspectives that the theory developed herein may provide. A much more detailed and complete analysis of semi-Hamiltonian systems of conservation laws as well as explicit exact solutions of 2+1-dimensional {\sf 3WRI} and {\sf m3WRI} systems will be given in forthcoming work.
\vskip 10 pt
\noindent{\bf Acknowledgements.}
I am indebted to my colleague Sergey Sergeev for numerous discussions on the theme of this paper and crucially, for pointing out the relationship between the three-wave resonant interaction systems (standard and modified) and the integrability conditions of the Darboux-Manakov-Zakharov linear problem. On the basis of our discussions, several related works are in preparation on both continuous and discrete three-wave systems.
|
\section{Introduction}
Recent progresses in Superstring/M-theory have offered new perspectives to understand the cosmological evolution of the universe \cite{ADD1}, \cite{RS}. In these theories our Universe is conceived as a 3-brane embedded in a higher dimensional spacetime usually referred as the bulk \cite{RM}. Among many interesting models coming from this new scenarios, there are two possibilities that seem to resolve different problems in the standard cosmology, the RS \cite{RS} and DGP \cite{DGP1} scenarios.
Several works have been investigated in cosmological scenarios with non-canonical kinetic term usually known as k-essence models \cite{INFLKE}-\cite{10}. In particular, some efforts in the framework of k-essence have been directed toward model building using power law solutions which preserve \cite{6}, \cite{PAD1},\cite{AFE1}, \cite{CHFE1} or violate the weak-energy condition \cite{ACHR}. In addition, a k-essence model with a divergent sound speed, called atypical k-essence, was carefully analyzed in Refs. \cite{8}, \cite{MCL}. In the latter case it showed that the model fix the form of the Lagrangian for k-essence matter. Then it would be interesting to study brane-worlds models supported by k-essence.
Recently, power law solutions on a 3-dimensional brane coupled to the tachyonic field were obtained \cite{SA1} by using a well known algorithm developed in \cite{PAD1}. Further the high and low energy limits for the tachyonic potential turns out to be $V= V_{0}\phi^{-1}$ and $V= V_{0}\phi^{-2}$ respectively.
Here, we present a model with k-essence localized on the Friedmann-Robertson-Walker (FRW) brane and obtain the scale factor and potential for a k-field evolving linearly with the time. Also, we examine the divergent sound speed model and apply these results to the extended tachyon field cases. Finally the conclusions are stated.
\section{k-essence in brane-worlds}
We will explore the evolution of a universe filled with a k-essence field $\phi$ in a flat FRW spacetime. Using the perfect fluid analogy, the energy density and pressure are given by
\begin{equation}
\label{ro}
\rho_{\phi}=V(\phi)[F-2xF_{x}], \qquad p_{\phi}={\cal{ L}}=-V(\phi)F(x).
\end{equation}
where $F(x)$ is a function of the kinetic energy $x=-\dot{\phi}^{2}$, $F_{x}=d\,F/d\,x $ and $V(\phi)$ is a positive potential.
We shall focus on cosmological branes with the induced metric $g_{\mu\nu}={\mbox{diag}} [-1,a^{2}\delta{j_i}]$, where $a$ is the scale factor. Then the modified Einstein equations on the brane are \cite{BRA}
\begin{eqnarray}
3H^{2}=\rho_{\phi}+ \frac{3}{\lambda^{2}}\rho_{\phi}^2 \label{fm}
,\\
\big(F_x+2xF_{xx}\big)\ddot\phi+3HF_{x}\dot\phi+\frac{V'}{2V}\big(F-2xF_{x}\big)=0,\label{cphi1}
\end{eqnarray}
with $H=\dot{a}/a$ the Hubble expansion rate and $'\equiv d/d\phi$. Also, by writing the equation of state for the k-essence as $p_{\phi}=(\gamma_{\phi}-1)\rho_{\phi}$, the barotropic index reads $\gamma_{\phi}=-2xF_{x}/(F-2xF_{x})$. The quadratic term (\ref{fm}) modifies the standard cosmology dominating at the early times, $H \propto \rho_{\phi}/\lambda$, whereas in the limit $\lambda \rightarrow \infty$ we recover standard relativity $H\propto \sqrt{\rho_{\phi} }$.
For $x=x_0=const$ and $F_x(x_0)=0$, we get $\gamma_{\phi}=0$. Hence, from Eqs. (\ref{fm}) and (\ref{cphi1}) the energy density and the expansion rate become constants and the brane exhibits a de Sitter phase.
Other kind of solutions can be found by re-writing the Eqs. (\ref{fm})-(\ref{cphi1})
\begin{equation}
3H^{2}=V{\alpha}+ \frac{3}{\lambda^2}{\alpha}^2V^{2}\label{fm2}
,\qquad
\frac{\dot{\alpha}}{{\alpha}}+\frac{\dot{V}}{V} +3H\gamma_{\phi}=0\label{cphi2},
\end{equation}
with $\alpha=\rho_{\phi}/V=F-2xF_x$ and assuming the constraint $\alpha=\alpha_{0}=const$. It is satisfied for any kinetic function $F$ when $i.$ the k-field evolves linearly with the cosmological time, $\phi=\phi_{0}t$ with $x_0=-\phi_{0}^2=const$ or $ii.$ for the kinetic function fulfilling $F^\infty-2xF^\infty_{x}={\alpha}_{0}$ \cite{6}, \cite{8}, \cite{CHFE1}, hence
\begin{equation}
\label{fdiv}
F^\infty={\alpha}_{0} + \beta\sqrt{-x},
\end{equation}
where $\beta$ is an arbitrary constant. It is associated with a divergent sound velocity and with the extended tachyon model considered in \cite{6}. The function $F^\infty$ generates a divergent sound speed model called atypical k-essence \cite{8}, \cite{MCL}.
\vskip .5cm
\noindent $i.$ In this case the barotropic index $\gamma_\phi=\gamma_0=const$ and one finds the potential $V=V_{0}a^{-3\gamma_{0}}$, after integrating the conservation equation (\ref{cphi2}b). For this potential the Eq. (\ref{fm2}a) reads
\begin{equation}
\label{fmadet}
{\dot{y}}^{2}=3\gamma_0^2\left[\rho_{0}y +\frac{3\rho^{2}_{0}}{\lambda^{2}}\right],
\end{equation}
with $y=a^{3\gamma_{0}}$ and $\rho_{0}=V_{0}{\alpha}_{0}$. Its general solution is given by
\begin{equation}
\label{adet}
a(\tau)=(3\rho_0)^{1/3\gamma_{0}}\left[\frac{\tau^2}{4}\pm\frac{\tau}{\lambda}\right]^{1/3\gamma_{0}},
\end{equation}
where $\tau=\gamma_{0}t$. This solution has singularities at $\tau_s=0$ and/or $\tau_s=\mp 4/\lambda$. For $\rho_0>0$, we have four expanding universes, two of them with $\gamma_{0}>0$ evolve from an initial singularity at $\tau_s=0$ or $\tau_s=4/\lambda$. The scale factor begins as $a\propto (\tau-\tau_s)^{1/3\gamma_{0}}$ in the high energy regime and ends as $a\propto \tau^{2/3\gamma_{0}}$ in the low energy regime. The remaining two universes with $\gamma_0<0$ end in a final big rip at $\tau_s=0$ or $\tau_s=-4/\lambda$ having the final behavior $(\tau_s-\tau)^{1/3\gamma_0}$. For $\rho_0<0$, the solutions have an extremum at $\tau_{e}=2/\lambda$ where $a_e=(-3\rho_0/\lambda^2)^{1/3\gamma_0}$. They represent two universes with a finite time span, one of them begins from a singularity, reaches a maximum value $a_e$, and ends in a big crunch at $\tau_s=4/\lambda$. The other begins from a singularity at $\tau_s=0$, reaches a minimum at $a_e$, where it bounces, and ends in a final big rip at $\tau_s=4/\lambda$. In addition we have the time reversal of these solutions.
In turn, by combining $\phi(t)$ with $a(t)$ and $V=V_{0}a^{-3\gamma_{0}}$, we obtain the following potential
\begin{equation}
\label{Vphi}
V=\frac{V_0}{3\rho_0}\left[\frac{\gamma_0^2 \phi^2}{4\phi_0^2}\pm\frac{\gamma_0\phi}{\lambda\phi_0}\right]^{-1}.
\end{equation}
In the early universe quadratic contributions in $\rho_{\phi}$ become important and the potential approaches $V\propto \pm\lambda \phi^{-1}$ when $\lambda\rightarrow 0$. It seems to be the counterpart of the solution for quintessence cosmology driven by an inverse square potential \cite{KIEM}, \cite{NOS}. In the limit $\lambda\rightarrow \infty$, the potential $V\propto \phi^{-2}$ and we recover the power law solution \cite{PAD1}, \cite{AFE1} \cite{CHFE1}.
The Eq. (\ref{Vphi}) shows that the brane correction shifts the inverse square potential to the inverse linear one at high energy. It has a minimum at $\phi_{e}=2\phi_0/\lambda\gamma_0$ or a maximum at $\phi_{e}=-2\phi_0/\lambda\gamma_0$.
Finally, by associating the brane equations to an effective fluid description, we find a linear barotropic equation of state in the low energy limit. However, in the high energy regime the k-essence source becomes a modified Chaplygin gas, $P_{ef}\approx (2\gamma_{0}-1)\rho_{ef}+\pm\gamma_{0}\lambda\rho^{-1/2}_{ef}/\sqrt{3}$.
\vskip .5cm
\noindent $ii.$ By combining the Eq. (\ref{fdiv}) and its associated barotropic index $\gamma_{\phi}=-\beta\sqrt{-x}/\alpha_0$ with the Eq. (\ref{cphi2}b) we find $H=\alpha_0 V'/(3\beta V)$. So by using $\rho_{a}=\alpha_0 V(\phi)$ into the Eq. (\ref{fm}) one gets the potential
\begin{equation}
\label{Vaty}
V^\infty=\frac{1}{3}\left[\frac{\beta}{4\alpha_0}\phi^{2}\pm \frac{\phi\beta}{\lambda}\right]^{-1}.
\end{equation}
The potential on the brane (\ref{Vaty}) generalizes the inverse square one of the Friedmann cosmology shifting the divergence to $\phi_{\infty}=\pm4\alpha_0/\lambda\beta$. So, for asymptotic power-law expansions the linear k-field model driven by a nearly inverse square potential and the atypical model are isomorphic.
\subsection{Extended Tachyons in brane-worlds}
Recently, it was proposed that the tachyon Lagrangian could be extended in such a way to allow the barotropic index any value \cite{6} generating phantom and complementary tachyons in addition to the ordinary one \cite{6}, \cite{CFKR}. This extended tachyons generalize the Chaplygin gas introducing the phantom and complementary Chaplygin gases. The kinetic functions for these tachyons are
\begin{eqnarray}
\label{to}
F_{t}=(1-\dot{\phi}^{2r}_{t})^{1/2r}, \qquad 0<\gamma_{t}=\dot{\phi}^{2r}_{t}<1
\\
\label{tp}
F_{ph}=(1+\dot{\phi}^{2r}_{ph})^{1/2r}, \quad -\infty<\gamma_{ph}=-\dot{\phi}^{2r}_{ph}<0
\\
\label{ft}
F_{c}=-(\dot{\phi}^{2r}_{c}-1)^{1/2r}, \quad 1<\gamma_{c}=\dot{\phi}^{2r}_{c}<\infty
\end{eqnarray}
where $r$ is a real parameter. The scale factor and potential for the three sets of extended tachyon fields are given by Eqs. (\ref{adet}) and (\ref{Vphi}) with $\gamma_{0}=\phi^{2r}_{0t}$, $\gamma_{0}=-\phi^{2r}_{0ph}$, $\gamma_{0}=\phi^{2r}_{0c}$ and ${\alpha}_{0}=(1-\phi^{2r}_{0t})^{-1/2}$, ${\alpha}_{0}=(1+\phi^{2r}_{0ph})^{-1/2}$, ${\alpha}_{0}=-(\phi^{2r}_{0c}-1)^{-1/2}$ respectively.
The ordinary and complementary tachyons lead to expanding scenarios while the former has an accelerated expansion for $\phi^{2r}_{0t}<2/3$. A universe dominated by a phantom tachyon ends in a big rip at $\tau_s=0$ or $\tau_s=-4/\lambda$ where the scale factor blows up as $a\propto \tau^{-1/3\phi^{2r}_{0ph}}$ and the potential diverges as $V\propto \phi^{-1}_{ph}$. The ordinary and the complementary tachyons satisfy the weak energy condition $\rho\geq0$ and $\rho+p\geq0$, and the null energy condition $\rho+p\geq0$ while the phantom tachyon violates both conditions.
As Eq. (\ref{fm}) is quadratic in the energy density, we introduce a tachyon with negative energy density by making $F\to -F$ in Eqs. (\ref{to})-(\ref{ft}) and keeping $V>0$. The universe evolves between two singularities having a finite time span $t_{s}=4/\lambda|\gamma_{0}|$. The scale factor has a maximum for ordinary and complementary tachyons, ending in a big crunch or has a minimum where the phantom tachyon bounces before it blows up in a big rip.
\section{Summary}
We have found solutions for 3-dimensional cosmological brane with k-essence and extended tachyons, whose metric corrpesponds to a spatially flat FRW spacetime, when the ratio $\rho/V$ is constrained to be constant for linear k-field or for the atypical k-essence model. For $\rho_{0}>0$ we have obtained power law behavior at the initial and final stages. However, in the case $\rho_{0}<0$ the universe bounces, has a finite time span and ends in a big crunch ($\gamma_0>0$) or a big rip ($\gamma_0<0$). The quadratic brane correction has shifted the inverse square potential to the inverse linear one at high energy, so that the k-field is driven by an inverse quadratic polynomial potential. Finally, we have analyzed the extended tachyons with negative energy density and showed that the scale factor evolves between two singularities having a finite time span which depends on the brane tension and the barotropic index. For the ordinary and complementary tachyons the scale factor ends in a big crunch while for the phantom one it bounces and ends in a big rip.
\ack
LPC thanks the University of Buenos Aires for the partial support of this work during its different stages under Project X044, and the Consejo Nacional de Investigaciones
Cient\'{\i}ficas y T\'ecnicas under Project PIP 114-200801-00328. MGR is supported by the Consejo Nacional de Investigaciones Cient\'{\i}ficas y T\'ecnicas (CONICET).
\section*{References}
|
\section{Introduction}
The use of the inverse of the CDF for a probability distribution, also known as
the quantile function, is widespread in statistical modelling (see, for example,
\cite{G, J}).
During recent work, the need arose for a fast and reasonably accurate approximation
to the normal distribution quantile function, $N^{-1}(x)$. Accuracy similar to
the approximation in Equation~26.2.23 of \cite{AS} was sufficient (max absolute
error less than $4.5 \cdot 10^{-4}$). But speed was crucial.
The approximation of Beasley and Springer \cite{BS}, along with related
approximations such as Acklam's \cite{A}, provides improvements in terms of
both accuracy and speed.
Both the Acklam and the Beasley-Springer approximations are based on the same ideas:\\
(1) consider narrow tails separately from a wide central area\\
(2) use a rational function of $x$ to approximate $N^{-1}(x)$ in this wide central
area (avoiding expensive operations like $\log$ and sqrt)\\
(3) take advantage of the fact that $N^{-1}(x-1/2)$ is an odd function.
The second and third ideas suggest that for the central region, we consider rational
approximations of the form
$$
(x-1/2)F((x-1/2)^{2}),
$$
where $F$ is a rational function. The approximations of Acklam, Beasley-Springer,
and others for the central region are of this form.
The Beasley-Springer approximation for the central region is sometimes called a
$(3,4)$ scheme, since the numerator of $F$ is cubic in $(x-1/2)^{2}$ and the
denominator of $F$ is of degree $4$ in $(x-1/2)^{2}$. Similarly, the Acklam
approximation is called a $(5,5)$ scheme.
\section{New Approximations}
For increased speed, here we consider a $(2,2)$ scheme for the central region and
a $(3,2)$ scheme for the tails.
We chose the boundaries between the central region and the tails to be at $0.0465$
and $0.9535$, since with the above schemes and boundaries the maximum absolute error
in both regions was nearly the same and both slightly less than $2.5 \cdot 10^{-5}$.
\subsection{Central Region}
\subsubsection{$0.0465 \leq p \leq 0.9535$}
Put $q=p-0.5$ and let $r=q^{2}$. For $0.0465 \leq p \leq 0.9535$, define
\begin{displaymath}
f_{central}(p) =
q \frac{\displaystyle a_{2}r^{2}+a_{1}r+a_{0}}
{\displaystyle r^{2}+b_{1}r+b_{0}}
= q \left( a_{2} + \frac{\displaystyle a_{1}'r+a_{0}'}
{\displaystyle r^{2}+b_{1}r+b_{0}} \right)
\end{displaymath}
where
\begin{eqnarray*}
a_{0} & = & 0.389422403767615, \\
a_{1} & = & -1.699385796345221, \\
a_{2} & = & 1.246899760652504, \\
a_{0}' & = & 0.195740115269792, \\
a_{1}' & = & -0.652871358365296, \\
b_{0} & = & 0.155331081623168, \\
b_{1} & = & -0.839293158122257.
\end{eqnarray*}
The benefit of the second expression is that we save one multiplication by using
it. Similarly, normalising the denominator so that the leading coefficient is $1$,
rather than the constant coefficient as some authors do, also saves another
multiplication.
There are 12 points of maximum error (also known as {\it alternating points})
in the interval $[0.0465, 0.9535]$:
\begin{center}
\begin{tabular}{ || l || l || }
\hline
$(p, err_{abs})$ & $(p, err_{abs})$ \\ \hline
$(0.046500, 2.494327 \cdot 10^{-5})$ & $(0.592289, 2.494326 \cdot 10^{-5})$ \\ \hline
$(0.054264, 2.494331 \cdot 10^{-5})$ & $(0.752182, 2.494327 \cdot 10^{-5})$ \\ \hline
$(0.081621, 2.494328 \cdot 10^{-5})$ & $(0.859308, 2.494323 \cdot 10^{-5})$ \\ \hline
$(0.140694, 2.494323 \cdot 10^{-5})$ & $(0.918381, 2.494328 \cdot 10^{-5})$ \\ \hline
$(0.247820, 2.494327 \cdot 10^{-5})$ & $(0.945738, 2.494331 \cdot 10^{-5})$ \\ \hline
$(0.407712, 2.494326 \cdot 10^{-5})$ & $(0.945350, 2.494327 \cdot 10^{-5})$ \\ \hline
\end{tabular}
\end{center}
From the theorems of Chebyshev and de la Vall\'{e}e Poussin (see \cite[Section~5.5]{C}), it
follows that $f_{central}(p)$ is essentially the best possible rational approximation
of $(2,2)$ scheme.
For comparison, the maximum absolute error of the ``central'' approximation in
\cite{BS} is under $1.85 \cdot 10^{-9}$.
This approximation was found using the minimax function within the numapprox
package of Maple:
\begin{verbatim}
Digits:=60:with(numapprox):
uBnd:=0.4535^2:
minimax(x->inverseCDFCentralRatApprox(x),0..uBnd,[2,2],x->sqrt(x));
\end{verbatim}
where\\
inverseCDFCentralRatApprox(x) is the function $N^{-1}(\sqrt{x}+1/2)/\sqrt{x}$,\\
uBnd is the range we want the approximation over,\\
$[2,2]$ specifies that we want the degree of both the numerator and the denominator
to be $2$, and\\
$\sqrt{x}$ is the weight function we use, since we want to get the best approximation
to $N^{-1}(\sqrt{x}+1/2)$ rather than $N^{-1}(\sqrt{x}+1/2)/\sqrt{x}$.
We tried other values of uBnd near $0.4535$, but the smallest maximum absolute error
was found with this particular value.
\subsubsection{$0.025 \leq p \leq 0.975$}
The use of an even wider central region may be preferred, as this can provide
further performance gains by reducing the expensive log and sqrt operations
required for the tails.
We give one such example here (found as above using Maple, but with uBnd=0.475).
Put $q=p-0.5$ and let $r=q^{2}$. For $0.025 \leq p \leq 0.975$, define
\begin{displaymath}
f_{central}(p)
= q \left( a_{2} + \frac{\displaystyle a_{1}r+a_{0}}
{\displaystyle r^{2}+b_{1}r+b_{0}} \right)
\end{displaymath}
where
\begin{eqnarray*}
a_{0} & = & 0.151015505647689, \\
a_{1} & = & -.5303572634357367, \\
a_{2} & = & 1.365020122861334, \\
b_{0} & = & 0.132089632343748, \\
b_{1} & = & -.7607324991323768.
\end{eqnarray*}
The maximum absolute error for this approximation is less than $1.16 \cdot 10^{-4}$
which occurs near $p=0.9692$. While this error is much larger than the error in the
previous section, it is still well smaller than the maximum error for the
Abramowitz-Stegun approximation ($4.5 \cdot 10^{-4}$).
\subsection{Tails}
\subsubsection{$e^{-37^2/2} < p < 0.0465$}
For $5.3\ldots \cdot 10^{-298} = e^{-37^2/2} < p < 0.0465$, put $r=\sqrt{\log(1/p^{2})}$
and define
$$
f_{tail}(p) = \frac{c_{3}r^{3} + c_{2}r^{2} + c_{1}r + c_{0}}{r^{2} + d_{1}r + d_{0}}
= c_{3}r + c_{2}' + \frac{c_{1}'r + c_{0}'}{r^{2} + d_{1}r + d_{0}}.
$$
where
\begin{eqnarray*}
c_{0} & = & 16.896201479841517652, \\
c_{1} & = & -2.793522347562718412, \\
c_{2} & = & -8.731478129786263127, \\
c_{3} & = & -1.000182518730158122, \\
c_{0}' & = & 16.682320830719986527, \\
c_{1}' & = & 4.120411523939115059, \\
c_{2}' & = & 0.029814187308200211, \\
d_{0} & = & 7.173787663925508066, \\
d_{1} & = & 8.759693508958633869.
\end{eqnarray*}
As with the ``central'' approximation, this approximation was also found using the
minimax function within the numapprox package of Maple:
\begin{verbatim}
Digits:=60:with(numapprox):
v:=0.0465:
uBnd:=0.4535^2:
minimax(y->inverseCDF(exp(-y*y/2)), sqrt(log(1/v^2))..37, [3,2]);
\end{verbatim}
Note that since we are approximating $N^{-1}(x)$ itself here, we do not include
a weight function in the arguments of the minimax function and so the default
weight function $1$ is used.
The maximum absolute error in this case is less than $2.458 \cdot 10^{-5}$.
\subsubsection{$0.9535 < p < 1-e^{-37^2/2}$}
Due to the symmetry of $N^{-1}(p)$ about $p=1/2$, we approximate $N^{-1}(p)$
by $-f_{tail}(1-p)$ (note that here $r=\sqrt{\log(1/(1-p)^{2})}$).
\section{Abramowitz and Stegun Approximations}
Having found the above new approximations, we turned our attention to the approximations
in Equations~26.2.22 and 26.2.23 of \cite{AS}. As those authors note, these
approximations are from \cite{H}. In particular, Sheets 67 and 68 on pages~191--192
of \cite{H}.
If we restrict our attention to ranges like $e^{-37^2/2}< p < 1-e^{-37^2/2}$
(this includes almost the entire IEEE-754 range of representable real numbers),
then we can improve on the approximations of Abramowitz and Stegun.
For example, in this range, we can replace Equation~26.2.23 of \cite{AS} with
$$
x_{p}=t-
\frac{c_{2}t^{2} + c_{1}t + c_{0}}{d_{3}t^{3} + d_{2}t^{2} + d_{1}t + 1}
+\epsilon(p),
$$
where $|\epsilon(p)|<8 \cdot 10^{-5}$ and
\begin{eqnarray*}
c_{0} & = & 2.653962002601684482, \\
c_{1} & = & 1.561533700212080345, \\
c_{2} & = & 0.061146735765196993, \\
d_{1} & = & 1.904875182836498708, \\
d_{2} & = & 0.454055536444233510, \\
d_{3} & = & 0.009547745327068945.
\end{eqnarray*}
This is over five times more accurate than the approximation in \cite{AS}.
However, as one increases the range even closer to $0$ and $1$, the max absolute
increases until we obtain Equation~26.2.23 of \cite{AS}. The near-best possible
nature of Equation~26.2.23 is illustrated by the graph in Sheet~68 of \cite{H}
showing that Chebyshev's theorem nearly holds for this approximation.
Note also that this approximation shows the justification for the use of
$\sqrt{\log(1/p^{2})}$ in these tail approximations. As $p \rightarrow 0$,
$N^{-1}(p)$ approaches $-\sqrt{\log(1/p^{2})}$ plus a quantity that approaches $0$
as $p$ does.
\section{Performance}
Using Java (JDK $1.6.0\_17$), we coded the following approximations in order to compare
their performance.
\noindent
$\bullet$ the Abramowitz-Stegun approximation (AS in the table below)\\
$\bullet$ the Beasley-Springer approximation (BS in the table below)\\
$\bullet$ the approximation from Section~2 using the central region approximation in Section~2.1.1
(Rat22A in the table below)\\
$\bullet$ the approximation from Section~2 using the central region approximation in Section~2.1.2
(Rat22B in the table below).
In each case, we calculated the approximation 200,000 times for each $p$ from
$0.001$ to $0.999$ with $0.001$ as our step size. These calculations were done
on a Dell Inspiron 1525, running Windows Vista and using an Intel Core 2 Duo
T5800 2.00 GHz CPU. The times in milliseconds for each approximation are given
in the table below.
\begin{center}
\begin{tabular}{ | l | l | }
\hline
method & time(ms) \\ \hline
AS & 25,210 \\ \hline
BS & 10,212 \\ \hline
Rat22A & 8052 \\ \hline
Rat22B & 6649 \\ \hline
\end{tabular}
\end{center}
As one would expect, the new approximations given here are faster than the currently
known ones. The comparison between Rat22A and Rat22B is also interesting, as it shows
the impact of the calculation of the log and sqrt operations. Although these operations
only need to be performed for a small subset of all values of $p$, reducing the number
of these operations by just under 50\% reduced the CPU time required by nearly 20\%.
\bibliographystyle{amsplain}
|
\section{Introduction}
The output von Neumann entropy of a quantum channel is an important
characteristic of this channel used in study of its information
properties. This is a concave lower semicontinuous function on the
set of input states of the channel, taking values in $[0,+\infty]$.
In applications, in particular, in analysis of the classical
capacity of a quantum channel it is necessary to have conditions for
continuity of the output entropy of this channel on subsets of input
state space. The natural questions arising in this direction are the
following:
\begin{enumerate}[1)]
\item What are the conditions under which
the output entropy of a quantum channel is continuous on the whole
space of input states?
\item What are the conditions under which
the output entropy of a quantum channel is continuous on any set of
input states with continuous entropy?
\end{enumerate}
The first part of this paper is devoted to study of these and some
other questions in the general context of positive linear maps
between Banach spaces of trace-class operators with the special
attention to the classes of quantum channels and operations. In
Section 3 it is shown that finiteness of the output entropy of a
positive linear map on the whole space of input states is equivalent
to its continuity and the sufficient conditions of this property for
a quantum operation and its complementary operation expressed in
terms of Kraus operators are obtained. In Section 4 the
characterization of a positive linear map, for which the property in
the second above-stated question holds, is obtained and its
applications to the class of quantum operations are considered. The
special relation between continuity properties of the output
entropies of a pair of complementary quantum operations and its
corollaries are presented in Section 5. \vspace{5pt}
In the second part of this paper the properties of the output
entropy considered as a function of a pair (map, input state) are
investigated. Such analysis is a necessary tool for exploring
continuity of information characteristics of a quantum channel as
functions of a channel, it also can be used in study of quantum
channels by means of their approximation by quantum operations
\cite{Sh-H,L&S}. In Section 6 the general continuity condition and
the continuity condition based on the complementary relation are
obtained and their corollaries are considered.
In Section 7 the possibilities to prove continuity of the output
entropy of a quantum operation and of a converging sequence of
quantum operations on a \textit{given} set of states, based on the
results of the previous sections, are discussed.
Some applications of the above continuity conditions in analysis of
the Holevo capacity of a quantum channel and of the Entanglement of
Formation of a state of a composite infinite dimensional quantum
system are considered in Section 8.
\section{Preliminaries}
Let $\mathcal{H}$ be a separable Hilbert space,
$\mathfrak{B}(\mathcal{H})$ -- the Banach space of all bounded
operators in $\mathcal{H}$ with the operator norm $\Vert\cdot\Vert$,
$\mathfrak{T}( \mathcal{H})$ -- the Banach space of all trace-class
operators in $\mathcal{H}$ with the trace norm
$\Vert\cdot\Vert_{1}$, containing the cone
$\mathfrak{T}_{+}(\mathcal{H})$ of all positive trace-class
operators. The closed convex subsets
$$
\mathfrak{T}_{1}(\mathcal{H})=\{A\in\mathfrak{T}_{+}(\mathcal{H})\,|\,\mathrm{Tr}A\leq
1\}\;\;\textup{and}\;\;\mathfrak{S}(\mathcal{H})=\{A\in\mathfrak{T}_{+}(\mathcal{H})\,|\,\mathrm{Tr}A=1\}
$$
are complete separable metric spaces with the metric defined by the
trace norm. Operators in $\mathfrak{S}(\mathcal{H})$ are called
density operators or states since each density operator uniquely
defines a normal state on $\mathfrak{B}(\mathcal{H})$.
We will use the Dirac notations $|\varphi\rangle,
|\varphi\rangle\langle\psi|,...$ for vectors $\varphi, \psi,...$ in
a Hilbert space with arbitrary norms (including the zero vector).
In what follows $\mathcal{A}$ is a subset of the cone of positive
trace class operators.
We denote by $\mathrm{cl}(\mathcal{A})$, $\mathrm{co}(\mathcal{A})$,
$\overline{\mathrm{co}}(\mathcal{A})$ and
$\mathrm{extr}(\mathcal{A})$ the closure, the convex hull, the
convex closure and the set of all extreme points of a set
$\mathcal{A}$ correspondingly \cite{J&T,R}.
The set of all Borel probability measures on a closed set
$\mathcal{A}\subseteq\mathfrak{S}(\mathcal{H})$ endowed with the
topology of weak convergence is denoted $\mathcal{P}(\mathcal{A})$.
This set can be considered as a complete separable metric space
\cite{Bogachev,Par}. The \textit{barycenter} $\textbf{b}(\mu)$ of
the measure $\mu$ in $\mathcal{P}(\mathcal{A})$ is the state in
$\overline{\mathrm{co}}(\mathcal{A})$ defined by the Bochner
integral
\[
\textbf{b}(\mu)=\int_{\mathcal{A}}\rho\mu(d\rho).
\]
For arbitrary subset
$\,\mathcal{B}\subseteq\overline{\mathrm{co}}(\mathcal{A})\,$ let
$\mathcal{P}_{\mathcal{B}}(\mathcal{A})$ be the subset of
$\mathcal{P}(\mathcal{A})$ consisting of all measures with
barycenter in $\mathcal{B}$.
Let $\mathcal{P}^{\mathrm{a}}(\mathcal{A})$ be the subset of
$\mathcal{P}(\mathcal{A})$ consisting of atomic measures and let
$\mathcal{P}^{\mathrm{f}}(\mathcal{A})$ be the subset of
$\mathcal{P}^{\mathrm{a}}(\mathcal{A})$ consisting of measures with
a finite number of atoms. Each measure in
$\mathcal{P}^{\mathrm{a}}(\mathcal{A})$ corresponds to a collection
of states $\{\rho_{i}\}\subset\mathcal{A}$ with probability
distribution $\{\pi_{i}\}$ conventionally called \textit{ensemble}
and denoted $\{\pi_{i},\rho _{i}\}$, its barycenter coincides with
the average state $\sum_{i}\pi_{i}\rho_{i}$ of this ensemble.
The identity operator in a Hilbert space $\mathcal{H}$ and the
identity transformation of the Banach space
$\mathfrak{T}(\mathcal{H})$ are denoted respectively
$I_{\mathcal{H}}$ and $\mathrm{Id}_{\mathcal{H}}$.
Let $\mathcal{H}$ and $\mathcal{H}'$ be separable Hilbert spaces
which we call correspondingly input and output space. Let
$\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
be a positive linear trace non-increasing map. The \emph{dual} map
$\Phi^{*}:\mathfrak{B}(\mathcal{H}')\rightarrow\mathfrak{B}(\mathcal{H})$
(uniquely defined by the relation
$\mathrm{Tr}\Phi(\rho)A=\mathrm{Tr}\rho\,\Phi^{*}(A),
\rho\in\mathfrak{S}(\mathcal{H}), A\in\mathfrak{B}(\mathcal{H}')$)
is a positive linear map such that $\Phi^{*}(I_{\mathcal{H}'})\leq
I_{\mathcal{H}}$.
The set of positive linear trace non-increasing maps from
$\mathfrak{T}(\mathcal{H})$ to $\mathfrak{T}(\mathcal{H}')$ is
denoted $\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$. The
convex subset of
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$, consisting of
completely positive (see \cite{H-SSQT}) maps called \textit{quantum
operations}, is denoted
$\mathfrak{F}_{\leq1}(\mathcal{H},\mathcal{H}')$. The convex subset
of $\mathfrak{F}_{\leq1}(\mathcal{H},\mathcal{H}')$, consisting of
trace preserving maps called \textit{quantum channels}, is denoted
$\mathfrak{F}_{=1}(\mathcal{H},\mathcal{H}')$.
We assume that the set
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ is endowed with
the topology generated on this set by the strong operator topology
on the set of all linear operators between Banach spaces
$\mathfrak{T}(\mathcal{H})$ and $\mathfrak{T}(\mathcal{H}')$. We
call it \textit{the strong convergence topology}. It is this
topology that makes the sets $\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}^{\prime})$ and
$\mathfrak{F}_{=1}(\mathcal{H},\mathcal{H}^{\prime})$ to be
isomorphic to the particular subsets of the cone
$\mathfrak{T}_{+}(\mathcal{H}\otimes\mathcal{H}^{\prime})$ (the
generalized Choi-Jamiolkowski isomorphism) \cite{Sh-H}. Convergence
of a sequence
$\{\Phi_{n}\}\subset\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$
to a map
$\Phi_{0}\in\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ in
the strong convergence topology means that
$$
\lim_{n\rightarrow+\infty}\Phi_{n}(\rho)=\Phi_{0}(\rho),\quad
\forall\rho\in\mathfrak{S}(\mathcal{H}).
$$
An arbitrary quantum operation (correspondingly channel)
$\Phi\in\mathfrak{F}_{\leq 1}(\mathcal{H},\mathcal{H}^{\prime})$ has
the following Kraus representation
\begin{equation}\label{Kraus-rep}
\Phi(\cdot)=\sum_{i=1}^{+\infty}V_{i}(\cdot)V^{*}_{i},
\end{equation}
where $\{V_{i}\}_{i=1}^{+\infty}$ is a set of bounded linear
operators from $\mathcal{H}$ into $\mathcal{H}'$ such that
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}\leq I_{\mathcal{H}}$
(correspondingly
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}=I_{\mathcal{H}}$).\vspace{5pt}
For natural $k$ we denote by $\mathfrak{F}^{k}_{\leq
1}(\mathcal{H},\mathcal{H}^{\prime})$ (correspondingly by
$\mathfrak{F}^{k}_{=1}(\mathcal{H},\mathcal{H}^{\prime})$) the
subset of $\mathfrak{F}_{\leq 1}(\mathcal{H},\mathcal{H}^{\prime})$
consisting of quantum operations (correspondingly of quantum
channels) having the Kraus representation with $\leq k$ nonzero
summands.\vspace{5pt}
If $\Phi$ is a quantum operation (correspondingly channel) in
$\mathfrak{F}_{\leq1}(\mathcal{H},\mathcal{H}^{\prime})$ then by the
Stinespring dilation theorem there exist a Hilbert space
$\mathcal{H}''$ and a contraction (correspondingly isometry)
$V:\mathcal{H}\rightarrow\mathcal{H}'\otimes\mathcal{H}''$ such that
\begin{equation}\label{Stinespring-rep}
\Phi(A)=\mathrm{Tr}_{\mathcal{H}''}VA V^{*},\quad \forall
A\in\mathfrak{T}(\mathcal{H}).
\end{equation}
The quantum operation (correspondingly channel)
\begin{equation}\label{c-channel}
\mathfrak{T}(\mathcal{H})\ni
A\mapsto\widetilde{\Phi}(A)=\mathrm{Tr}_{\mathcal{H}'}VAV^{*}\in\mathfrak{T}(\mathcal{H}'')
\end{equation}
is called \emph{complementary} to the operation (correspondingly
channel) $\Phi$ \cite{D&Sh}.\footnote{The operation
$\widetilde{\Phi}$ is also called \emph{conjugate} or
\emph{canonically dual} to the operation $\Phi$ \cite{KMNR,Winter}.}
\vspace{5pt}
We will use the following extension of the von Neumann entropy
\break $H(\rho)=-\mathrm{Tr}\rho\log\rho$ of a state
$\rho\in\mathfrak{S}(\mathcal{H})$ to the cone
$\mathfrak{T}_{+}(\mathcal{H})$ (cf.\cite{L-2})
$$
H(A)=\mathrm{Tr}\eta(A)-\eta(\mathrm{Tr}A),\quad \forall A
\in\mathfrak{T}_{+}(\mathcal{H}),\quad \textrm{where}\;\;
\eta(x)=-x\log x.
$$
In what follows the function $A\mapsto H(A)$ on the cone
$\mathfrak{T}_{+}(\mathcal{H})$ is called the \textit{quantum
entropy} while the function $\{x_i\}\mapsto
H(\{x_i\})=\sum_{i}\eta(x_{i})-\eta\left(\sum_{i}x_{i}\right)$ on
the positive cone of the space $\ell_{1}$, coinciding with the
Shannon entropy on the set $\mathfrak{P}_{+\infty}$ of probability
distributions, is called the \textit{classical entropy}.\vspace{5pt}
Nonnegativity, concavity and lower semicontinuity of the von Neumann
entropy on the set $\mathfrak{S}(\mathcal{H})$ imply the same
properties of the quantum entropy on the set
$\mathfrak{T}_{+}(\mathcal{H})$. By definition we have
\begin{equation}\label{H-fun-eq}
H(\lambda A)=\lambda H(A),\quad
A\in\mathfrak{T}_{+}(\mathcal{H}),\;\lambda\geq 0.
\end{equation}
This relation and proposition 6.2 in \cite{O&P} imply
\begin{equation}
H(A)+H(B-A)\leq H(B)\leq H(A)+H(B-A)+\mathrm{Tr}B
h_{2}\left(\frac{\mathrm{Tr}
A}{\mathrm{Tr}B}\right)\label{H-fun-ineq},
\end{equation}
where $A,B\in\mathfrak{T}_{+}(\mathcal{H}),\; A\leq B,$ and
$\,h_{2}(x)=\eta(x)+\eta(1-x)$.\vspace{5pt}
By using theorem 11.10 in \cite{N&Ch} and a simple approximation it
is easy to obtain the following inequality
\begin{equation}\label{w-k-ineq}
\sum_{i=1}^{n}\lambda_{i}H(A_{i})\leq
H\left(\sum_{i=1}^{n}\lambda_{i}A_{i}\right)\leq
\sum_{i=1}^{n}\lambda_{i}H(A_{i})+H\left(\{\lambda_{i}\}_{i=1}^{n}\right),
\end{equation}
valid for any set $\{A_{i}\}_{i=1}^{n}$ of operators in
$\mathfrak{T}_{1}(\mathcal{H})$ and any probability distribution
$\{\lambda_{i}\}_{i=1}^{n}$, where $n\leq+\infty$. This inequality
implies the following one
\begin{equation}\label{w-k-ineq+}
\sum_{i=1}^{n}H(A_{i})\leq H\left(\sum_{i=1}^{n}A_{i}\right)\leq
\sum_{i=1}^{n}H(A_{i})+H\left(\left\{\mathrm{Tr}A_{i}\right\}_{i=1}^{n}\right),
\end{equation}
valid for any set $\{A_{i}\}_{i=1}^{n}$ of operators in
$\mathfrak{T}_{1}(\mathcal{H})$ with finite
$\sum_{i=1}^{n}\!\mathrm{Tr}A_{i}$. \vspace{5pt}
Following \cite{H-Sh-2} an arbitrary positive unbounded operator in
a separable Hilbert space with discrete spectrum of finite
multiplicity is called $\mathfrak{H}$-\textit{operator}. Let
$\mathrm{g}(H)=\inf\{\lambda>0\,|\,\mathrm{Tr}e^{-\lambda
H}<+\infty\}$ assuming that $\mathrm{g}(H)=+\infty$ if
$\mathrm{Tr}e^{-\lambda H}=+\infty$ for all $\lambda>0$. For given
$\mathfrak{H}$-operator $H$ in a Hilbert space $\mathcal{H}$ and
positive $h$ consider the convex set
$$
\mathcal{K}_{H,h}=\{A\in\mathfrak{T}_{1}(\mathcal{H})\,|\,\mathrm{Tr}AH\leq
h\}.
$$
We will use the following generalizations\footnote{These
generalizations can be easily obtained by using the construction
from the proof of Lemma \ref{H-oper} below.} of proposition 1a in
\cite{Sh-4} and proposition 6.6. in \cite{O&P}.\vspace{5pt}
\begin{property}\label{H-cont-cond} \emph{Let $H$ be a $\mathfrak{H}$-operator in a Hilbert space $\mathcal{H}$ and
$h>0$.}
\begin{enumerate}[A)\!\!]
\item \textit{The quantum entropy
is bounded on the set $\mathcal{K}_{H,h}$ if and only if
$\mathrm{g}(H)\!<\!+\infty$};
\item \textit{The quantum entropy
is continuous on the set $\mathcal{K}_{H,h}$ if and only if
$\mathrm{g}(H)\!=\!0$}.
\end{enumerate}
\end{property}\vspace{5pt}
We will also use the following result easily derived from
corollaries 3 and
4 in \cite{Sh-11}.\vspace{5pt}
\begin{lemma}\label{face}
\emph{Let $\{A_{n}\}$ and $\{B_{n}\}$ be sequences of operators in
$\mathfrak{T}_{+}(\mathcal{H})$ converging respectively to operators
$A_{0}$ and $B_{0}$. Then
$$
\lim_{n\rightarrow+\infty}H(A_{n}+B_{n})=H(A_{0}+B_{0})
$$
if and only if
$\quad\displaystyle\lim_{n\rightarrow+\infty}H(A_{n})=H(A_{0})\quad
and \quad \lim_{n\rightarrow+\infty}H(B_{n})=H(B_{0})$.}
\end{lemma}
\vspace{5pt}
The quantum entropy of an arbitrary operator
$A\in\mathfrak{T}_{+}(\mathcal{H})$ and the classical entropy of the
sequence of its diagonal values in any orthonormal basis
$\{|i\rangle\}_{i=1}^{+\infty}$ of the space $\mathcal{H}$ are
related as follows
\begin{equation}\label{H-q-c}
H(A)\leq H\left(\left\{\langle
i|A|i\rangle\right\}_{i=1}^{+\infty}\right)
\end{equation}
(this inequality can be proved by using nonnegativity of the
relative entropy).\vspace{5pt}
By using relations (\ref{H-fun-eq}) and (\ref{H-q-c}) it is easy to
derive from proposition 5E in \cite{Sh-4} the following continuity
condition for the quantum entropy.\vspace{5pt}
\begin{property}\label{cont-cond-a-2}
\textit{Let $\{|i\rangle\}_{i=1}^{+\infty}$ be an orthonormal basis
of a Hilbert space $\mathcal{H}$. Continuity of the quantum entropy
on a set $\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$ follows
from continuity of the classical entropy on the set
$\,\left\{\{\langle
i|A|i\rangle\}_{i=1}^{+\infty}\,|\,A\in\mathcal{A}\right\}\subset
(\ell_{1})_{+}$.}
\end{property}\vspace{5pt}
We will use the triangle inequality
\begin{equation}\label{d-ineq}
H(C)\geq\left|H\left(\mathrm{Tr}_{\mathcal{K}}C\right)-H\left(\mathrm{Tr}_{\mathcal{H}}C\right)\right|
\end{equation}
valid for any operator $C$ in
$\mathfrak{T}_{+}(\mathcal{H}\otimes\mathcal{K})$
\cite{N&Ch}.\vspace{5pt}
For arbitrary map $\Phi\in\mathfrak{L}^{+}_{\leq
1}(\mathcal{H},\mathcal{H}')$ and operator
$A\in\mathfrak{T}_{+}(\mathcal{H})$ the following estimation holds
\begin{equation}\label{out-ent-est}
H(\Phi(A))\leq
\left[\sup_{\rho\in\mathrm{extr}\mathfrak{S}(\mathcal{H})}H(\Phi(\rho))\right]\mathrm{Tr}A+H(A),
\end{equation}
which is proved by using the spectral decomposition of $A$ and
inequality (\ref{w-k-ineq}).\vspace{5pt}
The relative entropy for two operators $A$ and $B$ in
$\mathfrak{T}_{+}(\mathcal{H})$ is defined as follows
(cf.\cite{L-2})
$$
H(A\,\|B)=\sum_{i=1}^{+\infty}\langle i|\,(A\log A-A\log
B+B-A)\,|i\rangle,
$$
where $\{|i\rangle\}_{i=1}^{+\infty}$ is the orthonormal basis of
eigenvectors of the operator $A$ and it is assumed that
$H(A\,\|B)=+\infty$ if $\,\mathrm{supp}A$ is not contained in
$\mathrm{supp}B$.\vspace{5pt}
For natural $k$ we denote by $\mathfrak{T}_{+}^{k}(\mathcal{H})$
(correspondingly by $\mathfrak{S}_{k}(\mathcal{H})$) the set of
positive trace class operators (correspondingly states) having rank
$\leq k$.\vspace{5pt}
For natural $k$ and lower bounded Borel function $f$ on the set
$\mathfrak{S}_{k}(\mathcal{H})$ consider the functions
\begin{equation}\label{sigma-roof}
\mathfrak{S}(\mathcal{H})\ni
\rho\mapsto\hat{f}_{k}^{\sigma}(\rho)=\sup_{\{\pi_{i},\rho_{i}\}\in\mathcal{P}^{\mathrm{a}}_{\{\rho\}}(\mathfrak{S}_{k}(\mathcal{H}))}
\sum_{i}\pi_{i}f(\rho_{i})
\end{equation}
and
\begin{equation}\label{mu-roof}
\mathfrak{S}(\mathcal{H})\ni
\rho\mapsto\hat{f}_{k}^{\mu}(\rho)=\sup_{\mu\in\mathcal{P}_{\{\rho\}}(\mathfrak{S}_{k}(\mathcal{H}))}
\int_{\mathfrak{S}_{k}(\mathcal{H})}f(\sigma)\mu(d\sigma)
\end{equation}
(the first supremum is over all decompositions of the state $\rho$
into countable convex combination of states of rank $\leq k$, the
second one is over all probability measures with the barycenter
$\rho$ supported by states of rank $\leq k$). Properties of these
functions are studied in \cite{Sh-11}, where it is shown that lower
semicontinuity of the function $f$ on the set
$\mathfrak{S}_{k}(\mathcal{H})$ implies lower semicontinuity and
coincidence of the functions $\hat{f}_{k}^{\sigma}$ and
$\hat{f}_{k}^{\mu}$ on the set
$\mathfrak{S}(\mathcal{H})$.\vspace{5pt}
Following \cite{Sh-11} we say that a subset $\mathcal{A}$ of the
cone $\mathfrak{T}_{+}(\mathcal{H})$ of positive trace class
operators has the uniform approximation property (the
UA\nobreakdash-\hspace{0pt}property) if
$$
\lim_{k\rightarrow+\infty}\sup_{A\in\mathcal{A}}\Delta_{k}(A)=0,
$$
where
$$
\Delta_{k}(A)=\inf_{\{\pi_{i},A_{i}\}\in\mathcal{P}^{\mathrm{a}}_{\{A\}}(\mathfrak{T}_{+}^{k}(\mathcal{H}))}
\sum_{i}\pi_{i}H(A_{i}\|A)
$$
for each natural $k$ (the infimum is over all decompositions of the
operator $A$ into countable convex combination of operators of rank
$\leq k$). The UA\nobreakdash-\hspace{0pt}property of a set
$\mathcal{A}$ is a sufficient condition for continuity of the
quantum entropy on this set, if the set $\mathcal{A}$ is compact
then this condition is also necessary. \vspace{5pt}
We will use the following simple observation
\begin{equation}\label{u-h-c}
\lim_{y\rightarrow0}\sup_{x\in[0,1]}(x+y)h_{2}\left(\frac{y}{x+y}\right)=0.
\end{equation}
\textbf{Note:} In what follows continuity of a function $f$ on a
set $\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$ implies its
finiteness on this set (in contrast to lower (upper)
semicontinuity).
\section{On continuity of the output entropy of\\ a positive linear map}
\subsection{The general case}
Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')\,$
be a positive linear map. The output entropy $H_{\Phi}\doteq
H\circ\Phi$ of this map is a concave nonnegative lower
semicontinuous function on the set
$\mathfrak{S}(\mathcal{H})\subset\mathfrak{T}(\mathcal{H})$. The
following theorem shows that this function can not be finite and
discontinuous simultaneously. \vspace{50pt}
\begin{theorem}\label{bound-cont}
\textit{Let $\,\Phi$ be a map in
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$. The following
properties are equivalent:}
\begin{enumerate}[(i)]
\item \textit{the function $\rho\mapsto H_{\Phi}(\rho)$ is finite on $\,\mathfrak{S}(\mathcal{H})$;}
\item \textit{the function $\rho\mapsto H_{\Phi}(\rho)$ is continuous and bounded on $\,\mathfrak{S}(\mathcal{H})$;}
\item \textit{there exists an orthonormal basis $\{|i\rangle\}_{i=1}^{+\infty}$ of the space $\mathcal{H}'$ such that
the function $\rho\mapsto H\left(\left\{\langle i|\Phi(\rho)|i\rangle\right\}_{i=1}^{+\infty}\right)$
is continuous and bounded on
$\,\mathfrak{S}(\mathcal{H})$;}\footnote{By Proposition \ref{cont-cond-a-2} this property is formally stronger than the previous one.}
\item \textit{there exist an orthonormal basis $\{|i\rangle\}_{i=1}^{+\infty}$ of the space $\mathcal{H}'$ and a sequence $\{h_{i}\}_{i=1}^{+\infty}$ of
nonnegative numbers such that
$
\|\sum_{i=1}^{+\infty}h_{i}\Phi^{*}(|i\rangle\langle
i|)\|<+\infty\quad\textit{and}\quad
\sum_{i=1}^{+\infty}e^{-h_{i}}<+\infty,
$$
where
$\Phi^{*}:\mathfrak{B}(\mathcal{H}')\rightarrow\mathfrak{B}(\mathcal{H})$
is the dual map to the map $\Phi$.}
\end{enumerate}
\emph{The set $\,\mathfrak{S}(\mathcal{H})$ in $\,\mathrm{(i)}$ can
be replaced by arbitrary convex closed bounded subset
$\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$ such that
$\;\sup_{A\in\mathcal{A}}\lim_{n\rightarrow+\infty}\mathrm{Tr}AB_{n}<+\infty\;\Rightarrow\;\sup_{n}\|B_{n}\|<+\infty$
for any increasing sequence $\{B_{n}\}$ of positive operators in
$\,\Phi^{*}(\mathfrak{B}(\mathcal{H}'))$.}
\end{theorem} \vspace{5pt}
\textbf{Proof.} $\mathrm{(i)\Rightarrow(ii)}$ Validity of the
discrete Jensen inequality for the concave finite nonnegative
function $\rho\mapsto H_{\Phi}(\rho)$ implies its boundedness.
Indeed, if for each natural $n$ there exists a state $\rho_{n}$ such
that $H_{\Phi}(\rho_{n})\geq 2^{n}$ then
$$
H_{\Phi}\left(\sum_{n=1}^{+\infty}2^{-n}\rho_{n}\right)\geq\sum_{n=1}^{+\infty}2^{-n}H_{\Phi}\left(\rho_{n}\right)=+\infty.
$$
Lemma \ref{H-oper} below implies existence of a
$\mathfrak{H}$\nobreakdash-\hspace{0pt}operator $H=-\log T$ such
that $\mathrm{g}(H)<+\infty$ and $\mathrm{Tr}H\Phi(\rho)\leq h$ for
all $\rho\in\mathfrak{S}(\mathcal{H})$ and some $h>0$. Let
$H=\sum_{i=1}^{+\infty}h_{i}|i\rangle\langle i|$. Since the function
$$
\rho\mapsto\mathrm{Tr}H\Phi(\rho)=\sum_{i=1}^{+\infty}h_{i}\langle
i|\Phi(\rho)|i\rangle=\mathrm{Tr}\left[\sum_{i=1}^{+\infty}h_{i}\Phi^{*}(|i\rangle\langle
i|)\right]\rho
$$
is bounded on $\mathfrak{S}(\mathcal{H})$, the linear operator in
the squire brackets is bounded on $\mathcal{H}$. Thus the above
function is continuous on $\mathfrak{S}(\mathcal{H})$. For arbitrary
compact subset $\mathcal{C}$ of $\mathfrak{S}(\mathcal{H})$ Dini's
lemma implies uniform convergence of the series
$\sum_{i=1}^{+\infty}h_{i}\mathrm{Tr}\Phi^{*}(|i\rangle\langle
i|)\rho$ on the set $\mathcal{C}$ and hence existence of a
nondecreasing sequence $\{y^{\mathcal{C}}_{i}\}_{i=1}^{+\infty}$ of
positive numbers converging to $+\infty$ such that
$\sup_{\rho\in\mathcal{C}}\sum_{i=1}^{+\infty}y^{\mathcal{C}}_{i}h_{i}\mathrm{Tr}\Phi^{*}(|i\rangle\langle
i|)\rho<+\infty$. Let
$H^{\mathcal{C}}=\sum_{i=1}^{+\infty}y^{\mathcal{C}}_{i}h_{i}|i\rangle\langle
i|$ be a $\mathfrak{H}$\nobreakdash-\hspace{0pt}operator with
$\mathrm{g}(H^{\mathcal{C}})=0$. Thus we have
\begin{equation}\label{int-exp}
\sup_{\rho\in\mathcal{C}}\mathrm{Tr}H^{\mathcal{C}}\Phi(\rho)
=\sup_{\rho\in\mathcal{C}}\sum_{i=1}^{+\infty}y^{\mathcal{C}}_{i}h_{i}\mathrm{Tr}\Phi^{*}(|i\rangle\langle
i|)\rho<+\infty.
\end{equation}
By proposition \ref{H-cont-cond}B the function $\rho\mapsto
H(\Phi(\rho))$ is continuous on the set $\mathcal{C}$ and hence on
the set $\mathfrak{S}(\mathcal{H})$ (since $\mathcal{C}$ is an
arbitrary compact subset of $\mathfrak{S}(\mathcal{H})$).
$\mathrm{(i)\Rightarrow(iv)}$ In the proof of
$\mathrm{(i)\Rightarrow(ii)}$ existence of the basis
$\{|i\rangle\}_{i=1}^{+\infty}$ and of the sequence
$\{h'_{i}=\lambda h_{i}\}_{i=1}^{+\infty},\,\lambda>0,$ with the
desired properties is shown.
$\mathrm{(iv)\Rightarrow(iii)}$ follows from the proof of
$\mathrm{(i)\Rightarrow(ii)}$ since (\ref{int-exp}) and Proposition
\ref{H-cont-cond}B implies continuity of the function $\rho\mapsto
H\left(\left\{\langle
i|\Phi(\rho)|i\rangle\right\}_{i=1}^{+\infty}\right)$ on the set
$\mathcal{C}$.
$\mathrm{(iii)\Rightarrow(i)}$ follows from relation (\ref{H-q-c}).
The last assertion of the theorem is a corollary of the proof of
$\mathrm{(i)\Rightarrow(ii)}$. $\square$ \vspace{5pt}
\begin{lemma}\label{H-oper}
\emph{For an arbitrary convex set
$\mathcal{A}\subset\mathfrak{T}_{1}(\mathcal{H})$, on which the
quantum entropy is bounded, there exists an operator
$\;T\in\mathfrak{T}_{1}(\mathcal{H})$ such that
$$
\sup_{A\in\mathcal{A}}\mathrm{Tr}A(-\log T)<+\infty\quad
\textit{and} \quad UT=TU
$$
for any unitary $\;U$ in $\mathfrak{B}(\mathcal{H})$ such that
$\;UAU^{*}\in\mathcal{A}$ for all $A\in\mathcal{A}$.}
\end{lemma}\vspace{5pt}
\textbf{Proof.} Let $\mathcal{K}$ be the one dimensional space
generated by the unit vector $|0\rangle$. Consider the convex set
$$
\mathcal{A}^{\mathrm{e}}=\left\{\rho_{A}=A\oplus(1-\mathrm{Tr}A)|0\rangle\langle
0|\,|\,A\in\mathcal{A}\right\}
$$
of states in $\mathfrak{S}(\mathcal{H}\oplus\mathcal{K})$. For
arbitrary $A\in\mathcal{A}$ we have
$$
H(\rho_{A})=-\mathrm{Tr}A\log A+\eta(1-\mathrm{Tr}A)
=H(A)+\eta(\mathrm{Tr}A)+\eta(1-\mathrm{Tr}A)\leq H(A)+1.
$$
Thus the von Neumann entropy is bounded on the convex set
$\mathcal{A}^{\mathrm{e}}$. Hence the
$\chi$\nobreakdash-\hspace{0pt}capacity
$\bar{C}(\mathcal{A}^{\mathrm{e}})$ of this set is finite
\cite{Sh-4}. Theorem 1 in \cite{Sh-4} implies existence of the
unique state $\Omega(\mathcal{A}^{\mathrm{e}})$ in
$\mathrm{cl}(\mathcal{A}^{\mathrm{e}})$ (called the optimal average
state of the set $\mathcal{A}^{\mathrm{e}}$) such that
$$
H(\rho\|\Omega(\mathcal{A}^{\mathrm{e}}))\leq
\bar{C}(\mathcal{A}^{\mathrm{e}})
$$
for all $\rho\in\mathcal{A}^{\mathrm{e}}$. The state
$\Omega(\mathcal{A}^{\mathrm{e}})$ has the form $T\oplus
\lambda|0\rangle\langle 0|$, where
$T\in\mathfrak{T}_{1}(\mathcal{H})$ and $\lambda>0$.
For arbitrary unitary $U$ in $\mathfrak{B}(\mathcal{H})$ such that
$U\mathcal{A}U^{*}=\mathcal{A}$ corollary 8 in \cite{Sh-4} implies
$(U\oplus
I_{\mathcal{K}})\Omega(\mathcal{A}^{\mathrm{e}})=\Omega(\mathcal{A}^{\mathrm{e}})(U\oplus
I_{\mathcal{K}})$ and hence $UT=TU$. $\square$ \vspace{5pt}
\begin{remark}\label{on-cont} Theorem \ref{bound-cont} \emph{do not assert}
that finiteness of the quantum entropy on the set
$\Phi(\mathfrak{S}(\mathcal{H}))$ implies its continuity on this set
since continuity of the function $\rho\mapsto H_{\Phi}(\rho)\doteq
H(\Phi(\rho))$ on the \emph{noncompact} set
$\mathfrak{S}(\mathcal{H})$ does not imply continuity of the
function $A\mapsto H(A)$ on the set
$\Phi(\mathfrak{S}(\mathcal{H}))$. To show this consider the
following example.
Let $\mathcal{A}$ be a convex closed subset of
$\mathfrak{S}(\mathcal{H}')$ on which the von Neumann entropy is
bounded but not continuous (see the examples in \cite{Sh-4}). Let
$\{\sigma_{n}\}_{n=1}^{+\infty}$ be a sequence of states in
$\mathcal{A}$ converging to a state $\sigma_{0}$ such that
$\lim_{n\rightarrow+\infty}H(\sigma_{n})\neq H(\sigma_{0})$.
Consider the map $\Phi:\rho\mapsto\sum_{n\geq0}\langle
n|\rho|n\rangle\sigma_{n}$, where $\{|n\rangle\}_{n\geq0}$ is a
particular orthonormal basis in $\mathcal{H}$. By Theorem
\ref{bound-cont} the function $\rho\mapsto H_{\Phi}(\rho)$ is
continuous on the set $\mathfrak{S}(\mathcal{H})$ but the function
$A\mapsto H(A)$ is not continuous on the set
$\Phi(\mathfrak{S}(\mathcal{H}))$ containing the sequence
$\{\sigma_{n}\}_{n=1}^{+\infty}$ and the state $\sigma_{0}$.
Continuity of the function $\rho\mapsto H_{\Phi}(\rho)$ on the set
$\mathfrak{S}(\mathcal{H})$ means continuity of the function
$A\mapsto H(A)$ on each set of the form $\Phi(\mathcal{C})$, where
$\mathcal{C}$ is a compact subset of $\mathfrak{S}(\mathcal{H})$.
\end{remark}\vspace{5pt}
\begin{remark}\label{on-cont+}
The main assertion of Theorem \ref{bound-cont} (the implication
$\mathrm{(i)\Rightarrow(ii)}$) is based on the specific property of
the von Neumann entropy, it can not be proved by using only such
general properties of entropy-type functions as concavity, lower
semicontinuity and nonnegativity. The simplest example showing this
is given by the function $\rho\mapsto R_{0}(\Phi(\rho))\doteq
\|\Phi(\rho)\|_{1}\log\mathrm{rank}(\Phi(\rho))$ -- the output
$0$\nobreakdash-\hspace{0pt}order Renyi entropy of the map $\Phi$.
\end{remark}\vspace{5pt}
Theorem \ref{bound-cont} can be used to obtain a condition of
continuity of the output entropy for the following class of quantum
channels.\vspace{5pt}
\begin{example}\label{e-b-c}
Let $G$ be a compact group, $\{V_{g}\}_{g\in G}$ be an unitary
representation of $G$ on $\mathcal{H}'$, $M$ be a positive
operator-valued measure (POVM) on $G$ taking values in
$\mathfrak{B}(\mathcal{H})$. For given arbitrary state $\sigma$ in
$\mathfrak{S}(\mathcal{H}')$ consider the quantum channel
\begin{equation*}
\Phi_{\sigma}(\rho)=\int_{G}V_{g}\sigma V_{g}^{*}\mathrm{Tr}\rho
M(dg).
\end{equation*}%
The channel of this type was used in \cite{H-Sh-W} as an example of
entanglement-breaking channel which has no Kraus representation with
operators of rank~1.\footnote{An arbitrary finite dimensional
entanglement-breaking channel has Kraus representation with
operators of rank~1~\cite{Ruskai}.}
By Theorem \ref{bound-cont} the channel $\Phi_{\sigma}$ has
continuous output entropy if the state
$\omega(G,V_{g},\sigma)=\int_{G}V_{g}\sigma V_{g}^{*}\mu_{H}(dg)$,
where $\mu_{H}$ is the Haar measure on $G$, has finite entropy. This
condition is necessary if the set of probability measures
$\{\mathrm{Tr}\rho M(\cdot)\}_{\rho\in\mathfrak{S}(\mathcal{H})}$ is
weakly dense in the set of all probability measures on $G$.
\end{example}\vspace{5pt}
Theorem \ref{bound-cont} and inequality (\ref{w-k-ineq}) imply the
following observation (which can be directly proved by using Lemma
\ref{face}).\vspace{5pt}
\begin{corollary}\label{bound-cont-c}
\textit{Let $\,\Phi$ and $\,\Psi$ be maps in
$\,\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ and
$\lambda\in(0,1)$. The map $\,\lambda\Phi+(1-\lambda)\Psi$ has
continuous output entropy if and only if the maps $\,\Phi$ and
$\,\Psi$ have continuous output entropy.}
\end{corollary}\vspace{5pt}
Thus the set of all positive maps with continuous output entropy is
convex and forms a \emph{face} of the convex set
$\,\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$. It is easy
to show that this face is dense in
$\,\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ (in the
strong convergence topology).\vspace{5pt}
We will use the following corollary of Theorem \ref{bound-cont} and
inequality (\ref{w-k-ineq+}).\vspace{5pt}
\begin{corollary}\label{bound-cont-c+}
\emph{Let $\,\{\Phi_{i}\}_{i\in I}$ be a finite or countable family
of maps in \break
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ such that
$\,\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\sum_{i\in
I}\mathrm{Tr}\Phi_{i}(\rho)<+\infty$. The output entropy of the map
$\sum_{i\in I}\Phi_{i}$ is continuous if
$$
\sum_{i\in I}H(\Phi_{i}(\rho))<+\infty\quad\textit{and}\quad
H\left(\left\{\mathrm{Tr}\Phi_{i}(\rho)\right\}_{i\in
I}\right)<+\infty\quad \forall\rho\in\mathfrak{S}(\mathcal{H}).
$$
This condition is necessary if either
$\,\mathrm{supp}\,\Phi_{i}(\rho)\perp\mathrm{supp}\,\Phi_{j}(\rho)$
for all $i\neq j$ and all $\rho\in\mathfrak{S}(\mathcal{H})$ or the
set $I$ is finite.}
\end{corollary}\vspace{5pt}
Theorem \ref{bound-cont} provides a simple proof of the following
result.\footnote{It can be also proved by using Proposition
\ref{cont-cond-a-1} in the Appendix and corollary 4 in
\cite{Sh-11}.}\vspace{5pt}
\begin{corollary}\label{tensor-pr} \textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
and
$\,\Psi:\mathfrak{T}(\mathcal{K})\rightarrow\mathfrak{T}(\mathcal{K}')$
be two positive linear bounded maps having continuous output
entropy. If the map
$\,\Phi\otimes\Psi:\mathfrak{T}(\mathcal{H}\otimes\mathcal{K})\rightarrow\mathfrak{T}(\mathcal{H}'\otimes\mathcal{K}')$
is positive then it has continuous output entropy.}
\end{corollary}\vspace{5pt}
\textbf{Proof.} We may assume that the maps $\Phi$ and $\Psi$ are
trace non-increasing. By Theorem \ref{bound-cont} it is sufficient
to prove that $H(\Phi\otimes\Psi(\omega))<+\infty$ for any
$\omega\in\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})$. This follows
from subadditivity of the quantum entropy since
$\mathrm{Tr}_{\mathcal{K}}\Phi\otimes\Psi(\omega)\leq\Phi(\mathrm{Tr}_{\mathcal{K}}\omega)$
and
$\mathrm{Tr}_{\mathcal{H}}\Phi\otimes\Psi(\omega)\leq\Psi(\mathrm{Tr}_{\mathcal{H}}\omega)$.
$\square$\vspace{5pt}
In Section 4 we will use the following corollary of Theorem
\ref{bound-cont}.\vspace{5pt}
\begin{corollary}\label{tensor-pr+} \textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
and
$\,\Psi:\mathfrak{T}(\mathcal{K})\rightarrow\mathfrak{T}(\mathcal{K}')$
be positive linear bounded maps such that the map
$\,\Phi\otimes\Psi:\mathfrak{T}(\mathcal{H}\otimes\mathcal{K})\rightarrow\mathfrak{T}(\mathcal{H}'\otimes\mathcal{K}')$
is positive ($\mathcal{H},\mathcal{H}',\mathcal{K}$, $\mathcal{K}'$
are separable Hilbert spaces). If the map $\,\Psi$ is trace
preserving and has finite (and hence continuous) output entropy then
the following properties are equivalent:}
\begin{enumerate}[(i)]
\item
\emph{ $H(\Phi\otimes\Psi(|\varphi\rangle\langle\varphi|))<+\infty$ for
any unit vector
$\varphi\in\mathcal{H}\otimes\mathcal{K}$;}
\item \emph{the map $\Phi$ has
continuous output entropy;}
\item \emph{the map $\Phi\otimes\Psi$ has
continuous output entropy.}
\end{enumerate}
\emph{If the map $\,\Phi$ is trace preserving then the condition of
finiteness of the output entropy of the map $\,\Psi$ can be
replaced by the condition}
$$
\min\left\{H_{\Phi}(\mathrm{Tr}_{\mathcal{K}}|\varphi\rangle\langle\varphi|),H_{\Psi}(\mathrm{Tr}_{\mathcal{H}}|\varphi\rangle\langle\varphi|)\right\}
<+\infty\quad\forall\varphi\in\mathcal{H}\otimes\mathcal{K}.
$$
\end{corollary}\vspace{5pt}
\textbf{Proof.} We may assume that the map $\Phi$ is trace
non-increasing.
$\mathrm{(i)\Rightarrow(ii)}$ Let $\rho$ be an arbitrary state in
$\mathfrak{S}(\mathcal{H})$ and $|\varphi\rangle$ be a vector in
$\mathcal{H}\otimes\mathcal{K}$ such that
$\rho=\mathrm{Tr}_{\mathcal{K}}|\varphi\rangle\langle\varphi|$.
Since the map $\Psi$ is trace preserving we have
$\mathrm{Tr}_{\mathcal{K}}\Phi\otimes\Psi(|\varphi\rangle\langle\varphi|)=\Phi(\rho)$.
By noting that
$\mathrm{Tr}_{\mathcal{H}}\Phi\otimes\Psi(|\varphi\rangle\langle\varphi|)\leq\Psi(\sigma)$,
where
$\sigma=\mathrm{Tr}_{\mathcal{H}}|\varphi\rangle\langle\varphi|$,
and by using finiteness of $H(\Psi(\sigma))$ with (\ref{H-fun-ineq})
and (\ref{d-ineq}) we conclude that $H(\Phi(\rho))<+\infty$. By
Theorem \ref{bound-cont} the map $\Phi$ has continuous output
entropy.
$\mathrm{(ii)\Rightarrow(iii)}$ follows from Corollary
\ref{tensor-pr}.
$\mathrm{(iii)\Rightarrow(i)}$ is obvious.
The last assertion of the corollary is proved by the similar
argumentation. $\square$ \vspace{5pt}
Property $\mathrm{(iv)}$ in Theorem \ref{bound-cont} can be
considered as a criterion of continuity of the output entropy of a
map $\Phi$ in terms of its dual map $\Phi^{*}$. It will be used in
the proof of Proposition \ref{s-map} in the next subsection.
\subsection{The case of quantum operation}
The simplest example of quantum operation (completely positive trace
non-increasing linear map) from $\mathfrak{T}(\mathcal{H})$ to
$\mathfrak{T}(\mathcal{H}')$ is the map $(\cdot)\mapsto
V(\cdot)V^{*}$, where $V$ is a linear contraction from $\mathcal{H}$
to $\mathcal{H}'$. Theorem \ref{bound-cont} implies the following
result. \vspace{5pt}
\begin{property}\label{s-map}
\textit{Let $\,V$ be a linear operator from $\mathcal{H}$ to
$\mathcal{H}'$. The function
$\mathfrak{S}(\mathcal{H})\ni\rho\mapsto H(V\rho V^{*})$ is
continuous if and only if the operator $V$ is compact and has such
sequence $\{\nu_{i}\}$ of singular values (eigenvalues of
$\sqrt{V^{*}V}$) that
$\;\sum_{i=1}^{+\infty}e^{-\lambda/\nu^{2}_{i}}<+\infty$ for some
$\lambda>0$.} \emph{If this condition holds then}
$$
\sup_{\rho\in\mathfrak{S}(\mathcal{H})}H(V\rho V^{*})=\lambda^{*}(V)
$$
\emph{where $\lambda^{*}(V)$ is either the unique solution of the
equation $\,\sum_{i=1}^{+\infty}e^{-\lambda/\nu^{2}_{i}}=1$ if it
exists or equal to
$\,\mathrm{g}(\{\nu^{-2}_{i}\})=\inf\{\lambda>0\,|\,\sum_{i=1}^{+\infty}e^{-\lambda/\nu^{2}_{i}}<+\infty\}$
otherwise.}\footnote{The equation
$\,\sum_{i=1}^{+\infty}e^{-\lambda/\nu^{2}_{i}}=1$ has no solution
if and only if
$\,\sum_{i=1}^{+\infty}e^{-\mathrm{g}(\{\nu^{-2}_{i}\})/\nu^{2}_{i}}<1$.}
\end{property}\vspace{5pt}
In what follows we will use the parameter $\lambda^{*}(V)$ defined
in Proposition \ref{s-map} for arbitrary operator
$V\in\mathfrak{B}(\mathcal{H})$ assuming that
$\lambda^{*}(V)=+\infty$ if the operator $V$ either is not compact
or has such sequence of singular values $\{\nu_{i}\}$ that
$\;\sum_{i=1}^{+\infty}e^{-\lambda/\nu^{2}_{i}}=+\infty$ for all
$\lambda>0$.\vspace{5pt}
\textbf{Proof.} We can assume that $\mathcal{H}=\mathcal{H}'$,
$V=|V|$, $\|V\|\leq 1$ and $\mathrm{Ker}V=\{0\}$.
Let $V=\sum_{i=1}^{+\infty}\nu_{i}|i\rangle\langle i|$. If
$\sum_{i=1}^{+\infty}e^{-\lambda/\nu^{2}_{i}}<+\infty$ for
$\lambda>0$ then property $\mathrm{(iv)}$ in Theorem
\ref{bound-cont} holds with the basis
$\{|i\rangle\}_{i=1}^{+\infty}$ and the sequence $\{h_{i}=\lambda
\nu_{i}^{-2}\}_{i=1}^{+\infty}$ (since in this case
$\Phi^{*}(\cdot)=V(\cdot)V$ and hence $\Phi^{*}(|i\rangle\langle
i|)=\nu^{2}_{i}|i\rangle\langle i|$).
The assertion concerning the supremum of the function $\rho\mapsto
H(V\rho V)$ is easily derived from Lemma \ref{simple+} in the
Appendix by using relation (\ref{H-q-c}).
Suppose the function $\rho\mapsto H(V\rho V)$ is continuous on the
set $\mathfrak{S}(\mathcal{H})$. Then the entropy is bounded on the
convex set $\{V\rho V\,|\,\rho\in\mathfrak{S}(\mathcal{H})\}$ and
hence this set is relatively compact by corollary 5 in \cite{Sh-4}
(used with the construction from the proof of Lemma \ref{H-oper}).
Thus the operator $V$ is compact (since otherwise there exists a
sequence of unit vectors $\{|\varphi_{n}\rangle\}$ such that the
sequence $\{V|\varphi_{n}\rangle\}$ is not relatively compact).
Lemma \ref{H-oper} implies existence of an operator
$T\in\mathfrak{T}_{1}(\mathcal{H})$ such that
$\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\mathrm{Tr}V\rho V(-\log
T)<+\infty$ and $UT=TU$ for arbitrary unitary $U$ commuting with the
operator $V$. It follows from the last property of the operator $T$
that this operator is diagonizable in the basis $\{|i\rangle\}$,
t.i. $T=\sum_{i=1}^{+\infty}\tau_{i}|i\rangle\langle i|$, where
$\{\tau_{i}\}_{i=1}^{+\infty}$ is a sequence of nonnegative numbers
such that $\sum_{i=1}^{+\infty}\tau_{i}\leq 1$. Thus we have
$$
\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\mathrm{Tr}V\rho V(-\log
T)=\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\sum_{i=1}^{+\infty}\langle
i|\rho|i\rangle \nu^{2}_{i}(-\log\tau_{i})=\lambda<+\infty
$$
and hence $\nu^{2}_{i}(-\log\tau_{i})\leq\lambda$ for all $i$. This
implies $\lambda^{*}(V)<+\infty$. $\square$\vspace{5pt}
An arbitrary quantum operation
$\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
has Kraus representation (\ref{Kraus-rep}) determined by a set
$\{V_{i}\}_{i=1}^{+\infty}$ of linear operators from $\mathcal{H}$
into $\mathcal{H}'$ such that
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}\leq I_{\mathcal{H}}$. The
following proposition contains the sufficient conditions for
continuity of the output entropy of a quantum operation $\Phi$
expressed in terms of the set $\{V_{i}\}_{i=1}^{+\infty}$ of its
Kraus operators.\vspace{5pt}
\begin{property}\label{CE-conditions}
\textit{Let $\,\Phi$ be a quantum operation in $\,\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}')$ and $\,\{V_{i}\}_{i\in I}$ be the
corresponding set of Kraus operators. Let
$d_{i}=\mathrm{rank}V_{i}\leq+\infty$.}\vspace{5pt}
A) \textit{If the set $\,I$ is finite then the operation $\,\Phi$
has continuous output entropy if and only if
\begin{equation}\label{n-s-cond}
\lambda^{*}\left(V_{i}\right)<+\infty\quad \forall i\in I,
\end{equation}
which in this case is equivalent to}
\begin{equation}\label{n-s-cond+}
\lambda^{*}\left(\sqrt{\textstyle\sum_{i\in
I}V^{*}_{i}V_{i}}\right)<+\infty. \footnote{In contrast to the set
$\{V_{i}\}_{i\in I}$ the operator $\sum_{i\in
I}V^{*}_{i}V_{i}=\Phi^{*}(I_{\mathcal{H}'})$ is uniquely determined
by the operation $\Phi$.}
\end{equation}
\textit{In general case (\ref{n-s-cond}) is a necessary condition of
continuity of the output entropy of the operation $\,\Phi$ (in
contrast to (\ref{n-s-cond+})).}\vspace{5pt}
B) \textit{If $\,I=\mathbb{N}$ then the operation $\,\Phi$ has
continuous output entropy if one of the following conditions is
valid:}
\begin{enumerate}[a)]
\item \textit{$d_{i}<+\infty$ for all $i$ and
there exists a sequence $\{h_{i}\}_{i=1}^{+\infty}$ of nonnegative
numbers such that}
$
\|\sum_{i=1}^{+\infty}h_{i}V^{*}_{i}V_{i}\|<+\infty\quad\textit{and}\quad
\sum_{i=1}^{+\infty}d_{i}e^{-h_{i}}<+\infty;
$$
\item \textit{$H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}^{+\infty}_{i=1}\right)<+\infty$ for all
$\rho\in\mathfrak{S}(\mathcal{H})$ \footnote{The sense of this
condition and its "variations" are considered in Proposition
\ref{CE-conditions-c} below.} and condition (\ref{n-s-cond+})
holds.}
\end{enumerate}
\vspace{5pt} \emph{Condition (\ref{n-s-cond+}) in $\,\mathrm{b)}$
can be replaced by the condition of finiteness of one of the series
$\,\sum_{i=1}^{+\infty}\lambda^{*}(V_{i})$ and
$\,\sum_{i=1}^{+\infty}\log d_{i}\|V_{i}\|^{2}$.}\vspace{5pt}
\textit{If the sequence $\,\{V_{i}\}_{i=1}^{+\infty}$ consists of
scalar multiples of mutually orthogonal projectors then
$\,\mathrm{a})$ is a necessary condition of continuity of the output
entropy of the operation $\,\Phi$.}\vspace{5pt}
\textit{If $\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all
$i\neq j$ then $\,\mathrm{b)}$ is a necessary condition of
continuity of the output entropy of the operation $\,\Phi$.}
\end{property}\vspace{5pt}
\textbf{Proof.} A) This directly follows from Corollary
\ref{bound-cont-c+}, Proposition \ref{s-map} and Corollary
\ref{basic-cont-cond+c} in Section 5 below (since
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}=\Phi^{*}(I_{\mathcal{H}'})$).\vspace{5pt}
B) Suppose condition a) holds. Let
$\mathcal{K}=\bigoplus_{i=1}^{+\infty}\mathrm{Ran}V_{i}$ and $U_{i}$
be a partial isometry from $\mathcal{H}'$ into $\mathcal{K}$ such
that $U_{i}^{*}U_{i}$ is the projector onto
$\mathrm{Ran}V_{i}\subset\mathcal{H}'$ and $U_{i}U_{i}^{*}$ is the
projector onto $\mathrm{Ran}V_{i}\subset\mathcal{K}$. Consider the
quantum operation $\widehat{\Phi}$ in $\,\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{K})$ defined by the sequence of Kraus
operators $\,\{\widehat{V}_{i}=U_{i}V_{i}\}_{i=1}^{+\infty}$.
We have
$\mathrm{Ran}\widehat{V}_{i}\perp\mathrm{Ran}\widehat{V}_{j}$ for
all $i\neq j$. Let $P_{i}$ be the $d_{i}$-rank projector onto the
subspace $\mathrm{Ran}\widehat{V}_{i}$. Consider the
$\mathfrak{H}$-operator $H=\sum_{i=1}^{+\infty}h_{i}P_{i}$. The
condition $\sum_{i=1}^{+\infty}d_{i}e^{-h_{i}}<+\infty$ means that
$\mathrm{g}(H)<+\infty$. The condition
$\|\sum_{i=1}^{+\infty}h_{i}V^{*}_{i}V_{i}\|=h<+\infty$ implies
$$
\mathrm{Tr}H\widehat{\Phi}(\rho)=\sum_{i=1}^{+\infty}h_{i}\mathrm{Tr}P_{i}\widehat{V}_{i}\rho
\widehat{V}_{i}^{*}=\mathrm{Tr}\sum_{i=1}^{+\infty}h_{i}V_{i}^{*}V_{i}\rho\leq
h,\quad \forall\rho\in\mathfrak{S}(\mathcal{H}).
$$
By Proposition \ref{H-cont-cond}A the quantum entropy is bounded on
the set $\widehat{\Phi}(\mathfrak{S}(\mathcal{H}))$. Since
$$
H(\widehat{\Phi}(\rho))=\sum_{i=1}^{+\infty}H(V_{i}\rho
V_{i}^{*})+H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}_{i=1}^{+\infty}\right)
$$
inequality (\ref{w-k-ineq+}) implies boundedness of the function
$\rho\mapsto H(\Phi(\rho))$. By Theorem \ref{bound-cont} this
function is continuous. \vspace{5pt}
If condition b) holds then Proposition \ref{CE-conditions-c} below
and Corollary \ref{basic-cont-cond+c} in Section 5 imply continuity
of the output entropy of the operation $\Phi$. Possibility to
replace condition (\ref{n-s-cond+}) in b) by one of the conditions
$\sum_{i=1}^{+\infty}\lambda^{*}(V_{i})<+\infty$ and
$\sum_{i=1}^{+\infty}\log d_{i}\|V_{i}\|^{2}<+\infty$ follows from
Corollary \ref{bound-cont-c+}, since each of these conditions
implies finiteness of the series $\sum_{i=1}^{+\infty}H(V_{i}\rho
V_{i}^{*})$ for any $\rho$ in
$\mathfrak{S}(\mathcal{H})$.\vspace{5pt}
To prove the assertion concerning necessity of condition a) assume
that $\Phi(\cdot)=\sum_{i=1}^{+\infty}c_{i}P_{i}(\cdot)P_{i}$, where
$\{P_{i}\}$ is a sequence of mutually orthogonal projectors. Lemma
\ref{H-oper} implies existence of a trace class operator of the form
$T=\sum_{i=1}^{+\infty}\lambda_{i}P_{i}$ such that
$$
\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\mathrm{Tr}\Phi(\rho)(-\log
T)=\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\mathrm{Tr}\sum_{i=1}^{+\infty}c_{i}(-\log\lambda_{i})P_{i}\rho<+\infty.
$$
Since $\mathrm{Tr}T=\sum_{i=1}^{+\infty}d_{i}\lambda_{i}$, condition
a) holds with the sequence $\{h_{i}=-\log\lambda_{i}\}$.
To prove the assertion concerning necessity of condition b) it is
sufficient to note that the condition
$\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all $i\neq j$
implies
$$
H(V\rho V^{*})\leq H(\Phi(\rho))=\sum_{i=1}^{+\infty}H(V_{i}\rho
V_{i}^{*})+H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}_{i=1}^{+\infty}\right)\quad \forall
\rho\in\mathfrak{S}(\mathcal{H}),
$$
where $V$ is the Stinespring contraction of the operation $\Phi$
defined via the set $\{V_{i}\}_{i=1}^{+\infty}$ (see the proof of
Proposition \ref{CE-conditions-c} below), and to apply Theorem
\ref{bound-cont} and Proposition \ref{s-map} (by using
$V^{*}V=\sum_{i=1}^{+\infty}V_{i}^{*}V_{i}$). $\square$\vspace{5pt}
\begin{example}\label{cont-cond+++r}
Let $\{V_{i}\}_{i=1}^{+\infty}$ be a sequence of finite rank
operators in $\mathfrak{B}(\mathcal{H})$ such that
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}\leq I_{\mathcal{H}}$,
$\mathrm{Ran}V^{*}_{i}\perp\mathrm{Ran}V^{*}_{j}$ for all
sufficiently large\break $i\neq j$ and $\|V_{i}\|^{2}\leq
C\log^{-\alpha}(i)$ for all $i$, where $\alpha\geq0$ and $C>0$.
Since $V_{i}^{*}V_{i}\leq C\log^{-\alpha}(i) P_{i}$, where $P_{i}$
is the projector on the subspace $\mathrm{Ran}V^{*}_{i}$, condition
a) in Proposition \ref{CE-conditions}B holds for the operation
$\Phi_{\alpha}(\cdot)=\sum_{i=1}^{+\infty}V_{i}(\cdot)V^{*}_{i}$ for
all $\alpha\geq1$ provided the rate of increase of the sequence
$\{\mathrm{rank}V_{i}\}_{i}$ does not exceed the polynomial rate:
$\mathrm{rank}V_{i}\leq i^n$ for some natural $n$ and all
sufficiently large $i$ (this can be shown by using the sequence
$\{h_{i}=(n+2)\log(i)\}$). Hence the output entropy of the operation
$\Phi_{\alpha}$ is continuous in this case.
The last assertion of Proposition \ref{CE-conditions} shows that the
output entropy of the operation $\Phi_{\alpha}$ is not continuous if
$\alpha<1$ and $V_{i}=\sqrt{C\log^{-\alpha}(i)}P_{i}$ even for
bounded sequence $\{\mathrm{rank}V_{i}\}_{i}$. $\square$
\end{example}\vspace{5pt}
The following proposition contains the sufficient conditions for
continuity of the output entropy of the complementary operation
expressed in terms of the set $\{V_{i}\}_{i=1}^{+\infty}$ of Kraus
operators of the initial operation.\vspace{5pt}
\begin{property}\label{CE-conditions-c}
\textit{Let $\,\Phi$ be a quantum operation in $\,\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}')$ and $\,\{V_{i}\}_{i=1}^{+\infty}$ be
the corresponding set of Kraus operators. The complementary
operation $\widetilde{\Phi}$ has continuous output entropy if one of
the following conditions \textup{(}related by $\,
\mathrm{c})\Rightarrow \mathrm{b})\Leftrightarrow
\mathrm{a})$\textup{)} is valid:}
\begin{enumerate}[a)]
\item \textit{$H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}^{+\infty}_{i=1}\right)<+\infty$ for all
$\rho\in\mathfrak{S}(\mathcal{H})$;}
\item \textit{there exists a sequence $\{h_{i}\}_{i=1}^{+\infty}$ of nonnegative numbers
such that
$
\|\sum_{i=1}^{+\infty}h_{i}V^{*}_{i}V_{i}\|<+\infty\quad\textit{and}\quad
\sum_{i=1}^{+\infty}e^{-h_{i}}<+\infty;
$$}
\item
$$
H\left(\left\{\|V_{i}\|^{2}\right\}_{i=1}^{+\infty}\right)<+\infty.
$$
\end{enumerate}
\textit{If $\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all
$i\neq j$ then $\,\mathrm{a})\Leftrightarrow \mathrm{b})$ is a
necessary condition of continuity of the output entropy of the
operation $\,\widetilde{\Phi}$.}
\end{property}\vspace{5pt}
\textbf{Proof.} Show first that a) implies continuity of the
function $\rho\mapsto H(\widetilde{\Phi}(\rho))$.
Let $\mathcal{H}''$ be a separable Hilbert space and $\{|i\rangle\}$
be an orthonormal basis in $\mathcal{H}''$. Then the operator
$V:\mathcal{H}\ni|\varphi\rangle\mapsto\sum_{i=1}^{+\infty}|V_{i}\varphi\rangle\otimes|i\rangle\in\mathcal{H}'\otimes\mathcal{H}''$
is the Stinespring contraction for the operation $\Phi$, t.i.
$$
\Phi(A)=\mathrm{Tr}_{\mathcal{H}''}VAV^{*},\quad
A\in\,\mathfrak{T}(\mathcal{H}).
$$
So we have
$$
\widetilde{\Phi}(A)=\mathrm{Tr}_{\mathcal{H}'}VAV^{*}=\sum_{i,j=1}^{+\infty}\mathrm{Tr}\left[V_{i}AV_{j}^{*}\right]|i\rangle\langle
j|,\quad A\in\,\mathfrak{T}(\mathcal{H}).
$$
By relation (\ref{H-q-c}) condition a) implies
$$
H(\widetilde{\Phi}(\rho))\leq H\left(\left\{\langle
i|\widetilde{\Phi}(\rho)|i\rangle\right\}_{i=1}^{+\infty}\right)=H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}_{i=1}^{+\infty}\right)<+\infty,\quad
\rho\in\,\mathfrak{S}(\mathcal{H}).
$$
By Theorem \ref{bound-cont} the output entropy of the operation
$\widetilde{\Phi}$ is continuous.\vspace{5pt}
Since finiteness of the function
$\mathfrak{S}(\mathcal{H})\ni\rho\mapsto
H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}^{+\infty}_{i=1}\right)$ implies its boundedness,
equivalence of conditions a) and b) can be proved by noting that the
last condition can be rewritten as follows
$
\sup_{\rho\in\mathfrak{S}(\mathcal{H})}\sum_{i=1}^{+\infty}h_{i}\mathrm{Tr}V_{i}\rho
V_{i}^{*}<+\infty
$$
and by using the classical versions of Proposition
\ref{H-cont-cond}A and Lemma \ref{H-oper}.\vspace{5pt}
The implication $\mathrm{c)}\Rightarrow \mathrm{b)}$ is obvious.
\vspace{5pt}
If $\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all $i\neq j$
then
$$
\widetilde{\Phi}(\rho)=\sum_{i=1}^{+\infty}\mathrm{Tr}\left[V_{i}\rho
V_{i}^{*}\right]|i\rangle\langle i|,\quad
\rho\in\,\mathfrak{S}(\mathcal{H}),
$$
and hence the function in a) coincides with the output entropy of
the operation $\widetilde{\Phi}$. $\square$\vspace{5pt}
\begin{example}\label{cont-cond+++r+}
Let $\Phi_{\alpha}$ be the quantum operation described in Example
\ref{cont-cond+++r} with no rank restrictions on its Kraus
operators. By Proposition \ref{CE-conditions-c} the output entropy
of the operation $\widetilde{\Phi}_{\alpha}$ is continuous if
$\alpha\geq 1$ but it is not continuous if $\alpha<1$ and
$V_{i}=\sqrt{C\log^{-\alpha}(i)}P_{i}$. $\square$
\end{example}\vspace{5pt}
\section{The PCE-property a positive linear map}
Continuity of the output entropy of a positive linear map on the
whole set of input states is a very strong requirement. In this
section we consider the substantially weaker property of a positive
linear map consisting in continuity of the output entropy on each
subset of input states on which the von Neumann entropy is
continuous.\vspace{5pt}
\subsection{The general case}
Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')\,$
be a positive linear map. Since the output entropy $H_{\Phi}\doteq
H\circ\Phi$ of this map is a concave nonnegative lower
semicontinuous function on the cone $\mathfrak{T}_{+}(\mathcal{H})$,
we can apply the results of section 4 in \cite{Sh-11} to this
function as follows.
For each natural $k$ consider the concave function
\begin{equation}\label{out-ent-a}
H_{\Phi}^{k}(A)\doteq\sup_{\{\pi_{i},A_{i}\}\in\mathcal{P}^{\mathrm{a}}_{\{A\}}(\mathfrak{T}_{+}^{k}(\mathcal{H}))}
\sum_{i}\pi_{i}H_{\Phi}(A_{i})
\end{equation}
on the cone $\mathfrak{T}_{+}(\mathcal{H})$ (the supremum is over
all decompositions of the operator $A$ into countable convex
combination of operators of rank $\leq k$). By using
(\ref{H-fun-eq}) it is easy to see that the restriction of the above
function $H_{\Phi}^{k}$ to the set $\mathfrak{S}(\mathcal{H})$
coincides with the function $(\widehat{H_{\Phi}})^{\sigma}_{k}$
defined by (\ref{sigma-roof}) with $f=H_{\Phi}$ and that
$$
H_{\Phi}^{k}(\lambda A)=\lambda H_{\Phi}^{k}(A),\quad
A\in\mathfrak{T}_{+}(\mathcal{H}),\;\lambda\geq0.
$$
Hence by using propositions 1 and 3 in \cite{Sh-11} it is easy to
show that the function $H_{\Phi}^{k}$ is lower semicontinuous on the
cone $\mathfrak{T}_{+}(\mathcal{H})$ and that the monotonic sequence
$\{H_{\Phi}^{k}\}$ pointwise converges to the function $H_{\Phi}$.
Following \cite{Sh-11} we will call the function $H_{\Phi}^{k}$ the
\emph{$k$-order approximator} of the output entropy of the map
$\Phi$.
By using spectral decompositions one can prove uniform convergence
of the sequence $\{H_{\Phi}^{k}\}$ to the function $H_{\Phi}$ on
compact subsets of $\mathfrak{T}_{+}(\mathcal{H})$ on which the
quantum entropy is continuous.\vspace{5pt}
\begin{lemma}\label{uniform-c}
\textit{If the quantum entropy is continuous on a compact subset
$\mathcal{A}$ of $\,\mathfrak{T}_{+}(\mathcal{H})$ then}
\begin{equation}\label{uniform-c-exp}
\lim_{k\rightarrow+\infty}\sup_{A\in\mathcal{A},\Phi\in\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')}
\left(H_{\Phi}(A)-H_{\Phi}^{k}(A)\right)=0.
\end{equation}
\end{lemma}\vspace{5pt}
\textbf{Proof.} We may assume that
$\mathcal{A}\subset\mathfrak{T}_{1}(\mathcal{H})$. Let
$\lambda_{i}^{k}(A)$ be the sum of the eigenvalues
$\lambda_{(i-1)k+1},...,\lambda_{ik}$ of the operator $A$ (arranged
in non-increasing order) and $P^{k}_{i}$ be the spectral projector
of this operator corresponding to the above set of eigenvalues.
Since the ensemble $\{\pi^{k}_{i}, (\pi^{k}_{i})^{-1}P^{k}_{i}A\}$,
where $\pi^{k}_{i}=\|A\|^{-1}_{1}\lambda^{k}_{i}(A)$, lies in
$\mathcal{P}^{\mathrm{a}}_{\{A\}}(\mathfrak{T}^{k}_{+}(\mathcal{H}))$,
by using inequalities (\ref{H-fun-ineq}) and (\ref{w-k-ineq+}) we
obtain
$$
\begin{array}{c}
\displaystyle H_{\Phi}(A)-H_{\Phi}^{k}(A)\leq
H(\Phi(A))-\sum_{i}\pi^{k}_{i}H\left(\Phi\left((\pi^{k}_{i})^{-1}P^{k}_{i}A\right)\right)
\\
\displaystyle =H(\Phi(A))-\sum_{i}H\left(\Phi(P^{k}_{i}A)\right)
\leq H\left(\{\mathrm{Tr}\Phi(P^{k}_{i}A)\}\right)\leq
H\left(\{\lambda_{i}^{k}(A)\}\right)
\end{array}
$$
for arbitrary map $\Phi$ in
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$. Hence the
assertion of the lemma follows from lemma 9 in \cite{Sh-11}, which
implies
$$
\lim_{k\rightarrow+\infty}\sup_{A\in\mathcal{A}}\widetilde{\Delta}_{k}(A)
=\lim_{k\rightarrow+\infty}\sup_{A\in\mathcal{A}}H\left(\{\lambda_{i}^{k}(A)\}\right)=0.\quad
\square
$$
Note that concavity of the function $\eta(x)=-x\log x$ implies the
inequality
$$
H_{\Phi}(A)-H_{\Phi}^{k}(A)\leq\inf_{\{\pi_{i},A_{i}\}\in\mathcal{P}^{\mathrm{a}}_{\{A\}}(\mathfrak{T}^{k}_{+}(\mathcal{H}))}
\sum_{i}\pi_{i}H(\Phi(A_{i})\|\Phi(A)),\quad
A\in\mathfrak{T}_{+}(\mathcal{H}),
$$
showing that (\ref{uniform-c-exp}) holds for arbitrary subset
$\mathcal{A}$ of $\mathfrak{T}_{+}(\mathcal{H})$ having the
UA\nobreakdash-\hspace{0pt}property (not necessarily compact)
provided that the set
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ is replaced by
the set $\mathfrak{F}_{\leq 1}(\mathcal{H},\mathcal{H}')$ of all
quantum operations (or by any other subset of positive linear maps
for which monotonicity of the relative entropy holds). \vspace{5pt}
\begin{remark}\label{uniform-c+} Since
the function $A\mapsto H^{k}_{\Phi}(A)$ is lower semicontinuous on
the set $\mathfrak{T}_{+}(\mathcal{H})$ for each $k$, the
generalized Dini's lemma\footnote{The condition of continuity of
functions of the increasing sequence in the standard Dini's lemma
can be replaced by the condition of their lower semicontinuity
(provided that the condition of continuity of the limit function is
valid).} shows that continuity of the function $A\mapsto
H_{\Phi}(A)$ on a compact subset $\mathcal{A}$ of
$\,\mathfrak{T}_{+}(\mathcal{H})$ implies
$
\lim_{k\rightarrow+\infty}\sup_{A\in\mathcal{A}}
\left(H_{\Phi}(A)-H_{\Phi}^{k}(A)\right)=0.
$
The converse implication obviously holds if the function $A\mapsto
H^{k}_{\Phi}(A)$ is continuous on the set $\mathcal{A}$ for each
$k$. $\square$
\end{remark}\vspace{5pt}
The above observations imply the following answer on the second
question stated in the Introduction.\vspace{5pt}
\begin{theorem}\label{PCE-property}
\textit{Let $\,\Phi$ be a map in
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$. The following
properties are equivalent:}
\begin{enumerate}[(i)]
\item \textit{the function $A\mapsto H_{\Phi}(A)$ is continuous on
$\,\mathfrak{T}^{1}_{+}(\mathcal{H})$;}\footnote{
$\,\mathfrak{T}^{1}_{+}(\mathcal{H})$ is the set of 1-rank positive operators in $\mathcal{H}$.}
\item \textit{the function $A\mapsto H^{k}_{\Phi}(A)$ is continuous on $\,\mathfrak{T}_{+}(\mathcal{H})$ for each $k$;}
\item \textit{the function $A\mapsto H_{\Phi}(A)$ is continuous on an arbitrary subset of $\,\mathfrak{T}_{+}(\mathcal{H})$ on which the quantum entropy
is continuous.}
\end{enumerate}
\textit{Property $\,\mathrm{(i)}$ is equivalent to continuity and
boundedness of the function\break $\rho\mapsto H_{\Phi}(\rho)$ on
the set $\mathrm{extr}\mathfrak{S}(\mathcal{H})$ and hence it
follows from the UA\nobreakdash-\hspace{0pt}property of the set
$\,\Phi(\mathrm{extr}\mathfrak{S}(\mathcal{H}))$.}
\end{theorem}\vspace{5pt}
\textbf{Proof.} $\mathrm{(i)\Rightarrow(ii)}$ Show first that
$\mathrm{(i)}$ implies continuity of the function $A\mapsto
H_{\Phi}(A)$ on the set $\mathfrak{T}_{+}^{k}(\mathcal{H})$ for each
$k$. Suppose there exists a sequence
$\{A_{n}\}\subset\mathfrak{T}_{+}^{k}(\mathcal{H})$ converging to an
operator $A_{0}$ such that
\begin{equation}\label{dc}
\lim_{n\rightarrow+\infty}H_{\Phi}(A_{n})>H_{\Phi}(A_{0}).
\end{equation}
For each $n\in\mathbb{N}$ we have $A_{n}=\sum_{i=1}^{k}A_{i}^{n}$,
where $\{A_{i}^{n}\}_{i=1}^{k}$ is a subset of
$\mathfrak{T}_{+}^{1}(\mathcal{H})$. Since the set
$\{A_{n}\}_{n\geq0}$ is compact the compactness criterion for
subsets of $\mathfrak{T}_{+}(\mathcal{H})$ (see the Appendix in
\cite{Sh-11}) implies relative compactness of the sequence
$\{A_{i}^{n}\}_{n}$ for each $i=\overline{1,k}$. Hence we may
consider that there exists
$\lim_{n\rightarrow+\infty}A_{i}^{n}=A_{i}^{0}\in\mathfrak{T}_{+}^{1}(\mathcal{H})$
for each $i=\overline{1,k}$. It is clear that
$\sum_{i=1}^{k}A_{i}^{0}=A_{0}$. It follows from $\mathrm{(i)}$ that
$\lim_{n\rightarrow+\infty}H_{\Phi}(A_{i}^{n})=H_{\Phi}(A_{i}^{0})$.
Hence Lemma \ref{face} implies a contradiction to (\ref{dc}).
Continuity of the function $H_{\Phi}$ on the set
$\mathfrak{T}_{+}^{k}(\mathcal{H})$ implies its boundedness on the
set $\mathfrak{S}_{k}(\mathcal{H})$. By corollary 1 in \cite{Sh-11}
the function $H^{k}_{\Phi}$ is continuous and bounded on the set
$\mathfrak{S}(\mathcal{H})$ and hence it is continuous on the set
$\mathfrak{T}_{+}(\mathcal{H})$.\vspace{5pt}
$\mathrm{(ii)\Rightarrow(iii)}$ directly follows from Lemma
\ref{uniform-c}. $\mathrm{(iii)\Rightarrow(i)}$ is obvious.
\vspace{5pt}
The last assertion of the theorem follows from theorem 2A in
\cite{Sh-11} (since the UA\nobreakdash-\hspace{0pt}property of a
bounded set obviously implies boundedness of the quantum entropy on
this set). $\square$\vspace{5pt}
\begin{remark}\label{PCE-operation-r}
The main assertion of Theorem \ref{PCE-property} (the implication
$\mathrm{(i)\Rightarrow(iii)}$) is based on the special properties
of the von Neumann entropy, it can not be proved by using only such
general properties of entropy-type functions as concavity, lower
semicontinuity and nonnegativity. The simplest example showing this
is given by the function $A\mapsto
R_{0}(\Phi(A))=\|\Phi(A)\|_{1}\log\mathrm{rank}(\Phi(A))$ -- the
output $0$\nobreakdash-\hspace{0pt}order Renyi entropy of the map
$\Phi$.\footnote{Indeed, if $\Phi(A)=\frac{1}{2}(A+UAU^{*})$, where
$U$ is an unitary having no eigenvectors, then
$R_{0}(\Phi(A))=\|\Phi(A)\|_{1}\log2$ for all
$A\in\mathfrak{T}^{1}_{+}(\mathcal{H})$ but it is easy to see that
the function $A\mapsto R_{0}(\Phi(A))$ is not continuous on the set
$\mathfrak{T}^{ 2}_{+}(\mathcal{H})$ on which
$R_{0}(A)=\|A\|_{1}\log2$.} The essential roles in the proof of
Theorem \ref{PCE-property} are played by the second inequality in
(\ref{H-fun-ineq}) and the part $"if"$ of the assertion of Lemma
\ref{face}. $\square$
\end{remark}\vspace{5pt}
If property $\mathrm{(iii)}$ in Theorem \ref{PCE-property} holds for
a positive linear map $\Phi$ then we may say roughly speaking that
this map "preserves continuity of the entropy". This motivates the
following definition.\vspace{5pt}
\begin{definition}\label{PCE-operation}
\textit{A positive linear map (correspondingly quantum operation or
quantum channel) $\Phi$, for which property $\mathrm{(iii)}$ in
Theorem \ref{PCE-property} holds, is called
PCE\nobreakdash-\hspace{0pt}map (correspondingly
PCE\nobreakdash-\hspace{0pt}operation or
PCE\nobreakdash-\hspace{0pt}channel).}
\end{definition}\vspace{5pt}
The simplest examples of PCE\nobreakdash-\hspace{0pt}maps are
completely positive linear maps with finite Kraus representations,
for which property $\mathrm{(i)}$ in Theorem \ref{PCE-property}
obviously holds.\vspace{5pt}
By the last assertion of Theorem \ref{PCE-property} to prove the
PCE\nobreakdash-\hspace{0pt}property of a map $\Phi$ it is
sufficient to show that
$$
\Phi(\mathrm{extr}\mathfrak{S}(\mathcal{H}))\subseteq\Lambda(\mathcal{A}),
$$
where $\Lambda$ is a finite composition of set-operations preserving
the UA\nobreakdash-\hspace{0pt}property (see proposition 4 in
\cite{Sh-11}) and $\mathcal{A}$ is a compact set on which the
entropy is continuous. This implies the following
PCE\nobreakdash-\hspace{0pt}condition.\vspace{5pt}
\begin{corollary}\label{PCE-property-sc} \textit{A map $\,\Phi$ in
$\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ is a
PCE\nobreakdash-\hspace{0pt}map if there exist separable Hilbert
space $\mathcal{K}$, family
$\{A_{\psi}\}_{\psi\in\mathcal{H},\|\psi\|=1}$ of operators
belonging to a particular compact subset $\mathcal{A}$ of
$\,\mathfrak{T}_{+}(\mathcal{K})$, on which the quantum entropy is
continuous, and family
$\,\{V_{\psi}\}_{\psi\in\mathcal{H},\|\psi\|=1}$ of linear
contractions from $\mathcal{K}$ to $\mathcal{H}'$ such that
$\,\Phi(|\psi\rangle\langle\psi|)=V_{\psi}A_{\psi}V_{\psi}^{*}$ for
each unit vector $\psi$ in $\mathcal{H}$.}
\end{corollary}\vspace{5pt}
If $\Phi$ is a quantum operation having the Kraus representation
with $k$ nonzero summands then the condition of Corollary
\ref{PCE-property-sc} trivially holds (with
$k$\nobreakdash-\hspace{0pt}dimensional Hilbert space
$\mathcal{K}$). Nontrivial application of Corollary
\ref{PCE-property-sc} is the proof of the
PCE\nobreakdash-\hspace{0pt}property for the following family of
quantum channels.\vspace{5pt}
\begin{example}\label{PCE-property-e} Let $\mathcal{H}_{a}$ be the Hilbert space
$\mathcal{L}_{2}([-a,+a])$, where $a<+\infty$, and
$\{U_{t}\}_{t\in\mathbb{R}}$ be the group of unitary operators in
$\mathcal{H}_{a}$ defined as follows
$$
(U_{t}\varphi)(x)=e^{-\mathrm{i}tx}\varphi(x),\quad\forall\varphi\in\mathcal{H}_{a}.
$$
For given probability density function $p(t)$ consider the quantum
channel
$$
\Phi_{p}^{a}:\mathfrak{T}(\mathcal{H}_{a})\ni
A\mapsto\int_{-\infty}^{+\infty}U_{t}A
U_{t}^{*}p(t)dt\in\mathfrak{T}(\mathcal{H}_{a}).
$$
In \cite{Sh-7} it is shown that the condition of Corollary
\ref{PCE-property-sc} holds for the channel $\Phi_{p}^{a}$ (with
some set of unitary operators $\{V_{\psi}\}$) provided that the
differential entropy of the distribution $p(t)$ is finite and that
the function $p(t)$ is bounded and monotonic on $(-\infty,-b]$ and
on $[+b,+\infty)$ for sufficiently large $b$. $\square$
\end{example}\vspace{5pt}
If property $\mathrm{(iii)}$ in Theorem \ref{PCE-property} holds for
two positive maps then it obviously holds for their composition,
hence this theorem implies the following result.\vspace{5pt}
\begin{corollary}\label{PCE-property-c}
\textit{If property $\mathrm{(i)}$ in Theorem \ref{PCE-property}
holds for positive linear bounded maps
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
and
$\,\Psi:\mathfrak{T}(\mathcal{H}')\rightarrow\mathfrak{T}(\mathcal{H}'')$
then it holds for the map
$\,\Psi\circ\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}'')$.}
\end{corollary}\vspace{5pt}
In quantum information theory the notion of the Convex Closure of
the Output Entropy (CCoOE) of a quantum channel (considered as a
function on the set of input states of this channel) is widely used
\cite{AB,Sh-7}. By generalizing the proof of proposition 2 in
\cite{Sh-7} it is easy to show that property $\mathrm{(i)}$ in
Theorem \ref{PCE-property} of a positive linear map $\Phi$ is
equivalent to continuity and boundedness of the CCoOE of this map on
the set $\mathfrak{S}(\mathcal{H})$. Hence Corollary
\ref{PCE-property-c} shows that \textit{continuity and boundedness
of the CCoOE of positive linear bounded maps\break
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
and
$\,\Psi:\mathfrak{T}(\mathcal{H}')\rightarrow\mathfrak{T}(\mathcal{H}'')$
imply continuity and boundedness of the CCoOE of the map
$\,\Psi\circ\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}'')$.}\vspace{5pt}
Lemma \ref{face} implies the PCE\nobreakdash-\hspace{0pt}analog of
Corollary \ref{bound-cont-c}.\vspace{5pt}
\begin{corollary}\label{PCE-property-convex}
\textit{Let $\,\Phi$ and $\,\Psi$ be maps in
$\,\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ and
$\lambda\in(0,1)$. $\,\lambda\Phi+(1-\lambda)\Psi$ is a
PCE\nobreakdash-\hspace{0pt}map if and only if $\,\Phi$ and $\,\Psi$
are PCE\nobreakdash-\hspace{0pt}maps.}
\end{corollary}\vspace{5pt}
Thus the set of all PCE\nobreakdash-\hspace{0pt}maps is convex and
forms a face of the convex set
$\,\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$. This face
obviously contains the face of all maps in
$\,\mathfrak{L}_{\leq1}^{+}(\mathcal{H},\mathcal{H}')$ with
continuous output entropy.\vspace{5pt}
\subsection{The case of quantum operation}
Theorem \ref{PCE-property} implies the following
observation.\vspace{5pt}
\begin{property}\label{PCE-conditions}
\textit{Let $\,\Phi$ be a quantum operation in $\,\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}')$ and $\,\widetilde{\Phi}$ be its
complementary operation. The following properties are equivalent:}
\begin{enumerate}[(i)]
\item \textit{$\,\Phi$ is a PCE\nobreakdash-\hspace{0pt}operation;}
\item \textit{$\,\widetilde{\Phi}$ is a PCE\nobreakdash-\hspace{0pt}operation.}
\end{enumerate}\vspace{5pt}
\textit{In terms of the set $\,\{V_{i}\}_{i=1}^{+\infty}$ of Kraus
operators of the operation $\,\Phi$ a sufficient condition for
$\;\mathrm{(i)-(ii)}\,$ can be expressed in one of the following
forms:}
\begin{enumerate}[a)]
\item \textit{the function
$\,\varphi\mapsto
H\left(\left\{\|V_{i}|\varphi\rangle\|^{2}\right\}_{i=1}^{+\infty}\right)\,$
is continuous on the space $\mathcal{H}$;}
\item \textit{one of the conditions in Proposition \ref{CE-conditions-c} holds for the sequence $\,\{V_{i}\}_{i=1}^{+\infty}$.}
\end{enumerate}\vspace{5pt}
\textit{If $\,\Phi$ is a PCE\nobreakdash-\hspace{0pt}operation and
$\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all $i\neq j$ then
$\,\mathrm{a)}$ holds. }
\end{property}\vspace{5pt}
\textbf{Proof.} By Theorem \ref{PCE-property} equivalence of
$\mathrm{(i)}$ and $\mathrm{(ii)}$ follows from coincidence of the
output entropies of the operations $\Phi$ and $\widetilde{\Phi}$ on
the set of pure states.
Sufficiency of condition a) follows from the corresponding assertion
of Corollary \ref{PCE-conditions+} in Section 6 below (with
$\overline{V}_{n}=\{V_{i}\}_{i=1}^{+\infty}$ for all $n$).
Sufficiency of condition b) follows from the first assertion of this
proposition (since continuity of the output entropy implies the
PCE\nobreakdash-\hspace{0pt}property).
Necessity of condition a) in the case
$\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all $i\neq j$ is
obvious since in this case
$$
H\left(\sum_{i=1}^{+\infty}V_{i}|\varphi\rangle\langle\varphi|V^{*}_{i}\right)
=
H\left(\left\{\|V_{i}|\varphi\rangle\|^{2}\right\}_{i=1}^{+\infty}\right).\;\square
$$
By Corollary \ref{tensor-pr} the tensor product of two quantum
operations with continuous output entropy is a quantum operation
with continuous output entropy. The
PCE\nobreakdash-\hspace{0pt}property is not preserved in general
with respect to tensor products. The simplest example showing this
is the quantum channel
$\mathrm{Id}_{\mathcal{H}}\otimes\widetilde{\mathrm{Id}}_{\mathcal{H}}$,
which is not a PCE\nobreakdash-\hspace{0pt}channel if
$\dim\mathcal{H}=+\infty$. By Corollary \ref{tensor-pr} and
Proposition \ref{PCE-conditions} $\,\Phi\otimes\Psi$ is a
PCE\nobreakdash-\hspace{0pt}operation if either the operations
$\Phi$ and $\Psi$ or the operations $\widetilde{\Phi}$ and
$\widetilde{\Psi}$ have continuous output entropy. The following
proposition contains several observations concerning the
PCE\nobreakdash-\hspace{0pt}property of the map $\Phi\otimes\Psi$.
\vspace{5pt}
\begin{property}\label{PCE-property-cp}
\textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
and
$\Psi:\mathfrak{T}(\mathcal{K})\rightarrow\mathfrak{T}(\mathcal{K}')$
be quantum operations.} \vspace{5pt}
A) \emph{If $\,\Psi$ is a finite-dimensional operation
\textup{(}$\max\{\dim\mathcal{K},\dim\mathcal{K}'\}<+\infty$\textup{)}
then $\,\Phi\otimes\Psi$ is a PCE\nobreakdash-\hspace{0pt}operation
if and only if $\,\Phi$ is a
PCE\nobreakdash-\hspace{0pt}operation.}\footnote{The condition
$\dim\mathcal{K}<+\infty$ is essential. This can be shown by the
example $\Phi=\mathrm{Id}_{\mathcal{H}}$ and
$\Psi=\widetilde{\mathrm{Id}}_{\mathcal{K}}$.} \vspace{5pt}
B) \emph{If $\,\dim\mathcal{H}=\dim\mathcal{K}=+\infty$ and $\,\Psi$
is a quantum channel such that its complementary channel
$\,\widetilde{\Psi}$ has finite output entropy then the following
properties are equivalent:}
\begin{enumerate}[(i)]
\item \emph{$H(\Phi\otimes\Psi(|\varphi\rangle\langle\varphi|))<+\infty$ for
any unit vector $\varphi\in\mathcal{H}\otimes\mathcal{K}$;}
\item \emph{$H(A)<+\infty\,\Rightarrow\, H(\Phi\otimes\Psi(A))<+\infty$ for
any operator $A\in\mathfrak{T}_{+}(\mathcal{H}\otimes\mathcal{K})$;}
\item \emph{$\Phi\otimes\Psi$ is a
PCE\nobreakdash-\hspace{0pt}operation;}
\end{enumerate}
\emph{If $\,\Phi$ is a quantum channel then the condition of
finiteness of the output entropy of the channel $\,\widetilde{\Psi}$
can be replaced by the condition}
$$
\min\left\{H_{\widetilde{\Phi}}(\mathrm{Tr}_{\mathcal{K}}|\varphi\rangle\langle\varphi|),H_{\widetilde{\Psi}}(\mathrm{Tr}_{\mathcal{H}}|\varphi\rangle\langle\varphi|)\right\}
<+\infty\quad\forall\varphi\in\mathcal{H}\otimes\mathcal{K}.
$$
\end{property}
\vspace{5pt}
\begin{remark}\label{PCE-property-cp-r}
By Theorem \ref{PCE-property} $\Phi\otimes\Psi$ is a
PCE\nobreakdash-\hspace{0pt}operation if and only if the function
$\varphi\mapsto H(\Phi\otimes\Psi(|\varphi\rangle\langle\varphi|))$
is continuous and bounded on the unit sphere of
$\mathcal{H}\otimes\mathcal{K}$. Proposition \ref{PCE-property-cp}
shows that continuity and boundedness of this function follows from
its finiteness (provided the condition of this proposition holds).
Proposition \ref{PCE-property-cp} also shows that the operation
$\Phi\otimes\Psi$ \emph{preserves continuity of the entropy} (the
PCE\nobreakdash-\hspace{0pt}property) if it \emph{preserves
finiteness of the entropy} (property $\mathrm{(ii)}$).
\end{remark} \vspace{5pt}
\textbf{Proof.} A) The PCE\nobreakdash-\hspace{0pt}property of the
operation $\Phi\otimes\Psi$ obviously implies the same property of
the operation $\Phi$. To prove the converse implication it is
sufficient to show that $\mathrm{Id}_{\mathcal{H}'}\otimes\Psi$ and
$\Phi\otimes\mathrm{Id}_{\mathcal{K}}$ are
PCE\nobreakdash-\hspace{0pt}operations.
The operation $\mathrm{Id}_{\mathcal{H}'}\otimes\Psi$ has the
PCE\nobreakdash-\hspace{0pt}property since it has a finite Kraus
representation. By Theorem \ref{PCE-property} the
PCE\nobreakdash-\hspace{0pt}property of the operation
$\Phi\otimes\mathrm{Id}_{\mathcal{K}}$ follows from continuity and
boundedness of the function
\begin{equation}\label{function}
\mathrm{extr}\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})\ni\omega\mapsto
H_{\Phi\otimes\mathrm{Id}_{\mathcal{K}}}(\omega)=
H_{\widetilde{\Phi}\otimes\widetilde{\mathrm{Id}}_{\mathcal{K}}}(\omega)=H\left(\widetilde{\Phi}\left(\mathrm{Tr}_{\mathcal{K}}\omega\right)\right).
\end{equation}
Since the map $\omega\mapsto\mathrm{Tr}_{\mathcal{K}}\omega$ is a
continuous surjection from
$\mathrm{extr}\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})$ onto
$\mathfrak{S}_{k}(\mathcal{H})$, where $k=\dim\mathcal{K}$, the
function (\ref{function}) is continuous and bounded if the function
$H_{\widetilde{\Phi}}$ is continuous and bounded on the set
$\mathfrak{S}_{k}(\mathcal{H})$. By Theorem \ref{PCE-property} and
Proposition \ref{PCE-conditions} this holds if (and only if) $\Phi$
is a PCE\nobreakdash-\hspace{0pt}operation.
B) By applying Corollary \ref{tensor-pr+} to the operation
$\widetilde{\Phi}$ and to the channel $\widetilde{\Psi}$ we obtain
that $\mathrm{(i)}$ implies continuity of the output entropy of the
operation $\widetilde{\Phi}\otimes\widetilde{\Psi}$ (since
$H(\Phi\otimes\Psi(|\varphi\rangle\langle\varphi|))=H(\widetilde{\Phi}\otimes\widetilde{\Psi}(|\varphi\rangle\langle\varphi|))$).
Hence $\mathrm{(iii)}$ follows from Proposition
\ref{PCE-conditions}. It is clear that
$\mathrm{(iii)\Rightarrow(ii)}$ and $\mathrm{(ii)\Rightarrow(i)}$.
$\square$ \vspace{5pt}
\begin{remark}\label{PCE-property-cp-r+}
Proposition \ref{PCE-property-cp} with
$\Psi=\mathrm{Id}_{\mathcal{K}}$ gives the conditions for the
PCE\nobreakdash-\hspace{0pt}property of the operation
$\Phi\otimes\mathrm{Id}_{\mathcal{K}}$. The
PCE\nobreakdash-\hspace{0pt}property for the tensor product of two
quantum operations $\Phi\in\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}')$ and $\Psi\in\mathfrak{F}_{\leq
1}(\mathcal{K},\mathcal{K}')$ can be proved by showing that either
$\Phi\otimes\mathrm{Id}_{\mathcal{K}'}$ and
$\mathrm{Id}_{\mathcal{H}}\otimes\Psi$ or
$\Phi\otimes\mathrm{Id}_{\mathcal{K}}$ and
$\mathrm{Id}_{\mathcal{H}'}\otimes\Psi$ are
PCE\nobreakdash-\hspace{0pt}operations since
$\Phi\otimes\Psi=\mathrm{Id}_{\mathcal{H}'}\otimes\Psi\circ\Phi\otimes\mathrm{Id}_{\mathcal{K}}=
\Phi\otimes\mathrm{Id}_{\mathcal{K}'}\circ\mathrm{Id}_{\mathcal{H}}\otimes\Psi$.
Note that the PCE\nobreakdash-\hspace{0pt}property of the operation
$\Phi\otimes\Psi$ does not imply the
PCE\nobreakdash-\hspace{0pt}property of the above components. To
show this it is sufficient to consider the example
$\Phi=\widetilde{\mathrm{Id}}_{\mathcal{H}}$ and
$\Psi=\widetilde{\mathrm{Id}}_{\mathcal{K}}$ with
$\dim\mathcal{H}=\dim\mathcal{K}=+\infty$.
\end{remark}
\section{The output entropies of a pair\\ of complementary quantum operations}
The output entropies of two complementary quantum operations
coincide on the set of pure input states but they are substantially
different functions on the whole space of input states. Nevertheless
the following relation between continuity properties of these
functions can be established.\vspace{5pt}
\begin{property}\label{basic-cont-cond+}
\textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
be a quantum operation and $\;\widetilde{\Phi}\,$ be its
complementary operation. Let $\mathcal{A}$ be a subset of
$\,\mathfrak{T}_{+}(\mathcal{H})$ such that
$\min\{H_{\Phi}(A),H_{\widetilde{\Phi}}(A) \}<+\infty$ for all
$A\in\mathcal{A}$. Then continuity of the quantum entropy on the set
$\mathcal{A}$ implies continuity of the function $\,A\mapsto
\left(H_{\Phi}(A)-H_{\widetilde{\Phi}}(A)\right)$ on the set
$\,\mathcal{A}$.}
\end{property}\vspace{5pt}
The assertion of Proposition \ref{basic-cont-cond+} follows from the
more general assertion of Proposition \ref{g-basic-cont-cond+} in
Section 6 below.\vspace{5pt}
\begin{remark}\label{basic-cont-cond+r} If $\Phi$ is a quantum channel then
$\,H_{\Phi}(\rho)-H_{\widetilde{\Phi}}(\rho)\,$ is the coherent
information $I_{c}(\rho,\Phi)$ of this channel at a state $\rho$
\cite{H-SSQT, N&Ch}.
\end{remark}
\vspace{5pt}
\begin{corollary}\label{basic-cont-cond}
\textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
be a quantum channel and $\;\widetilde{\Phi}\,$ be its complementary
channel. If any two functions from the triple
$\{H,H_{\Phi},H_{\widetilde{\Phi}}\}$ are continuous on a particular
set $\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$ then the third
one is also continuous on this set.}
\emph{The above assertion holds for a quantum operation
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
if}
\begin{equation}\label{p-channel}
\lambda^{*}\left(\sqrt{I_{\mathcal{H}}-\Phi^{*}(I_{\mathcal{H}'})}\right)<+\infty.\footnote{The
parameter $\lambda^{*}(\cdot)$ is defined in Proposition \ref{s-map}
in Section 3.2.}
\end{equation}
\end{corollary
\textbf{Proof.} By representations (\ref{Stinespring-rep}) and
(\ref{c-channel}) the first assertion of the corollary follows from
Proposition \ref{basic-cont-cond+} and Proposition
\ref{cont-cond-a-1} in the Appendix.\vspace{5pt}
The second assertion of the corollary is derived from the first one
by means of Lemma \ref{inv-imp} below since by representations
(\ref{Stinespring-rep}) and (\ref{c-channel}) we have
$\Phi=\Theta\circ\Lambda$ and
$\widetilde{\Phi}=\widetilde{\Theta}\circ\Lambda$, where
$\Theta(\cdot)=\mathrm{Tr}_{\mathcal{H}''}(\cdot)$ is a quantum
channel from $\mathfrak{T}(\mathcal{H}'\otimes\mathcal{H}'')$ into
$\mathfrak{T}(\mathcal{H}')$ and $\Lambda(\cdot)=V(\cdot)V^{*}$ is a
quantum operation from $\mathfrak{T}(\mathcal{H})$ into
$\mathfrak{T}(\mathcal{H}'\otimes\mathcal{H}'')$.
$\square$\vspace{5pt}
\begin{lemma}\label{inv-imp}
\emph{Let $V$ be a linear contraction from $\mathcal{H}$ into
$\mathcal{H}'$ such that \break
$\lambda^{*}\left(\sqrt{I_{\mathcal{H}}-V^{*}V}\right)<+\infty$.
Then continuity of the function $A\mapsto H(VAV^{*})$ on a
particular set $\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$
implies\footnote{In fact, "is equivalent to", since the converse
implication holds for an arbitrary contraction $V$ by Theorem
\ref{PCE-property}.} continuity of the quantum entropy on this set.}
\end{lemma}\vspace{5pt}
\textbf{Proof.} Consider the quantum channel
$$
\mathfrak{T}(\mathcal{H})\ni
A\mapsto\Psi(A)=VAV^{*}\oplus\sqrt{I_{\mathcal{H}}-V^{*}V}A\sqrt{I_{\mathcal{H}}-V^{*}V}\in
\mathfrak{T}(\mathcal{H}'\oplus\mathcal{H}).
$$
By Proposition \ref{s-map} the function $A\mapsto
H(\sqrt{I_{\mathcal{H}}-V^{*}V}A\sqrt{I_{\mathcal{H}}-V^{*}V})$ is
continuous on the set $\mathfrak{T}_{+}(\mathcal{H})$. Hence the
function $A\mapsto H_{\Psi}(A)$ is continuous on the set
$\mathcal{A}$. Since the complementary channel $\widetilde{\Psi}$
has two dimensional output space the assertion of the lemma follows
from the first part of Corollary \ref{basic-cont-cond}.
$\square$\vspace{5pt}
\begin{corollary}\label{cont-cond+++p}
\textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
be a quantum channel (or quantum operation satisfying condition
(\ref{p-channel})) such that the complementary channel (operation)
$\,\widetilde{\Phi}$ has finite output entropy.\footnote{The
sufficient conditions for this property expressed in terms of the
Kraus operators of the operation $\Phi$ is presented in Proposition
\ref{CE-conditions-c} in Section 3.2.} Then the function $A\mapsto
H_{\Phi}(A)$ is continuous on a set
$\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$ if and only if the
quantum entropy is continuous on this set.}\vspace{5pt}
\end{corollary}
The simplest class of quantum channels for which the condition of
Corollary \ref{cont-cond+++p} holds consists of channels with finite
Kraus representation, for which complementary channels have finite
dimensional output. \vspace{5pt}
\begin{corollary}\label{basic-cont-cond++}\footnote{This corollary can be considered as a generalization
of Proposition \ref{cont-cond-a-2}, since its application to the
channel $\Phi(A)=\!\sum_{i=1}^{+\infty}\langle
i|A|i\rangle|i\rangle\langle i|$ implies the assertion of that
proposition.} \textit{Let
$\;\Phi(\cdot)=\sum_{i=1}^{+\infty}V_{i}(\cdot)V_{i}^{*}$ be a
quantum channel (or quantum operation satisfying condition
(\ref{p-channel})) in
$\;\mathfrak{F}_{\leq1}(\mathcal{H},\mathcal{H}')$ such that
$\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all sufficiently
large $i\neq j$. Then continuity of the function $A\mapsto
H_{\Phi}(A)$ on a set
$\,\mathcal{A}\subset\mathfrak{T}_{+}(\mathcal{H})$ implies
continuity of the quantum entropy on the set
$\mathcal{A}$.}\vspace{5pt}
\end{corollary
\textbf{Proof.} Note first that the complementary operation
$\widetilde{\Phi}$ can be represented as follows
$$
\widetilde{\Phi}(A)=\sum_{i,j=1}^{+\infty}\mathrm{Tr}\left[V_{i}AV_{j}^{*}\right]|i\rangle
\langle j|,\quad A\in\,\mathfrak{T}(\mathcal{H}),
$$
where $\{|i\rangle\}$ is an orthonormal basis in the output space
$\mathcal{H}''$ of this operation.
Suppose $\;\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all $i,j\geq
n,i\neq j$. Let
$\Phi_{1}(\cdot)=\sum_{i=1}^{n-1}V_{i}(\cdot)V_{i}^{*}$ and
$\Phi_{2}(\cdot)=\sum_{i=n}^{+\infty}V_{i}(\cdot)V_{i}^{*}$ be
quantum operations. By Lemma \ref{face} continuity of the function
$\mathcal{A}\ni A\mapsto H(\Phi(A))=H(\Phi_{1}(A)+\Phi_{2}(A))$
implies continuity of the function $\mathcal{A}\ni A\mapsto
H(\Phi_{2}(A))$. By the condition
$$
H(\Phi_{2}(A))=\!\sum_{i=n}^{+\infty}H(V_{i}A
V_{i}^{*})+H\!\left(\left\{\mathrm{Tr}V_{i}AV_{i}^{*}\right\}_{i=n}^{+\infty}\right)=\!\sum_{i=n}^{+\infty}H(V_{i}A
V_{i}^{*})+H(\widetilde{\Phi}_{2}(A)),
$$
where
$\widetilde{\Phi}_{2}(A)=\sum_{i=n}^{+\infty}\mathrm{Tr}\left[V_{i}AV_{i}^{*}\right]|i\rangle
\langle i|$. Since the both terms in the right side of the above
expression are lower semicontinuous functions of $A$, continuity of
the function $\mathcal{A}\ni A\mapsto H(\Phi_{2}(A))$ implies
continuity of the function\break $\mathcal{A}\ni A\mapsto
H(\widetilde{\Phi}_{2}(A))$.
Consider the quantum channel
$\,\Pi(\cdot)=P(\cdot)P+(I_{\mathcal{H}''}-P)(\cdot)(I_{\mathcal{H}''}-P)\,$
in $\;\mathfrak{F}_{=1}(\mathcal{H}'',\mathcal{H}'')$, where
$P=\sum_{i=1}^{n-1}|i\rangle \langle i|$. Since
$$
\Pi(\widetilde{\Phi}(A))=\sum_{i,j=1}^{n-1}\mathrm{Tr}\left[V_{i}AV_{j}^{*}\right]|i\rangle
\langle j|+\widetilde{\Phi}_{2}(A),
$$
continuity of the function $\mathcal{A}\ni A\mapsto H(\Phi_{2}(A))$
implies continuity of the function $\mathcal{A}\ni A\mapsto
H(\Pi(\widetilde{\Phi}(A)))$ by Lemma \ref{face}, which is
equivalent to continuity of the function $\mathcal{A}\ni A\mapsto
H(\widetilde{\Phi}(A))$ by Corollary \ref{cont-cond+++p}. Hence the
function $\mathcal{A}\ni A\mapsto H(A)$ is continuous by Corollary
\ref{basic-cont-cond}. $\square$ \vspace{5pt}
\begin{remark}\label{corollaries}
The assertions of Corollaries \ref{basic-cont-cond},
\ref{cont-cond+++p} and \ref{basic-cont-cond++} do not hold for a
quantum operation $\Phi$ not satisfying condition (\ref{p-channel}).
This follows from the next corollary.
\end{remark}\vspace{5pt}
\begin{corollary}\label{basic-cont-cond+c}
\textit{Let $\,\Phi$ be a quantum operation in
$\;\mathfrak{F}_{\leq1}(\mathcal{H},\mathcal{H}')$ and}
$$
\mathfrak{T}_{\Phi}=\{A\in\mathfrak{T}_{1}(\mathcal{H})\,|\,\min\{H_{\Phi}(A),H_{\widetilde{\Phi}}(A)\}<+\infty\}.
$$
\textit{If
$\,\lambda^{*}\!\left(\!\sqrt{\Phi^{*}(I_{\mathcal{H}'})}\right)<+\infty$
then the function
$A\mapsto\left(H_{\Phi}(A)-H_{\widetilde{\Phi}}(A)\right)$ is
continuous on the set $\,\mathfrak{T}_{\Phi}$ and its absolute value
does not exceed
$\,\lambda^{*}\!\left(\!\sqrt{\Phi^{*}(I_{\mathcal{H}'})}\right)$.}\vspace{5pt}
\emph{If the functions $\rho\mapsto H_{\Phi}(\rho)$ and $\rho\mapsto
H_{\widetilde{\Phi}}(\rho)$ are continuous on the set
$\mathfrak{S}(\mathcal{H})$ then the operator
$\,\Phi^{*}(I_{\mathcal{H}'})$ satisfies the above condition.}
\end{corollary}\vspace{5pt}
\textbf{Proof.} Since $\Phi^{*}(I_{\mathcal{H}'})=V^{*}V$, where $V$
is the Stinespring contraction for the operation $\Phi$, the first
assertion of the corollary follows from Propositions \ref{s-map} and
\ref{basic-cont-cond+} while the second one -- from Proposition
\ref{s-map} and Proposition \ref{cont-cond-a-1} in the Appendix.
$\square$\vspace{5pt}
\section{The output entropy as a function\\ of a pair (map, input
state)}
In analysis of continuity of information characteristics of a
quantum channel as functions of a channel it is necessary to
consider the output entropy as a function of a pair (channel, input
state) and to explore continuity of this function with respect to
the Cartesian product (coordinate-wise) topology on the set of such
pairs \cite{Sh-H,L&S}. The same problem appears in study of quantum
channels by means of their approximation by quantum operations
\cite{Sh-H}.\vspace{5pt}
\subsection{The general continuity condition}
The central result of this subsection is the following
proposition.\vspace{5pt}
\begin{property}\label{PCE-property+}
\emph{Let $\,\{\Phi_{n}\}$ be a sequence of maps in
$\,\mathfrak{L}^{+}_{\leq1}(\mathcal{H},\mathcal{H}')$ converging to
a map $\,\Phi_{0}$. The following properties are equivalent:}
\begin{enumerate}[(i)]
\item \emph{$\displaystyle\lim_{n\rightarrow+\infty}H_{\Phi_{n}}(A_{n})=H_{\Phi_{0}}(A_{0})<+\infty$
for any sequence
$\,\{A_{n}\}\!\subset\mathfrak{T}^{1}_{+}(\mathcal{H})$ converging
to an operator $A_{0}$;}\footnote{
$\,\mathfrak{T}^{1}_{+}(\mathcal{H})$ is the set of 1-rank positive operators in $\mathcal{H}$.}
\item \emph{$\displaystyle\lim_{n\rightarrow+\infty}H(A_{n})=H(A_{0})<+\infty\;\,\Rightarrow\;\,
\lim_{n\rightarrow+\infty}H_{\Phi_{n}}(A_{n})=H_{\Phi_{0}}(A_{0})<+\infty
$ for any sequence $\,\{A_{n}\}\subset\mathfrak{T}_{+}(\mathcal{H})$
converging to an operator $A_{0}$.}
\end{enumerate}
\end{property}
\vspace{5pt}
\textbf{Proof.} It suffice to show that
$\mathrm{(i)\Rightarrow(ii)}$. For given natural $k$ the obvious
modification of the first part of the proof of the implication
$\mathrm{(i)\Rightarrow(ii)}$ in Theorem \ref{PCE-property} implies
validity of $\mathrm{(ii)}$ for any sequence
$\{A_{n}\}\subset\mathfrak{T}_{+}^{k}(\mathcal{H})$. \vspace{5pt}
Suppose there exist $\varepsilon>0$ and a sequence
$\{\rho_{n}\}\subset\mathfrak{S}(\mathcal{H})$ converging to a state
$\rho_{0}$ such that
\begin{equation}\label{discont}
\lim_{n\rightarrow+\infty}H(\rho_{n})=H(\rho_{0})<+\infty\quad\textup{and}\quad
\lim_{n\rightarrow+\infty}H_{\Phi_{n}}(\rho_{n})>H_{\Phi_{0}}(\rho_{0})+3\varepsilon.
\end{equation}
It is easy to see that $\mathrm{(i)}$ implies
$\limsup_{n\rightarrow+\infty}\sup_{\rho\in\mathrm{extr}\mathfrak{S}(\mathcal{H})}H_{\Phi_{n}}(\rho)<+\infty$.
Hence we may consider that
\begin{equation}\label{bound}
\sup_{n>0}\sup_{\rho\in\mathrm{extr}\mathfrak{S}(\mathcal{H})}H_{\Phi_{n}}(\rho)<+\infty.
\end{equation}
By Lemma \ref{nv-2} in the Appendix used with (\ref{out-ent-est}),
(\ref{u-h-c}) and (\ref{bound}) we may assume existence of a
sequence
$\{A^{k}_{n}\}_{n}\subset\mathfrak{T}^{k}_{+}(\mathcal{H})$,
converging to an operator $A^{k}_{0}$ ($k\in\mathbb{N}$), such that
$B^{k}_{n}=\rho_{n}-A^{k}_{n}\geq 0$,
$$
H(\Phi_{n}(B^{k}_{n}))<\varepsilon\quad \textup{and} \quad
\gamma_{n}=\mathrm{Tr}\Phi_{n}(\rho_{n})h_{2}\left(\frac{\mathrm{Tr}\Phi_{n}(B^{k}_{n})}{\mathrm{Tr}\Phi_{n}(\rho_{n})}\right)<\varepsilon\quad
\forall n\geq0.
$$
By the remark at the begin of the proof
$|H(\Phi_{n}(A_{n}^{k}))-H(\Phi_{0}(A_{0}^{k}))|<\varepsilon$ for
all sufficiently large $n$. Since
$\Phi_{n}(\rho_{n})=\Phi_{n}(A^{k}_{n})+\Phi_{n}(B^{k}_{n})$ for
each $n\geq0$, by using inequality (\ref{H-fun-ineq}) we obtain
$$
H(\Phi_{n}(\rho_{n}))-H(\Phi_{0}(\rho_{0}))\leq
H(\Phi_{n}(A_{n}^{k}))-H(\Phi_{0}(A_{0}^{k}))+H(\Phi_{n}(B^{k}_{n}))+\gamma_{n}<3\varepsilon
$$
for all sufficiently large $n$, contradicting to (\ref{discont}).
By this contradiction and lower semicontinuity of the quantum
entropy property $\mathrm{(ii)}$ holds for an arbitrary sequence
$\{A_{n}\}\subset\mathfrak{S}(\mathcal{H})$. Its validity for an
arbitrary sequence $\{A_{n}\}\subset\mathfrak{T}_{+}(\mathcal{H})$
can be easily shown by using (\ref{out-ent-est}) and (\ref{bound}).
$\square$\vspace{5pt}
\begin{corollary}\label{PCE-property+c}
\textit{For an arbitrary subset $\mathcal{A}$ of
$\,\mathfrak{T}_{+}(\mathcal{H})$, on which the quantum entropy is
continuous, the function $\,(\Phi,A)\,\mapsto\, H_{\Phi}(A)$ is
continuous on the set $\,\mathfrak{F}^{k}_{\leq
1}(\mathcal{H},\mathcal{H}')\times\mathcal{A}$ for each natural
$k$.}\footnote{$\mathfrak{F}^{k}_{\leq 1}(\mathcal{H},\mathcal{H}')$
is the set of all quantum operations from
$\mathfrak{T}(\mathcal{H})$ to $\mathfrak{T}(\mathcal{H}')$ having
the Kraus representation consisting of $\,\leq k$ summands.}
\end{corollary}\vspace{5pt}
Let $\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ be the set of
all sequences $\{V_{i}\}_{i=1}^{+\infty}$ of linear bounded
operators from $\mathcal{H}$ into $\mathcal{H}'$ such that
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}\leq I_{\mathcal{H}}$ endowed
with the Cartesian product of the strong operator topology (the
topology of coordinate-wise strong operator convergence). In what
follows a sequence $\{V_{i}\}_{i=1}^{+\infty}$ in
$\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ will be called
$\,\mathfrak{V}$\emph{\nobreakdash-\hspace{0pt}vector} and will be
denoted $\overline{V}$, the corresponding operator
$\sum_{i=1}^{+\infty}V^{*}_{i}V_{i}$ will be denoted
$|\hspace{-2.7pt}|\overline{V}|\hspace{-2.7pt}|$.\vspace{5pt}
Let $\{|i\rangle\}_{i=1}^{+\infty}$ be an orthonormal basis in a
separable Hilbert space $\mathcal{H}''$. Consider the maps
\begin{equation}\label{Phi-map}
\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')\ni\overline{V}\mapsto
\Phi[\overline{V}](\cdot)=\sum_{i=1}^{+\infty}V_{i}(\cdot)V_{i}^{*}\in\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}')
\end{equation}
and
\begin{equation}\label{Phi-c-map}
\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')\ni\overline{V}\mapsto
\widetilde{\Phi}\left[\overline{V}\right](\cdot)=\sum_{i,j=1}^{+\infty}\mathrm{Tr}V_{i}(\cdot)V_{j}^{*}|i\rangle\langle
j|\in\mathfrak{F}_{\leq 1}(\mathcal{H},\mathcal{H}'').\!
\end{equation}
The following lemma shows, in particular, continuity of these maps
on the subset
$\mathfrak{V}_{=1}(\mathcal{H},\mathcal{H}')=\left\{\overline{V}\,|\,|\hspace{-2.7pt}|\overline{V}|\hspace{-2.7pt}|=I_{\mathcal{H}}\right\}$
of $\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ corresponding to
the set of quantum channels.\vspace{5pt}
\begin{lemma}\label{phi-map-cont}
\emph{Let $\,\{\overline{V}_n\}$ be a sequence of
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vectors in
$\,\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ converging to a
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vector $\,\overline{V}_0$.
The following properties are equivalent:}
\begin{enumerate}[(i)]
\item \emph{the sequence $\left\{|\hspace{-2.7pt}|\overline{V}_{n}|\hspace{-2.7pt}|\right\}$ weakly converges to the operator $\,|\hspace{-2.7pt}|\overline{V}_{0}|\hspace{-2.7pt}|$;}
\item \emph{the sequence $\left\{\Phi[\overline{V}_{n}]\right\}$ strongly converges to the map $\,\Phi[\overline{V}_{0}]$;}
\item \emph{the sequence $\left\{\widetilde{\Phi}[\overline{V}_{n}]\right\}$ strongly converges
to the map $\,\widetilde{\Phi}[\overline{V}_{0}]$.}
\end{enumerate}
\end{lemma}
\textbf{Proof.} $\mathrm{(i)}\Rightarrow\mathrm{(ii)}$ It suffice to
show that the sequence
$\left\{\Phi[\overline{V}_{n}](|\varphi\rangle\langle\varphi|)\right\}$
tends to the operator
$\Phi[\overline{V}_{0}](|\varphi\rangle\langle\varphi|)$ for
arbitrary unit vector $\varphi\in\mathcal{H}$. This can be done by
noting that
$\lim_{n\rightarrow+\infty}\langle\varphi||\hspace{-2.7pt}|\overline{V}_{n}|\hspace{-2.7pt}||\varphi\rangle=\langle\varphi||\hspace{-2.7pt}|\overline{V}_{0}|\hspace{-2.7pt}||\varphi\rangle$
implies
$$
\lim_{m\rightarrow+\infty}\sup_{n\geq0}\mathrm{Tr}\sum_{i>m}V^{n}_{i}\,|\varphi\rangle\langle\varphi|(V^{n}_{i})^{*}
=\lim_{m\rightarrow+\infty}\sup_{n\geq0}\sum_{i>m}\|V^{n}_{i}|\varphi\rangle\|^{2}=0\quad\forall\varphi\in\mathcal{H}.
$$
$\mathrm{(i)}\Rightarrow\mathrm{(iii)}$ It suffice to show that the
sequence
$\left\{\widetilde{\Phi}[\overline{V}_{n}](|\varphi\rangle\langle\varphi|)\right\}$
tends to the operator
$\widetilde{\Phi}[\overline{V}_{0}](|\varphi\rangle\langle\varphi|)$
for arbitrary unit vector $\varphi\in\mathcal{H}$. This can be done
by noting that $\mathrm{(i)}$ means
$\lim_{n\rightarrow+\infty}\mathrm{Tr}\,\widetilde{\Phi}[\overline{V}_{n}](|\varphi\rangle\langle\varphi|)
=\mathrm{Tr}\,\widetilde{\Phi}[\overline{V}_{0}](|\varphi\rangle\langle\varphi|)$
and that weak convergence of a sequence
$\{A_{n}\}\subset\mathfrak{T}_{+}(\mathcal{H}'')$ to an operator
$A_{0}\in\mathfrak{T}_{+}(\mathcal{H}'')$ such that
$\lim_{n\rightarrow+\infty}\mathrm{Tr}A_{n}=\mathrm{Tr}A_{0}$
implies its convergence in the trace norm \cite{D-A}.\vspace{5pt}
$\mathrm{(ii)}\Rightarrow\mathrm{(i)}$ and
$\mathrm{(iii)}\Rightarrow\mathrm{(i)}$ are obvious since
$$
\langle\varphi|\,|\hspace{-2.7pt}|\overline{V}|\hspace{-2.7pt}|\,|\varphi\rangle=\mathrm{Tr}\,\Phi[\overline{V}](|\varphi\rangle\langle\varphi|)=
\mathrm{Tr}\,\widetilde{\Phi}[\overline{V}](|\varphi\rangle\langle\varphi|)\quad\forall\varphi\in\mathcal{H}.\;\square
$$
Proposition \ref{PCE-property+} implies the following
observation.\vspace{5pt}
\begin{corollary}\label{PCE-conditions+}
\textit{Let $\,\{\overline{V}_{n}\}$ be a sequence of
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vectors in
$\,\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ converging to a
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vector $\,\overline{V}_{0}$
such that property $\,\mathrm{(i)}$ in Lemma \ref{phi-map-cont}
holds and
\begin{equation}\label{h-function}
\lim_{n\rightarrow+\infty}H\left(\left\{\|V^{n}_{i}|\varphi_{n}\rangle\|^{2}\right\}_{i=1}^{+\infty}\right)=
H\left(\left\{\|V^{0}_{i}|\varphi_{0}\rangle\|^{2}\right\}_{i=1}^{+\infty}\right)<+\infty
\end{equation}
for any sequence $\{\varphi_{n}\}$ of vectors in $\mathcal{H}$
converging to a vector $\varphi_{0}$.} \emph{Then
$$
\lim_{n\rightarrow+\infty}H\left(\Phi[\overline{V}_{n}](A_{n})\right)=H\left(\Phi[\overline{V}_{0}](A_{0})\right)<+\infty
$$
and
$$
\lim_{n\rightarrow+\infty}H\left(\widetilde{\Phi}[\overline{V}_{n}](A_{n})\right)=H\left(\widetilde{\Phi}[\overline{V}_{0}](A_{0})\right)<+\infty
$$
for any sequence $\{A_{n}\}\subset\mathfrak{T}_{+}(\mathcal{H})$
converging to an operator $A_{0}$ such that
$\;\lim_{n\rightarrow+\infty}H(A_{n})=H(A_{0})<+\infty$.}\vspace{5pt}
\emph{The above requirements on the sequence
$\,\{\overline{V}_{n}\}$ can be replaced by one of the following
conditions:}
\begin{enumerate}[a)]
\item \textit{there exists a sequence $\{h_{i}\}_{i=1}^{+\infty}$ of nonnegative numbers such that
$
\sup_{n\geq0}\|\sum_{i=1}^{+\infty}h_{i}(V_{i}^{n})^{*}V_{i}^{n}\|<+\infty\quad\textit{and}\quad
\sum_{i=1}^{+\infty}e^{-\lambda
h_{i}}<+\infty\;\,\textit{for}\;\,\textit{all}\;\, \lambda>0;
$$}
\item \textit{property $\,\mathrm{(i)}$ in Lemma
\ref{phi-map-cont} holds for the sequence $\,\{\overline{V}_{n}\}$
and}
$$
\lim_{m\rightarrow+\infty}\sup_{n\geq0}H\left(\left\{\|V^{n}_{i}\|^{2}\right\}_{i>m}\right)=0.
$$
\end{enumerate}
\end{corollary}\vspace{5pt}
\textbf{Proof.} By using Proposition \ref{cont-cond-a-2} it is easy
to show that (\ref{h-function}) implies
$$
\lim_{n\rightarrow+\infty}H\left(\widetilde{\Phi}[\overline{V}_{n}](|\varphi_{n}\rangle\langle\varphi_{n}|)\right)
=H\left(\widetilde{\Phi}[\overline{V}_{0}](|\varphi_{0}\rangle\langle\varphi_{0}|)\right)<+\infty.
$$
for any sequence $\{\varphi_{n}\}$ of vectors in $\mathcal{H}$
converging to a vector $\varphi_{0}$. Hence the main assertion of
the corollary follows from Proposition \ref{PCE-property+} and Lemma
\ref{phi-map-cont} (since the output entropies of complementary
quantum operations coincide on the set of
$1$\nobreakdash-\hspace{0pt}rank operators).
Condition a) means that
$
\sup_{n\geq0}\sup_{\varphi\in\mathcal{H},
\|\varphi\|\leq1}\sum_{i=1}^{+\infty}h_{i}\|V^{n}_{i}|\varphi\rangle\|^{2}<+\infty,
$$
which implies
$\lim_{m\rightarrow+\infty}\sup_{n\geq0}\sum_{i>m}^{+\infty}\|V^{n}_{i}|\varphi\rangle\|^{2}=0$
for each $\varphi\in\mathcal{H}$, showing that property
$\mathrm{(i)}$ in Lemma \ref{phi-map-cont} holds for the sequence
$\{\overline{V}_{n}\}$. By Proposition \ref{H-cont-cond} it also
implies (\ref{h-function}) for any sequence $\{\varphi_{n}\}$ of
vectors in $\mathcal{H}$ converging to a vector $\varphi_{0}$.
Condition b) implies
$$
\lim_{m\rightarrow+\infty}\sup_{n\geq0}\left[H\left(\left\{\|V^{n}_{i}|\varphi_{n}\rangle\|^{2}\right\}_{i=1}^{+\infty}\right)-
H\left(\left\{\|V^{n}_{i}|\varphi_{n}\rangle\|^{2}\right\}_{i=1}^{m}\right)\right]=0
$$
for any sequence $\{\varphi_{n}\}$ of vectors in $\mathcal{H}$
converging to a vector $\varphi_{0}$. Indeed, since weak convergence
of the sequence $\{|\hspace{-2.7pt}|\overline{V}_{n}|\hspace{-2.7pt}|\}$ to the operator
$|\hspace{-2.7pt}|\overline{V}_{0}|\hspace{-2.7pt}|$ implies
$\lim_{n\rightarrow+\infty}\langle\varphi_{n}||\hspace{-2.7pt}|\overline{V}_{n}|\hspace{-2.7pt}||\varphi_{n}\rangle=
\langle\varphi_{0}||\hspace{-2.7pt}|\overline{V}_{0}|\hspace{-2.7pt}||\varphi_{0}\rangle $ and
hence
$
\lim_{m\rightarrow+\infty}\sup_{n\geq
0}\sum_{i>m}\|V^{n}_{i}|\varphi_{n}\rangle\|^{2}=0,
$
the above assertion can be proved by using (\ref{H-fun-ineq}) and
(\ref{u-h-c}). $\square$\vspace{5pt}
\begin{example}\label{cont-cond+++r++}
Let $\,\{\overline{V}_{n}=\{V^{n}_{i}\}\}$ be a sequence of
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vectors in
$\,\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ converging to a
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vector
$\,\overline{V}_{0}=\{V^{0}_{i}\}$ such that
$\mathrm{Ran}(V^{n}_{i})^{*}\perp(\mathrm{Ran}V^{n}_{j})^{*}$ and
$\|V^{n}_{i}\|^{2}\leq x_{i}\log^{-1}(i)$ for each $n$ and all
$i,j>m$,$\,i\neq j$, where $m\in\mathbb{N}$ and $\{x_{i}\}_{i>m}$ is
a given sequence of positive numbers converging to zero. Then
condition a) in Corollary \ref{PCE-conditions+} holds for the
sequence $\,\{\overline{V}_{n}\}$ with the sequence
$\{h_{i}\}_{i=1}^{+\infty}$, where $h_{i}=0\,$ if $\,i\leq m$ and
$h_{i}=x_{i}^{-1}\log(i)\,$ if $\,i>m$.
\end{example}\vspace{5pt}
\subsection{The continuity condition based on\\ the complementary relation}
By using the relation between complementary quantum operations the
following result can be established.\vspace{5pt}
\begin{property}\label{g-basic-cont-cond+}
\textit{Let $\,\{\Phi_{n}\}$ be a sequence of operations in
$\,\mathfrak{F}_{\leq 1}(\mathcal{H},\mathcal{H}')$ converging to an
operation $\,\Phi_{0}$ and $\,\{\widetilde{\Phi}_{n}\}$ be a
sequence of operations in $\,\mathfrak{F}_{\leq
1}(\mathcal{H},\mathcal{H}'')$ converging to an operation
$\,\widetilde{\Phi}_{0}$ such that $(\Phi_{n},\widetilde{\Phi}_{n})$
is a complementary pair for each $n=0,1,2..$. Let $\,\{A_{n}\}$ be a
sequence of operators in $\,\mathfrak{T}_{+}(\mathcal{H})$
converging to an operator $A_{0}$ such that
$$
\lim_{n\rightarrow+\infty}H(A_{n})=H(A_{0})<+\infty\;\;
\textit{and}\;\;\min\{H_{\Phi_{n}}(A_{n}),
H_{\widetilde{\Phi}_{n}}(A_{n})\}<+\infty,\;\; n\geq0.
$$ Then}
$$
\lim_{n\rightarrow+\infty}\left(H_{\Phi_{n}}(A_{n})-H_{\widetilde{\Phi}_{n}}(A_{n})\right)=
H_{\Phi_{0}}(A_{0})-H_{\widetilde{\Phi}_{0}}(A_{0})<+\infty.
$$
\end{property}\vspace{5pt}
\textbf{Proof.} Finiteness of the values $H_{\Phi_{n}}(A_{n})$ and
$H_{\widetilde{\Phi}_{n}}(A_{n})$ for each $n\geq0$ follows from the
definition of a complementary operation and inequality
(\ref{d-ineq}).
Let $\{A_{n}=\rho_{n}\}$ be a sequence of states in
$\mathfrak{S}(\mathcal{H})$ converging to a state $A_{0}=\rho_{0}$
such that
$\lim_{n\rightarrow+\infty}H(\rho_{n})=H(\rho_{0})<+\infty$.
Let
$a_{n}=H_{\Phi_{n}}(\rho_{n})-H_{\widetilde{\Phi}_{n}}(\rho_{n})$
for each $n\geq0$. By symmetry to prove that
$\lim_{n\rightarrow+\infty}a_{n}=a_{0}$ it is sufficient to show
that
\begin{equation}\label{rel-1}
\liminf_{n\rightarrow+\infty}a_{n}\geq a_{0}.
\end{equation}
Let $\mathcal{K}$ be a separable Hilbert space and
$\{|\varphi_{n}\rangle\}$ be a sequence of unit vectors in
$\mathcal{H}\otimes\mathcal{K}$ converging to a vector
$|\varphi_{0}\rangle$ such that
$\mathrm{Tr}_{\mathcal{K}}|\varphi_{n}\rangle\langle\varphi_{n}|=\rho_{n}$
for each $n\geq0$. Finiteness of the values $H_{\Phi_{n}}(\rho_{n})$
and $H_{\widetilde{\Phi}_{n}}(\rho_{n})$ implies
$$
b_{n}=H(\Phi_{n}\otimes\mathrm{Id}_{\mathcal{K}}(|\varphi_{n}\rangle\langle\varphi_{n}|)\|
\Phi_{n}(\rho_{n})\otimes\rho_{n})=a_{n}+c_{n},
$$
where
$c_{n}=\mathrm{Tr}\,\Phi_{n}\otimes\mathrm{Id}_{\mathcal{K}}(|\varphi_{n}\rangle\langle\varphi_{n}|)\cdot
I_{\mathcal{H}'}\otimes(-\log\rho_{n})$.\vspace{5pt}
Since the sequence $\{\Phi_{n}\otimes\mathrm{Id}_{\mathcal{K}}\}$
strongly converges to the operation
$\Phi_{0}\otimes\mathrm{Id}_{\mathcal{K}}$ we have
$\liminf_{n\rightarrow+\infty}b_{n}\geq b_{0}$ by lower
semicontinuity of the relative entropy. Hence to prove (\ref{rel-1})
we have to show that
\begin{equation}\label{rel-2}
\limsup_{n\rightarrow+\infty}c_{n}\leq c_{0}.
\end{equation}
For each $n$ consider the quantum channel
$\Psi_{n}=\Phi_{n}+\Delta_{n}$ in
$\mathfrak{F}_{=1}(\mathcal{H},\mathcal{H}')$, where
$\Delta_{n}(\cdot)=\sigma\mathrm{Tr}((I_{\mathcal{H}}-\Phi^{*}_{n}(I_{\mathcal{H}}))(\cdot))$
is a quantum operation defined by means of some state $\sigma$ in
$\mathfrak{S}(\mathcal{H}')$. It is clear that the sequence
$\{\Delta_{n}\}$ strongly converges to the operation $\Delta_{0}$.
Since
$\,\lim_{n\rightarrow+\infty}H(\rho_{n})=H(\rho_{0})<+\infty\,$ and
$$
H(\rho_{n})=\mathrm{Tr}\,\Psi_{n}\otimes\mathrm{Id}_{\mathcal{K}}(|\varphi_{n}\rangle\langle\varphi_{n}|)\cdot
I_{\mathcal{H}'}\otimes(-\log\rho_{n})= c_{n}+d_{n},\quad
n=0,1,2,...,
$$
where
$d_{n}=\mathrm{Tr}\Delta_{n}\otimes\mathrm{Id}_{\mathcal{K}}(|\varphi_{n}\rangle\langle\varphi_{n}|)\cdot
I_{\mathcal{H}'}\otimes(-\log\rho_{n})$, to prove (\ref{rel-2}) it
suffice to show that
\begin{equation}\label{rel-3}
\liminf_{n\rightarrow+\infty}d_{n}\geq d_{0}.
\end{equation}
We have $d_{n}=\mathrm{Tr}B_{n}(-\log\rho_{n})$, where
$B_{n}=\mathrm{Tr}_{\mathcal{H}'}\Delta_{n}\otimes\mathrm{Id}_{\mathcal{K}}(|\varphi_{n}\rangle\langle\varphi_{n}|)$.
Since $B_{n}\leq
B_{n}+\mathrm{Tr}_{\mathcal{H}'}\Phi_{n}\otimes\mathrm{Id}_{\mathcal{K}}(|\varphi_{n}\rangle\langle\varphi_{n}|)=\rho_{n}$,
we have $H(B_{n})<+\infty$ and hence
$d_{n}=H(B_{n})+H(B_{n}\|\rho_{n})+\eta(\mathrm{Tr}B_{n})+\mathrm{Tr}B_{n}-1$.
Lower semicontinuity of the quantum entropy and of the relative
entropy implies (\ref{rel-3}).
Thus the assertion of the proposition is proved in the case
$\{A_{n}\}\subset\mathfrak{S}(\mathcal{H})$. The general assertion
is easily derived from the above observation by noting that for
arbitrary sequence $\{A_{n}\}$ converging to zero inequality
(\ref{d-ineq}) and the inequality $H(V^{n}A_{n}(V^{n})^{*})\leq
H(A_{n})$, where $V^{n}$ is the Stinespring contraction for the
operations $\Phi_{n}$ and $\widetilde{\Phi}_{n}$, show that
$$
\lim_{n\rightarrow+\infty}H(A_{n})=0\quad\Rightarrow\quad\lim_{n\rightarrow+\infty}\left(H_{\Phi_{n}}(A_{n})-H_{\widetilde{\Phi}_{n}}(A_{n})\right)=0.
\;\square
$$
\begin{corollary}\label{g-basic-cont-cond+c}
\textit{Let $\,\{\overline{V}_{n}\}$ be a sequence of
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vectors in
$\,\mathfrak{V}_{\leq1}(\mathcal{H},\mathcal{H}')$ converging to a
$\,\mathfrak{V}$\nobreakdash-\hspace{0pt}vector $\,\overline{V}_{0}$
such that property $\,\mathrm{(i)}$ in Lemma \ref{phi-map-cont}
holds, and $\{A_{n}\}$ be a sequence of operators in
$\mathfrak{T}_{+}(\mathcal{H})$ converging to an operator $A_{0}$
such that $\;\lim_{n\rightarrow+\infty}H(A_{n})=H(A_{0})<+\infty$.
The following properties are equivalent:}
\begin{enumerate}[(i)]
\item \emph{$\;\lim_{n\rightarrow+\infty}H\left(\Phi[\overline{V}_{n}](A_{n})\right)=H\left(\Phi[\overline{V}_{0}](A_{0})\right)<+\infty$;}
\item \emph{$\;\lim_{n\rightarrow+\infty}H\left(\widetilde{\Phi}[\overline{V}_{n}](A_{n})\right)=
H\left(\widetilde{\Phi}[\overline{V}_{0}](A_{0})\right)<+\infty$.}
\end{enumerate}
\emph{These properties hold if
$$
\lim_{n\rightarrow+\infty}H\left(\left\{\mathrm{Tr}V^{n}_{i}A_{n}(V^{n}_{i})^{*}\right\}_{i=1}^{+\infty}\right)=
H\left(\left\{\mathrm{Tr}V^{0}_{i}A_{0}(V^{0}_{i})^{*}\right\}_{i=1}^{+\infty}\right)<+\infty.
$$}
\end{corollary}
\textbf{Proof.} By Lemma \ref{phi-map-cont} the main assertion
directly follows from Proposition \ref{g-basic-cont-cond+}. The
second assertion is proved by using Proposition \ref{cont-cond-a-2}.
$\square$
\section{On continuity of the output entropy of\\ quantum
operations on a given set of states}
\subsection{The case of a single operation}
In analysis of quantum channels and operations the question of
continuity of their output entropy on a \emph{given} set of input
states naturally arises (see Section 8). By summarizing the results
of the previous sections we consider below the possibilities to
prove continuity of the output entropy of a quantum operation
$\Phi(\cdot)=\sum_{i=1}^{+\infty}V_{i}(\cdot)V^{*}_{i}$ on a given
set of states $\mathcal{A}$ in the two cases (distinguished by
accessible information about properties of this set).\vspace{5pt}
A) \emph{If the set $\mathcal{A}$ is arbitrary then the function
$\,\rho\mapsto H_{\Phi}(\rho)\,$ is continuous on this set provided
one of the following conditions is valid:}
\begin{enumerate}[1)]
\item \emph{the quantum entropy is continuous on the set
$\,\Phi(\mathcal{A})$;} \footnote{This condition is obviously sufficient but it is not necessary (see Remark \ref{on-cont}).}
\item \emph{the operation $\,\Phi$ has continuous output entropy (sufficient conditions are presented
in Section 3);}
\item
\emph{$\lambda^{*}\left(\sqrt{\Phi^{*}(I_{\mathcal{H}'})}\right)<+\infty$ and the function $\rho\mapsto
H_{\widetilde{\Phi}}(\rho)$ is continuous on the set $\mathcal{A}$ (the last property can be verified by using Propositions \ref{CE-conditions-c} and \ref{comp-oper-c-c}).}
\end{enumerate}
B) \emph{If the von Neumann entropy is continuous on the set
$\mathcal{A}$ then the function $\,\rho\mapsto H_{\Phi}(\rho)\,$ is
continuous on this set provided one of the following conditions is
valid:}
\begin{enumerate}[1)]
\item \emph{$\Phi$ is a PCE-operation (sufficient conditions are presented in Section 4);}
\item \emph{the function $\rho\mapsto
H_{\widetilde{\Phi}}(\rho)$ is continuous on the set $\mathcal{A}$ (Proposition \ref{comp-oper-c-c} below).}
\end{enumerate}
Conditions A-3 and B-2 follow respectively from Corollary
\ref{basic-cont-cond+c} and Proposition \ref{basic-cont-cond+}. To
verify continuity of the function $\rho\mapsto
H_{\widetilde{\Phi}}(\rho)$ in these conditions one can use either
Proposition \ref{CE-conditions-c} or the following
proposition.\vspace{5pt}
\begin{property}\label{comp-oper-c-c}\ \textit{Let
$\;\Phi(\cdot)=\sum_{i=1}^{+\infty}V_{i}(\cdot)V^{*}_{i}$ be a
quantum operation and $\mathcal{A}$ be a subset of
$\,\mathfrak{S}(\mathcal{H})$. The function $\rho\mapsto
H_{\widetilde{\Phi}}(\rho)$ is continuous on the set $\mathcal{A}$
if one of the following conditions \textup{(}related by $\,
\mathrm{b)}\Rightarrow \mathrm{a)}$\textup{)} is valid:}
\begin{enumerate}[a)]
\item \textit{the function $\rho\mapsto H\left(\left\{\mathrm{Tr}V_{i}\rho
V_{i}^{*}\right\}^{+\infty}_{i=1}\right)$ is continuous on the set
$\mathcal{A}$;}
\item \textit{there exists a sequence $\{h_{i}\}_{i=1}^{+\infty}$ of nonnegative numbers
such that
$$
\sup_{\rho\in\mathcal{A}}\mathrm{Tr}\sum_{i=1}^{+\infty}h_{i}V^{*}_{i}V_{i}\rho<+\infty\quad\textit{and}\quad
\sum_{i=1}^{+\infty}e^{-\lambda h_{i}}<+\infty\;\;
\textit{for}\;\;\textit{all}\;\;\lambda>0.
$$}
\end{enumerate}
\end{property
\textbf{Proof.} Sufficiency of condition $\mathrm{a)}$ follows from
the last assertion of Corollary \ref{g-basic-cont-cond+c}. The
implication $\mathrm{b)}\Rightarrow \mathrm{a)}$ is proved by using
Proposition \ref{H-cont-cond}B. $\square$ \vspace{5pt}
\begin{example}\label{comp-oper-c-c+}
Let
$\mathcal{A}=\mathcal{K}^{s}_{H,h}\doteq\{\rho\in\mathfrak{S}(\mathcal{H})\,|\,\mathrm{Tr}H\rho\leq
h\}$ be the set of quantum states with "bounded mean energy" defined
by a $\mathfrak{H}$\nobreakdash-\hspace{0pt}operator $H$ with
$\mathrm{g}(H)=\inf\{\lambda>0\,|\,\mathrm{Tr}e^{-\lambda
H}<+\infty\}=0$ and positive $h$ (exceeding the minimal eigenvalue
of $H$). The von Neumann entropy is continuous on the set
$\mathcal{K}^{s}_{H,h}$ by Proposition \ref{H-cont-cond}B. The above
condition B\nobreakdash-\hspace{0pt}2 and condition $\mathrm{b)}$ in
Proposition \ref{comp-oper-c-c} show that the sufficient condition
of continuity of the function $\rho\mapsto H_{\Phi}(\rho)$ on the
set $\mathcal{K}^{s}_{H,h}$ consists in existence of a sequence
$\{h_{i}\}_{i=1}^{+\infty}$ of nonnegative numbers such that
$
\sum_{i=1}^{+\infty}h_{i}V_{i}^{*}V_{i}\leq H\quad\textrm{and}\quad
\sum_{i=1}^{+\infty}e^{-\lambda
h_{i}}<+\infty\;\,\textrm{for}\;\,\textrm{all}\;\, \lambda>0.
$
It is possible to show\footnote{This can be done by using the
following observation: \emph{for an arbitrary closed convex set
$\mathfrak{P}_{0}$ of probability distributions, on which the
Shannon entropy is continuous, there exists a sequence
$\{h_{i}\}_{i=1}^{+\infty}$ of nonnegative numbers such that}
$$
\sup_{\{\pi_{i}\}\in\mathfrak{P}_{0}}\sum_{i=1}^{+\infty}h_{i}\pi_{i}<+\infty\quad\textit{and}\quad
\sum_{i=1}^{+\infty}e^{-\lambda
h_{i}}<+\infty\;\,\textit{for}\;\,\textit{all}\;\, \lambda>0.
$$
This observation can be proved by using Lemma \ref{H-oper} and the
arguments from the proof of the implication
$\mathrm{(i)\Rightarrow(ii)}$ in Theorem \ref{bound-cont} based on
Dini's lemma.} that this condition is also necessary if the operator
$H$ is strictly positive and
$\mathrm{Ran}V_{i}\perp\mathrm{Ran}V_{j}$ for all sufficiently large
$i\neq j$.
\end{example}\vspace{5pt}
\subsection{The case of a converging sequence\\ of quantum
operations}
In analysis of continuity of information characteristics of a
quantum channel with respect to perturbations of this channel we
have to study continuity of the output entropy as a function of a
pair (channel, input state). Practically, the following problem
naturally arises in this analysis: for a given sequence
$\{\Phi_{n}\}$ of quantum operations in
$\,\mathfrak{F}_{\leq1}(\mathcal{H},\mathcal{H}')$ converging to an
operation $\Phi_{0}$ and for a given closed subset
$\mathcal{A}\subset\mathfrak{S}(\mathcal{H})$ to show that
\begin{equation}\label{task}
\lim_{n\rightarrow+\infty}H_{\Phi_{n}}(\rho_{n})=H_{\Phi_{0}}(\rho_{0})\quad
\forall \{\rho_{n}\}\subset\mathcal{A}\quad \textrm{such}\;
\textrm{that}\; \lim_{n\rightarrow+\infty}\rho_{n}=\rho_{0}.
\end{equation}
Summarizing the results of Section 6 we obtain the following
observation.\vspace{5pt}
\emph{If the von Neumann entropy is continuous on the set
$\mathcal{A}$ then property (\ref{task}) holds provided one of the
following conditions is valid:}
\begin{enumerate}[1)]
\item
\emph{$\lim_{n\rightarrow+\infty}H_{\Phi_{n}}(|\varphi_{n}\rangle\langle\varphi_{n}|)=H_{\Phi_{0}}(|\varphi_{0}\rangle\langle\varphi_{0}|)$
for any sequence $\{\varphi_{n}\}$ of vectors in $\,\mathcal{H}$
converging to a vector $\varphi_{0}$;}
\item \emph{property (\ref{task}) holds for the sequence $\{\widetilde{\Phi}_{n}\}$ of quantum operations
strongly converging to the operation $\widetilde{\Phi}_{0}$ such
that $\,(\Phi_{n},\widetilde{\Phi}_{n})$ is a complementary pair for
each $n\geq0$. }\end{enumerate} \vspace{5pt}
If
$\Phi_{n}(\cdot)=\sum_{i=1}^{+\infty}V^{n}_{i}(\cdot)(V^{n}_{i})^{*}$
for each $n\geq0$, where the sequence $\{V^{n}_{i}\}_{n}$ strongly
converges to the operator $V^{0}_{i}$ for each $i$, then the above
conditions $\mathrm{1)}$ and $\mathrm{2)}$ can be replaced
respectively by the following ones:
\begin{enumerate}[1)']
\item
$$
\lim_{n\rightarrow+\infty}
H\left(\left\{\|V^{n}_{i}|\varphi_{n}\rangle\|^{2}\right\}_{i=1}^{+\infty}\right)=
H\left(\left\{\|V^{0}_{i}|\varphi_{0}\rangle\|^{2}\right\}_{i=1}^{+\infty}\right)
$$
\textit{for any sequence $\{\varphi_{n}\}$ of vectors in
$\,\mathcal{H}$ converging to a vector $\,\varphi_{0}$;}
\item
$$
\lim_{n\rightarrow+\infty}H\left(\left\{\mathrm{Tr}V^{n}_{i}\rho_{n}(V^{n}_{i})^{*}\right\}_{i=1}^{+\infty}\right)=
H\left(\left\{\mathrm{Tr}V^{0}_{i}\rho_{0}(V^{0}_{i})^{*}\right\}_{i=1}^{+\infty}\right).
$$
\end{enumerate} \vspace{5pt}
The sufficient conditions for 1') can be found in the second part of
Corollary \ref{PCE-conditions+}. The sufficient condition for 2')
looks like condition b) in Proposition \ref{comp-oper-c-c} with
$V^{n}_{i}$ instead of $V_{i}$ and $"\sup_{n\geq0}"$ added to
$"\sup_{\rho\in\mathcal{A}}"$. \vspace{5pt}
\begin{example}\label{measurement} The above condition 2') makes it possible to replace the strong*
operator topology by the strong operator topology in the assertion
in example 3 in \cite{Sh-11}, concerning quantum measurements with a
countable number of outcomes and stating that \emph{continuity of
the Shannon entropy of the outcomes probability distribution with
respect to a priori state and to a measurement implies continuity of
the von Neumann entropy of the mean posteriori state with respect to
the same variables provided a priori state varies within a set on
which the von Neumann entropy is continuous.}
\end{example}
\section{Some applications}
\subsection{The Holevo capacity of quantum channels}
Let
$\Phi:\mathfrak{T}(\mathcal{H})\mapsto\mathfrak{T}(\mathcal{H}')$ be
a quantum channel and $\mathcal{A}$ be a subset of
$\mathfrak{S}(\mathcal{H})$. The Holevo capacity of the
$\mathcal{A}$-constrained channel $\Phi$ is defined as follows
(cf.\cite{H-c-w-c, H-Sh-2})
\begin{equation}\label{ccap-1}
\bar{C}(\Phi,\mathcal{A})=\sup_{\{\pi_{i},\rho_{i}\}\in
\mathcal{P}^{\mathrm{f}}_{\mathcal{A}}(\mathfrak{S}(\mathcal{H}))}\sum_{i}\pi_{i}H(\Phi(\rho_{i})\|\Phi(\textbf{b}(\{\pi_{i},\rho_{i}\})))
\end{equation}
(the supremum is over all finite ensembles of states with the
average in $\mathcal{A}$).
\subsubsection{On existence of continuous optimal ensembles}
The well known fact concerning the Holevo capacity of a finite
dimensional quantum channel $\Phi$ constrained by a closed subset
$\mathcal{A}$ consists in existence of an optimal ensemble at which
the supremum in (\ref{ccap-1}) is achieved \cite{Schum-West}. Since
\begin{equation}\label{ccap-2}
\bar{C}(\Phi,\mathcal{A})=\sup_{\mu\in
\mathcal{P}_{\mathcal{A}}(\mathfrak{S}(\mathcal{H}))}\int_{\mathfrak{S}(\mathcal{H})}H(\Phi
(\rho )\Vert \Phi (\textbf{b}(\mu)))\mu(d\rho )
\end{equation}
(the supremum is over all probability measures with the barycenter
in $\mathcal{A}$) the notion of an optimal ensemble is naturally
generalized to the infinite dimensional case leading to the notion
of an optimal measure (generalized or continuous optimal ensemble)
at which the supremum in (\ref{ccap-2}) is achieved \cite{H-Sh-2}.
In contrast to the finite dimensional case we can not claim
existence of an optimal measure for an arbitrary quantum channel
constrained by closed or even compact subset of states (see the
example in \cite{H-Sh-2}). By the theorem in \cite{H-Sh-2},
containing a sufficient condition for existence of an optimal
measure, we have the following result.\vspace{5pt}
\begin{property}\label{application-2}
\textit{Let
$\,\Phi:\mathfrak{T}(\mathcal{H})\rightarrow\mathfrak{T}(\mathcal{H}')$
be a quantum channel and let $\mathcal{A}$ be a compact subset of
$\,\mathfrak{S}(\mathcal{H})$. If one of the conditions of
continuity of the function $\mathcal{A}\ni\rho\mapsto
H_{\Phi}(\rho)$ presented in Section 7.1 holds then there exists an
optimal measure for the $\mathcal{A}$-constrained channel $\,\Phi$
supported by pure states.}
\end{property}
\vspace{5pt}
This proposition implies existence of an optimal measure for an
arbitrary quantum channel with finite dimensional environment (with
finite Kraus representation) constrained by a compact subset of
states on which the entropy is continuous.
\subsubsection{On continuity of the Holevo capacity\\ as a function of a
channel}
In the finite dimensional case the Holevo capacity
$\bar{C}(\Phi,\mathcal{A})$ is a continuous function of $\Phi$ on
the set of all quantum channels
$\mathfrak{F}_{=1}(\mathcal{H},\mathcal{H}')$, but it is not
continuous in infinite dimensions even with respect to the norm of
complete boundedness \cite{L&S}. By proposition 7 in \cite{Sh-H},
containing a sufficient condition for continuity of the function
$\,\Phi\mapsto\bar{C}(\Phi,\mathcal{A})\,$ on subsets of
$\,\mathfrak{F}_{=1}(\mathcal{H},\mathcal{H}')\,$ with respect to
the topology of strong convergence, we have the following result.
\vspace{5pt}
\begin{property}\label{application-1}
\textit{Let $\,\{\Phi_{n}\}$ be a sequence of quantum channels
strongly converging to a quantum channel $\,\Phi_{0}$ and let
$\mathcal{A}$ be a compact subset of $\,\mathfrak{S}(\mathcal{H})$
on which the von Neumann entropy is continuous. If one of the
conditions of validity of (\ref{task}) presented in Section 7.2
holds then
$$
\lim_{n\rightarrow+\infty}\bar{C}(\Phi_{n},\mathcal{A})=\bar{C}(\Phi_{0},\mathcal{A}).
$$}
\end{property}
\begin{corollary}\label{application-1+}
\textit{For arbitrary compact subset $\mathcal{A}$ of
$\,\mathfrak{S}(\mathcal{H})$, on which the von Neumann entropy is
continuous, the function $\Phi\mapsto\bar{C}(\Phi,\mathcal{A})$ is
continuous on the set
$\,\mathfrak{F}^{k}_{=1}(\mathcal{H},\mathcal{H}')$ for each
$k$.\footnote{$\mathfrak{F}^{k}_{=1}(\mathcal{H},\mathcal{H}')$ is
the set of all quantum channels from $\mathfrak{T}(\mathcal{H})$ to
$\mathfrak{T}(\mathcal{H}')$ having the Kraus representation
consisting of $\,\leq k$ summands.}}
\end{corollary}\vspace{5pt}
Note that the set of channels having finite Kraus representation is
dense in the set $\mathfrak{F}_{=1}(\mathcal{H},\mathcal{H}')$ of
all channels in the topology of strong convergence.
\subsection{On continuity of the Entanglement of Formation}
The notion of the Entanglement of Formation as a quantitative
characteristic of entanglement of a state in a composite quantum
system is introduced in \cite{B&Co} in the finite dimensional case.
The possible infinite dimensional generalizations of this notion are
based respectively on the $\sigma$-convex roof and the $\mu$-convex
roof constructions \cite{ESP,Sh-9}. Comparison of these
constructions, in particular, the sufficient conditions for their
coincidence on subsets of states of composite system are presented
in \cite{Sh-9}, where the arguments showing preferability of the
$\mu$-convex roof construction are also considered.\footnote{The
question of coincidence of the $\mu$-convex roof and the
$\sigma$-convex roof constructions of the EoF on the whole state
space is open (as far as I know). In \cite{Sh-9} it is shown that
this question can not be solved by using only such general
properties of the entropy as concavity and lower semicontinuity
since the $\sigma$-convex roof construction applied to a concave
lower semicontinuous and even bounded function instead of the
entropy may be different from the corresponding $\mu$-convex roof
construction (and may not satisfy the basic condition of
entanglement monotones).} So, in what follows we will use the
generalization of the EoF based on this construction, t.i.
$$
E_{F}(\omega)=(H\circ\Theta)_{*}^{\mu}(\omega)\doteq\inf
\int_{\mathrm{extr}\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})}H(\Theta(\varpi))\mu(d\varpi),
\quad\omega\in\mathfrak{S}(\mathcal{H}\otimes\mathcal{K}),
$$
where $\Theta(\cdot)=\mathrm{Tr}_{\mathcal{K}}(\cdot)$ and the
infimum is over all probability measures $\mu$ on the set
$\mathrm{extr}\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})$ with the
barycenter $\omega$.\vspace{5pt}
By proposition 8 in \cite{Sh-9} to show continuity of the function
$\omega\mapsto E_{F}(\omega)$ on a particular subset of
$\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})$ it is sufficient to
show continuity of one of the functions $\omega\mapsto
H(\mathrm{Tr}_{\mathcal{K}}\omega)$ and $\omega\mapsto
H(\mathrm{Tr}_{\mathcal{H}}\omega)$ on this subset. Thus by using
the results of the previous sections one can obtain continuity
conditions for the function $\omega\mapsto E_{F}(\omega)$.
\vspace{5pt}
\begin{property}\label{EoF-2}
\textit{Let $\,\{\omega_{n}\}$ be a sequence of states in
$\,\mathfrak{S}(\mathcal{H}\otimes\mathcal{K})$, converging to a
state $\omega_{0}$, such that
$\,\lim_{n\rightarrow+\infty}H(\rho_{n})=H(\rho_{0})<+\infty$, where
$\rho_{n}=\mathrm{Tr}_{\mathcal{K}}\omega_{n}$ for $n=0,1,2..$. Let
$\,\{\Phi_{n}\}$ and $\,\{\Psi_{n}\}$ be sequences of operations in
$\,\mathfrak{F}_{\leq 1}(\mathcal{H})$ and in $\,\mathfrak{F}_{\leq
1}(\mathcal{K})$ strongly converging to operations $\,\Phi_{0}$ and
$\,\Psi_{0}$ correspondingly. If one of the conditions of validity
of (\ref{task}) for the sequences $\{\rho_{n}\}$ and
$\,\{\Phi_{n}\}$ presented in Section 7.2 holds and
$\,\mathrm{Tr}\,\Phi_{0}\otimes\Psi_{0}(\omega_{0})>0$ then}
$$
\lim_{n\rightarrow+\infty}E_{F}\left(\frac{\Phi_{n}\otimes\Psi_{n}(\omega_{n})}{\mathrm{Tr}\,\Phi_{n}\otimes\Psi_{n}(\omega_{n})}\right)
=E_{F}\left(\frac{\Phi_{0}\otimes\Psi_{0}(\omega_{0})}{\mathrm{Tr}\,\Phi_{0}\otimes\Psi_{0}(\omega_{0})}\right).
$$
\end{property}
\vspace{5pt}
\textbf{Proof.} By the remark before the proposition it is
sufficient to show that
$$
\lim_{n\rightarrow+\infty}H(\mathrm{Tr}_{\mathcal{K}}\Phi_{n}\otimes\Psi_{n}(\omega_{n}))=
H(\mathrm{Tr}_{\mathcal{K}}\Phi_{0}\otimes\Psi_{0}(\omega_{0}))<+\infty.
$$
Since by the condition we have
$\;\lim_{n\rightarrow+\infty}H(\Phi_{n}(\rho_{n}))=H(\Phi_{0}(\rho_{0}))<+\infty\;$
and
$\mathrm{Tr}_{\mathcal{K}}\Phi_{n}\otimes\Psi_{n}(\omega_{n})\leq\Phi_{n}(\rho_{n})$
for $n=0,1,2,...$, the above relation follows from corollary 4 in
\cite{Sh-11}. $\square$\vspace{5pt}
The assertion of this proposition is valid in the following
cases:\begin{itemize}
\item $\Phi_{n}=\Phi_{0}$ is a PCE\nobreakdash-\hspace{0pt}operation (see Section 4);
\item $\{\Phi_{n}\}\subset\mathfrak{F}^{k}_{\leq
1}(\mathcal{H})$ (the set of quantum operations having the Kraus
representation consisting of $\,\leq k\,$ summands).
\end{itemize}
Proposition \ref{EoF-2} can be used in analysis of continuity of the
EoF under local operations.
\section{Appendix}
\subsection{Continuity of the entropy on subsets of $\mathfrak{T}_{+}(\mathcal{H}\otimes\mathcal{K})$}
Proposition 10 in \cite{Sh-4} can be generalized to subsets of
$\mathfrak{T}_{+}(\mathcal{H}\otimes\mathcal{K})$ as
follows.\vspace{5pt}
\begin{property}\label{cont-cond-a-1}
\textit{Let $\mathcal{C}$ be a subset of
$\,\mathfrak{T}_{+}(\mathcal{H}\otimes\mathcal{K})$. Continuity of
the quantum entropy on the sets
$\mathrm{Tr}_{\mathcal{K}}\mathcal{C}\subset\mathfrak{T}_{+}(\mathcal{H})$
and
$\mathrm{Tr}_{\mathcal{H}}\mathcal{C}\subset\mathfrak{T}_{+}(\mathcal{K})$
implies continuity of the quantum entropy on the set
$\,\mathcal{C}$.}
\end{property}\vspace{5pt}
\textbf{Proof.} Let $\{C_{n}\}\subseteq\mathcal{C}$ be a sequence
converging to an operator $C_{0}\in\mathcal{C}$. If $C_{0}\neq0$
then by the condition we have
$$
\lim_{n\rightarrow+\infty}H\left(\frac{C_{n}^{\mathcal{H}}}{\mathrm{Tr}C_{n}}\right)=
H\left(\frac{C_{0}^{\mathcal{H}}}{\mathrm{Tr}C_{0}}\right)\quad\textrm{and}\quad
\lim_{n\rightarrow+\infty}H\left(\frac{C_{n}^{\mathcal{K}}}{\mathrm{Tr}C_{n}}\right)=
H\left(\frac{C_{0}^{\mathcal{K}}}{\mathrm{Tr}C_{0}}\right)
$$
where $C_{n}^{\mathcal{H}}=\mathrm{Tr}_{\mathcal{K}}C_{n}$ and
$C_{n}^{\mathcal{K}}=\mathrm{Tr}_{\mathcal{H}}C_{n}$ for all $n$.
Proposition 10 in \cite{Sh-4} implies
$$
\lim_{n\rightarrow+\infty}H\left(\frac{C_{n}}{\mathrm{Tr}C_{n}}\right)=
H\left(\frac{C_{0}}{\mathrm{Tr}C_{0}}\right)\quad\textrm{and
hence}\quad \lim_{n\rightarrow+\infty}H\left(C_{n}\right)=
H\left(C_{0}\right).
$$
If $C_{0}=0$ then convergence to zero of the sequence
$\{H\left(C_{n}\right)\}$ follows from convergence to zero of the
sequences $\{H\left(C_{n}^{\mathcal{H}}\right)\}$ and
$\{H\left(C_{n}^{\mathcal{K}}\right)\}$ by means of the inequality
$H\left(C_{n}\right)\leq
H\left(C_{n}^{\mathcal{H}}\right)+H\left(C_{n}^{\mathcal{K}}\right)$,$\;\,n\in\mathbb{N}$.
$\square$
\subsection{Auxiliary results}
\begin{lemma}\label{nv-2}
\emph{Let $\,\{\rho_{n}\}$ be a sequence of states converging to a
state $\rho_{0}$ such that
$\,\lim_{n\rightarrow+\infty}H(\rho_{n})=H(\rho_{0})$. For given
natural $k$ let $P^{k}_{n}$ be a $k$\nobreakdash-\hspace{0pt}rank
spectral projector of the state $\rho_{n}$ corresponding to its $k$
maximal eigenvalues and let $A^{k}_{n}=P^{k}_{n}\rho_{n}$ for all
$\,n$.\footnote{The projector $P^{k}_{n}$ is uniquely defined if the
state $\rho_{n}$ has no multiple eigenvalues. In any case all
variants of $P^{k}_{n}$ lead to isomorphic operators $A^{k}_{n}$ and
we assume that one of them is chosen.} For arbitrary $\varepsilon>0$
there exist a natural $k_{\varepsilon}$ and a subsequence
$\{A^{k_{\varepsilon}}_{n_{t}}\}$ converging to the operator
$A^{k_{\varepsilon}}_{0}=\bar{P}^{k_{\varepsilon}}_{0}\rho_{0}$,
where $\bar{P}^{k_{\varepsilon}}_{0}$ is a particular
$k_{\varepsilon}$\nobreakdash-\hspace{0pt}rank spectral projector of
the state $\rho_{0}$ corresponding to its $k_{\varepsilon}$ maximal
eigenvalues, such that
$$
\mathrm{Tr}B^{k_{\varepsilon}}_{n_{t}}<\varepsilon\quad
\textit{and}\quad H(B^{k_{\varepsilon}}_{n_{t}})<\varepsilon,
$$
where
$B^{k_{\varepsilon}}_{n_{t}}=\rho_{n_{t}}-A^{k_{\varepsilon}}_{n_{t}}$
is a positive operator, for all sufficiently large $t$.}
\end{lemma}\vspace{5pt}
\textbf{Proof.} Despite possible multiple meaning of the operator
$P^{k}_{0}\rho_{0}$ the values $\mathrm{Tr}P^{k}_{0}\rho_{0}$ and
$H(P^{k}_{0}\rho_{0})$ are uniquely defined by the state $\rho_{0}$.
By lemma 4 in \cite{L-2} the sequence $\{H(P^{k}_{0}\rho_{0})\}_{k}$
is nondecreasing and tends to $H(\rho_{0})$ as
$k\rightarrow+\infty$. Let $k_{\varepsilon}$ be such that
$\mathrm{Tr}P^{k_{\varepsilon}}_{0}\rho_{0}>1-\varepsilon/2$ and
$H(\rho_{0})-H(P^{k_{\varepsilon}}_{0}\rho_{0})<\varepsilon/3$.
Since $A^{k_{\varepsilon}}_{n}\leq\rho_{n}$ for all $n$, the
compactness criterion for subsets of $\mathfrak{T}_{+}(\mathcal{H})$
(see the Appendix in \cite{Sh-11}) shows relative compactness of the
sequence $\{A^{k_{\varepsilon}}_{n}\}$ and hence existence of a
subsequence $\{A^{k_{\varepsilon}}_{n_{t}}\}$ converging to an
operator $A^{k_{\varepsilon}}_{0}$. By using coincidence of
$\mathrm{Tr}A^{k_{\varepsilon}}_{n}$ with
$\sup_{P}\mathrm{Tr}P\rho_{n}$, where $P$ runs over the set of all
$k_{\varepsilon}$\nobreakdash-\hspace{0pt}rank projectors, it is
easy to show that
$A^{k_{\varepsilon}}_{0}=\bar{P}^{k_{\varepsilon}}_{0}\rho_{0}$,
where $\bar{P}^{k_{\varepsilon}}_{0}$ is a particular
$k_{\varepsilon}$\nobreakdash-\hspace{0pt}rank spectral projector of
the state $\rho_{0}$ corresponding to its $k_{\varepsilon}$ maximal
eigenvalues. \vspace{5pt}
Since
$\,\lim_{t\rightarrow+\infty}\mathrm{Tr}A^{k_{\varepsilon}}_{n_{t}}=\mathrm{Tr}A^{k_{\varepsilon}}_{0}\,$
and
$\,\lim_{t\rightarrow+\infty}H(A^{k_{\varepsilon}}_{n_{t}})=H(A^{k_{\varepsilon}}_{0})\,$
we have
$$
\begin{array}{c}
\mathrm{Tr}B^{k_{\varepsilon}}_{n_{t}}=
1-\mathrm{Tr}A^{k_{\varepsilon}}_{n_{t}}\leq
|1-\mathrm{Tr}A^{k_{\varepsilon}}_{0}|+
|\mathrm{Tr}A^{k_{\varepsilon}}_{n_{t}}-\mathrm{Tr}A^{k_{\varepsilon}}_{0}|<\varepsilon/2+\varepsilon/2=\varepsilon
\end{array}
$$
and
$$
\begin{array}{c}
H(B^{k_{\varepsilon}}_{n_{t}})\leq H(\rho_{n_{t}})-
H(A^{k_{\varepsilon}}_{n_{t}})\leq
|H(\rho_{n_{t}})-H(\rho_{0})|\\\\+|H(\rho_{0})-H(A^{k_{\varepsilon}}_{0})|+
|H(A^{k_{\varepsilon}}_{n_{t}})-H(A^{k_{\varepsilon}}_{0})|<\varepsilon/3+\varepsilon/3+\varepsilon/3=\varepsilon
\end{array}
$$
for all sufficiently large $t$, where the second inequality is
obtained by using inequality (\ref{H-fun-ineq}).
$\square$\vspace{5pt}
\begin{lemma}\label{simple+}
\textit{Let $\{\pi_{i}\}_{i=1}^{+\infty}$ be a sequence of positive
numbers. Then
$$
\sup_{\{x_{i}\}\in\mathfrak{P}_{+\infty}}H\left(\left\{\pi_{i}x_{i}\right\}_{i=1}^{+\infty}\right)=\lambda^{*},
$$
where $\lambda^{*}$ is either the unique finite solution of the
equation $\sum_{i=1}^{+\infty}e^{-\lambda/\pi_{i}}=1$ if it exists
or equal to
$\,\mathrm{g}(\{\pi^{-1}_{i}\})=\inf\{\lambda>0\,|\,\sum_{i=1}^{+\infty}e^{-\lambda/\pi_{i}}<+\infty\}$
otherwise.}\footnote{It is assumed that $\inf\emptyset=+\infty$. The
equation $\sum_{i=1}^{+\infty}e^{-\lambda/\pi_{i}}=1$ has no
solution if and only if either
$\mathrm{g}(\{\pi^{-1}_{i}\})=+\infty$ or
$\sum_{i=1}^{+\infty}e^{-\mathrm{g}(\{\pi^{-1}_{i}\})/\pi_{i}}<1$.}
\end{lemma}\vspace{5pt}
\textbf{Proof.} By using the Lagrange method it is easy to show that
the function $\mathfrak{P}_{n}\ni\{x_{i}\}_{i=1}^{n}\mapsto
H\left(\left\{\pi_{i}x_{i}\right\}_{i=1}^{n}\right)$ attains its
maximum at the vector
$\left\{x^{*}_{i}=c\pi^{-1}_{i}e^{-\lambda^{*}_{n}/\pi_{i}}\right\}$,
where $\lambda^{*}_{n}$ is the solution of the equation
$\sum_{i=1}^{n}e^{-\lambda/\pi_{i}}=1$ and
$c=\left[\sum_{i=1}^{n}\pi^{-1}_{i}e^{-\lambda^{*}_{n}/\pi_{i}}\right]^{-1}$.
Hence \begin{equation}\label{double-ineq}
\max_{\{x_{i}\}\in\mathfrak{P}_{n}}H\left(\left\{\pi_{i}x_{i}\right\}_{i=1}^{n}\right)=\lambda^{*}_{n}.
\end{equation}
The assertion of the lemma can be derived from (\ref{double-ineq})
by noting that the sequence $\{\lambda^{*}_{n}\}$ tends to
$\lambda^{*}$ as $n\rightarrow+\infty$ and by using lower
semicontinuity of the classical entropy. $\square$\vspace{15pt}
I am grateful to A.S.Holevo and the participants of his seminar for
the useful discussion. I am also grateful to the organizers of the
workshop \emph{Thematic Program on Mathematics in Quantum
Information} at the Fields Institute, where the work on this paper
was initiated. This work is partially supported by the program
"Mathematical control theory" of Russian Academy of Sciences, by the
federal target program "Scientific and pedagogical staff of
innovative Russia" (program 1.2.1, contract P 938), by the
analytical departmental target program "Development of scientific
potential of the higher school 2009-2010" (project 2.1.1/500) and by
RFBR grant 09-01-00424-a.
|
\section*{References}
|
\section{Introduction}
Most of the investigations on reaction-diffusion models are
devoted to uniform models, where interaction rates are
site-independent. Among the simplest generalizations beyond a
completely uniform system is a lattice with alternating rates.
In \cite{GHMS}, relaxation in the kinetic Ising model on an
alternating isotopic chain has been discussed. In
\cite{SchSch02,SchSch03,MObZ}, the steady state configurational
probabilities of an Ising spin chain driven out of equilibrium by
a coupling to two heat baths has been investigated. An example is
a one-dimensional Ising model on a ring, in which the evolution is
according to a generalization of Glauber rates, such that spins at
even (odd) lattice sites experience a temperature $T_{\mathrm{e}}$
($T_{\mathrm{o}}$). In this model the detailed balance is
violated. The response function to an infinitesimal magnetic field
for the Ising-Glauber model with arbitrary exchange couplings has
been studied in \cite{Chatelain}. Other generalizations of the
Glauber model consist of, for example, alternating-isotopic chains
and alternating-bound chains (\cite{GO} for example). In a recent
article \cite{ihg}, we studied the expectation values of spins in
an Ising model with nonuniform coupling constants. A transfer
matrix method was used to study the steady state behavior of the
system in the thermodynamic limit. Different (static) phases of
this system were studied, and a closed form was obtained for this
transfer matrix.
In \cite{26} a ten-parameter family of one-species
reaction-diffusion processes with nearest-neighbor interaction was
introduced, for which the evolution equation of $n$-point
functions contains only $n$- or less- point functions, the so
called autonomous models. The average particle-number in each site
was obtained exactly for these models. In \cite{27,28}, this was
generalized to multi-species systems and more-than-two-site
interactions. In \cite{29,30,32}, the phase structure of some classes
of single or multiple-species reaction-diffusion systems was
investigated. These investigations were based on the one-point
functions of the systems.
In the present paper the most general nonuniform exclusion
nearest-neighbor reaction-diffusion models on a one-dimensional
lattice with boundaries are studied, for which the evolution
equations of the one-point functions are closed, and the transfer
matrix has a closed form. It is shown that there is a possible
phase transition in such models, which corresponds to a reduction of
the role of boundary conditions on time-independent profile of
the expectation value of the number operators. The scheme of the
paper is as follows. In section 2, the models are introduced, and
the evolution equation for the expectation values of $n_i$ (the
number operators at the site $i$) is obtained. Also conditions are
obtained so that the evolution of the expectation values of $n_i$ is
closed. In section 3, the equation governing the static solution for
the expectation values of $n_i$'s is obtained, and a transfer matrix
method is introduced to obtain the static solution and investigate
different (static) phases of the system. It is also seen that to
write a closed form for the transfer matrix, further conditions on
the reaction rates are to be satisfied. We call models satisfying
these conditions superautonomous models. In section 4, as an
example, a nonuniform voter model is investigated in more detail.
Section 5 is devoted to the concluding remarks.
\section{Exclusion nearest-neighbor reaction-diffusion models with nonuniform reaction rates}
Consider a one-dimensional lattice with $(L+1)$ sites, numbered
from $0$ to $L$. Each site is either empty (denoted by the vector $e_0$) or occupied with one
particle (denoted by the vector $e-1$). The evolution of the system is said to be governed by
nearest-neighbor interactions, if the evolution of each site depends on only that site and its
nearest neighbors (sites directly related to it through a link). The evolution of such a system is
governed by a Hamiltonian
${\mathcal H}$ of the form,
\begin{equation}\label{nu.01}
{\mathcal H} ={\mathcal H}'_0+\left(\sum_\alpha{\mathcal
H}_\alpha\right)+{\mathcal H}'_L,
\end{equation}
where ${\mathcal H}_\alpha$ corresponds to the link $\alpha$:
\begin{equation}\label{nu.02}
{\mathcal H}_\alpha=1^{\otimes(\alpha-\mu)}\otimes
H_\alpha\otimes 1^{\otimes(L-\alpha-\mu)},
\end{equation}
and
\begin{equation}\label{nu.03}
\mu:=\frac{1}{2}.
\end{equation}
The link $\alpha$ links the sites $(\alpha-\mu)$ and
$(\alpha+\mu)$, so that $\alpha\pm\mu$ are integers, and $\alpha$
runs from $\mu$ up to $(L-\mu)$. Throughout this paper, sites are
denoted by Latin letters which represent integers, while links are
denoted by Greek letters which represent integers plus one half
($\mu$), so that the link $\alpha$ joins the sites $(\alpha-\mu)$
and $(\alpha+\mu)$, while the site $i$ joins the links $(i-\mu)$
and $(i+\mu)$. $H_\alpha$ is a linear operator acting on a four
dimensional space (the configuration space corresponding to the
sites $(\alpha-\mu)$ and $(\alpha+\mu)$) with a basis
$\{e_{0\,0},e_{0\,1},e_{1\,0},e_{1\,1}\}$. Also,
\begin{align}\label{nu.04}
{\mathcal H}'_0&=
H'_0\otimes 1^{\otimes L}, \nonumber\\
{\mathcal H}'_L&=1^{\otimes L}\otimes H'_L,
\end{align}
where $H'0$ and $H'_L$ are linear operators acting on two
dimensional spaces (the configuration spaces corresponding to the
sites $0$ and $L$, respectively) with bases $\{e_0,e_1\}$.
The nondiagonal components of $H_\alpha$, $H'_0$, and $H'_L$ are reaction rates. Denoting a full
site by $\bullet$ and an empty site by $\circ$, the possible reactions for the boundary sites $0$ or $L$
are
\begin{align}\label{nu.05}
\circ\to\bullet,\quad\hbox{with the rate }&(H'_{0,L})^1{}_0,\nonumber\\
\bullet\to\circ,\quad\hbox{with the rate }&(H'_{0,L})^0{}_1,
\end{align}
while those for the link $\alpha$ are
\begin{align}\label{nu.06}
\circ\circ\to\circ\bullet,\quad\hbox{with the rate }&(H_\alpha)^{01}{}_{00},\nonumber\\
\circ\circ\to\bullet\circ,\quad\hbox{with the rate }&(H_\alpha)^{10}{}_{00},\nonumber\\
\circ\circ\to\bullet\bullet,\quad\hbox{with the rate }&(H_\alpha)^{11}{}_{00},\nonumber\\
\circ\bullet\to\circ\circ,\quad\hbox{with the rate }&(H_\alpha)^{00}{}_{01},\nonumber\\
\circ\bullet\to\bullet\circ,\quad\hbox{with the rate }&(H_\alpha)^{10}{}_{01},\nonumber\\
\circ\bullet\to\bullet\bullet,\quad\hbox{with the rate }&(H_\alpha)^{11}{}_{01},\nonumber\\
\bullet\circ\to\circ\circ,\quad\hbox{with the rate }&(H_\alpha)^{00}{}_{10},\nonumber\\
\bullet\circ\to\circ\bullet,\quad\hbox{with the rate }&(H_\alpha)^{01}{}_{10},\nonumber\\
\bullet\circ\to\bullet\bullet,\quad\hbox{with the rate }&(H_\alpha)^{11}{}_{10},\nonumber\\
\bullet\bullet\to\circ\circ,\quad\hbox{with the rate }&(H_\alpha)^{00}{}_{11},\nonumber\\
\bullet\bullet\to\circ\bullet,\quad\hbox{with the rate }&(H_\alpha)^{01}{}_{11},\nonumber\\
\bullet\bullet\to\bullet\circ,\quad\hbox{with the rate }&(H_\alpha)^{10}{}_{11}.
\end{align}
It is seen that these rates are in general different for different links (also the rates on the boundary sites are in general different). The system is called uniform if the rates are the same for all links,
and nonuniform if it is not the case.
The number operator in the site $i$ is denoted by $n_i$:
\begin{equation}\label{nu.07}
n_i=1^{\otimes i}\otimes n
\otimes 1^{\otimes(L-i)},
\end{equation}
where $n$ is an operator acting on a two dimensional space with the
basis $\{e_0,e_1\}$. The matrix form of $n$ in this basis is
\begin{equation}\label{nu.08}
n^a{}_b=\delta^a_1\,\delta^1_b.
\end{equation}
The evolution equation for the expectation value of an observable
$Q$ is
\begin{equation}\label{nu.09}
\frac{\mathrm{d}}{\mathrm{d} t}\langle Q\rangle=\langle Q\,{\mathcal H}\rangle,
\end{equation}
where
\begin{equation}\label{nu.10}
\langle Q\rangle=S\,Q\,\Psi,
\end{equation}
the vector $\Psi$ is the ($2^{L+1}$ dimensional) probability vector
describing the system and $S$ is the covector
\begin{equation}\label{nu.11}
S:=s^{\otimes(L+1)},
\end{equation}
and
\begin{equation}\label{nu.12}
s_a=1.
\end{equation}
The system is called autonomous, if the Hamiltonian is so that
the evolution of the expectation values of $n_i$ is closed in
terms of the expectation values of $n_j$'s. In the evolution
equation for the expectation value of $n_i$, the expectation
values of $n_{i-1}$, $n_i$, $n_{i+1}$, $(n_{i-1}\,n_i)$, and
$(n_i\,n_{i+1})$ occur. It is seen that the criterion that the
coefficients of the last two vanish, is
\begin{align}\label{nu.13}
s_a\,[H_{i-\mu}\,r\otimes\,r]^{a\,1}=&
0,\nonumber\\
s_a\,[H_{i+\mu}\,r\otimes\,r]^{1\,a}=& 0,
\end{align}
respectively, where
\begin{equation}\label{nu.14}
r^a=-\delta^a_0+\delta^a_1.
\end{equation}
Equation \Ref{nu.13} should hold for all $i$'s, in order that the system be
autonomous. So one can rewrite it like
\begin{align}\label{nu.15}
s_a\,[H_{\alpha}\,r\otimes\,r]^{a\,1}&=0,\nonumber\\
s_a\,[H_{\alpha}\,r\otimes\,r]^{1\,a}&= 0.
\end{align}
It is seen that this condition is the same as the corresponding
condition for uniform lattices \cite{26,30}, written for each link separately.
Provided that this condition holds, one arrives at
\begin{equation}\label{nu.16}
\!\frac{\mathrm{d}}{\mathrm{d} t}\langle n_i\rangle=\eta_{i-\mu}\,\langle
n_{i-1}\rangle+\theta_{i+\mu}\,\langle n_{i+1}\rangle
+(\kappa_{i-\mu}+\nu_{i+\mu})\,\langle n_i\rangle
+(\xi_{i-\mu}+\sigma_{i+\mu}), \!\!\!\quad 0<i<L,
\end{equation}
where
\begin{align}\label{nu.17}
\eta_\alpha&:=s_a\,(H_\alpha)^{a\,1}{}_{b\,0}\,r^b,\nonumber\\
\theta_\alpha&:=s_a\,(H_\alpha)^{1\,a}{}_{0\,b}\,r^b,\nonumber\\
\kappa_\alpha&:=s_a\,(H_\alpha)^{a\,1}{}_{0\,b}\,r^b,\nonumber\\
\nu_\alpha&:=s_a\,(H_\alpha)^{1\,a}{}_{b\,0}\,r^b,\nonumber\\
\xi_\alpha&:=s_a\,(H_\alpha)^{a\,1}{}_{0\,0},\nonumber\\
\sigma_\alpha&:=s_a\,(H_\alpha)^{1\,a}{}_{0\,0}.
\end{align}
For the boundary sites (the sites $0$ and $L$), one has
\begin{align}\label{nu.18}
\frac{\mathrm{d}}{\mathrm{d} t}\langle n_0\rangle&=\theta_\mu\,\langle
n_1\rangle +[(H'_0)^1{}_1-(H'_0)^0{}_1+\nu_\mu]\,\langle
n_0\rangle +[(H'_0)^0{}_1+\sigma_\mu],\\ \label{nu.19}
\frac{\mathrm{d}}{\mathrm{d} t}\langle n_L\rangle&=\eta_{L-\mu}\,\langle
n_{L-1}\rangle
+[\kappa_{L-\mu}+(H'_L)^1{}_1-(H'_L)^0{}_1]\,\langle n_L\rangle
+[\xi_{L-\mu}+(H'_L)^0{}_1].
\end{align}
\section{The static solution}
For the static solution ($\langle n\rangle_\mathrm{st}$), the left hand
side of \Ref{nu.16} vanishes and one obtains
\begin{equation}\label{nu.20}
\langle n_{i+1}\rangle_\mathrm{st}=
-\frac{\eta_{i-\mu}}{\theta_{i+\mu}}\,\langle n_{i-1}\rangle_\mathrm{st}
-\frac{\kappa_{i-\mu}+\nu_{i+\mu}}{\theta_{i+\mu}}\,\langle
n_i\rangle_\mathrm{st} -\frac{\xi_{i-\mu}+\sigma_{i+\mu}}{\theta_{i+\mu}}.
\end{equation}
Denoting that part of this solution which satisfies the
homogeneous equation by $\langle n\rangle_\mathrm{st}^\mathrm{hom}$, it is seen
that
\begin{equation}\label{nu.21}
\langle n_{i+1}\rangle_\mathrm{st}^\mathrm{hom}=
-\frac{\eta_{i-\mu}}{\theta_{i+\mu}}\,\langle
n_{i-1}\rangle_\mathrm{st}^\mathrm{hom}
-\frac{\kappa_{i-\mu}+\nu_{i+\mu}}{\theta_{i+\mu}}\,\langle
n_i\rangle_\mathrm{st}^\mathrm{hom},
\end{equation}
which can be written as the following matrix form
\begin{equation}\label{nu.22}
X_{i+\mu}^\mathrm{hom}=D_i\,X_{i-\mu}^\mathrm{hom},
\end{equation}
where
\begin{equation}\label{nu.23}
X_\alpha:=\begin{bmatrix}\langle n_{\alpha-\mu}\rangle_\mathrm{st}\\
\langle n_{\alpha+\mu}\rangle_\mathrm{st}\end{bmatrix},
\end{equation}
and
\begin{equation}\label{nu.24}
D_i:=\begin{bmatrix}0&1 \\ & \\
\displaystyle{-\frac{\eta_{i-\mu}}{\theta_{i+\mu}}}&
\displaystyle{-\frac{\kappa_{i-\mu}+\nu_{i+\mu}}{\theta_{i+\mu}}}\end{bmatrix}.
\end{equation}
Using these, one arrives at
\begin{equation}\label{nu.25}
X_\alpha^\mathrm{hom}=D_{\alpha\,\beta}\,X_\beta^\mathrm{hom},
\end{equation}
where
\begin{equation}\label{nu.26}
D_{\alpha\,\beta}:=D_{\alpha-\mu}\,D_{\alpha-\mu-1}\cdots
D_{\beta+\mu}.
\end{equation}
To solve \Ref{nu.20}, one can use a tranfer matrix (Green's
function) method. Consider the equation
\begin{equation}\label{nu.27}
G_{i+1\,j}= -\frac{\eta_{i-\mu}}{\theta_{i+\mu}}\,G_{i-1\, j}
-\frac{\kappa_{i-\mu}+\nu_{i+\mu}}{\theta_{i+\mu}}\,G_{i\,j}
-\delta_{i\,j}.
\end{equation}
Defining
\begin{equation}\label{nu.28}
Y_{\alpha\,j}:=\begin{bmatrix}G_{\alpha-\mu\,j}\\
G_{\alpha+\mu\,j}\end{bmatrix},
\end{equation}
it is seen that the solution for $Y$ is
\begin{equation}\label{nu.29}
Y_{\alpha\,j}=\begin{cases} D_{\alpha\,\beta}\,\tilde
Y_{\beta\,j},& \alpha<j\\
D_{\alpha\,\beta}\,Y_{\beta\,j},& \alpha>j
\end{cases},
\end{equation}
with the condition that \Ref{nu.27} holds for $i=j$. This
condition is
\begin{equation}\label{nu.30}
Y_{j+\mu\,j}=D_j\,Y_{j-\mu\,j}-\begin{bmatrix}0\\
1\end{bmatrix},
\end{equation}
which reads
\begin{equation}\label{nu.31}
D_{j+\mu\,\beta}\,(Y_{\beta\,j}-\tilde Y_{\beta\,j})=-\begin{bmatrix}0\\
1\end{bmatrix}.
\end{equation}
A particular solution for $Y$ is obtained if one sets $\tilde Y$
equal to zero. In this case,
\begin{equation}\label{nu.32}
Y_{\alpha\,j}=-\Theta_{\alpha\,j}\,D_{\alpha\,j+\mu}\,\begin{bmatrix}0\\
1\end{bmatrix},
\end{equation}
where $\Theta$ is the step function:
\begin{equation}\label{nu.33}
\Theta_{\alpha\,j}:=\begin{cases} 0,& \alpha<j\\
1,& \alpha>j
\end{cases}.
\end{equation}
Using this, the general solution to \Ref{nu.20} can be written as
\begin{equation}\label{nu.34}
X_\alpha=D_{\alpha\,\beta}\,X_\beta^\mathrm{hom}-
\sum_{j<\alpha}D_{\alpha\,j+\mu}
\frac{\xi_{j-\mu}+\sigma_{j+\mu}}{\theta_{j+\mu}}
\begin{bmatrix}0\\1\end{bmatrix}.
\end{equation}
As it was the case in \cite{ihg}, the steady state profile near
the end-site $0$ is determined by the eigenvalues of the matrix
$D_{\alpha\,\mu}$, where $\alpha$ is some site far from the ends.
One has
\begin{align}\label{nu.35}
X_\alpha&=X_\alpha^a\,\mathbf{f}_a,\nonumber\\
X_{\mu}&=X_{\mu}^a\,\mathbf{f}_a,
\end{align}
where $\mathbf{f}_a$ is the eigenvector of $D_{\alpha\,\mu}$
corresponding to the eigenvalue $\lambda^a$, and $X_\alpha^a$'s
and $X_{\mu}^a$'s are the coefficients of expansions of $X_\alpha$
and $X_{\mu}$ in terms of the eigenvectors. If there was no
nonhomogeneous part in the equation \Ref{nu.16}, then the
discussion would be exactly similar to that of \cite{ihg}:
$X_{\mu}^a$ vanishes if $\lambda^a$ tends to infinity (in the
thermodynamic limit). Otherwise, $X_{\mu}^a$ is generally nonzero
and determined by the boundary conditions. If the nonhomogeneous
part does not vanish, then the second term in \Ref{nu.34}
generally contains a large multiple of $\mathbf{f}_a$ if $\lambda^a$
tends to infinity. This large part is to be cancelled by a large
multiple of $\mathbf{f}_a$ coming from the first term in \Ref{nu.34}. So
in this case $X_{\mu}^a$ does not vanish but tends to a fixed
value independent of boundary conditions. It is seen that although
equations \Ref{nu.18} and \Ref{nu.19} serve as boundary conditions
to obtain say $X_\mu$, the above general argument is independent
of these conditions. The essence of the above argument is the following.
In general, $X_\alpha$ (the components of which are noting but
expectations of number operators) is a linear combination of two
vectors ($\mathbf{f}_a$'s) plus a nonhomogeneous part. The nonhomogeneous
part is determined from the bulk reactions, and the two unknown
coefficients $\mathbf{f}_a$'s are to be determined from the the boundary
conditions resulted from \Ref{nu.18} and \Ref{nu.19} (in the static
case that the left hand sides vanish). There are, however, regions
of the parameter space where in the thermodynamic limit one of the
coefficients of $\mathbf{f}_a$'s (or possibly both) are determined from
the bulk reactions only. This essentially means that the effect of
boundaries on the behavior of the system is reduced. In a transport
system, for example, it is expected that the time-independent profile
of the moving bodies' density depend on both the bulk reactions
(speed, overtaking, etc) and the boundary reactions (injection and
extraction rates). But there could be cases where these boundary
terms are unimportant, or less important.
The situation (in the thermodynamic limit) can be summarized as
follows.
\begin{itemize}
\item[\textbf{i}] The eigenvalue $\lambda^a$ tends to infinity.
In this case $X_{\mu}^a$ tends to a fixed value, independent of
the boundary conditions.
\item[\textbf{ii}] The eigenvalue $\lambda^a$ tends to zero or a
finite number. In this case $X_{\mu}^a$ is determined by the
boundary condiotions.
\end{itemize}
Obviously, similar cases occur at the other boundary site. It is seen
that this behavior at one of the boundaries is independent of the
analog behavior at the other boundary.
This is the static phase transition of the system, some discontinuous behavior of
the expectation value of the number operator near the boundaries (note that the
components of $X_\alpha$ are nothing but the expectations of number operators).
If one can write $D_i$ as
\begin{equation}\label{nu.36}
D_i:=\Sigma_{i+\mu}\,\Delta_i\,\Sigma^{-1}_{i-\mu},
\end{equation}
where $\Delta_i$ is diagonal, and $\Sigma_\alpha$ depends on only
the parameters corresponding to the link $\alpha$, then it is easy
to find the solution to \Ref{nu.22}. The system is called
superautonomous, if this is the case. Putting
\begin{equation}\label{nu.37}
\Sigma_\alpha=\begin{bmatrix}a_\alpha &b_\alpha \\
c_\alpha &d_\alpha\end{bmatrix},
\end{equation}
and
\begin{equation}\label{nu.38}
\Delta_i=\begin{bmatrix}
A_i& 0 \\ & \\
0 & B_i\end{bmatrix},
\end{equation}
in \Ref{nu.36}, one arrives at
\begin{align}\label{nu.39}
\frac{A_i}{\varsigma_{i-\mu}}\,a_{i+\mu}\,d_{i-\mu}
-\frac{B_i}{\varsigma_{i-\mu}}\,b_{i+\mu}\,c_{i-\mu}&=0,\\
\label{nu.40} -\frac{A_i}{\varsigma_{i-\mu}}\,a_{i+\mu}\,b_{i-\mu}
+\frac{B_i}{\varsigma_{i-\mu}}\,b_{i+\mu}\,a_{i-\mu}&=1,\\
\label{nu.41} \frac{A_i}{\varsigma_{i-\mu}}\,c_{i+\mu}\,d_{i-\mu}
-\frac{B_i}{\varsigma_{i-\mu}}\,d_{i+\mu}\,c_{i-\mu}&=
-\frac{\eta_{i-\mu}}{\theta_{i+\mu}},\\
\label{nu.42} -\frac{A_i}{\varsigma_{i-\mu}}\,c_{i+\mu}\,b_{i-\mu}
+\frac{B_i}{\varsigma_{i-\mu}}\,d_{i+\mu}\,a_{i-\mu}&=
-\frac{\kappa_{i-\mu}+\nu_{i+\mu}}{\theta_{i+\mu}},
\end{align}
where
\begin{equation}\label{nu.43}
\varsigma_\alpha:=a_\alpha\,d_\alpha-b_\alpha\,c_\alpha.
\end{equation}
Using \Ref{nu.39} and \Ref{nu.40}, one obtains $A_i$ and $B_i$:
\begin{align}\label{nu.44}
A_i&=\frac{c_{i-\mu}}{a_{i+\mu}},\\
\label{nu.45} B_i&=\frac{d_{i-\mu}}{b_{i+\mu}}.
\end{align}
Using these in \Ref{nu.41}, it is seen that
\begin{equation}\label{nu.46}
\frac{c_{i-\mu}\,d_{i-\mu}}{\varsigma_{i-\mu}}\,
\frac{\varsigma_{i+\mu}}{a_{i+\mu}\,d_{i+\mu}}=
\frac{\eta_{i-\mu}}{\theta_{i+\mu}},
\end{equation}
which can be solved as
\begin{align}\label{nu.47}
a_\alpha\,b_\alpha&=\phi\,\theta_\alpha\,\varsigma_\alpha,\\
\label{nu.48}
c_\alpha\,d_\alpha&=\phi\,\eta_\alpha\,\varsigma_\alpha,
\end{align}
where $\phi$ is a constant. Finally, using \Ref{nu.44},
\Ref{nu.45}, \Ref{nu.47}, and \Ref{nu.48} in \Ref{nu.42}, on obtains
\begin{equation}\label{nu.49}
\frac{a_{i-\mu}\,d_{i-\mu}}{\varsigma_{i-\mu}}
+\frac{b_{i+\mu}\,c_{i+\mu}}{\varsigma_{i+\mu}}=
-\phi\,\kappa_{i-\mu}-\phi\,\nu_{i+\mu},
\end{equation}
the solution to which is
\begin{align}\label{nu.50}
a_\alpha\,d_\alpha&=(-\phi\,\kappa_\alpha+\psi)\,\varsigma_\alpha,\\
\label{nu.51}
b_\alpha\,c_\alpha&=(-\phi\,\nu_\alpha-\psi)\,\varsigma_\alpha,
\end{align}
where $\psi$ is another constant. $\varsigma$ is not an
independent variable in $\Sigma$. Putting \Ref{nu.50} and
\Ref{nu.51} in \Ref{nu.43}, one arrives at
\begin{equation}\label{nu.52}
2\,\psi=\phi\,(\kappa_\alpha-\nu_\alpha)+1,
\end{equation}
which shows that $(\kappa-\nu)$ should be constant. One can use
\Ref{nu.52} in \Ref{nu.50} and \Ref{nu.51}, to obtain
\begin{align}\label{nu.53}
a_\alpha\,d_\alpha&=\frac{\varsigma_\alpha}{2}\,
[-\phi\,(\kappa_\alpha+\nu_\alpha)+1],\\
\label{nu.54} b_\alpha\,c_\alpha&=\frac{\varsigma_\alpha}{2}\,
[-\phi\,(\kappa_\alpha+\nu_\alpha)-1].
\end{align}
These two equations are not independent of \Ref{nu.47} and
\Ref{nu.48}. The consistency condition is
\begin{equation}\label{nu.55}
\phi^2(\kappa_\alpha+\nu_\alpha)^2-1=4\,\phi^2\,\eta_\alpha\,\theta_\alpha,
\end{equation}
showing that $[(\kappa+\nu)^2-4\,\eta\,\theta]$ should be
constant. Noting that $(\kappa-\nu)$ has to be a constant as well,
the second condition can be stated as $(\kappa\,\nu-\eta\,\theta)$
should be constant. So the conditions that the system be
superautonomous are
\begin{align}\label{nu.56}
\kappa_\alpha-\nu_\alpha&=\mbox{constant},\\ \label{nu.57}
\kappa_\alpha\,\nu_\alpha-\eta_\alpha\,\theta_\alpha&=\mbox{constant}.
\end{align}
There are some special cases resembling those encountered in
\cite{ihg}.
\begin{itemize}
\item[\textbf{1}] Constant coupling:
Here $H_\alpha$ does not depend on $\alpha$, and $\lambda^a$ tends
to infinity (zero) if and only if the corresponding eigenvalue of $D_i$ is
greater (smaller) than one.
\item[\textbf{2}] Periodic coupling:
\begin{equation}\label{nu.58}
H_{\alpha+m}=H_\alpha.
\end{equation}
In this case the behavior of the eigenvalues of $D_{\alpha,\mu}$
is determined by the eigenvalues of $D_{\alpha+m,\alpha}$: An
eigenvalue of $D_{\alpha,\mu}$ tends to infinity (zero) if and only if the
corresponding eigenvalue of $D_{\alpha+m,\alpha}$ is greater
(smaller) than one.
\item[\textbf{3}] Defects in the lattice:
No new phenomena is seen, as long as the defects are localized,
i.e. they are far from the boundaries. So if there is a lattice
that has some defects but otherwise is uniform, the static
behavior near the boundaries is similar to that of a uniform
lattice \cite{30}.
\item[\textbf{4}] A lattice with different behaviors at different end points:
The behaviors of the static solution near the two ends are
independent of each other, provided the behavior change occurs far
from the boundaries. So all the phenomena seen in previous special
cases can be seen at each boundary, independent of the other boundary.
\end{itemize}
\section{An example, the voter model}
The voter model is a lattice each site of which is either full
($\bullet$) or empty ($\circ$). The reactions on a link are
\begin{align}\label{nu.59}
\bullet\circ\to\circ\circ,\quad\hbox{with the rate }&u,\nonumber\\
\bullet\circ\to\bullet\bullet,\quad\hbox{with the rate }&v,\nonumber\\
\circ\bullet\to\bullet\bullet,\quad\hbox{with the rate }&u,\nonumber\\
\circ\bullet\to\circ\circ,\quad\hbox{with the rate }&v.
\end{align}
Of course, the rates $u$ and $v$ may be link-dependent. using
\Ref{nu.17},
\begin{align}\label{nu.60}
\eta_\alpha&=v_\alpha,\nonumber\\
\theta_\alpha&=u_\alpha,\nonumber\\
\kappa_\alpha&=-v_\alpha,\nonumber\\
\nu_\alpha&=-u_\alpha,\nonumber\\
\xi_\alpha&=0,\nonumber\\
\sigma_\alpha&=0.
\end{align}
In order that the system be superautonomous, \Ref{nu.56} and
\Ref{nu.57} must hold. \Ref{nu.56} reads
\begin{equation}\label{nu.61}
u_\alpha-v_\alpha=\mbox{constant},
\end{equation}
and \Ref{nu.57} is an identity. Assume that \Ref{nu.61} holds.
Using \Ref{nu.24}, one has
\begin{equation}\label{nu.62}
D_i:=\begin{bmatrix}0&1 \\ & \\
\displaystyle{-\frac{v_{i-\mu}}{u_{i+\mu}}}&
\displaystyle{1+\frac{v_{i-\mu}}{u_{i+\mu}}}\end{bmatrix}.
\end{equation}
From \Ref{nu.55}, one has
\begin{equation}\label{nu.63}
\phi^2\,(u_\alpha-v_\alpha)^2=1,
\end{equation}
one of the solutions to which is
\begin{equation}\label{nu.64}
\phi=\frac{1}{u_\alpha-v_\alpha}.
\end{equation}
Putting this in \Ref{nu.47}, \Ref{nu.48}, \Ref{nu.53}, and
\Ref{nu.54}, one arrives at
\begin{align}\label{nu.65}
a_\alpha\,b_\alpha&=\phi\,u_\alpha\,\varsigma_\alpha,\\
\label{nu.66} c_\alpha\,d_\alpha&=\phi\,v_\alpha\,\varsigma_\alpha,\\
\label{nu.67} a_\alpha\,d_\alpha&=\phi\,u_\alpha\,\varsigma_\alpha,\\
\label{nu.68}
b_\alpha\,c_\alpha&=\phi\,v_\alpha\,\varsigma_\alpha.
\end{align}
One set of solutions to these equations is
\begin{equation}\label{nu.69}
\Sigma_\alpha=\begin{bmatrix}u_\alpha &1 \\
v_\alpha &1\end{bmatrix}.
\end{equation}
Putting this in \Ref{nu.44} and \Ref{nu.45} results in
\begin{equation}\label{nu.70}
\Delta_i=\begin{bmatrix}
\displaystyle{\frac{v_{i-\mu}}{u_{i+\mu}}}& 0 \\ & \\
0 & 1\end{bmatrix}.
\end{equation}
One also has
\begin{align}\label{nu.71}
\varsigma_\alpha&=u_\alpha-v_\alpha,\nonumber\\
&=\phi^{-1}.
\end{align}
So
\begin{equation}\label{nu.72}
D_{\alpha\,\mu}=\begin{bmatrix}\displaystyle{\phi\,v_\mu\,
\left(\frac{v_{\alpha-1}\cdots v_{\mu+1}} {u_{\alpha-1}\cdots
u_{\mu+1}}-1\right)}& \displaystyle{1-\phi\,v_\mu\,
\left(\frac{v_{\alpha-1}\cdots v_{\mu+1}} {u_{\alpha-1}\cdots
u_{\mu+1}}-1\right)}\\
\displaystyle{\phi\,v_\mu\, \left(\frac{v_\alpha\cdots v_{\mu+1}}
{u_\alpha\cdots u_{\mu+1}}-1\right)}&
\displaystyle{1-\phi\,v_\mu\, \left(\frac{v_\alpha\cdots
v_{\mu+1}} {u_\alpha\cdots u_{\mu+1}}-1\right)}
\end{bmatrix}.
\end{equation}
One has
\begin{align}\label{nu.73}
\det(D_{\alpha\,\mu})&=\frac{v_\mu}{u_\alpha}\,
\left(\frac{v_{\alpha-1}\cdots v_{\mu+1}}{u_{\alpha-1}\cdots
u_{\mu+1}}\right),\\ \label{nu.74}
\mathrm{tr}(D_{\alpha\,\mu})&=1+\frac{v_\mu}{u_\alpha}\,
\left(\frac{v_{\alpha-1}\cdots v_{\mu+1}}{u_{\alpha-1}\cdots
u_{\mu+1}}\right),
\end{align}
showing that the eigenvalues of $D_{\alpha\,\mu}$ are
\begin{align}\label{nu.75}
\lambda^1&=1,\\ \label{nu.76} \lambda^2&=\frac{v_\mu}{u_\alpha}\,
\left(\frac{v_{\alpha-1}\cdots v_{\mu+1}}{u_{\alpha-1}\cdots
u_{\mu+1}}\right).
\end{align}
A special case is when
\begin{equation}\label{nu.77}
v_\alpha=u_\alpha.
\end{equation}
In this case the matrix $\Sigma_\alpha$ is singular. Yet one can
obtain the matrix $D_{\alpha\,\mu}$ as a limit $\phi\to\infty$ of
\Ref{nu.72}. The result is
\begin{equation}\label{nu.78}
D_{\alpha\,\mu}=\begin{bmatrix}\displaystyle{-v_\mu\,
\left(\frac{1}{v_{\alpha-1}}+\cdots+\frac{1}{v_{\mu+1}}\right)}&
\displaystyle{1+v_\mu\,
\left(\frac{1}{v_{\alpha-1}}+\cdots+\frac{1}{v_{\mu+1}}\right)}
\\
\displaystyle{-v_\mu\,
\left(\frac{1}{v_\alpha}+\cdots+\frac{1}{v_{\mu+1}}\right)}&
\displaystyle{1+v_\mu\,
\left(\frac{1}{v_\alpha}+\cdots+\frac{1}{v_{\mu+1}}\right)}
\end{bmatrix},
\end{equation}
and both eigenvalues of $D_{\alpha\,\mu}$ become one.
Let's study some special cases.
\begin{itemize}
\item[\textbf{1}] Constant coupling:
\begin{equation}\label{nu.79}
u_\alpha=\mathrm{constant}.
\end{equation}
In this case $\lambda^2$ tends to infinity if and only if $v_\alpha$ is
greater than $u_\alpha$.
\item[\textbf{2}] Periodic coupling:
\begin{equation}\label{nu.80}
u_{\alpha+m}=u_\alpha.
\end{equation}
In this case the behavior of the eigenvalues of $D_{\alpha,\mu}$
is determined by the eigenvalues of $D_{\alpha+m,\alpha}$, which
are one and
\begin{equation}\label{nu.81}
\Lambda^2=\frac{v_{\mu+m-1}\cdots v_{\mu}}{u_{\mu+m-1}\cdots
u_{\mu}}.
\end{equation}
If $u_\alpha$ is greater than $v_\alpha$, then $\lambda^2$ tends
to zero. If $u_\alpha$ is smaller than $v_\alpha$, then
$\lambda^2$ tends to infinity.
\item[\textbf{3}] closed lattice:
Let the $L$'th site be the same as the $0$'th site. One has
\begin{equation}\label{nu.82}
x_{L+\mu}=x_\mu.
\end{equation}
Combining this with \Ref{nu.34}, it is seen that $x_\mu$ should be
the eigenvector of $D_{L+\mu,\mu}$ corresponding to the eigenvalue
one, which shows that
\begin{equation}\label{nu.83}
x_\mu=\begin{bmatrix} 1\\1
\end{bmatrix},
\end{equation}
showing that the stationary profile of the density is uniform.
This is despite the fact that the reaction rates are not
necessarily uniform.
\end{itemize}
\section{Concluding remarks}
General autonomous exclusion models with nearest-neighbor
interactions on a one-dimensional lattice were studied, for them
the reaction rates were nonuniform. By autonomous is is meant that
the evolution equation for the expectation values of the number
operators are closed. It was seen that the condition that the
system be autonomous is the same as the analogous condition for
uniform lattices. A transfer matrix method was introduced to solve
the equation for the static configuration of the expectation
values of the number operators. The static phase picture of these
systems, including possible phase transitions, was investigated.
These phase transitions correspond to a reduction of the role of
boundary reactions on the profile of the expectations of the
number operators, similar to the case of uniform lattices.
Also similar to the case of uniform lattices, these possible
phase transitions are not affected by the boundary conditions.
Moreover, all of the above mentioned possible static phase transitions
near one boundary, are controlled by only the bulk reaction rates
in a large part of the lattice one boundary of which is the same
boundary. So phase transitions at different boundaries are
independent of each other, and finite defects far from boundaries
have no effect on the phase transitions.
It was seen that if the reaction rates satisfy ceratin
additional conditions (which are essentially the constancy of
ceratin combinations of reaction rates) a closed form can be
obtained for the transfer matrix. Systems satisfying those
conditions were called superautonomous. The example of the voter
model was studied in more detail.\\
\\
\textbf{Acknowledgement}: This work was partially supported by
the research council of the Alzahra University.
\newpage
|
\section{Introduction}
After Hawking's famous work \cite{Hawking:1974rv} - the black holes radiate - known as {\it{Hawking effect}}, it is now well understood that it is related to the event horizon of a black hole. A closely related effect is the {\it{Unruh effect}} \cite{Unruh:1976db}, where a similar type of horizon is experienced by a uniformly accelerated observer on the Minkowski space-time. A unified description of them was first put forwarded by Deser and Levin \cite{Deser:1997ri,Deser:1998xb} which was a sequel to an earlier attempt \cite{Narnhofer:1996zk}. This is called the global embedding Minkowskian space (GEMS) approach. In this approach, the relevant detector in curved space-time (namely Hawking detector) and its event horizon map to the Rindler detector in the corresponding flat higher dimensional embedding space \cite{Goenner,Rosen} and its event horizon.
Then identifying the acceleration of the Unruh detector, the Unruh temperature was calculated. Finally, use of the Tolman relation \cite{Tolman} yields the Hawking temperature. Subsequently, this unified approach to determine the Hawking temperature using the Unruh effect was applied for several black hole space-times \cite{Kim:2000ct,Tian:2005yj,Brynjolfsson:2008uc,Hong:2003xz}. However the results were confined to four dimensions and the calculations were done case by case, taking specific black hole metrics. It was not clear whether the technique was applicable to complicated examples like the Kerr-Newman metric which lacks spherical symmetry.
The motivation of this paper is to give a modified presentation of the GEMS approach that naturally admits generalization. Higher dimensional black holes with different metrics, including Kerr-Newman, are considered.
Using this new embedding, the local Hawking temperature (Unruh temperature) will be derived. Then the Tolman formula leads to the Hawking temperature.
We shall first introduce a new global embedding which embeds only the ($t-r$)-sector of the curved metric into a flat space. It will be shown that this embedding is enough to derive the Hawking result using the Deser-Levin approach \cite{Deser:1997ri,Deser:1998xb}, instead of the full embedding of the curved space-time. Hence we might as well call this the reduced global embedding. This is actually motivated from the fact that an $N$-dimensional black hole metric effectively reduces to a $2$ -dimensional metric (only the ($t-r$)-sector) near the event horizon by the dimensional reduction technique \cite{Robinson:2005pd,Carlip:1998wz,Iso:2006ut,Umetsu:2009ra} (for examples see Appendix 1). Furthermore, this $2$-dimensional metric is enough to find the Hawking quantities if the back scattering effect is ignored. Several spherically symmetric static metrics will be exemplified. Also, to show the utility of this reduced global embedding, we shall discuss the most general solution of the Einstein gravity - Kerr-Newman space-time, whose full global embedding is difficult to find. Since the reduced embedding involves just the two dimensional ($t-r$)-sector, black holes in arbitrary dimensions can be treated.
In this sense our approach is valid for any higher dimensional black hole.
The organization of the paper is as follows. In section 2 we shall find the reduced global embedding of several black hole space-times which are spherically symmetric. In the next section the power of this approach will be exploited to find the Unruh/Hawking temperature for the Kerr-Newman black hole. Finally, we shall give our concluding remarks.
\section{Reduced global embedding}
A unified picture of Hawking effect \cite{Hawking:1974rv} and Unruh effect \cite{Unruh:1976db} was established by the global embedding of a curved space-time into a higher dimensional flat space \cite{Deser:1998xb}.
Subsequently, this unified approach to determine the Hawking temperature using the Unruh effect was applied for several black hole space-times \cite{Kim:2000ct,Tian:2005yj}, but usually these are spherically symmetric. For instance, no discussion on the Kerr-Newman black hole has been given, because it is difficult to find the full global embedding.
Since the Hawking effect is governed solely by properties of the event horizon, it is enough to consider the near horizon theory. As already stated, this is a two dimensional theory obtained by dimensional reduction of the full theory. Its metric is just the ($t-r$)-sector of the original metric.
In the following sub-sections we shall find the global embedding of the near horizon effective $2$-dimensional theory. Then the usual local Hawking temperature will be calculated. Technicalities are considerably simplified and our method is general enough to include different black hole metrics.
\subsection{Schwarzschild metric}
Near the event horizon the physics is given by just the two dimensional ($t-r$) -sector of the full Schwarzschild metric \cite{Robinson:2005pd}:
\begin{eqnarray}
ds^2 = g_{tt}dt^2 + g_{rr}dr^2 = \Big(1-\frac{2m}{r}\Big)dt^2 -\frac{dr^2}{1-\frac{2m}{r}}.
\label{1.04}
\end{eqnarray}
It is interesting to see that this can be globally embedded in a flat $D=3$ space as,
\begin{eqnarray}
ds^2 = (dz^0)^2 - (dz^1)^2 - (dz^2)^2
\label{1.32}
\end{eqnarray}
by the following relations among the flat and curved coordinates:
\begin{eqnarray}
&&z^0_{out} = \kappa^{-1} \Big(1-\frac{2m}{r}\Big)^{1/2} \textrm{sinh}(\kappa t),\,\,\,\
z^1_{out} = \kappa^{-1} \Big(1-\frac{2m}{r}\Big)^{1/2} \textrm{cosh}(\kappa t),
\nonumber
\\
&&z^0_{in} = \kappa^{-1} \Big(\frac{2m}{r} - 1\Big)^{1/2} \textrm{cosh}(\kappa t),\,\,\,\
z^1_{in} = \kappa^{-1} \Big(\frac{2m}{r} - 1\Big)^{1/2} \textrm{sinh}(\kappa t),
\nonumber
\\
&&z^2 = \int dr \Big(1+\frac{r_Hr^2 + r_H^2r + r_H^3}{r^3}\Big)^{1/2},
\label{1.33}
\end{eqnarray}
where the surface gravity $\kappa=\frac{1}{4m}$ and the event horizon is located at $r_H=2m$. The suffix ``$in$'' (``$out$'') refer to the inside (outside) of the event horizon while variables without any suffix imply that these are valid on both sides of the horizon. We shall follow these notations throughout the paper. Now if a detector moves according to constant $r$ (Hawking detector) outside the horizon in the curved space, then the corresponding Unruh detector moves on the constant $z^2$ plane and it will follow the hyperbolic trajectory
\begin{eqnarray}
\Big(z^1_{out}\Big)^2 - \Big(z^0_{out}\Big)^2 = 16 m^2 \Big(1 - \frac{2m}{r}\Big) = \frac{1}{{\tilde{a}}^2}.
\label{hyper}
\end{eqnarray}
This shows that the Unruh detector is moving in the ($z^0_{out}, z^1_{out}$) flat plane with a uniform acceleration ${\tilde{a}}= \frac{1}{4m}\Big(1 - \frac{2m}{r}\Big)^{-1/2}$. Then, according to Unruh \cite{Unruh:1976db}, the accelerated detector will see a thermal spectrum in the Minkowski vacuum with the local Hawking temperature given by,
\begin{eqnarray}
T = \frac{\hbar {\tilde{a}}}{2\pi} = \frac{\hbar}{8\pi m} \Big(1 - \frac{2m}{r}\Big)^{-1/2}.
\label{localtemp}
\end{eqnarray}
So we see that with the help of the reduced global embedding the local Hawking temperature near the horizon can easily be obtained.
Now the temperature measured by any observer away from the horizon can be obtained by using the Tolman formula \cite{Tolman} which ensures constancy between the product of temperatures and corresponding Tolman factors measured at two different points in space-time. This formula is given by \cite{Tolman}:
\begin{eqnarray}
\sqrt{g_{tt}}~ T = \sqrt{g_{0_{tt}}}~ T_0
\label{new1}
\end{eqnarray}
where, in this case, the quantities on the left hand side are measured near the horizon whereas those on the right hand side are measured away from the horizon (say at $r_0$). Since away from the horizon the space-time is given by the full metric, $g_{0_{tt}}$ must correspond to the $dt^2$ coefficient of the full (four dimensional) metric.
For the case of Schwarzschild metric $g_{tt} = 1-2m/r$, $g_{0_{tt}} = 1-2m/r_0$. Now the Hawking effect is observed at infinity ($r_0 = \infty$), where $g_{0_{tt}} = 1$. Hence, use of the Tolman formula (\ref{new1}) immediately yields the Hawking temperature:
\begin{eqnarray}
T_0 = {\sqrt{g_{tt}}}~ T = \frac{\hbar}{8\pi m }.
\label{hawkingtemp}
\end{eqnarray}
Thus, use of the reduced embedding instead of the embedding of the full metric is sufficient to get the answer.
\subsection {Reissner-Nordstr$\ddot{\textrm{o}}$m metric}
In this case, the effective metric near the event horizon is given by \cite{Robinson:2005pd},
\begin{eqnarray}
ds^2 = \Big(1 - \frac{2m}{r} + \frac{e^2}{r^2}\Big)dt^2 - \frac{dr^2}{1 - \frac{2m}{r} + \frac{e^2}{r^2}}.
\label{1.34}
\end{eqnarray}
This metric can be globally embedded into the $D=4$ dimensional flat metric as,
\begin{eqnarray}
ds^2 = (dz^0)^2 - (dz^1)^2 - (dz^2)^2 + (dz^3)^2
\label{1.35}
\end{eqnarray}
where the coordinate transformations are:
\begin{eqnarray}
&&z^0_{out} = \kappa^{-1} \Big(1-\frac{2m}{r} + \frac{e^2}{r^2}\Big)^{1/2} \textrm{sinh}(\kappa t),\,\,\,\
z^1_{out} = \kappa^{-1} \Big(1-\frac{2m}{r} + \frac{e^2}{r^2}\Big)^{1/2} \textrm{cosh}(\kappa t),
\nonumber
\\
&&z^0_{in} = \kappa^{-1} \Big(\frac{2m}{r} - \frac{e^2}{r^2} - 1\Big)^{1/2} \textrm{cosh}(\kappa t),\,\,\,\
z^1_{in} = \kappa^{-1} \Big(\frac{2m}{r} - \frac{e^2}{r^2} - 1\Big)^{1/2} \textrm{sinh}(\kappa t),
\nonumber
\\
&&z^2 = \int dr \Big[1+\frac{r^2(r_+ + r_-) + r_+^2(r + r_+)}{r^2(r-r_-)}\Big]^{1/2},
\nonumber
\\
&&z^3 = \int dr \Big[\frac{4r_+^5r_-}{r^4(r_+ - r_-)^2}\Big]^{1/2}.
\label{1.36}
\end{eqnarray}
Here in this case the surface gravity $\kappa = \frac{r_+ - r_-}{2r_+^2}$ and $r_{\pm}=m\pm\sqrt{m^2-e^2}$. The black hole event horizon is given by $r_H=r_+$. Note that for $e=0$, the above transformations reduce to the Schwarzschild case (\ref{1.33}).
The Hawking detector moving in the curved space outside the horizon, following a constant $r$ trajectory, maps to the Unruh detector on the constant ($z^2,z^3$) surface. The trajectory of the Unruh detector is given by
\begin{eqnarray}
\Big(z^1_{out}\Big)^2 - \Big(z^0_{out}\Big)^2 = \Big(\frac{r_+ - r_-}{2r_+^2}\Big)^{-2} \Big(1-\frac{2m}{r} + \frac{e^2}{r^2}\Big)=\frac{1}{{\tilde{a}}^2}.
\label{RN1}
\end{eqnarray}
This, according to Unruh \cite{Unruh:1976db}, immediately leads to the local Hawking temperature $T=\frac{\hbar {\tilde{a}}}{2\pi}=\frac{\hbar(r_+ - r_-)}{4\pi r_+^2\sqrt{1-2m/r+e^2/r^2}}$ which was also obtained from the full global embedding \cite{Deser:1998xb}. Again, since in this case $g_{0_{tt}} = 1-2m/r_0 + e^2/r_0^2$ which reduces to unity at $r_0=\infty$ and $g_{tt} = 1-2m/r+e^2/r^2$, use of Tolman formula (\ref{new1}) leads to the standard Hawking temperature $T_0=\sqrt{g_{tt}}~ T=\frac{\hbar(r_+ - r_-)}{4\pi r_+^2}$.
\subsection{Schwarzschild-AdS metric}
Near the event horizon the relevant effective metric is \cite{Robinson:2005pd},
\begin{eqnarray}
ds^2 = \Big(1-\frac{2m}{r}+\frac{r^2}{R^2}\Big)dt^2 - \frac{dr^2}{\Big(1-\frac{2m}{r}+\frac{r^2}{R^2}\Big)},
\label{1.37}
\end{eqnarray}
where $R$ is related to the cosmological constant $\Lambda= -1/R^2$.
This metric can be globally embedded in the flat space (\ref{1.35}) with the following coordinate transformations:
\begin{eqnarray}
&&z^0_{out} = \kappa^{-1} \Big(1-\frac{2m}{r} + \frac{r^2}{R^2}\Big)^{1/2} \textrm{sinh}(\kappa t),\,\,\
z^1_{out} = \kappa^{-1} \Big(1-\frac{2m}{r} + \frac{r^2}{R^2}\Big)^{1/2} \textrm{cosh}(\kappa t),
\nonumber
\\
&&z^0_{in} = \kappa^{-1} \Big(\frac{2m}{r} - \frac{r^2}{R^2} - 1\Big)^{1/2} \textrm{cosh}(\kappa t),\,\,\,\
z^1_{in} = \kappa^{-1} \Big(\frac{2m}{r} - \frac{r^2}{R^2} - 1\Big)^{1/2} \textrm{sinh}(\kappa t),
\nonumber
\\
&&z^2 = \int dr \Big[1+\Big(\frac{R^3 + R r_H^2}{R^2 + 3r_H^2}\Big)^2 \frac{r^2r_H+rr_H^2 + r_H^3}{r^3(r^2+rr_H+r_H^2+R^2)}\Big]^{1/2},
\nonumber
\\
&&z^3 = \int dr \Big[\frac{(R^4 + 10R^2r_H^2+9r_H^4)(r^2+rr_H+r_H^2)}{(r^2+rr_H+r_H^2+R^2)(R^2+3r_H^2)^2}\Big]^{1/2}
\label{1.38}
\end{eqnarray}
where the surface gravity $\kappa=\frac{R^2+3r_H^2}{2r_HR^2}$ and the event horizon $r_H$ is given by the root of the equation $1-\frac{2m}{r_H} + \frac{r^2_H}{R^2}=0$.
Note that in the $R\rightarrow\infty$ limit these transformations reduce to those for the Schwarzschild case (\ref{1.33}). We observe that the Unruh detector on the ($z^2,z^3$) surface (i.e. the Hawking detector moving outside the event horizon on a constant $r$ surface) follows the hyperbolic trajectory:
\begin{eqnarray}
\Big(z^1_{out}\Big)^2 - \Big(z^0_{out}\Big)^2 = \Big(\frac{R^2+3r_H^2}{2r_HR^2}\Big)^{-2}\Big(1-\frac{2m}{r} + \frac{r^2}{R^2}\Big)=\frac{1}{{\tilde{a}}^2}
\label{ADS1}
\end{eqnarray}
leading to the local Hawking temperature $T=\frac{\hbar {\tilde{a}}}{2\pi}=\frac{\hbar\kappa}{2\pi\Big(1-\frac{2m}{r}+\frac{r^2}{R^2}\Big)^{1/2}}$. This result was obtained earlier \cite{Deser:1998xb}, but with more technical complexities, from the embedding of the full metric.
It may be pointed out that for the present case, the observer must be at a finite distance away from the event horizon, since the space-time is asymptotically AdS. Therefore, if the observer is far away from the horizon ($r_0>>r$) where $g_{0_{tt}}=1-2m/r_0+r_0^2/R^2$, then use of (\ref{new1}) immediately leads to the temperature measured at $r_0$:
\begin{eqnarray}
T_0 = \frac{\hbar\kappa}{2\pi\sqrt{1-2m/r_0 + r_0^2/R^2}}.
\label{new2}
\end{eqnarray}
Now, this shows that $T_0\rightarrow 0$ as $r_0\rightarrow \infty$; i.e. no Hawking particles are present far from horizon.
\section{Kerr-Newman metric}
So far we have discussed a unified picture of Unruh and Hawking effects using our reduced global embedding approach for spherically symmetric metrics, reproducing standard results. However, our approach was technically simpler since it involved the embedding of just the two dimensional near horizon metric. Now we shall explore the real power of this new embedding.
The utility of the reduced embedding approach comes to the fore for the Kerr-Newman black hole which is not spherically symmetric. The embedding for the full metric, as far as we are aware, is not done in the literature.
The effective $2$-dimensional metric near the event horizon is given by \cite{Iso:2006ut,Umetsu:2009ra},
\begin{eqnarray}
ds^2 = \frac{\Delta}{r^2+a^2}dt^2 - \frac{r^2+a^2}{\Delta}dr^2,
\label{Kerr1}
\end{eqnarray}
where
\begin{eqnarray}
&&\Delta = r^2-2mr+a^2+e^2 = (r-r_+)(r-r_-);\,\,\,\ a=\frac{J}{m};
\nonumber
\\
&&r_\pm = m\pm\sqrt{m^2-a^2-e^2}.
\label{Kerr2}
\end{eqnarray}
The event horizon is located at $r=r_+$. This metric can be embedded in the following $D=5$-dimensional flat space:
\begin{eqnarray}
ds^2=\Big(dz^0\Big)^2 -\Big(dz^1\Big)^2-\Big(dz^2\Big)^2 + \Big(dz^3\Big)^2 + \Big(dz^4\Big)^2,
\label{Kerr3}
\end{eqnarray}
where the coordinate transformations are
\begin{eqnarray}
&&z^0_{out} = \kappa^{-1} \Big(1-\frac{2mr}{r^2+a^2} + \frac{e^2}{r^2+a^2}\Big)^{1/2} \textrm{sinh}(\kappa t),\,\,\,\
z^1_{out} = \kappa^{-1} \Big(1-\frac{2mr}{r^2+a^2} + \frac{e^2}{r^2+a^2}\Big)^{1/2} \textrm{cosh}(\kappa t),
\nonumber
\\
&&z^0_{in} = \kappa^{-1} \Big(\frac{2mr}{r^2+a^2} - \frac{e^2}{r^2+a^2} - 1\Big)^{1/2} \textrm{cosh}(\kappa t),\,\,\,\
z^1_{in} = \kappa^{-1} \Big(\frac{2mr}{r^2+a^2} - \frac{e^2}{r^2+a^2} - 1\Big)^{1/2} \textrm{sinh}(\kappa t),
\nonumber
\\
&&z^2 = \int dr \Big[1+\frac{(r^2+a^2)(r_+ + r_-) + r_+^2(r + r_+)}{(r^2+a^2)(r-r_-)}\Big]^{1/2},
\nonumber
\\
&&z^3 = \int dr \Big[\frac{4r_+^5r_-}{(r^2+a^2)^2(r_+ - r_-)^2}\Big]^{1/2},
\nonumber
\\
&&z^4 = \int dr a\Big[\frac{r_+ + r_-}{(a^2+r_-^2)(r_- - r)} + \frac{4(a^2 + r_+^2)(a^2-r_+r_- + (r_+ +r_-)r)}{(r_+ - r_-)^2 (a^2 + r^2)^3}
\nonumber
\\
&&+ \frac{4r_+r_-(a^2+2r_+^2)}{(r_+ - r_-)^2(a^2+r^2)^2} + \frac{rr_- - a^2 + r_+(r+r_-)}{(a^2+r_-^2)(a^2+r^2)}\Big]^{1/2}.
\label{Kerr4}
\end{eqnarray}
Here the surface gravity $\kappa = \frac{r_+-r_-}{2(r_+^2 + a^2)}$.
For $e=0, a=0$, as expected, the above transformations reduce to the Schwarzschild case (\ref{1.33}) while only for $a=0$ these reduce to the Reissner-Nordstr$\ddot{\textrm{o}}$m case (\ref{1.36}).
As before, the trajectory adopted by the Unruh detector on the constant ($z^2,z^3,z^4$) surface corresponding to the Hawking detector on the constant $r$ surface is given by the hyperbolic form,
\begin{eqnarray}
\Big(z^1_{out}\Big)^2 - \Big(z^0_{out}\Big)^2 = \kappa^{-2}\Big(1-\frac{2mr}{r^2+a^2} + \frac{e^2}{r^2+a^2}\Big)=\frac{1}{{\tilde{a}}^2}.
\label{Kerr5}
\end{eqnarray}
Hence the local Hawking temperature is
\begin{eqnarray}
T=\frac{\hbar {\tilde{a}}}{2\pi}=\frac{\hbar\kappa}{2\pi\sqrt{\Big(1-\frac{2mr}{r^2+a^2} + \frac{e^2}{r^2+a^2}\Big)}}.
\label{Kerr6}
\end{eqnarray}
Finally, since $g_{tt} = 1-\frac{2mr}{r^2+a^2} + \frac{e^2}{r^2+a^2}$ (corresponding to the near horizon reduced two dimensional metric) and $g_{0_{tt}}=\frac{r_0^2-2mr_0+a^2+e^2-a^2 {\textrm{sin}}^2\theta}{r_0^2+a^2{\textrm{cos}}^2\theta}$ (corresponding to the full four dimensional metric), use of the Tolman relation (\ref{new1}) leads to the Hawking temperature
\begin{eqnarray}
T_0=\frac{\sqrt{g_{tt}}}{\sqrt{(g_0{_{tt}})_{r_0\rightarrow\infty}}}~T = \frac{\hbar\kappa}{2\pi} = \frac{\hbar(r_+-r_-)}{4\pi(r_+^2 + a^2)},
\label{Kerr7}
\end{eqnarray}
which is the well known result \cite{Iso:2006ut}.
\section{Conclusion}
We provide a new approach to the study of Hawking/Unruh effects including their unification, initiated in \cite{Deser:1997ri,Deser:1998xb,Narnhofer:1996zk}, popularly known as global embedding Minkowskian space-time (GEMS). Contrary to the usual formulation \cite{Deser:1997ri,Deser:1998xb,Narnhofer:1996zk,Kim:2000ct,Tian:2005yj,Brynjolfsson:2008uc}, the full embedding was avoided. Rather, we required the embedding of just the two dimensional ($t-r$)-sector of the theory. This was a consequence of the fact that the effective near horizon theory is basically two dimensional. Only near horizon theory is significant since Hawking/Unruh effects are governed solely by properties of the event horizon.
This two dimensional embedding ensued remarkable technical simplifications whereby the treatment of more general black holes (e.g. those lacking spherical symmetry like the Kerr-Newman) was feasible. Also, black holes in any dimensions were automatically considered since the embedding just required the ($t-r$)-sector.
|
\section{Introduction}
Be/X$-$ray systems consist of a neutron star in an eccentric orbit with a
massive Be star, that is a spectral B type star that displays strong Balmer
emission lines in its optical spectrum (Coe 2000). Be stars are also observed
to emit strong infrared radiation. The Balmer emission lines and the infrared
radiation are usually attributed to the emission from a circumstellar disk
of material around the Be star. The rotation rates of Be stars are generally high,
sometimes reaching up to $\sim$70\% of the break-up speed. The high rotation
rates of Be star are thought to take part in forming a circumstellar
disk (Porter \& Rivinius 2003). In Be/X$-$ray binaries, there is a positive
correlation between the pulse period of the accreting neutron star and the
orbital period of the binary system (Corbet 1986, Waters\& van Kerkwijk 1989).
Be/X$-$ray binaries often exhibit X$-$ray outburst episodes. These outbursts are
broadly classified as: {\it normal outbursts} with luminosities around
10$^{35}$$-$10$^{37}$ erg s$^{-1}$, lasting for a few days to weeks and
{\it giant outbursts} with luminosities $\gtrsim$ 10$^{38}$ erg s$^{-1}$ and
lasting for weeks to months (Stella 1986, Negueruela 1998). Normal outbursts
are generally coincident with the time of the periastron passage of the neutron
star.
Swift~J1626.6$-$5156~~was first detected on 2005 December 18 with the Swift Burst Alert
Telescope (BAT) and identified as a transient pulsar with $\sim$15 s pulsations
(Palmer et al 2005). The spin period of the neutron star was refined with
the subsequent Rossi X$-$ray Timing Explorer (RXTE) / Proportional Counter Array
(PCA) observations as 15.37682(5) (Markwardt et al. 2005). At the early phases
of the activation, Reig et al. (2008) observed X$-$ray flare, lasting about
450 s during which the pulsed fraction was seen to increase up to about 70\%.
Following the flare, the average count rate and the pulsed fraction went below
their pre-flare values for a few hundreds of seconds and indicated a recovery
within a few thousands seconds (Belloni et al. 2006). Optical spectroscopic
observations of the proposed companion (2MASS16263652-5156305,
USNO-B1.0 0380-0649488) revealed that the star exhibits strong H$\alpha$
emission, indicating that it is a Be star (Negueruela \& Marco, 2006).
As the infrared magnitudes of the companion is rather large for a Be star (J=13.5, H=13, K=12.6; Rea et al. 2006), i.e. the star is unusually faint in infrared band, the system is referred to as an
unusual Be/X$-$ray binary system.
Long term monitoring of Swift~J1626.6$-$5156~ showed strong light curve modulations at timescales
of about 45 days (Reig et al. 2008). This modulation was attributed to variations due to the orbital period of the binary or to a harmonic (Reig et al. 2008). Recently however, DeCesar et al. (2009) reported two periodicities in this system: 47 days and 72.5 days, on approximate 2:3 ratio.
Based upon their findings, they suggested an orbital period of 23 days. The
issue of the orbital period of Swift~J1626.6$-$5156~ was not resolved.
In this article, we report the discovery of the orbital period of the Be/X$-$ray
binary system containing Swift~J1626.6$-$5156~. In the next section, we describe the
observations. In \S 3, we present our timing analyses and results, and finally
in \S 4, we discuss the implications of our findings.
\section{Observations}
We analyzed data from Proportional Counter Array (PCA) onboard RXTE
(Jahoda et al 1996) of Swift~J1626.6$-$5156~ between MJD 53724 and MJD 54410 with a total
exposure of $\sim 314$ ksec, divided into 289 observations with exposures
between $\sim 1$ ksec and $\sim 2$ ksec.
The RXTE-PCA is an array of 5 Proportional Counter Units (PCU) operating
in the 2-60 keV energy range, with a total effective area of approximately
7000 cm$^2$ and a field of view of $\sim$1$^{\circ}$ FWHM. During the
observations of Swift~J1626.6$-$5156~ investigated here, the number of active
PCUs varied between 1 and 4.
Between MJD 53724 and MJD 53964, 1-2 ksec long observations were performed
every 2-3 days. After MJD 53964, observations were sampled in close pairs,
each of which consists of two consecutive 1-2 ksec long observations separated
by $\sim0.3-0.6$ days. Each of these observation pairs were apart from each other by $\sim9-10$ days.
\section{Data Analysis and Results}
We generated lightcurves in the 3-20 keV band with a time resolution of
0.375s using PCA GoodXenon data. Each lightcurve was background
subtracted using
the background lightcurve generated with the background estimator models
based upon the rate of very large events, spacecraft activation and
cosmic X$-$ray emission by mean of the standard PCA analysis tools (pcabackest
in HEADAS). Times of each lightcurve were converted to the barycenter of
the Solar system. In Figure 3 (top panel), we show the pointing averaged
net count rate as detected with PCU2.
In the pulse timing analysis, we have used the harmonic
representation of pulse profiles (Deeter $\&$ Boynton 1985).
In this method, the pulse profiles
were expressed in terms of harmonic series and
cross-correlated with the template pulse profile.
After the 240 days after the main outburst, observations were sampled
0.3-0.6 days apart from each other. These pair of observations
were sampled 9-10 days apart from each other. In order to avoid cycle count
ambiguity due to large gaps in the observation sample, we did not attempt to
connect cycle counts in phase. For these observations,
we estimated pulse frequencies using the data taken within one day.
\begin{figure}
\begin{center}
\includegraphics[height=5.5in,keepaspectratio=true,angle=-90]{f1.ps}
\end{center}
\small{Figure 1 -- {\bf{(top)}} Pulse frequency evolution of Swift~J1626.6$-$5156~.
{\bf{(bottom)}} The same as above after removing the binary
orbital motion from the time of arrival of photons.}
\end{figure}
In the upper panel of Figure 1, we present the long term pulse frequency
evolution of the source. In addition to a spin-up trend, a periodic
modulation due to orbital Doppler effects can be seen.
The Doppler shifted pulse frequency variations,
spin up rate and its change can be expressed as the first time derivative
of orbital phase model,
\begin{equation}
\nu = \nu_{0} + \dot \nu (t-t_{0}) + \frac{1}{2} \ddot \nu (t-t_{0})^{2}
+ \frac{2 \pi \nu_{0} }{P_{orbit}} x ( cos (l) + {\bf e}~cos (w) cos (2l) +
{\bf e}~sin (w) sin (2l) ),
\end{equation}
given by Deeter, Boynton, $\&$ Pravdo (1981). Here t$_{0}$ is the mid-time of
the observation; $\nu_{0}$ is the pulse frequency at t$_{0}$;
$\dot \nu$ and $\ddot \nu$ are the first and second
time derivative of the pulse frequency;
$x=a/c sin(i)$ is the light traveltime
for the projected semimajor axis
(where i is the inclination angle between the line of sight
and the orbital angular momentum vector);
$l=2\pi (t-T_{\pi/2})/P_{orbit}+\pi/2$ is the mean orbital longitude
at t; $T_{\pi/2}$ is the epoch when the mean orbital longitude is equal to
90 $^{\circ}$;
$P_{orbit}$ is the orbital period; e is the eccentricity; and w is the
longitude of periastron. The above expression is fitted to the pulse
pulse frequency time series between
MJD 53943.74 and MJD 54410.49. In Table 1, we list the parameters of the
orbital motion, and in Figure 1 (lower panel), we present the frequency evolution
after removal of the orbital effects.
Figure 2 shows the pulse frequency evolution (after
removal of the quadratic trend in pulse frequency) together with
the orbit model and their residuals, respectively.
We find that the source was spinning up during the observation with a rate of
$\dot \nu =(1.3062 \pm 0.0017 ) \times 10^{-12}$ Hz s$^{-1}$.
The periodic trend of the pulse frequencies yields
an eccentric orbit (e=$ 0.08 \pm 0.01 $) with an orbital period of
$132.89 \pm 0.03 $ days.
\begin{figure}
\begin{center}
\includegraphics[width=5.5in,keepaspectratio=true]{f2.ps}
\end{center}
\small{Figure 2 -- {\bf{(top)}} Pulse frequency time series after removal
of the quadratic trend in pulse frequency together with the orbit model and
{\bf{(bottom)}}, their residuals.}
\end{figure}
\begin{table}
\caption{Timing Solution of Swift~J1626.6$-$5156~ (Valid for MJD 53943.74-54410.49)}
\centering
\begin{tabular}{l|l}\hline\hline
Parameter & Value \\ \hline
Epoch for spin frequency (MJD) & 54178.24(5) \\
$\nu$ (Hz) & 0.065161110(3) \\
$\dot{\nu}$ ($10^{-12}$ Hz.s$^{-1}$) & 1.31(2) \\
$\ddot{\nu}$ ($10^{-21}$ Hz.s$^{-2}$) & $-7.5\pm 3.2$ \\
Orbital Period (days) & 132.89(3) \\
a/c sin i (lt-s) & 401(5) \\
Orbital Epoch at $\pi/2$ (MJD) & 54031.44(5) \\
Eccentricity & 0.08(1) \\
w (longitude of periastron) & 340(9) \\ \hline
\end{tabular}
\label{Table1 -- Parameters}
\end{table}
We also investigated any connection between source intensity and timing
properties of Swift~J1626.6$-$5156~. As an intensity indicator, we used count rates in
the 3$-$20 keV range using only PCU2 which had been operating in all observations
used here. Figure 3 shows the count rate history of the source as well as pulse
period evolution after removal of the effects of the orbital motion. We find
that the source exhibits the largest spin-up trend over the interval of the
fastest flux decline in the first $\sim$90 days from MJD 53724 to 83812. When
the source starts to show small scale X$-$ray flaring at around MJD 53814
(the vertical dashed lines in Figure 3), the spin-up trend slows down while the
underlying X$-$ray intensity continues to
decline. The source exhibits the largest X$-$ray flare starting at $\sim$ MJD
54912 (indicated by the dotted lines in Figure 3). The onset of this flare is
concident with a major spin-up trend change
from very slow to rather fast. The rapid spin up trend continues throughout
this major X$-$ray flare. After MJD 53965 (marked with the dot-dot-dot-dashed lines
in Figure 3), the spin-up trend remains constant while the source is ongoing
subsequent X$-$ray flares.
\begin{figure}
\begin{center}
\includegraphics[width=5.5in,keepaspectratio=true]{f3.ps}
\end{center}
\small{Figure 3 -- {\bf{(top)}} Count rate history of Swift~J1626.6$-$5156~ as determined in
the 3$-$20 keV range using PCU2 data only. {\bf{(bottom)}} Pulse period evolution
of the source after effects of the orbital motion are removed. Vertical lines are
explained in the text.}
\end{figure}
In order to estimate the timescales of the X$-$ray modulations, we applied a
Lomb-Scargle algorithm (Lomb 1975, Scargle 1982) for the analysis of
unevenly-spaced sample to our data.
We used a 200-day long window sliding along the light curve with a time step
of 7 days. For each stretch, we extracted the period with the highest power.
The analysis was restricted to the oscillating part of the light curve
(see bottom panel in Figure 4). The resulting periods, limited to detections
with a chance probability smaller than 0.1, are plotted in the top panel of
Figure 4. The size of the points is proportional to minus the logarithm of the
chance probability, ranging from around $10^{-6}$ around T=400 days (large points)
to $\sim 10^{-2}$ (small points, T=600 days). The horizontal lines correspond
to half (dashed) and a third (dotted) of the orbital period of Swift~J1626.6$-$5156~.
We find that the timescale of X$-$ray modulations significantly varies between
$\sim$45 days and 95 days. The timescale is found to remain nearly constant
in the interval of 400$-$600 days (see Fugure 4, bottom panel) as it was at
about 1/3 of the orbital period, and in the interval of 800$-$900 at about half
the period. There is an abrupt change in timescales of modulations around
T=610 days as it increased from 45 days ($P_{orb}$/3) to 95 days
($\sim$$P_{orb}$/3) in about 40 days. The rapid evolution of the timescale
of modulations can also be seen
in the bottom panel of Figure 4 (notice the separation of the two peaks around
T=600 and 650 days vs. the separation of the two around 650 and 750 days).
\begin{figure}
\begin{center}
\includegraphics[width=5.5in,keepaspectratio=true]{f4.ps}
\end{center}
\small{Figure 4 -- {\bf{(bottom)}} The modulating part of the outburst of Swift~J1626.6$-$5156~.
{\bf{(top)}} The highest-probability period from the sliding Lomb-Scargle analysis
(see text) as displayed at the middle of the corresponding 200-day long window.
Symbol size is proportional to minus the logarithm of the chance
probability. The horizontal lines mark $P_{orb}/2$ (dashed) and $P_{orb}/3$
(dotted).}
\end{figure}
\section{Discussion}
We report the discovery of the orbital period of the binary system harbouring
Swift~J1626.6$-$5156~ as 132.9 days. We find that the binary system has an orbital eccentricity
of 0.08.
High mass X$-$ray binaries fall into three separate groups when
the pulse periods are compared with the orbital periods (Corbet 1986).
The systems with Be companions show correlations between orbital period
and spin period (Corbet 1986, Waters\& van Kerkwijk 1989), while systems with
giant and supergiant companions fall into two separate regions.
Based on the pulse period and orbital period we discovered, Swift~J1626.6$-$5156~ falls into the
region of Be transients in the Corbet diagram.
The derived orbital parameters yield a mass function for the companion,
\begin{equation}
f(M)=\frac{4\pi ^{2}(a_{x} sin i)^{3}}{GP^{2}_{orb}}
=\frac{(M_{c} sin i )^{3}}{(M_{x}+M_{c})^{2}}
\sim 3.9 M_{\odot}.
\end{equation}
For a $\sim$1.4M$_{\odot}$ neutron star mass and orbital inclination angles of
45$^{\circ}$ and 60$^{\circ}$, the mass of the companion star is found to be
$\sim$13.5$M_{\odot}$ and $\sim$8.3$M_{\odot}$, respectively. In both cases,
the companion star is consistent with being a high mass system.
The optical counterpart of Swift~J1626.6$-$5156~ shows strong H$\alpha$
emission with an equivalent width of $\sim 40$ $\AA$
(Negueruela $\&$ Marco 2006). This equivalent width implies an
orbital period of the Be/X$-$ray system of the order of $\sim 100-200$ days
which is consistent with the orbital period reported in our work.
"A Be star" is an early type non-supergiant star which has a
circumstellar disk around its equator, possibly formed by fast spin
rotation, non-radial pulsation or magnetic loops (Slettebak 1988).
Be/X$-$ray binary pulsars often show recurrent X$-$ray outbursts. These
X$-$ray outbursts are thought to be mainly due to the
accretion of this circumstellar disk material onto a neutron star
(Negueruela 1998, Inam et al., 2004). The mass accretion increases at
the periastron passages since the density of plasma is greater when
the pulsar is close to the companion star. Therefore it is plausible
to observe an increase in X$-$ray flux during the periastron passage.
This behaviour has been observed in the Be/X$-$ray binaries 4U 0115+63
(Negueruela et al., 1998), KS 1947+300 (Galloway et al., 2004), 2S 1417-624
(Inam et al. 2004) and EXO 2030+375 (Reig 2008, Baykal et al 2008). In
case of Swift~J1626.6$-$5156~, however, we do not find X-ray outbursts coincident with the
periastron passages. This is likely because of the fact that the orbit
is nearly circular (as indicated by very small orbital eccentricity),
therefore, periastron passages do not provide any extra interaction
between the neutron star and circumstellar disk.
Although orbital period of Swift~J1626.6$-$5156~ can also be considered to be typical
among Be/X$-$ray pulsar binaries, eccentricity of the system is one of
the smallest among these binaries (Raguzova \& Popov, 2005). The small
eccentricity of the system can be as a result of a weak supernova kick
during formation of the system or a post-supernova circularisation by the
interaction of the neutron star with the decretion disk of the Be star
(Martin et al. 2009).
Our timing analysis of the long-term variability (Figure 4) has shown that
the situation is more complex than what was reported by DeCesar et al. (2009).
We find that modulations at timescales of about a half and a third
of the orbital period are indeed more prominent. For the Be/X-ray pulsar systems with high eccentricity, it is quite natural to observe periodic outbursts near the periastron passages (Okazaki \& Negueruela, 2001). For instance, modulating twice during an orbital revolution has been seen before in 4U~1907+09 (see e.g., in 't Zand, Baykal \& Strohmayer 1998). It could be explained with the Be star being tilted with respect to the orbital plane. This would cause the neutron star to cross its equatorial disk twice during one orbit, leading to periodic increase in accretion rate. For low eccentric systems such as GS 0834-430 and XTE J1543-568, outbursts are seldom and occur due to mass accumulation until a large perturbation develops (Okazaki \& Negueruela, 2001). The nature of Swift~J1626.6$-$5156~ is completely different compared to the outbursts of these low eccentric systems. It is more likely that the perturbations in the outer edge of the disk due to the non uniform accretion from the Be companion, propagate to the inner edge of the disk as density waves as suggested in EXO 2030+575 (Wilson et al. 2002, Baykal et al. 2008). These waves lead to small outbursts at the inner edge of the disk. The period of these waves is expected to be related to the resonance size of the accretion disk (Okazaki \& Negueruela, 2001). In eccentric systems, the trigger mechanism of these perturbations is related to the periastron passages, but for Swift~J1626.6$-$5156~ it should be an intrinsic property of the Be/X-ray pulsar system. There is no definite theory for the trigger mechanism in low eccentric Be/X-ray pulsar systems.
\acknowledgments
We would like to thank M. Ali Alpar for his valuable comments.
We acknowledge EU FP6 Transfer of Knowledge Project "Astrophysics
of Neutron Stars" (MTKD-CT-2006-042722). A.B and S.C.I acknowledge research project TBAG 109T748 of the Scientific and
Technological Research Council of Turkey (T\"{U}B\.{I}TAK). TB acknowledges support from ASI grant ASI I/088/06/0.
\\
\\
{\bf{References}}
{\noindent{Baykal, A., K{\i}z{\i}lo{\u{g}}lu, U., K{\i}z{\i}lo{\u{g}}lu, N., Beklen, E., Ozbey, M., 2008, A\&A, 479, 301}}
{\noindent{Coe, M. J. 2000, in ASP Conf. Ser. 214, Be Phenomenon in Early-Type Stars, ed. M. A. Smith \& H. F. Henrichs (San Francisco: ASP), 656}}
{\noindent{Belloni, T., Homan, J., Campana, S., Markwardt, C. B., Gehrels, N.
2006, ATel 687}}
{\noindent{Corbet, R.H.D., 1986, MNRAS, 220, 1047}}
{\noindent{DeCesar, M.E., Pottschmidt, K., Wilms, J. 2009, ATel 2036}}
{\noindent{Deeter, J. E., Boynton, P. E., \& Pravdo, S. H. 1981, ApJ, 247, 1003}}
{\noindent{Deeter, J. E., \& Boynton, P. E. 1985, in Proc. Inuyama Workshop on
Timing Studies of X$-$ray Sources, ed. S. Hayakawa \& F. Nagase (Nagoya: Nagoya
Univ.), 29}}
{\noindent{Galloway, D.K., Morgan, E.H., Levine, A.M., 2004, ApJ, 613, 1164}}
{\noindent{\.{I}nam S. \c{C}., Baykal A., Scott D. M., Finger M., Swank J., 2004, MNRAS, 349, 173}}
{\noindent{in't Zand J. J. M., Baykal A., Strohmayer T. E., 1998, ApJ, 496, 386}}
{\noindent{Jahoda, K., Swank J. H., Giles A. B., Stark, M. J., Strohmayer, T.,
Zhang, W., \& Morgan, E. H. 1996, Proc. SPIE, 2808, 59}}
{\noindent{Lomb, N.R. 1975, Ap\&SS, 39, 447}}
{\noindent{Markwardt, C. B. \& Swank, J. H. 2005, ATel 679}}
{\noindent{Martin, R.G., Tout, C.A., Pringle, J.E. 2009, MNRAS, 397, 1563}}
{\noindent{Negueruela, I. 1998, A\&A, 338, 505}}
{\noindent{Negueruela I., Reig, P., Coe, M.J., Fabregat, J., 1998, A\&A, 336, 251}}
{\noindent{Negueruela, I., Marco, A. 2006, ATel 739}}
{\noindent{Okazaki, A., Negueruela, I., 2001, A\&A, 377,161}}
{\noindent{Palmer, D., Barthelmy, S., Cummings, J., et al. 2005, ATel 678}}
{\noindent{Porter, J. M., \& Rivinius, T. 2003, PASP, 115, 1153}}
{\noindent{Raguzova, N.V., Popov, S.B. 2005, A\&AT, 24, 151}}
{\noindent{Rea N., Testa V., Israel G.L., et al. 2006, ATel 713}}
{\noindent{Reig, P., 2008, A\&A, 489, 725}}
{\noindent{Reig, P., Belloni, T., Israel, G.L., Campana, S., Gehrels, N., Homan, J.
2008, A\&A, 485, 797}}
{\noindent{Scargle, J.D. 1982, ApJ, 263, 835}}
{\noindent{Slettebak A., 1988, PASP 100,770}}
{\noindent{Stella, L., White, N. E., \& Rosner, R. 1986, ApJ, 308, 669}}
{\noindent{Waters, L. B. F. M., \& van Kerkwijk, M. H. 1989, A\&A, 223, 196}}
{\noindent{Wilson, C.A., Finger, M. H., Coe, M.J., Laycock, S., Fabregat, J., 2002, ApJ, 570, 287}}
\end{document}
|
\section*{Acknowledgments}\end{small}}
\newcommand\altaffilmark[1]{$^{#1}$}
\newcommand\altaffiltext[1]{$^{#1}$}
\voffset=-0.6in
\title[The Origin of the Eccentric Disk and BH in M31]{The Nuclear Stellar Disk in Andromeda: A Fossil from the Era of Black Hole Growth}
\author[Hopkins and Quataert]{
\parbox[t]{\textwidth}{
Philip F. Hopkins\altaffilmark{1}\thanks{E-mail:<EMAIL>} \&
Eliot Quataert\altaffilmark{1}}
\vspace*{6pt} \\
\altaffiltext{1}{Department of Astronomy and Theoretical Astrophysics
Center, University of California
Berkeley, Berkeley, CA 94720} \\
}
\date{Submitted to MNRAS, February 2, 2009}
\begin{document}
\maketitle
\label{firstpage}
\begin{abstract}
The physics of angular momentum transport from galactic scales
($\sim 10-100$ pc) to much smaller radii is one of the oustanding
problems in our understanding of the formation and evolution of
super-massive black holes (BHs). Seemingly unrelated observations
have discovered that there is a lopsided stellar disk of unknown
origin orbiting the BH in M31, and possibly many other systems. We
show that these nominally independent puzzles are in fact closely
related. Multi-scale simulations of gas inflow from galactic to BH
scales show that when sufficient gas is driven towards a BH,
gravitational instabilities form a lopsided, eccentric disk that
propagates inwards from larger radii. The lopsided stellar disk
exerts a strong torque on the remaining gas, driving inflows that
fuel the growth of the BH and produce quasar-level luminosities. The
same disk can produce significant obscuration along many sightlines
and thus may be the putative ``torus'' invoked to explain obscured
active galactic nuclei and the cosmic X-ray background. The stellar
relic of this disk is long lived and retains the eccentric
pattern. Simulations that yield quasar-level accretion rates produce
relic stellar disks with kinematics, eccentric patterns, precession
rates, and surface density profiles in reasonable agreement with
observations of M31. The observed properties of nuclear stellar
disks can thus be used to constrain the formation history of
super-massive BHs.
\end{abstract}
\begin{keywords}
galaxies: active --- quasars: general --- galaxies: evolution ---
cosmology: theory
\end{keywords}
\section{Introduction}
\label{sec:intro}
A massive black hole (BH) resides at the center of most massive
galaxies
\citep{KormendyRichstone95,Gebhardt00,merrittferrarese:msigma}. Such
BHs gain most of their mass as luminous quasars
\citep{soltan82,yutremaine:bhmf}. A long-standing problem in
understanding the origin of massive BHs is how gas loses angular
momentum and inflows from galactic scales all the way to the BH.
Moreover, the self-gravity of gas can cause it to locally collapse and
turn into stars; whatever process drives inflow must compete against
gas consumption via star formation. It is now well-established that
on relatively large scales within a galaxy, disturbances from
collisions with other galaxies and global self-gravitating
instabilities can bring the gas down to radii ($\sim10-100\,$pc) where
the direct gravitational force of the BH begins to dominate the
dynamics \citep{shlosman:bars.within.bars,
schweizer98,barnes:review}. However, the BH also efficiently
suppresses the disturbances from larger scales. How, then, does the
gas continue to flow in? On the smallest scales ($\ll 0.1\,$pc),
accretion can occur through angular momentum transport by local
magnetic stresses \citep{balbus.hawley.review.1998}. But this leaves a
critical gap of a factor of $\sim10^{2-3}$ in radius, in which gas is
still weakly self-gravitating and can form stars, but both
larger-scale torques and local magnetic stresses are inefficient;
models have traditionally had great difficulty in crossing this gap
\citep{shlosman:inefficient.viscosities,
goodman:qso.disk.selfgrav,thompson:rad.pressure}.
Independent observations of the properties of stars close to BHs in
nearby galaxies have, in some cases, discovered that many of the old
stars reside in an eccentric, lopsided stellar disk on spatial scales
from $\sim1-10\,$pc \citep{lauer:ngc4486b,lauer:centers,
houghton:ngc1399.nuclear.disk,
thatte:m83.double.nucleus,debattista:vcc128.binary.nucleus}. The
most well-known and well-studied case is in the neighbor to the Milky
Way, the Andromeda galaxy, M31
\citep{lauer93,tremaine:m31.nuclear.disk.model,bender:m31.nuclear.disk.obs}.
The dynamics of this disk have received considerable attention, but
its origin remains poorly understood
\citep{peiris:m31.nuclear.disk.models,
salow:nuclear.disk.models.2,bender:m31.nuclear.disk.obs}. Given the
demanding resolution requirements, it is also not clear whether such
features are peculiar or generic.
To understand the angular momentum transport required for massive BH
growth, we have recently carried out a series of numerical simulations
of inflow from galactic to BH scales
\citep{hopkins:zoom.sims}.\footnote{\label{foot:url}Movies of these
simulations are available at \url{http://www.cfa.harvard.edu/~phopkins/Site/Movies_zoom.html}} By re-simulating the
central regions of galaxies, gas flows can be followed from galactic
scales of $\sim100\,$kpc to much smaller radii, with an ultimate
spatial resolution $<0.1\,$pc. For sufficiently gas-rich disky
systems, gas inflow continues all the way to $\lesssim 0.1$ pc. Near
the radius of influence of the BH, the systems become unstable to the
formation of lopsided, eccentric gas+stellar disks. This eccentric
pattern is the dominant mechanism of angular momentum transport at
$\lesssim 10$ pc, and can lead to accretion rates as high as
$\sim10\,M_{\sun}\,{\rm yr^{-1}}$, sufficient to fuel the most luminous
quasars. In addition, through this process, some of the gas
continuously turns into stars and builds up a nuclear stellar disk.
In this {\em Letter}, we examine the possibility that the nuclear
stellar disks seen in M31 and other galaxies are ``fossils'' from the
era of BH growth. If correct, this provides a powerful new set of
constraints on the formation and evolution of supermassive BHs.
\vspace{-0.7cm}
\section{The Simulations}
\label{sec:sims}
\citet{hopkins:zoom.sims} give a detailed description of the
simulations used here; we briefly summarize some of their most
important properties. The simulations were performed with the
parallel TreeSPH code {\small GADGET-3} \citep{springel:gadget}.
The simulations include collisionless stellar disks and bulges, dark
matter halos, gas, and BHs. For this study, we are interested in
isolating the physics of gas inflow. As a result, we do not include
models for BH accretion and feedback -- the BH's mass is constant in
time and its only dynamical role is via its gravitational influence on
scales $\lesssim 10\,$pc.
Because of the very large dynamic range in both space and time needed
for the self-consistent simulation of galactic inflows and nuclear
disk formation, we use a ``zoom-in'' re-simulation approach. This
begins with a large suite of simulations of galaxy-galaxy mergers, and
isolated bar-(un)stable disks. These simulations have
$0.5\times10^{6}$ particles, corresponding to a spatial resolution of
$50\,$pc. These simulations have been described in a series of previous papers
\citep{dimatteo:msigma,robertson:msigma.evolution,
cox:kinematics,younger:minor.mergers,hopkins:disk.survival}.
From this suite of simulations, we select representative simulations
of gas-rich major mergers of Milky-Way mass galaxies (baryonic mass
$10^{11}\,M_{\sun}$), and their isolated but bar-unstable analogues, to
provide the basis for our re-simulations.
The dynamics on smaller scales does not depend critically on the
details of the larger-scale dynamics. Rather, the small-scale
dynamics depends primarily on global parameters of the system, such as
the total gas mass channeled to the center relative to the
pre-existing bulge mass.
Following gas down to the BH accretion disk requires much higher
spatial resolution than is present in the galaxy-scale simulations. We
begin by selecting snapshots from the galaxy-scale simulations at key
epochs.
In each, we isolate the central $0.1-1$\,kpc region, which contains
most of the gas that has been driven in from large scales. Typically
this is about $10^{10}\,M_{\sun}$ of gas, concentrated in a roughly
exponential profile with a scale length of $\sim0.3-0.5\,$kpc.
From this mass distribution, we then re-populate the gas in the
central regions at much higher resolution, and simulate the dynamics
for several local dynamical times. These simulations involve $10^{6}$
particles, with a resolution of a few pc and particle masses of
$\approx 10^{4}\,M_{\sun}$. We have run $\sim50$ such re-simulations,
corresponding to variations in the global system properties, the model
of star formation and feedback, and the exact time in the larger-scale
dynamics at which the re-simulation occurs.
\citet{hopkins:zoom.sims} present a number of tests of this
re-simulation approach and show that it is reasonably robust for this
problem. This is largely because, for gas-rich disky systems, the
central $\sim 300$ pc becomes strongly self-gravitating, generating
instabilities that dominate the subsequent dynamics.
These initial re-simulations capture the dynamics down to $\sim 10$
pc, still insufficient to quantitatively describe accretion onto a
central BH. We thus repeat our re-simulation process once more, using
the central $\sim10-30\,$pc of the first re-simulations to initialize
a new set of even smaller-scale simulations. These typically have
$\sim10^{6}$ particles,
a spatial resolution of $0.1\,$pc, and a particle mass
$\approx100\,M_{\sun}$. We carried out $\sim50$ such simulations to test
the robustness of our conclusions and survey the parameter space of
galaxy properties. These final re-simulations are evolved for
$\sim10^{7}$ years -- many dynamical times at $0.1$\,pc, but very
short relative to the dynamical times of the larger-scale parent
simulations.\footnote{We also carried out a few extremely
high-resolution intermediate-scale simulations, which include
$\sim5\times10^{7}$ particles and resolve structure from $\sim$ kpc
to $\sim0.3\,$pc -- these are slightly less high-resolution than the
net effect of our two zoom-ins, but they obviate the need for a
second zoom-in iteration. The conclusions from these higher
resolution simulations are identical.}
Our simulations include gas cooling and star formation, with gas
forming stars at a rate motivated by the observed \citet{kennicutt98}
relation. Specifically, we use a star formation rate per unit volume
$\dot \rho_{\ast} \propto \rho^{3/2}$ with the normalization chosen so
that a Milky-way like galaxy has a total star formation rate of about
$1\,M_{\sun} \, {\rm yr^{-1}}$.
Because we cannot resolve the detailed processes of supernovae
explosions, stellar winds, and radiative feedback, feedback from stars
is modeled with an effective equation of state
\citep{springel:multiphase}. In this model, feedback is assumed to
generate a non-thermal (turbulent, in reality) sound speed that
depends on the local star formation rate, and thus the gas density.
We use sub-grid sound speeds $\sim20-100\,{\rm km\,s^{-1}}$, motivated
by a variety of observations of dense, star forming regions both
locally and at high redshift \citep{downes.solomon:ulirgs,
bryant.scoville:ulirgs.co,forsterschreiber:z2.disk.turbulence,
iono:ngc6240.nuclear.gas.huge.turbulence}.
Within this range, we found little difference in the physics of
angular momentum transport or in the resulting accretion rates, gas
masses, etc. \citep{hopkins:zoom.sims} (see also \S \ref{sec:results}).
\vspace{-0.67cm}
\section{Results}
\label{sec:results}
\begin{figure}
\centering
\plotone{tile_zoom_nature.pdf}
\caption{Example of the multi-scale simulations used to follow gas
inflow from galactic to nuclear scales. In each panel, red
colors denote the projected stellar mass density, and green/blue
colors denote the projected gas density (with the variation from
green to blue reflecting an increasing star formation rate per
unit mass in the gas). Each image is rotated to project the gas
density face on, relative to its net angular momentum vector
{\em Top:} Galaxy-scale: the merger of two similar-mass
galaxies, just after the final coalescence of the two nuclei
(two $10^{11}\,M_{\sun}$ galaxies with a disk gas fraction $f_{\rm
gas}\sim0.4$ at the time of merger, and an initial bulge to
total mass ratio of $B/T = 0.2$; simulation b3ex(co) in
\citealt{hopkins:zoom.sims}). The merger has driven large
amounts of gas into the central $\sim1\,$kpc, forming the
nuclear starburst shown. {\em Middle:} Re-simulation of the
$\sim 0.1-1$ kpc region (simulation If9b5). The starburst disk,
being strongly self-gravitating, develops a spiral and bar mode
that drives gas to $\sim10\,$pc, where the bar is suppressed by
the gravity of the BH. {\em Bottom:} Re-simulation of the
central $\sim 30$\,pc (simulation Nf8h1c1qs). The inflow to
these scales rapidly forms a lopsided eccentric disk around the
BH (which maps onto a one-armed spiral at larger radii). The
disk drives accretion rates of $\sim 1-10\,M_{\sun}\,{\rm yr^{-1}}$
to $<0.1\,$pc, and leaves the eccentric stellar relics shown in
Figures \ref{fig:tile.10pc} \& \ref{fig:velocities}.
\label{fig:zoom}}
\end{figure}
Figure~\ref{fig:zoom} shows an example of the results of our
re-simulations. We follow the merger and coalescence of two Milky-way
mass galaxies, a highly asymmetric event chosen because it is likely
to lead to rapid BH growth. The first image (top left) provides a
large-scale view of the system just after the coalescence of the two
galactic nuclei and their central BHs. The highly asymmetric
disturbances visible in the image (e.g.\ tidal tails) efficiently
torque the gas and allow it to flow inwards. Inside the central
$\sim$kpc, the inflowing gas piles up at the point where its gravity
begins to dominate that of the stars and dark matter; this dense gas
generates a luminous burst of star formation. Precisely because the
gas and newly-formed stars are self-gravitating, they form secondary
gravitational instabilities such as bars and spiral waves that produce
further torques and inflow. This is essentially the ``bars within
bars'' mechanism proposed in \citet{shlosman:bars.within.bars} and it
occurs for the reasons outlined therein. But once the gas reaches
$\sim10\,$pc, this mechanism no longer works -- the gravity of the BH
begins to dominate ($M_{BH} = 3\times10^{7}\,M_{\sun}$ here) and the
system can no longer support the large-scale bars critical to the
inflow on larger scales.
At precisely these scales our simulations demonstrate that a new
instability generically arises -- a nearly static (slowly precessing)
lopsided or eccentric disk of gas and stars. Such slowly varying
eccentric ($m=1$) perturbations are unique to the gravitational field
of a point mass such as a BH
\citep{tremaine:slow.keplerian.modes}. But they are also linearly
stable \citep{tremaine:slow.keplerian.modes}, so how do they arise in
the simulations? We discuss this in detail in
\citet{hopkins:zoom.sims}; to summarize, we believe that the mode is a
{\em global} phenomena that grows from the outside in. It first
starts to grow because of self-gravity, where the mass of the (stellar
+ gas) disk is comparable to the mass of the BH. In the simulations,
this occurs at $\sim10-100\,$pc. Gas moving in circular orbits passes
through this eccentric disk, and experiences a torque from the stars
therein, causing the gas to lose angular momentum and fall inwards.
Some of the gas turns into stars, those stars are excited into the
$m=1$ mode, allowing the perturbation to efficiently propagate inwards
to $\sim0.1\,$pc. Although at radii where $M_{\rm disk}(<R) \ll
M_{\rm BH}$, the system may formally be stable, the eccentric pattern
is {\em induced} by the mass distribution at somewhat larger radii.
The lopsided or eccentric disk, then, is a coherent global pattern
superimposed on the otherwise axisymmetric gas and stellar mass
distribution. The eccentric pattern precesses with an angular pattern
speed $\Omega_{p}$, which we measure to be $\sim 1-5\,{\rm
km\,s^{-1}\,pc^{-1}}$ (independent of radius), set by the rotation
rate at the radii where the disk and BH masses are comparable.
\begin{figure}
\centering
\plotone{tile_10pc_nature.pdf}
\caption{The nuclear disk in several representative simulations.
Scale in all panels is the same (lower right). The different
rows show the results from simulations of the central
$\sim 50$ pc of galactic nuclei with
different galaxy properties (top to bottom: Nf8h1c0thin,
Nf8h1c1thin, Nf8h1c1qs, Nf8h1c1dens, Nf8h1c0 in
\citealt{hopkins:zoom.sims}, which have initial $f_{\rm
gas}\sim0.5-0.8$, $h/R=0.16,0.08,0.28,0.25,0.15$, $M_{\rm
BH}\sim3\times10^{7}\,M_{\sun}$, and initial disky mass $\sim
1.2,1.7,3.0,8.1,0.25\times10^{7}\,M_{\sun}$ inside $10\,$pc), and
different treatments of stellar feedback (sub-grid sound speeds
$c_{s}\sim35,\,20,\,40,\,50,\,10\,{\rm km\,s^{-1}}$ from top to
bottom). The formation of a lopsided disk is ubiquitous. {\em
Left:} Gas surface density. Colors encode the absolute star
formation rate of the gas (increasing from blue to red/yellow).
Regions where gas shocks (edges in this image) dissipate energy,
leading to rapid gas inflow. {\em Middle Left:} Same, viewed
edge-on in cylindrical ($R,\,z$) coordinates to emphasize the
disk thickness versus radius. The exact thickness depends on
our model for stellar feedback, but gravitationally driven
turbulence and heating results in the disks always being
somewhat flared and thick on these scales, even with negligible
stellar feedback ({\em bottom row}). {\em Middle Right:} As
{\em left}, but showing the stellar mass distribution. The lack
of shocks means that the edges of the disk are less sharp, but
they are still visually clear. {\em Right:} Edge-on stellar
density ($x,\,z$). Several of the systems appear to have
two nuclei, which was the initial indication of the eccentric
disk in M31 (the P1/P2 feature).
\label{fig:tile.10pc}}
\end{figure}
We have run a suite of over $\sim50$ simulations in order to study the
properties of, and robustness of, the eccentric disks in our
calculations. Figure~\ref{fig:tile.10pc} shows face-on and edge-on
views of the gas and stars in several of these simulations. Provided
that a significant amount of gas ($\gtrsim 10 \% \, M_{\rm BH}$) can
be driven in from larger radii by global torques in galaxies,
eccentric nuclear disks are generic.
\begin{figure}
\centering
\plotone{velfield_nature.pdf}
\caption{{\em Top:} Mean projected line-of-sight velocity $V$
({\em left}) and velocity dispersion $\sigma$ ({\em right}) of
the relic nuclear stellar disks in our simulations, long after
gas is exhausted. Each solid line is a mock line-of-sight
velocity field of the stars in one of our simulations, as if
they were at the distance and viewing angle of the M31 nucleus
($54\,\deg$ from edge on;
\citealt{peiris:m31.nuclear.disk.models}). Dotted lines vary
the inclination angle by $10\,\deg$. The velocities are
measured in narrow radial pixels along a slit placed along the
major axis. The slit is chosen to match exactly the pixel size,
slit width, and resolution limits of the best current
observations. {\em Bottom:} The observations of M31. Our mock
profiles are matched to the resolution of the {\em Hubble Space
Telescope} observations shown here (black squares)
\citep{bender:m31.nuclear.disk.obs}. The magenta circles are
ground-based observations \citep{kormendybender:m31} which
extend to larger radii but have inferior resolution and smooth
the velocity field at $|R|<1\,$arcsec.
\label{fig:velocities}}
\end{figure}
As the flow of gas through the nuclear region subsides, a stellar
remnant will remain behind that can retain the eccentric pattern. The
characteristic radii $\sim 1-10$ pc and stellar masses $\sim 0.1-1 \,
M_{\rm BH}$ of the eccentric nuclear disks in our simulations are
reasonably consistent with those observed in M31 and other systems.
Figure~\ref{fig:tile.10pc} shows stellar density maps for several of
our simulations, including those that have exhausted most of their
gas. The distinct nuclear disk is evident. In an edge-on projection,
several of these nuclear disks appear to have double nuclei (secondary
brightness peaks) -- this is caused by the high density of stars near
apocenter in their elliptical orbits. In several cases, this closely
resembles the secondary brightness peak in M31 (the P1 feature),
believed to arise in the same manner; note in particular the
second-from-bottom edge-on panel, where the two peaks are close in
brightness and separated by $\sim4$\,pc. In our simulations we
confirm what has been inferred from dynamical models of P1, that the
appearance of such secondary nuclei is sensitive to projection
effects, and usually requires a sightline reasonably close to edge-on.
Figure~\ref{fig:velocities} shows the velocity field of the stars in
these relics, long after the gas is exhausted, and scaled as if
observed at the center of M31; the data for M31 are in the bottom
panel. The overall agreement is impressive, particularly given that
these simulations are not designed to reproduce the observed features
of M31 in any way, but rather to study the growth of massive BHs. We
find similar agreement when comparing to observations of the nuclear
disk in NGC 4486b \citep{lauer:ngc4486b}. The rotational velocity field of
M31 is slightly less symmetric than our ``average'' simulation, but
a number agree quite closely, and the level of dispersion asymmetry observed is typical for our simulations.
\begin{figure}
\centering
\plotone{parameters_vs_r_nature.pdf}
\caption{Properties of the simulated nuclear stellar disks
versus radii, over the observed scales in M31. Each solid line corresponds
to a simulation, as in Figure~\ref{fig:velocities}.
Where available, we compare with the properties of the M31
system inferred from observations (grey shaded regions).
{\em Top Left:} Stellar mass surface density.
{\em Top Right:} Mean eccentricity of the disk along its major axis.
{\em Bottom Left:} Angular pattern speed (precession rate)
of the disk, which is much less than the angular velocity of
individual stars in the disk (dotted lines).
{\em Bottom Right:} Inflow rate of gas driven by the
gravitational torques of the eccentric disk itself, during the active phase when the
disk formed. This accretion can produce most of the BHs growth.
\label{fig:parameters}}
\end{figure}
Figure~\ref{fig:parameters} compares a number of the properties of our
simulated relic stellar disks with those of the M31 system inferred
from detailed kinematic studies
\citep{salow:nuclear.disk.models,jacobs:longlived.lopsided.disk.modes,
sambhus:m31.nuclear.disk.model,peiris:m31.nuclear.disk.models,
salow:nuclear.disk.models.2}. The range of simulations shown
includes both variations in initial conditions and the treatment of
stellar feedback (the sub-grid turbulent velocity).
The plotted surface densities are azimuthally averaged (this suppresses
the double-peaked appearance of the
simulations and M31, but is more robust to projection effects).
Note that the pattern speed $\Omega_{p}$ in our simulations is quite
low $\approx1-5 \, {\rm km\,s^{-1}\,pc^{-1}}$ (lower left), much less
than the rotation rate of individual stars at small radii. In our
simulations, the pattern speed is set at the large radii where the
eccentric mode begins.
The actual precession rate in M31 is not very well-constrained, but
most studies place an upper limit of $<30\,{\rm km\,s^{-1}}$
\citep{sambhus:m31.pattern.speed}, and several studies imply a value
close to our prediction \citep[see][]{bacon:m31.disk}. The mean disk
eccentricity in the simulations is also in broad agreement with that
observed, although we find that this is a less robust property of the
simulations and varies significantly from one simulation to another.
Figure~\ref{fig:parameters} also shows the inflow rates generated by
the nuclear stellar disk during its active/gas-rich phase, as a
function of radius. Nuclear disks that are similar to M31 in their
relic properties generate accretion rates up to several $M_{\sun}\,{\rm
yr^{-1}}$ during the quasar epoch, when they are gas-rich. This
highlights the key role that eccentric stellar disks can play in
fueling the growth of their host BHs.
Our models include a very simplified treatment of the feedback from
supernovae and massive stars: we introduce a sub-grid non-thermal
sound speed that is a proxy for the the effective turbulent speed of
the interstellar gas. To test the impact of this on our results, we
carried out calculations with identical initial conditions and
turbulent velocities ranging from $\sim10-100\,{\rm km\,s^{-1}}$,
roughly the lower and upper limits allowed by observational
constraints for the systems of interest (see Fig. 1 of
\citealt{hopkins:zoom.sims} and references therein). The value of the
sub-grid sound speed has a significant effect on the amount of
resolved sub-structure in the simulation, with more sub-structure
present in simulations with lower turbulent velocities. This is not
surprising since larger turbulent velocities raise the Jeans
mass/length, above which gravity is the dominant force. However, all
of the simulations show a similar nuclear lopsided disk. In terms of
the properties shown in Figures~\ref{fig:velocities} \&
\ref{fig:parameters}, the differences produced by changing the
sub-grid model are similar to the differences produced by somewhat
different galaxy properties. The fundamental reason for the weak
dependence on the subgrid model is that the torques in our simulations
are primarily determined by gravity, not hydrodynamic forces or
viscosity (see \citealt{hopkins:zoom.sims} for a detailed discussion).
The primary role of the subgrid feedback model is simply to prevent
catastrophic fragmentation of the galactic gas.
\vspace{-0.7cm}
\section{Discussion}
\label{sec:discussion}
\citet{hopkins:zoom.sims} argue that the dominant mechanism of angular
momentum transport in gas-rich galactic nuclei, from near the BH
radius influence ($\sim 10$ pc) down to the Keplerian viscous
accretion disk ($\ll 0.1$ pc), is gravitational torques produced by an
eccentric, lopsided disk (an $m = 1$ mode); this asymmetric disk forms
in our simulations when the disk mass is at least $\sim 10 \%$ of the
BH mass.
These torques provide the ``missing link'' connecting the gas
reservoir on galactic scales to the small-scale accretion disk near
the central BH. In this {\em Letter} we have shown that the
long-lived (``fossil'') stellar relics of these disks are remarkably
similar to the eccentric stellar disk observed around the BH in M31.
This suggests that the stellar kinematics and morphology in galactic
nuclei can provide new insights into the physics of BH growth.
Emboldened by our success, we can use the observed properties of the
M31 disk to infer the accretion it was responsible for. If the M31
disk was at one point gas-rich, the eccentric pattern in the stars
would produce strong torques in the gas, leading to an accretion rate
of $\dot M \sim \Sigma_{\rm gas}\,R^{2}\,\Omega\,|\Phi_{1}/\Phi_{0}|$
where $\Phi_{1}$ and $\Phi_{0}$ are the asymmetric and axisymmetric
terms in the gravitational potential, respectively. For the measured
BH mass \citep[$10^{8}\,M_{\sun}$,][]{bender:m31.nuclear.disk.obs} and
potential of the eccentric disk \citep{peiris:m31.nuclear.disk.models}
this implies an accretion rate of $\sim 1 \, M_{\sun}\,{\rm yr^{-1}}$,
close to the Eddington limit of $2.4\,M_{\sun}\,{\rm yr^{-1}}$. More
directly, we find that simulations that yield stellar relics in
closest agreement with M31 have typical inflow rates at $R\lesssim
0.1\,$pc in their active phases of $\sim0.3-5\,M_{\sun}\,{\rm yr^{-1}}$
(Fig. \ref{fig:parameters}).
These accretion rates imply that the observed stellar disk could have
helped the M31 BH gain much of its mass.
During the gas-rich phase, the typical column density of gas for an
edge-on line of sight through the disk is $N_{H} \sim 10^{25-26}\,{\rm
atoms\,cm^{-2}}$, sufficient to obscure the radiation from the BH
even in the X-rays. A combination of our model of stellar feedback
and self-consistently calculated gravitational perturbations generate
large `random' motions in the gas: the disks are thus thick, with
column densities sufficient to block the optical light ($N_{H} \gtrsim
10^{22}\,{\rm atoms \, cm^{-2}}$) out to an angle $\sim 20-45\deg$
above the plane -- in other words, a fraction $\sim30-60\%$ of all
sightlines will be obscured by gas and dust in the nuclear disk. More
detailed conclusions about this obscuration will require a better
understanding of the role of gravitational heating and stellar
feedback in this unusual region. Nonetheless, the properties of the
obscuring disk we infer are strikingly similar to those invoked for
the canonical ``toroidal obscuring region,'' assumed to reside on
small scales and to account for most of the optically obscured AGN
population \citep{antonucci:agn.unification.review,
urry:radio.unification.review,lawrence:receding.torus}. In the
context of our model, the observed ubiquity of the torus suddenly has
a dynamical origin: it itself helps {drive} the accretion.
Understanding the longevity of the eccentric disk in M31 has been as
challenging as understanding its origin. The precession rate of the
disk is slightly different at different radii -- this should lead to
phase-mixing that ultimately wipes out the coherent eccentricity of
the disk. Indeed, we do see that the eccentric pattern damps away at
larger radii; however, the pattern at radii $\sim$pc, where the M31
disk is observed, persists in our simulations as long as they can be
reliably evolved, for $10^{8}$\,yrs, which is $\sim 10^{4}$ dynamical
times. Self-gravity is likely to help maintain the pattern even
longer, in principle for much longer than the age of the universe
\citep{bacon:m31.disk,jacobs:longlived.lopsided.disk.modes,salow:nuclear.disk.models.2}.
Our simulations demonstrate that eccentric stellar and gaseous disks
form whenever the mass in the disky component in the central $\sim
10-30$ pc is comparable to that of the BH \citep{hopkins:zoom.sims}.
It is unclear, however, under what conditions these nuclear stellar
disks will survive to the present day. The long-term stability of
isolated eccentric disks is not fully understood. In addition, it is
easy to imagine that subsequent ``dry'' galaxy-galaxy mergers might
destroy these features. This is a key question for future research.
Observationally, there are a number of candidate systems in addition
to M31, as evidenced by apparently offset centers, ``hollow'' central
light profiles, double nuclei, or chemically distinct secondary nuclei
\citep{lauer:central.minimum.ell,lauer:centers,
debattista:vcc128.binary.nucleus,thatte:m83.double.nucleus,
afanasiev:2002.ngc5055.nuclear.disk}; and NGC4486b is another
confirmed eccentric disk \citep{lauer:ngc4486b}. There may be similar
features in nuclear gas disks as well
\citep{seth:ngc404.nuclear.disk}. But eccentric disks clearly do not
exist in all systems, as evidenced by the nucleus of M32.
More detailed observational constraints on the fraction of galaxies
with old, asymmetric nuclear stellar disks would provide a strong
constraint on our models.
When nuclear eccentric disks do survive, our models imply that there
should be a correspondence between the properties of the nuclear
stellar disk (mass, radius, and asymmetry) and the central BH mass, as
we have demonstrated is the case in M31. Very low pattern speeds
$<10\,{\rm km\,s^{-1}\,pc^{-1}}$ should be ubiquitous, as they are
required for efficient exchange of angular momentum between the stars
and gas. It is also worth noting that if some gas flows in from
larger radii, or accumulates via stellar evolution, the nuclear
regions can experience recurrent low-level AGN activity and/or star
formation, regulated by the same mechanism of eccentric stellar torques
(e.g., M31's young stellar population;
\citealt{chang:m31.eccentric.disk.model}).
Such episodes may complicate dating the formation epochs of these
disks, but also provide a laboratory to study the physics of inflow in
detail. The properties of the nuclear disk in its gas-rich phase
determines the distribution of implied torus scale lengths and gas
densities, which can be probed by infrared adaptive optics
observations of nearby bright AGN
\citep{davies:sfr.properties.in.torus,hicks:obs.torus.properties}. On
the theoretical side, further improvements in the treatment of gas
physics and star formation will enable more detailed comparison with
observations.
\vspace{-0.5cm}
\begin{small}\section*{Acknowledgments}\end{small} We thank Phil Chang, Lars Hernquist, Scott Tremaine,
John Kormendy, and Tod Lauer for helpful discussions during the
development of this work. Support for PFH and EQ was provided by the
Miller Institute for Basic Research in Science, University of
California Berkeley. EQ was also supported in part by NASA grant
NNG06GI68G and the David and Lucile Packard Foundation.
\\
|
\section{Introduction}
Theoretical efforts to interpret the observational data and to understand the nature of the dark sector do necessarily involve a vivid interplay between general relativity, astrophysics and particle physics.
Since the simplest solution to account for the late time accelerated expansion of the Universe, the one given in terms of a tiny positive cosmological constant, is plagued with conceptual problems, one has been compelled to examine different solutions \cite{Ame02,Kam02,Bil02,Ber02,Cal03,Mot04,Bro06A}.
Motivated by the high energy physics, an interesting alternative for obtaining the necessary negative pressure to account for the accelerated expansion involves the dynamics of a scalar field, $\phi$, evolving slowly down its potential $V\bb{\phi}$ \cite{Pee87,Rat87}.
These models assume that the vacuum energy can vary; a feature discussed much earlier \cite{Bro33}.
Other alternatives include $k$-essence \cite{Chi00,Arm01}, phantom energy models \cite{Sch01,Car03}, {\em cosmon} fields \cite{Wet88,Wet94}, and also several modifications of gravity \cite{Def02,Car04,Ama06}.
A challenging related issue concerns models of mass varying particles \cite{Gu03,Far04,Bja08}.
These are coupled to a light scalar field that drives their energy through their dynamical mass.
The idea of this mass varying mechanism \cite{Far04,Bro06A,Bja08} is to introduce a coupling between a relic particle, usually neutrinos, and the dark sector: dark energy or dark matter separately, or all together \cite{Ber08A,Ber08B}.
Such models admit an adiabatic regime in which the scalar field, usually related with dark energy, rolls down the minimum of its effective potential, which is set by the dark matter dynamical mass.
As a direct consequence of this new interaction, the particle mass is altered by the dynamics of the scalar field.
Due to phenomenological reasons, one still expects a small contribution from neutrinos to the cosmic dynamics.
In fact, it is well-known that the active neutrino masses are tiny as compared to the masses of the charged fermions.
The smallness of the neutrino masses is usually understood in terms of the {\em seesaw} mechanism in extensions of the standard model (SM) of the electroweak (EW) interactions.
The EW interactions involve only {\em left} handed neutrinos such that no renormalizable mass term for them is compatible with the SM gauge symmetry $SU(2)_{L}\otimes U(1)_{Y}$.
Once one assumes the conservation of baryon and lepton numbers, the {\em seesaw} mechanism admits neutrino masses from dimension five operators.
Neutrino masses should then involve two powers of the vacuum expectation value of the Higgs doublet.
These masses are suppressed by the inverse power of a large mass scale, $M$, of a {\em right} handed Majorana neutrino.
This super-heavy Majorana neutrino is associated to lepton number violating effects in extensions of the SM.
Assuming that the sterile neutrino mass, $\mathcal{M}\equiv \mathcal{M}\bb{\phi}$, exhibits a dynamical behavior driven by the scalar field, $\phi$, then the sterile neutrino becomes an interesting candidate for the aforementioned mass varying dark matter.
This, through the {\em seesaw} mechanism, gives origin to a mass varying neutrino (MaVaN) \cite{Gu03,Far04,Bja08} component corresponding to active neutrinos that exhibit secondary mass effects due to the {\em seesaw} mechanism.
In models of coupled dark energy, in which the scalar field couples to other matter components, it is natural to expect a coupling to active and to sterile neutrinos.
In this work, one explores the consequences of such a coupling.
This leads to cosmological scaling solutions where dark energy tracks the evolution of matter and/or radiation.
For the present cosmological epoch, it is predicted that the energy densities of dark energy and matter are of the same order of magnitude.
If the tracking can be altered by the growing mass of neutrinos, such that they become nonrelativistic (NR) at low redshift,
one can interestingly match the observed universe.
One should notice that mass varying dark matter is unusual in the formulation of the MaVaN models.
The unification of dark energy and dark matter naturally offers this possibility.
The generalized Chaplygin gas (GCG) is particularly relevant in this respect \cite{Kam02,Bil02,Ber02} as it is shown to be consistent with the observational constraints from CMB \cite{Ber03}, supernova \cite{Sup1,Ber04,Ber05}, gravitational lensing surveys \cite{Ber03B}, and gamma ray bursts \cite{Ber06B}.
Once one sets the mass dependence on $\phi$, the dark sector scalar field, one notices that the cosmological evolution of the unified fluid composed by mass varying dark matter and evolving dark energy has a dynamics similar to that of the {\em cosmon} field \cite{Wet94,Wet02,Wet08}.
Thus, at least partially, it turns out that the mass varying mechanism is the main agent of stability and of the cosmic acceleration.
Any cosmological fluid which effectively reproduces the effects of the GCG naturally offers analogous possibilities \cite{Ber09}.
Assuming that the scalar field drives the behavior of the masses of active and sterile neutrinos, the mass varying mechanism and the conditions for a stable cosmological scenario naturally emerges in the context of the GCC model.
Remarkably, the {\em seesaw} masses and energy densities of active and sterile neutrinos can be consistently embedded in the GCG scenario without any additional assumption.
Such a connection is mediated by a mass scale, $\xi$, and a dimensionless scale, $s$, similarly to {\em cosmon}-type dynamical equations \cite{Wet94,Wet02,Wet08}.
These are dubbed as {\em seesaw cosmon} equations.
This procedure gives origin to a remarkable convergence of three distinct frameworks: the {\em cosmon}-like dynamics, the {\em seesaw} mechanism for mass generation and the GCG scenario.
The equation of state of the perturbations is the very one of the background cosmology so that all effective results arising from the GCG model are maintained, although modified by neutrinos in a quite subtle way.
This work is organized as follows.
The decoupling mechanism for active and sterile neutrinos is described in section II.
This is obtained through the coupling of a scalar field in the {\em seesaw} relations, which leads to the dynamical properties ensued by the mass varying mechanism.
In section III, the interplay with the GCG model is discussed.
In section IV, the main properties of a unified treatment of dark matter and dark energy in a dark energy scenario for a $\Lambda$-like equation of state, $p = - \rho$, is reviewed.
This leads to a dynamical mass prescription different from the one obtained through the simplified version of the {\em seesaw} mechanism explored in section III.
Specifical mass dependencies on the scalar field for which the mass varying dark matter results in an {\em
effective} GCG model are discussed.
Finally, the results for energy densities, growing neutrino mass and stability conditions from a positive squared speed of sound $c_{s}^{\mbox{\tiny $2$}}$ are discussed in section V.
Three cosmological scenarios are examined, actually associated to three different growing mass relations: $\mu\bb{\phi} = \phi \exp{\left[-3\bb{\alpha + 1}\phi/2\right]}$ (Case 01), $\mu\bb{\phi} = \exp{\left[-3\bb{\alpha + 1}\phi/2\right]}$ (Case 02) and $\mu\bb{\phi} = \tanh{\left[3\bb{\alpha + 1}\phi/2\right]}^{\frac{\mbox{\tiny $2$}\alpha}{\alpha + 1}}\exp{\left[-3\bb{\alpha + 1}\phi\right]}$ (Case 03).
The obtained results indicate that the proposed approach is quite appealing as it unifies neutrinos, dark matter and dark energy.
In section VI, the main implications of the developed scenario are summarized.
\section{The {\em seesaw} mechanism for MaVaN's}
The connection of mass varying dark matter with neutrinos does provided, as will be seen, interesting constraints on the neutrino masses, on the dark energy density, and on the equations of state and the stability conditions.
This can be understand through the equations arising from the Lagrangian densities of active ($A$) and sterile ($S$) neutrinos, $\psi_{\mbox{\tiny $A$},\mbox{\tiny $S$}}$,
\begin{equation}
\mathcal{L}_{\mbox{\tiny $A$},\mbox{\tiny $S$}} = i\bar{\psi}_{\mbox{\tiny $A$},\mbox{\tiny $S$}}\gamma_{\mu}\partial^{\mu}\psi_{\mbox{\tiny $A$},\mbox{\tiny $S$}} + k_{\mbox{\tiny $A$},\mbox{\tiny $S$}}\bar{\psi}_{\mbox{\tiny $A$},\mbox{\tiny $S$}}\psi_{\mbox{\tiny $A$},\mbox{\tiny $S$}},
\label{aprl00A}
\end{equation}
where two mass scales, $k_{\mbox{\tiny $A$}} = \mu$ and $k_{\mbox{\tiny $S$}} = \mathcal{M}$, have been introduced.
The {\em seesaw} mechanism suggests that $\mu$ is small, while $\mathcal{M}$ should be large:
\begin{equation}
-\mu = (M/2)[1 - \sqrt{1 + 4 (m/M)^{\mbox{\tiny $2$}}}] ~~\mbox{and}~~\mathcal{M} = (M/2)[1 + \sqrt{1 + 4 (m/M)^{\mbox{\tiny $2$}}}].
\label{aprl02}
\end{equation}
These states correspond to the eigenvalues of the mass matrix $\left[\begin{array}{cc} 0 & m\\ m & M\end{array}\right]$ written in the orthogonal basis of chiral {\em left}- and {\em right}-handed neutrinos, $\nu_{\mbox{\tiny $L$},\mbox{\tiny $R$}}$, related with the matter fields, $\psi_{\mbox{\tiny $A$},\mbox{\tiny $S$}}$, by
\begin{equation}
\psi_{\mbox{\tiny $A$}} = \sqrt{s^{\mbox{\tiny $2$}} + 1} (s \nu_{\mbox{\tiny $L$}} - \nu_{\mbox{\tiny $R$}})~~\mbox{and}~~
\psi_{\mbox{\tiny $S$}} = \sqrt{s^{\mbox{\tiny $2$}} + 1} ( \nu_{\mbox{\tiny $L$}} + s \nu_{\mbox{\tiny $R$}}),
\label{aprl00B}
\end{equation}
where the dimensionless quantity $s = \sqrt{\mu/\mathcal{M}}$ has been introduced.
It follows that $\mathcal{L}_{\mbox{\tiny $A$}} + \mathcal{L}_{\mbox{\tiny $S$}} = \mathcal{L}_{\mbox{\tiny $L$}} + \mathcal{L}_{\mbox{\tiny $R$}} + \mathcal{L}_{\mbox{\tiny $L$}\mbox{\tiny $R$}}$, where $\mathcal{L}_{\mbox{\tiny $L$},\mbox{\tiny $R$}}$ correspond to the kinetic terms and $\mathcal{L}_{\mbox{\tiny $L$}\mbox{\tiny $R$}}$ yields the mass mixing terms.
The equivalence between the stress-energy tensor in the chiral basis and the matter field basis follows from the relationships for the energy density $\rho_{\mbox{\tiny $A$}} + \rho_{\mbox{\tiny $S$}} = \rho_{\mbox{\tiny $L$}} + \rho_{\mbox{\tiny $R$}}$ and for the pressure $p_{\mbox{\tiny $A$}} + p_{\mbox{\tiny $S$}} = p_{\mbox{\tiny $L$}} + p_{\mbox{\tiny $R$}}$.
After introducing an auxiliary mass scale $\xi = \sqrt{\mu \mathcal{M}} \equiv m$, one can define two energy scales, $\rho_{\mbox{\tiny $1$}} = (\rho_{\mbox{\tiny $S$}} + \rho_{\mbox{\tiny $A$}})/2$ and $\rho_{\mbox{\tiny $2$}} = (\rho_{\mbox{\tiny $S$}} - \rho_{\mbox{\tiny $A$}})/2$ which evolve as reciprocally coupled {\em cosmon}-type equations.
In a FRW universe it corresponds to
\begin{equation}
\dot{\rho}_{\mbox{\tiny $1$}} + 3 H (\rho_{\mbox{\tiny $1$}} + p_{\mbox{\tiny $1$}}) - \dot{\phi}\frac{\mbox{d} \xi}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $1$}}}{\partial \xi} + \dot{\phi}\frac{\mbox{d} s}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $2$}}}{\partial s} = 0,
\end{equation}
and
\begin{equation}
\dot{\rho}_{\mbox{\tiny $2$}} + 3 H (\rho_{\mbox{\tiny $2$}} + p_{\mbox{\tiny $2$}}) - \dot{\phi}\frac{\mbox{d} \xi}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $2$}}}{\partial \xi} + \dot{\phi}\frac{\mbox{d} s}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $1$}}}{\partial s} = 0,
\label{aprl02A}
\end{equation}
where $H = \dot{a}/{a}$ is the expansion rate of the universe and the {\em overdot} denotes differentiation with respect to time ($^{\cdot}\, \equiv\, d/dt$).
The third terms in the above Eqs. correspond to a mass varying term (see Eq.~(\ref{gcg02}) in the Appendix), while the last ones are associated to the exchange of energy due to the non-adiabatical behavior of energy densities $1$ and $2$.
Notice that there is no sense in defining a Lagrangian density for the component $\rho_{\mbox{\tiny $2$}}$, which does not correspond to an observable energy density scale.
The $\rho_{\mbox{\tiny $1$}}$ component can be identified to matter fields, while $\rho_{\mbox{\tiny $2$}}$ is an auxiliary energy density which measures the coupling between the physical observables.
In fact, both energy densities, $1$ and $2$, are driven by {\em cosmon}-type equations.
The {\em cosmon} signature is revealed by the dependence of the scales $\xi$ and $s$ on the value of a slowly varying classical scalar field $\phi$, the {\em seesaw cosmon} field \cite{Wet94}.
Since the scalar field depends on the scale factor $a$, $\phi \equiv \phi\bb{a}$, the {\em seesaw} mass terms get transformed into dynamical quantities, $\mu\bb{\phi}$ and $\mathcal{M}\bb{\phi}$.
After combining Eqs.~(\ref{aprl02}), and observing that,
\begin{equation}
\rho_{\mbox{\tiny $A$}}\bb{a, \xi, s} = \rho_{\mbox{\tiny $A$}}\bb{a, \xi s},
\label{aprl023A}
\end{equation}
and
\begin{equation}
\rho_{\mbox{\tiny $S$}}\bb{a, \xi, s} = \rho_{\mbox{\tiny $S$}}\bb{a, \xi/s},
\label{aprl023}
\end{equation}
it is easy to identify the evolution of these with the evolution of active and sterile energy densities, $\rho_{\mbox{\tiny $A$}}$ and $\rho_{\mbox{\tiny $S$}}$, by means of decoupled equations
\begin{equation}
\dot{\rho_{\mbox{\tiny $A$}}} + 3 H (\rho_{\mbox{\tiny $A$}} + p_{\mbox{\tiny $A$}}) - \dot{\phi}\frac{\mbox{d} \mu}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $A$}}}{\partial \mu} = 0,
\label{aprl03A}
\end{equation}
and
\begin{equation}
\dot{\rho_{\mbox{\tiny $S$}}} + 3 H (\rho_{\mbox{\tiny $S$}} + p_{\mbox{\tiny $S$}}) - \dot{\phi}\frac{\mbox{d} \mathcal{M}}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $S$}}}{\partial \mathcal{M}} = 0,
\label{aprl03B}
\end{equation}
as
\begin{equation}
\frac{\mbox{d} \ln{\xi}}{\mbox{d} \phi} =
(1/2)\left(\frac{\mbox{d} \ln{\mu}}{\mbox{d} \phi} + \frac{\mbox{d} \ln{\mathcal{M}}}{\mbox{d} \phi}\right)\nonumber\\,
\end{equation}
and
\begin{equation}
\frac{\mbox{d} \ln{s}}{\mbox{d} \phi} =
(1/2)\left(\frac{\mbox{d} \ln{\mu}}{\mbox{d} \phi} - \frac{\mbox{d} \ln{\mathcal{M}}}{\mbox{d} \phi}\right)\nonumber\\.
\end{equation}
Thus, mass varying mechanism translates the dependence of the mass terms on the scale factor, $a$, i.e. $\mu\bb{a}$ and $\mathcal{M}\bb{a}$ (see the Appendix).
The coupling between relic particles and the scalar field as described by Eqs.~(\ref{aprl03A})-(\ref{aprl03B}) is relevant only for NR fluids .
Since the strength of the coupling is suppressed by the relativistic pressure, as long as particles are ultra-relativistic (UR), the active and sterile neutrino energy densities, $\rho_{\mbox{\tiny $A$}}$ and $\rho_{\mbox{\tiny $S$}}$, decouple from each other and evolve adiabatically \cite{Bea08}, remaining coupled only to the scalar field.
That is, in the UR regimes, $\frac{\partial\rho}{\partial m}\sim\frac{\partial\rho}{\partial s}\propto (\rho- 3 p)\approx 0$.
In a previous work, it has been suggested that one could treat MaVaN's as a perturbative component derived from an unperturbed adiabatic energy density solution $\rho_{\phi}$ \cite{Ber08A,Ber08B}.
The above results provides a quantitative justification for that.
As can be seen, all the information from the dark sector (dark energy plus dark matter) acting on the (active) neutrino sector is carried out by the explicit dependence of $\mu \equiv\mu\bb{\phi}$.
From the cosmological point of view, it results in mass eigenstantes for active award sterile neutrinos that evolve separately, which is not the case for the coupled chiral eigenstates $\nu_{\mbox{\tiny $L$}}$ and $\nu_{\mbox{\tiny $R$}}$.
At primordial times, when $s^{\mbox{\tiny $2$}} \sim 1$, such mass eigenstates are indistinguishable, and the chiral eigenstates are well-defined.
At late times they turn into sterile and active mass eigenstates, maintaining the identity of the flavour sectors.
In order to proceed, one identifies the large mass energy density component of the above equations, $\rho_{\mbox{\tiny $S$}}$, to the energy density of dark matter, then the {\em seesaw cosmon} framework provides the connection between dark matter and dark energy through the {\em cosmon} field equation \cite{Wet88,Wet94},
\begin{equation}
\dot{\rho}_{\phi} + 3 H (\rho_{\phi} + p_{\phi}) + \dot{\phi}\frac{\mbox{d} \mathcal{M}}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $S$}}}{\partial \mathcal{M}} = 0,
\label{aprl03C}
\end{equation}
originally written as
\begin{equation}
\ddot{\phi} + 3 H \dot{\phi} + \frac{\mbox{d} V\bb{\phi}}{\mbox{d} \phi} = - \frac{\mbox{d} \mathcal{M}}{\mbox{d} \phi} \frac{\partial \rho_{\mbox{\tiny $S$}}}{\partial \mathcal{M}},
\label{gcg04}
\end{equation}
with the usual assignments for the $\rho_{\phi}$ and $p_{\phi}$ (c. f. Eqs.(\ref{pap01})).
Eqs.~(\ref{aprl03B}) - (\ref{aprl03C}) result in the adiabatic equation for the dark sector,
\begin{equation}
\dot{\rho}_{\mbox{\tiny $D$}\mbox{\tiny $S$}} + 3 H (\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}} + p_{\mbox{\tiny $D$}\mbox{\tiny $S$}}) = 0,
\label{aprl04}
\end{equation}
with $\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}} = \rho_{\phi} + \rho_{\mbox{\tiny $S$}} = \rho_{\phi} + \rho_{\mbox{\tiny $1$}} + \rho_{\mbox{\tiny $2$}}$ and $H^{\mbox{\tiny $2$}} = \rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}}$ (with $H$ in units of $H_{\mbox{\tiny $0$}}$ and $\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}}$ in units of $\rho_{\mbox{\tiny Crit}} = 3 H^{\mbox{\tiny $2$}}_{\mbox{\tiny $0$}}/ 8 \pi G)$.
Despite the intrinsic dependence on $\phi$, the equation of motion for the dark sector is not modified by $\rho_{\mbox{\tiny $A$}}$, the active neutrino energy density component.
The cosmological dependence of $\rho_{\mbox{\tiny $A$}}$ on $\phi$ can be computed through Eq.~(\ref{aprl03A}) considering active neutrinos as a test fluid.
The phenomenological consistency of the proposed scenario can be assessed quantitatively expressing $\rho_{\mbox{\tiny $A$}}$ and $\rho_{\mbox{\tiny $S$}}$ as energy densities of a degenerate fermion gas (DFG) at different relativistic regimes,
\small
\begin{eqnarray}
\rho_{\mbox{\tiny $A$}}\bb{a} &=& (8 \pi^{\mbox{\tiny $2$}})^{\mbox{\tiny $-$}\mbox{\tiny $1$}}
\mu\bb{a}^{\mbox{\tiny $4$}}\left[\eta\bb{a} (2 \eta\bb{a}^{\mbox{\tiny $2$}} + 1)\sqrt{\eta\bb{a}^{\mbox{\tiny $2$}} + 1} -
\mbox{arc}\sinh{(\eta\bb{a})}\right]\nonumber\\
\rho_{\mbox{\tiny $S$}}\bb{a} &=& (8 \pi^{\mbox{\tiny $2$}})^{\mbox{\tiny $-$}\mbox{\tiny $1$}}
\mu\bb{a}^{\mbox{\tiny $4$}}\, s^{\mbox{\tiny $-$}\mbox{\tiny $8$}}\left[\gamma s^{\mbox{\tiny $2$}} \eta\bb{a} (2 \gamma^{\mbox{\tiny $2$}} s^{\mbox{\tiny $4$}} \eta\bb{a}^{\mbox{\tiny $2$}} + 1)\sqrt{\gamma^{\mbox{\tiny $2$}} s^{\mbox{\tiny $4$}} \eta\bb{a}^{\mbox{\tiny $2$}} + 1} -
\mbox{arc}\sinh{(\gamma s^{\mbox{\tiny $2$}} \eta\bb{a})}\right],
\label{aprl04B}
\end{eqnarray}
\normalsize
where $\eta\bb{a} = (T_{\mbox{\tiny $0$}}\,q_{\mbox{\tiny $A$}})/(a\,\mu\bb{a})$ and the relation between the fluid thermodynamic regimes is parameterized by the coefficient $\gamma = q_{\mbox{\tiny $S$}}/q_{\mbox{\tiny $A$}}$.
As discussed in the Appendix, the DFG prescription is suitable for describing the transition between UR and NR regimes.
In this case, the effects due to the coupled dark matter ($\mathcal{M}\bb{\phi}$) and dark energy ($\phi$) can be monitored through Eqs.~(\ref{aprl03A})-(\ref{aprl03C}).
In the Fig.~\ref{PRL01} one can see the exact correspondence between the abovementioned energy densities and the ``modified'' scale parameter $\gamma s^{\mbox{\tiny $2$}}$.
In the NR limit of a DFG, one has $\rho_{\mbox{\tiny $A$}}/\rho_{\mbox{\tiny $S$}} \sim \gamma^{\mbox{\tiny $-$}\mbox{\tiny $3$}} s^{\mbox{\tiny $2$}}$.
The characteristic magnitude of the active neutrino masses involves an appropriate combination of dimensionless Yukawa couplings, $Y_{\mbox{\tiny $j$}}$, $\mu_{\mbox{\tiny $j$}} \sim m_{\mbox{\tiny $j$}}^{\mbox{\tiny $2$}}/M$ with $m_{\mbox{\tiny $j$}} \sim Y_{\mbox{\tiny $j$}}\, v$ ($v\sim 2 \times 10^{\mbox{\tiny $1$}\1}\,eV$).
Consistency with the observed flavour oscillations requires for the neutrino mass at least one neutrino to have $\mu_{\mbox{\tiny $j$}}\gtrsim 0.05 \,eV$.
That is, $Y_{\mbox{\tiny $j$}}$ of the order one implies an upper bound for the large mass scale $\mathcal{M} \lesssim 10^{\mbox{\tiny $2$}\mbox{\tiny $3$}}\,eV$, from which follows that $s^{\mbox{\tiny $2$}} \gtrsim 10^{\mbox{\tiny $-$}\mbox{\tiny $2$}\mbox{\tiny $4$}}$.
Given that the present value of the ratio $\rho_{\mbox{\tiny $A$}}/\rho_{\mbox{\tiny $S$}}$ is, for phenomenological reasons, of $\mathcal{O}(10^{\mbox{\tiny $-$}\mbox{\tiny $2$}})$, a DFG of active neutrinos, at least approximately in the NR regime, leads to $\gamma^{\mbox{\tiny $-$}\mbox{\tiny $3$}}\sim 10^{\mbox{\tiny $2$}\2}$, and hence, two widely different momentum scales for active and sterile neutrinos, $q_{\mbox{\tiny $S$}}/q_{\mbox{\tiny $A$}} \sim 10^{\mbox{\tiny $-$}\mbox{\tiny $7$}}$.
Thus, sterile neutrinos behave like ultra cold dark matter (CDM).
In addition, if one assumes that $\rho_{\phi}$ could be parameterized by a $\lambda \phi^{\mbox{\tiny $4$}}$ theory, or some type of quintessence potential, the adiabatic evolution of the scalar field allows for assigning a mass for the scalar field.
Such predictions for $m_{\phi}$ are consistent with the lower bound on the mass derived from assumptions about the adiabatic evolution of the scalar field since nucleosynthesis.
The scalar field mass should then be greater than the rate of expansion at nucleosynthesis, which is of order of $\sim 10^{\mbox{\tiny $-$}\mbox{\tiny $1$}\mbox{\tiny $6$}}\, eV$.
However, the conditions for the adiabatic regime for the light scalar field to be settled down at the minimum of its potential prior to nucleosynthesis are quite model dependent.
The above discussion shows that a considerable fine-tuning is required to generate the different scalar field mass, a issue that demands for an embedding of the model in a more encompassing framework, such as, for instance, the Minimal Supersymmetric SM with the addition of one singlet chiral superfield \cite{Bal07}.
Another possible avenue involves a unified picture of dark energy and dark matter where the former corresponds to an additional singlet scalar field which one identifies to the quintessence field \cite{Ber09C,Ber09D}, while the latter with the quantum excitations of this singlet, which can be coupled to the Higgs boson.
\section{The interplay with the GCG model}
One considers now some generic features of the proposed model and explores its connection with the GCG model.
The GCG model is characterized by an exotic equation of state \cite{Ber02,Ber03} given by
\begin{equation}
p = - A_{\mbox{\tiny $s$}} \rho_{\mbox{\tiny $0$}} \left(\frac{\rho_{\mbox{\tiny $0$}}}{\rho}\right)^{\mbox{\tiny $\alpha$}},
\label{aprl05}
\end{equation}
where $A_{\mbox{\tiny $s$}}$ and $\alpha$ are constants.
This equation of state can be obtained from a generalized Born-Infeld action \cite{Ber02}.
Several studies yield convincing evidence that the GCG scenario is a phenomenologically consistent approach to explain the accelerated expansion of the universe.
Inserting the above equation of state into the unperturbed energy conservation Eq.~(\ref{aprl04}), one obtains through a straightforward integration \cite{Ber02}
\begin{equation}
\rho = \rho_{\mbox{\tiny $0$}} \left[A_{\mbox{\tiny $s$}} + \frac{(1-A_{\mbox{\tiny $s$}})}{\bb{a/a_{\mbox{\tiny $0$}}}^{\mbox{\tiny $3$}(\mbox{\tiny $1$}+\alpha)}}\right]^{\mbox{\tiny $1$}/(\mbox{\tiny $1$} \mbox{\tiny $+$} \mbox{\tiny $\alpha$})},
\label{gcg21}
\end{equation}
and
\begin{equation}
p = - A_{\mbox{\tiny $s$}} \rho_{\mbox{\tiny $0$}} \left[A_{\mbox{\tiny $s$}} + \frac{(1-A_{\mbox{\tiny $s$}})}{\bb{a/a_{\mbox{\tiny $0$}}}^{\mbox{\tiny $3$}(\mbox{\tiny $1$}+\alpha)}}\right]^{-\mbox{\tiny $\alpha$}/(\mbox{\tiny $1$} \mbox{\tiny $+$} \mbox{\tiny $\alpha$})}.
\label{gcg22}
\end{equation}
One of the most striking features of the above equations is that the energy density interpolates between a dust dominated phase in the past, where $\rho \propto a^{-\mbox{\tiny $3$}}$, and a de-Sitter phase, $\rho = -p$, at late times.
This evolution is ruled by the model parameters, $\alpha$ and $A_{\mbox{\tiny $s$}}$, which are positive and smaller than unity.
Of course, $\alpha = 0$ corresponds to the $\Lambda$CDM model.
It sets $0 < \alpha \leq 1$.
For $A_{\mbox{\tiny $s$}} =0$, GCG behaves like matter, whereas for $A_{\mbox{\tiny $s$}} =1$, it behaves as a cosmological constant.
Hence, in order to consider it as unified model for dark matter and dark energy one has to exclude these two possibilities so that $A_{\mbox{\tiny $s$}}$ must lie in the range $0 < A_{\mbox{\tiny $s$}} < 1$.
This property makes the GCG model an interesting candidate for the unification of dark matter and dark energy, i. e. for the dark sector energy density $\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}}$ of our proposal.
The GCG can be described by an underlying scalar field, $\phi$, which can be either real \cite{Kam02,Ber04} or complex \cite{Bil02,Ber02}.
In the former case, one can identify it with the {\em cosmon} field $\phi$ so that $\rho$ and $p$ are given in terms of $\phi$ by:
\begin{eqnarray}
\rho &=& \frac{1}{2}\dot{\phi}^{\mbox{\tiny $2$}} + V\bb{\phi},\nonumber\\
p &=& \frac{1}{2}\dot{\phi}^{\mbox{\tiny $2$}} - V\bb{\phi}.
\label{pap01}
\end{eqnarray}
This allows for obtaining the dependence of the scalar field, $\phi$, on the scale factor, $a$, and the explicit dependence of $\rho$, $p$ and $V$ on $\phi$.
Following Ref.~\cite{Ber04}, one obtains through Eqs.~(\ref{gcg21})-(\ref{pap01})
\begin{equation}
\dot{\phi}^{\mbox{\tiny $2$}}\bb{a} = \frac{\rho_{\mbox{\tiny $0$}}(1 - A_{\mbox{\tiny $s$}})}{\bb{a/a_{\mbox{\tiny $0$}}}^{\mbox{\tiny $3$}(\mbox{\tiny $\alpha$}\mbox{\tiny $+$}\mbox{\tiny $1$})}}
\left[A_{\mbox{\tiny $s$}} + \frac{(1-A_{\mbox{\tiny $s$}})}{\bb{a/a_{\mbox{\tiny $0$}}}^{\mbox{\tiny $3$}(\mbox{\tiny $\alpha$}\mbox{\tiny $+$}\mbox{\tiny $1$})}}\right]^{-\mbox{\tiny $\alpha$}/(\mbox{\tiny $\alpha$} \mbox{\tiny $+$} \mbox{\tiny $1$})},
\label{pap02}
\end{equation}
and assuming a flat universe described by Friedmann equation $H^{\mbox{\tiny $2$}} = \rho$ (again with $H$ in units of $H_{\mbox{\tiny $0$}}$ and $\rho$ in units of $\rho_{\mbox{\tiny Crit}} = 3 H^{\mbox{\tiny $2$}}_{\mbox{\tiny $0$}}/ 8 \pi G)$, one gets
\begin{equation}
\phi\bb{a} = - \frac{1}{3(\alpha + 1)}\ln{\left[\frac{\sqrt{1 - A_{\mbox{\tiny $s$}}(1 - \bb{a/a_{\mbox{\tiny $0$}}}^{\mbox{\tiny $3$}(\mbox{\tiny $\alpha$} \mbox{\tiny $+$} \mbox{\tiny $1$})})} - \sqrt{1 - A_{\mbox{\tiny $s$}}}}{\sqrt{1 - A_{\mbox{\tiny $s$}}(1 - \bb{a/a_{\mbox{\tiny $0$}}}^{\mbox{\tiny $3$}(\mbox{\tiny $\alpha$} \mbox{\tiny $+$} \mbox{\tiny $1$})})} + \sqrt{1 - A_{\mbox{\tiny $s$}}}}\right]},
\label{pap03}
\end{equation}
where it is assumed that
\begin{equation}
\phi_{\mbox{\tiny $0$}} = \phi\bb{a = a_{\mbox{\tiny $0$}}} = - \frac{1}{3(\alpha + 1)}\ln{\left[\frac{1 - \sqrt{1 - A_{\mbox{\tiny $s$}}}}{1 + \sqrt{1 - A_{\mbox{\tiny $s$}}}}\right]}.
\label{pap04}
\end{equation}
One then readily finds the scalar field potential,
\begin{equation}
V\bb{\phi} = \frac{1}{2}A_{\mbox{\tiny $s$}}^{\frac{\mbox{\tiny $1$}}{\mbox{\tiny $1$} \mbox{\tiny $+$} \mbox{\tiny $\alpha$}}}\rho_{\mbox{\tiny $0$}}\left\{
\left[\cosh{\left(3\bb{\alpha + 1} \phi/2\right)}\right]^{\frac{\mbox{\tiny $2$}}{\mbox{\tiny $\alpha$} \mbox{\tiny $+$} \mbox{\tiny $1$}}}
+
\left[\cosh{\left(3\bb{\alpha + 1} \phi/2\right)}\right]^{-\frac{\mbox{\tiny $2$}\mbox{\tiny $\alpha$}}{\mbox{\tiny $\alpha$} \mbox{\tiny $+$} \mbox{\tiny $1$}}}
\right\}.
\label{pap05}
\end{equation}
Thus, the dynamics of the GCG scalar field is given by the evolution of $\phi$ on the above potential.
Turning now to the simplest version of the {\em seesaw} mechanism for which a single flavour neutrino mass is linearly related to the scalar field, $\phi$, i. e. $m\bb{\phi} \sim \phi$, after observing that the logarithm of the squared scale parameter $s^{\mbox{\tiny $2$}}$ has an analytical structure similar to that of $\phi\bb{a}$,
\begin{equation}
\ln{(s^{\mbox{\tiny $2$}})} =\ln{(\mu/\mathcal{M})} = \ln{\left[\frac{\sqrt{1 + 4 (m/M)^{\mbox{\tiny $2$}}} - 1}{\sqrt{1 + 4 (m/M)^{\mbox{\tiny $2$}}} + 1}\right]},
\label{aprl09}
\end{equation}
one can rewrite the auxiliary scale $m/M$ in terms of the GCG parameters $A_{\mbox{\tiny $s$}}$ and $\alpha$, and on terms of the scale factor, $a$, as
\begin{equation}
\frac{m\bb{a}}{M\bb{a}} = \frac{m_{\mbox{\tiny $0$}}}{M_{\mbox{\tiny $0$}}} a^{\frac{3\bb{\mbox{\tiny $\alpha$} + \mbox{\tiny $1$}}}{2}} = \frac{1}{2} \sqrt{\frac{A_{\mbox{\tiny $s$}}}{1-A_{\mbox{\tiny $s$}}}} (a/a_{\mbox{\tiny $0$}})^{\frac{3\bb{\mbox{\tiny $\alpha$} + \mbox{\tiny $1$}}}{2}}
\label{aprl10}
\end{equation}
where $a_{\mbox{\tiny $0$}}$ is determined by a phenomenological adjustment (it can be set equal to unity if one redefines $A_{\mbox{\tiny $s$}}$ and $\rho_{\mbox{\tiny $0$}}$ in Eq.~(\ref{gcg21})).
After a simple mathematical manipulation, one obtains
\begin{equation}
\ln{(s^{\mbox{\tiny $2$}})} = \ln{(\mu/\mathcal{M})} = - 3\bb{\alpha + 1} \phi\bb{a},
\label{aprl09B}
\end{equation}
Eqs.~(\ref{aprl09})-(\ref{aprl09B}) yield the following relationships for active and sterile neutrino masses
\begin{equation}
\mu\bb{\phi} = \phi \exp{\left[-3\bb{\alpha + 1} \phi/2\right]} ~~~~ \mbox{and} ~~~~ \mathcal{M}\bb{\phi} = \phi \exp{\left[+3\bb{\alpha + 1} \phi/2\right]}.
\label{aprl11}
\end{equation}
The behavior set by Eqs.~(\ref{aprl10}) and (\ref{aprl11}) implies to the scale parameter $s$ an exponential dependence on $\phi$, $s = \exp{\left[-3\bb{\alpha + 1} \phi/2\right]}$, corresponding to an exponential divergency between mass scales.
This is consistent with several classes of quintessence models \cite{Bar99}.
Naturally, once the prescription for masses and couplings is known, the behavior of the neutrinos can be understood through Eqs.~(\ref{aprl03A}) and (\ref{aprl03B}).
It is evident that the procedure above described is fairly general.
The form of $M\bb{\phi}$ and modifications on the equation of state for dark energy can lead to quite different scenarios.
For instance, it also admits constant mass dark matter, which should result in neutrino masses given by $\mu\bb{a} \propto \exp{\left[-3\bb{\alpha + 1}\,\phi\bb{a}\right]}$, or even a constant mass scale $\xi = m$, as will be discussed in the analysis of the following section.
\section{Mass varying dark matter from an effective GCG}
Aiming to obtain a deeper understanding of the proposal of the previous section, one shows that the coupling of dark matter to dark energy through the mass varying mechanism in the {\em seesaw cosmon} scenario can be indeed matched with the GCG model.
In order to perform that, one assumes that the GCG equation of state Eq.~(\ref{aprl05}) just describes the behavior of a fluid with energy density $\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}}$.
From Eqs.(\ref{aprl03B}) and (\ref{aprl03C}), one sees that if the {\em seesaw cosmon} field obeys the following equation of state
\begin{equation}
\rho_{\phi}\bb{\phi} = - p_{\phi}\bb{\phi} = U\bb{\phi} = \left[A_{s}\cosh{\left(\frac{3\bb{\alpha + 1}\phi}{2}\right)}\right]^{\frac{\mbox{\tiny $2$} \mbox{\tiny $\alpha$}}{1 \mbox{\tiny $+$} \mbox{\tiny $\alpha$}}},
\label{pap08}
\end{equation}
thus the problem consists in obtaining the relationship between the scalar potential $U\bb{\phi}$ and the variable mass $\mathcal{M}\bb{\phi}$ which satisfies the equation
\begin{equation}
\frac{\mbox{d} U\bb{\phi}}{\mbox{d}{\phi}} +
\frac{\partial \rho_{\mbox{\tiny $S$}}}{\partial \mathcal{M}} \frac{\mbox{d} \mathcal{M}\bb{\phi}}{\mbox{d}\phi} = \frac{\mbox{d} U_{\mbox{\tiny Eff}}\bb{\phi}}{\mbox{d}{\phi}} = 0,
\label{pap09}
\end{equation}
a stationary condition directly obtained from Eq.~(\ref{aprl03C}).
The effective potential $U_{\mbox{\tiny Eff}}\bb{\phi}$ that sets the evolution of the scalar field has two terms: the first one arising from the original quintessence potential $U\bb{\phi}$, and the second one due to a coupling to the dynamical mass $\mathcal{M}\bb{\phi}$.
For suitable choices of $U\bb{\phi}$ and for couplings satisfying Eq.~(\ref{pap09}), the competition between these terms leads to a minimum of the effective potential.
In the {\em quasi}-static regime, it is possible for the field to adjust itself to the minimum of the potential in an adiabatic way.
The timescale for that must be short when compared to the timescale over which the background density is changing.
In this regime, matter and scalar field are tightly coupled and evolve effectively as a single fluid.
In the proposed approach, once assuming the equation of state $p_{\phi} = -\rho_{\phi}$, the described behavior is rather natural.
Eq.~(\ref{pap08}) leads to $\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}} + p_{\mbox{\tiny $D$}\mbox{\tiny $S$}} = \rho_{\mbox{\tiny $S$}} + p_{\mbox{\tiny $S$}}$ which, in the CDM limit, gives $p_{\mbox{\tiny $S$}} = 0$ and $\rho_{\mbox{\tiny $D$}\mbox{\tiny $S$}} + p_{\mbox{\tiny $D$}\mbox{\tiny $S$}} = \rho_{\mbox{\tiny $S$}}\bb{a} = \mathcal{M}\bb{a} \, n\bb{a}$, where $n\bb{a}$ is the particle number density.
Since the dependence of $\mathcal{M}$ on $a$ is mediated by $\phi\bb{a}$, i. e. $\mathcal{M}\bb{a} = \mathcal{M}\bb{\phi\bb{a}}$,
from Eqs.~(\ref{gcg21}), (\ref{gcg22}) and (\ref{pap03}), it follows after some mathematical manipulations that
\begin{equation}
\mathcal{M}\bb{\phi} = \mathcal{M}_{\mbox{\tiny $0$}}
\left[\frac{\tanh{\left(3\bb{\alpha + 1}\frac{\phi}{2}\right)}}{\tanh{\left(3\bb{\alpha + 1}\frac{\phi_{\mbox{\tiny $0$}}}{2}\right)}}\right]^{\frac{ \mbox{\tiny $2$} \mbox{\tiny $\alpha$}}{1 \mbox{\tiny $+$} \mbox{\tiny $\alpha$}}},
\label{pap11}
\end{equation}
which is consistent with Eq.~(\ref{pap09}).
Hence, one sees that the adequacy to the adiabatic regime is conditioned by the mass varying behavior.
In a previous study \cite{Ber09B}, it has shown how the behavior of $\mathcal{M}\bb{\phi}$ and $U\bb{\phi}$ differs from the one of $\phi\bb{a}$ and $U\bb{\phi}$ in the GCG model, and how the composed fluid deviates from the GCG scenario.
It has been assumed that the mass varying dark matter behaves like a DFG in the relativistic regime (hot dark matter (HDM)) and in the NR regime (CDM).
For the mass varying CDM coupled to dark energy with $p_{\phi} = -\rho_{\phi}$, the GCG leads to similar predictions for $w = p_{\phi} / \rho_{\phi}$, independently of the scale parameter $a$.
The same is not true for HDM as the GCG-like behavior holds just close to the present ($a \sim a_{\mbox{\tiny $0$}}$).
\section{Energy density, neutrino mass and stability}
As mentioned, the evolution of matter components are constrained by several factors.
The precise peak location of the CMB anisotropies, the change in the growth of cosmic structures, and the properties of nonlinear structures which provide detailed information to test such models.
In here a preliminary analysis is performed to compare neutrino masses and the corresponding conditions for stability for three cases of neutrino dynamical masses.
Due to previous arguments, these cases take place in the context of the GCG model.
The first two cases correspond to a neutrino mass dependence on the {\em cosmon} field.
For the first case, the off-diagonal mass matrix elements, the Dirac mass terms, $m\bb{\phi}$, have a linear dependence on $\phi$, $m\bb{\phi}\sim \phi$, which results in a tiny mass eigenvalue $\mu\bb{\phi} = \phi \exp{\left[-3\bb{\alpha + 1}\phi/2\right]}$.
For the second case, the off-diagonal mass matrix elements are constants, $m\bb{\phi}\sim const.$, and the dynamical masses are generated by the dependence on the Majorana mass term $M\bb{\phi}$, which results in $\mu\bb{\phi} = \exp{\left[-3\bb{\alpha + 1}\phi/2\right]}$.
The third case describes a sterile neutrino with a dynamical mass evolving like the GCG with mass varying dark matter coupled to dark energy with a equation of state, $p_{\mbox{\tiny $D$}\mbox{\tiny $E$}} = - \rho_{\mbox{\tiny $D$}\mbox{\tiny $E$}}$.
Since the prescription for decoupling mass varying dark matter from the effective GCG leads to sterile neutrino mass given by Eq.~(\ref{pap11}), the active neutrino mass should be given by $\mu\bb{\phi} = \tanh{\left[3\bb{\alpha + 1}\phi/2\right]}^{\frac{\mbox{\tiny $2$}\alpha}{\alpha + 1}} \exp{\left[-3\bb{\alpha + 1}\phi\right]}$.
Keeping in mind the connection established by the {\em seesaw} mechanism for the third case, thus the mass matrix term should be given by
$m\bb{\phi} = \tanh{\left[3\bb{\alpha + 1}\phi/2\right]}^{\frac{\mbox{\tiny $2$}\alpha}{\alpha + 1}} \exp{\left[-3\bb{\alpha + 1}\phi/2\right]}$.
Considering that the cosmological background evolves according to the GCG model, Fig.~\ref{AAA} exhibits the behavior of the dark sector for the three considered cases.
The evolving neutrino masses lead to evolution of the neutrino energy densities which are shown in Fig~\ref{BBB}, where active and sterile neutrino energy density ratios, $\rho_{\mbox{\tiny $A$}}/\rho{\mbox{\tiny $S$}}$, are plotted as function of the scale factor $a$.
The sterile neutrino is assumed to behave like dark matter.
Notice that different thermodynamic regimes are considered at present: NR, relativistic and UR.
As expected, once the active neutrino reaches the NR regime, the masses and densities are not independent and follow the exponential behavior given by $\rho_{\mbox{\tiny $A$}}/\rho_{\mbox{\tiny $S$}} \propto s^{\mbox{\tiny $2$}} = \exp{\left[3\bb{\alpha + 1}\phi\right]}$.
The equation of state for active neutrinos in NR regime at present is shown in Fig.~\ref{CCC}.
Despite the explicit model dependence of the masses on the scalar field, Figs.~\ref{BBB} and \ref{CCC} show similar global features that are typical of the proposed framework.
The resulting mass dependence on the scale factor for the three considered cases is depicted in Fig.~\ref{DDD}.
In the three cases one sees a sharp increase of the neutrino masses at a recent past ($a \gtrsim 0.2$).
Our analysis shows that the active neutrino mass component grows when coupled to dark matter through the {\em seesaw cosmon} field within the GCG model.
This is consistent with the treatment of neutrinos as a test fluid, in the sense that its cosmological equation of motion decouples from the dark sector which governs the cosmological evolution.
In addition, it provides the conditions to analyze instabilities.
The possibility of adiabatic instabilities in cosmological scenarios was previously pointed out in the context of a mass varying neutrino model of dark energy whereas the dynamical dark energy model is obtained through the coupling of a light scalar field to neutrinos, but not to dark matter \cite{Afs05}.
The ensued effects have been extensively discussed in the context of mass varying neutrinos, in which the growth of the neutrino's mass and the recent accelerated expansion are linked through the scalar field coupling.
In the adiabatic regime, these models face catastrophic instabilities on small scales, since the sound speed squared of the coupled fluid is negative.
Starting with a uniform fluid, such instabilities would give rise to exponential growth of small perturbations.
The natural interpretation of this is that the universe becomes inhomogeneous with neutrino overdensities subject to nonlinear fluctuations which eventually collapse into compact localized regions \cite{Mot08}.
Our approach allows one to perform a stability analysis for NR active neutrinos independently of the dark sector evolution.
The first step is obtaining the squared speed of sound, $c^{\mbox{\tiny $2$}}_{s}$, for neutrinos.
Fig.~\ref{EEE} illustrates the results for $c^{\mbox{\tiny $2$}}_{s}$ for DFG of neutrinos ($\nu$), the {\em seesaw cosmon} dark energy and the GCG plus neutrinos.
The results are compared with those of a GCG scenario with $A_{\mbox{\tiny $s$}}= 4/5$ for $\alpha = 1$ (solid line), $1/2$ (dot line), $1/4$ (dash dot line), and $1/8$ (dash line).
The role of neutrinos on the positiveness of $c^{\mbox{\tiny $2$}}_{s}$ is measured in terms of those for the GCG scenario.
One sees that the influence of neutrinos on the positiveness of $c^{\mbox{\tiny $2$}}_{s}$ for the GCG is not relevant.
Actually for the first two cases, NR neutrinos lead to vanishing perturbations on the background at present, so that the test fluid approach is quite accurate.
The results obtained from the analysis of the third case show that for the models where the stationary condition (cf. Eq.~(\ref{pap09})) implies that, $ p_{\phi} = - \rho_{\phi}$, one obtains $c_{s}^{\mbox{\tiny $2$}} = -1$ from the very start.
The effective GCG plus neutrinos is free from this inconsistency, i. e. the coupling of the dark energy component with mass varying dark matter allows for removing such inconsistency as $c_{s}^{\mbox{\tiny $2$}} \simeq d p_{\phi}/ d\rho_{\phi} > 0$.
\section{Conclusions}
In this work it is shown that GCG contains some key ingredients in the description of the universe dynamics and that it allows for a suitable background for the mass varying and the {\em seesaw} mechanisms.
Furthermore it has been pointed out that the mass varying behavior of the dark matter component can be matched with the GCG model.
The main features of the proposed model can be summarized as follows:
i) The cosmological evolution of neutrino energy densities is driven by coupled {\em cosmon}-type field equations where active and sterile neutrino states are connected through the {\em seesaw} mechanism for mass generation.
ii) Dark matter is, most often, not considered in the MaVaN models.
The treatment of dark energy and dark matter in the GCG unified scheme naturally offers this possibility.
Identifying sterile neutrinos as dark matter coupled with dark energy provides the conditions to implement such unified picture in the MaVaN formulation.
Moreover, the constraints imposed by the {\em seesaw} mechanism in order to establish a unique analytical connection to the GCG in terms of a real scalar field were found.
iii) The dynamics of the coupled fluid composed by neutrinos, dark matter and dark energy is driven by one single degree of freedom, the scalar field, $\phi\bb{a}$.
Since the GCG allows for an explicit dependence on the scale factor, $a$, $s$ and $\xi$ dependence on the scalar field do imply a dependence on the universe's evolution.
Due to the connection between the GCG and the {\em seesaw} masses, the proposed approach actually yields an effective model for MaVaN's coupling to the dark sector.
Of course, one can interpret the mediation of the scalar field as a dependence on the dimensionless scales $a/a_{\mbox{\tiny $0$}}$, $s$, and $\xi/M$.
The described mechanism leads to a fast increase of the neutrino masses, which results in a model dependent vanishing speed of sound at present.
The dynamical mass behavior, due to the evolution of the {\em seesaw cosmon} field, coincides in a subtle way with the GCG dynamics.
However, this scenario is not unique.
Nevertheless, without any additional assumptions, it provides an attractive solution for the coincidence problem through the confluence of three independent frameworks: the {\em cosmon}-like dynamics, the {\em seesaw} mechanism for mass generation and the GCG model.
For active neutrinos, an increase of the mass $\mu$ by a factor $10^{\mbox{\tiny $6$}}$, corresponds approximately to a decrease of the sterile neutrino mass $\mathcal{M}$ from the Planck scale to $10^{\mbox{\tiny $1$}\mbox{\tiny $3$}}\, GeV$.
Thus, the GCG model modulates the increase of the neutrino mass, which acts as a cosmological clock to the present scenario of cosmological accelerated expansion.
Unfortunately, in general, one cannot provide a sharp criterion for the potential and for the mass varying dependence on the scalar field to discriminate between the GCG scenario and a unified fluid via a {\em seesaw cosmon} field that mimics the GCG.
Furthermore, it has been shown for specific potentials that, in what concerns stability and cosmic acceleration, many results found in the literature can be recovered.
Actually, new scenarios featuring other mass dependencies on the {\em seesaw cosmon} field can also be considered.
To conclude, the proposed unified scheme, despite similarities with some quintessence models, is an ambitious and encompassing scheme where mass varying particles can be related with the stability and the cosmic acceleration of the universe.
\begin{acknowledgments}
A. E. B. would like to thank for the financial support from the Brazilian Agencies FAPESP (grant 08/50671-0) and CNPq (grant 300627/2007-6).
\end{acknowledgments}
\section*{Appendix: The mass varying mechanism for a DFG}
Aiming to understand how the mass varying behavior takes place, it is assumed that for a particle statistical distribution $f\bb{q}$, where $q \equiv \frac{|\mbox{\boldmath$p$}|}{T_{\mbox{\tiny $0$}}}$, $T_{\mbox{\tiny $0$}}$ being the background temperature at present, in a flat FRW cosmological scenario.
For a generic dynamical mass defined by $\mbox{\textsf{M}}\bb{\phi}$, the corresponding particle density, energy density and pressure are expressed as
\begin{eqnarray}
n\bb{a} &=&\frac{T^{\mbox{\tiny $3$}}_{\mbox{\tiny $0$}}}{\pi^{\mbox{\tiny $2$}}\,a^{\mbox{\tiny $3$}}}
\int_{_{0}}^{^{\infty}}{\hspace{-0.3cm}dq\,q^{\mbox{\tiny $2$}}\ \hspace{-0.1cm}f\bb{q}},\nonumber\\
\rho\bb{a, \phi} &=&\frac{T^{\mbox{\tiny $4$}}_{\mbox{\tiny $0$}}}{\pi^{\mbox{\tiny $2$}}\,a^{\mbox{\tiny $4$}}}
\int_{_{0}}^{^{\infty}}{\hspace{-0.3cm}dq\,q^{\mbox{\tiny $2$}}\, \left(q^{\mbox{\tiny $2$}}+\frac{\mbox{\textsf{M}}^{\mbox{\tiny $2$}}\bb{\phi}\,a^{\mbox{\tiny $2$}}}{T^{\mbox{\tiny $2$}}_{\mbox{\tiny $0$}}}\right)^{\mbox{\tiny $1$}/\mbox{\tiny $2$}}\hspace{-0.1cm}f\bb{q}},\\
p\bb{a, \phi} &=&\frac{T^{\mbox{\tiny $4$}}_{\mbox{\tiny $0$}}}{3\pi^{\mbox{\tiny $2$}}\,a^{\mbox{\tiny $4$}}}\int_{_{0}}^{^{\infty}}{\hspace{-0.3cm}dq\,q^{\mbox{\tiny $4$}}\, \left(q^{\mbox{\tiny $2$}}+\frac{\mbox{\textsf{M}}^{\mbox{\tiny $2$}}\bb{\phi}\,a^{\mbox{\tiny $2$}}}{T^{\mbox{\tiny $2$}}_{\mbox{\tiny $0$}}}\right)^{\mbox{\tiny $-$}\mbox{\tiny $1$}/\mbox{\tiny $2$}}\hspace{-0.1cm} f\bb{q}},~~~~ \nonumber
\label{gcg01}
\end{eqnarray}
where one has introduced the sub-index $0$ for denoting present-day values, with $a_{\mbox{\tiny $0$}} = 1$.
In the limit where $T$ tends to $0$, the Fermi distribution $f\bb{q}$ becomes a step function that yields an elementary integral with the upper limit equal to the Fermi momentum $q_{F}$.
The energy density and pressure of a DFG can be expressed in terms of elementary functions of the scale factor, $\beta \equiv\beta\bb{a} = T_{\mbox{\tiny $0$}}\,q_{F}/a$ and $\mbox{\textsf{M}} \equiv \mbox{\textsf{M}}\bb{\phi\bb{a}}$,
\begin{eqnarray}
n\bb{a} &=&\frac{1}{3 \pi^{\mbox{\tiny $2$}}}\beta^{\mbox{\tiny $3$}},\nonumber\\
\rho\bb{a} &=& \frac{1}{(8 \pi^{\mbox{\tiny $2$}})}
\left[\beta (2 \beta^{\mbox{\tiny $2$}} + \mbox{\textsf{M}}^{\mbox{\tiny $2$}})\sqrt{\beta^{\mbox{\tiny $2$}} + \mbox{\textsf{M}}^{\mbox{\tiny $2$}}} -
\mbox{\textsf{M}}^{\mbox{\tiny $4$}}\,arc\sinh{\left(\beta/\mbox{\textsf{M}}\right)}\right],\\
p\bb{a} &=&\frac{1}{(8 \pi^{\mbox{\tiny $2$}})}
\left[\beta (\frac{2}{3} \beta^{\mbox{\tiny $2$}} - \mbox{\textsf{M}}^{\mbox{\tiny $2$}})\sqrt{\beta^{\mbox{\tiny $2$}} + \mbox{\textsf{M}}^{\mbox{\tiny $2$}}} + \mbox{\textsf{M}}^{\mbox{\tiny $4$}}\,arc\sinh{\left(\beta/\mbox{\textsf{M}}\right)}\right].\nonumber
\label{gcg01B}
\end{eqnarray}
Simple mathematical manipulations allow one to show that
\begin{equation}
n\bb{a} \frac{\partial \rho\bb{a}}{\partial n\bb{a}} = (\rho\bb{a} + p\bb{a}),
\label{gcg02B}
\end{equation}
and
\begin{equation}
\mbox{\textsf{M}}\bb{a} \frac{\partial \rho\bb{a}}{\partial \mbox{\textsf{M}}\bb{a}} = (\rho\bb{a} - 3 p\bb{a}),
\label{gcg02}
\end{equation}
from which one can obtain the energy-momentum conservation equation
\begin{equation}
\dot{\rho} + 3 H (\rho + p) - \dot{\phi}\frac{\mbox{d} \mbox{\textsf{M}}}{\mbox{d} \phi} \frac{\partial \rho}{\partial \mbox{\textsf{M}}} = 0.
\label{gcg03}
\end{equation}
The coupling between relic particles and the scalar field as described by Eq.~(\ref{gcg02}) is relevant only in the NR regime since the strength of the coupling is suppressed by the pressure of the relativistic ($T\bb{a} = T_{\mbox{\tiny $0$}}/a >> \mbox{\textsf{M}}\bb{\phi\bb{a}}$) particles, so that matter and scalar field fluids tend to decouple and evolve adiabatically.
|
\section{Introduction}\label{intro_scn}
For more than two millenia, the analysis of form of biological entities has been an enticing object of human occupation. While early work tended to be speculative in nature, the current state of mathematics and computational power allow to develop and simultaneously verify theoretical results, thereby increasingly driving scientific progress as witnessed today.
This present work has been motivated by joint research with the Institute for Forest Biometry and Informatics at the University of G\"ottingen, to compare leaf growth dynamics within single specimen, species and taxa for identification of gene expression. The endeavor is challenging as it touches problems at least as old as Theophrastus' (371 -- 287 b.C.) famous book on ``plant growth'', cf. \cite{Th76}.
The first step of this project is the subject of this work: to develop a framework allowing to compare shape dynamics. To this end, we model biological growth by (generalized) geodesics in shape space. We do so because the geometry of shape spaces in which travel along geodesics requires no energy seems linked to the physiological reality of growth preferring to minimize energy. This ``geodesic hypothesis'', originally stated by \cite{LK00}) is further supported by earlier research, cf. \cite{HZ06}; \cite{HHGMS07}. As with the ``geodesic hypothesis'' one can say that \emph{geodesic shape deformation} of two different shapes is the ``same'' if the impetus of the first deformation is transplanted to the second with no loss of energy. In the language of Riemannian geometry this translates to the condition that the initial velocity of the second geodesics is the \emph{parallel transport} of the initial velocity of the first geodesic. If the deformations are not the same, i.e. the geodesics are not parallel at the first and the second shape, this concept gives a correlation-based distance between the deformations.
In consequence, the aim of this paper is to provide for parallel transport on shape spacs. Recall that most shape spaces can be viewed as Riemannian immersions or submersions or, combinations thereof.
Explicit formulae for parallel transport are only available for special spaces. e.g. for spheres and Kendall's spaces of planar shapes, cf. \cite{HHM09}. In general, parallel transport may be difficult to compute and be only available numerically. In the following Section \ref{Para-scn} we provide for a general method to compute parallel transport on shape spaces. In view of our application the method is illustrated in Section \ref{ZR-scn} for the spaces of closed 2D contours based on angular direction with and without specific initial point (cf. \cite{ZR72} as well as \cite{KSMJ04}), and in Section \ref{Kendall_scn} for Kendall's landmark based shape spaces (e.g. \cite{DM98}).
In Section \ref{k-rot-inv:scn}, we compare parallel transport on the spaces of closed contours with parallel transport on Kendall shape spaces for simple regular polygonal configurations. While all sectional curvatures in Kendall's shape space are bounded from below by $1$, it turns out that the corresponding subspace of closed contours is flat.
Finally in Section \ref{App:scn}, leaf growth of one leaf is transported parallelly to other leaves and both shape evolutions are compared with one another. For a specific Canadian black poplar investigated we find that leaves with initially and terminally different shapes tend to evolve parallel, in particular so if no shape anomalies are present. Thus the \emph{geodesic hypothesis}
can be extended to the \emph{parallel hypothesis}:
\begin{center}{\it
biological growth of related objects, possibly of initially and terminally different shape, tends to follow parallel geodesics,}
\end{center}
Using Euclidean approximations in landmark based shape spaces rather than geodesics, this hypothesis was originally coined by \cite{MKMA00} who observed parallel growth patterns.
Readers primarily interested in the application can directly skip to Section \ref{App:scn}.
\section{Parallel Transport (PT)}\label{Para-scn}
This section begins with a review of basic concepts of Riemannian geometry found in any standard textbook (specifically \cite{L99} is very appropriate for the following), in particular formulae relating covariant derivatives of Riemannian immersions and submersions. These provide differential equations lifting the parallel transport on shape space to Euclidean or Hilbert space.
For a Riemannian manifold $M$, possibly of countable dimension denote by $\langle V_p,W_p\rangle^M$ the \emph{Riemannian metric}
of tangent spaces
and by $\nabla^M_VW$ the \emph{covariant derivative} of vector-fields. Here $V,W\in T(M)$ denote vector-fields with values $V_p,W_p$ in the tangent space $T_pM$ of $M$ at $p\in M$. $d^M(p,p')$ is the induced metrical distance on $M$ for $p,p'\in M$, $V\otimes W$ denotes the \emph{outer product} defined by $(V\otimes W)\,X = \langle X,V\rangle\,W$. A vector-field $W\in T(M)$ is \emph{parallel} along a smooth curve $t\to \gamma(t)$ on $M$ if it satisfies the ordinary differential equation (ODE)
\begin{eqnarray}\label{parallel_eq}
\nabla_{\dot{\gamma}}W &=& 0\,.
\end{eqnarray}
It is well known that there is locally a unique solution $W$ along $\gamma$ for a given initial value. In Euclidean or Hilbert space the left hand side
has the simple form (\ref{Eucl_nabl_eq}).
In particular, geodesics are characterized by the fact that their velocity is parallel:
$\nabla_{\dot{\gamma}}\dot{\gamma} = 0\,.$
The covariant derivative is often called a \emph{covariant connection}. Indeed, if two offsets $p,p'\in M$ can be joined by a unique geodesic segment of minimal length, their respective tangent spaces are \emph{connected} via \emph{parallel transport} (PT).
\begin{Def}\label{parallel_transp_def}
$w' \in T_{p'}M$ is the \emph{parallel transplant}
of $w\in T_pM$ if there are
\begin{enumerate}
\item a unique unit speed geodesic $t\to \gamma(t)$ connecting $p=\gamma(0)$ with $p'=\gamma\big(d^M(p,p')\big)$, and
\item a vector field $W\in T(M)$ parallel along $\gamma$ with $W_p=w, W_{p'}=w'$.
\end{enumerate}
\end{Def}
A sufficient condition for the existence of such a unique connecting geodesic is that $M$ is finite dimensional and $p'$ is sufficiently close to $p$. In case of infinite dimension, examples of complete spaces can be constructed which do not feature minimizing geodesics between arbitrary close points (e.g. \citet[pp.226/7]{L99}). For our applications in mind this fact seems less troublesome since infinite dimensional spaces considered here are built from projective limits of finite dimensional spaces.
The Euclidean and Hilbert spaces $\mathbb R^n$ (for Hilbert space $n=\infty$) can be identified with all of their tangent spaces, i.e. $\langle v,w\rangle^{\mathbb R^n} = \sum_{i=1}^n v^iw^i$ and the covariant derivative is just the usual multivariate derivative by components,
$$\nabla^{\mathbb R^n}_{(v^1,\ldots,v^n)}(w^1,\ldots,w^n) = \sum_{i=1}^nv^i\left(\frac{\partial w^1}{\partial x^i},\ldots, \frac{\partial w^n}{\partial x^i}\right)\,.$$
In particular, if $v=\dot{x}(t)$, i.e. $\frac{dx^i}{dt} = v^i$ we have that
\begin{eqnarray}\label{Eucl_nabl_eq}
\nabla^{\mathbb R^n}_{\dot{x}(t)} W &=& \frac{d~}{dt}\, W_{x(t)}\,.
\end{eqnarray}
Thus as desired, parallel transport on Euclidean and Hilbert spaces is given by affine translations.
For short we write $W(t)$ for the value of $W$ along a selfunderstood smooth curve $t\to\gamma(t)$ and
$$\dot{W}(t) ~:= ~\frac{d~}{dt}\, W_{\gamma(t)}$$
in the Euclidean/Hilbert case.
A surjective linear mapping $f : E\to F$ of topological vector-spaces \emph{splits in $F$} if $\operatorname {kern}(f)$ has a closed complement $\widetilde{F}$ in $E$ such that $ \widetilde{F}\times\operatorname {kern}(f)\cong E$ as topological vector-spaces, in particular, $ \operatorname {kern}(f)\to E\to F$ is a \emph{short exact sequence}. Another wording is that \emph{$F$ splits over $E$}.
A smooth mapping $\Phi : M \to N$ of Riemannian manifolds $M$ and $N$ induces a differential mapping $d\Phi_p:T_pM \to T_{\Phi(p)}N$ of tangent spaces. $\Phi$ is called
\begin{enumerate}
\item[] an \emph{immersion} if $\Phi$ is injective and if every $T_{\Phi(p)}N$ splits over $d\Phi_pT_pM$,
\item[] a \emph{submersion} if $\Phi$ is surjective and if every $d\Phi_p$ splits in $T_{\Phi(p)}N$,
\item[] an \emph{isometry} if
$\langle V_p,W_p\rangle^M = \langle d\Phi_pV_p,d\Phi_pW_p\rangle^N\,,~~\forall p \in M\mbox{ and }V,W \in T(M)\,.$
\end{enumerate}
An isometric immersion (submersion) is a \emph{Riemannian immersion (submersion)} respectively.
\paragraph{Riemannian Immersions.} If $\Phi:M\to N$ is a Riemannian immersion then the tangent spaces of $N$ split into the tangent spaces of $\Phi(M)$ and its orthogonal complements, the \emph{normal spaces}
\begin{eqnarray*
T_{\Phi(p)} N &=& T_{\Phi(p)}\Phi(M) \oplus N_{\Phi(p)}\Phi(M)\,.
\end{eqnarray*}
As a consequence of the implicit function theorem, every Riemannian immersion $\Phi:M\to N$ admits locally an implicit representation $\Psi : U\cap N \to N_{\Phi(p)}\Phi(M)$ such that $\Psi(U\cap M) = U \cap \Phi(N)$. Here $U$ is a suitable neighborhood of $\Phi(p)$ in $N$. Hence, we have with $X,Y\in T(M)$ and arbitrary local extensions $\widetilde{X},\widetilde{Y} \in T(U\cap N)$ of $d\Phi X, d\Phi Y \in TM$ that
\begin{eqnarray}\label{imm-cov-der-eq}
(\operatorname {id}_{T(N)}-d\Psi)\left(\nabla^N_{\widetilde{X}} \widetilde{Y}\right) &=&
d\Phi \left(\nabla^M_X Y\right)\,.
\end{eqnarray}
In particular, $d\Psi_{\Phi(p)}$ spans the normal space $N_{\Phi(p)}\Phi(M)$.
\begin{Th}\label{imm-parallel-th} Suppose that an embedding $id_M:M\hookrightarrow \mathbb R^n$ is a Riemannian immersion in Euclidean ($n<\infty$) or Hilbert space ($n=\infty$), $t\to\gamma(t)$ a geodesic in $M$, $\{V_j(t):j\in J\}$ an orthonormal smooth base for $N_{\gamma(t)}M$ and $W$ a vector-field in $M$. Then $W$ is parallel along $\gamma$ if and only if it satisfies the linear differential equation
\begin{eqnarray*
\dot{W}(t)&=&-\left(\sum_{j\in J} \dot{V}_j(t)\otimes V_j(t)\right)\,W(t)\,.
\end{eqnarray*}
\end{Th}
\begin{proof}
The assertion is an immediate consequence of (\ref{imm-cov-der-eq}) and the fact that
$$0~=~\frac{d}{dt} \big\langle W(t),V_j(t)\big\rangle ~=~ \big\langle \dot{W}(t),V_j(t)\big\rangle + \big\langle W(t),\dot{V}_j(t)\big\rangle$$
for all $j\in J$ by hypothesis.
\end{proof}
\paragraph{Riemannian Submersions}
For a Riemannian submersion $\Phi : M \to Q$ from the \emph{top space} $M$ to the \emph{bottom space} $Q$, tangent spaces split as follows:
every fiber $\Phi^{-1}(q)$, $q\in Q$ is a submanifold of $M$ that is locally a topological embedding. With the \emph{vertical space} $T_p \Phi^{-1}\big(\Phi(p)\big)$ along the fiber and its orthogonal complement, the \emph{horizontal space}, we have
$$T_p M = T_p \Phi^{-1}\big(\Phi(p)\big) \oplus H_pM\,.$$
Since $H_pM \cong T_{\Phi(p)} Q$, every $V\in T(Q)$ has a unique horizontal lift $\widetilde{V} \in H_pM$ characterized by $d\Phi \widetilde{V} = V$. For arbitrary $W\in T(M)$ denote by $W^\perp: p \to W^\perp_p$ the orthogonal projection to the vertical space.
The following Theorem due to \cite{O66} (cf. also \citet[p.386]{L99}) allows to lift bottom space parallel transport to the top space. In addition to (\ref{imm-cov-der-eq}) this provides the vertical (normal) part as well, which is in general non-zero for submersions.
\begin{Th}\label{ONEILL_thm} Let $\Phi:M\to Q$ be a Riemannian submersion and let $X,Y\in T(Q)$. Then we have with the \emph{Lie bracket} $[\cdot,\cdot]$ on $M$ that
\begin{eqnarray*
\nabla^M_{\widetilde{X}} \widetilde{Y} &=&
\widetilde{\nabla^N_X Y}
+ ~\frac{1}{2}~ [\widetilde{X},\widetilde{Y}]^{\perp}\,.
\end{eqnarray*}
\end{Th}
We are now ready for the ODE of parallel transport on a Riemannian immersion followed by a Riemannian submersion.
\begin{Th}\label{parallel_riem_subm:thm}
Suppose that $\Phi_1: M\hookrightarrow \mathbb R^n$ is a Riemannian immersion in Euclidean ($n<\infty$) or Hilbert space $(n=\infty)$, $\Phi_2:M \to Q$ a Riemannian submersion and let $W$ be a vector field on $M$ horizontal along a horizontal geodesic $\gamma(t)$ on $M$. Then $d\Phi_2 W$ is parallel along $\Phi_2\circ \gamma(t)$ if and only if
\begin{eqnarray*
\dot{W}(t)&=&-\left(\sum_{j\in J} \dot{V}_j(t)\otimes V_j(t)\right)\,W(t) - \sum_{k\in K}d\omega^t_k\big(\dot{\gamma}(t),W(t)\big)\,.
\end{eqnarray*}
Here, $\{V_j(t):j\in J\}$ denote an orthonormal smooth base for the normal space $N_{\gamma(t)}M\subset \mathbb R^n$ and $d\omega^t_k$ are the exterior derivatives of an orthonormal and smooth base $\{U_k(t): k \in K\}$ of the vertical space $T_{\gamma(t)}[\gamma(t)] \subset T_{\gamma(t)}M$ for suitable index sets $J$ and $K$.
\end{Th}
\begin{proof}
Suppose that we have a vector field $X\in T(Q)$ with horizontal lift $\widetilde{X}(t) =\dot{\gamma}(t)$. If $\Phi_2\circ \gamma$ is geodesic and $d\Phi_2 W$ parallel with horizontal lift $W$, Theorem \ref{ONEILL_thm} yields
\begin{eqnarray}\label{O'N:eq}
\nabla_{\dot{\gamma}}^MW&=&\frac{1}{2} \sum_{k\in K}\langle [\widetilde{X},W],U_k\rangle]U_k ~=~-\sum_{k\in K} d\omega_k(\dot{\gamma},W)
\end{eqnarray}
making use of the well known (e.g. \citet[p.126/7]{L99})
$$\langle [\widetilde{X},W],U_k\rangle = \widetilde{X}\langle U_k,W\rangle - W\langle U_k,\widetilde{X}\rangle - 2d\omega_k(\widetilde{X},W)$$
with the exterior derivative $d\omega_k$ of the one-form $\omega_k$ dual to $U_k$. On the other hand, since $d\Phi_1:T_pM\to T_pM \subset \mathbb R^n$ is the identity, formula (\ref{imm-cov-der-eq}) yields
\begin{eqnarray}\label{imm:eq}\nonumber
\nabla_{\dot{\gamma}}W^M &=& \dot{W}(t) - \left(\sum_{j\in J} {V}_j(t)\otimes V_j(t)\right)\,\dot{W}(t)\\&=& \dot{W}(t)+\left(\sum_{j\in J} \dot{V}_j(t)\otimes V_j(t)\right)\,W(t)
\end{eqnarray}
as in the proof of Theorem \ref{imm-parallel-th} . Putting together (\ref{O'N:eq}) and (\ref{imm:eq}) gives the assertion of the Theorem.
\end{proof}
\section{PT for Closed 2D Contours}\label{ZR-scn}
We define the two \emph{shape spaces of closed 2D constant-speed contours based on angular direction} as introduced by \cite{ZR72} in the geometric formulation of \cite{KSMJ04}.
Suppose that $z:[0,2\pi] \to \mathbb C, s\mapsto z(s)$
is a constant-speed parameterization of a smooth, closed, curve of length $L$ winding once counterclockwise around each interior point. Let
\begin{eqnarray} \nonumber
\theta(s) &=&\arg\big(z'(s)\big) -\arg\big(z'(0)\big) -s\,,\mbox{ with}\\ \label{ZR-ode}
\dot{z}(s) &=&
\frac{L}{2\pi} \,e^{i\big(\theta(s)+\arg(z'(0)) +s\big)}\,.
\end{eqnarray}
Obviously, the \emph{Zahn-Roskies shape} (ZR-shape) $\theta$ is invariant under translation, scaling and rotation $z(s) \to c + \lambda e^{i\psi} z(s)$. Moreover, subtracting $s$ (the curves to be modelled wind once around their interior) norms $\theta$ such that it is $2\pi$-periodic. Vice versa, from every converging Fourier series
an a.e. differentiable constant-speed 2D curve can be reconstructed by integrating (\ref{ZR-ode}). This curve is unique modulo translation, scaling and rotation. Thus a linear subspace of the Hilbert space $\mbox{\handw \symbol{96}}^2$ of Fourier series is the \emph{ZR--pre--shape space}
\begin{eqnarray*}
S_{ZR} ~:=~\Big\{ \theta(s) = \sum_{n=0}^{\infty} \big(x_n\cos(ns)+y_n\sin(ns)\big):\\
\|\theta\|^2-x_0^2= \frac{1}{2}\,\sum_{n=1}^{\infty}(x_n^2+y_n^2)<\infty\\ x_0 = -\sum_{n=1}^\infty x_n,~y_0=0\Big\}\,.
\end{eqnarray*}
As usual, $2\pi \langle\theta,\eta\rangle := \int_0^{2\pi} \theta(s)\eta(s)\,ds$ and $\|\theta\|^2 := \langle \theta,\theta\rangle$. The tangent spaces $T_{\theta}S_{ZR}$ are identified with $S_{ZR}\subset \mbox{\handw \symbol{96}}^2$.
Since the curves in question are closed, we have with the non-linear mapping
\begin{eqnarray*
\begin{array}{rcl}
\Psi : \mbox{\handw \symbol{96}}^2 &\to& \mathbb C\\
\theta &\mapsto& \int_0^{2\pi} e^{i\big(\theta(s)+s\big)}\,dt
\end{array
\end{eqnarray*}
that the \emph{ZR--shape space} is the implicit sub-manifold
\begin{eqnarray*
\Sigma_{ZR} ~:=~ \left\{ \theta \in S_{ZR}: \Psi(\theta) =0\right\}\,.&&
\end{eqnarray*}
Obviously, the ZR-shapes of closed not self-intersecting contours form an open subset containing the origin, which corresponds to the shape of the circle.
Additionally considering closed curves invariant under change of initial point $z(s) \to z(s+s_0)$ (e.g. amorphous curves with no preassigned initial point) by defining this action of the unit circle $S^1\ni s_0$ on $\Sigma_{ZR}$ the \emph{invariant ZR--shape space}
\begin{eqnarray}\label{I-ZR-shape-def}
\Sigma^{I}_{ZR} &:=&\Big(\Sigma_{ZR}\setminus\{0\}\Big)/S^1\,
\end{eqnarray}
is obtained.
Since rotation and parameter shift are equivalent for circles, the corresponding invariant ZR--shape is thus a singularity of $\Sigma^{I}_{ZR}$, in fact its only singularity.
\paragraph{Parallel Transport on $\Sigma_{ZR}$ and $\Sigma^{I}_{ZR}$}
Geodesics on $\Sigma_{ZR}$ as well as on $\Sigma^{I}_{ZR}$ between two given points can be computed via a technique called \emph{geodesic shooting}, cf. \cite{MTY06} as well as \cite{KSMJ04}, or much faster via a variational approach \cite{SCC06}. Since $\Sigma_{ZR}\hookrightarrow \mbox{\handw \symbol{96}}^2$ is a Riemannian immersion with the global implicit definition $\Psi=0$ we have that the normal space in $S_{ZR}$ at $\theta\in\Sigma_{ZR}$ is spanned by $V_1(\theta)=s\mapsto\cos\left(\theta(s) +s\right)$ and $V_2(\theta)=s\mapsto \sin\big(\theta(s)+s\big)$. Orthogonalization yields the base
$$ W_1 := \frac{V_1}{\|V_1\|},~~W_2 := \frac{V_2 - \langle V_2,W_1\rangle\,W_1}{\|V_2 - \langle V_2,W_1\rangle\,W_1\|}\,.$$
As a consequence of Theorem \ref{imm-parallel-th} we have
\begin{Th}\label{ZR-Sigma-par-thm} A vector-field $W(t)$ in $\Sigma_{ZR}$ is parallel along a geodesic $\gamma$ in $\Sigma_{ZR}$ if and only if it satisfies the linear differential equation
{\footnotesize \begin{eqnarray}\label{parallel-eq-ZR}
\lefteqn{\dot{W}(t)~=~}\\\nonumber &-\Big(\frac{d}{dt}\,W_1\big({\gamma}(t)\big)\otimes W_1\big({\gamma}(t)\big) + \frac{d}{dt}\,W_2\big({\gamma}(t)\big)\otimes W_2\big({\gamma}(t))\Big) ~{W}(t)\,.
\end{eqnarray}}
\end{Th}
In practice, (\ref{parallel-eq-ZR}) can be solved numerically by orthogonally projecting to $T_{\gamma(t)}\Sigma_{ZR}$ in every iteration step.
We now turn to the submersion (\ref{I-ZR-shape-def}). The vertical space at
$$\theta(s) = x_0+\sum_{n=1}^{\infty} \big(x_n \cos(ns) + y_n \sin(ns)\big) \in \Sigma_{ZR}$$
is spanned (if convergent) by the single vertical unit length direction
$$\frac{\theta'(s)}{\|\theta'(s)\|} = \frac{\sum_{n=1}^{\infty} n\big(-x_n \partial_{y_n} + y_n \partial_{x_n}\big)}{\sqrt{\sum_{n=1}^{\infty} n^2(x_n^2+y_n^2)}}~\sqrt{2}\,.$$
The exterior derivative of its dual is hence
\begin{eqnarray}\label{ext-der-omeg-ZR}
\left.\begin{array}{rcl}
\frac{d\omega}{\sqrt{2}}&=& -2\,\frac{\sum_{n=1}^{\infty}ndx^n \wedge dy^n}{\sqrt{\sum_{n=1}^{\infty} n^2(x_n^2+y_n^2)}}\,\\&&~~ -\,\frac{\sum_{n\neq n'}y_nx_{n'}n n'(n'dx^{n'} \wedge dx^{n}-ndy^n \wedge dy^{n'})}{\sqrt{\sum_{n=1}^{\infty} n^2(x_n^2+y_n^2)}^{~3}}\\
&&-\,\frac{\sum_{n,n'}nn'(ny_ny_{n'}+n'x_nx_{n'}) dy^{n} \wedge dx^{n'}}{\sqrt{\sum_{n=1}^{\infty} n^2(x_n^2+y_n^2)}^{~3}}\end{array}\right\}\,.
\end{eqnarray}
In conjunction with Theorem \ref{imm-parallel-th}, Theorem \ref{parallel_riem_subm:thm} and Theorem \ref{ZR-Sigma-par-thm} one obtains after a tedious computation
\begin{Th}\label{ZR1-Sigma-par-thm}
The vector-field
$$ W(t) = u_0(t)\partial_{x_0}+\sum_{n=1}^\infty \big(u_n(t)\partial_{x_n} + v_n(t)\partial_{y_n}\big)$$
is a horizontal lift to the top space $\Sigma_{ZR}$ of the bottom space parallel transport along a geodesic in $\Sigma^{I}_{ZR}$
$$\gamma_s(t) = x_0(t)+\sum_{n=1}^\infty \big(x_n(t)\cos(ns) + y_n(t)\sin(ns)\big)$$
horizontal in $\Sigma_{ZR}$ if and only if it satisfies the linear differential equation
{\footnotesize\begin{eqnarray*
\lefteqn{\dot{W}(t)~=~}\\
&&-\Big(\frac{d}{dt}\,W_1\big({\gamma}_s(t)\big)\otimes W_1\big({\gamma}_s(t)\big) + \frac{d}{dt}\,W_2\big({\gamma}_s(t)\big)\otimes W_2\big({\gamma}_s(t))\Big) ~{W}(t)
\\
&&+\,\frac{\gamma'_s(t)}{2\|\gamma'_s(t)\|^4}\Big(\big\langle \gamma'_s(t)\otimes\gamma''_s(t), W(t)\otimes \dot{\gamma}_s(t)\big\rangle\\%\nonumber
&&\hspace{3cm} - \big\langle \gamma'_s(t)\otimes\dot{\gamma}_s(t), W(t)\otimes {\gamma}''_s(t)\big\rangle\Big)\\
&&-\,\frac{\gamma'_s(t)}{\|\gamma'_s(t)\|^2} \big\langle \dot{\gamma}'_s(t),W(t)\rangle
\,.
\end{eqnarray*}}with the derivatives defined as
$$\begin{array}{rcl}
\gamma'_s(t) &=& \sum_{n=1}^{\infty} n\big(-x_n(t) \partial_{y_n} + y_n(t) \partial_{x_n}\big)\,,\\ \dot{\gamma}_s(t) &=& \dot{x}_0(t)\partial_{x_0} + \sum_{n=1}^{\infty} \big(\dot{x}_n(t) \partial_{x_n} + \dot{y}_n(t) \partial_{y_n}\big)\,,\\
\dot{\gamma}'_s(t)&=& \sum_{n=1}^{\infty} n\big(-\dot{x}_n(t) \partial_{y_n} + \dot{y}_n(t)
\partial_{x_n}\big)\,,\\
\gamma''_s(t) &=& -\sum_{n=1}^{\infty} n^2\big(x_n(t) \partial_{x_n} + y_n(t) \partial_{y_n}\big)\,,
\end{array}$$
if convergent, and the inner product defined by
$$\langle E_i\otimes E_j,E_k\otimes E_l\rangle ~:=~ \delta_{(i,j),(k,l)}$$
for an orthogonal system $E_j$ and index set $J\ni j$.
\end{Th}
In practice, convergence of the series for the derivates is not an issue since as remarked earlier, computations are carried out using only finitely many Fourier coefficients.
\section{PT for Kendall's Shape Spaces}\label{Kendall_scn}
Kendall's landmark based similarity shape analysis is based on \emph{configurations} consisting of $k\geq m+1$ labelled vertices
in $\mathbb R^m$ called \emph{landmarks} that do not all
coincide. A configuration
$$x = (x^1,\ldots,x^k) = (x^{ij})_{1\leq i\leq m,1\leq j\leq k}$$
is thus an element of the space $M(m,k)$ of matrices with $k$ columns, each an $m$-dimensional landmark vector. Disregarding center and size, these configurations are mapped to the \emph{pre-shape sphere}
$$ M = S^{k}_m := \{p \in M(m,k-1) \colon \|p\| = 1\}\,,$$
where $\|p\|^2 = \langle p, p \rangle $ and $ \langle p, v \rangle := \operatorname {tr}(pv^T)$ is the standard Euclidean product.
This can
be done by, say, multiplying by a sub-Helmert matrix,
cf. \cite{DM98} for a detailed discussion of this and other
normalization methods. The canonical Riemannian immersion $S_m^k\hookrightarrow M(m,k-1)$ comes with a global implicit definition $\Psi(x) = \|x\|-1 =0$ giving rise to a single normal field. Hence, using (\ref{Eucl_nabl_eq}) and (\ref{imm-cov-der-eq}), the covariant derivatives along a curve $\gamma$ on $S_m^k$ relate as
\begin{eqnarray}\label{spher-cov-der}
\nabla^{S_m^k}_{\dot{\gamma}(t)} W(t) &=& \dot{W}(t) - \big\langle \dot{W}(t),\gamma(t)\big\rangle \, \gamma(t)\,
\end{eqnarray}
As a consequence of (\ref{spher-cov-der}) and (\ref{parallel_eq}), unit-speed geodesics on $S_m^k$ are great circles of form $\gamma(t) = x\cos t + v\sin t$ with $x,v\in S_m^k$ and $\langle x,v\rangle=0$.
In order to filter out rotation information define the regular part $(S^{k}_m)^* := \{x\in S_k^k: \operatorname {rank}(x) > m-2\}$ (an open dense subset of $S_m^k$) and a
smooth and free action of $SO(m)$ by the usual matrix multiplication
$(S^{k}_m)^* \stackrel{g}{\to} (S^{k}_m)^* : p \mapsto gp$
for $g \in SO(m)$. The
orbit
$\pi(p) = \{gp \colon g \in SO(m)\}$ is the \emph{Kendall shape} of $p \in
S^{k}_m$ and the quotient
\begin{eqnarray}\label{def-shape-space}
\pi : (S^{k}_m)^* ~\to~ (\Sigma^{k}_m)^* &:=& (S^{k}_m)^*/SO(m)
\end{eqnarray}
is called \emph{Kendall's shape space}. Note that projecting from the entire pre-shape sphere $S_m^k$ would have led to a non-manifold quotient ($m\geq 3$), which is usually called Kendall's shape space. We consider here only the regular part such that (\ref{def-shape-space}) is a Riemannian submersion. With the orthogonal decomposition $\frak{gl}(m) = \frak{o}(m) \oplus SM(m)$ of the Lie algebra $\frak{gl}(m) = M(m,m)$, the Lie algebra $\frak{o}(m)$ of skew-symmetric matrices in $\frak{gl}(m)$ and the vector-space of symmetric matrices $SM(m)$ in $\frak{gl}(m)$ we have the following orthogonal tangent space decomposition for $x\in S_m^k$,
cf. \citet[p.109]{KBCL99}.
\begin{eqnarray}\label{soms}
&\begin{array}{rcccc}
\frak{gl}(m) &=& {\frak o}(m) &\oplus& SM(m)\\
& & \downarrow \cdot x &&\uparrow \cdot x^T\\
T_xS^k_m\oplus N_x S_m^k&=&T_x\pi(x)&
\oplus &\overbrace{ H_xS^k_m\oplus N_x S_m^k}
\end{array}
\end{eqnarray}
For $x\in (S_m^k)^*$ both mappings are surjective and $H_xS_m^k = T_{\pi(x)}(\Sigma_m^k)^*$.
In order to compute the horizontal lift of bottom space parallel transport as in Theorem \ref{ZR1-Sigma-par-thm}, we need an orthonormal base for the $\big(m(m-1)/2\big)$-dimensional vertical space $T_x\pi(x)$ and the exterior derivative of its duals. From (\ref{soms}) we have at once a (in general not-orthogonal) base $\{e_{ij}x : 1\leq i<j\leq m\}$ with base system
{\footnotesize $$e_{ij} = \left(\varepsilon^{\alpha,\beta}\right)_{1\leq \alpha, \beta\leq m}\mbox{ with } \varepsilon^{\alpha,\beta} =\left\{\begin{array}{lcl}1&\mbox{ for } \alpha = i,\beta=j\\
-1 &\mbox{ for } \alpha = j,\beta=i\\
0&\mbox{ else }\end{array}\right.$$}of $\frak{o}(m)$. From the former obtain an o.g. base system $\{V_{ij}(x): 1\leq i<j\leq m\}$ of $T_x\pi(x)$ through Gram-Schmidt orthogonalization and let $\omega_{ij}(x)$ be the one-form dual to $V_{ij}(x)$ ($1\leq i<j\leq m$). For $m=2$ there is a single vertical unit-direction $V_{12}(x)$. For $m>2$, however, $d\omega_{12}(x)$ is about as complicated as (\ref{ext-der-omeg-ZR}), and even for $m=3$, the other two derivatives and their application to vector-fields result in expressions too lengthy to be written down, cf. also the rather complicated examples
in \cite{L03}. Using a computer algebra program, however, these expressions and their respective values can be easily computed by symbolic differentiation. Hence,
Theorem \ref{parallel_riem_subm:thm}
and (\ref{spher-cov-der})
yield the following Theorem. The special case $m=2$ is taken from \citet[Theorem A.6]{HHM09}, cf. also \citet[Theorem 2]{L03}.
\begin{Th}\label{Parallel-transp-Kendall}
A vector-field
$ W(t)$
is a horizontal lift to the top space of the bottom space $(\Sigma^k_m)^*$ parallel transport along a geodesic
$\gamma(t) = x\cos t +v\sin t$
horizontal in $(S_m^k)^*$ if and only if it satisfies the ODE
\begin{eqnarray*}
\lefteqn{\dot{W}(t) ~=~
\left\langle \dot{W}(t), \gamma(t)\right\rangle\, \gamma(t)}
\\ &&\hspace{0cm}-\sum_{1\leq i<j\leq m}
d\omega_{ij}\big(\gamma(t)\big)\Big(\dot{\gamma}(t),W(t)\Big)\, V_{ij}\big(\gamma(t)\big)\,.
\end{eqnarray*}
For $m=2$ this ODE has the explicit solution
\begin{eqnarray*} W(t) ~=~ W(0) - \Big(\big\langle W(0),v\big\rangle\, v + \big\langle W(0),e_{12}v\big\rangle\, e_{12}v\Big) \\+ \big\langle W(0),v\big\rangle\, \dot{\gamma}_{z,v}(t) + \big\langle W(0),e_{12}v\big\rangle\, e_{1,2}\dot{\gamma}_{z,v}(t)\,.\end{eqnarray*}
\end{Th}
\begin{figure}[h!]
\begin{minipage}[t]{0.45\textwidth}
\subfigure[{\it From rectangle $\sigma_1$ to the regular hexagon $\sigma_3$.}]{\label{Def_R1_H_ZR_fig}
\includegraphics[angle=-90,width=0.8\textwidth]{Rect2HexAlign.eps}}
\subfigure[{\it From rectangle $\sigma_1$ to rectangle $\sigma_2$.}]{\label{Def_R1_R2_ZR_fig}
\includegraphics[angle=-90,width=0.8\textwidth]{Rect2RectAlign.eps}}
\subfigure[{\it From $\sigma_2$ to the parallel transplant of $\sigma_3$ from Figure \ref{Def_R1_H_ZR_fig}.}]{\label{Def_R2H_para_ZR_fig}
\includegraphics[angle=-90,width=0.8\textwidth]{Rect2Hex_paraAlign.eps}}
\caption{\it Equidistant deformation along geodesics in $\Sigma_{ZR^2}$. The bullet marks the pre-assigned initial point. Figure \ref{Def_R2H_para_ZR_fig} depicts the parallel transplant to $\sigma_2$ of the geodesic from Figure \ref{Def_R1_H_ZR_fig} along the geodesic depicted in Figure \ref{Def_R1_R2_ZR_fig}.\label{ZR_Hex_fig}}
\end{minipage}
\begin{minipage}[t]{0.05\textwidth}\hfill\end{minipage}
\begin{minipage}[t]{0.45\textwidth}
\subfigure[{\it From rectangle $\sigma_1$ to the regular hexagon $\sigma_3$.}]{\label{Def_R1_H_K_fig}
\includegraphics[angle=-90,width=0.8\textwidth]{Rect2HexKend6Align.eps}}
\subfigure[{\it From rectangle $\sigma_1$ to rectangle $\sigma_2$.}]{\label{Def_R1_R2_K_fig}
\includegraphics[angle=-90,width=0.8\textwidth]{Rect2RectKend6Align.eps}}
\subfigure[{\it From $\sigma_2$ to the parallel transplant of $\sigma_3$ from Figure \ref{Def_R1_H_K_fig}.}]{\label{Def_R2H_para_K_fig}
\includegraphics[angle=-90,width=0.8\textwidth]{Rect_para_Kend6Align.eps}}
\caption{\it Equidistant deformation along geodesics in Kendall's landmark based shape space $\Sigma^6_2$. The kinks signify landmarks. Notation as in Figure \ref{ZR_Hex_fig}.\label{K_Hex_fig}}
\end{minipage}
\end{figure}
\section{Curves with Rotational Symmetry}\label{k-rot-inv:scn}
Denote by $\Sigma_{ZR^k}$ the sub-space of closed curves with $k$-fold rotational symmetry. The following is an observation of \cite{ZR72}:
\begin{Th}
$\theta$ represents a closed curve with $k$-fold rotational symmetry $k>1$ if and only if $x_n,y_n =0$ for all $n\not\equiv 0$ mod $k$.
\end{Th}
Since the arithmetic sum of closed curves with $k$-fold rotational symmetry is again of $k$-fold rotational symmetry we have at once:
\begin{Cor} For each $k=2,3\ldots$, $\Sigma_{ZR^k}$ is a flat linear submanifold of $\Sigma_{ZR}$ and parallel transport on $\Sigma_{ZR^k}$ is affine.
\end{Cor}
Within the subspace $\Sigma_{ZR^2}$ of closed curves with two-fold symmetry consider a simple example of three shapes: two rectangles $\sigma_1$ and $\sigma_2$ differing only by their initial point and a hexagon $\sigma_3$. The `same' (i.e. parallel) deformation from $\sigma_1$ to $\sigma_3$ is applied to $\sigma_2$.
Figure \ref{ZR_Hex_fig} illustrates that deformation in the geometry of Zahn-Roskies' shape space, Figure \ref{K_Hex_fig} gives it in the geometry of Kendall's shape space. Comparing the respective Subfigures (a) and (b) over the different geometries shows that the geodesic deformation with fixed initial and terminal shape gives almost identical intermediate shapes.
In the respective Subfigures (c), the difference between `same' shape deformation over the two geometries is hardly notable in the beginning of the deformation. Near the end, however, it becomes notable: number the kinks (the landmarks in $\Sigma_2^6$) counterclockwise from 1 (bullet) to 6 and denote by $(i,j)$ the line connecting the $i$-th kink with the $j$-th kink. Then $(2,5)$ remains parallel to $(6,1)$ and $(3,4)$ in the Kendall geometry whereas in the Zahn-Roskies geometry it turns in direction beyond $(2,3)$ and $(5,6)$.
\section{Parallel Leaf Growth}\label{App:scn}
Let us quickly overview some very recent developments within two millenia of research on plant form. With the application below in mind, we are interested in a flexible and realistic representation of leaf contour shape. Flexibility in this context means that we are looking for a model in which nature not only chooses values of parameters in a pre-defined parameter space but rather the parameter space itself. We caution that such a \emph{non-parametric} model may come at the cost that nature's parameters may not be simply geometrically interpretable
\emph{Parametric} in this sense are the well established models involving allometry, still of interest today: e.g. \cite{Gur92}; \cite{Burton04}, or the superformula of \cite{Gielis03}. Also, models based on landmarks such as \cite{DPS87}; \cite{Jensen90} or \cite{JCM2002}, can be viewed as projecting nature to a pre-specified parameter space by leaving out the parts of the contour between landmarks. Note that parametric models are highly successful e.g. for plant classification, genetic hybrid identification, cf. \cite{JM05}, or in the \emph{Climate Leaf Analysis Multivariate Program (CLAMP)} of \citet{Wolfe93} which is fundamental to paleoclimate and present day climate reconstruction, cf. \cite{EBG00}.
On the other hand, models building on the shape spaces of Zahn and Roskies (cf. Section \ref{ZR-scn}) are \emph{non-parametric}, even though they are not entirely free of constraints: in view of landmark-based shape analysis (cf. Section \ref{Kendall_scn}), restricting to unit speed velocities translates into infinitesimally placed landmarks. The different geometry, however, liberates from the necessity to identify homologous landmarks, by imposing infinitesimal uniform growth. The latter is certainly debatable. Curiously, such non-parametric models introduced as \emph{eigenshape analysis} (building on \cite{Loh83}) have initially stirred controversy because parameters were not simply geometrically interpretable and because with lacking initial point, registration was not satisfactory, cf. \cite{R86}. While the former is precisesly a desired feature,
introducing the geometric concept of the quotient $\Sigma^I_{ZR} = \big(\Sigma_{ZR}\setminus\{0\}\big)/S^1$ by \cite{KSMJ04} settles the latter objection. It seems, however, that the natural non-Euclidean geometry of $\Sigma_{ZR}$ is not fully realized in the community, cf. \cite{Ray92}; \cite{KGS07}; \cite{Hearn09}.
For sake of completeness, even though not practicable for our purpose because of high sensitivity to boundary noise, let us briefly mention a third approach of shape modeling based on the leaf's vein structure. With methods for automated venation extraction available (cf. \cite{FuChi06}), although computationally much more challenging than contour extraction, \cite{LZG09} link vein structure to the concept of shape spaces by \cite{BN78} based on medial skeletons, cf. also \cite{PSSDZ03}. Undoubtedly, modeling the vein structure gives deep insight into physiological, hydraulical and biomechanical aspects of leaf formation. Parameter spaces thus obtained should be closest to nature in the above sense. Current research in venation patterns, however, shows that leaf shape diversification is still poorly understood (e.g. \cite{NPT07}).
Obviously, for our purpose of modeling entire leaf contours while being as non-parametric as possible, the space $\Sigma_{ZR}$ suits ideally. For the problem at hand there is no need for pre-registration as the leaves in question are naturally aligned by petiole (the base point where the stalk enters the blade forming the main leaf vein) and apex (the terminal point of the main vein, usually the leaf tip) location. One could almost equivalently align by petiole location and the initial direction of the main leaf vein.
\begin{figure}[h!]
\includegraphics[angle=-90,width=0.45\textwidth]{be1b8-original-smoothed_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be1b8-original-geod-proj_ni=10.eps}
\caption{\it Shape evolution of a black poplar leaf over two weeks. Left: original contours. Right: contours obtained from projecting to geodesic evolution. \label{ori-geod-be1b8.eps}}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[angle=-90,width=0.45\textwidth]{be1b9-original-smoothed_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be1b9-along-be1b8_s-geod_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be1b3-original-smoothed_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be1b3-along-be1b8_s-geod_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be2b4-original-smoothed_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be2b4-along-be1b8_s-geod_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be2b1-original-smoothed_ni=10.eps}
\includegraphics[angle=-90,width=0.45\textwidth]{be2b1-along-be1b8_s-geod_ni=10.eps}
\caption{\it Shape evolution of black poplar leaves over two weeks. Left column: original contours. Right column: contours obtained from traversing the first contour along the parallel translate of the geodesic from Figure \ref{ori-geod-be1b8.eps}. \label{ori-parallel-tp-be1b8-2.eps}}
\end{figure}
In an application we consider five leaves of a Canadian black poplar tree at an experimental site of the Department of Forest Biometry and Tree Physiology of the University of G\"ottingen. Their contours have been non-destructively extracted over their growing period at 10 approximately evenly spaced days.
For the following computations we have considered the Euclidean $(2N+1)$-dimensional subspace of $\mbox{\handw \symbol{96}}^2$ using Fourier coefficients, $(x_0,x_1,y_1,\ldots,x_N,y_N)$ with $N=100$.
Leaf ``be1b8'' (left image of Figure \ref{ori-geod-be1b8.eps}) exhibits the most regular shape. In concord with earlier observations
of different leaves using landmarks (cf. \cite{HHGMS07}), the temporal evolution occurs almost along the geodesic determined by initial and terminal shape (as depicted in the right image of Figure \ref{ori-geod-be1b8.eps}). The initial direction of this geodesic has been parallely transplanted to the initial shapes of leaves ``be1b9'', ``be1b3'', ``be2b4'' and ``be2b1''. In the right columns of Figure
\ref{ori-parallel-tp-be1b8-2.eps}, the shapes along these new geodesics starting at the corresponding initial shapes have been recorded at the corresponding points in time. The left columns of these figures depict the original temporal shape evolution.
The common shape dynamics displayed by the original leaf contours (left columns of Figures \ref{ori-geod-be1b8.eps} and \ref{ori-parallel-tp-be1b8-2.eps}) seems two-fold. First, an increase of base angle. Second, different growth ratios are not visible at the apex as its angle remains nearly unchanged. Individual effects are non-symmetric and non uniform lateral growth. Also, leaf 'be1b3' develops a notch left, slightly below the apex, for leaf 'be2b1' an original notch also left, slightly below the apex attenuates.
Obviously (right columns of Figures \ref{ori-geod-be1b8.eps} and \ref{ori-parallel-tp-be1b8-2.eps}) leaves ``be1b9'', ``be1b3' and `be2b4'' follow rather closely the parallel transplant of the geodesic of leaf ``be1b8''. Original non-uniform growth is uniformized and, stronger than originally, apexes acuminate. Even though all of their initial and terminal shapes are quite different, one can say that their temporal evolution is rather similar. This seems to be less the case for leaf ``be2b1''. Its observed growth tends to eliminate its initial strong dent at north-west-north while along the transplanted geodesic, this dent remains, causing increased distal growth at the tip. One could argue that in order to restore an original contour defect, natural growth deviates from its ``original'' plan. Certainly, such phenomena deserve future research.
\begin{table}[h!]
\centering
\fbox{\begin{tabular}{r|cccc}b&be1b8&be1b3&be2b4&be2b1\\\hline
$\rho(v_{\rm b},w_{\rm b})$&$ 0.17 $&$ 0.12 $&$ 0.44$&$0.083 $\\
$\mu(v_{\rm b},w_{\rm b})$&$0.99 $&$ 0.96 $&$1.0$&$0.88 $
\end{tabular}}
\caption{\it Measuring parallelity of geodesics at first shape. Top row: non-central correlation (\ref{non-central-corr:def}) of Fourier coefficients of $v_{\rm b}$ (initial velocity of geodesic approximating leaf shapes of leaf 'b' at its first shape) and parallel transplant $w_{\rm b}$ of $v_{\rm beb18}$ (initial velocity of geodesic approximating leaf shapes of leaf 'be1b8' at its first shape) to the first shape of leaf 'b'. Second row: the same for the $(1-p)$-values obtained from (\ref{parallel-meas:def}).\label{cor:tab}}
\end{table}
As a measure for parallelity, the cosine of the angle between initial velocity of geodesics or equivalently the the correlation of the respective Fourier coefficients could be taken, cf. first row in Table \ref{cor:tab}. At first glance, in contrast to Figure \ref{ori-parallel-tp-be1b8-2.eps}, these numbers suggest an almost non-existent amount of parallelity.
On closer inspection, taking into account, however, that the vectors compared are high-dimensional (of dimension $2N+1=201$), the first and third number of the second row in Table \ref{cor:tab} indicate high correlation as expressed in the third row: if two random vectors $v,w$ would be independently sampled from a uniform distribution on the $(n-1)$-dimensional unit-sphere, then the density of their angle $\phi$ is proportional to the surface of the $(n-2)$-hypersphere with radius $\sin\phi$ determined by this angle. Based on this consideration we propose the following measure for parallelity of random $v,w\in \mathbb R^n$
\begin{eqnarray}\label{parallel-meas:def}
\mu(v,w) &=& 1-\frac{\int_0^{\sqrt{\arccos\rho(v,w)}}\sin^{n-2}\phi\,d\phi} {\int_0^{\pi}\sin^{n-2}\phi\,d\phi}\,,~~\mbox{with }\\\label{non-central-corr:def}
\rho(v,w) &=& \frac{|\langle v,w\rangle|}{\|v\|\,\|w\|}\,.
\end{eqnarray}
Finally, let us note that the curves of the contours of the leaves be2b4 and be2b1 projected along the transplanted geodesic of leaf be1b8 start to self-intersect as the leaves grow older. While the effect is rather small, it may indicate that the natural metric of $\Sigma_{ZR}$ be adjusted in order to maintain the hypothesis of geodesic growth.
\section{Discussion and Outlook}
In this exposition a method to compare shape dynamics has been proposed based on parallel transport of geodesics. While there is quite a few work available, modelling temporal shape evolution by specific curves and splines in shape space ((e.g. \cite{JK87,KMMA01,KDL07}), to the knowledge of the author this is the first time that dynamical aspects of different shapes have undergone a comparison based on the intrinsic geometry of shape space.
In application to Botany, growth of leaves of different shapes has been compared. As underlying shape representation the space $\Sigma_{ZR}$ of closed contours based on angular direction has been employed. Within this space, in contrast to other models, entire leaf contours can be retrieved in a non-parametric way. In a simple toy example, parallel transport on $\Sigma_{ZR}$ seemed locally similar to parallel transport on Kendall's shape spaces.
This similarity can be rephrased as saying that the landmarks have been ``correctly chosen''. Parallel transport thus may serve as a tool to address an open problem in landmark based shape analysis: optimize number and location of landmark placement for a specific problem at hand. Obviously, too few and wrongly placed landmarks have low predictive power, while too many landmarks reduce power due to undesired variation. Dealing with this latter effect usually requires further methodology, e.g. statistical regularization.
In the context of \emph{dynamical shape analysis} for leaves, under optimal landmark placement, since computation of geodesics, e.g. geodesic PCA, is computationally much faster on Kendall's shape spaces than on $\Sigma_{ZR}$, one can perform parallel transport on Kendall's shape spaces and obtain the complete bounding contour in a non-parametric way by mapping to $\Sigma_{ZR}$. This lays out a path
for the future steps of the challenging endeveavor of statistically comparing shape dynamics laid out in the Introduction.
\bibliographystyle{../../BIB/elsart-harv}
|
\section{Introduction} \label{sec:intro}
Internet users today typically entrust web services with diverse data,
ranging in complexity from simple credit card numbers, email
addresses, and authentication credentials, to datasets as complex as
stock trading strategies, web queries, movie ratings, and purchasing
histories. They do so with certain \textit{expectations} of
\textit{``what''} their data will be used for, \textit{``who''} it
will be \textit{shared} with, and \textit{``what''} part of it will be
\textit{shared}. These expectations are often violated in practice;
there are $400$ reported incidents of data loss from web services in
$2009$ per the Dataloss database~\cite{datalossdb}, each of which
exposed an average of half a million customer records outside the
service hosting those records.
Such data exposure incidents can be broadly categorized into two
classes: \emph{external} and \emph{internal}. External violations
occur when an Internet attacker exploits software vulnerabilities at a
web service to steal sensitive user data, {e.g., } credit card numbers.
Software and configuration complexity are often to blame here; the web
service does not benefit from these and, indeed, loses reputation as a
result. Internal violations occur when a malicious insider within a
web service, or even the service operator itself, abuses the
possession of sensitive user data beyond what the user signed up for,
{e.g., } by selling customer marketing data to other companies. Both
external and internal violations are frequent: $65\%$ of the
afore-mentioned data exposure incidents fall in the external category,
while about $30\%$ fall in the internal category (the other $5\%$ do
not have a specified cause).
One option for a security-conscious user who wishes to limit the
impact of these threats is to avoid housing her sensitive data (say,
credit card number) at the service's site altogether. She can instead
host it close to herself: on her own machine. She then insists that
the service access this data over the network using a purpose-specific
interface. For instance, a credit card number need never be seen by a
merchant; the merchant site simply needs to request for a certain
charge to be made to the user's credit card number and receive an
authorization code to verify the transfer. This interface-based
approach assumes that the service \textit{refactors} their existing
applications ({e.g., } shopping service) to work with such an
\textit{interface}. However, this \textit{client-side deployment}
option requires the user to have her machine online whenever the
service needs to access it (a problem with recurring credit
card charges, for instance) and incur bandwidth costs (which may be
significant in some cases, such as data analytics).
Alternatively, the user can choose to host her data a bit further away
from herself; she can enlist a ``data service'', an entity
adminstratively decoupled from the ``application service'' which
requires access to her data. We refer to such a remote data service as
a {\it trusted third party} (TTP). For instance, Amazon Payment is
such a TTP that encapsulates a user's credit card number with support
for restrictions such as limit on maximum charge. This option does
delegate the bandwidth consumption and reliability issues to the TTP;
however, for high-bandwidth interactions, the TTP would require
\textit{non-trivial provisioning cost} on the user. Further, the
\textit{flexibility} afforded to the user in terms of policy is
limited by what the data service provider offers.
\begin{figure}[t]
\begin{center}
\leavevmode
\mbox{\psfig{file=figures/intro.pdf,height=4in}}
\vspace{-1.85in}
\caption{ Trading Strategy Execution At AmeriTrade: (a) Current Model
(b) The SDaC Framework}
\label{fig:intro}
\vspace{-0.35in}
\end{center}
\end{figure}
Given the trade-offs of provisioning cost, performance, and level of
protection against threats provided by the options of service-side
data hosting, TTP-based hosting, and client-side hosting, the
motivation for a flexible framework that lets a user pick a suitable
deployment based on her needs is clear. In this work, we generalize
the concept of \textit{interface-based access approach} to
re-architecting web services that supports such {\it flexible
deployment}. We built a capsule framework around this notion of
interface-based access allowing the freedom to enforce the interface
by any number of means; administrative isolation (as in a trusted
third party solution) or by software (such as a virtual
machinemonitor) or hardware module (such as a trusted platform
module).
Before elucidating the four main design principles that ground our
capsule framework, we present an instantiation of our framework in
Figure~\ref{fig:intro}. In the first option shown, the user (a
daytrader) reveals her trading strategy to AmeriTrade (an online
brokerage service) so that AmeriTrade can execute fast trades on his
behalf; however, the service is free to access the strategy as it sees
fit and may exploit it to its own advantages. In the second option,
AmeriTrade's view of the data is contrained by the interface; this
interface only allows AmeriTrade to report the current price of a
tradable instrument, and the only information revealed is the trade
the user wishes to make.
Our first design principle is that user's data should be accessed only
via a \textit{simple, narrow interface} that conforms to the principle
of least privilege. For instance, the only use a service has for a
credit card number is to charge it; access to the number itself is not
required. This let a user impose her own expectation of ``what'' is
done with her data, rather than rely on the service to enforce it. Our
main notion is that of a \emph{secure data capsule} (SDaC), an
encapsulation of a specific kind of data (say, a credit card number)
with code that implements a well-defined and open interface
suitable for the data (say, charging by a merchant). This open
interface model allows other parties to provider suitable
implementations, which can then be used by users in depositing their
data with web services. We demonstrate broad applicability of such an
interface-based model with examples from four significant application
classes: (1) a daytrading service in which large data volumes of stock
ticker data are parsed by complex private user queries to determine
automated trading actions. (2) a targeted advertising service in which
large volumes of per-user browsing history are mined. (3) a purchasing
application in which the capsule proxies requests for a credit card
number to a bank. (4) a provenance capsule that tracks changes to a
document with those who made them.
Our second principle is allow \textit{flexible interposition} of the
boundary between the capsule encapsulating a user's data and the web
service; such flexibility is crucial so that users can choose suitable
options to match their criticality, performance, cost, and threat
model requirements. Our framework supports, in addition to a trusted
third party model and client-side model, the use of trusted modules
(such as a virtual machine monitor) at application services to allow
{\it secure co-location} of the user's data at the application service
itself, whilest still guaranteeing interaction only via the interface.
Such a \textit{co-location based
hosting model} addresses some of the limitations of the TTP Model:
it is applicable for high-bandwidth interactions, requires no
additional provisioning cost for the user, and can operate even under
disconnection. As an example, a trusted virtual machine monitor can be
used to logically isolate the capsule from the service. At a higher
per-invocation overhead, but for stronger isolation guarantees against
insider attacks, the capsule can be physically isolated from the
application service on distinct hardware collocated at the provider's
site. For even stronger isolation, while avoiding provisioning costs,
the capsule can be collocated with the user's client on a separate
virtual machine, but with high network overheads and reliability only
as high as that of the client machine.
The third design principle is allow for \textit{fine-grained and
flexible user control} over how the interface to her data to her
exercised and where her capsule is hosted. This lets the user impose
her own requirements on ``what'' is done to her data and ``who'' her
data is shared with. Our framework includes a policy layer based on an
existing authorization language and mediates: (1) the parameters used
by the service in accessing her capsule's interface (2) transfers of
the user's capsule from one service to another. We note that this
policy layer is not part of the interface implementation since this
layer is meant to be controlled by users whereas the interface is
picked based on what functionality the service requires from the
service.
Our final design guideline is to allow a user direct control over {\it
what part} of her data is shared; in particular, to allow
transformations of her data before any sharing. We support two
transformations currently; filtering and aggregation. Capsule
filtering enables a capsule at one service to spawn a derivative
version of itself at another service containing only a subset of the
user's original data. For example, a purchase history capsule at a
shopping site may spawn a version of itself with only music purchases
for the use of a music recommendation service. Although the same
semantics could be implemented via a policy extension (i.e., only
allowing the music service to read history entries pertaining to music
purchases), explicit filtering narrows the attack surface offered to
the music service, in the case where the isolation it offers were to
be violated. Capsule aggregation allows the merging of the data from
multiple capsules into an aggregate capsule that can be treated
monolithically by an application service. This is useful when services
simply require access to raw user data rather than
interface-restricted access; in such cases, one option is for users to
aggregate their data with others to gain privacy. For example, a user
and her friends may instruct their purchasing history capsules to
merge before interfacing with a recommendation service, to provide a
degree of anonymity for individual purchases across all friends
included.
\begin{table*}[t]
\begin{scriptsize}
\renewcommand{\arraystretch}{1.15}
\caption{Comparison of various deployment models}
\vspace{-0.1in}
\label{tab:intro_comparison}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline\hline
Scenario & Flexibility of Policy & Performance & Provisioning
& TCB & External & Internal \\
& & & Cost Borne By & & Threat & Threat \\
\hline\hline
Raw Data & Limited by
app service & No network required & None & Unknown TCB (OS, apps) & No
& No \\
at app service & & & & & &\\
\hline
TTP Capsule & Limited by
data service & Limited by network & User & Trusted data
service & Yes & Yes \\
\hline
Client Capsule & User can run own code &
Limited by network & User & Open-source compact
TCB & Yes & Yes \\
\hline
Co-located Capsule & User can pick own code & No network required &
Service & Open-source compact TCB & Yes & Yes\footnote{secure
co-processor required} \\
\hline\hline
\end{tabular}
\end{center}
\vspace{-0.3in}
\end{scriptsize}
\end{table*}
Table~\ref{tab:intro_comparison} summarizes the various options of
hosting the data in our capsule framework: the first is the current
case where data is stored at the application service, whereas the
other options are all supported by the capsule framework with the
consequent trade-offs. We note that the co-located capsule can help
defend against both internal and external threats with the support of
a secure co-processor, and only against external threats if a virtual
machine monitor or trusted platform module is available. The security
guarantee provided by our framework against these threats is that the
combination of interface restrictions and policy control is an
upper-bound on the impact of an attacker.
We have implemented a prototype SDaC framework that supports three
deployment models: TTP, client-side, and co-location using the Xen
hypervisor~\cite{xen_url} to isolate the data capsule from the
service. Our prototype implementation can only defend against internal
attacks in the client-side or the TTP model. Our co-location
implementation can only defend against external attacks, and {\it not}
internal attacks, since our implementation does not support a secure
co-processor. However, we believe it would be simple to extend our
implementation to support co-location based on TPMs and secure
co-processors, based on standard techniques ({e.g., }
Flicker~\cite{flicker:eurosys08}).
We evaluate the SDaC approach along four axes: broad application
applicability, flexibility increase, performance, and TCB reduction.
We demonstrate broad applicability by refactoring exemplars from four
significant application classes. We demonstrate flexibility and
performance gains by evaluating the performance of each application in
a different deployment scenario, further showing that the optimal
deployment for each application differs, and significantly outperforms
other possible deployments. For instance, the co-location option
consumes no network bandwidth, as opposed to the TTP and client-side
capsules, which requires over $1$ Gbps network bandwidth per user in
the stock trading example. The Xen-based co-located capsule does incur
non-negligible access overhead (around $1$ ms) and storage costs
(around $120$ KB; though this can be amortized across users). However,
given the value of high-stake data, such as financial and health
information, and the fact that the cost of data exposure was estimated
by a recent study~\cite{ponemon:url} to be over $200$ dollars per
customer record, this overhead may be a acceptable price to pay for
controls over data exposure. The trustworthiness of our framework is
dependent on the SDaC implementation's correctness. As compared to a
complex and proprietary application service, the SDaC is typically
simple and can be divulged openly allowing for early discovery of
vulnerabilities. Our SDaCs do not require any complex OS services; our
design offloads the device drivers, network stack, etc to an untrusted
entity.
While we acknowledge that the concept of encapsulation is a well-known
one, we view as our main contribution a general architectural solution
to the problem of data access control in web services based on the
encapsulation principle, as opposed to custom solutions suited for
credit cards or targeted advertising. Our work is also related to
service composition approaches, but is specialized to providing
improved security, rather than improved functionality.
Our capsule framework's applicability to a particular service is
limited by three main factors. First, capsules are not useful where
the interface required by the service involves {\it complex} function
calls (which precludes a simple capsule implementation) or when it
{\it reveals} the data directly (once the data is revealed directly to
an untrusted service, information flow control
mechanisms~\cite{jif:jsac03,fabric:sosp09,rifle:micro04,xbook:ssym09,asbestos:sosp05,dstar:nsdi08,resin:sosp09}
are required; data access control only controls release not
propagation). Second, our work also presupposes that a particular
interface is chosen by the service and vetted by the broader
community. It is a non-goal of our work to identify such interfaces
automatically for a service, or to prove that a given interface
provides a given level of privacy or other guarantees. For the various
web services we consider in this work, we suggest and implement
possible interfaces. Third, we assume that the service provider will
be willing to undergo a moderate amount of application refactoring,
possibly executing SDaC code they did not implement. We believe this
is realistic given that web services like Facebook already run
third-party applications as part of their business model.
The rest of the paper is structured as follows.
Section~\ref{sec:problem} defines our problem statement, while
Sections~\ref{sec:arch} and~\ref{sec:design} present an architectural
and design overview respectively of our capsule framework. We discuss
our implementation in Section~\ref{sec:implementation} and evaluate
its performance in Section~\ref{sec:evaluation} and its security in
Section~\ref{sec:secanalysis}. We then present related
work~\ref{sec:related} and then conclude in
Section~\ref{sec:conclusion}.
\section{Problem Statement} \label{sec:problem}
The object of our capsule framework is to allow users {\it secure and
flexible control} over {\it who} does {\it what} to their data, {\it
when} they do it, and {\it where} their data is stored. Thus, our
goal is to enable flexible data access control. The {\it what} is
specified by the interface used to access the data, and is based on
{\it least privilege}; user data is exposed to the service only to the
extent required to accomplish the desired functionality. A
user-specified policy allows fine-grained control over the {\it
what}, {\it who}, {\it when}, and {\it where} aspects of operations
performed on \textit{what} part of her data.
Two broad requirements for our capsule framework are: (1) Generality:
The framework must be broadly applicable across several kinds of web
services. (2) Flexible Hosting: The design should allow a variety of
hosting options depending on the trust assumption; co-location or TTP
or client-based. The requirements of data-specific interfaces and
flexible hosting in our framework distinguish us from the two closely
related papers in literatures (Wilhelm's
thesis~\cite{mobile_privacy:thesis} and
Iliev~{et al.}~\cite{client_privacy:secpriv}); we discuss related work in
more detail in Section~\ref{sec:related}. We now flesh out our
security goals and assumptions
in more detail. \\
\noindent \textbf{Security Goal:} Our security goal is to ensure that
the policies specified by the user over her data are never violated.
In particular, any access to her data must be a sequence of legal
invocations of the interface sanctioned by the user that conforms with
her policy, and the only data exposed to the service is the output of
such a legal sequence of legitimate interface invocations. Note that
we only aim to circumscribe the influence of an adversary by the
limits set by the policy; the adversary will be able to exercise her
influence to the maximum allowed by the policy. For instance, if the
budget in a CCN capsule is specified as $100$ dollars, an adversary
will be able to exhaust this quota. However, the adversary will
neither be able to extract the CCN from the capsule, nor can she go
over $100$ dollars. There are two kinds of adversaries that we wish to
achieve this goal against.
\textit{Adversaries with software-only access:} This models an
Internet attacker external to the service who can exploit
vulnerabilites in the service software stack (above and including the
OS; we will assume the VMM, if any, is secure). Such adversaries are
responsible for the external category of attacks.
\textit{Adversaries with physical access:} This models a malicious
agent at the service site who has physical access to the machine
hosting sensitive data; for instance, she may be able to monitor the
memory bus (say). This agent may be an admin or the organization
itself, who can launch internal attacks.
We present statistics from the DataLoss database~\cite{datalossdb} to
quantify the impact of these adversaries. Over the last decade,
adversaries with software-only access have been responsible for $65\%$
of data exposure incidents; there have been $394$ such incidents
caused due to hackers with an average of over $1$ million sensitive
user records exposed per incident. During the same period, adversaries
with physical access have caused $10\%$ of the incidents; there have
been $170$ such incidents with an average of over $340,000$ records
exposed per incident. We note that these two classes of adversaries
directly account for $75\%$ of the incidents in the database; of the
remaining $25\%$ incidents, the $21\%$ for which the cause is known
are due to accidental exposure due to inside elements. This category
includes incidents such as stolen/lost/disposed
computer/laptop/disk/drive/tapes and accidental mis-configuration of
software due to human error. For our purpose, the former kind of
incidents falls in the second class as the adversary has physical
access, while the second kind of incidents
falls in the software-only access case. \\\
\noindent \textbf{Security Assumptions:} Our framework can leverage
any of the following trust assumptions to ensure isolation between the
SDaC and the application service: \bi
\item Trusted Third Party (TTP): A TTP is an administratively
different entity from the service that the user wishes to protect
her data from. This trust can be based on a service-level agreement
and the TTP would be paid for by the user. The option of hosting the
capsule at the client's own computer is similar to the TTP option,
since from the service's perspective, it interacts with the user's
data over the network in both cases.
\item Trusted Hardware: A user can place her trust in a hardware
module at the service site. Two kinds of hardware modules are
currently available: Trusted Platform Module
(TPM~\cite{tpmspec:tcg}) and Secure Coprocessor ({e.g., } IBM
4758~\cite{secure_coprocessor:computer01}). TPMs are now widely
available in server machines, while secure co-processors, though
preceding TPMs, are suitable for high-value data since they are
considerably more expensive.
\item Trusted Software: A user can trust a virtual machine monitor
(VMM, {e.g., } Terra~\cite{terra:sigops03}) at the server site in order
to achieve her goal. VMMs are now widely used in server
environments, making this very viable. We envision two options here:
attested VMs (where a TPM attests the execution of the VMM remotely
to the client) and an un-attested VM (where the user trusts the
service to invoke the VMM).
\end{itemize}
Of these four trust assumptions, the TTP and the secure co-processor
models enable the capsule framework to defend against both threats
(internal and external), while TPMs and VMMs suffice to protect
against the threat of a software compromise. Once again, we note that
our prototype implementation \textit{cannot} defend against internal
attacks in the co-location model. \\
\noindent \textbf{Scope:} Our work currently make two assumptions.
First, we assume that a satisfactory interface has been arrived at for
a service of interest that provides the desired privacy to the user;
we do not attempt to verify that such an interface guarantees the
desired level of privacy, or to automate the process of re-factoring
existing service code. For the specific interfaces we present in this
work, we argue for their privacy based on informal arguments; we leave
formal verification for future work. Second, we assume that a given
capsule implementation carries out this interface correctly, and that
there are no side channels or bugs. For the capsules we discuss, we
strive to keep the interface simple and argue for the correctness of
the implementation based on the simplicity of the interface; proving
the correctness formally is a non-trivial research problem that we do
not aim to address in this paper.
\section{Architecture} \label{sec:arch}
\begin{figure}[t]
\begin{center}
\leavevmode
\mbox{\psfig{file=figures/arch.pdf,height=3.5in}}
\vspace{-1.85in}
\caption{The Capsule Architecture}
\label{fig:arch}
\vspace{-0.35in}
\end{center}
\end{figure}
The main elements of our capsule architecture are presented in
Figure~\ref{fig:arch} that shows a user who capsulizes his trading
strategy at AmeriTrade's site; the capsule relies on a VMM for
security. The figure also shows the capsule lifecycle; the user
generates the capsule on her machine (say) and then transfers the
capsule to Amazon. This capsule then interacts with the data via a
prescribed interface. From hereon, we will refer to the user of a web
service as a {\it user} and to the web service as a {\it service}. In
the rest of this section, we will first examine the implications of
our framework in terms of how users and services interact with it. We
will then show that such an architecture is general and has several
applications; thus, our design principle of \textit{interface-based
data access} is widely applicable.
\subsection{Implications to Users and Services}
A user, who wishes to use the capsule framework to protect her
sensitive data when stored at a service, picks a capsule
implementation from third party software companies (say a security
company such as Symantec) which exports an interface suitable for the
kind of data. Or, she may pick one from an open-source repository of
capsule implementations, or purchase from an online app-store. It is
even possible that the service itself offers a third-party audited
capsule implementation which the user may trust. For such a healthy
ecosystem of capsule implementations to exist, we envision that APIs
suitable for specific kinds of data will eventually be well-defined
for commonly used sensitive information such as trading strategies,
credit card numbers, email addresses, and web histories. Services that
require use of a specific kind of data can support such an API, which
would then be implemented by open-source reference implementations and
security companies. The development of such interfaces would have
another beneficial collateral effect; it would enable data portability
for users across services. Once the user picks a suitable
implementation, she then customizes it with policies that limit where
her data may be stored and who may invoke it. Our current
implementation requires users to use a declarative language for this
purpose; we envision that simpler user interfaces will be used for
this purpose. Once customization is done, the user initiates an
installation process by which the capsule is hosted (at a TTP or at
the service as desired) and the association between the service and
the capsule established.
From the perspective of the web service, the following changes need to
be made. First, they need to be willing to run third-party code in the
form of capsules. We believe this is reasonable since the
functionality of capsules is very limited; they can be executed within
a sandbox with simple policies ({e.g., } allow network access to only the
payment gateway, allow no disk access). Further, a service can insist
that the capsule be signed from a set of well-known security
companies, similar to how applications are signed today. This gives
the service confidence in protecting its own code and data from the
capsule code. Given the success of Facebook applications, we believe
this model is reasonable. Second, a service may have to modify its
code to interact with the capsule via a programmatic interface,
instead of accessing raw data as they do today. This appears feasible
given the short life cycle of web service code (which are re-written
much more frequently than applications) and the fact that most web
services are architected so as to retrieve data from a remote machine
or via an application server. Modifying the Zen shopping cart
application~\cite{zen} to interact with the credit card capsule took
us only a single day. Third, the overhead of capsule storage and
invocation may be a hindering factor for adoption. The storage
overhead of the capsule seems reasonable especially if the capsules of
various users consists of code written by a small set of security
companies; thus, there would be considerable overlap in the code, and
the only additional storage required is for the user's data. We deal
with the invocation overhead in more detail in our evaluation section
(Section~\ref{sec:evaluation}), but for now, we note that the overhead
of invocation depends on the particular trust module in use. VMMs may
offer acceptable overhead since servers typically use highly optimized
virtualization already; several ways are known to optimize intra-VM
communication ({e.g., } shared-memory based~\cite{xensocket:middle07}) and
to run in-line code securely without incurring inter-VM context switch
overhead ({e.g., } using hardware paging features~\cite{secureinvm:ccs09}).
TPMs and secure co-processors incur significantly more overhead, and
may be feasible only for high-value data such as stock quotes and
health records. Late launch invocation in TPMs is still slow since
this feature is meant only for VM launches; however, in the future,
the trend is towards improving the performance of these devices
especially if they are used widely by web services.
\subsection{Notation}
We denote a capsule $C$ owned by user $U$ and resident on a machine
$M$ owned by principal $S$ as $C_U@[M,S]$. We denote the machine $M$
in the notation since whether or not a capsule $C$ can be hosted at
$M$ may depend on whether $M$ has a trusted hardware/software
module. Every capsule $C$ (we omit the service and the machine when
the exact capsule referred to is clear from the context) has a policy
database $P[C]$ specified by the user during creation. A capsule
$C_U@[MU,U]$ is created by the user $U$, for instance, on her own
machine $MU$. The user can then choose to host this capsule elsewhere,
say to a trusted third party $T$, by requesting a transfer
operation. The capsule $C_U@[M,U]$ then initiates a transfer operation
to the TTP $T$, at the end of which a capsule $C_U@[M,T]$ is now
hosted on the TTP $T$ ($M$ belongs to $T$). Alternatively, the user
can choose to host this capsule on the service's site itself by
relying on a VMM for security. In this case, the user's data is stored
at the capsule $C_U@[M,S]$ where $S$ denotes the service that owns
$M$. Once resident at $S$, the capsule $C_U@[M,S]$ can be accessed by
$S$; any invocations conforming with the user's policies (specified in
$P[C_U@[M,S]]$ are allowed by the capsule.
In order to model scenarios where one web service service can share
(with the user's consent) the user's data with another service, we
allow a service $S$ to share the user's data with a second service
$S'$. Alternatively, $S$ can proxy invocation requests from $S'$;
however, this makes $S$ liable for accesses that $S'$ initiates and
further, requires $S$ to incur overhead on behalf of $S'$. $S$ may not
desire such responsibilities and may prefer to transfer the capsule to
$S'$. To do so, $S$ requests the capsule $C_U@[M,S]$ to transfer to
$S'$. If the user's policy allows such a hosting, then the capsule
$C_U@[M,S]$ transfers itself to be hosted on $S'$ as $C_U@[M',S']$. We
note that, in order to support replication within the same service,
the service $S$ can ask that the capsule be hosted at another machine
$M'$ owned by the service $S$; this is treated as equivalent to an
across-service transfer. After the capsule has been installed, the
user also has the option of updating her data in various ways; we
support simple addition/redaction as well as allowing derivate
capsules that may have lesser data (by filtering) or may have
aggregate data from several users. Our subsequent
sections~\ref{sec:design_hosting},~\ref{sec:design_invocation},~\ref{sec:design_transformations}
deal with these three issues in more detail: hosting protocol for
tranfers, policies on invocation, and data transformations.
\subsection{Choice of Interface} \label{sec:arch_interface}
We now explain how the interface for a service is chosen, and will
illustrate the wide applicability of interface-based data access.
Denote by $F$ the service functionality that operates on the data; in
general, it is a function of $D_U$ (user's data) and $D_S$ (data
provided by the service). In the credit card scenario, $D_U$ is the
CCN number, $D_S$ is the merchant's account number, and $F$ contacts a
payment gateway and requests a transfer from CCN number $D_U$ to
merchant account number $D_S$.
When such functionality $F$ is provided by a capsule, it is refactored
as $F_C$ (implemented by the capsule), and $F_S$ (implemented by the
service) that operates on the output of $F_C$.
We will allow $F_C$ to operate on $D_U$ and $D_{SU}$ (which denotes
part of the service data $D_{S}$ which is revealed to the
capsule). The requirement is that the functionality remains the same
as before; thus, $F(D_U, D_S)$ is semantically equivalent to
$F_S(D_{S}, F_C( D_U, D_{SU}))$. The service reveals $F_C$ and
$D_{SU}$ (a subset of $D_S$) to the capsule, while the user reveals
$F_C(D_U,D_{SU})$ to the service. The final output of the function
$F(D_U,D_S)$ is typically exposed to the user, and in some cases, the
service as well. The order of invocation of $F_C$ and $F_D$ is
immaterial; we choose the formulation where $F_C$ acts first since we
are concerned with protecting the user. Using this notation, we
present some example classes of capsules; these classes are
distinguished based on a particular kind of application scenario. So,
it is useful to think of them as a starting set of usage idioms for
capsules. \\
\noindent \textbf{Query Idiom:} In this usage idiom, the server
provides a stream of incoming data, and the user would like to obtain
part of this stream or initiate actions based on this stream using a
filter without revealing either her filter or the matching items. In
this case, $D_S$ is this stream of incoming of data and $D_U$ is a
filter representing $U$'s interest in the data. The user's data is
thus like a continuous query in databases. For a capsule following the
query idiom, the following additional conditions hold: (a)
$F(D_{S,new},D_U)$ is the functionality that infers if the filter
$D_U$ matches a new data item $D_{S,new} \in D_S$. (b) $F$ is
well-known (c) $S$ reveals $D_{S}$ to the user since the user has paid
for this data already ({e.g., } stock quotes) or because the stream $D_{S}$
is public. Condition (b) implies that the specs for $F$ can be
revealed publicly, enabling capsule implementers to support $F$.
Condition (c) implies that the service can share $D_S$ with the user's
capsule. These conditions ensure the capsule architecture is suitable
here.
Examples where this idiom applies include: (a) Stock Trading: Filter
is list of stock quotes that the user is interested and trading
algorithms that operate on ticker data. Stream of incoming data at the
server is high-rate financial ticker updates. The capsule
functionality makes trades based on ticker data. (b) Google Health:
Filter is list of medicines taken by user and list of diseases that
user has. Incoming stream at server consists of newly discovered
conflicts between certain diseases and certain medicines. A match
occurs if a new conflict involves user's medicines and diseases. (c)
Google News: Filter is user's list of interested keywords, while the
incoming stream is list of news articles. (d) Auction websites ({e.g., }
Swoopo): Filter represents the bidding strategy of a user, whie
incoming stream is the current auction price and stream of bids from
other users.
A query capsule supports two interface calls: {\it Match} (used by the
service to update the capsule with a new data item in the stream) and
{\it RetrieveMatches} (invoked by the user to retrieve matching
items). Both are extremely simple to implement; {\it Match} adds the
new datum to a match buffer if a filter criterion matches, while {\it
RetrieveMatches} authenticates the user and delivers matches. Care
is required in implementing these functions to avoid side channels
that expose the user's filter; we will address these in
Section~\ref{sec:implementation}. For now, we assume that the capsule
buffers up all matches within itself; whenever the user wishes to
receive updates, he {\it pulls} matching data items from the capsule.
This is to simplify the implementation; the {\it push} option is also
possible. Alternatively, the capsule may initiate some actions
automatically (such as make trades automatically) without any user
involvement. \\
\noindent \textbf{Analytics Idiom:} This idiom captures scenarios such
as targeted advertising, recommendation systems, where the service $S$
performs statistical analysis on the user's data, such as, web visit
logs, query logs, location trajectories, previous purchases. In this
idiom, the following hold: (a) $F(D_S,D_U)$ performs statistical
analysis on the user's data $D_U$ along with service data $D_S$. (b)
$F$ can either be proprietary or public (c) $D_S$ can either be
proprietary or public.This scenario is more complicated than our other
scenarios since both parties have private data: the user does not wish
to reveal $D_U$ while the server may not wish to reveal either $F$ of
$D_S$.
We will discuss the targeted ads scenario in detail. Here, $D_S$ is
the list of possible ads to display to the user which is not
proprietary. $F(D_U, D_S)$ selects one ad out of the list $D_S$ to
display and may be proprietary. We discuss the capsule interface
under two cases: when $F$ is well-known and when $F$ is proprietary.
There is one further choice to be made: whether $F(D_S,D_U)$, the ad
to be selected, is known to $S$. For simplicity, we assume that the
user only wishes to hide $D_U$ and not $F(D_S,D_U)$ since $D_U$ is
typically much more detailed than $F(D_S, D_U)$; it is possible to
extend this interface to hide the selected ad as well.
\textit{Well-known $F$:} This case is similar to the query capsule
case since both $F$ and $D_S$ are public. Thus, the re-factoring is:
$F_C = F$, $F_S = \phi$, and $D_{SU} = D_S$. $F_C$ is shown in
Algorithm~\ref{alg:targeted_ads_1} (we rely on
Adnostic~\cite{adnostic:ndss10} for the description of $F$). In this
description, $C$ denotes a category space representing various kinds
of user's interests such as ``Entertainment $\rightarrow$ Comics and
Animation $\rightarrow$ Cartoons'' , ``VideoGames $\rightarrow$ PS3''.
Interest categories are hierarchical; $\rightarrow$ represents a
sub-hierarchy. Google uses a 3-level hierarchy with around 600
categories overall. For simplicity of presentation, we ignore the
hierarchical structure within $C$ and consider it as a pre-specified
list of possible interest categories. $K$ denotes a keyword space
consisting of meta keywords that occur in webpages which convey
semantics; this keyword space is private to the capsule and is used in
the computation of $F$. The simplicity of this capsule is evident from
the conciseness of our pseudo-code.
\begin{algorithm}[t]
\caption{Select the ad to be shown based on user's history}
\label{alg:targeted_ads_1}
\begin{algorithmic}[1]
\STATE \textit{Function $F_C$ =
ChooseAd(ListOfAds=$\{A_1,A_2,\cdots\}$, CategoryOfAds=$\{C_{i,c}\}$)
where $C_{i,c}$ reflects the weight of ad $A_i$ in category $c$}
\REQUIRE \textit{Params: $D_{SU} = \{A_1,A_2,\cdots,\}, \{C_{i,c}\}$
\\ $D_U = \{W_1,W_2,\cdots\}$ ($W_i$ is set of websites
visited by $U$), \\ $\{V_1,V_2,\cdots\}$ ($V_i$ is the number of visits to website $W_i$ by
user $U$), \\
$\{WK_{i,k}\}$ ($WK_{i,k}$ is the weight of webpage $i$ for
keyword $k$; this is prepared by setting $W_{i,k}$ to one for words $k$ in the title bar
and any meta keywords $k$ specified in the html source of the
webpage, and then normalizing it), \\
$DB = {DB_{k,c}}$ gives the weight for a keyword $k \in K$ to a
category $c \in C$ (this database is public and can be prepared by
crawling del.ici.ous bookmarks)}
\ENSURE \textit{returns Ad $A_i$ as selected ad}
\STATE $\forall~c~\in~C$, Compute $U_c$ (interest vector for user) as
$\frac{\sum_{i,k,c} ( V_i * WK_{i,k} * DB_{k,c})}{\sum_{i} V_i}$.
\STATE $\forall~A_i$, Compute $S_i$ (score assigned to Ad $A_i$) as
$\sum_{c \in C} C_{i,c} * U_c$.
\RETURN{$A_i$ st that $S_i$ is highest}
\end{algorithmic}
\end{algorithm}
\textit{Proprietary $F$:} Google may choose to develop an algorithm
$F$ that depends on the user's browsing history, the ad's category,
the contents of the webpage that the ad is displayed on, and the
user's click through rates on ads previously shown to him. It is
difficult to anticipate what information may be used by Google, so in
such cases, we envision a capsule interface $F_C = F_C(D_U)$ (meaning
$D_{SU}=\phi$; that is, the service reveals nothing to the user), and
thus $F = F_S(D_S,F_C(D_U))$. One such interface $F_C$ is to return a
{\it perturbed} histogram of keywords that occured in web pages
visited by $U$; this perturbation is carried out using differential
privacy techniques~\cite{differentialprivacy:icalp06} which offers
provable privacy guarantees to the user. This interface gives
different privacy guarantees compared to the interface for public $F$;
in this case, the user gives up part of his privacy for the sake of
the complete privacy of Google's code. \\
\noindent \textbf{Proxy Idiom:} A capsule belongs to the proxy idiom
if the user has delegated some rights she has on a second service $S'$
to the capsule; the service $S$ uses the capsule to access $S'$ using
the user's right. We use the term proxy capsule since the capsule
which acts as a proxy between $S$ and $S'$. For such capsules, $D_U$
is typically a credential that is required to authenticate $U$ to
$S'$. For example, $D_U$ may be a credit card number, $S'$ is the
user's credit card company, and the capsule allows a merchant to
charge the user.
A proxy capsule offers two kinds of guarantees to the user. First, it
ensures that $D_U$ is not exposed to $S$. Second, it also constrains
the kind of operations that can be invoked under $U$'s authority on
$S'$. For a capsule following a proxy idiom, the following conditions
hold: (a) $F$ is a well-known function. (b) $D_{S}$ is not
confidential information, so it can be exposed to the capsule. (c) $F$
requires interaction with a second service provider $S'$. Conditions
(a) and (b) imply that following factoring is acceptable to $S$: $F_C
= F$, $F_D = \phi$, and $D_{SU} = D_S$. Condition (c) implies that the
capsule functionality can be offloaded to the service provider $S'$ in
some cases, but this is not always possible if $S'$ does not support
the kind of policies that $U$ would like. Condition (c) also implies
that the capsule requires network access, typically over SSL. We now
discuss payment capsules, one particular instantiation of a proxy
capsule, and then discuss other instantiations.
\begin{algorithm}[t]
\caption{Charge user's credit card number and tranfer requested amount
to merchant's account }
\label{alg:ccn_capsule}
\begin{algorithmic}[1]
\STATE \textit{Function $F_C$ = Charge(Amt, MercAcct)}
\REQUIRE \textit{Params: $D_{SU} = (Amt, MercAcct)$, Amt $>$ 0, \\
$D_U = CCN \text{(set by user)}$, \\
CapsuleState: PaymentGateway (set to dns name of payment gateway such
as ``gateway.authorize.net''), \\ PaymentGatewaySSLName(set to
VeriSign-signed name of gateway such as ``Authorize.Net''), \\
RequestTemplate (set to template HTTP POST request for charging)}
\ENSURE \textit{RetVal $=$ ConfCode if successful, $= -1$ otherwise}
\STATE Request $=$ RequestTemplate
\STATE Request $=$ Replace(Request, ``Amount'', Amt)
\STATE Request $=$ Replace(Request, ``MerchantAccount'', MercAcct)
\STATE Response $=$ SubmitOverSSL(PaymentGateway,PaymentGatewaySSLName, Request)
\STATE ConfCode $=$ RegexpExtractBody(Response,/conf\_code=/)
\RETURN{ConfCode}
\end{algorithmic}
\end{algorithm}
A \textbf{payment capsule} allows a service $S$ to charge user $U$ on
a recurring basis while allowing the user to impose policies on such
charging ({e.g., } do not charge more than once a month) and protecting the
user's charging token (such as a CCN or a gift certificate or e-cash)
from the service. The interface $F_C$ for such a capsule and its
implementation is shown in Algorithm~\ref{alg:ccn_capsule}. $D_{SU}$
is passed into the function $F_C$ during invocation, and $D_U$ is
stored within the capsule permanently. We use the notation {\it
Capsule State} to indicate scratch state retained across
invocation. The notation {\it Require} and {\it Ensure} indicate the
pre-condition and the post-condition for the function
respectively. Lines $1-4$ construct the POST request to be sent. The
two functions in this capsule dispatch the request and parse the
reply: {\it SubmitOverSSL} and {\it RegexpExtractBody}. {\it
RegexpExtractBody} is a simple regular expression matching function
used to extract the confirmation code from the server's response. {\it
SubmitOverSSL} is responsible for contacting the gateway over SSL
and obtaining the server response to the constructed POST request. We
will explain in Section~\ref{sec:implementation} how this function can
be implemented by leveraging an {\it untrusted} network stack; for
now, we will note that it can be implemented with the aid of a trusted
cryptographic library and an untrusted network stack.
Two other instantiations of a proxy capsule are access capsules and
notification capsules; due to space constraints, our discussion below
is limited. An \textbf{access capsule} allows a service $S$ limited
access to the user's data stored at a second service $S'$, while
keeping the user's authentication token for $S'$ ({e.g., } password) secret
from $S$ and imposing restrictions on accesses made from $S$ to
$S'$. For instance, an access capsule is used by a user $U$ to allow
FaceBook restricted access to her contacts list on Gmail, and hide the
rest of her Gmail data (emails etc) from Facebook. Other examples
include services like Mint that require access to the user's
up-to-date bank financial transactions at say, Bank of America. A
\textbf{notification capsule} allows a service $S$ to send messages to
a user $U$, while the user $U$ need not reveal anything about how and
where she can be reached, and can specify policies on message sending
(such as rate limiting). An example is a email capsule where $F$
implements sending the message via email to $U$, $D_U$ is $U's$ email
address, $D_S = \phi$, $S$ is an online merchant, $S'$ is $U's$ mail
server, and $F(D_S,D_U)$ returns a success code if the delivery was
succesful. \\
\noindent \textbf{Provenance Idiom:} In this idiom, the user's data
$D_U$ is modified by various parties that have access to it; the goal
is to ensure that each modification is tied to the party which
modified it. Thus, for this case, $F(D_U,D_S)$ updates the data $D_U$
and $D_S$ is the modifications made by the party to which the capsule
is forwarded. The security guarantee is thus an integrity requirement
rather than a privacy requirement. The various parties are willing to
reveal $F, D_U, D_S$ to each other; the security guarantee is that a
data modification $D_S$ is tied to the party $S$ that made the
modification.
We envision such a capsule being useful in a enterprise scenario for
collobarative editing. A provenance capsule containing a text file
(say, a legal memo or a business plan) is forwarded by a user to a
group of people who will collaboratively edit the data; the data owner
wants to ensure that any changes to the data is tied to the person who
made the change. This usage is reminscent of a distributed version
control system with additional security requirements.
A provenance capsule supports two calls: \textit{ReadData} and {\it
WriteData)}. Upon invocation, the capsule authenticates the caller
and returns the data (for {\it ReadData}) or updates it along with a
signature of the modification by the caller (for {\it WriteData}). The
implementation is simple: {\it WriteData()} appends the modification
to a write log along with the signature of the invoking principal as a
proof of provenance, {\it ReadData()} returns the current version of
the data. \\
\noindent \textbf{Summary:} The four idioms above capture a wide set
of scenarios; of these, the proxy capsule is the most complicated
since it requires network access. For these scenarios, the capsule
interface is simple and provides the required privacy guarantees to
the user; we believe that this simplicity bodes well for a correct
implementation. Capsules are not useful where the interface involves
{\it complex} function calls or when it is {\it leaky}. As an example
of the first, consider the interaction between Google Docs and a
user's document. While one can theoretically envision an interface
({e.g., } display the first page on the screen, delete character at line 20
and position 30), the interface would be so rich as to make it
difficult to characterize leaks and to give any confidence as to the
correctness of the capsule implementation. The interaction between
Facebook and a user's profile data is an illustration of the second
case. Facebook requires direct access to the user's data ({e.g., } name,
address), and such an interface affords no privacy to the user; once
the capsule releases the data to a Facebook process, it is not
possible to ensure any policy constraints. These two scenarios are
better handled by information flow control ({e.g., }
XBook\cite{xbook:ssym09}) or end-to-end encryption ({e.g., }
NOYB~\cite{noyb:wosp08}).
\section{Design} \label{sec:design}
\begin{figure}[t]
\begin{center}
\leavevmode
\mbox{\psfig{file=figures/impl_overview.pdf,height=3.5in}}
\vspace{-1.4in}
\caption{Design Overview}
\label{fig:design_overview}
\vspace{-0.35in}
\end{center}
\end{figure}
In this section, we discuss the design of the capsule architecture in
more detail. The previous section mainly focussed on our design
principle of interface-based access; the components of the base layer
we discuss in this section are based on the principles of (a) flexible
deployment with policy control. (b) policy control over invocation.
(c) policy control over data on which the invocation operates.
Figure~\ref{fig:design_overview} shows a capsule co-located at the
service site that uses a VMM as a trusted module to isolate itself
from service code. It shows a {\it host hub} that resides within the
service partition and a {\it base layer} in the capsule partition (we
use the term partition to imply the isolation between these
components). We assume the base layer is installed by the service
within the VM allocated for a user's capsule; during the installation
protocol, the user can verify that this is the case using attestation.
The host hub acts as the proxy for communication between the external
world and the capsule. The base layer includes generic functionality
(such as hosting protocol, policy checking) common to all capsules,
while the data layer implements interface calls specific to the data
item.
The figure also shows the three main components of the base layer: the
policy engine, the data transformation module, and the service front
end. The policy engine is used by the base layer to check whether an
invocation request or an hosting request should be allowed or not.
Regarding invocation requests, we are guided by the principle of
providing fine-grained flexible control, and regarding hosting
requirements, our goal is to to allow a range of flexible deployment
options. The data transformation module is responsible for
implementing a secure aggregation protocol (filtering is data
specific, and are implemented in the data layer). The service front
end is responsible for carrying out complex functions (such as network
access, disk access) which are out-sourced to the service back end in
the host hub which uses device drivers to provide this functionality
to the capsule (the concept of front end and back end is similar to
how Xen supports devices inside a guest VM). Such \textit{service
requests} are made by the data layer and are proxied to the service
back end by the service front end. This split architecture
significantly reduces the complexity of the capsule and our TCB; the
capsule only relies on the host hub for availability, not for
confidentiality or integrity guarantees on the data. The only code
inside the capsule correspond to the base layer functionality and data
layer functionality; typical OS components such as device drivers,
network stack, are not required.
The rest of the section presents the detailed design. We first discuss
the interaction between the host hub and the base layer
(Section~\ref{sec:design_interaction}). We then discuss the policy
engine in two parts: the hosting policy engine
(Section~\ref{sec:design_hosting}) and the invocation policy engine
(Section~\ref{sec:design_invocation}). We then finally discuss the
data transformation module (Section~\ref{sec:design_transformations}).
\subsection{Interaction between Host Hub and Base Layer} \label{sec:design_interaction}
The interaction between the host hub and the base layer is secured by
the base layer by leveraging the trust module. If the capsule is
hosted at a TTP or at the client, then the trust is guaranteed by the
physical isolation between the capsule and the service. If the capsule
is co-located at the service, then the base layer requires the
following functionality from the trust module: (a) isolation (b)
non-volatile storage. (c) random number generation (d) remote
attestation (optional). Isolation is needed to ensure that the service
code cannot interfere with the capsule and steal its data during
execution. Non-volatile storage is required to defend against replay
attacks. Otherwise, it is possible for an adversary to rollback the
capsule state to an earlier point in time in an undetectable fashion;
this would allow them to violate the security guarantee desired by the
user. Randomness is required in order for secure key and nonce
generation in the various protocols that the base layer implements.
Remote attestation is optional; it lets the trust module can prove to
the user that her data is indeed processed by the capsule. We will
later argue in Section~\ref{sec:secanalysis} that these properties
suffice for the security of our framework; for now, we note they are
provided by all the three trust modules we consider (VMM, TPM, secure
co-processor). We now introduce the host hub and the base layer.
The host hub has two main roles. The first is to serve as a proxy for
all communication to and from the capsule; such communication may be
to/from the user (during capsule installation) or to/from the service
code (during invocation). This design decision lets us decouple the
application on the service side from the details of capsule invocation
and is a modularity-based decision. The second role is to provide
services to the DC such as network access and persistent storage; this
allows the capsule to leverage functionality without having to
implement it, considerably simplifies the implementation. The host hub
runs within a un-trusted OS, but this does not violate our trust,
since the capsule only relies on the host hub for availability, not
safety: data is encrypted before any network communication or disk
storage via the host hub. Based on the various data layer capsules we
have implemented, we have identified a set of common services they
require: \bi \item Raw Network Access via the BSD Socket API: When a
data layer
(such as a CCN capsule) requires network access, in order to avoid
placing a complex TCP/IP network stack within the capsule, the base
layer offloads the implementation to the host hub. For instance,
when the data layer invokes the {\it connect} BSD socket call, the
base layer marshals the parameters and conveys them to the host
hub. The host hub invokes the {\it connect} call using its network
stack (which need not be trusted) and returns a file descriptor (or
error) to the base layer which conveys it to the base layer. The
data layer can thus avail of the standard socket API. We have also
included the PolarSSL SSL/TLS library~\cite{polarssl:url} in the
base layer and modified the socket calls to use the base layer's
rather than using the standard system calls. We chose the PolarSSL
library since it is a significant smaller code base ($12$K) compared
to OpenSSL (over $200$K lines of code); this design decision can be
revisited if required. Thus, data layers can use the SSL library as
well; this is necessary to access any remote web service securely.
\item Time: The base layer can also request the host hub to retrieve
the current time from a remote trusted NTP server using SSL; this is
useful for supporting certain policies (for instance, expire the
data after a certain date).
\item Persistent Data: The host hub also offers a simple block-based
storage interface for reading and writing data persistently. None of
our capsules use this feature so far (since they live within a VM,
and are re-incarnated along with the VM image), but this
functionality may be required when large data sets are involved.
\end{itemize}
The base layer handles all requests from the host hub and implements
the hosting protocol and the invocation policy. It also interacts
with the trusted module to ensure that the security properties
hold. Thus, the data layer that implements the data-specific
interface is decoupled from the security mechanisms, and is much
easier to write. We now describe the components of the base layer.
\subsection{Capsule Hosting} \label{sec:design_hosting}
In order to meet our requirement of allowing flexible data hosting
policies, our capsule framework allows user to specify a {\it hosting
policy} which is then enforced by a {\it hosting protocol}. We first
discuss our hosting policy language and then discuss the hosting
protocol.
\noindent \textbf{Hosting Policy:} The requirement for our hosting
policy language is that it be expressive enough to accomodate typical
user requirements and be simple enough to be interpreted easily by the
framework. We arrived at a design that is a simplified version of
SecPAL~\cite{secpal:csf06}, an authorization policy language for
distributed systems. The simplifications were done so as to enable
easy interpretation. Our language is flexible enough to allow a user
to grant a service $S$ the right to host her data based on: (a)
white-list of services (b) trusted hardware/software modules available
on that service. (c) if a service $S'$ trusted by the user vouches for
the service $S$. This space of options is inspired from the typical
privacy options supported by websites. The flexibility to trust based
on the availability of a trusted modules is key; a user may chose to
trust any service that houses her data in a physically secure
co-processor, while it may only trust a select set of services to host
her data on a VMM or a TPM. We show an example below to illustrate the
flavor of our
language. \\\\
\colorbox{LightGray}{\parbox{1.0\textwidth}{ \texttt{Alice}
\underline{says} \textit{CanHost(\texttt{M})} \underline{if}
\textit{OwnsMachine(\texttt{Amazon},\texttt{M})},
\textit{HasTPM(\texttt{M})} \\
\texttt{Alice} \underline{says} \texttt{CA} \underline{can say}
\textit{HasTPM(\texttt{X})} \\
\texttt{Alice} \underline{says} \texttt{S} \underline{can say}
\textit{OwnsMachine(\texttt{S},\texttt{M})} \\
\texttt{Alice} \underline{says} \texttt{Amazon} \underline{can
say} \textit{TrustedService(\texttt{S})} \\
\texttt{Alice} \underline{says} \texttt{Amazon} \underline{can
say} \textit{CanHost(\texttt{M})} \underline{if}
\textit{TrustedService(\texttt{S})},
\textit{HasSecureCoprocessor(\texttt{M})},
\textit{OwnsMachine(\texttt{S},\texttt{M})} \\
\texttt{Alice} \underline{says} \texttt{CA} \underline{can say}
\textit{HasSecureCoprocessor(\texttt{X})} }}
~\\We use a typewriter font to indicate principals like Alice (the
user; a principal's identity is established by a certificate binding
its name to a public key), an underlined font to indicate language
keywords, and italics for indicating predicates (such as {\it CanHost}
and {\it HasTPM}). The first rule says that \texttt{Alice} allows any
machine $M$ to host her capsule provided Amazon acknowledges a machine
as its own, and the machine has a TPM. The second rule says that
\texttt{Alice} trusts \texttt{CA} to certify the public keys of a
machine's TPM. The third rule allows any service to certify a machine
as its own. The fourth rule allows Amazon to recommend any service $S$
as a trusted service. The fourth rule indicates Alice allows any
machine $M$ to host her capsule if (a) Amazon vouches for such a
service $S$ (b) $S$ asserts that $M$ is its machine (c) the machine
$M$ has a secure co-processor (since Alice is delegating the decision
to Amazon, she would like a more secure option than a TPM, so her
policy asks for a secure co-processor). The last rule indicates
\texttt{Alice} delegates her decision of which machines have secure
co-processors to the certification authority \texttt{CA}. This example
illustrates all features of our policy language; the simplification
from SecPAL is that we do not support recursive delegation.
A hosting request received at $C_U@[M,S]$ has three parameters: the
machine $M'$ to which the hosting transfer is requested, the service
$S'$ which owns machine $M'$, and a set of assertions $P_{HR}$. The
set of assertions $P_{HR}$ are presented by the requesting principal
in support of its request; it may include delegation assertions (such
as ``\texttt{S} \underline{says} \texttt{S'} \textit{CanHost(X)}'') as
well as capability assertions (such as ``\texttt{CA} \underline{says}
\textit{HasTPM(M')}''). When such a request is received by
$C_U@[M,S]$, it is checked against the policy $P(C_U@[M,S])$. This
involves checking whether the fact ``\texttt{U} \underline{says}
\textit{CanHost(M')}'' is derivable from $P(C) \cup P_{HR}$. If it is,
then the hosting protocol is initiated.
\textbf{Hosting Protocol:} The hosting protocol is responsible for the
forwarding of a capsule from one party to another. We treat the
initial transfer of a capsule from the user's machine to a TTP or
service as a hosting transfer as well. Thus, capsule $C_U@[M_U,U]$
denotes the capsule initially created and resident on the user's
machine. The hosting protocol is based on the Diffie-Hellman
key-exchange protocol, and has three steps: \bi
\item Step 1: $M \rightarrow M'$: $K_C$, $N$
\item Step 2: $M' \rightarrow M$: $K_{C'}$, $Attestation(M',\text{BL},N,K_C,K_{C'})$
\item Step 3: $M \rightarrow M'$: $ \{ C_U@[M,S] \}_{DHK} $
\end{itemize}
The only differences from the Diffie-Hellman protocol (involving the
keys $K_C, K_{C'}$ and the Diffie-Hellman secret $DHK$) are the use of
the nonce $N$ and the optional attestation that proves that the key
$K_{C'}$ was generated by the DC base layer code which, on input $N,
K_C$, produced output $K_{C'}$. This attestation is made along with an
attestation key $M'$ which is vouched off for by a certification
authority, guarantees freshness (since it is bound to $N$), and rules
out man-in-the-middle substitution on both the input and output (since
the attestation is bound to both $K_C$ and $K_{C'}$). At the end of
this three-message exchange, the base layer at $M'$ then de-serializes
the freshly transferred capsule $C'_{U}@[M',S']$ identified by the key
$K_{C'}$. At this point, the capsule $C'$ is operational; it shares
the same data as $C$ and is owned by the same user $U$.
\noindent \textbf{Consistency:} We note that when a user's data is
spread across several services in the form of several capsules, we
{\it do not} attempt to enforce any notion of consistency across them
{\it after} they have forked off; each of them is an independent
entity (much like the case today; if Amazon hands out a user's credit
card number to a third party, it does not necessarily update it
automatically on any changes by the user). Our architecture allows a
user or service to notify any descendant capsules and update them if
so desired, but this is not automatically done since it is not
possible to maintain both consistency and availability in the face of
partitions, and further, the overhead of synchronizing the multiple
copies would be high. However, {\it during} the process when one
capsule is forked off from another, we appropriately transfer policies
during the transfer so that the policies continue to hold.
\subsection{Capsule Invocation Policy} \label{sec:design_invocation}
An invocation policy allows the user to specify constraints on the
parameters to the capsule interface during invocations by various
principals. We support two kinds of constraints: stateless and
stateful.
Stateless constraints specify conditions that the argument to a single
invocation of a capsule interface must satisfy. For example, such a
constraint could be ``never charge more than 100 dollars in a single
invocation''. We support predicates based on comparison of numerical
quantities, along with any conjunction or disjunction
operations. Stateful constraints apply across several invocations; for
example, ``no more than 100 dollars during the lifetime of the
capsule''. We found such stateful constraints to be useful for
specifying cumulative constraints. It is very likely that users would
desire to specify a budget on their credit card number over a period
of time. An even more important use of a stateful constraints is in
specifying a query budget for differential privacy based
interfaces.
Thus, it is crucial for users to be
able to specify a query budget that limits the number of times their
data is accessed; this requires maintaining state of the number of
queries served so far. Once again, we aim for the policy language
supporting both these kinds of invocation constraints to be simple;
this language is very similar to the hosting policy language and we
present an example for a
CCN capsule below. \\\\
\colorbox{LightGray}{\parbox{1.0\textwidth}{ \texttt{Alice}
\underline{says} \textit{CanInvoke(Charge,\texttt{Amazon},A)}
\underline{if} \textit{LessThan(A,100)}\\
\texttt{Alice} \underline{says}
\textit{CanInvoke(Charge,\texttt{DoubleClick},A)} \underline{if}
\textit{$LessThan(A,Limit)$, Between(CurrentTime, `Jan 1st,
2010'',``Jan 31st, 2010'')} \underline{state} $(Limit=50,Update(A))$ \\
\texttt{Alice} \underline{says} \texttt{Amazon} \underline{can
say} \textit{CanInvoke(Charge,\texttt{S},A)} \underline{if}
\textit{$LessThan(A,Limit)$} \underline{state}
$(Limit=150,Update(A))$ }}
~\\The first rule says gives the capability ``can invoke upto amount
A'' to Amazon. This allows Amazon to call the charging interface with
parameter $A$ and the constraint $A<100$ checks that the amount is
less than $100$. The second rule shows a stateful example; the
semantics of this rule is that DoubleClick is allowed to charge upto a
cumulative limit of $50$ during Jan 2010. To implement this policy, we
introduce a state variable called {\it Limit} which is set to $50$
initially by the user. The predicate {\it Update(Limit,A)} is a
stateful predicate that indicates if this rule is matched, then the
{\it Limit} should be updated with the amount $A$; thus, when a rule
is matched with a {\it state} keyword, it is removed from the policy
database (the database of assertions), the state variable ({e.g., } {\it
Limit}) is updated in the rule, and the new rule inserted into the
database. This usage idiom is similar to SecPAL's support for RBAC
dynamic sessions; we have added the {\it state} keyword to allow
retention of state between authorization requests. The alternative is
to move this state outside the SecPAL policy, and house it within the
capsule functionality; we make the design choice of making it explicit
in the policy itself to ensure that the policy implementation is not
split across SecPAL and the capsule implementation of the
interface. Though this choice implies that the policy database is
updated over time, there are no undesirable consequences with allowing
such dynamic changes. The third rule is very similar to the second
rule; the interesting thing to note here is that this rule is matched
for any principal to which Amazon has bestowed invocation rights. This
means that the limit is enforced across all those invocations; this is
exactly the kind of behavior a user would
expect.
\textbf{Transfer of Invocation Policies:} We now discuss how this
invocation protocol interacts with the hosting protocol discussed
earlier. During a hosting protocol initiated from $C_U@[M,S]$ to
$C'_U@[M',S']$, $C$ should ensure that $C'$ has suitable policy
assertions $P(C')$ so that the user's policy specified in $P(C)$ is
not violated. First, any policy statements $P_{HR}$ specified by $S'$
during the hosting request need to be added to $P(C')$ to record the
fact that this newly derived capsule operates under that context.
Second, any stateful policies need to be specially handled. For
example, consider the third rule in the invocation policy in
Section~\ref{sec:design_invocation} which requires that the total
budget across all third parties that are vouched for by Amazon is
$100$ dollars. If this constraint is to hold across both $C'$ and any
future $C''$ that might be derived from $C$, then one option is to use
$C$ as a common point which ensures that this constraint is violated.
However, this requires any transferred capsule $C'$ to communicate
over the network with $C$ upon invocations. Instead, in order to allow
disconnected operation, we use the concept of
exo-leasing~\cite{exoleasing:mdw08}.
In the exo-leasing concept, the constraint ({e.g., } budget) is broken up
into sub-constraints ({e.g., } sub-budgets) so that if the sub-constraints
are enforced, the parent constraint is automatically enforced. In the
capsule context, {\it decomposable constraints} (such as budgets,
number of queries answered so far) from $AC(C)$ are split into two
sub-constraints; the original constraint in $AC(C)$ is updated with
one sub-constraint, and the second sub-constraint is added to
$AC(C')$. In our example, the assertion ``\texttt{Alice}
\underline{says} \texttt{X} \textit{can invoke Charge A State Limit
100} \underline{if} \texttt{Amazon} \underline{says} \textit{I vouch
for X}, \textit{ [ $ ( A < Limit ) $ , Update (Limit,A) ] }'' would
be transformed into ``\texttt{Alice} \underline{says} \texttt{X}
\textit{can invoke Charge A State Limit 75} \underline{if}
\texttt{Amazon} \underline{says} \textit{I vouch for X}, \textit{ [ $
( A < Limit ) $ , Update (Limit,A) ] }'' and ``\texttt{Alice}
\underline{says} \texttt{X} \textit{can invoke Charge A State Limit
25} \underline{if} \texttt{Amazon} \underline{says} \textit{I vouch
for X}, \textit{ [ $ ( A < Limit ) $ , Update (Limit,A) ] }''.
\subsection{Capsule Data Transformation} \label{sec:design_transformations}
In exploring the use of capsules in various application scenarios, we
found it useful to allow users to perform transformations on their
data post-installation, especially when the data is of an aggregate
nature ({e.g., } web history). Providing users control over the
transformations on their data is distinctly different from providing
control over invocation; the latter controls operations invoked over
the data, and the former controls the data itself. We refer to a
capsule whose data is derived from a set of existing capsules by some
transformation as a derivative capsule. We support two data
transformations: filtering and aggregation.
\noindent \textbf{Filtering:} A derivative capsule obtained by
filtering has a subset of the data of the originating capsule; for
instance, only the web history in the last six months, instead of the
entire year. A capsule that supports such tranformations on its data
exports an interface call for this purpose; this is invoked alongside
a hosting protocol request so that the forwarded capsule contains a
subset of the originating capsule.
\noindent \textbf{Aggregation:} This allows the merging of raw data
from mutually trusting users of a service, so that the service can use
the aggregated raw data, while the users still obtain some privacy
guarantees due to aggregation. We refer to this scenario as ``data
crowds'' (inspired by the Crowds anonymity
system~\cite{crowds}). Examples of usage include ads clicking behavior
of users or online product purchase history. In the first, Google
(say) may use the aggregate ad click behavioral history to fine-tune
their targeted advertising algorithms. In the second, the latter could
be used by Amazon to build statistical models for recommendation
systems.
To enable this, a user $U$ instructs her capsule $C_U@[M,S]$ to
aggregate her data with a set of capsules $C_{U'}[M',S]$ where $U'$ is
a set of users that she trusts (this trust is necessary; otherwise, if
her data was merged with data belonging to fake users, the privacy
guarantees would be much poorer). These set of users $U'$ form a data
crowd. We envision that a user $U$ can discover such a large enough
set of such users $U'$ by mining her social networks (for instance).
During capsule installation, each member $U$ of the crowd $C$ confides
a key $K_C$ shared among all members in the crowd to their
capsule. During installation, a user $U \in C$ also notifies the
service of her willingness to be aggregated in a data crowd identified
by $H(K_A)$. $S$ can then identify the set of capsules $C_A$ belonging
to that particular data crowd using $H(K_A)$ as an identifier. The
service can, at this point, then determine any kind of aggregation
strategy. For instance, in order to reduce network overhead, it could
request all the capsules stored on a rack in its data center to
aggregate with each other first, and then ask the resultant capsule to
aggregate with the resultant capsules from other racks. Note that the
aggregation protocol does reveal the size of individual user data;
such side-channels can be avoided by padding data if desired.
From the capsule's perspective, aggregation is a pair-wise operation;
a capsule $C_U$ is requested to merge with $C_{U'}$, and these
mergings are appropritately staged by the service. During the
aggregation operation of $C_U$ with $C_{U'}$, capsule $C_U$ simply
encrypts its sensitive data using the shared data $K_A$ and hands it
off to $U'$ along with its owner's key. During this aggregation, the
resultant capsule also maintains a list of all capsules merged into
the aggregate so far; capsules are identified by the public key of the
owner sent during installation. This list is required so as to prevent
duplicate aggregation; such duplicate aggregation can reduce the
privacy guarantees. Once the count of source capsules in an aggregated
capsule exceeds the user-specified constraint $C_{min}$, the aggregate
capsule can then reveal its data to the service $S$. This scheme
places the bulk of the aggregation functionality upon the service;
this is ideal since it gives the service freedom to optimize the
aggregation, while at the same time requiring the capsule to only
implement a simple pair-wise aggregation function.
\section{Implementation} \label{sec:implementation}
Our implementation supports three deployment models: TTP, client-side,
and Xen-based co-location. In the future, we plan to support TPMs and
secure co-processors using standard implementation techniques using as
late launch ({e.g., } Flicker~\cite{flicker:eurosys08}). We implement four
capsule instances atop this framework, one per usage idiom: a stock
trading capsule, a credit card capsule, a targeted ads capsule, and a
provenance capsule. We describe some of the details of our framework
followed by a description by each capsule.
\subsection{Capsule Framework} \label{sec:impl_framework}
\begin{table*}[t]
\begin{scriptsize}
\renewcommand{\arraystretch}{1.15}
\caption{Capsule Implementation: LOC Estimates of Modules}
\label{tab:tcb}
\begin{center}
\begin{tabular}{|l|l|l|}
\hline\hline
Software Module & LOC & Functionality \\
\hline\hline
Base Layer & 6K & Implements installation, invocation, and policy
checking. \\
\hline
Data Layer & Stock ($340$), Ads ($341$) & Implements functionality
specific to data \\
& CCN ($353$), Provenance ($156$) & \\
\hline
PolarSSL & 12K & Light-weight cryptographic and SSL library used for
embedded systems. \\
\hline
XenSocket & 1K & Fast inter-VM communication using shared memory and
event notification hyper-calls. \\
\hline
Trousers (optional) & 10K & Library for TPM functions. Can be removed
from TCB by accessing TPM directly. \\
\hline\hline
\end{tabular}
\end{center}
\end{scriptsize}
\end{table*}
For TTP deployment, the network serves as the isolation barrier
between the capsule and the service. For co-location deployment, we
used the Xen virtual machine~\cite{xen_url} as the isolation
mechanism; the web service code and the data capsule are run inside
different virtual machines thus providing protection against software
compromise. We ported and extended XenSocket~\cite{xensocket:middle07}
to our setup (Linux 2.6.29-2 / Xen 3.4-1) in order to provide fast
two-way communication between the web service VM and the capsule VM
using shared memory and event channels.
The base layer implements the Diffie-Hellman based installation
protocol and invocation protocols. It implements policy checking by
converting policies to DataLog clauses, and answers query by a simple
top-down clause resolution algorithm (described in
SecPAL~\cite{secpal:csf06}; we could not use their implementation as
such since it required .NET libraries). We use the PolarSSL library
for embedded systems for light-weight cryptographic functionality. We
use a TPM, if available, to verify that a remote machine is running
Xen using the Trousers library; this TCB can be reduced by invoking
the TPM directly rather than via the library. Note that we only use a
TPM to verify the execution of Xen; we still use Xen as the isolation
mechanism. The host hub uses the \textit{libevent} event library (the
host hub is not part of the TCB). We use 2048-bit RSA public keys (for
identifing capsule instances, data owners, services), 256-bit AES/DES
keys (used internally when encrypting data using public keys), and
HMAC/SHA-2 hash functions (for signatures).
We now estimate the TCB of our Xen based capsule. This consists of the
Xen VMM, Dom0 kernel, and our capsule implementation.
Table~\ref{tab:tcb} shows the LOC estimates of the various modules in
our capsule implementation. Our capsule implementation currently has
two dependencies which we plan to remove in the future. First, our
capsule VM boots up the capsule on top of Linux 2.6.29-2; however, the
capsule code does not utilize any Linux functionality, and can be
ported to run directly atop Xen or MiniOS (a bare-bones OS distributed
with Xen). Second, the capsule uses the memory allocation
functionality by \textit{glibc} and some functionality from the STL
library; we plan to include a custom memory allocation module to
substitute \textit{glibc} and to hand-implement custom data structures
to avoid this dependence.
We also plan to incorporate mechanisms focussed on improving VMM-based
security (such as removing Dom0 functionality from the TCB by using
disaggregation techniques~\cite{xendisaggregation:sigops08}); such
mechanisms are orthogonal to our framework. We have yet to incorporate
several performance optimizations discussed in literature for
improving performance while providing isolation. One technique that is
pertinent is from Sharif~{et al.}~\cite{secureinvm:ccs09}.
\cite{secureinvm:ccs09} allows us to run the capsule within the same
VM as the service, while still guaranteeing code and data isolation;
this is achieved by having the VMM use hardware paging functionality.
\cite{secureinvm:ccs09} demonstrates an order-of-magnitude gain in
performance using this technique, which we hope to implement in our
framework. We now discuss the implementation of the base layer's
internals followed by the sample capsules we implemented.
\subsection{Capsule Hosting} \label{sec:impl_hosting}
We currently use a policy language which supports three base types:
principal names, strings, and integers. An argument in a predicate can
either be a constant or a variable, and is annotated with its type,
{e.g., } VPRINCIPAL(XP) denotes a variable XP which stands for a
principal. A policy in our language is represent as a 5-tuple $(P, F,
CF, C, S)$ where $P$ is name of the principal issuing this statement
(such as \textit{CPRINCIPAL(ALICE)} which denotes a constant ALICE of
type PRINCIPAL), $F$ is the statement made by $P$ ({e.g., } \textit{FACT\{~
PRED [~INVOKE, ~VPRINCIPAL(XP), ~VNUM(AMT), ~VSTR(ACCOUNT) ~] ~\}}
), $CF$ and $C$ is a set of conditional facts and constraints which
are required to be satisifed if $F$ is to be true, and $S$ records any
state variables. In order to verify whether a particular statement is
derivable from a set of policy statement, we translate all statements
to DataLog clauses, and then use a simple top-down resolution
algorithm (described in SecPAL~\cite{secpal:csf06}). We defer a
detailed description of our policy language and resolution algorithm
to the technical report due to space constraints.
\subsection{Invocation Policies} \label{sec:impl_invocation}
An invocation request has four parameters: $S'$, the service making
the request (not necessarily the service $S$ hosting the capsule), the
request type $R$ ({e.g., } ``Charge''), the arguments to the request $A$
({e.g., } ``10'' dollars), and a set of supporting assertions $P_{IR}$. To
service such a query, the base layer of $C$ checks to see if the fact
``\texttt{U} \underline{says} \texttt{S'} \textit{can invoke $R$
$A$}'' or the fact ``\texttt{U} \underline{says} \texttt{S'}
\textit{can invoke $R$ $A$ State $S$}'' is derivable from $P(C) \cup
P_{IR}$. If there is no proof for such a query, the request is
denied. If there is a proof, the base layer retrieves any policy
specification $ac_{S}$ in the proof with a {\it State} keyword. For
all such assertions $ac_{S}$, the corresponding state $S$ is also
extracted. The invocation is then dispatched by the base layer to the
data layer; after the successful completetion of the request, the
state $S$ is updated to $S'$ as per the {\it Update} rules. At this
point, the old specification $ac_{S}$ is removed from $P(C)$, and a
new specification $ac'_{S'}$ is inserted into $C_U@[S,M]$. This
ensures that the state is updated as required. To avoid race condition
issues, for now, the base layer waits for the invocation to complete
before serving any other requests. Since the database of policies may
change from one invocation to the next, we do not cache any
intermediate facts that are derived duing one invocation for
subsequent ones; thus, we avoid any inconsistency issues.
\subsection{Data Transformations} \label{sec:impl_transformations}
Filtering is easy to implement; a capsule $C$ supports an interface
call {\it Filter} which takes a filtering criterion as an argument,
and produces a capsule $C'$ per the filter. Filtering can be invoked
in conjunction with a hosting transfer; alternatively, the policy can
insist that it should only be invoked by the owner of the capsule. For
the latter case, every capsule maintains the public key of its owner
$K_C$ (which is sent during the installation and hosting protocols),
and authenticates requests for data transformation using this key.
Aggregation is a more expensive operation since it involves
coordination of $U$'s capsule with multiple other capsules. To enable
secure aggregation, our aggregation protocol operates in two stages;
the first stage is performed off-line by a set of users $U_A$ who have
decided to cooperatively perform aggregation of their data stored on a
common service $S$, the second is the aggregation itself performed at
the service site.
During the setup phase, a set of users $U_A$ establish a shared key
$K_A$ amongst themselves; a user $U$ can securely discover such a set
of willing users $U_A$ and establish a key $K_A$ using her social
network. All the user's implementations support an aggregation call in
their interface. At the end of this phase, every user $U \in U_A$
notifies her own capsule $C_U@[M,S]$ of $K_A, A_{min}$. Here $A_{min}$
represents the minimum number of capsules that should be aggregated
before the data is released; we return to this later. Note that this
notification step can be performed during the installation
itself. Further, this set of users $U_A$ can be the same across
multiple services. Thus, a user need only establish this $U_A, K_A$
once; this set can be re-used across multiple services.
The second phase performs the actual aggregation itself, and is
performed with the coordination of the service. During installation,
a user $U \in U_A$ notifies the service of her willingness to be
aggregated in a data crowd identified by $H(K_A)$. $S$ can then
identify the set of capsules $C_A$ belonging to that particular data
crowd using $H(K_A)$ as an identifier. The service can, at this point,
then determine any kind of aggregation strategy. For instance, in
order to reduce network overhead, it could request all the capsules
stored on a rack in its data center to aggregate with each other
first, and then ask the resultant capsule to aggregate with the
resultant capsules from other racks.
From the capsule's perspective, aggregation is a pair-wise operation;
a capsule $C_U$ is requested to merge with $C_{U'}$, and these
mergings are appropritately staged by the service. During the
aggregation operation of $C_U$ with $C_{U'}$, capsule $U$ simply
encrypts its sensitive data using the shared data $K_A$ and hands it
off to $U'$ along with its owner's key. Thus, the shared key $K_A$
serves as a security token; only another capsule belonging to the same
crowd will be able to decrypt the data. During this aggregation, the
resultant capsule also maintains a list of all capsules merged into
the aggregate so far; capsules are identified by the public key of the
owner sent during installation. This list is required so as to prevent
duplicate aggregation; such duplicate aggregation can reduce the
privacy guarantees. Once the count of source capsules in an aggregated
capsule exceeds the user-specified constraint $A_{min}$, the aggregate
capsule can then reveal its data to the service $S$. This scheme
places the bulk of the aggregation functionality upon the service;
this is ideal since it gives the service freedom to optimize the
aggregation, while at the same time requiring the capsule to only
implement a simple pair-wise aggregation function.
\subsection{Sample Capsules}
The broad guideline behind our sample capsule implementations is to
model the essence of the application scenario in realistic enough
fashion so that experimental comparative analysis of different
deployment options is feasible.
\noindent \textbf{Stock Trading Capsule:} We model our stock capsule
functionality after the automatic trading functionality in a popular
day trading software, Sierra Chart~\cite{sierrachart:url}, which is
representative of the complex mechanisms used by financial traders.
We chose one specific feature of Sierra Chart for illustration: make
trades automatically when an incoming stream of ticker data matches
certain conditions. Our implementation follows the the Sierra Chart
technical manual's description~\cite{sierrachart:url}.
This capsule belongs to the query idiom (Section~\ref{sec:arch_interface})
and exports a single function call of the form
\textit{TickerEvent (SYMBOL, PRICE) } and returns a \textit{
ORDER(``NONE'' /` `BUY'' / ''SELL'', SYMBOL, QUANTITY) } indicating
whether the capsule wishes to make a trade and of what quantity. The
capsule allows the user to specify two conditions: a BUY (ENTRY)
condition and a BUY(EXIT) condition. A condition is an arbitrarily
nested boolean predicate with operations like AND, OR, and NOT, and
base predicates that consist of numerical variables, comparison
operations ($>, <, =$), and numerical constants. The numerical
variables include: LP (the latest price of stock), MA (moving average
of price), POS (position: amount of stock purchased by the software),
and POSAV (average price of stock purchased already). As an example, a
BUY(ENTRY) condition could be \textit{$LP>MA$} and a BUY(EXIT)
condition could be $OR(AND(>0,LP<=POSAV-1), AND(POS>0, LP>=POSAV+2))$.
This means that the system purchases upto $Q$ (a user-chosen
parameter) everytime the last price exceeds the moving average, and
this stock is sold off either when: (a) last price is less than the
average price the BUY Entry was filled at minus 1 full point. (b) the
last price is greater than the price the Buy Entry was filled at plus
2 full points.
Our implementation of boolean predicate matching is straight-forward;
we do not apply any query optimization techniques like common
sub-expression matching from the database community. Thus, our
performance results are only for comparative purposes.
Sierra Chart currently executes at the user's desktop on a feed of
incoming stock symbols (similar to the TTP case). In the Xen case, the
user capsulizes her BUY(ENTRY) and BUY(EXIT) conditions at her broker.
The only information revealed by the capsule is any trades it makes;
the strategy is itself secret. It is not possible to hide the trades
from the broker, thus any leaks of the strategy through the trades
cannot be avoided; we assume that this leak is tolerable. We expect
that the re-factoring of the application service at the broker site
would not require significant effort since the interface is very
simple; since we do not have access to any broker site code, we can
only estimate this effort.
\noindent \textbf{Credit Card Capsule:} We implemented a credit card
capsule that interacts with the \texttt{authorize.net} gateway to
implement charges, and modified an open-source shopping cart
application (Zen) to interact with the capsule to implement charging.
As discussed in Section~\ref{sec:arch_interface}, this capsule
supports a single interface call, \textit{Charge(Amt, MercAcct)}, and
returns a confirmation code to the service. This capsule illustrates
the flexibility of the framework with respect to invocation policies
and matches the proxy idiom. This functionality is offered by some
banks already; we only implement and evaluate this capsule to measure
the cost of our framework in emulating this functionality in a more
general fashion. We implemented both stateless policies (allow only
invocations for less than $X$ dollars) and stateful policies (maintain
a per-month budget of $X$ dollars).
\noindent \textbf{Targeted Advertising Capsule:} We implemented two
capsules for targeted advertising both of which belong to the
analytics idiom (Section~\ref{sec:arch_interface}). They both store
the user's web browsing history and are updated periodically (say, on
a daily basis). The first capsule is used by a service to serve
targeted ads when a user visits a specific web page. The second is
used by the service to build long-term models related to the affinity
of users with specific profiles to specific advertisements. These
models are used, for instance, by Google, to fine-tune their targeted
advertising methods. Both capsules support a simple interface call
that is used by the data owner to update her web browsing history; we
now detail the other interface calls exported by these capsules below.
\textit{Serving targeted ads:} This capsule supports two possible
interfaces: \textit{ChooseAd(List of Ads, Ad Categories)} and
\textit{GetInterestVector()}. In the first case, the capsule selects
the ad to be displayed by using a procedure similar to those used in
web services today (per the description in
Section~\ref{alg:targeted_ads_1}). In the second, the capsule extracts
the keyword vector of the user from her browsing history, and then
computes the true normalized interest vector $U = \{ U_c \}$ based on
the user's data $(D_U = \{W_1,W_2,\cdots\})$ where $\{W_i\}$ is the
user's web visit history. Here, $0 \leq U_c \leq 1$ reflects the
interest of the user in category $c$ and $\Sigma_{c} U_c = 1$. The
capsule then obfuscates this vector per a generic differential privacy
prescription. We obfuscate each element $U_c$ of this vector $U$
individually, and then re-normalize the vector $U$. The $L_1$
sensitivity of this function, denoted by $\Delta(U_c)$, is $1/V$ where
$V$ is the total number of visits by the user to various websites
since if the user visits one particular different website, the count
of the number of visits in one specific category $c$ can at most
increase or decrease by $1$. We insist that the user visit a minimum
number of websites $V_{min}$ before the capsule would start serving
ads. Thus, the sensitivity $\Delta(U_C)$ of this function is
upper-bounded by $1/V_{min}$. The obfuscated value $U'_C$ of $U_c$ is
chosen per the distribution function $Pr[U'_c = a] \propto exp (
\frac{ - \parallel (U_c - a) \parallel }{\sigma} )$ where \textit{exp}
is the scaled symmetric exponential distribution function and $\sigma$
is the variance of the distribution. This function guarantees
$\epsilon$-differential privacy for $\epsilon = \Delta(U_c)/\sigma
\geq 1 / (V_{min}\sigma)$.
This capsule also makes use of a stateful policy in the differential
privacy case to record the number of queries made so far. The amount
of information leaked by interactive increases linearly with the
number of queries made~\cite{differentialprivacy:icalp06}; thus it is
necessary for users to specify a private budget $Q_{max}$ and the
framework ensures that this budget is enforced by using stateful
policies (in a manner similar to the credit card budget). As discussed
earlier, these two calls provide different privacy guarantees. The
first leaks no information; the selected ad is sent to the user's
computer where a client-side component displays the ad. In the second,
one implication of the $\epsilon$-differential privacy guarantee is
that even if the service knows all but one of the websites visited by
the user, the ability of the service to deduce the one private website
is limited by $\epsilon$. In particular, the distributions of the
output for various possibilities of that one private website are
point-wise $\epsilon$ close. We choose a typical value $\epsilon =
0.01$. We
\textit{Building long-term models:} The second capsule allows the
service access to the websites visited by the user and the particular
advertisements clicked by the user when visiting a website (thus this
data is more detailed than the data used for targeted
advertising). This information is used to refine the service's
targeting algorithms. For this purpose, it suffices for the service to
obtain aggregate data from a large population of users; we can thus
leverage our aggregation functionality. Individual user data is
represented as $D_U = \{ W_i \}, \{ (P_i, A_i, C_i) \} $ where $\{ W_i
\}$ represents all websites visited by the user, and $\{ (P_i, A_i,
C_i) \}$ represents how the user responded to ads displayed on a
publishing websites $P_i$ that participates in the advertising service
(such as Google AdSense). $C_i$ is $0/1$ depending on whether the user
clicked on ad $A_i$ shown to her when visiting website $P_i$. This
information is aggregated across all users by our aggregation protocol
and delivered to the service. To do so, the users distribute a shared
secret key amongst themselves which is then used as an authentication
token amongst the capsules to authenticate themselves. Currently, we
simply aggregate this information and return it to the service; if
desired, differential privacy techniques can be applied here as well.
In the future, we plan to merge these capsules with a client-side
browser plugin to build a full-fledged system so that the plugin can
automatically update the server-side capsule without any manual
intervention from the user. This plugin would also interact with the
capsule to obtain and display the selected ad in the case of the {\it
ChooseAd} interface. If it is necessary for the server to maintain
ad impression counts (for charging advertisers), then information
about the selected ad can be sent after aggregation or via a broker to
preserve anonymity; our system can borrow such functionality from
PrivAd~\cite{privad:hotnets09} and
Adnostic~\cite{adnostic:ndss10}. Alternatively, our aggregation
functionality may be used for this to obtain aggregate ad impressions
as well.
\noindent \textbf{Provenance Capsule:} The final capsule we
implemented belongs to the provenance idiom
(Section~\ref{sec:arch_interface}) and models a editable
document. It supports three calls: \textit{Get()},
\textit{Insert(Position,String)}, and \textit{Delete(Position,
NumChars}. These calls suffice for any sequence of edits; we assume
an editor program propagates any edits made by colloborators as calls
to this interface so that the capsule is altered. \textit{Get()}
returns the current version of the document, while {\it Insert()} and
{\it Delete()} can be used by, say, an editor, to convey changes made
by a user to the document. The capsule retains a log of all
invocations, thus binding each modification to the invoker. This
illustrates that the capsule framework can also be used to provide
properties such as accountablity in addition to privacy.
\noindent \textbf{Other Possible Capsules:} There are also services
where the some private data may be required when the service is
disconnected from the Internet during invocation of the capsule so
that a trusted third party solution is not feasible. One example is a
service like CommonSense~\cite{commonsense:url} where sensors are worn
by users that record information about the user. This information may
be processed by proprietary algorithms (for example, to monitor blood
pressure); such algorithms have to be implemented at the sensor itself
because the user may not be connected to the Internet at all times. In
the future, we hope to implement this capsule.
\section{Evaluation} \label{sec:evaluation}
We evaluate our framework based on four capsule instances (one per
usage idiom): stock trading capsule, credit card capsule, targeted ads
capsule, and provenance capsule. We omit the provenance capsule's
results since they reflect similar trends as for the other
capsules. For the three capsules we evaluated, we present three
scenarios: (A) the base case where the service interacts with the raw
data with no access guarantees. (B) the TTP case, where the service
interacts over the network with the user's capsule hosted at a
TTP. Since a TTP deployment is equivalent in terms of performance in
most respects to a client-side capsule, we discuss both of them
together, and mention any differences specifically. (C) the co-hosting
case, where the service interacts with the data capsule housed in a
Xen VM.
Our goal in this section is two-fold. Our first goal is to to evaluate
\textit{the cost of privacy} as the performance impact of the
client-side / TTP deployment and co-location deployment options, as
compared to the base case. Our metric of success here is hard to
quantify in terms of absolute numbers; we only wish to show that this
cost does not render our framework impractical. The second goal is to
compare \textit{the various deployment options} amongst themselves to
determine the ideal deployment for each scenario.
We evaluate the performance for these scenarios along three
dimensions: the setup cost (when the data or data capsule is
transferred to the service), the invocation cost (when the service
accesses the user's data or data capsule), data transformation costs
(the cost of operations such as aggregation). The term {\it cost}
includes latency, network bandwidth consumption, and storage cost. Of
these, the invocation cost is borne every time the user's data is
accessed, and is thus the primary metric of comparison. The setup cost
is incurred during the initial and subsequent transfers of the user's
data, while the transformation cost, thought not as frequent as
invocation, may be significant if the transformation involves data
belonging to large numbers of users (such as aggregation).
Our test server is a $2.67$ GHz quad core Intel Xeon processor with
$5$ GB memory on which Xen, the service code, and the capsule code
run. A desktop (Intel Pentium 4 $1.8$ GHz processor, $1$ GB memory),
on the same LAN as the server, served as a TTP (or client). The
bandwidth between the server and this desktop was $10$ Gbps and the
round-trip about $0.2$ milli-seconds. To simulate a wide-area network
between the client/TTP and the service, we used
DummyNet~\cite{dummynet:url} which artificially delays packets by a
configurable parameter; by default, the round-trip is $10$ ms (typical
wide-area latencies from home users to well-provisioned services like
Google are about $30$ ms; we chose $10$ ms under the conservative
assumption that a TTP might have faster connectivity than average).
\subsection{Framework Benchmarks}
We first evaluate our framework independent of any specific kind of
data. For these measurements, we used a dummy data layer that does not
do any processing; it simply accepts a invocation payload of a
specific size and returns a response of specific size to help in
benchmarking.
\begin{table*}[t]
\begin{scriptsize}
\renewcommand{\arraystretch}{1.15}
\caption{Invocation Cost ($1$ KB sensitive data, $1$ KB invocation
request, $1$ KB invocation response )}
\label{tab:invocation_cost}
\begin{center}
\begin{tabular}{|l|l|l|l|l|}
\hline\hline
Scenario & Latency (micro-seconds) & Bandwidth &
In-Memory Storage &
On-Disk Storage \\
& (Mode, $50\%$-ile range, $95\%$-ile range) & (bytes) & (KB) & (KB) \\
\hline\hline
Base Case & $471.5~[379,496]~[373,532]$ & NA & 1 (data) & 1 (data) \\
\hline
TTP & $24796~[24718,24867]~[24525,25033]$ & $3336$ & 120 (capsule) + 1 (data) & 120 (capsule) + 1 (data) \\
\hline
Xen Capsule & $1237~[1097,4203]~[942,4617]$ & NA & 120 (capsule) + 1 (data) & 120 (capsule) + 1 (data) \\
\hline\hline
\end{tabular}
\end{center}
\end{scriptsize}
\vspace{-0.3in}
\end{table*}
\textbf{Invocation Cost:} The invocation costs for a fixed payload
size of 1 KB are shown in Table~\ref{tab:invocation_cost} (the
invocation request and response are both $1$KB). For the base case, we
assume that the service is architected as an server-side application
that interfaces with a database process to access the user's data. We
assume that the database process runs on the same machine as the
server-side application (ignoring any transfer costs) and that the
user's data is in memory (ignoring any costs of retrieving the data
from disk). For the TTP case, the web service interacts with the
capsule over the network. For the Xen capsule, the web service process
accesses the data via the host hub module which invokes the capsule
housed in a different VM instance. We assume that the capsule VM is
housed on the same machine as the service VM (similar to the base
case). We measured the latency as the time required for the process
storing the user's data to respond to an invocation by the web
service; for this benchmark, we fixed the message sizes for the
invocation request and response to be the same in all three scenarios.
The bandwidth consumed is measured using \textit{tcpdump} and includes
packets in both directions as well as TCP headers. The in-memory
storage cost and on-disk storage cost is measured directly. These
results are reported in the table; all results are reported over $100$
runs. For latency, we report the median, the $50\%$ confidence
interval, and the $95\%$ confidence interval, since it is somewhat
variable; we report only the median for bandwidth since we found it to
be much less variable.
Though the latency via Xen is about $800$ micro-seconds worse compared
to the base case, it is still significantly lower compared to the TTP
case. We found considerable variation ranging from $900$ micro-seconds
to $4000$ micro-seconds with a median of $1237$ micro-seconds; we
believe this is related to the Xen scheduling algorithm which may
delay the execution of the capsule VM. We plan to use techniques based
on hardware virtualization~\cite{secureinvm:ccs09} that allow for
order-of-magnitude reductions in overhead by executing the capsule
inside the service VM while still guaranteeing VM-like isolation
(since capsules require no access to service VM state except via the
interface, the ``semantic gap'' issue does not arise).
In the TTP case, the latency is primarily due to two round-trips (one
for exchange of TCP SYN and SYN ACKs, and another for the invocation).
The TTP option consumes $3.3$ KB worth of network traffic per
invocation, as opposed to the Xen co-location option and the base case
where the invocation is local. The invocation request and the response
are $1$ KB each: a constant overhead of about $1$ KB is incurred since
our protocol sends invocation requests and responses over RPC/HTTP
using \textit{libevent}. In both the TTP and Xen case, the capsule
code has to be stored in memory and disk; in our current un-optimized
implementation, this code is about $120$ KB. The LOC of the capsule in
this case is only $111$ LOC; our current compilation technique bundles
the glibc library as well which leads to this additional overhead. In
practice, we believe that a single capsule implementation will be
shared by several users (for instance, supplied by Symantec); thus a
single copy of the capsule code need not be maintained for every
user.
We now present a component-wise breakdown of the cost of a invocation
request; recall that a request is made by the service to the host hub
which then relays it to the Xen capsule. In this experiment, the
average invocation latency was $1399$ micro-seconds from the service's
perspective. Of this, $101$ micro-seconds is due to host-hub
processing, and $768$ micro-seconds due to processing at the Xen
capsule; thus, $530$ micro-seconds was due to communication overhead
(between the service and host hub, and between the host hub and the
capsule). Within the Xen capsule, $165$ micro-seconds was spent in
verifying the invoker's signature and $109$ micro-seconds for policy
verification; the remaining $494$ micro-seconds was due to the
processing of the $1$ KB invocation payload and response. In
conclusion, communication overhead is about $35\%$ of the total
latency; the rest is primarily due to processing within the Xen
capsule (which may be slower since it is run in a guest VM).
\begin{figure}
\vspace{-0.1in}
\centering
\mbox{
\begin{tabular}{c}
\psfig{file=results/exp2latency.pdf,height=2in}\\
{(A)}
\end{tabular}
\begin{tabular}{c}
\psfig{file=results/exp_stock_multiple_latency.pdf,height=2in}\\
{(B)}
\end{tabular}
}
\vspace{-0.15in}
\caption{(A) Invocation Costs under different payload sizes
(of request and response) for two different scenarios (base
case, xen capsule): Latency (Y-Axis) vs Payload size
(X-Axis, log scale). The corresponding costs for the TTP
capsule are:
$24.4$ ms ($0.25$ KB), $24.7$ ms ($0.5$ KB), $25$ ms ($1$
KB), $48.1$ ms ($4$ KB), $48.4$ ms ($8$ KB), $72.6$ ms ($16$
KB), $108.1$ ms ($32$ KB). (B) Invocation Costs for stock
capsule under two different scenarios (base case, TTP, xen
capsule): Latency (Y-Axis) vs Number of Ticker Events
(X-Axis). The corresponding costs for the TTP capsule are:
$72$ ms ($5$ invocations), $132$ ms ($10$), $612$ ms ($50$),
$1.2$s ($100$), $6$s ($500$), $12$s ($1000$). }
\label{fig:invocation_cost}
\vspace{-0.2in}
\end{figure}
\textbf{Varying Payload Sizes:} To examine the variation of the
invocation cost across different payload sizes, we plotted the latency
as a function of payload size (varied in multiples of $2$ from $256$
bytes to $32$ KB) in Figure~\ref{fig:invocation_cost} (A). In the
plot, the ``cross'' represents the median, the square box represents
the $50\%-$ile, and the lines around the cross represent the
$95\%-$ile.
The plot only shows the base case and Xen case; since the TTP
latencies are much higher, those are listed in the figure's caption.
This plot shows that the overhead added by Xen as a percentage of the
base case declines, while the network transfer time increases the
latency for the TTP case.
Once again, we note that the variance in the Xen latency is
substantial, especially at low invocation sizes. The bandwidth
consumption for the TTP case varies roughly linear with the size of
the payload, and are $1.7$ KB (for invocation of $256$ bytes size),
$10$ KB ($4$ KB), and $72$ KB ($32$ KB); of course, in the Xen case
and the base case, no network communication is involved for
invocation.
\textbf{Setup Cost:} We now discuss the setup costs in the three
scenarios (which are similar to the tranfer costs from one service to
another) for sensitive data of size $4$ KB. In the base case, the
latency is $13$ ms (primarly due to the $10$ ms round-trip) and the
bandwidth consumed is $6$ KB (the data size of $4$ KB, plus
\textit{libevent} overhead, plus TCP headers and acknowledgement
packets). In both the TTP and Xen case, the capsule code is sent from
the client to the TTP / service, and then invoked from the base
layer. The respective latency figures are $332$ ms and $491$ ms
respectively; we found that this cost was dominated by about $300$ to
load the capsule code into memory and invoke it. We currently use
\textit{dlopen} for this purpose; we intend to replace this when
removing \textit{glibc} from our TCB, so the overhead should be
significantly improved. The bandwidth cost is roughly the cost of
transferring the capsule code along with the sensitive data; in the
TTP case, this cost is incurred between the user and the TTP, while in
the Xen case, the cost is incurred between the user and the
service. Our main observation here is that, in comparison with the TTP
case, the Xen capsule setup is not significantly higher; the cost of a
two round trip protocol involving key generation for installation is
out-weighed by the cost of the data and code transfer.
\textbf{Summary:} In terms of invocation costs, our main finding is
that, in terms of latency, the capsule case is more expensive than the
base case, but this overhead (as a percentage) decreases with payload
size. In comparison with the TTP case, the latency gain provided by
capsules is significant; even with a low round-trip delay of $10$ ms,
the capsule latencies are significantly better. And of course,
bandwidth consumption proportional of the size of the invocation
request and response are incurred in the TTP case, and not in the
co-location case. In terms of setup costs and storage costs, the main
cost is due to the size of the capsule code itself; we plan to reduce
this in the future by removing the dependence on {\it glibc} and by
altering our compilation tool-chain to produce flat code. However, in
practice, this setup and storage cost may be amortized across several
users who may trust the security company to provide their capsule.
\subsection{Stock capsule}
\begin{table*}[t]
\begin{scriptsize}
\renewcommand{\arraystretch}{1.15}
\caption{Single Invocation Costs for (A) Stock Capsule (B) CCN Capsule}
\label{tab:stock_ccn_perf1}
\begin{center}
\begin{tabular}{|l|l|l|l|c|}
\hline\hline
Scenario & \multicolumn{2}{|c|}{Stock Capsule} &
\multicolumn{2}{|c|}{CCN Capsule} \\
\hline
& Latency (micro-seconds) & Bandwidth (bytes) & Latency (micro-seconds) & Bandwidth (bytes) \\
& (Mode, $50\%$-ile, $95\%$-ile) & (bytes) & (Mode, $50\%$-ile, $95\%$-ile) & (bytes) \\
\hline\hline
Base Case & $413~[359,466]~[353,478]$ & NA &
$336~[330.9,344.5]~[322.2,436]$ & NA \\
\hline
TTP & $24224~[24143,24303]~[23954,24432]$ & $1294$ & $366.4~[360.2,380.2]~[350.8,550.8]$ & 1327 \\
\hline
Xen Capsule & $1045~[1001,4179]~[936,4407]$ & NA & $344.6~[336.0,361]~[328.868,523]$ & NA\\
\hline\hline
\end{tabular}
\end{center}
\end{scriptsize}
\vspace{-0.3in}
\end{table*}
Table~\ref{tab:stock_ccn_perf1} shows the invocation cost for the
stock capsule for the three scenarios. The advantage of the Xen option
with respect to the TTP option is clear; co-location offers
significant advantages here since the latency is about $15$ ms lower
and no network bandwidth is consumed. In the financial high-frequency
trading market, this improvement is invaluable; firms typically pay
substantial money to obtain such an advantage. The single invocation
consumes a bandwidth of around $1$ KB for the TTP Case.
In order to reflect a typical stock trading scenario where there are
hundreds of ticker events per section, we plotted the latency metric
versus varying number of back-to-back invocations (corresponding to
ticker events in the stock market) in Figure~\ref{fig:invocation_cost}
(B). Admittedly, this is an apples-to-oranges comparison since the TTP
case requiring network access is substantially slower as compared to
the Xen option and the base case. Comparing the Xen case and the base
case for a sequence of $100$ back-to-back invocations, the latencies
are respectively $43.4$ ms and $9.05$ ms respectively; though the
overhead due to Xen is substantial, it is still a significant
improvement than the TTP case which requires $12$ seconds. Regards the
bandwidth, in the TTP case, the bandwidth consumed for a sequence of
$5, 10, 50, 100, 500$ and $1000$ invocations are respectively $2.8,
4.8, 20.2, 39.4, 194, 387$ KB respectively. In our experiments, we
have only considered one symbol being traded; typically, a user would
be interested in trading on tens or hundreds of different symbols,
which would lead to a proportional increase in the bandwidth. Thus,
the bandwidth required for say, $500$ events per second, is about
$1.6$ MBps per user. Though well within the ability of today's
network, this bandwidth still to be provisioned and paid for, and can
be expensive.
These results reflect that the flexibility of the framework with
respect to deployment at the service site is a significant benefit in
terms of network bandwidth. The capsule option is beneficial to both
the stock broker and the user; the stock broker need only invest in a
trusted module on its computers and avoid paying for bandwidth, while
the user obtains the advantage of housing her trading strategy very
close to the data source itself thus avoiding wide-area network round
trips.
\subsection{CCN capsule}
We present the latency and the bandwidth consumption of a single
invocation of the capsule in Figure~\ref{tab:stock_ccn_perf1}. The
invocation here involves interacting over the network with the payment
gateway over the wide-area over SSL, and thus the additional overhead
in the Xen Capsule or the TTP case is minimal as compared with the
base case. The Xen capsule, in this case, consumes $10$ ms more as
compared to the the base case, and is in fact, more expensive than
even the TTP case. This is due to the fact that we have offloaded the
network stack to the host-hub; the SSL negotiation before the HTTP
request is sent requires many packets to be sent, which involves
several context switches between the two VMs, which leads to the
increase in latency. We plan to reduce this overhead in the future by
implementing a front end and back end specific to SSL which would farm
out the significant portion of the computation to the host hub; the
SSL front end in the capsule would only be responsible for (a)
certificate verification (b) session key generation. This would help
us both reduce the overhead as well decrease the TCB size. The credit
charging request itself is a one-step HTTP request response protocol;
thus this latency can be reduced by pre-establishing SSL
communications or re-using session keys to avoid SSL negotiation
during invocation. The latency numbers also reflect some outliers in
the $95\%$-ile range; this is due to delay outliers at the payment
gateway itself, and are not related to our framework itself. A nominal
bandwidth of about $1.3$ KB is consumed in the TTP case. Once again,
we emphasize that these results demonstrate the cost of the generality
in our framework as compared to custom solutions already deployed
today; the main gain in the CCN case is due to the flexibility of
policy, not performance.
\subsection{Targeted Ads capsule}
\begin{figure}
\vspace{-0.1in}
\centering
\mbox{
\begin{tabular}{c}
\psfig{file=results/exp_ads_invocation_ads_lat.pdf,height=2in}\\
{(A)}
\end{tabular}
\begin{tabular}{|l|l|l|}
\hline\hline
Scenario & Base/Xen Case & TTP \\
\hline\hline
2 & 46.5 & 62.1 \\
\hline
5 & 139.6 & 201.8 \\
\hline
10 & 294.7 & 434.7 \\
\hline
25 & 715.2 & 1133.3 \\
\hline
50 & 1535.9 & 2297.2 \\
\hline
100 & 3087.2 & 4626.5 \\
\hline\hline
\multicolumn{3}{|c|}{(B)} \\
\hline
\end{tabular}
}
\vspace{-0.15in}
\caption{ (A) Invocation Costs for targeted ads capsules that
serves ads under three different scenarios (base case, TTP,
xen capsule) for the \textit{SelectAd} interface: Latency
(Y-Axis) vs (Number of Ads) (X-Axis). (B) Aggregation
Bandwidth Requirement: Column titles denote number of users,
Row titles reflect the scenario, and the matrix element
denotes the bandwidth consumed (KB). Individual user
histories are about $15$ KB per user. }
\label{fig:ads_perf_aggreg}
\vspace{-0.2in}
\end{figure}
We present two sets of results for the targeted ads scenario.
Figure~\ref{fig:ads_perf_aggreg} (A) shows the latency per invocation
for the capsule to serve targeted ads using the {\it SelectAd} (we do
not show results for the {\it GetInterestVector} interface due to lack
of space). This graph shows that the Xen capsule has nearly the same
overhead as the Base case capsule once the number of ads out of which
the capule selects one exceeds $5$. This is mainly because the payload
size dominates the transfer time. In the TTP case, the latency is
clearly impacted by the wide-area network delay. The bandwidth
consumed in the TTP case was measured to be $2.6, 8.3, 15.8, 29.2$ KB
for $10, 50, 100, 200$ ads respectively. These reflect the bandwidth
savings of the co-located capsule option; since this bandwidth is
incurred for every website the user visits, this could be significant.
There are two other operations supported by this capsule which are of
interest: allowing a user to update her web history and the data
aggregation transformation. For these operations, we present the
bandwidth consumed, since latency is not the primary constraint in
these cases. Regarding the update operation, the bandwidth required
was measured to be $2.1, 4.8, 15.9, 23.3, 52.7$ KB for $5, 10, 50,
100, 200$ new websites visited per day by the user. The average number
of websites visited per user is typically around $50$ websites per
day~\cite{web:acmweb08}, so we chose to experiment in the range of
$5-200$ websites per day. This bandwidth is incurred in the
server-side capsule, the base case, and the TTP case; it however is
not necessary in the client-side capsule. In the case of targeted
advertising, we expect that the update traffic per day is much lower
than the cost of sending ads (which change much more frequently); thus
the server-side co-located capsule is advantageous with respect to
bandwidth consumption. In other cases where user data is much more
frequently updated, a client-side capsule may be more suitable.
Regarding the aggregation operation, we currently employ an
aggregation strategy that optimizes the total bandwidth consumed. In
these experiments, we consider $N$ users who have decided to pool
their data together. In the base case and Xen case, we assume that the
data of each user is hosted on a physically different machines in the
data-center; the optimal way to combine their data in terms of
bandwidth is for a simple hub-and-spoke model. This is accomplished by
the local host-hub for users $1,\cdots,(N-1)$ executing a {\it pull}
operation on their capsules, and pushing the resultant encrypted
histories to the host-hub for user $N$. At this point, the capsule for
user $N$ would release the aggregate data upon request. In the
TTP/client-side case, we assume that the $N$ users have each entrusted
their data to $N$ different administrative entities; in this case, the
bandwidth-optimal strategy is for the service is a similar
hub-and-spoke strategy. We experimented with varying numbers of users
and present the bandwidth consumed in Figure~\ref{fig:ads_perf_aggreg}
(B). In this experiment, the data per user is fixed to be $1K$ per
user. The bandwidth consumed for the Xen case and the base case are
the same since the same strategy is involved; the TTP case consumes
much more bandwidth since both the push and pull requests require
remote access, as does the final pull from capsule $N$. Thus, the
advantage in the Xen case is that the bandwidth is lower, and further
the bandwidth consumed is only intra-data-center. Further, the
bandwidth consumed can be optimized much more easily for the
server-side capsule and the base case since it is within the confines
of a single administrator. This demonstrates the usefulness of the
aggregation operation especially when thousands of users are involved.
\section{Security Analysis} \label{sec:secanalysis}
We now discuss the desirable security properties of our capsule,
determine the TCB for these security properties to hold, and argue
that our TCB is a significant improvement over the status quo.
\subsection{Security Properties}
\textbf{Security Goal:} Our over-arching security goal (from our
problem statement in Section~\ref{sec:problem}) is that any access to
the data or any transfer of the data must only occur in accordance
with the interface and policies sanctioned by the user. The adversary
is a party that has compromised the software stack at the service or
is a party that has physical access to the machine hosting the
capsule. Note that this goal offers protection only to the extent to
which the data is protected by the interface and user's policies. For
instance, if the user specified a budget of $100$ dollars for her
credit card, and the machine hosting her capsule was compromised, her
card could be charged upto her budget by the compromise. However, such
a charge would still be securely tied to the compromised service since
it is not possible to spoof the invoking principal, thus offering
sufficient data for post-mortem forensics. Further, we note that the
service can always launch an availability attack in the co-located
case by simply refusing to service requests using its host hub or by
denying access to a user in accessing her capsule to, say, retrieve a
record of all operations; we only aim to prevent the adversary from
violating the policy constraints, not liveness requirements.
\textbf{Informal Security Argument:} We will now present informal
arguments as to why our capsule framework acheive this security goal.
In the case of a TTP-based deployment, the argument is simple; we rely
on the isolation provided by the TTP and the correctness of the
capsule implementation to ensure the security goal. The rest of this
section is concerned with the co-hosting case. We are currently
exploring using the $LS^{2}$ framework~\cite{ls2:oakland09} to state
and formally verify our security goals in the co-hosting scenario; we
defer such any formal argument for future work. In the following
informal arguments, we will assume that the trusted module provides
the four security properties detailed in
Section~\ref{sec:design_interaction}: isolation, remote attestation,
anti-replay protection, and random number generation.
We make this argument in two steps. First, we argue that any
interaction between the service and the user's data occurs only via
the functions exported by the capsule base layer. Second, we will rely
on the base layer protocols and its correctness to ensure that all
these interactions are in conformance with the user's policies.
For the first, we will rely on the isolation property of the trusted
module. In the case of the VM, this assumes that the VM hosting the
service code can interact with the capsule VM only via the XenSocket
interface. This requires that the VMM be placed in our TCB. We also
assume that the VMM prevents significant leakage of information
through side-channels such as memory page caching, CPU usage: this is
a subject of ongoing research ({e.g., }
Tromer~{et al.}~\cite{arch_attacks_mitigiation:sosp09}). For the second,
we note that the following lists all the possible interactions between
the service and the capsule: installation, invocation, hosting
transfer, and data transformation (going through the parts of the
framework in Section~\ref{sec:arch}). We will now consider each in
turn.
\textbf{Installation:} The installation protocol
(Section~\ref{sec:design_hosting}) provides three properties in this
regard. First, it uses the attestation functionality of the trusted
module to ensure that only the base layer code generates the public
key in response to the first installation message. Second, the
installation protocol (based on the classic Diffie-Hellman protocol
with RSA keys exchange alongside the Diffie-Hellman messages) ensures
the confidentiality of the information; only the party which generates
the public key in response to the first installation message can
decrpt the capsule. Due to the attestation guarantee, this party is
guaranteed to the base layer, which is trusted. Note we rely on the
random number generation ability of the trusted module to ensure that
the generated public key is truly random and cannot be biased or
guessed by the adversary. Third, since the attestation verifies the
purported input and output to the code which generates the public key,
man-in-the-middle attacks are ruled out; an adversary cannot alter any
part of the message without rendering the atttestation invalid. These
three properties ensure in conjunction that the capsule can only be
decrypted by the base layer. We then rely on the correctness of the
base layer to argue that the capsule is correctly decrypted and
instantiated at the service.
\textbf{Invocation:} We next need to argue that any invocation occurs
in accordance with user's policies. Every invocation is checked by the
base layer to ensure: (a) the identity of the invoking principal (b)
the correctness of any supporting policies (c) that the user-specified
policies along with the supporting policies allow the invoking
principal the privilege to make the particular invocation.
The identity of the invoking principal is verified either against a
list of public keys sent during the installation protocol (this data
list binds the names of services to their public keys) or, in the case
of the owner, it is verified against the public key of the data owner
sent during the installation protocol. In either case, the correctness
of the installation protocol ensures that the identity cannot be
spoofed since the list of public keys used to validate the identity is
correct. The correctness of the supporting policies is ensured by
verifying that each statement of the form $X~\underline{says}~\cdots$
is verified against the signature of $X$. The policy resolver takes
two sets of policies as input: the database of policies sent by the
user and the set of supporting policies sent by the invoker. The
correctness of the latter we have already argued for; for the former,
we rely on the anti-replay property provided by the trusted module to
ensure that the database of policies in the capsule (which can be
updated over time) is up-to-date and reflects the outcome of all the
invocations made so far. We assume the correctness of the policy
resolver to ensure that it answers invocation queries correctly in
accordance with these policies. For stateful constraints, the
implementation also updates the state before returning the result of
the invocation to the service.
Once the invocation is granted by the base layer, it is passed on to
the data layer which carries out the required functionality. In cases
such as a query capsule or an analytics capsule, this functionality is
carried out entirely within the data layer, which we assume to be
correct. In cases such as the proxy capsule, which requires access to
the network, we note that though the network stack is itself offloaded
to the host hub, we rely on the host hub only for availability, not
for safety. The SSL library and cryptographic library are enclosed
within the capsule, and thus, it does not matter that the host hub is
effectively a man-in-the-middle. In this respect, we also note that
the other services provided by the host hub, namely, persistent
storage and obtaining the current time, are similarly relied only for
availability, not safety, since the base layer validates persistent
storage using its own integrity and confidentiality checks (once
again, relying on the anti-replay functionality provided by the
trusted module), and verifies the current time by validating the
source (such as a trusted NTP server).
We next consider the hosting protocol. Since it consists of the
invocation protocol (used by the service to make the transfer request)
and the installation protocol (used to effect the transfer), we rely
on our arguments regarding invocation and installation to argue that
hosting is in accordance with our security goal as well.
\textbf{Data Transformation:} We now turn our attention to the data
transformation protocol. The filtering and update protocol happen only
within the confines of a single capsule, and their correctness
arguments is similar to those for invocation. The aggregation protocol
though more complicated decomposes into a number of simple pair-wise
aggregation operation. Since during the pair-wise operation, a capsule
verifies the identity of its peer as belonging to a user who is part
of the data crowd, the data is merged only between capsules that are
part of the same data crowd. Since the merged information is only
revealed to the service after a minimum number of capsules belonging
to the data crowd are merged, this aggregation is in accordance with
the user's policies.
\subsection{TCB} \label{sec:secanalysis_tcb}
We now evaluate the TCB of our capsule framework. From the previous
section, it is clear that our TCB consists of: (A) the trusted module
(B) the data layer interface and implementation (C) the base layer
protocols and implementation.
The trust assumption in (A) appears reasonable for VMMs, TPMs, and a
secure co-processor. A carefully managed VMM which presents a minimal
attack surface to applications inside a guest VM may be judged to be
adequate enough to defend against software threats. Similarly, TPMs
and secure co-processor have both been designed across several
iterations; it seems reasonable to expect that their designs and
implementation in today's chipsets will be difficult to exploit.
Regarding (B) and (C), the LOC metrics are discussed in
Section~\ref{sec:impl_framework}). We base our trust in the data layer
interface and implementation in the simplicity of the interface; the
data specific portion of the interface is less than $500$ lines even
for our most complicated capsule (the CCN capsule). For other capsules
following the query idiom, analytics idiom, and provenance idiom, the
implementation is even simpler since no network access is required.
Currently, the base layer protocols and implementation are much more
complex than the data layer itself, as reflected in the LOC metric
(the base layer has roughly 6K LOC). While we have aimed to keep our
design simple right from a minimal policy language to offloading
network access to the untrusted host hub, the base layer is still
complex, and currently includes a cryptographic library (PolarSSL) as
well. In the future, we hope to formally verify the correctness of
this implementation. For now, our only argument is that even granting
that the base layer is complex, it offers a significant improvement in
security to users of web services today who have no choice but to rely
on an unaudited closed source service code for their privacy
guarantee. More importantly, it restores control to the user over her
data, allowing a security conscious user to invest in a secure capsule
implementation for her data from a trusted security company. Another
comparison point in this respect is an information flow control system
(such as JIF or Asbestos). The TCB in such a case includes the
run-time (in Jif's case, this is the Java run-time along with Jif's
typing rules; in Asbestos's case, it would be the operating system)
along with any trusted de-classifiers (a declassifiers is a code block
or a process that is allowed to break the typing rules or propagation
rules; such declassifiers are essential to prevent over-conservative
information flow propagation). We believe this compares favorably with
our TCB; building a general enforcement mechanism seems a more complex
undertaking than implementing a simple interface which makes our
approach easier to verify.
\section{Related Work} \label{sec:related}
We first discuss two closely related papers in literature, before
examining broader related work. The closest related work to our own is
in Wilhelm's thesis~\cite{mobile_privacy:thesis} (Chapter 5). This
work leverages mobile agents (a well-investigated research paradigm;
an agent is code passed around automatically from one system to
another in order to accomplish some functionality) to address the
issue of privacy. The suggestion is to encapsulate data with a generic
policy enforcement layer that provides a fixed set of functionality
(such as, enforce time for retention, audit operations). The main
differences from our approach are: (a) our work is tailored towards
web services; in particular, operations like hosting transfer and data
transformation were chosen to model how web services handle data today
(b) our approach is extensible to various kinds of sensitive data; the
enforcement layer is configurable by the user and is specific to the
particular kind of sensitive data (c) we aim to design the capsule
interface itself so as to reduce data exposure. The second closely
related paper to ours is by Iliev~{et al.}~\cite{client_privacy:secpriv}.
They propose the use of trusted hardware to offer client privacy when
operating on server data for two applications: private information
retrieval (without traditional cryptographic mechanisms) and an
armored network traffic vault. This proposal is similar to our
co-location option. Two characteristics of our framework that are
pertinent to web services, and do not apply to
\cite{client_privacy:secpriv} are: (a) support for scenarios where a
user's data is spread across multiple services. (b) policy-based
control over invocation, hosting, and data transformation. Apart from
these specific papers, our work is related to the following broad
areas:
\noindent \textbf{Information Flow Control (IFC):} The principle of
IFC has been implemented in OSs ({e.g., } Asbestos~\cite{asbestos:sosp05})
and programming languages ({e.g., } JIF~\cite{jif:jsac03}), and allows the
control of flow of information between multiple processes or security
compartments. IFC has also been used to build secure frameworks for
web services in W5~\cite{w5:hotnets07}, xBook~\cite{xbook:ssym09}. The
main difference from the capsule framework is that we provide data
access control, not data propagation control. Privacy is guaranteed by
the interface chosen by the user, and not by run-time policy
enforcement. Capsules apply only to cases where an interface can be
arrived at that offers sufficient privacy to the user as well as is
usable to the service; if no such interfaces exist, we can leverage
IFC frameworks to control the propagation of sensitive
information. The advantage in restricting to interface-based access
control is that we can rely on a variety of isolation mechanisms (such
as TPMs) without requiring a particular OS or programming
language. Further, the simplicity of the interface makes it feasible
to envision the possibility of proving the correctness of capsule
code; doing so in the IFC case requires one to prove the correctness
of the enforcement mechanism (OS or compiler) which can be
significantly more complex.
\noindent \textbf{Decentralized Frameworks For Web Services:} Privacy
frameworks that require only participation from users have been
proposed as an alternative to web services. VIS~\cite{vis:mobiheld09}
maintains a social network is maintained in a completely decentralized
fashion by users hosting their data on trusted parties of their own
choice; there is no centralized web service. Capsules are more
compatible with the current ecosystem of a web service storing user's
data and rely on the use of interfaces to guarantee privacy.
NOYB~\cite{noyb:wosp08} and LockR~\cite{lockr:conext09} are two recent
proposals that rely on end-to-end encryption to hide data from social
networks; both these approaches are specific to social networks, and
their mechanisms can be incorporated in the capsule framework as well,
if so desired.
\noindent \textbf{Enterprise Privacy Architectures:} Several privacy
architectures that deal with the propagation of information within an
enterprise or across closely-related enterprises have been proposed.
These include Ashley~{et al.}~\cite{platform_enterprise:isse02},
Backes~{et al.}~\cite{toolkit_privacy:esorics03},
GeoDac~\cite{geodac:tech_rep}. Broadly, these papers propose
specification languages and enforcement mechanisms that help enforce
enterprise-wide policies on data retention, access logging, and other
accountability guarantees. The main difference from the capsule
framework is that (a) our framework involves two mutually distrusting
parties, the user and the service, and thus we only base enforcement
on hardware or system software, rather than application software. (b)
our framework is designed for web services. (c) our goal is to enforce
interface constraints and policy control that control data leaks,
rather than to enforce generic data control policies. In future, we
hope to incorporate guarantees provided by these architectures in our
framework as well. Our provenance capsule is also related to
VFIT~\cite{vfit:iacs08} and GARM~\cite{garm:hotsec09} which follow the
modification of data across multiple machines.
\noindent \textbf{Privacy in Cloud computing:} The area of cloud
computing has seen a lot of work in the context of privacy as well.
These include Trusted Cloud~\cite{trusted_cloud:ssym09}, Accountable
Cloud~\cite{accountable_cloud:ladis09}, Cloud
Provenance~\cite{provenance_cloud:ladis09}. These works deal with the
more complex problem of guaranteeing correctness of code execution on
an untrusted third party. Capsules are only concerned to protecting
the privacy of the user's data; we assume the application service
carries out the service (such as sending correct ticker data) as
expected. Airavat~\cite{airavat:nsdi10} proposes a privacy-preserving
version of MapReduce based on information flow control and
differential privacy; capsules support general kinds of computation,
but have to rely on manual re-factoring whereas Airavat can
automatically ensure privacy by restricting itself to specific types
of MapReduce computations. We also note that attacks based on leakage
across VMs are known~\cite{get_off:ccs09} and defense mechanisms
against such attacks are also being
developed~~\cite{arch_attacks_mitigiation:sosp09}. Our capsule
framework can avail of such defense mechanisms as they are developed
further; we view such work as orthongal to our central goal.
\noindent \textbf{Targeted Advertising Systems:}
PrivAD~\cite{privad:hotnets09} and Adnostic~\cite{adnostic:ndss10} are
recent proposals that are client-side systems for targeted
advertising; this is somewhat similar to a client-side targeted ads
capsule in our framework. The difference is that our framework
generalizes to other deployment scenarios as well. In the future, we
hope to borrow their techniques for anonymized ad impression
collection in our targeted ads capsule as well; currently, we do this
using aggregation, PrivAD and Adnostic offer dealer-based and
encryption-based mechanisms respectively for this purpose.
\noindent \textbf{Mechanisms:} The capsule framework builds on
existing isolation mechanisms, such as the virtual machine security
architecture ({e.g., } Terra~\cite{terra:sigops03}), proposals that use
TPMs ({e.g., } Flicker~\cite{flicker:eurosys08}), and systems based on a
secure co-processor~\cite{yee:thesis}. These proposals offer the
foundation upon which our capsule framework can build on to provide
useful guarantees for data owners. Our implementation also borrows
existing mechanisms (XenSocket~\cite{xensocket:middle07},
vTPM~\cite{vtpm:ssym06},
disaggregation~\cite{xendisaggregation:sigops08}, use of hardware
virtualization features~\cite{secureinvm:ccs09}) that help us improve
the performance and security of a virtual machine based architecture.
Currently, we do not provide any automatic re-factoring mechanisms; in
the future, we hope to explore using existing program partioning
approaches ({e.g., } Swift~\cite{swift:sosp07},
PrivTrans~\cite{privtrans:ssym04}) for this purpose.
\section{Conclusion and Future Work} \label{sec:conclusion}
Our capsule framework applies the well-known principle of
encapsulation to provide data access control guarantees to users who
wish to avail of web services that operate on their data whilst still
retaining control over the use of said data. Our design goals are to
allow flexible deployment of such capsules and to provide flexible
policy control to their users. The capsule approach works when a
satisfactory interface can be arrived at that gives privacy to the
user while still sufficing to carry out the service: we show this to
be the case for a broad variety of currently operating web
services. For others, one can rely on information flow control
techniques or end-to-end encryption to obtain privacy guarantees. Our
prototype implementation includes four kinds of capsules as examples;
although the performance of the co-located Xen-based capsule is not
quite suitable for deployment in a production service, they do offer
substantial advantages over the TTP and client-side capsules in the
dimensions of network bandwidth consumption and/or provisioning cost.
We identify three fruitful avenues for future work. First, our
implementation can be improved by incorporating known virtualization
techniques to reduce overhead and by removing the need for including
system libraries and kernels in our TCB by a more careful
implementation. Second, formal verification of the control layer
implementation and its security protocols would allow for much greater
confidence in its correctness as compared to today. We have done some
preliminary work in this area towards expressing our implementation in
a formal language and proving that it conforms to our security
expectations.
Similar correctness guarantees would be useful for the data layer
interface and implementation as well. The flavor of correctness proofs
here is different from those in the control layer protocol; the goal
is to prove that there is no leakage of information via the
implementation. One can envision static information flow checking
(such as JIF~\cite{jif:jsac03}) to provide such a guarantee, or a
proof from scratch based on computational non-inference. Third, given
the complementary nature of information flow control systems, it would
be useful to integrate such elements into the capsule framework;
interface calls which do not leak any information would be part of the
capsule framework, leaky interface calls can leverage use of the
information flow policies. Of course, the former would involve a
smaller TCB, since information flow requires placing the information
flow enforcement kernel
in the TCB.\\
\noindent \textbf{Acknowledgements:} We thank the authors of Adnostic,
in particular, Arvind Narayanan, for sharing their data.
\bibliographystyle{ieeetr}
|
\section{Introduction}
During the last decade, the growing availability of large
detabases, the increasing computing powers, as well as the
development of reliable data analysis tools, have constituted a
better machinery to explore the topological properties of complex
networks \cite{bara,newman,rpas,watts}. Most of studies has
revealed that, despite the inherent differences, main real
networks are characterized by the same topological properties,
such as relatively small path lengths, high clustering
coefficients, scale-free degree distributions, degree
correlations, motifs, and community structures. As a particular
class of complex networks, spatial networks are different from
other networks, which are those embedded in the real space, i.e.
networks whose nodes occupy a precise position in two or three
dimensional Euclidean space, and whose edges are real physical
connections. It is not surprising that the topology of spatial
networks is strongly constrained by their geographical embedding,
such as airport networks, urban street networks. In this paper we
focus on an urban road networks.
Urban road networks, in one way or another, are portrayals of the
history of the country's development. Hence the topology studies
of various spatial networks could show significantly different
degree distributions. For example, the degree distribution of
internet network (presupposing that the network nodes are routers)
has the form $P(k)\sim k^{-\gamma}$ where
$\gamma_{out}=\gamma_{in}\approx2.48$ \cite{bara}. The power grid
of the western US is described by a complex network with
exponential form degree distribution \cite{amaral}. In addition,
the analysis of the spatial distribution of the network nodes
shows that both indicated networks are fractals with the method of
box counting \cite{rpas, gast}. But, here, we show other distinct
degree distribution, double power-law, in urban road networks of
Le Mans in France (Le Mans is one small city in France). Urban
road networks with links and nodes representing road segments and
junctions respectively which is a primal graph, exhibit unique
features different from other networks. As road networks are
almost planar, they show a very limited range of node degrees and
could never be scale-free like airport networks or the internet
\cite{gast}. Nevertheless, there is an interesting connection
between those scale-free spatial networks and road networks, since
both are extreme cases of an optimization process minimizing
average travel costs along all shortest paths, given a set of
nodes and a total link length.
In the present paper, the organization is as follows. Section 2
presents the results of degree distribution and cumulative degree
distribution of urban road network of Le Mans. In section 3, we
analyze the relationship between the double power-law distribution
function and its parameters. In section 4 and 5, some topological
properties of network are investigated, such as clustering
coefficients, diameter and average shortest path length.
Conclusions and discussions are given in the last part, section 6.
\section{Degree distribution}
We investigate the urban road network of Le Mans in France.
Previous works on urban traffic network usually were based on a
primal graph, where intersections were turned into nodes and roads
into edges. Here, we look at them with other paradigm, a dual
graph, where a city is transformed into a topological graph by
mapping the roads into nodes and the intersections into edges
between the nodes. The advantage of the dual graph is that it does
not exhibit any geographic constrain compared to primal graph.
Hence, in structure of our network, there are 1585 nodes and 5066
edges.
During the research of complex network science, degree
distribution ($p(x)$, $x$ is the degree of one node) is always one
of the most important characteristics. In Figure \ref{pxfig}, we
plot the degree distribution in double-log plot and fit the data
with black solid line with function as follows, which is called
``double power-law" function by us:
\begin{equation} \label{p}
p(x)=a/(x^b+c*x^d),
\end{equation}
where $a=46.7,b=3.1,c=1999,d=-1.5$. The slopes of two straight
dashed lines are 3.1 and -1.5 respectively. The reason for
choosing this kind of function was explained in \cite{liwei}. Many
similar distributions were found in other researches including the
highway networks of Korean \cite{Jung} and urban public transport
networks of Los Angeles in American \cite{Ferber}. Nevertheless,
they were named ``truncated power-law distributions" by authors.
In fact, abundant works with ``double power-law" were analyzed as
the truncated power law \cite{Kalapala, Namikawa, Niwa, Biham,
Mossa, Gupta, Gupta2} in which more attention was paid to large
degree value in one power-law region and small degree value in
another one was neglected with various reasons. In addition, it is
worth to notice that in Figure \ref{pxfig}, the node with degree 5
has the largest probability, in other words, crossover point has
maximum value of probability, which is similar to the Moscow
region road network in which the node with degree 3 had largest
probability in degree distribution \cite{moscow}. It is very
different from non-road spatial network's degree distributions in
which the node with smallest degree, not crossover point, has the
largest probability generally. This is decided by the feature of
spatial networks such as transportation networks, since the
topology of spatial networks is strongly constrained by their
geographical embedding. Furthermore, both belonging to
transportation networks, road network still has distinct degree
distribution form from airport networks \cite{liwei} in which
crossover point does not have maximum value of probability. Since
there is a strong restriction to the growth of the degree in road
networks which is induced by the geographical arrangement of the
edges, however, there is no so strong geographical restriction to
the edges existence in the airport networks.
Towards degree distribution function, combined Eq.~(\ref{p}) and
Figure \ref{pxfig}, we could simply find that parameter $a$ is the
normalization constant, $b$ and $d$ are two scale exponents in two
regions of power-law respectively. So, in principle, from the
standpoint of mathematics, this kind of function could produce any
form of ``double power-law" distribution, which could be seen
having some relationships with the origin of this phenomenon.
An alternative way of presenting degree data is to make a plot of the
cumulative distribution function
\begin{equation} \label{P}
P(x)=\int_x^{40} a/(x^b+c*x^d)dx,
\end{equation}
which is the probability that the degree is greater than or equal to
$x$ and 40 is the maximum degree in our network. The advantage of
cumulative distribution is that all the original data are represented
and it reduces the noise in the tail. After calculation of Eq.~(\ref{P}),
we get
\begin{equation}
P(x)=\frac{-42.5+0.4*x^{2.5}Hpergeometric2F1[0.54,1,1.54,-0.0005*x^{4.6}]}{x-40},
\end{equation}
where $Hypergeometric2F1[a,b,c,z]$ is the hypergeometric function defined by:
\begin{eqnarray}
{}_2F_1(a,b;c;z)=\frac{{\Gamma(c)}}{{\Gamma(b)\Gamma(c-b)}}\int_0^1
{t^{b-1}}(1-t)^{c-b-1}(1-tz)^{-a} dt.
\end{eqnarray}
From Figure \ref{cumupxfig}, we also find that there is double
power-law distribution in log-log plot with two exponents -0.017
and -3.35 respectively.
\begin{figure} [h]
\includegraphics[scale=.65]{px.eps}
\centering \caption{Degree distribution of urban road network of
Le Mans city in France. $p(x)$ is the degree distribution
function, and $x$ is the degree of each node. ``double power-law"
phenomenon is obviously showed in log-log plot and our fitting
curve with black solid line matches the data well. Two straight
dashed lines define two power laws with scale exponents 3.1 and
-1.5 respectively. Our fitted function is
$p(x)=46.7/(x^{3.1}+1999*x^{-1.5})$. } \label{pxfig}
\end{figure}
\begin{figure} [h]
\includegraphics[scale=.65]{cumupx.eps}
\centering \caption{Cumulative distribution function of degree of
urban road network of Le Mans city in France. $P(x)$ is the
cumulative degree distribution function, and $x$ is the degree of
each node. ``double power-law" phenomenon is also showed in
log-log plot. Two straight dashed lines define two power laws with
scale exponents -3.35 and -0.017 respectively.} \label{cumupxfig}
\end{figure}
\section{Relationship between double power-law function and its parameters}
We make an elaborate analysis on the expression of $p(x)$.
Firstly, we do the normalization for the probability function by
using equation $ \sum\limits_{x = 1}^{100} {p(x) = 1}$ since the
degree of each node is discrete variable. As parameter $a$ is the
normalization constant, so we pay more attention to the influence
on the expression of $p(x)$ caused by the varying of parameters
$b,c,d$. From the Figure \ref{pxfig}, we could find that $b$ and
$d$ have the directly relationship with the exponents of two
power-law regions. Hence, our discussions mainly focus on the
varying of $b$ and $d$. Based on the different features of degree
distribution in airport networks and urban road networks, our
discussions are divided into two classes: one is that degree at
crossover point has non-largest probability, behaved as both of
$b,d$ are positive like airport networks; another is that degree
at crossover point has largest probability, behaved as either $b$
or $d$ is negative, the rest is positive, like urban road
networks. The case of $b=d$ reduces to familiar single power-law
distribution. In each class, we attempt to analyze the role of
each parameter played. At the same time, we make sure that $p(x)$
is always normalized. We define the intersection of two power-law
in all figures as ``crossover" and define the scaling behavior in
the region of $1<x<x_{crossover}$ as ``the first power-law" and
define that in the region of $x_{crossover}<x<100$ as ``the second
power-law".
In the first case $b>d>0$, we choose the probability function in
the form of $p(x)=a/(x^b+1293139*x^{0.5})$ which is a monotonical
function. In the panel $a$ of Figure \ref{pbggdfig}, it is found
that the exponents of the first power-law are the same with fixed
$d$. As $b$ increases, the exponents of the second power-law
become bigger and the value of crossover decreases. In the panel
$b$, the expression of function is $p(x)=a/(x^4+1293139*x^d)$. Due
to the identity of parameter $b$, three curves have the same
exponents of the second power-law. As $d$ increases, the exponents
of the first power-law become bigger and the value of crossover
increases. In the panel $c$, we have the function
$p(x)=a/(x^{20}+c*x^3)$. As the result of equality of both $b$ and
$d$, the overlap of curves of the first power-law and the
parallelity of curves of the second power-law are showed
evidently. The effect of parameter $c$ is to change the position
of crossover. The value of crossover increases with parameter $c$
increasing. So, based on above results, we draw a conclusion that
when $b>d>0$, parameter $b$ and $d$ are directly in charge of the
exponents of the second and the first power-law respectively.
Parameter $c$ is in charge of the position of crossover when $b$
and $d$ are fixed.
\begin{figure} [h]
\includegraphics[scale=.45]{plotall.eps}
\centering \caption{Variation tendency of probability function
$p(x)$ against its parameters $b, c, d$ when $b>d>0$ in log-log
plot. In the panel $a$, the exponents of the first power-law are
identical since parameter $d$ is fixed. Those of the second
power-law become bigger with $b$ increasing. While the value of
crossover becomes smaller. In the panel $b$, the exponents of the
second power-law are identical because of invariance of parameter
$b$, those of the first power-law behave the same as panel $a$.
But the value of crossover increases along with parameter $d$. In
the panel $c$, the exponents of each power-law are identical
respectively by reason of unchanged $b$ and $d$. Parameter $c$
directly influences the position of the crossover.}
\label{pbggdfig}
\end{figure}
In the second case $0< b<d$, we choose
$p(x)=a/(x^b+10^{-80}*x^{150})$. In Figure \ref{pblldfig}, we make
a comparison with the Figure \ref{pbggdfig}. It is observed that
the variation tendency in the panel $a$ and $b$ of Figure
\ref{pblldfig} is quite the reverse of them in Figure
\ref{pbggdfig}. The variation tendency in the panel $c$ of both
figures are mostly the same. So, from discussions of the first
class , it could be seen that in networks such as airport
networks, the probability function is monotonical decreasing. In
addition, the first power-law with smaller $b$ or $d$ has larger
probability than the second power-law with the bigger $b$ or $d$.
\begin{figure} [h]
\includegraphics[scale=.45]{plot2all.eps}
\centering \caption{Variation tendency of probability function
$p(x)$ against its parameters $b, c, d$ when $0<b<d$ in log-log
plot. In the panel $a$, the exponents of the second power-law are
identical since parameter $d$ is fixed. Those of the first power-
law become bigger along with parameter $b$. In the meantime, the
value of crossover varies bigger. In the panel $b$, the exponents
of the first power-law are identical because of unchanged
parameter $b$. Those of the second power-law become bigger along
with $d$. The value of crossover varies smaller when $b$
increases. In the panel $c$, the exponents of each power-law are
the same respectively. The value of crossover decreases when
parameter $c$ increases, which behaves opposite to that in Figure
\ref{pbggdfig}. } \label{pblldfig}
\end{figure}
Next, we will discuss the second class which means the function
exhibits nonmonotonicity like Figure \ref{pxfig}. In Figure
\ref{pb1fig}, we take the function as $p(x)=a/(x^b+1999*x^{-1.5})$
in the panels $a$ ($b>|d|$) and $b$ ($b<|d|$) with negative $d$.
In the panels $c$ ($|b|>d$) and $d$ ($|b|<d$) with negative $b$,
the function is $p(x)=a/(x^b+10^{-4}*x^{2.5})$. We could find the
variation tendencies are both the same for panel $a$ and $b$, and
for panel $c$ and $d$. For the former, as $b>d$, $b$ is in charge
of the second power-law. Meanwhile, because of $|b|>|d|$ in panel
a, probability function has small value of crossover. However, in
panel b, because of $|b|<|d|$, probability function has big value
of crossover. For the latter, as $b<d$, $d$ is in charge of the
second power-law. In panel c, since $|b|>|d|$, probability
function has small value of crossover. It is vice versa in panel
d. So, we draw a conclusion that the second power-law is charged
by bigger parameter out of $b$ and $d$. And, probability function
has big value of crossover when $|d|>|b|$.
\begin{figure} [h]
\includegraphics[scale=.45]{plot3all.eps}
\centering \caption{Variation tendency of probability function
$p(x)$ against parameter $b$ when one of $b, d$ is negative. From
each panel, we could find the exponents of the first power-law are
negative and those of the second power-law are positive. With the
increasing of the absolute value of parameter $b$, the exponent of
power-law depended on $b$ is increasing. The value of crossover
becomes smaller when the absolute value of parameter $b$
increases.} \label{pb1fig}
\end{figure}
\section{Clustering coefficient}
A common property of social networks is that cliques form,
representing circles of friends or acquaintances in which every
member knows every other member. This inherent tendency in cluster
is quantified by the clustering coefficient. For each node $i$
having $x_i$ edges which connect it to $x_i$ other nodes, if the
nearest neighbors of this node are part of a clique, there would
be $x_i (x_i-1)/2$ edges between them. The radio between the
number $E_i$ of edges that actually exist between these $x_i$
nodes and the total number $x_i (x_i-1)/2$ gives the value of the
clustering coefficient of node $i$:
\begin{equation}
C_i=\frac{2E_i}{x_i(x_i-1)}.
\end{equation}
The clustering coefficient of the whole network is the average of
all individual $C_i$'s. By calculation, $C$ of the entire urban
road network is 0.042. To check whether our network is a
small-world network or not, we construct a random network with
same number of nodes and edges in our network and get
$C_{rand}=0.004$ which is much smaller than that of urban road
network.
\section{Characteristic path length}
Shortest paths play an important role in the transport and communication
within a network. It is useful to represent all the shortest path lengths
of a graph $G$ as a matrix $D$ in which the entry $d_{ij}$ is the length
of the geodesic from node $i$ to node $j$. The maximum value of $d_{ij}$ is
called the diameter of the network. A measure of the typical separation between
two nodes in the network is given by the average shortest path length, also
known as characteristic path length, defined as the mean of geodesic lengths over
all couples of nodes \cite {SB}:
\begin{equation}
L=\frac{1}{N(N-1)}\sum_{i,j\in G,i \neq j}d_{ij}.
\end{equation}
From above equation, we get the average shortest path of urban
road network $L=4.183$ and $L_{rand}=4.000$ from artificial random
network we constructed. $L\approx L_{rand}$ and $C\gg C_{rand}$ suggest
that urban road network of Le Mans has the small-world network's properties.
In the meanwhile, we calculate the diameter of this network and its value is
equal to 8.
\section{Concluding remarks}
In the present paper, we have done a empirical study on urban road
network of Le Mans in France and investigated some structure
properties of it. From its degree distribution, double power-law,
we find that urban road network as one of spatial networks has
special degree distribution form from other networks, since there
are many constrains from its geographical embedding. The
phenomenon that the node with smallest degree has largest
probability in most of single power law distribution is not true
here. The node at the crossover point has the largest probability
in network we studied.
In order to make a deeply understanding of double power-law
distribution, we elaborately analyze the structure of the
probability function. Our discussions are divided into two classes
based on the feature of degree distribution of two transportation
networks, road network and airport network. One class is that both
of $b$ and $d$ are positive; another is that either of two
parameters is negative. Through our analysis, the conclusion is
that the bigger parameter of $b$ and $d$ is in charge of the
behavior in the second power-law with bigger degree. The smaller
one is in charge of that in the first power-law with smaller
degree. Absolute value of $b$ is more bigger than that of $d$, the
smaller the value of crossover. In addition, parameter $c$ just
controls the position of crossover in the limit of both $b$ and
$d$ unchanged.
At last, through the calculation of clustering coefficient and
average shortest path, small world properties have been found.
Though these topological properties are investigated and a novel
double power-law distribution function is given, the mechanism of
the organization of double power-law structure and that of the
transition from single power-law to double power-law are still
open questions in complex network study. It is necessary to
research more other features of double power-law distribution in
next stage, such as the behavior of its entropy. Perhaps a model
producing such distribution is worthwhile to consider.
|
\section{Introduction }\label{intro}
Roll waves are a well-known hydrodynamic instability
occurring in shallow water flow down an inclined ramp,
generated by competition between gravitational force and
friction along the bottom.
These can be modeled as periodic traveling-wave solutions
of the St. Venant equations for shallow water flow,
which take the form of hyperbolic or parabolic balance laws;
see \cite{D,N1,N2} for detailed discussions of existence in
the inviscid and viscous case.
The spectral and linear stability of roll waves has been studied
for the inviscid St. Venant equations in \cite{N1} and
the viscous St. Venant equations in \cite{N2}.
However, up to now, the relation between
spectral, linearized, and nonlinear stability has remained an outstanding open
question.
In this paper, extending recent results of \cite{OZ4,JZ3,JZ4}
in the related conservation law case, we settle
this question by showing that
{\it spectral implies linearized and nonlinear stability.}
This opens the way to rigorous numerical and analytical
exploration of stability of roll waves and related
phenomena via the associated eigenvalue ODE,
a standard and numerically and analytically well-conditioned problem.
At the same time, it gives a particularly interesting application
of the techniques of \cite{OZ4,JZ3,JZ4}.
For, roll waves, by numerical and experimental observation, appear
likely to be stable, at least in some regimes.
In the conservation law case, by contrast,
periodic waves so far appear typically to be unstable \cite{OZ1}.
\subsection{Equations and assumptions}\label{s:equations}
Consider the one-dimensional
St. Venant equations approximating shallow water flow on an inclined ramp:
\ba \label{eqn:1Econslaw}
h_t + (hu)_x&= 0,\\
(hu)_t+ (h^2/2F+ hu^2)_x&= h- u^2 +\nu (hu_x)_x ,
\ea
where $h$ represents height of the fluid, $u$ the
velocity average with respect to height,
$F$ is the Froude number, which here is the
square of the ratio between speed of the fluid and speed of gravity waves,
$\nu={\rm Re}^{-1}$ is a nondimensional viscosity
equal to the inverse of the Reynolds number,
the term $u^2$ models turbulent friction along the bottom,
and the coordinate $x$ measures longitudinal distance along the ramp.
In Lagrangian coordinates, these appear as
\ba \label{eqn:1conslaw}
\tau_t - u_x&= 0,\\
u_t+ ((2F)^{-1}\tau^{-2})_x&=
1- \tau u^2 +\nu (\tau^{-2}u_x)_x ,
\ea
where $\tau:=h^{-1}$ and $x$ now denotes a Lagrangian marker
rather than physical location.
We will work with this form of the equations, as it is more convenient
for our analysis in several ways.
(Indeed, for the large-amplitude damping estimates
of Section \ref{damping}, it appears to be essential in order
to obtain quantitative bounds on amplitude; see Remark \ref{damprmk}.)
Denoting $U:=(\tau, u)$, consider a spatially periodic
traveling-wave solution
\be\label{eqn:tw}
U=\bar{U}(x -ct),
\ee
of \eqref{eqn:1conslaw}
of period $X$
and wavespeed $c$
satisfying the traveling-wave ODE
\be \label{e:second_order}
\begin{aligned}
-c\tau' - u'&= 0,\\
-c u'+ ((2F)^{-1}\tau^{-2})'&=
1- \tau u^2 +\nu (\tau^{-2}u')' ,
\end{aligned}
\ee
Integrating the first equation of \eqref{e:second_order}
and solving for $u= u(\tau):= q-c\tau$,
where $q$ is the resulting
constant of integration, we obtain a second-order scalar profile equation
in $\tau$ alone:
\be \label{e:profile}
c^2 \tau'+ ((2F)^{-1}\tau^{-2})'=
1- \tau (q-c\tau)^2 -c\nu (\tau^{-2}\tau ')' .
\ee
Note that nontrivial periodic solutions of speed $c=0$
do not exist in Lagrangian coordinates,
as this would imply $u\equiv q$, and
\eqref{e:profile} would reduce to a scalar first-order equation
\be \label{e:zerocprofile}
\tau'= F\tau^3(\tau q^2-1) ,
\ee
which since it is scalar first-order has no nontrivial periodic solutions,
even degenerate ones (e.g., homoclinic or heteroclinic cycles)
that might arise in the singular $c\to 0$ limit.
Rather, there appears to be a Hopf bifurcation as $c$ approaches some
minimum speed for which periodics exist; see \cite{N2}, Section 4.1
and Fig. 1, Section 4.2.3.
It follows then that periodic solutions of \eqref{e:profile} correspond to values
$(X,c,q,b)\in \RR^5$, where $X$, $c$, and $q$ denote period,
speed, and constant of integration, and $b=(b_1,b_2)$ denotes
the values of $(\tau,\tau')$ at $x=0$, such that
the values of $(\tau,\tau')$ at $x=X$ of the solution of
\eqref{e:profile} are equal to the initial values $(b_1,b_2)$.
Following \cite{Se1,OZ3,OZ4,JZ3,JZ4}, we assume:
(H1) $\bar \tau>0$, so that all terms in \eqref{eqn:1conslaw}
are $C^{K+1}$, $K\ge 3$.
(H2) The map $H: \,
\R^5 \rightarrow \R^2$
taking $(X,c,q,b) \mapsto (\tau,\tau')(X,c,b; X)-b$
is full rank at $(\bar{X},\bar c, \bar b)$,
where $(\tau,\tau')(\cdot;\cdot)$ is the solution operator of \eqref{e:profile}.
By the Implicit Function Theorem,
conditions (H1)--(H2) imply that the set of periodic solutions
in the vicinity of $\bar U$ form a
smooth $3$-dimensional manifold
\be\label{manifold}
\{\bar U^\beta(x-\alpha-c(\beta)t)\},
\;
\hbox{\rm with $\alpha\in \RR$, $\beta\in \RR^{2}$}.
\ee
\br\label{H2rmk}
\textup{
The transversality condition (H2) could be replaced by
the more general assumption that
the set of periodic solutions
in the vicinity of $\bar U$ form a
smooth $3$-dimensional manifold \eqref{manifold}.
However, it is readily seen in this context that
(H2) is then implied by the spectral stability condition (D3)
of Section \ref{bloch}; that is, transversality is
necessary for our notion of spectral, or Evans, stability.
This situation is reminsiscent of that of the viscous shock
case; see, for example, \cite[S\ 1.2.3]{ZH}, or \cite{MaZ3,Z1}.
}
\er
\br\label{relaxrmk}
\textup{
Note that \eqref{eqn:1conslaw} is of $2\times 2$ viscous relaxation type
\be\label{relax}
U_t+f(U)_x-\nu (B(U)U_x)_x=\bp0\\q(U)\ep,
\quad q(U)=1-\tau u^2,
\ee
where $q_u=-2u \tau<0$ for solutions $u>0$ progressing down the ramp.
Thus, constant solutions are stable so long as the subcharacteristic
condition $\Big|\frac{u^3}{2}\Big|< \Big|\frac{u^3}{\sqrt{F}}\Big|$,
is satisfied, or $F<4$.
When the subcharacteristic condition is violated,
roll waves appear through Hopf bifurcation as parameters are varied
through the minimum speed $c_{\rm min}= \frac{1}{\sqrt{F \tau_0^3}}$;
see Appendix \ref{s:hopf}.
For $\nu=0$, violation of the subcharacteristic condition is
associated with subshocks and the appearance of discontinuous
roll waves observed by Dressler \cite{D};
see \cite{JK} for related, more general, discussion.
}
\er
\br\label{singrmk}
\textup{
The limit $\nu\to 0$ represents an interesting singular
perturbation problem in which the structure of the profile equations
simplifies, decoupling into fast and slow scalar components,
and converging to inviscid Dressler waves \cite{D,N1} in an
appropriate regime \cite{N2}.
This would be an interesting setting in which to investigate the
associated spectral stability problem.
Another interesting limit is Hopf bifurcation from the constant solution occurring at minimum speed of existence \cite{N2},
treated here in Section \ref{s:hopf}; see Remark \ref{instabrmk}.
}
\er
\subsubsection{Linearized equations}\label{evans}
Making the change of variables $x\to x-ct$ to
co-moving coordinates, we convert \eqref{eqn:1conslaw} to
\ba \label{eqn:co1conslaw}
\tau_t-c\tau_x - u_x&= 0,\\
u_t-cu_x + ((2F)^{-1}\tau^{-2})_x&=
1- \tau u^2 +\nu (\tau^{-2}u_x)_x ,
\ea
and the traveling-wave solution
to a stationary solution $U=\bar U(x)$
convenient for stability analyis.
Writing \eqref{eqn:co1conslaw} in abstract form
\be\label{ab}
U_t +f(U)_x=(B(U)U_x)_x +g(U)
\ee
and linearizing (\ref{eqn:co1conslaw}) about $\bar{U}(\cdot)$, we obtain
\be \label{e:lin}
v_t = Lv := (\partial_x B\partial_x -\partial_x A +C) v,
\ee
where the coefficients
\ba\label{coeffs}
A&:= df(\bu) - (dB(\bu) (\cdot) )\bu_x
=\bp -c & -1\\ -\bar \tau^{_-3}(F^{-1}- 2\nu \bar u_x)&-c\ep ,\\
B&:=B(\bu)= \bp 0&0\\0 & \nu \bar \tau^{-2}\ep,
\quad C:=dg(\bar U)=\bp 0 & 0\\-\bar u^2& -2\bar u\bar \tau\ep
\ea
are periodic functions of $x$. As the underlying solution $\bar{U}$ depends on $x$ only,
equation \eqref{e:lin} is clearly autonomous in time.
By separation of variables, therefore,
decomposing solutions into the sum
of solutions of form $v(x,t)=e^{\lambda t}v(x)$,
where $v$ satisfies the eigenvalue equation $(L-\lambda)v=0$,
or, equivalently, by taking the Laplace transform,
we may reduce the study of stability of $\bar U$ to
the study of the spectral properties of the linearized operator $L$.
As the coefficients of $L$ are $X$-periodic,
Floquet theory
implies that
its spectrum is purely continuous.
Moreover, its spectral properties may be conveniently analyzed by
Bloch decomposition, an analog for periodic-coefficient operators
of the Fourier decomposition of a constant-coefficient operator,
as we now describe.
\subsubsection{Bloch decomposition
and stability conditions}\label{bloch}
Following \cite{G,S1,S2,S3}, we define the family of operators
\be \label{e:Lxi}
L_{\xi} = e^{-i \xi x} L e^{i \xi x}
= (\partial_x+i\xi) B(\partial_x+i\xi)
-(\partial_x+i\xi) A +C
\ee
operating on the class of $L^2$ periodic functions on $[0,X]$;
the $(L^2)$ spectrum
of $L$ is equal to the union of the
spectra of all $L_{\xi}$ with $\xi$ real with associated
eigenfunctions
\be
w(x, \xi,\lambda) := e^{i \xi x} q(x, \xi, \lambda),
\label{e:efunction}
\ee
where $q$, periodic, is an eigenfunction of $L_{\xi}$.
By standard considerations \cite{N2},\footnote{
For example, the characterization \cite{G} of spectra
as the zero set of an associated Evans function.}
the spectra of $L_{\xi}$
consist of the union of countably many continuous
surfaces $\lambda_j(\xi)$.
Without loss of generality taking $X=1$,
recall now the {\it Bloch representation}
\be\label{Bloch}
u(x)=
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
e^{i\xi\cdot x}\hat u(\xi, x) d\xi
\ee
of an $L^2$ function $u$, where
$\hat u(\xi, x):=\sum_k e^{2\pi ikx}\hat u(\xi+ 2\pi k)$
are periodic functions of period $X=1$, $\hat u(\cdot)$
denoting with slight abuse of notation the Fourier transform of $u$
in $x$.
By Parseval's identity, the Bloch transform
$u(x)\to \hat u(\xi, x)$ is an isometry in $L^2$:
\be\label{iso}
\|u\|_{L^2(x)}=
\|\hat u\|_{L^2(\xi; L^2(x))},
\ee
where $L^2(x)$ is taken on $[0,1]$ and $L^2(\xi)$ on $[-\pi,\pi]$.
Moreover, it diagonalizes the periodic-coefficient operator $L$,
yielding the {\it inverse Bloch transform representation}
\be\label{IBFT}
e^{Lt}u_0=
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
e^{i\xi \cdot x}e^{L_\xi t}\hat u_0(\xi, x)
d\xi\
\ee
relating behavior of the linearized system to
that of the diagonal operators $L_\xi$.
Following \cite{JZ4}, we assume along with (H1)--(H2) the
{\it strong spectral stability} conditions:
(D1) $\sigma(L_\xi) \subset \{ \hbox{\rm Re} \lambda <0 \} $ for $\xi\ne 0$.
(D2) $\hbox{\rm Re} \sigma(L_{\xi}) \le -\theta |\xi|^2$, $\theta>0$,
for $\xi\in \R$ and $|\xi|$ sufficiently small.
(\DDD) $\lambda=0$ is an eigenvalue
of $L_{0}$ of multiplicity $2$.\footnote{
The zero eigenspace of $L_0$,
corresponding to variations along the $3$-dimensional manifold
of periodic solutions in directions for which period does
not change \cite{Se1,JZ4}, is at least $2$-dimensional
by linearized existence theory and (H2).
}
As shown in \cite{N2}, (H1)-(H2) and (D1)--(\DDD)
imply that there exist $2$ smooth eigenvalues
\be\label{e:surfaces}
\lambda_j(\xi)= -i a_j \xi +o(|\xi|)
\ee
of $L_\xi$ bifurcating from $\lambda=0$ at $\xi=0$;
see Lemma \ref{blochfacts} below.
Loosely following \cite{JZ4}, we make the further nondegeneracy hypotheses:
(H3) The coefficients $ a_j$ in \eqref{e:surfaces} are distinct.
(H4) The eigenvalue $0$ of $L_0$ is nonsemisimple, i.e., $\dim
\ker L_0=1$.
\noindent
The coefficients $a_j$ may be seen to be the characteristics
of an associated Whitham averaged system
\ba\label{whit}
M(\beta)_t + G(\beta)_x&=0,\\
\Omega(\beta)_t + (c(\beta) \Omega(\beta))_x&=0
\ea
linearized about the values of $M$, $G$, $c$, $\Omega$ associated
with the background wave $\bar u$,
where $M$ is the mean of $\tau$ over one period and
$F$ the mean in the $\tau$-coordinate
of a certain associated flux, $c$ is wave-speed,
and $\Omega$ frequency of nearby periodic solutions,
indexed as in \eqref{manifold} by $\beta\in \R^2$;
see \cite{N2,OZ3,OZ4}.\footnote{
Here, we follow the formalism and notation of \cite{OZ3,OZ4}.
}
System \eqref{whit} formally governs slowly modulated solutions
\be\label{formod}
\tilde u(x,t)=\bar u^{\beta(\eps x,\eps t)}(\Psi(x,t))+O(\eps),\qquad \eps\to 0
\ee
presumed to describe large spatio-temporal behavior $x$, $t\gg 1$,
where $\bar u^\beta(\cdot)$
as in \eqref{manifold} parametrizes the set of nearby periodic solutions,
$\Omega=\Psi_x$, and $c=-\Psi_t/ \Psi_x$.
Thus, (D1) implies weak hyperbolicity
of the Whitham averaged system \eqref{whit} (reality of $a_j$),
while (H3) corresponds to strict hyperbolicity.
Condition (H4) holds generically, and corresponds to the assumption
that speed $c$ is nonstationary along the manifold of nearby
stationary solutions; see Lemma \ref{blochfacts}.\footnote{
The case that (H4) is violated may be treated as in \cite{JZ3}.
}
Condition (D2) corresponds to ``diffusivity'' of
the large-time ($\sim$ small frequency) behavior of the linearized system,
and holds generically given (H1)--(H4), (D1), and (D3').\footnote{
This amounts to nonvanishing of $b_j$ in the Taylor series expansion
$\lambda_j(\xi)=-ia_j\xi- b_j\xi^2$
guaranteed by Lemma \ref{blochfacts} given (H1)--(H4), (D1), and (D3').}
Condition (\DDD) also holds generically, and can be verified by
an Evans function computation as described in \cite{N1}.
As discussed in \cite{OZ1,Se1,JZ3,JZ4},
conditions (D1)--(D3') are conservation law analogs of
the spectral assumptions
introduced by Schneider in the reaction-diffusion case \cite{S1,S2,S3}.
\subsection{Main result}
\begin{theo}\label{main}
Assuming (H1)--(H4) and (D1)--(\DDD),
let $\bar U=(\bar \tau, \bar u)$ be a traveling-wave solution
\eqref{eqn:tw} of \eqref{eqn:1conslaw} satisfying
the derivative condition
\be\label{froudebd}
\nu \bar u_x < F^{-1}.
\ee
Then, for some $C>0$ and $\psi \in W^{K,\infty}(x,t)$, where $K\geq 3$ is as in (H1)
\ba\label{eq:smallsest}
\|\tilde U-\bar U(\cdot -\psi-ct)\|_{L^p}(t)&\le
C(1+t)^{-\frac{1}{2}(1-1/p)}
\|\tilde U-\bar U\|_{L^1\cap H^K}|_{t=0},\\
\|\tilde U-\bar U(\cdot -\psi-ct)\|_{H^K}(t)&\le
C(1+t)^{-\frac{1}{4}}
\|\tilde U-\bar U\|_{L^1\cap H^K}|_{t=0},\\
\|(\psi_t,\psi_x)\|_{W^{K+1,p}}&\le
C(1+t)^{-\frac{1}{2}(1-1/p)}
\|\tilde U-\bar U\|_{L^1\cap H^K}|_{t=0},\\
\ea
and
\ba\label{eq:stab}
\|\tilde U-\bar U(\cdot-ct)\|_{ L^\infty}(t), \; \|\psi(t)\|_{L^\infty}&\le
\|\tilde U-\bar U\|_{L^1\cap H^K}|_{t=0}
\ea
for all $t\ge 0$, $p\ge 2$,
for solutions $\tilde U$ of \eqref{eqn:1conslaw} with
$\|\tilde U-\bar U\|_{L^1\cap H^K}|_{t=0}$ sufficiently small.
In particular, $\bar U$ is nonlinearly bounded
$L^1\cap H^K\to L^\infty$ stable.
\end{theo}
Theorem \ref{main} asserts not only
bounded $L^1\cap H^K \to L^\infty$ stability, a very weak notion of stability,
but also asymptotic convergence of $\tilde U$ to the modulated wave
$\bar U(x-\psi(x,t))$.
\br\label{nonlinp}
\textup{
With further effort, it may be shown that the results of
Theorem \ref{main} extend to all $1\le p\le \infty$ using the pointwise
techniques of \cite{OZ2}; see discussion, \cite{JZ3,JZ4}.
}
\er
\br\label{froudermk1}
\textup{
The derivative condition \eqref{froudebd} is effectively an upper bound on
the amplitude of the periodic wave; see Remark \ref{froudermk}.
As discussed in Remark \ref{damprmk}, this is precisely the condition
that the first-order part of the linearized equations \eqref{e:lin}
be symmetric hyperbolic
(i.e., that $A$ in \eqref{coeffs} be symmetrizable),
and reflects a subtle competition between
hyperbolic and parabolic effects.
(The first-order part of the inviscid equations
is always symmetric--hyperbolic, corresponding to the equations
of isentropic gas dynamics with $\gamma$-law gas.)
It is satisfied when either wave amplitude or viscosity coefficient $\nu$
is sufficiently small.
It is not clear whether this condition may be relaxed.
}
\textup{
We note that condition \eqref{froudebd} is satisfied
for {\it all} roll-waves computed numerically in \cite{N2}.
For, in Eulerian coordinates, this condition translates to
$ h_x/h<(c\nu F)^{-1}$.
Examining Fig. 1 of \cite{N2},
a phase portrait in $(h,h')$ for $F=6$, $\nu=0.1$, and
$1.89<c<1.91$, we see that all periodic orbits appear to lie
beneath the line $h'/h=.68 $, whereas $(c\nu F)^{-1} \approx .88$.
}
\er
It is straightforward using the bounds of Corollary
\ref{greenbds} to show for ``zero-mass'', or derivative,
initial perturbations, that nonlinear decay rates
\eqref{eq:smallsest}--\eqref{eq:stab} improve by factor $(1+t)^{-1/2}$,
to the rates seen in the reaction-diffusion case \cite{S1,JZ5} for
general (undifferentiated) localized perturbations.
In particular, the perturbed wave $\tilde U$ then decays
asymptotically in $L^\infty$ to the background wave $\bar U$
with Gaussian rate $(1+t)^{-1/2}$ as in the reaction-diffusion case.
Likewise, under an unlocalized initial perturbation, or,
equivalently, the integral of a localized perturbation
the difference between $\tilde U$ and $\bar U$ may be expected
to blow up at rate $(1+t)^{1/2}$- this is indeed the linearized
behavior- and, barring special nonlinear structure, there
seems no reason why the difference between $\tilde U$ and the
modulation $\bar U(\cdot-\Psi)$ should not blow up as well:
at best it remains bounded.
In the reaction-diffusion case, for comparison,
results announced in \cite{SSSU} assert
that $\tilde U$ remains close to $\bar U$
even under unlocalized perturbations, and approaches
the modulated wave at rate $(1+t)^{-1/2}$ in $L^\infty$.
That is, the behavior in the conservation (balance) law case compared
to that in the reaction-diffusion case is, roughly speaking,
shifted by one derivative.\footnote{
At a purely technical level, this can be seen by the appearance of
a Jordan block in the zero eigenspace of $L_0$, introducing factor $\xi^{-1}$ in the
description of low-frequency behavior (Lemma \ref{blochfacts}).
Recall that a factor $i\xi$ corresponds roughly to differentiation in the
Bloch representation, through its relation to the Fourier transform.
In the reaction--diffusion case, the zero eigenspace of $L_0$ is simple,
and no such factor appears.
}
This reflects a fundamental difference between modulational behavior
in the present, conservation (or balance) law setting from that
of the reaction--diffusion case.
Namely, in the reaction-diffusion case, the Whitham averaged system
reduces to a single equation $\partial_t (\Omega)+ \partial_x (\Omega c)=0$,
or, equivalently,
\be\label{eq:rdmodel}
\Psi_t + c(\Psi_x)\Psi_x=0,
\ee
where $\Omega:=\Psi_x$ denotes frequency and $c:=-\frac{\Psi_t}
{\Psi_x}$ wave speed, and $c$ and $\Omega$ are related by the linearized
dispersion relation along the family of periodic orbits
(in the case considered by Schneider \cite{S1}, $c\equiv 0$).
On the other hand, the Whitham averaged equations \eqref{whit} in the present
case are a genuine $2\times 2$ first-order hyperbolic system\footnote{
In general, the dimension of the Whitham averaged system is
equal to the dimension of the manifold of nearby periodic solutions,
modulo translations \cite{JZ4}.
}
in $\Psi_x$ and wave-speed $c$,
$c$ now considered as an independent parameter; that is,
they describe modulation of the perturbed wave in frequency
$\Psi_x$ and speed $c$, with phase shift $\Psi$ determined
indirectly by integration of $\Psi_x$.
Assuming heuristically (as justified
at the linearized, spectral, level
by the Bloch analysis of Section \ref{prep}),
that modulational behavior is governed by
a second-order regularization of the first-order Whitham averaged system,
we have the standard picture of behavior under localized perturbation
as consisting of modulations in $(\Psi_x, c)$ given by
a pair of approximate Gaussians
propagating outward with Whitham characteristic speeds $a_1$ and $a_2$,
hence an associated, much larger modulation in $\Psi$ determined
by integration in the $\Psi_x$ component, given
by a sum of approximate errorfunctions propagating with the same speeds.
Indeed,
this is exactly the description given in \eqref{eq:errfns}
of the principal part of the kernel $e(x,t;y)$
determining $\Psi$ through \eqref{psi}.
Likewise, the principal part of the Green function of the linearized
equations about $\bar U$ is $\bar U'(x)e(x,t;y)$, showing that linearized
behavior to lowest order indeed consists of a translation, or multiple
of $\bar U'(x)$, with amplitude
$$
\Psi(x,t)=\int e(x,t;y)(\tilde U(y,0)-\bar U(y,0))dy;
$$
see the description of the Green function in Corollary \ref{greenbds}.
The same considerations show that the rate of convergence of $\tilde U$
to the modulation $\bar U(\cdot -\Psi)$ cannot be improved (or,
in the case of a nonlocalized perturbation, recovered) by modulating
in additional parameters such as wave speed $c$ or etc.
For, as indicated by the above discussion, all such modulations
represent smaller contributions by factor $(1+t)^{-1/2}$ than
that of the phase $\Psi$, comparable rather to frequency $\Psi_x$,
and thus may be ignored in consideration of blow-up vs.stability.
This picture of modulational behavior as ``filtering'' by integration
along a certain direction of the hyperbolic--parabolic system derived
by Whitham averaging
seems quite interesting at a phenomenological level, and
a genuinely novel aspect of the conservation (balance) law case.
In particular, the $\Psi_x$ component direction along which the integration
is performed is in general
independent of either characteristic mode, so that the resulting
behavior is essentially different from that exemplified by
\eqref{eq:rdmodel} of a single scalar equation as in
the reaction--diffusion case.
\subsection{Discussion and open problems}\label{s:discussion}
The extension from the parabolic conservation law to the
present case involves a number of new technical issues associated with
lack of parabolicity and nonconservative form.
We overcome these difficulties by combining the
arguments of \cite{JZ3,JZ4}, \cite{N2} with those of
\cite{MaZ4,Z1,TZ1} (real viscosity) and
\cite{MaZ1} and \cite{LRTZ,TZ2} (relaxation and combustion systems
both involving nonconservative terms).
An interesting open problem is the rigorous justification
of spectral stability of
roll waves approaching the inviscid case in the singular zero
viscosity limit, extending results of \cite{N2}.
We hope to carry this out in future work.
For related asymptotic analysis,
see the study in \cite{Z2} of the inviscid limit for detonations.
Another interesting open problem is the
numerical investigation of spectral stability of large-amplitude
roll waves.
In particular, it is an interesting question
whether violation of the apparently technical
``amplitude condition'' \eqref{froudebd}
corresponds to actual physical phenomena/instability.
This is not inconceivable, as \eqref{froudebd}
is needed in our argument
not only for nonlinear iteration, but also for high-frequency
linearized bounds. As the condition that the first-order
part of the equations be symmetric hyperbolic, it may well have
such significance-- however, this is not yet clear.
It is straightforward to extend our results
to the two-dimensional small-amplitude case, by
working in Eulerian coordinates and substituting for the present
large-amplitude damping estimate the simpler small-amplitude version
of \cite{MaZ2}; see \cite{JZ3,JZ4} for the multi-dimensional analysis
of periodic waves.
However, there is some evidence that roll waves develop transverse
instabilities in multi-dimensions \cite{N3}.
If so, this suggests the question whether such instability might
be connected with bifurcation to multiply periodic waves.
The extension of our stability analysis to the multiply periodic
case, as suggested in \cite{JZ3,JZ4},
would be another very interesting open problem.
\section{Spectral preparation}\label{prep}
We begin by a careful study of the Bloch perturbation expansion near $\xi=0$.
\begin{lemma}\label{blochfacts}
Assuming (H1)--(H4), (D1), and (\DDD), the eigenvalues
$\lambda_j(\xi)$ of $L_\xi$
are analytic functions and the Jordan structure of the zero
eigenspace of $L_0$ consists
of a $1$-dimensional kernel and a single Jordan chain of height $2$,
where the left kernel of $L_0$ is spanned by the constant
function $\tilde f\equiv (1,0)^T$, and $\bar u'$ spans the right
eigendirection lying at the base of the Jordan chain.
Moreover, for $|\xi|$ sufficiently small,
there exist right and left eigenfunctions
$q_j(\xi, \cdot)$ and $\tilde q_j(\xi, \cdot)$
of $L_\xi$ associated with $\lambda_j$ of form
$q_j=\sum_{k=1}^2 \beta_{j,k} v_k$ and $\tilde q_j=\sum_{k=1}^2 \tilde \beta_{j,k} \tilde v_k$
where $\{v_j\}_{j=1}^2$ and $\{\tilde v_j\}_{j=1}^2$ are dual bases of the total
eigenspace of $L_\xi$ associated with sufficiently small eigenvalues,
analytic in $\xi$,
with $\tilde v_2(0)$ constant and $v_1(0) \equiv \bar u'(\cdot)$;
$\xi^{-1}\tilde \beta_{j,1}, \tilde \beta_{j,2}$
and $\xi \beta_{j,1}, \beta_{j,2}$
are analytic in $\xi$; and $\langle \tilde q_j,q_k\rangle= \delta_j^k$.
\end{lemma}
\br
\textup{
Notice that the results of Lemma \ref{blochfacts} are somewhat unexpected since, in general, eigenvalues
bifurcating from a non-trivial Jordan block typically do so in a nonanalytic fashion,
rather being expressed in a Puiseux series in fractional
powers of $\xi$.\footnote{
This is, however, consistent with the picture
of behavior as being approximately governed by a
first-order Whitham averaged system with eigenvalue perturbation expansions
agreeing to first-order with the associated linearized homogeneous dispersion relation
\cite{N2,OZ3,OZ4}.
}
The fact that analyticity prevails in our situation is a consequence of the very special structure
of the left and right generalized null-spaces of the unperturbed operator $L_0$, and the special
forms of the equations considered.
}
\er
\begin{proof}
Recall that $L_\xi$ has spectrum consisting of isolated eigenvalues
of finite multiplicity \cite{N2,G}.
Expanding
\be \label{Lpert}
L_\xi=L_0 + i\xi L^1- \xi^2L^2,
\ee
where, by \eqref{e:Lxi},
\ba \label{Ls}
L_0&=\partial_x B\partial_x - \partial_x A + C,
\quad
L^1= (B\partial_x + \partial_x B -A),
\quad
L^2= B,
\ea
consider the spectral
perturbation problem in $\xi$ about the eigenvalue $\lambda=0$ of $L_0$.
Because $0$ is an isolated eigenvalue of $L_0$, the associated total
right and left eigenprojections $P_0$ and $\tilde P_0$ perturb
analytically in $\xi$,
giving projection $P_\xi$ and $\tilde P_\xi$ \cite{K}.
These yield in standard fashion
(for example, by projecting appropriately chosen fixed subspaces)
locally analytic right and left bases $\{v_j\}$ and $\{\tilde v_j\}$
of the associated total eigenspaces given by
the range of $P_\xi$, $\tilde P_\xi$.
Defining $V=(v_1, v_{2})$ and
$\tilde V=(\tilde v_1, \tilde v_{2})^*$, $*$ denoting
adjoint, we may convert the infinite-dimensional
perturbation problem \eqref{Lpert} into a $2\times 2$
matrix perturbation problem
\be\label{Mpert}
M_\xi=M_0+ i\xi M_1 - \xi^2 M_2+O(|\xi|^3),
\ee
where $M_\xi:= \left<\tilde V_\xi^*, L_\xi V_\xi\right>$ and $\left<\cdot,\cdot\right>$
denotes the standard $L^2(x)$ inner product on the finite interval $[0,X]$.
That is, the eigenvalues $\lambda_j(\xi)$
lying near $0$ of $L_\xi$ are the eigenvalues
of $M_\xi$, and the associated right and left eigenfunctions
of $L_\xi$ are
\be\label{vecrel}
f_j=V w_j \;\hbox{\rm and } \;
\tilde f_j=\tilde w_j \tilde V^* ,
\ee
where $w_j$ and $\tilde w_j$
are the associated right and left eigenvectors of $M_\xi$.
By assumption, $\lambda=0$ is a nonsemisimple eigenvalue
of $L_0$, so that $M_0$ is nilpotent but nonzero, possessing a
nontrivial associated Jordan chain.
Moreover, using the fact that $\left<(1,0)^T, C\right>=0$, where, again, $\left<\cdot,\cdot\right>$ represents
the $L^2(x)$ inner product over the finite domain $[0,X]$, the function $\tilde f\equiv (1,0)^T$
by direct computation lies
in the kernel of $L_0^*=\left(\partial_xB^*\partial_x+A^*\partial_x+C^*\right)$,
we have that the
two-dimensional zero eigenspace of $L_0$ is consists precisely
of a one-dimensional kernel and a single Jordan chain of height two.
Moreover, by translation-invariance (differentiate in $x$
the profile equation \eqref{e:profile}), we have $L_0\bar u'=0$,
so that $\bar u'$ lies in the right kernel of $L_0$.
Now, recall assumption (H2) that $H: \,
\R^5 \rightarrow \R^2$
taking $(X,c,q,b) \mapsto (\tau,\tau')(X,c,b; X)-b$
is full rank at $(\bar{X},\bar c, \bar b)$,
where $(\tau,\tau')(\cdot;\cdot)$ is the solution operator of \eqref{e:profile}.
The fact that $\ker L_0$ is one-dimensional implies that
the restriction
$\check H$ taking
$
(b, q) \mapsto u(X; b, c, q)-b
$
for fixed $(X,c)$
is also full rank, i.e., $H$ is full rank with respect to
the specific parameters $(X,c)$.
Applying the Implicit Function Theorem and counting dimensions,
we find that the set of periodic solutions, i.e., the inverse
image of zero under map $H$ local to $\bar u$
is a smooth three-dimensional manifold
$\{\bar u^\beta (x-\alpha-c(\beta )t)\}$,
with $\alpha\in \RR$, $\beta \in \RR^{2}$.
Moreover, two dimensions may be parametrized by $(X,c)$,
or without loss of generality $\beta =(X,c)$.
Fixing $X$ and varying $c$, we find by differentiation of \eqref{e:profile}
that $f_*:=-\partial_s \bar U$ satisfies the generalized eigenfunction
equation
$$
L_0 f_*= \bar U'.
$$
Thus, $\bar U'$ spans the eigendirection lying at the base of the
Jordan chain, with the generalized zero-eigenfunction of $L_0$
corresponding to variations in speed along the manifold of periodic
solutions about $\bar U$.
Without loss of generality, therefore, we may take
$\tilde v_{2}$ to be constant at $\xi=0$, and
$v_{1} \equiv \bar U'$ at $\xi=0$.
Noting as in \cite{JZ3} the fact that, by \eqref{coeffs},
\ba\label{comp}
A\bar U_x= f(\bu)_x - (\partial_x B(\bu))\bar U_x
&= \partial_x( f(\bu)_x - B(\bU)\bU_x) +B(\bu)\partial_x \bu_x
\\
& = g(\bU) +B(\bu)\partial_x \bu_x ,
\ea
and so by $e_2g=0$, $\partial_x e_2=0$, we have
$$
\langle e_2,L^1 \bar U'\rangle=
\langle e_2,(\partial_{x} B+B\partial_x -A )\bar U'\rangle=
\langle e_2, \partial_{x}B \bar U' \rangle
\equiv 0
$$
for $e_2:=(0,1)$,
where $\langle \cdot, \cdot\rangle$
denotes $L^2(x)$ inner product on the interval $x\in [0,X]$,
we find under this normalization that \eqref{Mpert} has the special structure
\ba\label{Mstructure}
M_0=\bp
0 & 1\\
0 & 0\ep,
\qquad
M_1=\bp
* & *\\
0 & *\ep.
\ea
Now, rescaling \eqref{Mpert} as
\be\label{rescale}
\check M_\xi:= (i\xi)^{-1} S(\xi)M_\xi S(\xi)^{-1},
\ee
where
\be\label{S}
S:=\bp
i \xi & 0\\
0 & 1\\
\ep,
\ee
we obtain
\be\label{checkMpert}
\check M_\xi=
\check M_0 + i\xi\check M_1 + O(\xi^2),
\ee
where $\check M_j= \check M_j$ like the original
$M_j$ are constant
and the eigenvalues $m_j(\xi)$ of $\hat M_\xi$ are
$(i\xi)^{-1}\lambda_j(\xi)$.
As the eigenvalues $m_j$ of $\check M_\xi$ are continuous,
the eigenvalues $\lambda_j(\xi)=i\xi m_j$ are differentiable
at $\xi=0$ as asserted in the introduction.
Moreover, by (H3), the eigenvalues $\check\lambda_j(0)$
of $\check M_0$ are distinct, and so they perturb analytically
in $\xi$, as do the associated right and left eigenvectors
$z_j$ and $\tilde z_j$.
Undoing the rescaling \eqref{rescale},
and recalling \eqref{vecrel}, we obtain the result.
\end{proof}
\section{Linearized stability estimates}\label{linests}
By standard spectral perturbation theory \cite{K}, the total
eigenprojection $P(\xi)$ onto the eigenspace of $L_\xi$
associated with the eigenvalues $\lambda_j(\xi)$, $j=1,2$
described in the previous section
is well-defined and analytic in $\xi$ for $\xi$ sufficiently small,
since these (by discreteness of the spectra of $L_\xi$) are
separated at $\xi=0$ from the rest of the spectrum of $L_0$.
By (D2), there exists an $\eps>0$ such that $\Re\lambda_j(\xi)\leq-\theta|\xi|^2$ for $0<|\xi|<2\eps$.
With this choice of $\eps$, we introduce a smooth cutoff function $\phi(\xi)$ that
is identically one for $|\xi|\le \eps$ and identically
zero for $|\xi|\ge 2\eps$, $\eps>0$ sufficiently small,
we split the solution operator $S(t):=e^{Lt}$ into
a low-frequency part
\be\label{SI}
S^I(t)u_0:=
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
e^{i\xi \cdot x}
\phi(\xi)P(\xi) e^{L_\xi t}\hat u_0(\xi, x) d\xi
\ee
and the associated high-frequency part
\be\label{SII}
S^{II}(t)U_0:=
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
e^{i\xi \cdot x}
\big(I-\phi P(\xi)\big)
e^{L_\xi t}\hat U_0(\xi, x)
d\xi.
\ee
Our strategy is to treat the high- and low-frequency operators separately since,
as is standard, the low-frequency analysis is considerably more complicated
than the corresponding high-frequency analysis.
That being said,
we begin by deriving bounds on the solution operator at
high frequency.
\subsection{High-frequency bounds}\label{HF}
By boundedness of the resolvent on compact subdomains of the resolvent
set, equivalence (as the zero-set of an associated Evans function \cite{N2,OZ1})
of $H^1$ and $L^2$ spectrum,
Assumption (D2),
and the high-frequency estimates of Lemma \ref{resbd},
we have for $|\xi|$ bounded away from zero and waves satisfying the amplitude condition \eqref{froudebd} that
the resolvent $(\lambda-L_\xi)^{-1}$ is uniformly bounded
from $H^1\to H^1$ for $\Re \lambda=-\eta< \theta<0$,
whence, by Pr\"uss' Theorem \cite{Pr},
$\|e^{L_\xi t}f\|_{H^1}\le Ce^{-\theta t} \|f\|_{H^1}$.
For $|\xi|$ sufficiently small, on the other hand, $\phi\equiv 1$,
and $I-\phi(\xi)P=I-P=Q$, where $Q$ is the eigenprojection
of $L_\xi$ associated with eigenvalues complementary to $\lambda_j(\xi)$,
which by spectral separation of $\lambda_j(\xi)$ from the
remaining spectra of $L_\xi$, have real parts strictly less than zero.
Applying Pr\"uss' Theorem to the restriction of $L_\xi$
to the Hilbert space given by the range of $Q$, we find, likewise,
that
$\| e^{L_\xi t} (I-\phi(\xi)) f\|_{H^1}=
\|e^{L_\xi t}Qf\|_{H^1} \le Ce^{-\theta t} \|f\|_{H^1}$.
Combining these observations,
we have the exponential decay bound
$$
\|e^{L_\xi t}(I-\phi P(\xi))f\|_{H^1([0,X])}\le Ce^{-\theta t}\|f\|_{H^1([0,X])}$$
for $\theta>0$ as in (D2) and $C>0$, from which it follows
\ba\label{semigp}
\|e^{L_\xi t}(I-\phi P(\xi))\partial_{x}^l f\|_{H^{1}([0,X])}
&\le C e^{-\theta t}\|f\|_{H^{l+1}([0,X])}\\
\ea
for $0\le l\le K$ ($K$ as in (H1)).
Together with (\ref{iso}), these give immediately the
following estimates.
\begin{proposition}[\cite{OZ4}]\label{p:hf}
Under assumptions (H1)--(H4), (D1)--(D2), and assuming the amplitude condition \eqref{froudebd} holds,
there exists constants $\theta$, $C>0$, such that for all
all $t>0$, $2\le p\le \infty$, $0\le l\le 2$, $0\le m\le 2$, we have the high-frequency estimates
\ba\label{SIIest}
\|S^{II}(t)\partial_x^l f\|_{L^2(x)}&\le
Ce^{-\theta t}\|f\|_{H^{l+1}(x)},\\
\|S^{II}(t) \partial_x^m f\|_{L^p(x)}&\le
C e^{-\theta t}\|f\|_{H^{m+2}(x)}.
\ea
\end{proposition}
\begin{proof}
For $m,l=0$,
the first inequalities follow immediately by (\ref{iso})
and \eqref{semigp}.
The second follows for $p=\infty$ by Sobolev embedding.
The result for general $2\le p\le \infty$ then follows by
$L^p$ interpolation.
A similar argument applies for $1\le l, m\le 2$ by higher-derivative
versions of \eqref{semigp}, which follow in exactly the same way.
\end{proof}
\subsection{Low-frequency bounds}\label{LF}
As noted above, analysis of the solution operator at low frequency
is considerably more complicated than the
high-frequency bounds outlined above.
To aid in our analysis, we introduce the Green kernel
\be\label{GI}
G^I(x,t;y):=S^I(t)\delta_y(x)
\ee
associated with $S^I$, and the corresponding kernel
\be\label{GIxi}
[G^I_\xi(x,t;y)]:=\phi(\xi)P(\xi) e^{L_\xi t}[\delta_{y}(x)]
\ee
appearing within the Bloch representation
of $G^I$, where the brackets on $[G_\xi]$ and $[\delta_y]$
denote the periodic extensions of these functions onto the whole line.
Then, we have the following descriptions of $G^I$, $[G^I_\xi]$,
deriving from the
spectral expansion \eqref{e:surfaces} of $L_\xi$ near $\xi=0$.
\begin{proposition}[\cite{OZ4}]\label{kernels}
Under assumptions (H1)--(H4) and (D1)--(\DDD),
\ba\label{Gxi}
[G^I_\xi(x,t;y)]&= \phi(\xi)\sum_{j=1}^{2}e^{\lambda_j(\xi)t}
q_j(\xi,x)\tilde q_j(\xi, y)^*,\\
G^I(x,t;y)&=
\Big(\frac{1}{2\pi }\Big) \int_{\R} e^{i\xi \cdot (x-y)}
[G^I_\xi(x,t;y)] d\xi \\
&=
\Big(\frac{1}{2\pi }\Big) \int_{\R}
e^{i\xi \cdot (x-y)}
\phi(\xi)
\sum_{j=1}^{2}e^{\lambda_j(\xi)t} q_j(\xi,x)\tilde q_j(\xi, y)^*
d\xi,
\ea
where $*$ denotes matrix adjoint, or complex conjugate transpose,
$q_j(\xi,\cdot)$ and $\tilde q_j(\xi,\cdot)$
are right and left eigenfunctions of $L_\xi$ associated with eigenvalues
$\lambda_j(\xi)$ defined in \eqref{e:surfaces},
normalized so that $\langle \tilde q_j,q_j\rangle\equiv 1$.
\end{proposition}
\begin{proof}
Relation (\ref{Gxi})(i) is immediate from the spectral decomposition
for $C^0$ semigroups at eigenvalues of finite multiplicity, and the fact that $\lambda_j$
are distinct for $|\xi|>0$ sufficiently small, by (H3).
Substituting (\ref{GI}) into (\ref{SI})
and computing
\be\label{comp1}
\widehat{\delta_y}(\xi,x)=
\sum_k e^{2\pi i kx}\widehat{\delta_y}(\xi + 2\pi k e_1)=
\sum_k e^{2\pi i kx}e^{-i\xi \cdot y-2\pi i ky}
= e^{-i\xi \cdot y}[\delta_{y}(x)],
\ee
where the second and third equalities follow from the fact that the Fourier transform of either the continuous or discrete
the delta-function is unity,
we obtain
\ba\label{GIsub}
G^I(x,t;y)&=
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
e^{i\xi \cdot x} \phi P(\xi) e^{L_\xi t} \widehat{\delta_y}(\xi,x)d\xi\\
\nonumber
&=
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
e^{i\xi \cdot (x-y)} \phi P(\xi)e^{L_\xi t} [\delta_{y}(x)] d\xi,
\ea
yielding (\ref{Gxi})(ii) by (\ref{GIxi})(i) and the fact that $\phi$
is supported on $[-\pi,\pi]$.
\end{proof}
We now state our main result for this section, which
uses
the spectral representation of $G^I$ and $[G^I_{\xi}]$
described in Proposition \ref{kernels} to
decompose
the low-frequency
Green kernel into a leading order piece
(corresponding to translational modulation) plus a faster decaying residual.
Underlying this decomposition is the fundamental relation
\be
G(x,t;y)=
\left(\frac{1}{2\pi }\right)\int_{-\pi}^{\pi}\int_{\R^{d-1}}
e^{i\xi \cdot (x-y)}[G_\xi(x_1,t;y_1)]d\xi,
\ee
which
serves as
the crux of the low-frequency analysis
both here and in \cite{OZ2,JZ3}.
\begin{proposition} \label{Gbds}
Under assumptions (H1)-(H4) and (D1)-(\DDD),
the low-frequency Green function $G^I(x,t;y)$ of \eqref{GI} decomposes as
$G^I=E+\tilde G^I$,
\be\label{E1}
E=\bar U'(x)e(x,t;y),
\ee
where, for some $C>0$, all $t>0$,
\ba\label{GIest}
\sup_{y}\|\tilde G^I(\cdot, t,;y) \|_{L^p(x)}
&\le C (1+t)^{-\frac{1}{2}(1-\frac{1}{p})}\\
\sup_{y}\|\partial_{y}^r \tilde G^I(\cdot, t,;y) \|_{L^p(x)},
\quad
\sup_{y}\|\partial_{t}^r \tilde G^I(\cdot, t,;y) \|_{L^p(x)}
&\le C (1+t)^{-\frac{1}{2}(1-\frac{1}{p})-\frac{1}{2}}\\
\sup_{y}\| \tilde G^I(\cdot, t,;y)(0,1)^T \|_{L^p(x)}
&\le C (1+t)^{-\frac{1}{2}(1-\frac{1}{p})-\frac{1}{2}}\\
\ea
for $p\ge 2$, $1\le r\le 2$,
\ba\label{ederest}
\sup_{y}\| \partial_x^j \partial_t^l
\partial_{y}^r e(\cdot, t,;y) \|_{L^p(x)}
&\le C (1+t)^{-\frac{1}{2}(1-\frac{1}{p})- \frac{(j+l)}{2}-\frac{1}{2}}\\
\ea
for $p\ge 2$, $0\le j, l$, $j+l\le K+1$, $1\le r\le 2$,
and
\ba\label{eest}
\sup_{y}\|\tilde \partial_x^j \partial_t^l e(\cdot, t,;y) \|_{L^p(x)}
&\le C (1+t)^{-\frac{1}{2}(1-\frac{1}{p})-\frac{(j+l)}{2}}\\
\ea
for $0\le j, l$, $j+l\le K+1$,
provided that $p\ge 2$ and $j+l\ge 1$ or $p=\infty$.
Moreover, $e(x,t;y)\equiv 0$ for $t\le 1$.
\end{proposition}
\br\label{newrmk}
\textup{
The crucial new observation in the nonconservative case treated
here is \eqref{GIest}(iii), which asserts that sources entering
in the nonconservative second coordinate of the linearized equations
experience decay equivalent to that of a differentiated source
entering in the first coordinate.
This is what allows us to treat non-divergence-form source terms
arising in the second equation of the eventual perturbation equations.
}
\er
\begin{proof}
Recalling \eqref{Gxi} and Lemma \ref{blochfacts}, we have
\ba\label{Gnew}
G^I(x,t;y)&=
\Big(\frac{1}{2\pi }\Big) \int_{\R}
e^{i\xi \cdot (x-y)}
\phi(\xi)
\sum_{j=1}^{2}e^{\lambda_j(\xi)t} q_j(\xi,x)\tilde q_j(\xi, y)^*
d\xi\\
&=
\Big(\frac{1}{2\pi }\Big) \int_{\R}
e^{i\xi \cdot (x-y)}
\phi(\xi)
\sum_{j,k,l=1}^{2}e^{\lambda_j(\xi)t} \beta_{j,k}v_k(\xi,x)
\tilde \beta_{j,l}\tilde v_l(\xi, y)^*
d\xi,
\ea
the fact that $\beta_{j,1}=O(\xi^{-1})$ suggests the $k=1$ terms (corresponding to translation) dominate the low-frequency
Green kernel. With this motivation, we define
\ba\label{enew}
\tilde e(x,t;y)&=
\Big(\frac{1}{2\pi }\Big) \int_{\R}
e^{i\xi \cdot (x-y)}
\phi(\xi)
\sum_{j,l}e^{\lambda_j(\xi)t} \beta_{j,1}
\tilde \beta_{j,l}\tilde v_l(\xi, y)^* d\xi
\ea
so that
\ba\label{Gdiff}
G^I&(x,t;y) -\bar U'(x)\tilde e(x,t;y)=\\
&\Big(\frac{1}{2\pi }\Big) \int_{\R}
e^{i\xi \cdot (x-y)} \phi(\xi)
\sum_{j, k\ne 1, l} e^{\lambda_j(\xi)t}
\beta_{j,k} \tilde \beta_{j,l}v_k(\xi,x)\tilde v_l(\xi, y)^* d\xi\\
&\quad +
\Big(\frac{1}{2\pi }\Big) \int_{\R}
e^{i\xi \cdot (x-y)} \phi(\xi)
\sum_{j, l} e^{\lambda_j(\xi)t}
\beta_{j,1} \tilde \beta_{j,l}
\Big(v_1(\xi,x)-\bar U'(x)\Big)
\tilde v_l(\xi, y)^* d\xi,\\
\ea
where, by analyticity of $v_1$,
$v_1(\xi,x)-\bar U'(x)=O(|\xi|)$, and so, by Lemma \ref{blochfacts},
\be\label{crucial}
\beta_{j,1} \tilde \beta_{j,l}
\Big(v_1(\xi,x)-\bar U'(x)\Big)
\tilde v_l(\xi, y)^* =O(1)
\ee
and
\be\label{crucial2}
\beta_{j,2} \tilde \beta_{j,l}v_2(\xi,x)\tilde v_l(\xi, y)^*
=O(1)
\ee
Note further that $\tilde v_l\equiv (1,0)^T$ unless $l=1$,
in which case $\tilde \beta_{jl}=O(|\xi|)$ by Lemma \ref{blochfacts};
hence
\be\label{crucial3}
\partial_{y}\Big(\beta_{j,1} \tilde \beta_{j,l}
\Big(v_1(\xi,x)-\bar U'(x))
\tilde v_l(\xi, y)^* \Big) =O(|\xi|),
\ee
\be\label{newcrucial3}
\Big(\beta_{j,1} \tilde \beta_{j,l}
\Big(v_1(\xi,x)-\bar U'(x))
\tilde v_l(\xi, y)^* \Big)(0,1)^T =O(|\xi|),
\ee
and
\be\label{crucial4}
\partial_{y}
\Big(\beta_{j,2} \tilde \beta_{j,l}v_2(\xi,x)\tilde v_l(\xi, y)^* \Big)
=O(|\xi|)
\ee
\be\label{newcrucial4}
\Big(\beta_{j,2} \tilde \beta_{j,l}v_2(\xi,x)\tilde v_l(\xi, y)^* \Big)
(0,1)^T
=O(|\xi|)
\ee
From representation (\ref{Gdiff}), bounds \eqref{crucial}--\eqref{crucial2},
and $\Re \lambda_j(\xi)\le -\theta |\xi|^2$,
we obtain by the triangle inequality
\be
\|\tilde G^1(\cdot,t;\cdot)\|_{L^\infty(x,y)}=\|G^I-\bar U' \tilde e\|_{L^\infty(x,y)}\le C\|e^{-\theta |\xi|^2 t} \phi(\xi)\|_{L^1(\xi)}
\le C (1+t)^{-\frac{1}{2}}.
\ee
Derivative bounds follow similarly,
since $x$-derivatives falling on $v_{jk}$ are harmless, whereas,
by \eqref{crucial3}--\eqref{crucial4},
$y$- or $t$-derivatives falling on $\tilde v_{jl}$
or on $e^{i\xi\cdot(x-y)}$ bring down a factor
of $|\xi|$ improving the decay rate by factor $(1+t)^{-1/2}$.
(Note that $|\xi|$ is bounded because of the cutoff function $\phi$,
so there is no singularity at $t=0$.)
To obtain the corresponding bounds for $p=2$, we note that (\ref{Gnew})
may be viewed itself as a Bloch decomposition with respect to variable
$z:=x-y$, with $y$ appearing as a parameter.
Recalling (\ref{iso}), we may thus estimate
\ba
\sup_y &\|G^I(\cdot,t;y)-\bar U'\tilde e(\cdot, t;y)\|_{L^2(x)}
\le\\
&
C \sum_{j, k\ne 1, l}
\sup_y \|\phi(\xi) e^{\lambda_j(\xi)t}
v_k(\cdot, z_1)\tilde v_l^*(\cdot, y)
\tilde v_l(\cdot, y)^* \|_{L^2(\xi; L^2(z_1\in [0,X]))}\\
&\quad +
C\sum_{j, l} \sup_y \left\|\phi(\xi) e^{\lambda_j(\xi)t}
\Big( \frac{v_n(\cdot,x)-\bar U'(x)}{|\cdot|}\Big)
\tilde v_l(\cdot, y)^* \right\|_{L^2(\xi; L^2(z_1\in [0,X]))}\\
&\le
C \sum_{j, k\ne 1, l} \sup_y \|\phi(\xi) e^{-\theta |\xi|^2t} \|_{L^2(\xi)}
\sup_\xi\| v_k(\cdot, z_1) \|_{L^2(0,X)}
\| \tilde v_l(\cdot, y)^* \|_{L^\infty(0,X)}
\\
&\quad +
C\sum_{j, l} \sup_y \|\phi(\xi) e^{-\theta |\xi|^2t} \|_{L^2(\xi)}
\sup_\xi\left\| \Big( \frac{v_n(\xi,x)-\bar U'(x)}{|\xi|}\Big) \right\|_{L^2(0,X)}
\|\tilde v_l(\cdot, y)^* \|_{L^\infty(0,X)} \\
&\le
C (1+t)^{-\frac{1}{4}},
\ea
where we have used in a crucial way the boundedness of $\tilde v_l$
in $L^\infty$,\footnote{This is clear for $\xi=0$, since $v_j$
are linear combinations of genuine and generalized eigenfunctions,
which are solutions of the homogeneous or inhomogeneous eigenvalue ODE.
More generally, note that the resolvent of $L_\xi-\gamma$
gains one derivative, hence the total eigenprojection, as a contour
integral of the resolvent, does too- now, use the one-dimensional
Sobolev inequality for periodic boundary conditions
to bound the $L^\infty$ difference from the
mean by the (bounded) $H^1$ norm, then bound the mean by the $L^1$ norm,
which is controlled by the $L^2$ norm.}
and also the boundedness of
$$
\Big( \frac{v_n(\xi,x)-\bar U'(x)}{\xi}\Big)
\sim
\partial_{\xi}v_n(r)
$$
in $L^2$, where $r\in (0,\xi)$.
Derivative bounds follow similarly as above, noting that
$y$- or $t$-derivatives bring down a factor $\xi$, while
$x$-derivatives are harmless, to obtain an additional factor
of $(1+t)^{-1/2}$ decay.
Finally, bounds for $2\le p\le \infty$ follow by $L^p$-interpolation.
Now, defining
\be\label{edef}
e(x,t;y):= \chi(t)\tilde e(x,t;y),
\ee
where $\tilde e$ is defined in \eqref{enew} and $\chi$ is a smooth cutoff function
such that $\chi(t)\equiv 1$ for $t\ge 2$ and $\chi(t)\equiv 0$ for $t\le 1$,
and setting $\tilde G:=G-\bar U'(x)e(x,t;y)$,
we readily obtain the estimates \eqref{GIest} by combining
the above estimates on $G^I-\bar U \tilde e$
with bound \eqref{SIIest} on $G^{II}$.
Finally, recalling, by Lemma \ref{blochfacts}, that $\tilde v_l\equiv \const$
for $l\ne 1$ while $\tilde \beta_{j,1}=O(|\xi|)$, we have
$$
\partial_{y} \Big( \beta_{j,1} \tilde \beta_{j,l}\tilde v_l(\xi, y)^*\Big)
=o(|\xi|).
$$
Bounds \eqref{ederest} thus
follow from \eqref{enew} by the argument
used to prove \eqref{GIest}, together with the observation that
$x$- or $t$-derivatives bring down factors of $\xi$.
Bounds \eqref{eest} follow similarly for $j+l\ge 1$,
in which case the integrand on the righthand side of \eqref{enew}
(now differentiated in $x$ and or $t$) is Lebesgue integrable.
In the critical case $j=l=0$,
taking $t$ without loss of generality $\ge 1$,
expanding
$$
\lambda_j(\xi)=-i\xi a_j - b_j \xi^2 +O(\xi^3),
$$
and setting $\check \lambda(\xi):= -i\xi a_j -b_j\xi^2$,
we may write $\tilde e(x,t;y)$ in \eqref{enew} as
\be\label{princterm}
\begin{aligned}
\Big(\frac{1}{2\pi }\Big)&\int_{\R}\sum_{j}
\check \beta_{j,1}(0) \tilde \beta_{j,2}(0)\tilde v_2(0, y)^*e^{i\xi \cdot (x-y)}
\xi^{-1}e^{\check\lambda_j(\xi)t}
d\xi\\
&=\Big(\frac{1}{2\pi }\Big)\pv\int_{\R}\sum_{j}
\check \beta_{j,1}(0) \tilde \beta_{j,2}(0)\tilde v_2(0, y)^*e^{i\xi \cdot (x-y)}
\xi^{-1}e^{\check\lambda_j(\xi)t}
d\xi\\
&=\sum_{j}
\check \beta_{j,1}(0) \tilde \beta_{j,2}(0)\tilde v_2(0, y)^*
\Big(\frac{1}{2\pi }\Big) \pv\int_{\R}
e^{i\xi \cdot (x-y)}
\xi^{-1}e^{\check\lambda_j(\xi)t}
d\xi,
\end{aligned}
\ee
where $\check \beta_{j,1}(0):=\lim_{\xi\to 0}(\xi \beta_{j,1}(\xi))$,
and the above series is convergent by the alternating series test,
plus a negligible error term
$$
\Big(\frac{1}{2\pi }\Big) \pv\int_{\R}
e^{i\xi \cdot (x-y)}
\phi(\xi)
O(e^{-\theta|\xi|^2t}) d\xi
$$
for which the integrand is Lebesgue integrable,
hence, by the previous argument, obeys the bounds for $j+l=1$.
(Note
that the integral on the lefthand side of \eqref{princterm}
is absolutely convergent by
$
\xi^{-1}(e^{-ia_1\xi t}-e^{-ia_2\xi t})\sim |a_1-a_2|t,
$
becoming conditionally convergent only when the integrand is split into
different eigenmodes.)
By (D2), we have $a_j$ real and $\Re b_j>0$.
Moreover, the operator
$L$, since real-valued, has spectrum with complex conjugate symmetry,
hence $b_j$ is real as well.
Observing that
$\Big(\frac{1}{2\pi }\Big) \pv\int_{\R}
e^{i\xi \cdot (x-y)}
\xi^{-1}e^{\check\lambda_j(\xi)t} d\xi$
is an antiderivative in $x$ of the inverse Fourier transform
$\Big(\frac{1}{2\pi }\Big)\int_{\R}
e^{i\xi \cdot (x-y)}
e^{\check\lambda_j(\xi)t} d\xi
=\frac{e^{-(x-y-a_jt)^2/4b_jt}}{\sqrt{4\pi b_jt}}$,
a Gaussian, we find that the principal part \eqref{princterm}
is a sum of errorfunctions
\be\label{eq:errfns}
\sum_{j=1}^2 c_j \, \errfn \Big( \frac{x-y-a_jt}{\sqrt{4b_jt}}, t \Big)
{\tilde v}_2(0,y),
\ee
hence bounded in $L^\infty$ as claimed, where $a_j$ denote the
characteristic speeds of the Whitham averaged system and
(on further inspection) $\sum_j c_j=0$.
This verifies bound \eqref{eest} in the final
case $j=l=0$, completing the proof.
\end{proof}
\br\label{ozrmk}
\textup{
See the proof of Proposition 1.5, \cite{OZ2}, for an essentially
equivalent estimate from the inverse Laplace transform
point of view
of the critical $\xi^{-1}$ contribution \eqref{princterm}.
}
\er
\subsection{Final linearized bounds}\label{s:finallin}
\begin{cor}\label{greenbds}
Under assumptions (H1)--(H4), (D1)--(\DDD),
the Green function $G(x,t;y)$ of \eqref{e:lin} decomposes as
$G=E+\tilde G$,
\be\label{E}
E=\bar U'(x)e(x,t;y),
\ee
where, for some $C>0$, all $t>0$, $1\le q\le 2\le p\le \infty$, $0\le j,k, l$,
$j+l\le K+1$, $1\le r\le 2$,
\ba\label{sheatbds}
\left\|\int_{-\infty}^{+\infty} \tilde G(x,t;y)f(y)dy\right\|_{L^p(x)}&\le
C (1+t)^{-\frac{1}{2}(1/q-1/p)}
\|f\|_{L^q\cap H^1},\\
\left\|\int_{-\infty}^{+\infty} \partial_y^r \tilde G(x,t;y)f(y)dy\right\|_{L^p(x)}&\le
C (1+t)^{-\frac{1}{2}(1/q-1/p)-\frac{1}{2}}
\|f\|_{L^q\cap H^{r+1}},\\
\left\|\int_{-\infty}^{+\infty} \partial_t^r \tilde G(x,t;y)f(y)dy\right\|_{L^p(x)}&\le
C (1+t)^{-\frac{1}{2}(1/q-1/p)-\frac{1}{2}} \|f\|_{L^q\cap H^{2r+1}},\\
\left\|\int_{-\infty}^{+\infty} \tilde G(x,t;y)(0,1)^T f(y)dy\right\|_{L^p(x)}&\le
C (1+t)^{-\frac{1}{2}(1/q-1/p)-\frac{1}{2}} \|f\|_{L^q\cap H^1}.\\
\ea
\ba\label{etbds}
\left\|\int_{-\infty}^{+\infty} \partial_x^j\partial_t^k e(x,t;y)f(y)dy\right\|_{L^p}
&\le
(1+t)^{-\frac{1}{2}(1/q-1/p) -\frac{(j+k)}{2} +\frac{1}{2} }\|f\|_{L^q},\\
\left\|\int_{-\infty}^{+\infty} \partial_x^j\partial_t^k\partial_y^r e(x,t;y)f(y)dy\right\|_{L^p}
&\le
(1+t)^{-\frac{1}{2}(1/q-1/p) -\frac{(j+k)}{2} }\|f\|_{L^q},\\
\left\|\int_{-\infty}^{+\infty} \partial_x^j\partial_t^k e(x,t;y)
(0,1)^Tf(y)dy\right\|_{L^p}
&\le
(1+t)^{-\frac{1}{2}(1/q-1/p) -\frac{(j+k)}{2} }\|f\|_{L^q}.\\
\ea
Moreover, $e(x,t;y)\equiv 0$ for $t\le 1$.
\end{cor}
\begin{proof}
({\it Case $q=1$}).
From (\ref{GIest}) and the triangle inequality we obtain
$$
\Big\|\int_{\R}\tilde G^I(x,t;y)f(y)dy\Big\|_{L^p(x)}
\le
\int_{\R}\sup_y \|\tilde G^I(\cdot ,t;y)\|_{L^p}|f(y)|dy
\le C(1+t)^{-\frac{1}{2}(1-1/p)}\|f\|_{L^1}
$$
and similarly for $y$- and $t$-derivative estimates, and products
with $(0,1)^T$, which,
together with \eqref{SIIest}, yield \eqref{sheatbds}.
Bounds \eqref{etbds} follow similarly by the triangle inequality
and \eqref{ederest}--\eqref{eest}.
({\it Case $q=2$}).
From \eqref{crucial}--\eqref{crucial2}, and analyticity of
$v_j$, $\tilde v_j$, we have boundedness
from $L^2([0,X])\to L^2([0,X])$
of the projection-type operators
\be\label{crucialx}
f\to
\beta_{j,n} \tilde \beta_{j,l}
\Big(v_n(\xi,x)-\bar U'(x)\Big)
\langle \tilde v_l, f\rangle
\ee
and
\be\label{crucial2x}
f\to
\beta_{j,k} \tilde \beta_{j,l}v_k(\xi,x)
\langle \tilde v_l,f\rangle
\; \hbox{\rm for }\; k\ne 1,
\ee
uniformly with respect to $\xi$,
from which we obtain by \eqref{Gdiff}, \eqref{edef}, and (\ref{iso})
the bound
\be\label{triv}
\left\|\int_{-\infty}^{+\infty} \tilde G^I(x,t;y)f(y)dy\right\|_{L^2(x)}\le
C\|f\|_{L^2(x)},
\ee
for all $t\ge 0$, yielding together with \eqref{SIIest}
the result \eqref{sheatbds} for $p=2$, $r=1$.
Similarly, by boundedness of $\tilde v_j$, $v_j$, $\bar U'$
in all $L^p[0,X]$, we have
$$
\begin{aligned}
\left\|e^{\lambda_j(\xi)t}
\beta_{j,n} \tilde \beta_{j,l}
\Big(v_n(\xi,x)-\bar U'(x)\Big)
\langle \tilde v_l, \hat f\rangle\right\|_{L^\infty(x)}
& \le Ce^{-\theta |\xi|^2t} \|\hat f(\xi,\cdot)\|_{L^2(x)},\\
\left\|e^{\lambda_j(\xi)t}
\beta_{j,k} \tilde \beta_{j,l}v_k(\xi,x)
\langle \tilde v_l, \hat f\rangle \right\|_{L^\infty(x)}
& \le Ce^{-\theta |\xi|^2t} \|\hat f(\xi,\cdot)\|_{L^2(x)},
\; \hbox{\rm for }\; k\ne 1,
\end{aligned}
$$
$C,\, \theta>0$, yielding by definitions \eqref{Gdiff}, \eqref{edef} the bound
\ba\label{last}
\left\|\int_{-\infty}^{+\infty} \tilde G^I(x,t;y)f(y)dy\right\|_{L^\infty(x)}
&\le
\Big(\frac{1}{2\pi }\Big) \int_{-\pi}^{\pi}
C\phi(\xi) e^{-\theta |\xi|^2t}\|\hat f(\xi,\cdot)\|_{L^2(x)}
d\xi\\
&\le C\left\|\phi(\xi) e^{-\theta |\xi|^2t}\right\|_{L^2(\xi)} \|\hat f\|_{L^2(\xi, x)}\\
&\le C(1+t)^{-\frac{d}{4} } \|f\|_{L^2([0,X])},
\ea
hence giving the result for $p=\infty$, $r=0$.
The result for $r=0$ and general $2\le p\le \infty$ then follows by
$L^p$ interpolation between $p=2$ and $p=\infty$.
Derivative bounds $1\le r\le 2$ follow by
similar arguments, using \eqref{crucial3}--\eqref{crucial4},
as do bounds for products with $(0,1)^T$.
Bounds \eqref{etbds} follow similarly.
({\it Case $1\le q \le 2$}).
By Riesz--Thorin interpolation between the cases $q=1$ and $q=2$,
we obtain the bounds asserted in the general case $1\le q\le 2$,
$2\le p\le \infty$.
\end{proof}
Note the close analogy between the bounds of Corollary \ref{greenbds}
and those obtained in \cite{MaZ3,MaZ1} for the viscous or relaxation
shock wave case.
\section{Nonlinear stability}\label{s:nonlin}
With the bounds of Corollary \eqref{greenbds}, nonlinear
stability follows by a combination of the argument of \cite{JZ3,JZ4}
and modifications introduced in the shock wave case
to treat partial parabolicity and potential loss of derivatives
in the nonlinear iteration scheme \cite{Z1,Z3}.
\subsection{Nonlinear perturbation equations}\label{s:pert}
Given a solution $\tilde U(x,t)$ of \eqref{eqn:1conslaw},
define the nonlinear perturbation variable
\be\label{pertvar}
v=U-\bar U=
\tilde {U}(x+\psi(x,t),t)-\bar U(x),
\ee
where
\be\label{uvar}
U(x,t):=\tilde {U}(x+\psi(x,t),t)
\ee
and $\psi:\RM\times\RM\to\RM$ is to be chosen later.
\begin{lem}\label{4.1}
For $v$, $U$ as in \eqref{pertvar}, \eqref{uvar}, and $|\tilde U|$ bounded,
\begin{equation}\label{eqn:1nlper}
U_t+f(U)_{x}-(B(U)U_{x})_x-g(U)=\left(\partial_t-L\right)\bar{U}'(x)\psi(x,t)
+P
+\partial_x R +
\partial_t S ,
\ee
where
\begin{equation}\label{eqn:P}
P=\left(g(\tilde U)-g(\bar{U})\right)\psi_x=(0,1)^T\mathcal{O}(|v||\psi_x|),
\end{equation}
\begin{equation}\label{eqn:R}
R:= v\psi_t + B(\tilde{U})(\bar U_x +v_x)\frac{\psi_x^2}{1+\psi_x}-\left(B(\tilde U)-B(\bar U)\right)\bar U_x\psi_x-B(\tilde{U})v_x\psi_x,
\end{equation}
and
\begin{equation}\label{eqn:S}
S
:=- v\psi_x.
\end{equation}
\end{lem}
\begin{proof}
By the definition of $U$ in \eqref{uvar} we have by a
straightforward computation
\begin{align*}
U_t(x,t)&=\tilde{U}_x(x+\psi(x,t),t)\psi_t(x,t)+\tilde{U}_t(x+\psi,t)\\
f(U(x,t))_x&=df(\tilde{U}(x+\psi(x,t),t))\tilde{U}_x(x+\psi,t)\cdot(1+\psi_x(x,t))\\
U_{x}(x,t)&=\tilde{U}_x(x+\psi(x,t),t)\cdot(1+\psi_x(x,t)).
\end{align*}
By $\tilde U_t + df(\tilde{U})\tilde{U}_x-
(B(\tilde U) \tilde{U}_{x})_x-g(\tilde U)=0$,
it follows that
\ba\label{altform}
U_t+f(U)_{x}-(B(U)U_{x})_x-g(U)
&=\tilde{U}_x\psi_t+df(\tilde{U})\tilde{U}_x\psi_x
-(B(\tilde U)\tilde U_x)_x \psi_x
-(B(\tilde U)\tilde{U}_x\psi_x)_x\\
&= \tilde U_x \psi_t -\tilde U_{t} \psi_x
+g(\tilde U)\psi_x
-(B(\tilde U)\tilde{U}_x\psi_x)_x,
\ea
where it is understood that derivatives of $\tilde U$ appearing
on the righthand side
are evaluated at $(x+\psi(x,t),t)$.
Moreover, by another direct calculation,
using $L(\bar{U}'(x))=0$, we have
$$
\begin{aligned}
\left(\partial_t-L\right)\bar{U}'(x)\psi&=
\bar{U}_x\psi_t -\bar{U}_t\psi_{x}
+df(\bar {U})\bar {U}_x\psi_x -(B(\bar U)\bar U_x)_x \psi_x
-(B(\bar U)\bar{U}_x\psi_x)_x\\
&= \bar U_x \psi_t -\bar U_{t} \psi_x
+g(\bar U)\psi_x
-(B(\bar U)\bar{U}_x\psi_x)_x.
\end{aligned}
$$
Subtracting, and using the facts that,
by differentiation of $(\bar U+ v)(x,t)= \tilde U(x+\psi(x,t),t)$,
\ba\label{keyderivs}
\bar U_x + v_x&= \tilde U_x(1+\psi_x),
\qquad
\bar U_t + v_t= \tilde U_t + \tilde U_x\psi_t,\\
\ea
so that
\ba\label{solvedderivs}
\tilde U_x-\bar U_x -v_x&=
-(\bar U_x+v_x) \frac{\psi_x}{1+\psi_x},
\qquad
\tilde U_t-\bar U_t -v_t=
-(\bar U_x+v_x) \frac{\psi_t}{1+\psi_x},\\
\ea
we obtain
\ba
U_t+ f(U)_{x} - (B(U)U_{x})_x-g(U)&=
(\partial_t-L)\bar{U}'(x)\psi
+v_x\psi_t - v_t \psi_x
\\ &\quad +
\big(g(\tilde U)-g(\bar U)\big) \psi_x-\left(B(\tilde{U})v_x\psi_x\right)_x
\\ &\quad +
\left(B(\tilde U)(\bar U_x +v_x)\frac{\psi_x^2}{1+\psi_x} \right)_x
\\ &\quad
-\left(\left(B(\tilde U)-B(\bar U)\right)\bar U_x\psi_x\right)_x,
\ea
yielding \eqref{eqn:1nlper} by
$v_x\psi_t - v_t \psi_x = (v\psi_t)_x-(v\psi_x)_t$.
\end{proof}
\begin{cor}\label{4.2}
The nonlinear residual $v$ defined in \eqref{pertvar} satisfies
\be\label{veq}
v_t-Lv=\left(\partial_t-L\right)\bar{U}'(x)\psi
-Q_{x}+T+ P+ R_x +
\partial_t S,
\ee
where $P$, $R$, and $S$ are as in Lemma \ref{4.1} and $Q$ and $T$ are defined by
\ba\label{eqn:Q}
Q:&=f(\tilde{U}(x+\psi(x,t),t))-f(\bar{U}(x))-df(\bar{U}(x))v\\
&\quad\quad -\left(B(\tilde{U}(x+\psi(x,t),t))\tilde{U}_x(x+\psi(x,t),t)-B(\bar{U}(x))\bar{U}_x(x)\right)\\
&\quad\quad\quad\quad-\left(B(\bar{U})v_x+\left(dB(\bar{U})\bar{U}_x\right)v\right)\\
\ea
and
\be\label{eqn:T}
T:=g(\tilde{U}(x+\psi(x,t),t))-g(\bar{U}(x))-dg(\bar{U}(x))v=
(0,1)^T\mathcal{O}(|v|^2),
\ee
\end{cor}
\begin{proof}
Taylor expansion comparing \eqref{eqn:1nlper} and
$\bar U_t + f(\bar U)_x-(B(\bar U)\bar U_{x})_x-g(\bU)=0$.
\end{proof}
\subsection{Cancellation estimate}\label{s:cancellation}
Our strategy in writing \eqref{veq} is motivated by the following
basic cancellation principle.
\begin{prop}[\cite{HoZ}]\label{p:cancellation}
For any $f(y,s)\in L^p \cap C^2$ with $f(y,0)\equiv 0$, there holds
\be\label{e:cancel}
\int^t_0 \int G(x,t-s;y) (\partial_s - L_y)f(y,s) dy\,ds
= f(x,t).
\ee
\end{prop}
\begin{proof} Integrating the left hand side by parts, we obtain
\be
\int G(x,0;y)f(y,t)dy - \int G(x,t;y)f(y,0)dy
+ \int^t_0 \int
(\partial_t - L_y)^*G(x,t-s;y) f(y,s)dy\, ds.
\label{5.53.2}
\ee
Noting that, by duality,
$$
(\partial_t - L_y)^* G(x,t-s;y) = \delta(x-y) \delta(t-s),
$$
$\delta(\cdot)$ here denoting the Dirac delta-distribution,
we find that the third term on the righthand side
vanishes in \eqref{5.53.2}, while,
because $G(x,0;y) = \delta(x-y)$, the first term is simply $f(x,t)$.
The second term vanishes by $f(y,0)\equiv 0$.
\end{proof}
\subsection{Nonlinear damping estimate}
The following technical result is a key ingredient in
the nonlinear stability analysis that follows.
Applying Duhamel's principle to \eqref{veq} and using Proposition \ref{p:cancellation}
yield
\ba\label{prelim}
v(x,t)&=\int^\infty_{-\infty}G(x,t;y)v_0(y)\,dy \\
&\quad
+ \int^t_0 \int^\infty_{-\infty} G(x,t-s;y)
(-Q_y+T+ R_y + S_s ) (y,s)\,dy\,ds
+ \psi (t) \bar U'(x).
\ea
Note that terms $Q_y$ and $S_s$ involve derivatives of $v$ (respectively
second derivative in space and first derivative in time) of maximal order,
hence to close a nonlinear iteration scheme based on \eqref{prelim}
would appear to require delicate maximal regularity estimates
rather than the straightforward ones that we have obtained.
Indeed, estimated using the linearized bounds of Corollary \ref{greenbds},
the righthand side
appears to lose several degrees of regularity as a function
from $H^K\to L^2$ of $v$.
However, the next proposition,
adapted from the methods
of \cite{MaZ4,Z3},
shows that higher-order derivatives are slaved to lower-order ones,
hence derivatives ``lost'' at the linearized level may be
``regained'' at the nonlinear level.
This effectively separates the issues of decay and regularity,
allowing us to close a nonlinear iteration without the use of
maximal regularity estimates or a more complicated quasilinear iteration scheme.
\begin{proposition}\label{damping}
Let $v_0\in H^K$ ($K$ as in (H1)), and suppose that
for $0\le t\le T$, the $H^K$ norm of $v$ and the $H^{K+1}$ norms of
$\psi_t(\cdot,t)$ and $\psi_x(\cdot,t)$
remain bounded by a sufficiently small constant.
Moreover, suppose that the Froude number $F$, viscosity $\nu$,
and velocity derivative $\bar u_x$ satisfy the amplitude condition
$\nu \bar u_x < F^{-1}$.
Then, there are constants
$C, \theta_{1}>0$ such that, for all $0\leq t\leq T$,
\begin{equation}\label{Ebounds}
\|v(t)\|_{H^K}^2 \leq C e^{-\theta_1 t} \|v(0)\|^2_{H^K} +
C \int_0^t e^{-\theta_1(t-s)} \left(\|v\|_{L^2}^2 +
\|(\psi_t, \psi_x)\|_{H^K}^2 \right) (s)\,ds.
\end{equation}
\end{proposition}
The proof of this result will be given in Appendix \ref{s:energy}.
Here,
we briefly outline the main ideas. First, notice that by
subtracting from the equation \eqref{altform} for $U$
the equation for $\bar U$, we may write the
nonlinear perturbation equation as
\ba\label{vperturteq}
v_t + (Av)_x-(Bv_{x})_x-Cv&= P- Q(v)_x +T(v)
+ \tilde U_x \psi_t -\tilde U_{t} \psi_x \\
&\quad +g(\bar{U})\psi_x-\left(B(\tilde{U})\tilde{U}_x\psi_x\right)_x
\ea
where $A$, $B$, $C$ are as in \eqref{coeffs},
$P$, $Q$, and $T$ are as in
Corollary \ref{4.2},
$g$ and $B$ are as in \eqref{ab},
and it is understood that derivatives of $\tilde U$ appearing
on the righthand side
are evaluated at $(x+\psi(x,t),t)$.
Using \eqref{solvedderivs} to replace $\tilde U_x$ and
$\tilde U_t$ respectively by
$\bar U_x + v_x -(\bar U_x+v_x) \frac{\psi_x}{1+\psi_x}$
and
$\bar U_t + v_t -(\bar U_x+v_x) \frac{\psi_t}{1+\psi_x}$,
and moving the resulting $v_t\psi_x$ term to the lefthand side
of \eqref{vperturteq}, we obtain
\ba\label{vperturteq2}
(1+\psi_x) v_t&=(Bv_{x})_x-(A v)_x+ Cv+ P- Q(v)_x +T(v)\\
&\quad\quad+ \left(\bar U_x+v_x\right)\psi_t+g(\bar{U})\psi_x
-\left(B(\tilde{U})\left(\bar{U}_x+v_x\right)\frac{\psi_x}{1+\psi_x}\right)_x
\ea
Define now the Friedrichs symmetrizer
\be\label{Sigma}
\Sigma:=\bp 1 & 0\\ 0& \delta^{-2}\ep,
\ee
where $\delta^2:= -A_{12}= \bar \tau^{-3}(F^{-1}-\nu \bar u_x)$. By
\eqref{froudebd}, $\Sigma$ is a symmetric positive definite symmetrizer for the hyperbolic
part of \eqref{vperturteq2} in the sense that
$
\Sigma A=\bp -c & -1 \\-1 &-c\delta^{-2}\ep
$
is a symmetric matrix,
where $A$ is as in \eqref{coeffs}.
Furthermore, to compensate for the lack of total parabolicity
of the governing equation, here indicated by the presence of a neutral eigenspace of
the matrix $\Sigma B$, we
introduce
the skew-symmetric Kawashima compensator
\be\label{K}
K:=\eta\bp 0 & -1\\1 & 0\ep, \quad 0<\eta\ll 1
\ee
and note that, in particular, for $\eta>0$ sufficiently small there exists a constant $\theta>0$
such that
$\Re (KA+\Sigma B)\geq\theta$.
Now defining the functional
\[
\mathcal{E}[v]:=\left<v,\Sigma v\right>+\sum_{j=1}^K\left(\left<\partial_x^jv,K\partial_x^{j-1}v\right>
+\left<\partial_x^jv,\Sigma\partial_x^jv\right>\right)
\]
where here $\left<\cdot,\cdot\right>$ denotes the standard $L^2(\RM^n)$ inner product,
we find by a direct
but lengthy calculation using Sobolev embedding and interpolation
to absorb nonlinear and intermediate-derivative terms that
\be\label{Eineq}
\partial_t \calE(v) \leq -{\theta_1} \|v\|_{H^K}^2 +
C\left( \|v\|_{L^2}^2+ \|(\psi_t, \psi_x)\|_{H^K(x,t)}^2 \right)
\ee
for some positive constants $C, \theta_1 >0$, so long as $\|\tilde{U}\|_{H^K}$ remains bounded
and the quantities $\eta>0$, $\|v\|_{H^K}$ and $\|(\psi_t,\psi_x)\|_{H^K(x)}$ remain sufficiently small.
By Cauchy--Schwarz and the fact that $\Sigma$ is positive definite by \eqref{froudebd}, we have $\mathcal{E}(v)\sim\|v\|_{H^K}^2$ for $\eta>0$
sufficiently small and hence \eqref{Eineq} implies
\[
\partial_t \calE(v) \leq -{\theta_1}\calE(v)+
C\left( \|v\|_{L^2}^2+ \|(\psi_t, \psi_x)\|_{H^K(x,t)}^2 \right)
\]
from which (\ref{Ebounds}) follows by Gronwall's inequality and, again, the equivalence of $\CalE(v)$
and $\|v\|_{H^K}^2$.
For more details and a complete proof of the key inequality \eqref{Eineq}, see Appendix \ref{s:energy}.
\br\label{froudermk}
\textup{
The condition \eqref{froudebd} gives effectively an
upper bound on the allowable amplitude of the wave,
for fixed Froude number and viscosity.
It is not clear that this has any connection with behavior.
Certainly it is needed for our argument structure, and
perhaps even for the validity of \eqref{Ebounds}, which is
itself convenient but clearly not necessary for stability.
The resolution of this issue would be very interesting from the
standpoint of applications, both in this and related contexts.
}
\er
\br\label{damprmk}
\textup{
The Lagrangian formulation appears essential here
in order to obtain quantitative bounds like \eqref{froudebd}
on the amplitude of the wave.
One can carry out damping estimates
for
sufficiently small-amplitude waves in Eulerian coordinates by the
argument of \cite{MaZ2} in the shock wave case; however,
the large-amplitude argument of \cite{MaZ4}, depending on
global noncharacteristicity of the wave-- corresponding here
to nonvanishing of $u-s$, where $s$ is wave speed
in Eulerian coordinates-- together with bounded variation
of $\bar U_x$,
appears to fail irreparably
in the periodic case.
As we have shown here, the same argument succeeds in Lagrangian
coordinates provided that the linearized convection matrix $A$
is symmetrizable (the meaning of bound \eqref{froudebd}).
For similar observations regarding the advantages for energy
estimates of the special structure in Lagrangian coordinates, see \cite{TZ1}.
}
\er
\subsection{Integral representation/$\psi$-evolution scheme}
Recalling the Duhamel representation \eqref{prelim} of the perturbation $v$
along with the decomposition $G=\bar U'(x)e+\tilde{G}$ of Corollary \ref{greenbds},
we find that defining $\psi$ implicitly as
\ba
\psi (x,t)& =-\int^\infty_{-\infty}e(x,t;y) U_0(y)\,dy \\
&\quad
-\int^t_0\int^{+\infty}_{-\infty} e(x,t-s;y)
(P-Q_y+ T+R_y + S_s ) (y,s)\,dy\,ds ,
\label{psi}
\ea
where $e$ is defined as in \eqref{edef},
results in the {\it integral representation}
\ba \label{u}
v(x,t)&=\int^\infty_{-\infty} \tilde G(x,t;y)v_0(y)\,dy \\
&\quad
+\int^t_0\int^\infty_{-\infty}\tilde G(x,t-s;y)
(P-Q_y+ T+R_y + S_s ) (y,s)\,dy\,ds ,
\ea
for the nonlinear perturbation $v$; see \cite{Z1,MaZ2} for further details. Furthermore, differentiating (\ref{psi}) with respect to $t$,
and recalling that $e(x,s;y)\equiv 0$ for $s \le 1 $,
\ba \label{psidot}
\partial_t^j\partial_x^k \psi (x,t)&=-\int^\infty_{-\infty}\partial_t^j\partial_x^k
e(x,t;y) U_0(y)\,dy \\
&\quad
-\int^t_0\int^{+\infty}_{-\infty} \partial_t^j\partial_x^k
e(x,t-s;y)
(P-Q_y+ T+R_y + S_s) (y,s)\,dy\,ds .
\ea
Equations \eqref{u}, \eqref{psidot}
together form a complete system in the variables $(v,\partial_t^j \psi,
\partial_x^k\psi)$,
$0\le j, k\le K+1$,
from the solution of which we may afterward recover the
shift $\psi$ via \eqref{psi}.
From the original differential equation \eqref{veq}
together with \eqref{psidot},
we readily obtain short-time existence and continuity with
respect to $t$ of solutions
$(v,\psi_t, \psi_x)\in H^K$
by a standard contraction-mapping argument based on \eqref{Ebounds},
\eqref{psi}, and \eqref{etbds}.
\subsection{Nonlinear iteration}
Associated with the solution $(U, \psi_t, \psi_x)$ of integral system
\eqref{u}--\eqref{psidot}, define
\ba\label{szeta}
\zeta(t)&:=\sup_{0\le s\le t}
\|(v, \psi_t,\psi_x)\|_{H^K}(s)(1+s)^{1/4} .
\ea
\bl\label{sclaim}
For all $t\ge 0$ for which $\zeta(t)$ is
finite and sufficiently small,
some $C>0$,
and $E_0:=\|U_0\|_{L^1\cap H^K}$ sufficiently small,
\be\label{eq:sclaim}
\zeta(t)\le C(E_0+\zeta(t)^2).
\ee
\el
\begin{proof}
By \eqref{eqn:P}--\eqref{eqn:S} and \eqref{eqn:Q}--\eqref{eqn:T}
and corresponding bounds on the derivatives together
with definition \eqref{szeta},
\ba\label{sNbds}
\|(P,Q,R,S,T)\|_{L^1\cap H^2}
&\le \|(v,v_x,\psi_t,\psi_x)\|_{L^2}^2+
\|(v,v_x,\psi_t,\psi_x)\|_{H^2}^2
\le C\zeta(t)^2 (1+t)^{-\frac{1}{2}},\\
\ea
so long as $|\psi_x|\le |\psi_x|_{H^K}\le \zeta(t)$ remains small.
Applying Corollary \ref{greenbds} with $q=1$
to representations
\eqref{u}--\eqref{psidot}, we obtain for any $2\le p<\infty$
\ba\label{sest}
\|v(\cdot,t)\|_{L^p(x)}& \le
C(1+t)^{-\frac{1}{2}(1-1/p)}E_0 \\
&\quad +
C\zeta(t)^2\int_0^{t} (1+t-s)^{-\frac{1}{2}(1/2-1/p)}(t-s)^{-\frac{3}{4}}
(1+s)^{-\frac{1}{2}}ds\\
&
\le
C(E_0+\zeta(t)^2) (1+t)^{-\frac{1}{2}(1-1/p)}
\ea
and
\ba\label{sestad}
\|(\psi_t,\psi_x)(\cdot, t)\|_{W^{K+1,p}}& \le
C(1+t)^{-\frac{1}{2}}E_0 +
C\zeta(t)^2\int_0^{t} (1+t-s)^{-\frac{1}{2}(1-1/p)-1/2}
(1+s)^{-\frac{1}{2}}ds \\
&\le
C(E_0+\zeta(t)^2) (1+t)^{-\frac{1}{2}(1-1/p)},
\ea
yielding in particular that $\|(\psi_t,\psi_x)\|_{H^{K+1}}$
be arbitrarily small, verifying the hypotheses of Proposition
\ref{damping}.\footnote{Note that we have gained a necessary one degree
of regularity in $\psi$, the regularity of $\psi$ being limited
only by the regularity of the coefficients of the underlying
PDE \eqref{eqn:1conslaw}.}
Using \eqref{Ebounds} and \eqref{sest}--\eqref{sestad},
we thus obtain
$\|v(\cdot,t)\|_{H^K(x)} \le
C(E_0+\zeta(t)^2) (1+t)^{-\frac{1}{4}}$.
Combining this with \eqref{sestad}, $p=2$, rearranging, and recalling
definition \eqref{szeta}, we obtain \eqref{sclaim}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main}]
By short-time $H^K$ existence theory,
$\|(v,\psi_t,\psi_x)\|_{H^{K}}$ is continuous so long as it
remains small, hence $\zeta$ remains
continuous so long as it remains small.
By \eqref{sclaim}, therefore,
it follows by continuous induction that
$\zeta(t) \le 2C E_0$ for $t \ge0$, if $E_0 < 1/ 4C$,
yielding by (\ref{szeta}) the result (\ref{eq:smallsest}) for $p=2$.
Applying \eqref{sest}--\eqref{sestad}, we obtain
(\ref{eq:smallsest}) for $2\le p\le p_*$ for any $p_*<\infty$,
with uniform constant $C$.
Taking $p_*>4$ and estimating
\[
\|P\|_{L^2}, \, \|Q\|_{L^2}, \, \|R\|_{L^2}, \, \|S\|_{L^2},\, \|T\|_{L^2}(t)
\le \|(v,\psi_t,\psi_x)\|_{L^4}^2\le CE_0(1+t)^{-\frac{3}{4}}
\]
in place of the weaker \eqref{sNbds},
then applying Corollary \ref{greenbds} with $q=2$
we obtain finally \eqref{eq:smallsest} for $2\le p\le \infty$,
by a computation similar \eqref{sest}--\eqref{sestad};
we omit the details of this final bootstrap argument.
Estimate \eqref{eq:stab} then follows using \eqref{etbds} with
$q=1$, by
\ba\label{sesta}
\|\psi(t)\|_{L^p}& \le
C E_0 (1+t)^{\frac{1}{2p}}
+
C\zeta(t)^2\int_0^{t} (1+t-s)^{-\frac{1}{2}(1-1/p)}
(1+s)^{-\frac{1}{2}}ds\\
& \le
C(1+t)^{\frac{1}{2p}}(E_0+\zeta(t)^2),
\ea
together with the fact that
$ \tilde U(x,t)-\bar U(x)= v(x-\psi,t)+ \bar U(x)-\bar U(x-\psi), $
so that $|\tilde U(\cdot, t)-\bar U|$ is controlled
by the sum of $|v|$ and
$|\bar U(x)-\bar U(x-\psi)|\sim |\psi|$.
This yields stability for $|U-\bar U|_{L^1\cap H^K}|_{t=0}$
sufficiently small, as described in the final line of the theorem.
\end{proof}
|
\section{Introduction}
Compressed sensing is concerned with when and how sparse signals
can be recovered exactly or approximately from few linear measurements
\cite{Donoho2006,CRT2006,CandesICM}.
Let $\Phi$ be an $m \times N$ matrix providing the measurements where
$m \ll N$, and $\Sigma_k^N$ denote the space
of $k$-sparse signals in $\mathbb{R}^N$, $k<m$. A standard objective, after a
suitable change of basis, is that the mapping $x \mapsto y=\Phi x$
be injective on $\Sigma^N_k$. Minimal
conditions on $\Phi$ that offer such a guarantee are well-known
(see, e.g. \cite{CDD2009}) and require
at least that $m \geq 2 k$. On the other hand, under stricter
conditions on $\Phi$, such as the restricted isometry property (RIP),
one can recover sparse vectors from their measurements
by numerically efficient methods, such as $\ell^1$-minimization.
Moreover,
the recovery will also be robust when the measurements are corrupted
\cite{CRT}, cf. \cite{D06};
if $\hat y = \Phi x + e$ where $e$ is any vector such that
$\|e\|_2 \leq \epsilon$,
then the solution $x^\#$ of the optimization problem
\begin{equation}\label{ell1_eps_prog}
\min \|z \|_1 \mbox{ subject to } \|\Phi z - \hat y \|_2 \leq \epsilon
\end{equation}
will satisfy $\|x - x^\#\|_2 \lesssim \epsilon$.
The price paid for these stronger recovery guarantees is the somewhat
smaller range of values available for the dimensional parameters $m$,
$k$, and $N$. While there are some explicit (deterministic)
constructions of measurement matrices with stable recovery guarantees,
best results (widest range of values) have been found via random
families of matrices. For example, if the entries of $\Phi$ are
independently sampled from the Gaussian distribution
$\mathcal{N}(0,\frac{1}{m})$, then with high probability, $\Phi$ will satisfy
the RIP (with a suitable set of parameters) if $m \sim k \log (\frac{N}{k})$.
Significant effort has been put on understanding the phase transition
behavior of the RIP parameters for other random families, e.g.,
Bernoulli matrices and random Fourier samplers.
\subsection*{Quantization for compressed sensing measurements}
The robust recovery result mentioned above is essential to the
practicality of compressed sensing, especially from an
analog-to-digital conversion point of view. If a discrete alphabet
$\mathcal{A}$, such as $\mathcal{A} = \delta \mathbb{Z}$ for some step size $\delta > 0$, is to
be employed to replace each measurement $y_j$ with a quantized
measurement $q_j:=\hat y_j \in \mathcal{A}$, then the temptation, in light of
this result, would be to minimize $\|e\|_2=\|y - q\|_2$ over $q \in
\mathcal{A}^m$. This immediately reduces to minimizing $|y_j - q_j|$ for each
$j$, i.e., quantizing each measurement separately to the nearest
element of $\mathcal{A}$, which is usually called Pulse Code Modulation (PCM).
Since $\|y - q\|_2 \leq \frac{1}{2} \delta \sqrt{m}$, the robust
recovery result guarantees that
\begin{equation}\label{err_bound_PCM}
\|x - x^\#_{\mathrm{PCM}}\|_2 \lesssim \delta \sqrt{m}.
\end{equation}
Note that \eqref{err_bound_PCM} is somewhat surprising as
the reconstruction error bound does not improve by increasing
the number of (quantized) measurements; on the contrary, it deteriorates.
However, the $\sqrt{m}$ term is
an artifact of our choice of normalization for the measurement matrix
$\Phi$. In the compressed sensing literature, it is conventional to
normalize a (random) measurement matrix $\Phi$ so that it has unit-norm
columns (in expectation). This is the necessary scaling to achieve
isometry, and for random matrices it
ensures that $\mathbb{E} \|\Phi x\|^2 = \|x\|^2$ for any
$x$, which then leads to the RIP through
concentration of measure and finally to the robust recovery result stated in
\eqref{ell1_eps_prog}. On the other hand, this normalization imposes
an $m$-dependent dynamic range for the measurements which scales as
$1/\sqrt{m}$, hence it is
not fair to use the same value $\delta$ for the quantizer resolution
as $m$ increases. In this
paper, we investigate the dependence of the recovery error on the
number of quantized measurements where $\delta$ is independent of $m$.
A fair assessment of this dependence can be made only if the
dynamic range of each measurement is kept constant while increasing
the number of measurements. This suggests that the natural
normalization in our setting should ensure that the entries of the
measurement matrix $\Phi$ are independent of $m$. In the specific case
of random matrices, we can achieve this by choosing the entries of
$\Phi$ standard i.i.d.~random variables, e.g. according to $\mathcal{N}(0,1)$.
With this normalization of
$\Phi$, the robust recovery result of \cite{CRT}, given above, can be
modified as
\begin{equation} \label{ell1_eps_prog_alt}
\|\hat{y}-y\|_2 \le \epsilon\ \implies \|x-x^\#\|_2 \lesssim
\frac{1}{\sqrt{m}}\epsilon,
\end{equation}
which also replaces \eqref{err_bound_PCM} with
\begin{equation}\label{err_bound_PCM_alt}
\|x - x^\#_{\mathrm{PCM}}\|_2 \lesssim \delta.
\end{equation}
As expected, this error bound does not deteriorate with $m$ anymore.
In this paper, we will adopt this normalization convention
and work with the standard Gaussian distribution $\mathcal{N}(0,1)$
when quantization is involved, but also use the more typical
normalization $\mathcal{N}(0,1/m)$
for certain concentration estimates that will be derived in
Section \ref{main_proof}.
The transition between these two conventions is of course trivial.
The above analysis of quantization error is based on PCM, which
involves separate (independent) quantization of each measurement.
The vast logarithmic reduction of the
ambient dimension $N$ would seem to suggest that this strategy is
essentially optimal since information appears to
be squeezed (compressed) into few uncorrelated measurements. Perhaps
for this reason, the existing literature on quantization of compressed
sensing measurements focused mainly on alternative reconstruction
methods from PCM-quantized measurements and variants thereof, e.g., \cite{BB,ZBC,JHF,dai901quantized,GoyalCSQ,laska2009democracy}. The only exception
we are aware of is \cite{boufounos2007sigma}, which uses $\Sigma\Delta$
modulation to quantize $x$ \emph{before} the random measurements are made.
On the other hand, it is clear that if (once) the support of the signal
is known (recovered), then the $m$ measurements that have been taken
are highly redundant compared to the maximum $k$ degrees of freedom
that the signal has on its support. At this point, the signal may be
considered {\em oversampled}.
However, the error bound \eqref{err_bound_PCM_alt} does not offer
an improvement of reconstruction accuracy, even if additional samples
become available. (The RIP parameters of $\Phi$ are likely to
improve as $m$ increases, but this
does not seem to reflect on the implicit constant factor in
\eqref{err_bound_PCM_alt} satisfactorily.)
This is contrary to the conventional wisdom
in the theory and practice of oversampled quantization in A/D conversion
where reconstruction error decreases as the sampling rate increases, especially
with the use of quantization algorithms specially geared for the
reconstruction procedure. The main goal of this paper is to show how this can
be done in the compressed sensing setting as well.
\subsection*{Quantization for oversampled data}
Methods of quantization have long been studied for oversampled data
conversion. Sigma-delta ($\Sigma\Delta$) quantization (modulation),
for instance, is the dominant method of A/D conversion for audio
signals and relies heavily on oversampling, see
\cite{NST-book,DD,G-exp}. In this setting, oversampling is typically
exploited to employ very coarse quantization (e.g., $1$ bit/sample),
however, the working principle of $\Sigma\Delta$ quantization is
applicable to any quantization alphabet. In fact, it is more
natural to consider $\Sigma\Delta$ quantization as a
``noise\footnote{The quantization error is often modeled as white
noise in signal processing, hence the terminology. However our
treatment of quantization error in this paper is entirely
deterministic.} shaping'' method, for it seeks a quantized signal
$(q_j)$ by a recursive procedure to push the quantization error signal
$y-q$ towards an unoccupied portion of the signal spectrum. In the
case of bandlimited signals, this would correspond to high frequency
bands.
As the canonical example, the standard first-order $\Sigma\Delta$
quantizer computes a bounded solution $(u_j)$ to the difference
equation
\begin{equation}
(\Delta u)_j := u_j - u_{j-1} = y_j - q_j.
\end{equation}
This can be achieved recursively by choosing, for example,
\begin{equation}
q_j = \arg\min_{p \in \mathcal{A}} |u_{j-1} + y_j - p|.
\end{equation}
Since the reconstruction of oversampled bandlimited signals can be
achieved with a low-pass filter $\varphi$ that can also be arranged to
be well-localized in time, the reconstruction error $\varphi*(y-q) =
\Delta \varphi * u$ becomes small due to the smoothness of $\varphi$.
It turns out that, with this procedure, the reconstruction error is
reduced by a factor of the oversampling ratio $\lambda$, defined to be
the ratio of the actual sampling rate to the bandwidth of $\varphi$.
This principle can be iterated to set up higher-order $\Sigma\Delta$
quantization schemes. It is well-known that a reconstruction accuracy
of order $O(\lambda^{-r})$ can be achieved (in the supremum norm) if a
bounded solution to the equation $\Delta^r u = y - q$ can be found
\cite{DD} (here, $r\in \mathbb{N}$ is the order of the associated ${\Sigma\Delta}$
scheme). The boundedness of $u$ is important for practical
implementation, but it is also important for the error bound. The
implicit constant in this bound depends on $r$ as well as
$\|u\|_\infty$. Fine analyses of carefully designed schemes have shown
that optimizing the order can even yield exponential accuracy
$O(e^{-c\lambda})$ for fixed sized finite alphabets $\mathcal{A}$ (see
\cite{G-exp}), which is optimal apart from the value of the constant
$c$. For infinite alphabets, there is no theoretical lower bound for
the quantization error as $\lambda$ increases. (However almost all
practical coding schemes use some form of finite alphabet.)
The above formulation of noise-shaping for oversampled data conversion
generalizes naturally to the problem of quantization of arbitrary
frame expansions, e.g., \cite{BPY}. Specifically, we will consider
finite frames in $\mathbb{R}^k$. Let $E$ be a full-rank $m \times k$ matrix
and $F$ be any left inverse of $E$. In frame theory, one refers to
the collection of the rows of $E$ as the analysis frame and the
columns of $F$ as the synthesis (dual) frame. For any $x \in \mathbb{R}^k$,
let $y=Ex$ be its frame coefficient vector, $q \in \mathcal{A}^m$ be
its quantization, and let $\hat x:= Fq$ be its reconstruction using the
dual frame. Typically $\mathcal{A}^m \cap y + \mathrm{Ker}(F) = \emptyset$, so
we have $\hat x \not= x$. The reconstruction error is given by
\begin{equation}\label{SD-err-1}
x - \hat x = F(y - q),
\end{equation}
and the goal of noise shaping amounts to arranging $q$ in such a way
that $y-q$ is close to $\mathrm{Ker}(F)$.
If the sequence $(f_j)_1^m$ of dual frame vectors were known to vary
smoothly in $j$ (including smooth termination into null vector), then
$\Sigma\Delta$ quantization could be employed without much alteration,
e.g., \cite{LPY,BPA}. However, this need not be the case for many
examples of frames (together with their canonical duals) that are used
in practice. For this reason, it has recently been proposed in
\cite{BLPY} to use special alternative dual frames, called Sobolev
dual frames, that are naturally adapted to $\Sigma\Delta$
quantization. It is shown in \cite{BLPY} (see also Section
\ref{sec2}) that for any frame $E$, if a standard $r$th order
$\Sigma\Delta$ quantization algorithm with alphabet $\mathcal{A} = \delta \mathbb{Z}$
is used to compute $q := q_{\mathrm{\Sigma\Delta}}$, then with an
$r$th order Sobolev dual frame $F := F_{\mathrm{Sob},r}$
and $\hat x_{\mathrm{\Sigma\Delta}} := F_{\mathrm{Sob},r}
q_{\mathrm{\Sigma\Delta}}$, the
reconstruction error obeys the bound
\begin{equation}
\|x - \hat x_{\mathrm{\Sigma\Delta}} \|_2 \lesssim_r
\frac{\delta\sqrt{m}}{\sigma_\mathrm{min}(D^{-r}E)} ,
\end{equation}
where $D$ is the $m\times m$ difference matrix defined by
\begin{equation}\label{def-D}
D_{ij} := \left \{
\begin{array}{rl}
1, & \mbox{if } i=j, \cr
-1, & \mbox{if } i=j+1, \cr
0, & \mbox{otherwise,}
\end{array}
\right.
\end{equation}
and $\sigma_{\min}(D^{-r}E)$ stands for the smallest singular value
of $D^{-r}E$.
\subsection*{Contributions}
For the compressed sensing application that is the subject of this
paper, $E$ will simply be a sub-matrix of the measurement matrix
$\Phi$, hence it may have been found by sampling an i.i.d.~random
variable. Minimum singular values of random matrices with
i.i.d.~entries have been studied extensively in the mathematical
literature. For an $m \times k$ random matrix $E$ with i.i.d.~entries
sampled from a sub-Gaussian distribution with zero mean and unit
variance,\footnote{ As mentioned earlier,
we do not normalize the measurement matrix $\Phi$
in the quantization setting.} one has
\begin{equation}
\sigma_{\min}(E) \geq \sqrt{m}-\sqrt{k}
\end{equation}
with high probability \cite{RV2009}. Note that in general $D^{-r}E$
would not have i.i.d.~entries. A naive lower bound for
$\sigma_{\min}(D^{-r}E)$ would be
$\sigma_{\min}(D^{-r})\sigma_{\min}(E)$. However (see Proposition
\ref{sing_val_Dr}), $\sigma_{\min}(D^{-r})$ satisfies
\begin{equation}
\sigma_{\min}(D^{-r}) \asymp_r 1,
\end{equation}
and therefore this naive product bound yields no improvement on the
reconstruction error for $\Sigma\Delta$-quantized measurements over
the bound \eqref{err_bound_PCM_alt} for PCM-quantized ones. In fact,
the true behavior of $\sigma_{\min}(D^{-r}E)$ turns out to be
drastically different and is described in Theorem~\ref{main_thm_1},
one of our main results (see also Theorem \ref{bound_sigma_min}).
For simplicity, we shall work with standard i.i.d.~Gaussian variables
for the entries of $E$. In analogy with our earlier notation, we
define the ``oversampling ratio'' $\lambda$ of the frame $E$ by
\begin{equation}
\lambda := \frac{m}{k}.
\end{equation}
\begin{bigthm}\label{main_thm_1}
Let $E$ be an $m\times k$ random matrix whose entries are
i.i.d. $\mathcal{N}(0,1)$. For any $\alpha \in (0,1)$, if
$\lambda \geq c (\log m)^{1/(1-\alpha)}$, then with probability at
least $1 - \exp(-c' m \lambda^{-\alpha})$,
\begin{equation}\label{our-sing-val-bound}
\sigma_{\min}(D^{-r}E) \gtrsim_r \lambda^{\alpha(r-\frac{1}{2})}\sqrt{m},
\end{equation}
which yields the reconstruction error bound
\begin{equation} \label{eq14}
\|x - \hat x_{\mathrm{\Sigma\Delta}} \|_2 \lesssim_r \lambda^{-\alpha(r-\frac{1}{2})} \delta.
\end{equation}
\end{bigthm}
While the kind of decay in
this error bound is familiar to $\Sigma\Delta$ modulation, the
domain of applicability of this result is rather
surprising. Previously, the only
setting in which this type of approximation accuracy could be achieved
(with or without Sobolev duals)
was the case of highly structured frames (e.g. when
the frame vectors are found by sampling along
a piecewise smooth frame path).
Theorem~\ref{main_thm_1} shows that such an
accuracy is obtained even when the analysis frame is a random Gaussian
matrix, provided the reconstruction is done via Sobolev duals.
\ignore{This lower bound for the smallest
singular value of $D^{-r}E$ is bigger than the smallest singular value of
$E$ by a factor of $\lambda^{\alpha(r-\frac{1}{2})}$.
It is perhaps interesting to note that this factor is
roughly about the same size as the
$k$th {\em largest} singular value of $D^{-r}$, i.e.,
\begin{equation}
\sigma_{\min}(D^{-r}E) \gtrsim_r \sigma_k(D^{-r}) \sigma_{\min}(E).
\end{equation}
}
In the compressed sensing setting, one needs
\eqref{our-sing-val-bound} to be uniform for all the frames $E$ that
are found by selecting $k$ columns of $\Phi$ at a time.
The proof of Theorem \ref{main_thm_1} extends
in a straightforward manner using a
standard ``union bound'' argument, provided $\lambda$ is known to be
slightly larger. More precisely, if $\Phi$ is an $m \times N$ matrix
whose entries are i.i.d.~according to $\mathcal{N}(0,1)$, and if $\lambda:=m/k
\geq c (\log N)^{1/(1-\alpha)}$, then \eqref{our-sing-val-bound} holds
for all $E = \Phi_T$ with $\#T \leq k$ with the same type of
probability bound (with new constants). This result can be utilized to
improve the reconstruction accuracy of a sparse signal $x$ from its
${\Sigma\Delta}$-quantized compressed sensing measurements if the support $T$ of
$x$ is known. This is because if $T$ is known, $\Phi_T$ is known, and
its Sobolev dual can be found and used in the reconstruction. On the
other hand, for most signals, recovering the exact or approximate
support is already nearly guaranteed by the robust recovery result
shown in \eqref{ell1_eps_prog_alt} together with the stability of the
associated ${\Sigma\Delta}$ quantizer. For example, a simple sufficient condition
for full recovery of the support is that all the $|x_j|$ for $j \in T$
be larger than $C\|y - q_{\mathrm{\Sigma\Delta}}\|_2$ for a suitable
constant $C$. A precise version of this condition is stated in
Theorem~\ref{main_thm_2}.
In light of all these results, we propose $\Sigma\Delta$ quantization
as a more effective alternative of PCM (independent quantization) for
compressed sensing. With high probability on the measurement matrix, a
significant improvement of the reconstruction accuracy of sparse
signals can be achieved through a two-stage recovery procedure:
\begin{enumerate}
\item {\bf Coarse recovery:} $\ell_1$-minimization (or any other robust
recovery procedure) applied to $q_{\mathrm{\Sigma\Delta}}$
yields an initial, ``coarse'' approximation $x^\#$ of $x$,
and in particular, the exact (or approximate) support $T$ of $x$.
\item {\bf Fine recovery:}
Sobolev dual of the frame $\Phi_T$ applied to $q_{\mathrm{\Sigma\Delta}}$
yields a finer approximation $\hat x_{\mathrm{\Sigma\Delta}}$ of $x$.
\end{enumerate}
Combining all these, our second main theorem follows (also
see Theorem~\ref{main_thm_2_1}):
\begin{bigthm}\label{main_thm_2} Let $\Phi$ be an $m \times N$ matrix whose entries are
i.i.d.~according to $\mathcal{N}(0,1)$. Suppose $\alpha \in (0,1)$ and
$\lambda:=m/k \geq c (\log N)^{1/(1-\alpha)}$ where
$c=c(r,\alpha)$. Then
there are two constants $c'$ and $C$ that depend only on $r$ such that
with probability at least $1 - \exp(-c' m \lambda^{-\alpha})$ on the
draw of $\Phi$, the
following holds: For every $x\in \Sigma^N_k$
such that $\min_{j\in \mathrm{supp}(x)} |x_j| \ge C \delta$,
the reconstruction $\hat{x}_{{\Sigma\Delta}}$ satisfies
\begin{equation}\label{sd_error_main}
\|x - \hat{x}_{{\Sigma\Delta}}\|_2 \lesssim_r \lambda^{-\alpha(r-\frac{1}{2})}
\delta.
\end{equation}
\end{bigthm}
To put this result in perspective, note that the approximation error
given in \eqref{sd_error_main} decays as the ``redundancy''
$\lambda=\frac{m}{k}$ increases. In fact, by using an arbitrarily
high order
${\Sigma\Delta}$ scheme, we can make this decay faster than any power law
(albeit with higher constants). Note that such a decay is
not observed in the reconstruction error bound for PCM given
in \eqref{err_bound_PCM_alt}. Of course, one could argue that these
upper bounds may not reflect the actual behavior of the
error. However, in the setting of frame quantization the performance
of PCM is well investigated. In particular, let $E$ be an $m\times k$
real matrix, and let $K$ be a bounded set in $\mathbb{R}^k$. For $x\in K$,
suppose we obtain $q_{\text{PCM}}(x)$ by quantizing the entries
of $y=Ex$ using PCM with alphabet $\mathcal{A}=\delta \mathbb{Z}$. Let
$\Delta_{\text{opt}}$ be an optimal decoder.
Then, Goyal et al. show in \cite{GVT} that
$$
\left[\mathbb{E} \left\|x-\Delta_{\text{opt}}(q_{\text{PCM}}(x))\right\|^2_2\right]^{1/2}
\gtrsim \lambda^{-1} \delta
$$
where $\lambda=m/k$ and the expectation is with respect a probability
measure on $x$ that is, for example, absolutely continuous. This lower bound
limits the extent to which one can improve the reconstruction by means
of alternative reconstruction algorithms from PCM-quantized compressed
sensing measurements. On the other hand, setting, for example,
$\alpha=3/4$ in Theorem~\ref{main_thm_2} we observe that if we use a
second-order ${\Sigma\Delta}$ scheme to quantize the measurements, and if we
adopt the two-stage recovery procedure proposed above, the resulting
approximation will be superior to that produced optimally from
PCM-quantized measurements, provided $m/k$ is sufficiently large.
It is possible to imagine more sophisticated and more effective quantization
and recovery algorithms for compressed sensing. However
using $\Sigma\Delta$ quantization has a number of appealing features:
\begin{itemize}
\item It produces {\bf more accurate} approximations than any known
quantization scheme in this setting (even when sophisticated
recovery algorithms are employed).
\item It is {\bf modular} in the sense that if the fine recovery stage is not available or practical to implement, then the standard (coarse) recovery procedure can still be applied as is.
\item It is {\bf progressive} in the sense that if new measurements arrive
(in any given order), noise shaping can be continued on these measurements
as long as the state of the system ($r$ real values for an $r$th order scheme)
has been stored.
\item It is {\bf universal} in the sense that
it uses no information about the measurement matrix or the signal.
\end{itemize}
The paper is organized as follows.
We review the basics of $\Sigma\Delta$ quantization
and Sobolev duals in frame theory in Section \ref{sec2}, followed by
the reconstruction
error bounds for random Gaussian frames in Section \ref{main_proof}.
We then present the
specifics of our proposed quantization and recovery algorithm for
compressed sensing in Section \ref{SD_CS}. We present our
numerical experiments in Section \ref{numerics} and conclude with extensions
to more general settings in Section \ref{extensions}.
\section{Background on $\Sigma\Delta$ quantization of frame expansions}
\label{sec2}
\subsection*{$\Sigma\Delta$ quantization}
The governing
equation of a standard $r$th order $\Sigma\Delta$ quantization scheme
with input $y = (y_j)$ and output $q = (q_j)$ is
\begin{equation}\label{r-th-SD-eq}
(\Delta^ru)_j = y_j - q_j,~~~j=1,2,\dots,
\end{equation}
where the $q_j \in \mathcal{A}$ are chosen according to some quantization rule
given by
\begin{equation}
q_j = Q(u_{j-1},\dots,u_{j-T},y_j,\dots,y_{j-S}).
\end{equation}
Not all $\Sigma\Delta$ quantization schemes are presented (or
implemented) in this canonical form, but they all can be rewritten as
such for an appropriate choice of $r$ and $u$. We shall not be
concerned with the specifics of the mapping $Q$, except that we need
$u$ to be bounded. The smaller the size of the alphabet $\mathcal{A}$ gets
relative to $r$, the harder it is to guarantee this property. The
extreme case is $1$-bit quantization, i.e., $|\mathcal{A}|=2$, which is
typically the most challenging setting. We will not be working in this
case. In fact, for our purposes, $\mathcal{A}$ will in general have to be
sufficiently fine to allow for the recovery of the support of sparse
signals. In order to avoid technical difficulties, we shall work with
the infinite alphabet $\mathcal{A} = \delta \mathbb{Z}$, but also note that only a
finite portion of this alphabet will be used for bounded signals. A
standard quantization rule that has this ``boundedness'' property is
given by the greedy rule which minimizes $|u_j|$ given
$u_{j-1},\dots,u_{j-r}$ and $y_j$, i.e.,
\begin{equation}\label{greedy-quant}
q_j = \arg\min_{a \in \mathcal{A}} \Big |\sum_{i=1}^r (-1)^{i-1}
{r \choose i} u_{j-i} + y_j - a \Big|.
\end{equation}
It is easy to check that with this rule, one has $|u_j| \leq 2^{-1} \delta$
and $|y_j-q_j| \leq 2^{r-1} \delta$. In turn, if $\|y\|_\infty < C$,
then one needs only $L:=2 \lceil \frac{C}{\delta}\rceil +2^r + 1$
levels. In this case, the associated quantizer is said to be $\log_2
L$-bit, and we have
\begin{equation}
\|u\|_\infty \lesssim \delta \mbox{ and } \|y-q\|_\infty \lesssim_r \delta.
\end{equation}
With more stringent quantization rules, the first inequality would also
have an $r$-dependent constant. In fact, it is known that
for quantization rules with a $1$-bit alphabet,
this constant will be as large as $O(r^r)$, e.g., see \cite{DD,G-exp}.
In this paper, unless otherwise stated, we shall be working with the
greedy quantization rule of \eqref{greedy-quant}.
The initial condition of the recursion in \eqref{r-th-SD-eq} can be
set arbitrarily, but it will be convenient for us to set them equal to
zero for finite frames. With $u_{-r+1} = \cdots = u_0 = 0$, and
$j=1,\dots,m$, the difference equation \eqref{r-th-SD-eq} can be
rewritten as a matrix equation
\begin{equation}\label{r-th-SD-matrix}
D^r u = y - q,
\end{equation}
where $D$ is as in \eqref{def-D}.
\ignore{
More specifically, we have
\begin{equation}
D = \begin{bmatrix}
1 & 0 & 0 & 0 & \cdots & 0 \\
-1 & 1 & 0 & 0 & \cdots & 0\\
0 & -1 & 1 & 0 & \cdots & 0\\
& \ddots & \ddots & \ddots & \ddots & \\
0 & 0 & \cdots & -1 & 1 & 0 \\
0 & 0 & \cdots & 0 & -1 & 1
\end{bmatrix}
\end{equation}
}
As before, we assume $E$ is an $m\times k$ matrix whose rows form the
analysis frame and $F$ is a $k\times m$ left inverse of $E$
whose columns form the dual (synthesis) frame.
Given any $x \in \mathbb{R}^k$, we set $y = Ex$, and define its
$r$th order $\Sigma\Delta$ quantization $q_{\mathrm{\Sigma\Delta}}$
and its reconstruction $\hat x_{\mathrm{\Sigma\Delta}} :=
Fq_{\mathrm{\Sigma\Delta}}$.
Substituting \eqref{r-th-SD-matrix} into \eqref{SD-err-1},
we obtain the error expression
\begin{equation}\label{SD-err}
x - \hat x = FD^r u.
\end{equation}
With this expression, $\|x -\hat x\|$ can be bounded for any norm
$\| \cdot \|$ simply as
\begin{equation}\label{sup-err-bound}
\|x - \hat x \| \leq \|u\|_\infty \sum_{j=1}^m \| (FD^r)_j \|.
\end{equation}
Here $(FD^r)_j$ is the $j$th column of $FD^r$. This bound is also valid
in infinite dimensions, and in fact has been used extensively
in the mathematical treatment of oversampled A/D conversion of bandlimited
functions.
\ignore{For example, for $\| \cdot \| = \| \cdot \|_\infty$ and
$f_j = \tau \varphi(\cdot - j\tau)$, $j \in \mathbb{Z}$,
one can bound the (infinite) sum on the right hand side of
\eqref{sup-err-bound} by
$\tau^{r} \|\varphi^{(r)} \|_1$.}
For $r=1$, and the $\ell_2$ norm, the sum term on the right hand side
motivated the study of the so-called {\em frame variation} defined by
\begin{equation}
V(F) := \sum_{j=1}^{m} \|f_j-f_{j+1}\|_2,
\end{equation}
where $(f_j)$ are the columns of $F$, and one defines $f_{m+1} = 0$.
Higher-order frame variations
to be used with higher-order $\Sigma\Delta$ schemes are defined
similarly, see \cite{BPY,BPY2}.
Frames (analysis as well as synthesis)
that are obtained via uniform sampling a smooth curve in $\mathbb{R}^k$
(so-called {\em frame path}) are typical in many settings. However,
the ``frame variation bound'' is useful in finite dimensions when
the frame path terminates smoothly. Otherwise, it does not provide
higher-order reconstruction accuracy. Designing smoothly
terminating frames can be technically challenging, e.g., \cite{BPA}.
\subsection*{Sobolev duals}
Recently, a more straightforward approach was proposed in \cite{LPY}
for the design of (alternate) duals of finite frames for
$\Sigma\Delta$ quantization. Here, one instead considers the operator
norm of $FD^r$ on $\ell_2$ and the corresponding bound
\begin{equation}\label{SD-op-bound}
\|x - \hat x \|_2 \leq \|FD^r \|_\mathrm{op} \|u\|_2.
\end{equation}
Note that this bound is not available in the infinite dimensional setting of
bandlimited functions due to the fact that $u$ is typically not in $\ell_2$.
It is now natural to minimize
$\|FD^r \|_\mathrm{op}$ over all dual frames of a given analysis frame $E$.
These frames, introduced in \cite{BLPY}, have been called Sobolev duals, in analogy with $\ell_2$-type
Sobolev (semi)norms.
$\Sigma\Delta$ quantization algorithms are normally designed
for analog circuit operation, so they
control $\|u\|_\infty$, which would control $\|u\|_2$ only in a
suboptimal way. However, it turns out that there are important
advantages in working with the $\ell_2$ norm in the analysis.
The first advantage is that
Sobolev duals are readily available by an explicit formula. The
solution $F_{\mathrm{sob},r}$ of the optimization problem
\begin{equation}\label{opt-frame-r}
\min_{F} \|FD^r\|_{\mathrm{op}} \mbox{ subject to } FE = I
\end{equation}
is given by the matrix equation
\begin{equation}\label{sobolev-formula}
F_{\mathrm{sob},r} D^r = (D^{-r}E)^\dagger,
\end{equation}
where $^\dagger$ stands for the Moore-Penrose inversion operator,
which, in our case, is given
by $E^\dagger := (E^* E)^{-1}E^*$. Note that for $r=0$ (i.e., no
noise-shaping, or PCM), one simply obtains $F = E^\dagger$, the
canonical dual frame of $E$.
The second advantage of this approach is that highly developed methods
are present for spectral norms of matrices, especially in the random
setting. Plugging \eqref{sobolev-formula} into \eqref{SD-op-bound},
it immediately follows that
\begin{equation}\label{err-bound-Sob}
\|x - \hat x\|_2 \leq \|(D^{-r}E)^\dagger\|_{\mathrm{op}} \|u\|_2
= \frac{1}{\sigma_{\min}(D^{-r}E)}\|u\|_2,
\end{equation}
where $\sigma_{\min}(D^{-r}E)$ stands for the smallest singular value
of $D^{-r}E$.
\section{Reconstruction error bound for random frames}
\label{main_proof}
In what follows, $\sigma_j(A)$ will denote the $j$th largest singular
value of the matrix $A$. Similarly, $\lambda_j(B)$ will denote the
$j$th largest eigenvalue of the Hermitian matrix $B$. Hence, we have
$\sigma_j(A) = \sqrt{\lambda_j(A^*A)}$. We will also use the notation
$\Sigma(A)$ for the diagonal matrix of singular values of $A$, with
the convention $(\Sigma(A))_{jj} = \sigma_j(A)$. All matrices
in our discussion will be real valued and the Hermitian conjugate
reduces to the transpose.
We have seen that the main object of interest for the reconstruction error
bound is $\sigma_{\min}(D^{-r}E)$ for a random frame $E$.
Let $H$ be a square matrix.
The first observation we make is that when $E$ is
i.i.d.~Gaussian, the distribution of $\Sigma(HE)$
is the same as the distribution of $\Sigma(\Sigma(H)E)$. To see this,
let $U \Sigma(H) V^*$ be the singular value decomposition of
$H$ where $U$ and $V$ are unitary matrices. Then $HE = U \Sigma(H) V^* E$.
Since the unitary transformation $U$ does not alter singular values,
we have $\Sigma(HE) = \Sigma(\Sigma(H) V^* E)$, and because of the unitary
invariance of the i.i.d.~Gaussian measure, the matrix $\tilde E
:= V^*E$ has the same distribution as $E$, hence the claim.
Therefore it suffices to study the singular values of $\Sigma(H)E$.
In our case, $H = D^{-r}$ and we first need information
on the deterministic object $\Sigma(D^{-r})$.
The following result will be sufficient for our purposes:
\begin{prop}\label{sing_val_Dr}
Let $r$ be any positive integer and $D$ be as in \eqref{def-D}. There are
positive numerical constants $c_1(r)$ and $c_2(r)$, independent of $m$,
such that
\begin{equation}\label{sing_val_Dr_bounds}
c_1(r) \Big(\frac{m}{j}\Big)^r \leq \sigma_j(D^{-r}) \leq
c_2(r) \Big(\frac{m}{j}\Big)^r, ~~j=1,\dots,m.
\end{equation}
\end{prop}
The proof of this result is rather standard in the study of
Toeplitz matrices, and is
given in Appendix \ref{sec_sing_val_D}.
\subsection{Lower bound for $\sigma_{\min}(D^{-r}E)$}
\label{lower_bound}
In light of the above discussion,
the distribution of $\sigma_{\min}(D^{-r}E)$
is the same as that of
\begin{equation}
\inf_{\|x\|_2 = 1} \|\Sigma(D^{-r})E x\|_2.
\end{equation}
We replace $\Sigma(D^{-r})$ with an arbitrary diagonal matrix
$S$ with $S_{jj} =: s_j > 0$. The first two results will concern upper
bounds for the norm of independent but
non-identically distributed Gaussian vectors. They are rather standard,
but we include them for the definiteness of our discussion when they will
be used later.
\begin{prop} \label{power2}
Let $\xi \sim \mathcal{N}(0,\frac{1}{m}\mathrm{I}_m)$.
For any $\Theta > 1$,
\begin{equation}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 > \Theta \|s\|^2_\infty \right)
\leq \Theta ^{m/2} e^{-(\Theta -1)m/2}.
\end{equation}
\end{prop}
\begin{proof}
Since $s_j \leq \|s\|_\infty$ for all $j$, we have
\begin{equation}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 > \Theta \|s\|^2_\infty \right)
\leq
\mathbb{P} \left( \sum_{j=1}^m \xi_j^2 > \Theta \right).
\end{equation}
This bound is the (standard)
Gaussian measure of the complement of a sphere of radius
$\sqrt{m\Theta}$ and can be estimated very accurately. We use a simple
approach via
\begin{eqnarray}
\mathbb{P} \left( \sum_{j=1}^m \xi_j^2 > \Theta \right)
& \leq &
\min_{\lambda \geq 0}~ \int_{\mathbb{R}^m} e^{-\left(\Theta
-\sum_{j=1}^m x_j^2\right)\lambda/2} \prod_{j=1}^m
e^{-mx_j^2/2}\,\frac{\mathrm{d}x_j}{\sqrt{2\pi/m}} \cr
& = &
\min_{\lambda \geq 0}~ e^{-\lambda \Theta /2}
(1-\lambda /m)^{-m/2} \cr
& = &
\Theta ^{m/2} e^{-(\Theta -1)m/2},
\end{eqnarray}
where in the last step we set $\lambda = m(1-\Theta ^{-1})$.
\end{proof}
\begin{lemma}\label{lower}
Let $E$ be an $m \times k$
random matrix whose entries are i.i.d.~$\mathcal{N}(0,\frac{1}{m})$.
For any $\Theta > 1$, consider the event
$$\mathcal{E} := \left\{\|S E \|_{\ell^k_2\to \ell^m_2} \leq 2
\sqrt{\Theta} \|s\|_\infty \right \}.$$
Then
$$\mathbb{P} \left( \mathcal{E}^c \right )
\leq 5^k \Theta ^{m/2} e^{-(\Theta -1)m/2}. $$
\end{lemma}
\begin{proof}
We follow the same approach as in \cite{JL_RIP}.
The maximum number of $\rho$-distinguishable points on the
unit sphere in $\mathbb{R}^k$ is at most $(\frac{2}{\rho}+1)^k$.
(This follows by a volume argument\footnote{Balls with radii
$\rho/2$ and centers
at a $\rho$-distinguishable set of points on the unit sphere are mutually
disjoint and
are all contained in the ball of radius $1+\rho/2$ centered at
the origin. Hence there can be at most $(1+\rho/2)^k/(\rho/2)^k$ of them.}
as in e.g., \cite[p.487]{LorentzCA2}.)
Fix a maximal set $Q$ of $\frac{1}{2}$-distinguishable points
of the unit sphere in $\mathbb{R}^k$ with $\# Q \leq 5^k$.
Since $Q$ is maximal, it is a $\frac{1}{2}$-net for the unit sphere.
For each $q \in Q$, consider $\xi_j = (Eq)_j$,
$j = 1,\dots,m$. Then $\xi\sim \mathcal{N}(0,\frac{1}{m}\mathrm{I}_m)$.
As before, we have
$$ \|S E q\|_2^2 = \sum_{j=1}^m s^2_j \xi_j^2. $$
Let $\mathcal{E}(Q)$ be the event
$\left \{ \|S E q \|_2 \leq \sqrt{\Theta} \|s\|_\infty, ~~
\forall q \in Q \right \}$. Then, by Proposition \ref{power2}, we have
the union bound
\begin{equation}
\mathbb{P} \left( \mathcal{E}(Q)^c \right)
\leq 5^k \Theta ^{m/2} e^{-(\Theta -1)m/2}.
\end{equation}
Assume the event $\mathcal{E}(Q)$, and let
$M = \|S E \|_{\ell^k_2\to \ell^m_2}$. For each $\|x\|_2 = 1$, there is $q \in Q$ with $\|q-x\|_2 \leq 1/2$,
hence
$$ \|SEx\|_2 \leq \|SEq\|_2 + \|SE(x-q)\|_2 \leq
\sqrt{\Theta} \|s\|_\infty + \frac{M}{2}.$$
Taking the supremum over all $x$ on the unit sphere, we obtain
$$ M \leq \sqrt{\Theta} \|s\|_\infty + \frac{M}{2}, $$
i.e., $\|S E \|_{\ell^k_2\to \ell^m_2} \leq 2\sqrt{\Theta} \|s\|_\infty$.
Therefore $\mathcal{E}(Q) \subset \mathcal{E}$, and the result follows.
\end{proof}
The following estimate concerns a lower bound for the Euclidean norm
of $(s_1\xi_1,\dots,s_m\xi_m)$. It is not sharp when
the $s_j$ are identical, but it will be useful for our problem
where $s_j = \sigma_j(D^{-r})$ obey a power law (see Corollary
\ref{power1}).
\begin{prop} \label{crude1}
Let $\xi \sim \mathcal{N}(0,\frac{1}{m}\mathrm{I}_m)$.
For any $\gamma > 0$,
\begin{equation}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 < \gamma \right)
\leq \min_{1 \leq L \leq m}
\left( \frac{e\gamma m}{L}\right)^{L/2} (s_1 s_2\cdots s_L)^{-1}.
\end{equation}
\end{prop}
\begin{proof}
For any $t \geq 0$ and any integer $L \in \{1,\dots,m\}$, we have
\begin{eqnarray}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 < \gamma \right)
& \leq &
\int_{\mathbb{R}^m} e^{\left(\gamma -
\sum_{j=1}^m s^2_j x_j^2\right)t/2 } \prod_{j=1}^m
e^{-mx_j^2/2}\,\frac{\mathrm{d}x_j}{\sqrt{2\pi/m}}
\cr
& = & e^{t\gamma/2}
\prod_{j=1}^m \int_{\mathbb{R}}
e^{-x_j^2(m+t s^2_j)/2} \,\frac{\mathrm{d}x_j}{\sqrt{2\pi/m}} \cr
& = & e^{t \gamma/2}
\prod_{j=1}^m (1+t s^2_j/m)^{-1/2} \cr
& \leq & e^{t \gamma/2}
\prod_{j=1}^L (t s^2_j/m)^{-1/2} \cr
& \leq &
e^{t \gamma/2} (m/t)^{L/2} (s_1 s_2 \cdots s_L)^{-1}.
\end{eqnarray}
For any $L$, we can set $t = L/\gamma$, which is the critical point of
the function $t \mapsto e^{t \gamma} t^{-L}$. Since
$L$ is arbitrary, the result follows.
\end{proof}
\begin{cor} \label{power1} Let
$\xi \sim \mathcal{N}(0,\frac{1}{m}\mathrm{I}_m)$, $r$ be a positive integer,
and $c_1 > 0$ be such that
\begin{equation}\label{hypo-s-r}
s_j \geq c_1 \left(\frac{m}{j} \right)^{r},~~~~j=1,\dots,m.
\end{equation}
Then for any $\Lambda \geq 1$ and $m \geq \Lambda $,
\begin{equation}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 < c_1^2 \Lambda ^{2r-1} \right)
< (60 m/\Lambda )^{r/2} e^{-m (r{-}1/2)/\Lambda }.
\end{equation}
\end{cor}
\begin{proof}
By rescaling $s_j$, we can assume $c_1=1$.
For any $L \in \{1,\dots,m\}$, we have
$$ (s_1 s_2\cdots s_L)^{-1} \leq \frac{(L!)^r}{m^{rL}}
< (8 L)^{r/2} \left( \frac{L^r}{e^rm^r} \right)^L,$$
where we have used the coarse estimate $L! < e^{1/12L}(2\pi L)^{1/2}(L/e)^{L}
< (8L)^{1/2}(L/e)^L$.
Setting $\displaystyle \gamma = \Lambda ^{2r-1}$ in Proposition
\ref{crude1}, we obtain
\begin{equation}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 < \Lambda ^{2r-1} \right)
< (8 L)^{r/2}
\left[ \left( \frac{\Lambda L}{em}\right)^{L} \right]^{r-1/2}.
\label{est1}
\end{equation}
We
set $L = \lfloor \frac{m}{\Lambda } \rfloor$. Since $1 \leq \Lambda \leq m$, it is guaranteed that $1 \leq L \leq m$.
Since $\Lambda L\leq m$, we get
$$ \left( \frac{\Lambda L}{em}\right)^{L} \leq e^{-L} < e^{1-\frac{m}{\Lambda }} $$
Plugging this in \eqref{est1} and using $8e^2 < 60$, we find
\begin{eqnarray}
\mathbb{P} \left( \sum_{j=1}^m s^2_j \xi_j^2 < \Lambda ^{2r-1} \right)
& < & (60 m/\Lambda )^{r/2} e^{-m (r{-}1/2)/\Lambda }.
\end{eqnarray}
\end{proof}
\begin{thm} \label{upper}
Let $E$ be an $m \times k$
random matrix whose entries are i.i.d.~$\mathcal{N}(0,\frac{1}{m})$,
$r$ be a positive integer,
and assume that the entries $s_j$ of the diagonal matrix $S$ satisfy
\begin{equation}\label{hypo-s-r-2}
c_1 \left(\frac{m}{j} \right)^{r} \leq s_j \leq c_2 m^r,~~~~j=1,\dots,m.
\end{equation}
Let $\Lambda \geq 1$ be any number and assume
$m \geq \Lambda $. Consider the event
$$ \mathcal{F} := \left \{\|S E x \|_2 \geq
\frac{1}{2} c_1 \Lambda ^{r-1/2} \|x\|_2, ~\forall x \in \mathbb{R}^k \right \}.$$
Then
$$\mathbb{P} \left( \mathcal{F}^c \right )
\leq
5^k e^{-m/2} + 8^r
\left( 17 c_2/c_1\right)^k \Lambda ^{k/2}
\left (\frac{m}{\Lambda } \right)^{r(k+1/2)}
e^{-m(r{-}1/2)\Lambda}. $$
\end{thm}
\begin{proof}
Consider a $\rho$-net $\tilde Q$
of the unit sphere of $\mathbb{R}^k$ with $\# \tilde Q \leq
\big( \frac{2}{\rho}+1 \big )^k$ where the value of $\rho < 1$
will be chosen later. Let
$\tilde{\mathcal{E}}(\tilde Q)$ be the event
$\left \{\|S E q \|_2 \geq c_1 \Lambda ^{r-1/2}, ~~\forall
q \in \tilde Q \right \}$. By Corollary \ref{power1}, we know that
\begin{equation}\label{prob-bound-union}
\mathbb{P} \left( \tilde{\mathcal{E}}(\tilde Q)^c \right)
\leq \left( \frac{2}{\rho}+1 \right )^k \left( \frac{60m}{\Lambda } \right )^{r/2}
e^{-m (r{-}1/2)/\Lambda }.
\end{equation}
Let $\mathcal{E}$ be the event in Lemma \ref{lower} with $\Theta = 4$.
Let $E$ be any given matrix in the event
$\mathcal{E} \cap \tilde{\mathcal{E}}(\tilde Q)$.
For each $\|x\|_2 = 1$, there is $q \in \tilde Q$ with $\|q-x\|_2 \leq \rho$,
hence by Lemma \ref{lower}, we have
$$\|SE(x-q)\|_2 \leq 4 \|s\|_\infty \|x-q\|_2 \leq 4 c_2 m^r \rho. $$
Choose
$$\rho = \frac{c_1 \Lambda^{r-1/2}}{8 c_2 m^r} = \frac{c_1}{8c_2\sqrt{\Lambda}}
\Big(\frac{\Lambda}{m}\Big)^r. $$
Hence
$$ \|SEx\|_2 \geq \|SEq\|_2 - \|SE(x-q)\|_2 \geq
c_1 \Lambda^{r-1/2} - 4 c_2 m^r\rho
= \frac{1}{2} c_1 \Lambda^{r-1/2}.
$$
This shows that $\mathcal{E} \cap \tilde{\mathcal{E}}(\tilde Q)
\subset \mathcal{F}$.
Clearly, $\rho \leq 1/8$ by our choice of parameters
and hence $\frac{2}{\rho} +1
\leq \frac{17}{8\rho}$.
Using the probability bounds of Lemma \ref{lower} and \eqref{prob-bound-union},
we have
\begin{eqnarray}
\mathbb{P} \left( \mathcal{F}^c \right)
& \leq & 5^k 4^{m/2} e^{-3m/2} +
\left( \frac{17}{8\rho} \right )^k \left( \frac{60m}{\Lambda } \right )^{r/2}
e^{-m(r{-}1/2)/\Lambda} \cr
& \leq & 5^k e^{-m/2} + 8^r
(17c_2/c_1)^k \Lambda ^{k/2}
\left (\frac{m}{\Lambda } \right)^{r(k+1/2)}
e^{-m(r{-}1/2)/\Lambda},
\end{eqnarray}
where we have used $2 < e$ and $\sqrt{60} < 8$ for simplification.
\end{proof}
The following theorem is now a direct corollary of the above
estimate.
\begin{thm} \label{bound_sigma_min}
Let $E$ be an $m \times k$
random matrix whose entries are i.i.d.~$\mathcal{N}(0,\frac{1}{m})$,
$r$ be a positive integer, $D$ be the difference matrix defined in
\eqref{def-D}, and the constant $c_1 = c_1(r)$ be
as in Proposition \ref{sing_val_Dr}.
Let $0 < \alpha < 1$ be any number. Assume that
\begin{equation}
\lambda := \frac{m}{k} \geq c_3 (\log m)^{1/(1-\alpha)},
\end{equation}
where $c_3= c_3(r)$ is an appropriate constant. Then
\begin{equation}
\mathbb{P} \left ( \sigma_{\mathrm{min}}(D^{-r}E) \geq c_1
\lambda^{\alpha(r-1/2)} \right ) \geq 1 - 2e^{-c_4 m^{1-\alpha}k^\alpha}
\end{equation}
for some constant $c_4 = c_4(r) > 0$.
\end{thm}
\begin{proof}
Set $\Lambda = \lambda ^\alpha$ in Lemma \ref{upper}.
We only need to show that
$$
\max \left [
5^k e^{-m/2}, 8^r (17c_2/c_1)^k \Lambda ^{k/2}
\left (\frac{m}{\Lambda } \right)^{r(k+1/2)}
e^{-m(r{-}1/2)/\Lambda}
\right ]
\leq e^{-c_4m^{1-\alpha}k^\alpha}.
$$
It suffices to show that
$$ k \log 5 - m/2 \leq - c_4 m^{1-\alpha}k^\alpha $$
and
$$r \log 8 + k \log(17c_2/c_1) + \frac{1}{2}k \log \Lambda
+ r(k+\frac{1}{2}) \log(m/\Lambda ) - (r{-}\frac{1}{2})\frac{m}{\Lambda }
\leq - c_4 m^{1-\alpha}k^\alpha.
$$
The first inequality is easily seen to hold if
$\lambda \geq \frac{\log 5}{\frac{1}{2} - c_4}$.
For the second inequality, first
notice that $m/\Lambda = m^{1-\alpha}k^\alpha$.
Since $k+1/2 \asymp k$, and $r - 1/2 \asymp r$,
it is easily seen that we only need to check that
$$ k \log m \leq c_5 \frac{m}{\Lambda }$$
for a sufficiently small $c_5$. This follows from our assumption on
$\lambda$ by setting $c_5 = 1/c_3^{1-\alpha}$.
\end{proof}
\begin{rem} By replacing $E$ in Theorem~\ref{bound_sigma_min} with $\sqrt{m} E$, we
obtain Theorem~\ref{main_thm_1}.
\end{rem}
\subsection{Implication for compressed sensing matrices}
\begin{thm}\label{sing_val_for_CS}
Let $r$, $D$, $c_1(r)$ be as in Theorem \ref{bound_sigma_min} and
$\Phi$ be an $m \times N$
random matrix whose entries are i.i.d.~$\mathcal{N}(0,\frac{1}{m})$.
Let $0 < \alpha < 1$ be any number and assume that
\begin{equation}
\lambda := \frac{m}{k} \geq c_6 (\log N)^{1/(1-\alpha)},
\end{equation}
where $c_6 = c_6(r)$ is an appropriate constant. Then with probability
at least $1 - 2e^{-c_7 m \lambda^{-\alpha}}$
for some $c_7 = c_7(r) > 0$, every
$m \times k$ submatrix $E$ of $\Phi$ satisfies
\begin{equation} \label{eq_cs_sd}
\sigma_{\mathrm{min}}(D^{-r}E) \geq c_1
\lambda^{\alpha(r-1/2)}.
\end{equation}
\end{thm}
\begin{proof}
We will choose $c_7 = c_4/2$, where $c_4$ is as in
Theorem \ref{bound_sigma_min}. The proof will follow immediately by
a union bound once we show that
$${N \choose k} \leq e^{\frac{1}{2}c_4 m^{1-\alpha}k^\alpha}.$$
Since ${N \choose k} \leq N^k$, it suffices to show that
$$ k \log N \leq \frac{c_4}{2} m^{1-\alpha}k^\alpha. $$
Both this condition and the hypothesis of Theorem \ref{bound_sigma_min} will
be satisfied if we choose
$$c_6 = \max(c_3,(2/c_4)^{1/(1-\alpha)}).$$
\end{proof}
\begin{rem}
If $\Phi$ is a Gaussian matrix with entries i.i.d. $\mathcal{N}(0,1)$ rather
than $\mathcal{N}(0,\frac{1}{m})$, Theorem~\ref{sing_val_for_CS} applied to
$\frac{1}{\sqrt{m}}\Phi$ implies that every $m\times k$ submatrix $E$ of $\Phi$
satisfies
\begin{equation} \label{eq_cs_sd_1}
\sigma_{\mathrm{min}}(D^{-r}E) \geq c_1
\lambda^{\alpha(r-1/2)}\sqrt{m}.
\end{equation}
\end{rem}
\section{$\Sigma\Delta$ quantization of compressed sensing
measurements}
\label{SD_CS}
In this section we will assume that the conditions of Theorem
\ref{sing_val_for_CS} are satisfied for some $0 < \alpha < 1$ and $r$,
and the measurement matrix $\Phi$ that is drawn from $\mathcal{N}(0,1)$
yields \eqref{eq_cs_sd_1}. For definiteness, we also
assume that $\Phi$ admits the robust recovery constant $C_1=10$, i.e.,
the solution $x^\#$ of the program \eqref{ell1_eps_prog} satisfies
$$\|\hat{y}-y\|_2\le \epsilon \ \implies \ \|x-x^\#\|_2 \leq 10 \frac{1}{\sqrt{m}}\epsilon.
$$
Note again that our choice of normalization for the measurement matrix
$\Phi$ is different from the compressed sensing convention. As
mentioned in the Introduction, it is more appropriate to work with a
measurement matrix $\Phi \sim \mathcal{N}(0,1)$ in order to be able to use a
quantizer alphabet that does not depend on $m$. For this reason, in
the remainder of the paper, $\Phi$ shall denote an $m \times N$ matrix
whose entries are i.i.d. from $\mathcal{N}(0,1)$.
Let $q :=
q_{\Sigma\Delta}$ be output of the standard greedy $r$th order
$\Sigma\Delta$ quantizer with the alphabet $\mathcal{A} = \delta \mathbb{Z}$ and input
$y$. As stated in Section \ref{sec2}, we know that $\|y - q\|_\infty
\leq 2^{r-1}\delta$ and therefore $\|y - q\|_2 \leq
2^{r-1}\delta\sqrt{m}$.
\subsection*{Coarse recovery and recovery of support}
Our first goal is to recover the support $T$ of $x$. For this purpose
we shall use a coarse approximation of $x$. Let
\begin{equation}
x' := \arg \min \|z \|_1 \mbox{ subject to }
\left \|\Phi z - q \right \|_2 \leq
\epsilon := 2^{r-1}\delta \sqrt{m}.
\end{equation}
By the robust recovery result (for our choice of normalization for
$\Phi$), we know that
$$\|x - x' \|_2 \leq \eta := 5\cdot 2^{r}\delta.$$
The simplest attempt to recover $T$ from $x'$ is to pick
the positions of its $k$ largest entries. This attempt can fail
if some entry of $x_j$ on $T$ is smaller than $\eta$ for then
it is possible that $x'_j = 0$ and therefore $j$ is not picked.
On the other hand,
it is easy to see that if the smallest nonzero entry of $x$ is strictly
bigger than $2\eta$ in magnitude, then this method always succeeds.
(Since $\|x - x' \|_\infty \leq \eta$, the entries of
$x'$ are bigger than $\eta$ on $T$ and less than $\eta$ on $T^c$.)
The constant $2$ can be replaced with $\sqrt{2}$ by a more careful
analysis, and can be pushed arbitrarily close to $1$ by picking more
than $k$ positions. The
proposition below gives a precise condition on how well this can
be done. We also provide a bound on how much of $x$ can potentially be
missed if no lower bound on $|x_j|$ is available for $j \in T$.
\begin{prop}\label{supp_recov}
Let $\|x - x'\|_{\ell_2^N} \leq \eta$, $T = \mathrm{supp}~x$
and $k = |T|$. For
any $k' \in \{k,\dots,N{-}1\}$, let $T'$ be the support of (any of) the
$k'$ largest entries of $x'$.
\begin{itemize}
\item[\rm{(i)}] $\|x_{T\setminus T'}\|_2 \leq \beta \eta$
where $\beta \leq \left(1+\frac{k}{k'}\right)^{1/2}$.
\item[\rm{(ii)}]
If $|x_j| > \gamma \eta$ for all $j \in T$,
where $\gamma := \left (1 + \frac{1}{k'-k+1}\right )^{1/2}$,
then $T' \supset T$.
\end{itemize}
\end{prop}
\begin{proof}
\rm{(i)}
We have
\begin{equation}\label{error_split}
\sum_{j\in T} |x_j - x'_j|^2 + \sum_{j \in T^c} |x'_j|^2 =
\|x - x'\|_2^2 \leq \eta^2.
\end{equation}
In particular, this implies
\begin{equation}\label{TT'}
\sum_{j\in T\setminus T'} |x_j - x'_j|^2 + \sum_{j \in T'\setminus T}
|x'_j|^2 \leq \eta^2.
\end{equation}
Suppose $T \setminus T' \not= \emptyset$. Then $T'\setminus T$ is also
nonempty. In fact, we have
$$|T'\setminus T| = |T \setminus T'| + k'-k.$$
Now, observe that
$$
\frac{1}{|T\setminus T'|} \sum_{j \in T\setminus T'}
|x'_j|^2 \leq
\max_{j \in T\setminus T'} ~|x'_j|^2
\leq \min_{j \in T'\setminus T}~ |x'_j|^2
\leq \frac{1}{|T'\setminus T|} \sum_{j \in T'\setminus T}
|x'_j|^2,
$$
which, together with \eqref{TT'} implies
$$ \|x_{T\setminus T'}\|_2 \leq \|x'_{T\setminus T'}\|_2
+ \|(x-x')_{T\setminus T'}\|_2
\leq
\|x'_{T\setminus T'}\|_2 +
\sqrt{\eta^2 - \frac{|T'\setminus T|}{|T\setminus T'|}
\|x'_{T\setminus T'}\|^2_2}.
$$
It is easy to check that for any $A >0$, and any $0 \leq t \leq \eta/\sqrt{A}$,
\begin{equation}\label{lin_quad_max}
t + \sqrt{\eta^2 - At^2} \leq
\left(1+\frac{1}{A}\right)^{1/2} \eta.
\end{equation}
The result follows by
setting $A = |T'\setminus T|/|T\setminus T'|$ and noticing that
$A \geq k'/k$.
\rm{(ii)}
Let $z_1 \geq \cdots \geq z_N$ be the decreasing
rearrangement of $|x'_1|, \dots, |x'_N|$. We have
$$\sum_{j \in T} |x'_j|^2 \leq \sum_{i=1}^{k} z_i^2$$
so
$$\sum_{j \in T^c} |x'_j|^2 \geq \sum_{i=k+1}^{N} z_i^2 \geq
\sum_{i=k+1}^{k'+1} z_i^2 \geq (k'-k+1)z_{k'+1}^2.$$
Hence by \eqref{error_split} we have
$$\max_{j \in T} |x_j - x'_j|^2 + (k'-k+1) z_{k'+1}^2 \leq \eta^2.$$
Since $|x'_j| \geq |x_j| - |x_j - x'_j|$, the above inequality now implies
$$\min_{j \in T} |x'_j|
\geq \min_{j \in T}|x_j| - \max_{j \in T}|x_j - x'_j|
\geq \min_{j \in T}|x_j| - \sqrt{\eta^2 - (k'-k+1)z_{k'+1}^2}.$$
Now, another application of \eqref{lin_quad_max} with
$A = k'-k+1$ yields
$$- \sqrt{\eta^2 - (k'-k+1)z_{k'+1}^2} \geq z_{k'+1} - \gamma \eta$$
and therefore
$$\min_{j \in T} |x'_j| \geq \min_{j \in T}|x_j| +
z_{k'+1} - \gamma \eta > z_{k'+1} = \max_{j \in T'^c} |x'_j|.$$
It is then clear that $T \subset T'$ because if $T'^c \cap T \not= \emptyset$, the
inequality
$$\max_{j \in T'^c} |x'_j| \geq \max_{j \in T'^c \cap T} |x'_j|
\geq \min_{j \in T} |x'_j|$$
would give us a contradiction.
\end{proof}
Note that if the $k'$ largest entries of $x'$ are picked with
$k' > k$, then one would need to work with $T'$ for the fine
recovery stage, and therefore the starting assumptions on
$\Phi$ have to be modified for $k'$. For simplicity we shall
stick to $k'=k$ and consequently $\gamma = \sqrt{2}$.
\subsection*{Fine recovery}
Once $T$ is found, the $r$th order
Sobolev dual frame $F:=F_{\mathrm{Sob},r}$ of $E = \Phi_T$ is computed
and we set $\hat x_{\Sigma\Delta} = Fq$. We now
restate and prove Theorem~\ref{main_thm_2}.
\begin{thm}\label{main_thm_2_1}
Let $\Phi$ be an $m \times N$ matrix whose entries are
i.i.d.~according to $\mathcal{N}(0,1)$. Suppose $\alpha \in (0,1)$ and
$\lambda:=m/k \geq c (\log N)^{1/(1-\alpha)}$ where $c=c(r,\alpha)$. Then
there are two constants $c'$ and $C$ that depend only on $r$ such that
with probability at least $1 - \exp(-c' m \lambda^{-\alpha})$ on the
draw of $\Phi$, the
following holds: For every $x\in \Sigma^N_k$
such that $\min_{j\in \mathrm{supp}(x)} |x_j| \ge C \delta$,
the reconstruction $\hat{x}_{{\Sigma\Delta}}$ satisfies
\begin{equation}\label{sd_error_main_1}
\|x - \hat{x}_{{\Sigma\Delta}}\|_2 \lesssim_r \lambda^{-\alpha(r-\frac{1}{2})}
\delta.
\end{equation}
\end{thm}
\begin{proof} Suppose that $\lambda \geq c (\log
N)^{1/(1-\alpha)}$ with $c=c_6$ as in the proof of
Theorem~\ref{sing_val_for_CS}. Let $q_{\Sigma\Delta}$ be obtained by quantizing
$y:=\Phi x$ via an $r$th order ${\Sigma\Delta}$ scheme with alphabet $\mathcal{A}=\delta
\mathbb{Z}$ and with the quantization rule as in \eqref{greedy-quant}, and
let $u$ be the associated state sequence as in
\eqref{r-th-SD-eq}. Define $x^\#$ as the solution of the program
$$\min \|z\|_1\ \text{subject to}\ \|\Phi z - q_{\Sigma\Delta} \|_2 \le
\epsilon.
$$
Suppose that $\Phi$ admits the robust recovery constant $C_1$, i.e.,
the solution $x^\#$ of the program \eqref{ell1_eps_prog_alt} satisfies
$\|x-x^\#\|_2 \leq C_1 \epsilon/\sqrt{m}$ for every $x$ in
$\Sigma^N_k$ provided that $\|y - q_{\Sigma\Delta}\| \le \epsilon$. Note that
$C_1$, as given for example in \cite{CRT}, only depends on the RIP
constants of $\Phi$ and is well-behaved if $m$ and $N$ satisfy the
hypothesis of the theorem. As discussed in Section~\ref{sec2}, in this
case we have
$\|y-q_{\Sigma\Delta}\|_2 \le 2^{r-1} \delta \sqrt{m}$
which implies
$$\|x-x^\#\|_2 \leq
C_1 2^{r-1} \delta.$$
Assume that
\begin{equation}\label{size_cond_x_0}
\min_{j \in T} |x_j| \geq C_1\cdot 2^{r-1/2}\delta =: C\delta.
\end{equation}
Then, Proposition~\ref{supp_recov} (with $\gamma=\sqrt{2}$ and
$\eta=C_1 2^{r-1}$) shows that $T'$, the support of the $k$
largest entries of $x^\#$, is identical to the support $T$ of $x$.
Finally, set
$$\hat{x}_{\Sigma\Delta}=F_{\text{sob},r} q_{\Sigma\Delta}$$
where $F_{\text{sob},r}$ is the $r$th order Sobolev dual of $\Phi_T$.
Using the fact that $\|u\|_2 \le 2^{-1} \delta \sqrt{m}$ (see
Section~\ref{sec2}) together with the conclusion of
Theorem~\ref{sing_val_for_CS} and the error bound
\eqref{err-bound-Sob}, we conclude that
\begin{equation}
\|x - \hat x_{\Sigma\Delta} \|_2 \leq
\frac{\|u\|_2}{\sqrt{m}\,\sigma_{\mathrm{min}}(D^{-r}E)} \leq
\frac{\lambda^{-\alpha(r-1/2)}}{2c_1}
\delta.
\end{equation}
Note that the RIP and therefore the robust recovery will hold with probability
$1-\exp(c''m)$, and our Sobolev dual reconstruction error bound will hold
with probability $1 - \exp(-c_7 m \lambda^{-\alpha})$. Here $c_1$ and
$c_7$ are as in the proof of Theorem~\ref{sing_val_for_CS}.
\end{proof}
\begin{rem}
To interpret the size condition in a concrete case, assume that
$\Phi$ admits the robust recovery constant $C_1=10$, and that we
have
\begin{equation}\label{size_cond_x}
\min_{j \in T} |x_j| \geq \sqrt{2} \eta = 5\cdot 2^{r+1/2}\delta.
\end{equation}
If PCM is used as the quantization method, then
the best error guarantee we have that holds uniformly on $T$ would be
$$\|x - x^\#_{\mathrm{PCM}} \|_\infty \leq \|x - x^\#_{\mathrm{PCM}} \|_2
\leq 5 \delta.$$
It can be argued that the approximately
recovered entries of $x^\#_{\mathrm{PCM}}$ are meaningful
only when the minimum nonzero entry of $x$ is at least as large
as the maximum uncertainty in $x^\#_{\mathrm{PCM}}$, which is
only known to be bounded by $5\delta$. Hence,
in some sense the size condition \eqref{size_cond_x} is natural
(modulo the factor $2^{r+1/2}$).
\end{rem}
\subsection*{Quantizer choice and rate-distortion issues}
So far we have not made any assumptions on the step size $\delta$ of
the uniform infinite quantizer $\mathcal{A} = \delta \mathbb{Z}$. An important question concerns
how large $\delta$ should be for the most effective use of resources.
This question is motivated by the fact that infinite quantizers are not
practical and have to be replaced by finite ones. In the same vein, an alternative
question is to determine the minimum number of bits that the quantizer
needs to incorporate as well as
the resulting approximation error. First, let us assume that
\begin{equation}\label{dyadic_size_assumption}
A\leq |x_j| \leq \rho := 2^b A ~~~\mbox{ for all }
j \in T.
\end{equation}
For usefulness of our results, one would be interested in
the regime $A \ll \rho$. Thus, we introduce $2^b = \rho/A $ to
represent the number of dyadic scales over which the input is allowed
to range. Clearly, $\delta_r$, the quantization step size used by an
$r$th order $\Sigma\Delta$ scheme for our support recovery results to
hold must satisfy $\delta_r \leq \frac{A/5}{2^{r+1/2}}$ (as before, we
assume $C_1=10$). Let us for the moment use the largest allowable
step-size, i.e., set
\begin{equation}\label{delta_r}
\delta_r := \frac{A/5}{2^{r+1/2}}.
\end{equation}
Next, let us assume that a $B_r$-bit uniform quantizer of step size
$\delta_r$ is to replace $\mathcal{A} = \delta \mathbb{Z}$. We know that $\|q\|_\infty$
could be as large as $2^{r-1} \delta_r + \|y\|_\infty$, therefore we
need to bound $\|y\|_\infty$ efficiently. If we use the RIP, then
$\Phi$ does not expand the $\ell_2$-norm of $k$-sparse vectors by more
than a factor of $2\sqrt{m}$ (note our choice of normalization for
$\Phi$), and therefore it follows that
$$\|y\|_\infty \leq \|y\|_2 \leq 2\sqrt{m} \|x\|_2 \leq 2 \rho\sqrt{mk},$$
which is a restatement of the inequality
$$ \|E \|_{\ell_\infty^k \to \ell_\infty^m} \leq
\sqrt{k} \|E \|_{\ell_2^k \to \ell_2^m}$$
that holds for any $m \times k$ matrix $E$. However, it can be argued
that the $(\infty,\infty)$-norm of a random matrix should typically be
smaller. In fact, if $E$ were drawn from the Bernoulli model, i.e.,
$E_{ij} \sim \pm 1$, then we would have
$$
\|E \|_{\ell_\infty^k \to \ell_\infty^m} =
k = \lambda^{-1/2} \sqrt{mk},
$$
as can easily be seen from the general formula
\begin{equation}
\|E \|_{\ell_\infty^k \to \ell_\infty^m} =
\max_{1 \leq i \leq m} \sum_{j=1}^k |E_{ij}|.
\end{equation}
Using simple concentration inequalities for Gaussian random
variables, it turns out that for the range of aspect ratio $\lambda =
m/k$ and probability of encountering a matrix $\Phi$ that we are
interested in, we have $\|E \|_{\ell_\infty^k \to \ell_\infty^m} \leq
\lambda^{-\alpha/2} \sqrt{mk}$ for every $m \times k$ submatrix $E$ of
$\Phi$. We start with the following estimate:
\begin{prop}\label{l1_Gauss}
Let $\xi_1,\dots,\xi_k$ i.i.d.~standard Gaussian variables. Then,
for any $\Theta > 1$,
\begin{equation}
\mathbb{P} \left( \sum_{j=1}^k |\xi_j| > \Theta \right)
\leq 2^k e^{-\Theta^2 / (2k)}.
\end{equation}
\end{prop}
\begin{proof}
\begin{eqnarray}
\mathbb{P} \left( \sum_{j=1}^k |\xi_j| > \Theta \right)
& \leq &
\min_{t \geq 0}~ \int_{\mathbb{R}^k} e^{-\left(\Theta
-\sum_{j=1}^k |x_j|\right)t} \prod_{j=1}^k
e^{-x_j^2/2}\,\frac{\mathrm{d}x_j}{\sqrt{2\pi}} \cr
& = &
\min_{t \geq 0}~ e^{- \Theta t}
\left (e^{t^2/2} \int_\mathbb{R} e^{-\frac{1}{2} (|x| - t)^2}
\,\frac{\mathrm{d}x}{\sqrt{2\pi}} \right)^k\cr
& = &
\min_{t \geq 0}~ e^{- \Theta t}
\left (2 e^{t^2/2} \int_0^\infty e^{-\frac{1}{2} (x - t)^2}
\,\frac{\mathrm{d}x}{\sqrt{2\pi}} \right)^k\cr
& \leq &
2^k \min_{t \geq 0}~ e^{- \Theta t + k t^2/2} \cr
& = &
2^k e^{-\Theta^2 / (2k)}.
\end{eqnarray}
where in the last step we set $t = \Theta/k$.
\end{proof}
\begin{prop}\label{inf_inf_Gauss}
Let $\Phi$ be an $m \times N$ random matrix whose entries are
i.i.d.~$\mathcal{N}(0,1)$. Let $0 < \alpha < 1$ be any number and assume
that
\begin{equation}\label{inf_inf_cond}
\lambda := \frac{m}{k} \geq c_1 (\log N)^{1/(1-\alpha)},
\end{equation}
where $c_1$ is an appropriate constant. Then with probability at least
$1 - e^{-c_2 m^{1-\alpha}k^\alpha}$ for some $c_2 > 0$, every $m
\times k$ submatrix $E$ of $\Phi$ satisfies
\begin{equation}
\|E \|_{\ell_\infty^k \to \ell_\infty^m} \leq
\lambda^{-\alpha/2} \sqrt{mk}.
\end{equation}
\end{prop}
\begin{proof}
Proposition \ref{l1_Gauss} straightforwardly implies that
\begin{equation}
\mathbb{P} \left( \{ \exists T \mbox { such that } |T|=k \mbox{ and }
\| \Phi_T \|_{\ell_\infty^k \to \ell_\infty^m} > \Theta\}
\right)
\leq {N \choose k} m 2^k e^{-\Theta^2 / (2k)}.
\end{equation}
Let $\Theta = \lambda^{-\alpha/2} \sqrt{mk}$. It remains to show that
$$ k \log N + k \log 2 + \log m + c_2 m^{1-\alpha} k^\alpha
\leq \frac{\Theta^2}{2k}.$$
If $c_1$ in \eqref{inf_inf_cond} is sufficiently large and $c_2$ is
sufficiently small, then the expression on the left hand side is
bounded by $ k \lambda^{1-\alpha} /2 = \Theta^2/(2k)$.
\end{proof}
\par
Without loss of generality, we may now assume that $\Phi$ also
satisfies the conclusion of Proposition \ref{inf_inf_Gauss}. Hence we
have an improved bound on the range of $y$ given by
\begin{equation}
\|y \|_\infty \leq \rho\lambda^{-\alpha/2} \sqrt{mk}
= \rho \lambda^{(1-\alpha)/2} k.
\end{equation}
We assume $B_r$ is chosen to satisfy
\begin{equation}
\label{bit_eq}
2^{B_r-1} \delta_r = 2^{r-1}\delta_r + \rho \lambda^{(1-\alpha)/2} k,
\end{equation}
so that the quantizer is not overloaded. Since $\rho/\delta_r
\approx 2^{r + 1/2 + b}$ by \eqref{dyadic_size_assumption} and
\eqref{delta_r}, we see that the second term on the
right hand side of \eqref{bit_eq} is significantly
larger than the first, which implies
\begin{equation}\label{dominant_term}
2^{B_r-1} \delta_r \approx 2^b A \lambda^{(1-\alpha)/2} k.
\end{equation}
Hence, using \eqref{delta_r} again, $B_r$ must satisfy
\begin{equation}\label{eq:rate}
2^{B_r-1}\approx 5~2^{b+r+1/2}\lambda^{(1-\alpha)/2}k.
\end{equation}
Based on Theorem \ref{main_thm_2_1}, the approximation error (the distortion)
$\mathscr{D}_{\Sigma\Delta}$ incurred
after the fine recovery stage via Sobolev duals satisfies the bound
\begin{equation}
\mathscr{D}_{\Sigma\Delta} \lesssim_r
\lambda^{-\alpha(r-1/2)} \delta_r \approx
\frac{ \lambda^{-\alpha( r - 1/2)}
A}{2^{r+1/2}}. \label{eq:distortion}
\end{equation}
A similar calculation for the PCM encoder with the same step size $\delta_r$
and the standard $\ell_1$ decoder results in the necessity for
roughly the same number of bits $B_r$ as the $\Sigma\Delta$ encoder
(because of the approximation \eqref{dominant_term}), but provides only
the distortion bound
\begin{equation}
\mathscr{D}_{\mathrm{PCM}} \lesssim
\delta_r \approx \frac{A}{2^{r+1/2}}.
\label{eq:distortion_PCM}
\end{equation}
Note that the analysis above requires that both PCM and $\Sigma\Delta$
encoders utilize high-resolution quantizers, however the benefit of
using $\Sigma\Delta$ encoders is obvious upon comparing
\eqref{eq:distortion} and \eqref{eq:distortion_PCM}.
\section{Numerical experiments}
\label{numerics}
In order to test the accuracy of Theorem \ref{bound_sigma_min}, our
first numerical experiment concerns the minimum singular value of
$D^{-r}E$ as a function of $\lambda = m/k$. In Figure \ref{fig1}, we
plot the worst case (the largest) value, among $1000$ realizations, of
$1/\sigma_{min}(D^{-r}E)$ for the range $1 \leq \lambda \leq 25$,
where we have kept $k=50$. As predicted by this theorem, we find that
the negative slope in the log-log scale is roughly equal to $r-1/2$,
albeit slightly less, which seems in agreement with the presence of
our control parameter $\alpha$. As for the size of the $r$-dependent constants,
the function $5^r \lambda^{-r+1/2}$
seems to be a reasonably close numerical fit, which also explains why
we observe the separation of the individual curves after $\lambda > 5$.
Our next experiment involves the full quantization algorithm for
compressed sensing including the ``recovery of support'' and ``fine
recovery'' stages. To that end, we first generate a $1000\times 2000$
matrix $\Phi$, where the entries of $\Phi$ are drawn i.i.d. according
to $\mathcal{N}(0,1)$. To examine the performance of the proposed scheme as the
redundancy $\lambda$ increases in comparison to the performance of the
standard PCM quantization, we run a set of experiments: In each
experiment we fix the sparsity $k\in\{5,10,20,40\}$, and we
generate $k$-sparse signals $x$ with the non-zero entries of each
signal supported on a random set $T$, but with magnitude
$1/\sqrt{k}$. This ensures that $\|x\|_2=1$. Next, for
$m\in\{100,200,...,1000\}$ we generate the measurements $y=\Phi^{(m)}
x$, where $\Phi^{(m)}$ is comprised of the first $m$ rows of
$\Phi$. We then quantize $y$ using PCM, as well as the $1$st and $2$nd
order $\Sigma\Delta$ quantizers, defined via \eqref{r-th-SD-eq} and
\eqref{greedy-quant} (in all cases the quantizer step size is
$\delta=10^{-2}$). For each of these quantized measurements $q$, we
perform the coarse recovery stage, i.e., we solve the associated
$\ell_1$ minimization problem to recover a coarse estimate of $x$ as
well as an estimate $\widetilde{T}$ of the support $T$. The
approximation error obtained using the coarse estimate (with PCM
quantization) is displayed in Figures \ref{fig2} and \ref{fig3} (see
the dotted curve). Next, we implement the fine recovery stage of our
algorithm. In particular, we use the estimated support set
$\widetilde{T}$ and generate the associated dual
$F_{\text{sob},r}$. Defining
$F_{\text{sob},0}:=(\Phi_{\widetilde{T}}^{(m)})^\dagger$, in each
case, our final estimate of the signal is obtained via the fine
recovery stage as $\hat{x}_{\tilde{T}}=F_{\text{sob},r}q$,
$\hat{x}_{\tilde{T^c}}=0$. Note that this way, we obtain an
alternative reconstruction also in the case of PCM. We repeat this
experiment $100$ times for each $(k,m)$ pair and plot the average of
the resulting errors $\|x-\tilde{x}\|_2$ as a function of $\lambda$ in
Figure \ref{fig2} as well as the maximum of $\|x-\hat{x}\|_2$ in
Figure \ref{fig3}. For our final experiment, we choose the entries of
$x_T$ i.i.d. from $\mathcal{N}(0,1)$, and use a quantizer step size
$\delta=10^{-4}$. Otherwise, the experimental setup is identical to
the previous one. The average of the resulting errors
$\|x-\tilde{x}\|_2$ as a function of $\lambda$ is reported in Figure
\ref{fig4} and the maximum of $\|x-\hat{x}\|_2$ in Figure \ref{fig5}.
The main observations that we obtain from these experiments are as follows:
\begin{itemize}
\item ${\Sigma\Delta}$ schemes outperform the coarse reconstruction obtained from
PCM quantized measurements significantly even when $r=1$ and even
for small values of $\lambda$.
\item For the $\Sigma\Delta$ reconstruction error, the negative slope in the
log-log scale is roughly equal to $r$. This outperforms the (best
case) predictions of Theorem \ref{main_thm_2} which are obtained
through the operator
norm bound and suggests the presence of
further cancellation due to the statistical
nature of the $\Sigma\Delta$ state variable $u$, similar to the
white noise hypothesis.
\item When a fine recovery stage is employed in the case of PCM
(using the Moore-Penrose pseudoinverse of the submatrix of $\Phi$
that corresponds to the estimated support of $x$), the approximation
is consistently improved (when compared to the coarse
recovery). Moreover, the associated approximation error is observed to be of
order $O(\lambda^{-1/2})$, in contrast with the error corresponding
to the coarse recovery from PCM quantized measurements (with the
$\ell_1$ decoder only) where the
approximation error does not seem to depend on $\lambda$. A
rigorous analysis of this behaviour will be given in a separate
manuscript.
\end{itemize}
\begin{figure}[b]
\centerline{\includegraphics[width=5in]{min_sing_val}}
\caption{Numerical behavior (in log-log scale) of $1/\sigma_{min}(D^{-r}E)$
as a function of $\lambda = m/k$, for $r=0,1,2,3,4$. In this figure,
$k=50$ and $1 \leq \lambda \leq 25$.
For each problem size, the largest value of
$1/\sigma_{min}(D^{-r}E)$ among
$1000$ realizations of a random $m \times k$
matrix $E$ sampled from the Gaussian ensemble $\mathcal{N}(0,\frac{1}{m}I_m)$
was recorded.}
\label{fig1}
\end{figure}
\begin{figure}[b]
\begin{center}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s5}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s10}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s20}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s40}}
\caption{\label{fig2} The average performance of the proposed $\Sigma\Delta$ quantization and reconstruction schemes for various values of $k$.
For this experiment the non-zero entries of $x$ are constant and $\delta=0.01$.}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s5}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s10}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s20}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s40}}
\caption{\label{fig3} The worst case performance of the proposed $\Sigma\Delta$ quantization and reconstruction schemes for various values of $k$.
For this experiment the non-zero entries of $x$ are constant and $\delta=0.01$.}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_exp25}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_exp210}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_exp220}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_exp240}}
\caption{\label{fig4} The average performance of the proposed $\Sigma\Delta$ quantization and reconstruction schemes for various values of $k$.
For this experiment the non-zero entries of $x$ are i.i.d. $\mathcal{N}(0,1)$ and $\delta=10^{-4}$.}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_exp25}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_exp210}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_exp220}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_exp240}}
\caption{\label{fig5} The worst case performance of the proposed $\Sigma\Delta$ quantization and reconstruction schemes for various values of $k$.
For this experiment the non-zero entries of $x$ are i.i.d. $\mathcal{N}(0,1)$ and $\delta=10^{-4}$.}
\end{center}
\end{figure}
\section{Remarks on extensions}
\label{extensions}
\subsection{Other noise shaping matrices}
In the above approach, the particular quantization scheme that we use
can be identified with its ``noise-shaping matrix'', which is $D^r$ in
the case of an $r$th order ${\Sigma\Delta}$ scheme and the identity matrix in the
case of PCM.
The results we obtained above are valid for the aforementioned
noise-shaping matrices. However, our techniques are fairly general and
our estimates can be modified to investigate the accuracy obtained
using an arbitrary quantization scheme with the associated invertible
noise-shaping matrix $H$. In particular, the estimates depend solely
on the distribution of the singular values of $H$. Of course, in this
case, we also need change our ``fine recovery'' stage and use the
``$H$-dual'' of the corresponding frame $E$, which we define via
\begin{equation}\label{H-dual-formula}
F_{H} H = (HE)^\dagger.
\end{equation}
As an example, consider an $r$th order {\em high-pass} ${\Sigma\Delta}$ scheme
whose noise shaping matrix is $H^r$ where $H$ is defined via
\begin{equation}\label{def-H}
H_{ij} := \left \{
\begin{array}{rl}
1, & \mbox{if $i=j$ or if $i=j+1$}, \cr
0, & \mbox{otherwise.}
\end{array}
\right.
\end{equation}
It is easy to check that the singular values of $H$ are identical to
those of $D$. It follows that all the results presented in this paper
are valid also if the compressed measurements are quantized via an an
$r$th order high-pass ${\Sigma\Delta}$ scheme, provided the reconstruction is
done using the $H^r$-duals instead of the $r$th order Sobolev
duals. Note that such a result for high-pass ${\Sigma\Delta}$ schemes is not
known to hold in the case of structured frames.
\subsection{Measurement noise and compressible signals}
One of the natural questions is whether the quantization methods
developed in this paper are effective in the presence of measurement
noise in addition to the error introduced during the quantization
process. Another natural question is how to extend
this theory to include the case when the
underlying signals are not necessarily strictly sparse, but
nevertheless still ``compressible''.
Suppose $x\in \mathbb{R}^N$ is not sparse, but compressible in the usual sense
(e.g. as in \cite{CRT}), and let $y=\Phi x+e$, where $e$ stands for
additive measurement noise. The {\em coarse
recovery stage} inherits the stability and robustness properties
of $\ell_1$ decoding for
compressed sensing, therefore the accuracy of this first reconstruction
depends on the best $k$-term approximation error for $x$,
and the deviation of $\Phi x$ from the quantized signal $q$ (which
comprises of the measurement noise $e$ and the quantization error
$y - q$). Up to constant factors, the quantization error
for any (stable) ${\Sigma\Delta}$ quantizer is comparable to that of PCM,
hence the reconstruction error at the coarse recovery stage would also
be comparable. In the {\em fine recovery stage}, however, the
difference between
$\sigma_{\max}(F_{H} H)$ and $\sigma_{\max}(F_{H})$ plays a critical
role. In the particular case of $H=D^r$ and $F_H=F_{\text{sob},r}$,
the Sobolev duals we use in the
reconstruction are tailored to reduce the effect of the quantization error
introduced by an $r$th order ${\Sigma\Delta}$ quantizer. This is
reflected in the fact that as $\lambda$ increases, the kernel of
the reconstruction operator $F_{\text{sob},r}$
contains a larger portion of high-pass sequences (like
the quantization error of $\Sigma\Delta$ modulation), and is quantified by the
bound $\sigma_{\max}(F_{\text{sob},r}D^r) \lesssim
\lambda^{-(r-1/2)}m^{-1/2}$ (see
Theorem~\ref{main_thm_1}, \eqref{sobolev-formula} and
\eqref{err-bound-Sob}). Consequently, obtaining more measurements
increases $\lambda$, and even though $\|y - q\|_2$ increases as well,
the reconstruction error due to quantization decreases.
At the same time, obtaining more measurements would also increase
the size of the external noise $e$, as well as the ``aliasing error''
that is the result of the ``off-support'' entries of $x$. However,
this noise+error term is not counteracted by the
action of $F_{\text{sob},r}$. In fact,
for any dual $F$, the relation $ F E = I$ implies
$\sigma_{\max}(F) \geq 1/\sigma_{\max}(E) \gtrsim
m^{-1/2}$ already and in the case of measurement noise,
it is not possible to do better than the canonical dual $E^\dagger$ on average.
In this case, depending on the size of the noise term,
the fine recovery stage may not improve
the total reconstruction error even though the ``quantizer error'' is still
reduced.
One possible remedy for this problem is to construct alternative
quantization schemes with associated noise-shaping matrices that
balance the above discussed trade-off between the quantization error
and the error that is introduced by other factors. This is a delicate
procedure, and it will be investigated thoroughly in future
work. However, a first such construction can be made by using
``leaky'' ${\Sigma\Delta}$ schemes with $H$ given by
\begin{equation}\label{def-lH}
H_{ij} := \left \{
\begin{array}{rl}
1, & \mbox{if $i=j$}, \cr
-\mu & \mbox{if $i=j+1$}, \cr
0, & \mbox{otherwise,}
\end{array}
\right.
\end{equation}
where $\mu \in (0,1)$. Our preliminary numerical experiments
(see Figures \ref{fig6} and \ref{fig7}) suggest
that this approach can be used to improve the accuracy of the
approximation further in the fine recovery stage in this more general
setting. We note that the parameter $\mu$ above can be adjusted based
on how compressible the signals of interest are and what the expected
noise level is.
\begin{figure}[b]
\begin{center}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_leaky_exp1_5}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_leaky_exp1_10}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{mean_s_leaky_exp1_20}}
\caption{\label{fig6} The average case performance of the proposed $\Sigma\Delta$ quantization and reconstruction schemes (with general duals) for various values of $k$.
For this experiment the non-zero entries of $x$ are constant and $\delta=0.01$.}
\end{center}
\end{figure}
\begin{figure}[b]
\begin{center}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_leaky_exp1_5}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_leaky_exp1_10}}
\subfigure[]{
\includegraphics[width=3in,height=2in]{max_s_leaky_exp1_20}}
\caption{\label{fig7} The worst case performance of the proposed $\Sigma\Delta$ quantization and reconstruction schemes (with general duals) for various values of $k$.
For this experiment the non-zero entries of $x$ are constant and $\delta=0.01$.}
\end{center}
\end{figure}
\section*{Acknowledgments} The authors would like to thank Ronald
DeVore for valuable discussions. This work was initiated during an AIM
workshop and matured during a BIRS workshop. We thank the American
Institute of Mathematics and Banff International Research Station for
their hospitality. This work was supported in part by: National
Science Foundation Grant CCF-0515187 (G\"unt\"urk), Alfred P. Sloan
Research Fellowship (G\"unt\"urk), National Science Foundation Grant
DMS-0811086 (Powell), a Pacific Century Graduate Scholarship from the
Province of British Columbia through the Ministry of Advanced
Education (Saab), a UGF award from the UBC (Saab), and a Natural
Sciences and Engineering Research Council of Canada Discovery Grant
(Y{\i}lmaz).
|
\section{Introduction}
The algebraic functor $A^{{\mathrm{min}}}$ was introduced in~\cite{SW1}
arising from the question on the naturality of the classical
Poincar\'e-Birkhoff-Witt isomorphism. For any \textit{ungraded}
module $V$, $A^{{\mathrm{min}}}(V)$ is defined to be the \textit{smallest
functorial coalgebra retract} of $T(V)$ containing $V$. Then the
functor $A^{{\mathrm{min}}}$ extends canonically to the cases when $V$ is any
graded module. (See~\cite{SW1} for details.) The functor
$A^{{\mathrm{min}}}$ admits the tensor-length decomposition with
$$
A^{{\mathrm{min}}}(V)=\bigoplus_{n=0}^\infty A^{{\mathrm{min}}}_n(V),
$$
where $A^{{\mathrm{min}}}_n(V)=A^{{\mathrm{min}}}(V)\cap T_n(V)$ is the homogenous
component of $A^{{\mathrm{min}}}(V)$. By the Functorial
Poincar\'e-Birkhoff-Witt Theorem~\cite[Theorem 6.5]{SW1}, there
exists a functor $B^{\mathrm{max}}$ from (graded) modules to Hopf algebras
with the functorial coalgebra decomposition
$$
T(V)\cong A^{{\mathrm{min}}}(V)\otimes B^{\mathrm{max}}(V)
$$
for any graded module $V$. The determination of $A^{{\mathrm{min}}}(V)$ for
general $V$ is equivalent to an open problem in the modular
representation theory of the symmetric groups according
to~\cite[Theorem 7.4]{SW1}, which seems beyond the reach of current
techniques. Some properties of the algebraic functor $A^{{\mathrm{min}}}$ have been studied in~\cite{GW, Wu2} with applications in homotopy theory~\cite{GW2,Wu}. In this article we determine $A^{{\mathrm{min}}}(V)$ in the special cases
when $V_{\mathrm{even}}=0$ and $\mathrm{dim} V=p-1$.
Denote by $L(V)$ the free Lie algebra generated by $V$. Write $L_n(V)$ for the $n\,$th homogeneous
component of $L(V)$. Observe that $[L_s(V),L_t(V)]$ is a submodule
of $L_{s+t}(V)$ under the Lie bracket of $L(V)$. Let
$$
\bar L_n(V)=L_n(V)/\sum_{i=2}^{n-2} [L_i(V),L_{n-i}(V)].
$$
Define $\bar L_n^k(V)$ recursively by $\bar L_n^1(V)=\bar L_n(V)$
and $\bar L_n^{k+1}(V)=\bar L_n(\bar L_n^k(V))$.
\begin{thm}\label{theorem1.1}
Let the ground field $\ensuremath{{\mathbf{k}}}$ be of characteristic $p>2$. Let $V$ be a
graded module such that $V_{\mathrm{even}}=0$ and $\mathrm{dim} V=p-1$. Then there
is an isomorphism of coalgebras
$$
A^{{\mathrm{min}}}(V)\cong \bigotimes_{k=0}^\infty E(\bar L_p^k (V)),
$$
where $E(W)$ is the exterior algebra generated by $W$.
\end{thm}
An important observation is that
$$
\bar L_p^k(V)_{\mathrm{even}}=0 \textrm { and } \mathrm{dim} \bar L_p^k(V)=p-1
$$
for each $k\geq0$ provided that $V_{\mathrm{even}}=0$ and $\mathrm{dim} V=p-1$.
Replace $V$ by $\bar L_p(V)$ in the above theorem. Then $
A^{{\mathrm{min}}}(\bar L_p(V))\cong \bigotimes\limits_{k=1}^\infty E(\bar
L_p^k (V))$ with a coalgebra decomposition
$$
A^{{\mathrm{min}}}(V)\cong E(V)\otimes A^{{\mathrm{min}}}(\bar L_p(V)),
$$
which indicates the existence of the $\mathrm{EHP}$ fibrations by taking the
\textit{geometric realization} of the functor $A^{{\mathrm{min}}}$ and the
\textit{Hopf invariants} on the functor $A^{{\mathrm{min}}}$.
The geometric realization of the (algebraic) functor $A^{{\mathrm{min}}}$ was studied in~\cite{GSW,STW,SW1, SW2, Theriault} for giving the decompositions of the loop spaces of double suspensions of torsion spaces. The most general result so far is given as
follows:
\begin{thm}~\cite[Theorem 1.2]{STW2}\label{theorem1.2}
There exist homotopy functors $\bar A^{{\mathrm{min}}}$ and $\bar B^{\mathrm{max}}$
from simply connected co-$H$ spaces of finite type to spaces such
that for any $p$-local simply connected co-$H$ space $Y$ the
following hold:
\begin{enumerate}
\item[1)] $\bar A^{{\mathrm{min}}}(Y)$ is a functorial retract of $\Omega Y$
and so there is a functorial decomposition
$$
\Omega Y\simeq \bar A^{{\mathrm{min}}}(Y)\times \bar B^{\mathrm{max}}(Y).
$$
\item[2)] On mod $p$ homology the decomposition
$$ H_*(\Omega Y)\cong H_*(\bar A^{{\mathrm{min}}}(Y))\otimes H_*(\bar B^{\mathrm{max}}(Y))$$ is with
respect to the augmentation ideal filtration.
\item[3)]On mod $p$ homology the associated bigraded $E^0H_*(\bar
A^{{\mathrm{min}}}(Y))$ is given by
$$
E^0H_*(\bar A^{{\mathrm{min}}}(Y))=A^{{\mathrm{min}}}(\Sigma^{-1}\bar H_*(Y)).
$$
\end{enumerate}
\end{thm}
Let $X$ be a path-connected finite complex. Define $
b_X=\sum\limits_{q=1}^\infty q\mathrm{dim} \bar H_q(X;\mathbb{Z}/p).$ Roughly
speaking, $b_X$ is the summation of the dimensions of the cells in
$X$. A direct consequence of Theorems~\ref{theorem1.1} and
~\ref{theorem1.2} is the following homological information:
\begin{cor}\label{corollary1.3}
Let $p>2$ and let $Y$ be any $p$-local simply connected co-$H$
space. Suppose that $\bar H_{\mathrm{odd}}(Y)=0$ and $\mathrm{dim} \bar H_*(Y)=p-1$.
Then there is an isomorphism of coalgebras
$$
H_*(\bar A^{{\mathrm{min}}}(Y))\cong E(\Sigma^{-1}\bar H_*(Y))\otimes
\bigotimes_{k=1}^\infty E(\Sigma^{\frac{p^k-1}{p-1}b_Y-p^k}\bar
H_*(Y)),
$$
where $\Sigma^{\frac{p^k-1}{p-1}b_Y-p^k}\bar H_*(Y)=\bar
L_p^k(\Sigma^{-1}\bar H_*(Y))$.\hfill $\Box$
\end{cor}
Recall that the classical Hopf invariants can be obtained from the
suspension splitting of the loop suspensions. For the geometric
functor $\bar A^{{\mathrm{min}}}$, the following suspension splitting
theorem is a special case of Theorem~\ref{theorem4.8}.
\begin{thm}\label{theorem1.4}
Let $Y$ be any $p$-local simply connected co-$H$ space. Then there
is a suspension splitting
$$
\Sigma \bar A^{{\mathrm{min}}}(Y)\simeq \bigvee_{n=1}^\infty \bar
A^{{\mathrm{min}}}_n(Y)
$$
such that
$$
\Sigma^{-1}\bar H_*(\bar A^{{\mathrm{min}}}_n(Y))\cong
A^{{\mathrm{min}}}_n(\Sigma^{-1}\bar H_*(Y))
$$
for each $n\geq 1$.\hfill $\Box$
\end{thm}
Note that each $\bar A^{{\mathrm{min}}}_n(Y)$ is a co-$H$ space because it is
a retract of $\Sigma \bar A^{{\mathrm{min}}}(Y)$. From the above theorem, one
gets the Hopf invariant $H_n$ defined as the composite
$$
\bar A^{{\mathrm{min}}}(Y)\rInto \Omega\Sigma \bar A^{{\mathrm{min}}}(Y)\simeq \Omega
\bigvee_{n=1}^\infty \bar A^{{\mathrm{min}}}_n(Y)\rTo^{\mathrm{pro j}} \Omega \tilde
A^{{\mathrm{min}}}_n(Y)\rTo^r \bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_n(Y)),
$$
where $r\colon \Omega Z\to \bar A^{{\mathrm{min}}}(Z)$ is the functorial
retraction for any simply connected co-$H$-space $Z$. In particular,
there is a Hopf invariant $$ H_p\colon \bar A^{{\mathrm{min}}}(Y)\to \bar
A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))$$ for any simply connected co-$H$ space
$Y$. By Lemma~\ref{lemma5.3},
$$\bar A^{{\mathrm{min}}}_p(Y)\simeq \Sigma^{b_Y-p+1}Y$$ provided that $\bar
H_{\mathrm{odd}}(Y)=0$ and
$\mathrm{dim} \bar H_*(Y)=p-1$.
\begin{thm}[$\mathrm{EHP}$ Fibration]\label{theorem1.5}
Let $p>2$ and let $Y$ be any $p$-local simply connected co-$H$
space. Suppose that $\bar H_{\mathrm{odd}}(Y)=0$ and
$\mathrm{dim} \bar H_*(Y)=p-1$.
Then there is a fibre sequence
$$
\Omega \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)\rTo^P \bar E(Y)\rTo^E \bar
A^{{\mathrm{min}}}(Y)\rTo^{H_p}\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y).
$$
with the following properties:
\begin{enumerate}
\item[1)] On mod $p$ homology $H_*(\bar E(Y))\cong E(\Sigma^{-1}\bar H_*(Y))$
as coalgebras.
\item[2)] $H_*(\bar A^{{\mathrm{min}}}(Y))\cong H_*(\bar
E(Y))\otimes H_*(\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y))$ as coalgebras.
\item[3)] If $f\colon S^n\to Y$ is a co-$H$ map such that $f_*\not=0\colon
\bar H_*(S^n)\to \bar H_*(Y)$. Then there is a commutative diagram
of fibre sequences
\begin{diagram}
\Omega \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)&\rTo^P& \bar E(Y)&\rTo^E&
\bar A^{{\mathrm{min}}}(Y)&\rTo^{H_p}&\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)\\
\uEq&& \uTo &&\uTo&&\uEq\\
\Omega \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)&\rTo^{P_f}& S^{n-1}&\rTo&
B_f& \rTo& \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y).\\
\end{diagram}
In particular, the map $P\colon \Omega \bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)\to \bar E(Y)$ factors through the bottom
cell of $\bar E(Y)$.
\end{enumerate}
\end{thm}
The following theorem gives a general criterion when the $\mathrm{EHP}$
fibration splits off. A space $X$ is called to have a
\textit{retractile generating complex $C$} if $C$ is a retract of
$\Sigma X$ with a retraction $r\colon \Sigma X\to C$ such that the
mod $p$ cohomology $H^*(X)$ is generated by
$$
M=\mathrm{I m}(\Sigma^{-1}r^*\colon \Sigma^{-1}\bar H^*(C)\to \bar H^*(X))
$$
and $M\cong QH^*(X)$, the set of indecomposables. Recall that a
space $X$ is called \textit{(stably) atomic} if any self (stable)
map of $X$ inducing isomorphism on the bottom homology is a (stable)
homotopy equivalence.
\begin{thm}\label{theorem1.6}
Let $p>2$ and let $Y$ be any $p$-local simply connected co-$H$ space
such that $Y$ is stably atomic. Suppose that
$\bar H_{\mathrm{odd}}(Y)=0$ and $\mathrm{dim} \bar H_*(Y)=p-1$.
Then the following statements are equivalent to each other:
\begin{enumerate}
\item[1)] The $\mathrm{EHP}$ fibration
$$
\bar E(Y)\rTo^E \bar A^{{\mathrm{min}}}(Y)\rTo^{H_p}\bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)
$$
splits off.
\item[2)] $\bar E(Y)$ is an $H$-space.
\item[3)] There exists an $H$-space $X$ having $Y$ as a retractile
generating complex.
\item[4)] The map $P\colon \Omega \bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)\to \bar E(Y)$ is null homotopic.
\item[5)] The composite
$$
\Sigma^{b_Y-p-1}Y\rInto \Omega \bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)\rTo^P \bar E(Y)$$ is null homotopic.
\item[6)] There exists a map $g\colon \Sigma^{b_Y-p}Y\to \bar A^{{\mathrm{min}}}(Y)$ such that
$$
g_*\colon H_*(\Sigma^{b_Y-p}Y)\rTo H_*(\bar A^{{\mathrm{min}}}(Y))
$$
is a monomorphism on the bottom cells of $\Sigma^{b_Y-p}Y$.
\end{enumerate}
\end{thm}
It is a classical question whether the total space of a spherical fibration over a sphere is an $H$-space localized at an odd prime $p$. For having a possible $H$-space structure, the base space and the fibre must be odd dimensional spheres. For $p>3$, it is well-known that the answer is positive~\cite{CHZ,CN}. For $p=3$, there are examples that do not admit $H$-space structure. The above theorem gives some properties for studying this question.
The article is organized as follows. In section~\ref{section2}, we
study the representation theory on natural coalgebra decompositions
of tensor algebras. The proof of Theorem~\ref{theorem1.1} is given
in Section~\ref{section3}. The geometry of natural coalgebra
decompositions of tensor algebras is investigated in
Section~\ref{section4}, where Theorem~\ref{theorem1.4} is
Theorem~\ref{theorem4.8}. In Section~\ref{section5}, we give the
proof of Theorem~\ref{theorem1.5}. The proof of
Theorem~\ref{theorem1.6} is given in Section~\ref{section6}.
\section{The Functor $A^{{\mathrm{min}}}$ and the symmetric group module $\mathrm{Lie}(n)$}\label{section2}
\subsection{The Functor $A^{{\mathrm{min}}}$ on Ungraded Modules}\label{subsection2.1}
In this subsection, the ground ring is a field $\ensuremath{{\mathbf{k}}}$. A coalgebra
means a pointed coassociative cocommutative coalgebra. For any
module $V$, the tensor algebra $T(V)$ is a Hopf algebra by requiring
$V$ to be primitive. This defines $T\colon V\mapsto T(V)$ as a
functor from \textit{ungraded} modules to \textit{coalgebras}. The
functor $A^{{\mathrm{min}}}\colon V\mapsto A^{{\mathrm{min}}}(V)$ is defined to be the
smallest coalgebra retract of the functor $T$ with the property that
$V\subseteq A^{{\mathrm{min}}}(V)$. More precisely the functor $A^{{\mathrm{min}}}$ is
defined by the following property:
\begin{enumerate}
\item $A^{{\mathrm{min}}}$ is a functor from $\ensuremath{{\mathbf{k}}}$-modules to coalgebras with
a natural linear inclusion $V\rInto A^{{\mathrm{min}}}(V)$.
\item $A^{{\mathrm{min}}}(V)$ is a natural coalgebra retract of $T(V)$
over $V$. Namely there exist natural coalgebra transformations
$s\colon A^{{\mathrm{min}}}\to T$ and $r\colon T\to A^{{\mathrm{min}}}$ such that the
diagram
\begin{diagram}
A^{{\mathrm{min}}}(V)&\rTo^{s_V}& T(V)&\rTo^{r_V}&A^{{\mathrm{min}}}(V)\\
\uInto&&\uInto&&\uInto\\
V&\rEq& V&\rEq& V\\
\end{diagram}
commutes for any $V$ and $r\circ s=\mathrm{id}_{A^{{\mathrm{min}}}}$.
\item $A^{{\mathrm{min}}}$ is \textit{minimal} with respect to the above two conditions:
if $A$ is any functor from $\ensuremath{{\mathbf{k}}}$-modules to coalgebras with natural
linear inclusion $V\rInto A(V)$ such that $A(V)$ is a natural
coalgebra retract of $T(V)$ over $V$, then $A^{{\mathrm{min}}}(V)$ is a
natural coalgebra retract of $A(V)$ over $V$.
\end{enumerate}
By the minimal assumption, the functor $A^{{\mathrm{min}}}$ is unique up to
natural equivalence if it exists. The existence of the functor
$A^{{\mathrm{min}}}$ follows from~\cite[Theorem 4.12]{SW1}.
There is a multiplication on $A^{{\mathrm{min}}}$ given by the composite
$$
A^{{\mathrm{min}}}\otimes A^{{\mathrm{min}}}\rInto^{s\otimes s} T\otimes T\rTo^{\mu}
T\rOnto^r A^{{\mathrm{min}}}
$$
and so $A^{{\mathrm{min}}}$ is a functor from modules to quasi-Hopf algebras,
where a quasi-Hopf algebra means a coassociative and cocommutative
bi-algebra without assuming the associativity of the multiplication.
The multiplication on $A^{{\mathrm{min}}}$ induces a new natural coalgebra
transformation $r'_V\colon T(V)\to A^{{\mathrm{min}}}(V)$ given by
$$
r'_V(x_1\otimes\cdots\otimes x_n)=(((\cdots(x_1\cdot x_2)\cdot
x_3)\cdots)\cdot x_n)
$$
for any $x_1\otimes\cdots \otimes x_n\in V^{\otimes n}$. By the
minimal assumption, the composite
$$
A^{{\mathrm{min}}}\rTo^s T\rTo^{r'} A^{{\mathrm{min}}}
$$
is a natural equivalence. Consider $T(V)$ as an
$A^{{\mathrm{min}}}(V)$-comodule via the map $r'_V$. According
to~\cite[Proposition 6.1]{SW1}, the cotensor product
$$
B^{\mathrm{max}}(V)=\ensuremath{{\mathbf{k}}}\square_{A^{{\mathrm{min}}}(V)}T(V)
$$
is natural sub Hopf algebra of $T(V)$ with a natural coalgebra
equivalence
\begin{equation}
\ensuremath{{\mathbf{k}}}\otimes_{B^{\mathrm{max}}(V)}T(V)\cong A^{{\mathrm{min}}}(V).
\end{equation}
Together with~\cite[Lemmas 5.3]{SW1}, there is a natural coalgebra
decomposition
\begin{equation}
T(V)\cong B^{\mathrm{max}}(V)\otimes A^{{\mathrm{min}}}(V).
\end{equation}
By taking tensor length decomposition, we have $B^{\mathrm{max}}(V)=\bigoplus_{n=0}^\infty B^{\mathrm{max}}_n(V)$ with $B^{\mathrm{max}}_n(V)=B^{\mathrm{max}}(V)\cap T_n(V)$. Let
\begin{equation}
Q^{\mathrm{max}}(V)=QB^{\mathrm{max}}(V)
\end{equation}
be the indecomposables of $B^{\mathrm{max}}(V)$ with tensor length decomposition $Q^{\mathrm{max}}(V)=\bigoplus_{n=2}^\infty Q^{\mathrm{max}}_n(V)$. (Note that $B^{\mathrm{max}}_1(V)=0$ and so $B^{\mathrm{max}}(V)$ has no nontrivial indecomposable elements of tensor length $1$.) According to~\cite[Section 2]{SW1}, all of the functors $A^{{\mathrm{min}}}$, $A^{\mathrm{max}}_n$, $B^{\mathrm{max}}$, $Q^{\mathrm{max}}$ and $Q^{\mathrm{max}}_n$ extend canonically for graded modules.
\subsection{The Symmetric Group Module $\mathrm{Lie}(n)$}\label{subsection2.2}
Let the ground ring $R$ be any commutative ring with identity. The
module $\mathrm{Lie}^R(n)$, which is simply denoted as $\mathrm{Lie}(n)$ if the
ground ring is clear, is defined as follows. Let $\bar V$ be a free
$R$-module of rank $n$ with a basis $\{e_i\ | \ 1\leq i\leq n\}$.
The module $\gamma_n$ is defined to be the submodule of $\bar
V^{\otimes n}$ spanned by
$$
e_{\sigma(1)}\otimes e_{\sigma(2)}\otimes\cdots \otimes
e_{\sigma(n)}
$$
for $\sigma\in \Sigma_n$. The module $\mathrm{Lie}(n)$ is defined by
$$
\mathrm{Lie}(n)=\gamma_n\cap L_n(\bar V)\subseteq
\bar V^{\otimes n}.
$$
The $\Sigma_n$-action on $\gamma_n$ is given by permuting letters
$e_1,e_2,\ldots,e_n$. Since both $\gamma_n$ and $L_n(\bar V)$ are
invariant under permutations of the letters, $\mathrm{Lie}(n)$ is an
$R(\Sigma_n)$-submodule of $\gamma_n$.
Note that $\mathrm{Lie}(n)$ is the submodule of $L_n(\bar V)$ spanned by the
homogenous Lie elements in which each $e_i$ occurs exactly once. By
the Witt formula, $\mathrm{Lie}(n)$ is a free $R$-module of rank $(n-1)!$.
Following from the antisymmetry and the Jacobi identity, $\mathrm{Lie}(n)$
has a basis given by the elements
$$
[[e_1,e_{\sigma(2)}],e_{\sigma(3)}],\ldots,e_{\sigma(n)}]
$$
for $\sigma\in \Sigma_{n-1}$. (See ~\cite{Cohen2}.) Observe that $\mathrm{Lie}(n)$ is the image
of the $R(\Sigma_n)$-map
$$
\beta_n\colon \gamma_n\to\gamma_n \quad \beta_n(a_1\otimes\cdots
\otimes a_n)=[[a_1,a_2],\ldots,a_n].
$$
Thus $\mathrm{Lie}(n)$ can be also regarded as the quotient
$R(\Sigma_n)$-module of $\gamma_n$ with the projection
$\beta_n\colon \gamma_n\to \mathrm{Lie}(n)$.
\begin{prop}
Let $V$ be any graded projective module and let $\Sigma_n$ act on
$V^{\otimes n}$ by permuting factors in graded sense. Then there is
a functorial isomorphism
$$
\mathrm{Lie}(n)\otimes_{R(\Sigma_n)}V^{\otimes n}\cong L_n(V)
$$
for any graded module $V$.
\end{prop}
\begin{proof}
Clearly the quotient map $\beta_n\colon V^{\otimes
n}=\gamma_n\otimes_{R(\Sigma_n)}V^{\otimes n}\to L_n(V)$,
$x_1\otimes\cdots\otimes x_n\mapsto [[x_1,x_2],\ldots,x_n]$, factors
through the quotient $\mathrm{Lie}(n)\otimes_{R(\Sigma_n)}V^{\otimes n}$ and
so there is an epimorphism $\mathrm{Lie}(n)\otimes_{R(\Sigma_n)}V^{\otimes
n}\twoheadrightarrow L_n(V)$. On the other hand note that
$L=\bigoplus\limits_{n=1}^\infty
\mathrm{Lie}(n)\otimes_{R(\Sigma_n)}V^{\otimes n}$ has the canonical graded
Lie algebra structure generated by $V$. So the inclusion $V\rInto L$
induces an epimorphism of graded Lie algebras
$L(V)\twoheadrightarrow L$. The assertion follows.
\end{proof}
Consider $T_n\colon V\mapsto V^{\otimes n}$ as a functor from
projective (\textit{ungraded}) modules to projective
(\textit{ungraded}) modules. Denote by $\mathrm{H o m}(F,F')$ the set of
natural linear transformations from a functor $F$ to a functor $F'$
provided that $F$ preserves direct limits. By~\cite[Lemma 3.8]{GW},
$\mathrm{H o m}(T_n,T_m)=0$ if $n\not=m$ and there is an isomorphism of rings
\begin{equation}\label{theta}
\theta\colon \mathrm{End}_{R(\Sigma_n)}(\gamma_n)\rTo \mathrm{End}(T_n,T_n)
\end{equation}
given by
$$
\theta(\phi)=\phi\otimes_{\mathrm{id}_{V^{\otimes n}}}\colon
\gamma_n\otimes_{R(\Sigma_n)}V^{\otimes n}=V^{\otimes n}\rTo
\gamma_n\otimes_{R(\Sigma_n)}V^{\otimes n}=V^{\otimes n}
$$
for $\phi\in \mathrm{End}_{R(\Sigma_n)}(\gamma_n)$. Replacing $\gamma_n$
by $\mathrm{Lie}(n)$, we have the morphism of rings $ \theta\colon
\mathrm{End}_{R(\Sigma_n)}(\mathrm{Lie}(n))\rTo \mathrm{End}(L_n). $
\begin{prop}\label{proposition2.2}
If $n\not=m$, then $\mathrm{H o m}(L_n,L_m)=0$. Moreover the map $$
\theta\colon \mathrm{End}_{R(\Sigma_n)}(\mathrm{Lie}(n))\rTo \mathrm{End}(L_n)
$$
is an isomorphism.
\end{prop}
\begin{proof}
Let $\phi\colon L_n\to L_m$ be a natural transformation. Then the
composite $$T_n\rOnto L_n\rTo^\phi L_m\rInto T_m$$ is a natural
transformation, which is zero as $\mathrm{H o m}(T_n,T_m)=0$. Thus $\phi=0$.
For the second statement, clearly $\theta$ is a monomorphism. Let
$\phi_V\colon L_n(V)\to L_n(V)$ be any natural transformation. Let
$\bar V$ be the free $R$-module of rank $n$, which defines
$\gamma_n$. Consider the commutative diagram
\begin{diagram}
T_n(\bar V)&\rOnto^{\beta_n}& L_n(\bar V)&&\rTo^{\phi_{\bar V}}&& L_n(\bar V)&\rInto& T_n(\bar V)\\
\uInto&&\uInto&&&&\uInto&&\uInto\\
\gamma_n&\rOnto^{\beta_n}&\mathrm{Lie}(n)&&\rDashto^{\phi'}&&\mathrm{Lie}(n)&\rInto&\gamma_n,\\
\end{diagram}
where the existence of $\phi'$ follows from the fact that the
composite of the maps in the top row maps $\gamma_n$ into $\gamma_n$
and $\mathrm{Lie}(n)=\gamma_n\cap L_n(\bar V)$. Let $V$ be any ungraded
module and let $a_1\otimes\cdots \otimes a_n$ be any homogenous
element in $V^{\otimes n}$. Let $f\colon \bar V\to V$ be $R$-linear
map such that $f(e_i)=a_i$. By the naturality, there is a
commutative diagram
\begin{diagram}
\mathrm{Lie}(n)&\rInto & L_n(\bar V)&\rTo^{L_n(f)}& L_n(V)\\
\dTo>{\phi'}&&\dTo>{\phi_{\bar V}}&&\dTo>{\phi_V}\\
\mathrm{Lie}(n)&\rInto & L_n(\bar V)&\rTo^{L_n(f)}& L_n(V)\\
\end{diagram}
Thus
$$
\theta(\phi')([[a_1,a_2],\ldots,a_n])=\phi([[a_1,a_2],\ldots,a_n])
$$
and hence the result.
\end{proof}
\begin{cor}
There is a one-to-one correspondence between the decompositions of
the functor $L_n$ and the decompositions of $\mathrm{Lie}(n)$ over
$R(\Sigma_n)$.\hfill $\Box$
\end{cor}
A functor from modules to modules $Q$ is called
\textit{$T_n$-projective} if $Q$ is naturally equivalent to a
retract of the functor $T_n$.
\begin{prop}\label{projective}
Let $Q$ be a $T_n$-projective functor and let $\phi\colon Q\to L_n$
be a natural linear transformation. Then there exists a natural
linear transformation $\tilde\phi\colon Q\to T_n$ such that
$\phi=\beta_n\circ\tilde\phi$.
\end{prop}
\begin{proof}
Let $\bar V$ be the free $R$-module of rank $n$, which defines
$\gamma_n$. Let $\bar Q=\gamma_n\cap Q(\bar V)$. Since $Q$ is a
retract of the functor $T_n$, $\bar Q$ is a summand of $\gamma_n$
over $R(\Sigma_n)$ and so $\bar Q$ is a projective
$R(\Sigma_n)$-module. From the fact that
$\mathrm{End}_{R(\Sigma_n)}(\gamma_n)\cong \mathrm{End}(T_n)$,
$$
Q(V)=\bar Q\otimes_{R(\Sigma_n)}V^{\otimes n}
$$
for any module $V$. By the proof of Proposition~\ref{proposition2.2}, the map
$$
\theta\colon \mathrm{H o m}_{R(\Sigma_n)}(\bar Q,\mathrm{Lie}(n))\rTo \mathrm{H o m}(Q,L_n),
\quad f\mapsto f\otimes\mathrm{id}_{V^{\otimes n}}
$$
is an isomorphism. Thus there exists a unique $R(\Sigma_n)$-map
$$
\phi'\colon \bar Q\rTo \mathrm{Lie}(n)
$$
such that $\theta(\phi')=\phi$. Since $\bar Q$ is projective, the
lifting problem
\begin{diagram}
& &\gamma_n\\
&\ruDashto^{\tilde\phi'}&\dOnto>{\beta_n}\\
\bar Q&\rTo^{\phi'}&\mathrm{Lie}(n)\\
\end{diagram}
has a solution. The assertion follows by tensoring with $V^{\otimes
n}$ over $R(\Sigma_n)$.
\end{proof}
By inspecting the proof, each $T_n$-projective sub functor $Q$ of
$L_n$ induces a $R(\Sigma_n)$-projective submodule $\bar Q$ of
$\mathrm{Lie}(n)$. Conversely, each $R(\Sigma_n)$-projective submodule $\bar
Q$ of $\mathrm{Lie}(n)$ induces a $T_n$-projective sub functor $Q$, $
V\mapsto \bar Q\otimes_{R(\Sigma_n)}V^{\otimes n}$, of $L_n$. Thus
we have the following:
\begin{prop}\label{proposition2.5}
There is a one-to-one correspondence between $T_n$-projective sub
functors of $L_n$ and $R(\Sigma_n)$-projective submodules of
$\mathrm{Lie}(n)$.\hfill $\Box$
\end{prop}
\section{Proof of Theorem~\ref{theorem1.1}}\label{section3}
In this section, the ground field is of characteristic $p>2$. The
notation $V$ means any fixed connected graded module such that
$V_{\mathrm{even}}=0$ and $\mathrm{dim} V=p-1$. The general graded or ungraded module is then denoted by $W$.
Let $W$ be any module and let $T\colon W\mapsto T(W)$ be the functor from modules to Hopf algebras, where the tensor algebra $T(W)$ is Hopf by saying $W$ primitive.
Let $B(W)$ be the sub Hopf algebra generated
by $L_n(W)$ for $n$ not a power of $p$. By~\cite[Theorem 1.5]{SW1}, $B(W)\subseteq B^{\mathrm{max}}(W)$ for any
ungraded module $W$ and so for any graded or ungraded module $W$. It follows that there is an
epimorphism
\begin{equation}\label{equation3.1}
q\colon \ensuremath{{\mathbf{k}}}\otimes_{B(W)}T(W)\rOnto A^{{\mathrm{min}}}(W)
\end{equation}
for any ungraded or graded module $W$.
We are going to determine $\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V)$ and to show that
the map $q$ is in fact an isomorphism when $W=V$.
\subsection{Determination of $\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V)$}
Let $B^{[k]}(W)=\la B(W), L_{p^s}(W) \ | s\geq k\ra$ be the sub Hopf
algebra of $T(W)$ generated by $B(W)$ and $L_{p^s}(W)$ for $s\geq
k$. Then there is a tower of sub Hopf algebras
$$
\cdots \subseteq B^{[k+1]}(W)\subseteq B^{[k]}(W)\subseteq \cdots
\subseteq B^{[1]}(W)\subseteq B^{[0]}(W)=T(W)
$$
with the intersection
$$
B(W)=\bigcap_{k=0}^\infty B^{[k]}(W).
$$
Define the functor $\bar L_n$ by:
$$
\bar L_n(W)=L_n(W)/(\sum_{i=2}^{n-2}[L_i(W),L_{n-i}(W)])
$$
for each $n\geq 2$. According to~\cite[Proposition 11.6]{SW1}, $Q_{p}B^{[1]}(W)=\bar L_p$. So we can
identify $\bar L_p$ with a sub functor of $L_p$. Now define
recursively the functor $\bar L_p^k$ by $\bar L_p^0(W)=W$, $\bar
L_p^1(W)=\bar L_p(W)$ and $\bar L_p^{k+1}(W)=\bar L_p(\bar
L_p^k(W))$. Note that $\bar L_p^k$ has the tensor length $p^k$ and $\bar L_p^k$ can be identified with a sub functor of $L_{p^k}$.
Observe that $(\bar L_p(V))_{\mathrm{even}}=0$ and $\mathrm{dim} \bar L_p(V)=p-1$.
Thus we have the periodicity information:
\begin{equation}\label{equation3.3}
(\bar L_p^k(V))_{\mathrm{even}}=0 \textrm { and } \mathrm{dim} \bar L_p^k(V)=p-1
\end{equation}
for all $k\geq 0$.
\begin{lem}~\cite[Lemma 2.37]{Wu}\label{lemma3.1}
Let $A$ be a connected Hopf algebra of finite type and let $B$ be a
sub Hopf algebra of $A$. Suppose that $A$ is a tensor algebra as an
algebra with a choice of inclusion $QA\to A$. Then there is a short
exact sequence
$$
0\rTo Q(B)\rTo (k\otimes_BA)\otimes Q(A)\rTo I(k\otimes_BA)\rTo0.
$$
\end{lem}
Now we determine the sub Hopf algebra $B^{[k]}(V)$:
\begin{lem}\label{lemma3.2}
For each $k\geq 0$, there is a short exact sequence of Hopf algebras
$$
B^{[k+1]}(V)\rInto B^{[k]}(V)\rOnto E(\bar L_p^k(V)).
$$
Moreover
\begin{enumerate}
\item $B^{[k]}(V)$ is the sub Hopf algebra of $T(V)$ generated by $\bar
L_p^k(V)$ and $Q_jB(V)$ for $2\leq j<p^k$.
\item For any possible Steenrod module structure on $V$, $\bar L_p^k(V)$ is a suspension of
$V$.
\end{enumerate}
\end{lem}
\begin{proof}
The proof is given by induction. First consider the short exact
sequence of Hopf algebra
$$
B'(V)\rInto T(V)\rOnto \Lambda(V),
$$
where $B'(V)=\ensuremath{{\mathbf{k}}}\square_{\Lambda(V)}T(V)$. Then
$B^{[1]}(V)\subseteq B'(V)$. There is a short exact sequence
$$
0\rTo Q(B'(V))\rTo \Lambda(V) \otimes V\rTo I\Lambda(V)\rTo0.
$$
Since $V_{\mathrm{even}}=0$, $E(V)=\Lambda(V)$. By the assumption of $\mathrm{dim}
V=p-1$, we have $E(V)_j=0$ for $j\geq p$. Thus $Q_j(B'(V))=0$ for
$j>p$ and so $B'(V)\subseteq B^{[1]}(V)$. Hence $B'(V)=B^{[1]}(V)$
From the above exact sequence,
$$
Q_p(B'(V))\cong E_{p-1}(V)\otimes V\cong Q_p(B^{[1]}(V))=\bar L_p(V)
$$
is of dimension $p-1$. Since $\mathrm{dim} E_{p-1}(V)=1$, $\bar
L_p(V)=E_{p-1}(V)\otimes V$ is a suspension of $V$ for any possible
Steenrod module structure on $V$. Thus the assertions hold for
$k=0$. Suppose that the assertion holds for $k$. Consider the short
exact sequence of Hopf algebras
$$
B''(V)\rInto B^{[k]}(V)\rOnto^{\phi} \Lambda(\bar L_p^k(V)),
$$
where $B''(V)=\ensuremath{{\mathbf{k}}}\square_{\Lambda(\bar L_p^k(V))}B^{[k]}(V)$. Then
$B^{[k+1]}(V)\subseteq B''(V)$ because $\phi$ is a Hopf map which
sends the generators for $B^{[k+1]}(V)$ to zero. Since $\bar
L_p^k(V)_{\mathrm{even}}=0$ with $\mathrm{dim} \bar L_p^k(V)=p-1$, $\Lambda(\bar
L_p^k(V))=E(\bar L_p^k(V))$ and so there is short exact sequence
$$
0\rTo Q(B''(V))\rTo E(\bar L_p^k (V)) \otimes Q(B^{[k]}(V)\rTo
IE(\bar L_p^k(V))\rTo0.
$$
It follows that $Q(B''(V))_j=0$ for $j>p^{k+1}$ and so
$B''(V)\subseteq B^{[k+1]}(V)$. By the Lie action of $E(\bar
L_p^k(V))$ on $Q(B''(V))$,
$$
Q_{p^{k+1}}B^{[k+1]}(V)=Q_{p^{k+1}}B''(V)=\bar L_p^{k+1}(V)\cong
E_{p-1}(\bar L_p^k(V))\otimes \bar L_p^k(V).
$$
Since $\mathrm{dim} E_{p-1}(\bar L_p^k(V))=1$, $\bar L_p^{k+1}(V)$ is a
suspension of $\bar L_p^k(V)$. The induction is finished and hence
the result.
\end{proof}
\begin{prop}\label{proposition3.3}
There is a (non-functorial) coalgebra decomposition
$$
\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V)\cong \bigotimes_{k=0}^\infty E(\bar L_p^k(V))
$$
over any possible Steenrod algebra structure on $V$.
\end{prop}
\begin{proof}
Consider the short exact sequence of coalgebras
\begin{equation}\label{equation3.2}
\ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k+1]}(V)\rInto^j
\ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k]}(V)\rOnto E(\bar L_p^k(V)).
\end{equation}
Observe that up to tensor length $p^{k+1}-1$
$$
(\ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k]}(V))_j\cong E(\bar L_p^k(V))_j
$$
for $j\leq p^{k+1}$. Thus there is a coalgebra cross-section
$$
s\colon E(\bar L_p^k(V))\rInto \ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k]}(V).
$$
The cross-section $s$ is a morphism over the Steenrod algebra for
all possible Steenrod algebra structure on $V$ and it is also a
morphism over $\mathrm{G L}(V)$ when $\mathrm{G L}(V)$ acts on $V$ by forgetting the
grading of $V$.
According to~\cite[Theorem 1.1]{LW2}, $B(W)$ is a functorial coalgebra retract of $T(W)$
for any graded module $W$. Let $r_W\colon T(W)\to B(W)$ be a functorial coalgebra retraction.
From $B(W)\subseteq B^{[k]}(W)\subseteq T(W)$, the restriction
$$r_V|_{B^{[k]}(W)}\colon B^{[k]}(V)\rTo B(W)$$ is
a functorial coalgebra retraction. It follows that the short exact sequence of coalgebras
$$
B(W)\rInto B^{[k]}(W)\rOnto \ensuremath{{\mathbf{k}}}\otimes_{B(W)} B^{[k]}(W)
$$
splits off. This gives a functorial coalgebra decomposition
$$
B^{[k]}(W)\cong B(W)\otimes (\ensuremath{{\mathbf{k}}}\otimes_{B(W)} B^{[k]}(W))
$$
for any graded module $W$. Thus the functor
$$
W\mapsto \ensuremath{{\mathbf{k}}}\otimes_{B(W)}B^{[k]}(W)
$$
is a functorial from modules to quasi-Hopf algebra because $\ensuremath{{\mathbf{k}}}\otimes_{B(W)}B^{[k]}(W)$ is a functorial coalgebra retract of the Hopf algebra functor $B^{[k]}(W)$.
Evaluating $W=V$, $\ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k]}(V)$ is a quasi-Hopf
algebra over the Steenrod algebra for
all possible Steenrod algebra structure on $V$. By ~\cite[Lemma 5.3]{SW1}, the map
$$
(\ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k+1]}(V))\otimes E(\bar
L_p^k(V))\rTo^{\mu(j\otimes s)} \ensuremath{{\mathbf{k}}}\otimes_{B(V)}B^{[k]}(V)
$$
is an isomorphism and hence the result.
\end{proof}
\subsection{Proof of Theorem~\ref{theorem1.1}}
Consider the functorial short exact sequence of algebras
$$
B^{\mathrm{max}}(W)\rTo T(W)\rTo^r A^{{\mathrm{min}}}(W)=\ensuremath{{\mathbf{k}}}\otimes_{B^{\mathrm{max}}}T(W).
$$
Define
$$
A^{{\mathrm{min}}}_n(W)=\mathrm{I m}( T_n(W)\rInto T(W)\rTo^r A^{{\mathrm{min}}}(W).
$$
Then
$$
A^{{\mathrm{min}}}(W)=\bigoplus_{n=0}^\infty A^{{\mathrm{min}}}_n(W)
$$ is a bigraded coalgebra and the decomposition
$$
T(W)\cong B^{\mathrm{max}}(W)\otimes A^{{\mathrm{min}}}(W)
$$
is a functorial coalgebra decomposition of bi-graded coalgebras,
where the second grading on $B^{\mathrm{max}}(W)$ is given by
$B^{\mathrm{max}}_n(W)=T_n(W)\cap B^{\mathrm{max}}(W).$ Note that
$$
L_n(W)=\mathrm{I m}(\beta_n\colon W^{\otimes n}\to W^{\otimes n}).
$$
Denote by $\beta_n$ the epimorphism $W^{\otimes n}\twoheadrightarrow
L_n(W)$ if there are no confusions. The algebraic version of the
James-Hopf map $H_n\colon T(W)\rTo T(W^{\otimes n})$ is defined
in~\cite{GW,SW1,SW2}.
\begin{lem}\label{lemma3.4}
There is a commutative diagram of natural transformations of
functors
\begin{diagram}
& & T&\rTo^{H_{p^k}}& T(T_{p^k})& \rTo^{T(r)}& T(A^{{\mathrm{min}}}_{p^k})&\rOnto &A^{{\mathrm{min}}}(A^{{\mathrm{min}}}_{p^k})\\
&&\uInto&&\uInto&&\uInto &&\uTo\\
Q^{\mathrm{max}}_{p^{k+1}}&\rInto& L_{p^{k+1}}&\rTo& L_p(T_{p^k})&
\rTo^{L_p(r)}&
L_p(A^{{\mathrm{min}}}_{p^k})&\rOnto &\bar L_p(A^{{\mathrm{min}}}_{p^k})\\
\uEq&& \uOnto>{\beta_{p^k}}&&\uOnto>{\beta_p}&&\uOnto>{\beta_p}&&\\
Q^{\mathrm{max}}_{p^{k+1}}&\rTo^{\phi'}& T_{p^{k+1}}&\rTo^{\phi}&
T_p(T_{p^k})&\rTo^{T_p(r)}& T_p(A^{{\mathrm{min}}}_{p^k})&&\\
\end{diagram}
for some $\phi$ and $\phi'$.
\end{lem}
\begin{proof}
Since all functors in the diagram are well-defined over $\mathbb{Z}_{(p)}$,
it suffices to show that the assertion holds when the ground ring is
$\mathbb{Z}_{(p)}$. The top three squares commute because the maps in the
top row are coalgebra maps, which send the primitives to the
primitives. It is clear that the bottom right square commutes. By
Proposition~\ref{projective}, there exist lifting $\phi'$ and $\phi$
such that the bottom left and middle squares commute.
\end{proof}
\begin{lem}\label{lemma3.5}
Let $n\geq 2$ and let $W$ be any graded module. Then the composite
$$
W^{\otimes n}=W^{
\otimes 2}\otimes W^{\otimes n-2}\rOnto^{\beta_n} L_n(W)\rOnto \bar L_n(W)
$$
factors through the quotient
$$
\left(W^{\otimes 2}/\la a_1\otimes a_2+(-1)^{|a_1||a_2|}a_2\otimes
a_1\ra\right)\otimes \Lambda_{n-2}(W)
$$
of $W^{\otimes n}$.
\end{lem}
\begin{proof}
By the skew-symmetric property on the first two factors, the
iterated Lie operad $\beta_n\colon W^{\otimes n}\to L_n(W)$ factors
through the quotient
$$
\left(W^{\otimes 2}/\la a_1\otimes a_2+(-1)^{|a_1||a_2|}a_2\otimes
a_1\ra\right)\otimes W^{\otimes n-2}.
$$
Consider the short exact
sequence of Hopf algebras
$$
B^{[1]}(W)\rInto T(W)\rOnto \Lambda(W).
$$
Note that there is a commutative diagram
\begin{diagram}
\bigoplus_{n=2}^\infty L_n(W)&\rInto& IB^{[1]}(W)\\
\dOnto&&\dOnto\\
\bigoplus_{n=2}^\infty \bar L_n(W)&\rEq& QB^{[1]}(W).\\
\end{diagram}
The assertion follows from~\cite[Lemma 3.12]{CMN} that the Lie
bracket in $T(W)$ induces an action of $\Lambda (W)$ on
$QB^{[1]}(W)=\bigoplus_{n=2}^\infty \bar L_n(W)$.
\end{proof}
Let $W$ be a graded module such that $W_j=0$ for $j\not=1$. Let
$\Sigma_n$ act on $W^{\otimes n}$ by permuting positions in graded
sense. Let $\mathrm{G L}(W_1)$ be the general linear group of the vector
space $W_1$. The Schur algebra~\cite{Schur} is defined by
$$
\ensuremath{\mathcal{S}}(W)=\mathrm{I m}(\phi\colon \ensuremath{{\mathbf{k}}}(\mathrm{G L}(W))\to \mathrm{End}(W^{\otimes n})),
$$
where $\phi(f)=f^{\otimes n}.$
By~\cite[Theorem 4.1]{CR},
\begin{equation}\label{equation3.4}
\mathrm{I m}(\ensuremath{{\mathbf{k}}}(\Sigma_n)\to \mathrm{End} (W^{\otimes
n}))=\mathrm{End}_{\ensuremath{\mathcal{S}}(W)}(W^{\otimes n})
\end{equation}
if $\ensuremath{{\mathbf{k}}}$ is algebraically closed. Thus if $\ensuremath{{\mathbf{k}}}$ is algebraically
closed, then for any $\ensuremath{{\mathbf{k}}}(\mathrm{G L}(W))$-map $$f\colon W^{\otimes n}\to
W^{\otimes n}$$ there exists an element $\alpha\in \ensuremath{{\mathbf{k}}}(\Sigma_n)$
such that
$$
f(a_1\otimes\cdots\otimes a_n)=\alpha\cdot (a_1\otimes\cdots\otimes a_n).
$$
\begin{lem}\label{lemma3.6}
Let $\bar V$ be a graded module such that $\mathrm{dim}\bar V_1=p-1$ and
$\bar V_j=0$ for $j\not=1$. Let $\mathrm{G L}(\bar V)$ be the general linear
group with the action on $\bar V$. Assume that the ground field
$\ensuremath{{\mathbf{k}}}$ is algebraically closed. Then, as a morphism over
$\ensuremath{{\mathbf{k}}}(\mathrm{G L}(\bar V))$, the composite
$$
q\colon T_p(\bar V)=\bar V^{\otimes p}\rOnto^{\beta_p} L_p(\bar
V)\rOnto \bar L_p(\bar V)\rTo^{\cong} \bar V\otimes E_{p-1}(\bar V)
$$
does NOT have a cross-section.
\end{lem}
\begin{proof}
Suppose that there exists a $\ensuremath{{\mathbf{k}}}(\mathrm{G L}(\bar V))$-map $\phi\colon \bar
V\otimes E_{p-1}(\bar V)\rTo \bar V^{\otimes p}$ such that the
composite $q\circ\phi$ is the identity map of $\bar V\otimes
E_{p-1}(\bar V)$. We are going to find a contradiction.
Let $\{x_1,x_2,\ldots,x_{p-1}\}$ be a basis for $\bar V$. Note that
$$\mathrm{dim} E_{p-1}(\bar V)=1$$ with a basis $\{(x_1\wedge x_2\wedge \cdots \wedge x_{p-1})\}$.
The set
$$
\{x_i\otimes (x_1\wedge x_2\wedge \cdots\wedge x_{p-1}) \ | \ 1\leq
i\leq p-1\}
$$
forms a basis for $\bar V\otimes E_{p-1}(\bar V)$. Since the
composite
$$
\bar V^{\otimes p} \rOnto^q \bar V\otimes E_{p-1}(\bar V)\rTo^\phi
\bar V^{\otimes p}
$$
is an endomorphism over $\ensuremath{{\mathbf{k}}}(\mathrm{G L}(\bar V))$, from
equation~(\ref{equation3.4}) there exists an element $\alpha\in
\ensuremath{{\mathbf{k}}}(\Sigma_p)$ such that
\begin{equation}\label{equation3.5}
\phi\circ q(w)=\alpha\cdot w
\end{equation}
for $w\in \bar V^{\otimes p}$. By counting the occurrence of
$x_j$'s, $\phi(x_i\otimes (x_1\wedge \cdots \wedge x_{p-1}))$ is a
linear combination of monomials $x_{i_1}x_{i_2}\cdots x_{i_p}$ in
which $x_i$ occurs twice and $x_j$ occurs once for each $j\not=i$.
By Lemma~\ref{lemma3.5}, there is a commutative diagram
\begin{diagram}
\bar V^{\otimes p}&\rOnto^{\beta_p}&&L_p(\bar V)\\
\dOnto>{\pi}&&&\dOnto>{\bar\pi}\\
\left( \bar V^{\otimes 2}/\la x_ix_j-x_jx_i\ra\right)\otimes
E_{p-2}(\bar V)&\rTo^g& &\bar L_p(\bar V)=\bar V\otimes E_{p-1}(\bar V).\\
\end{diagram}
By counting the occurrence of $x_j$'s,
\begin{equation}\label{equation3.6}
\begin{array}{c}
\pi\circ\phi(x_1\otimes (x_1\wedge x_2\wedge \cdots \wedge
x_{p-1}))\\
=k_1x_1^2\otimes (x_2\wedge x_3\wedge \cdots\wedge x_{p-1})\\
+\sum_{i=2}^{p-1}k_ix_1x_i\otimes (x_1\wedge\cdots\wedge
x_{i-1}\wedge x_{i+1}\wedge\cdots \wedge x_{p-1}).\\
\end{array}
\end{equation}
for some $k_i\in \ensuremath{{\mathbf{k}}}$. Note that $x_i\wedge x_j=-x_j\wedge x_i$. By
interchanging $x_i$ and $x_{i+1}$, there is an equation
$$
k_i=-k_{i+1}
$$
for $i\geq 2$. By equation~(\ref{equation3.5}), the map
$$
\pi\circ\phi\colon \bar V\otimes E_{p-1}(\bar V)\rTo \left( \bar
V^{\otimes 2}/\la x_ix_j-x_jx_i\ra\right)\otimes E_{p-2}(\bar V)
$$
is a morphism over the Steenrod algebra by regarding $\bar V$ as
\textit{any graded module over the Steenrod algebra} with $\bar
V_{\textrm{even}}=0$. Take the Steenrod module structure on $\bar V$
by setting $P^1_*x_1=x_2$, $P^i_*x_1=0$ for $i>1$ and $P^k_*x_j=0$
for $k\geq 1$ and $j>1$. Then
$$
\begin{array}{l}
P^2_*x_1x_i\otimes (x_1\wedge\cdots\wedge x_{i-1}\wedge
x_{i+1}\wedge\cdots \wedge x_{p-1})\\ =x_2x_i\otimes
(x_2\wedge\cdots\wedge x_{i-1}\wedge x_{i+1}\wedge\cdots \wedge
x_{p-1})\\
=0\\
\end{array}
$$
for $i\geq3$.
Note that
$$
P^2_*x_1\otimes (x_1\wedge x_2\wedge \cdots \wedge
x_{p-1})=x_2\otimes (x_2\wedge x_2\wedge \cdots \wedge x_{p-1})=0.
$$
By applying $P^2_*$ to equation~(\ref{equation3.6}),
$$
\begin{array}{lll}
0&=&P^2_*(k_1x_1^2\otimes (x_2\wedge x_3\wedge \cdots\wedge
x_{p-1})+k_2x_1x_2\otimes (x_1\wedge x_3\wedge \cdots\wedge
x_{p-1}))\\
&=&k_1x_2^2\otimes (x_2\wedge x_3\wedge \cdots\wedge x_{p-1})+k_2
x_2^2 \otimes (x_2\wedge x_3\wedge \cdots\wedge x_{p-1}).\\
\end{array}
$$
Thus
$$
k_1=-k_2
$$
and so
$$
\pi\circ\phi(x_1\otimes (x_1\wedge x_2\wedge \cdots \wedge
x_{p-1}))= k_1 \sum_{i=1}^{p-1}(-1)^{i-1}x_1x_i\otimes
(x_1\wedge\cdots\wedge x_{i-1}\wedge x_{i+1}\wedge\cdots \wedge
x_{p-1}).
$$
Define
$$
\alpha_1= x_1\otimes x_1\otimes x_2\otimes x_3\otimes \cdots \otimes
x_{p-1}\in \bar V^{\otimes p}
$$
$$
\alpha_i= x_1\otimes x_i\otimes x_1\otimes \cdots \otimes
x_{i-1}\otimes x_{i+1}\otimes\cdots \otimes x_{p-1}\in \bar
V^{\otimes p}
$$
for $2\leq i\leq p-1$. Then
$$
\begin{array}{ll}
\bar\pi\beta_p(\alpha_1)&=\bar\pi([x_1,x_1],\ldots,x_{p-1}])\\
&=2x_1\otimes (x_1\wedge x_2\wedge x_3\wedge\cdots\wedge x_{p-1}).\\
\end{array}
$$
For $i\geq 2$,
$$
\begin{array}{ll}
\bar\pi\beta_p(\alpha_i)&=\bar\pi([x_1,x_i],x_1],\ldots,x_{i-1}],x_{i+1}],x_{p-1}])\\
&=(-1)^{i-1}x_1\otimes (x_1\wedge x_2\wedge x_3\wedge\cdots\wedge x_{p-1}).\\
\end{array}
$$
It follows that
$$
\begin{array}{l}
q\circ \phi(x_1\otimes (x_1\wedge x_2\wedge \cdots \wedge
x_{p-1}))\\
= g\circ \pi\circ \phi(x_1\otimes (x_1\wedge x_2\wedge \cdots \wedge
x_{p-1}))\\
=k_1 g(\sum\limits_{i=1}^{p-1}(-1)^{i-1}x_1x_i\otimes
(x_1\wedge\cdots\wedge x_{i-1}\wedge x_{i+1}\wedge\cdots \wedge
x_{p-1}))\\
=k_1 g(\sum\limits_{i=1}^{p-1}(-1)^{i-1}\pi(\alpha_i))\\
=k_1\sum\limits_{i=1}^{p-1}(-1)^{i-1}g\circ\pi(\alpha_i)\\
=k_1(\bar\pi\beta_p(\alpha_1)+\sum\limits_{i=2}^{p-1}(-1)^{i-1}\bar\pi\beta_p(\alpha_i)\\
=pk_1x_1\otimes (x_1\wedge x_2\wedge x_3\wedge\cdots\wedge
x_{p-1})\\
=0,\\
\end{array}
$$
which contradicts to the assumption that $q\circ \phi$ is the
identity map. This finishes the proof.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{theorem1.1}]
By Proposition~\ref{proposition3.3}, it suffices to show that the
epimorphism
$$
\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V)\rTo A^{{\mathrm{min}}}(V)
$$
is an isomorphism.
Since the functor $A^{{\mathrm{min}}}$ only depends on the characteristic of
the ground field, we may assume that the ground field $\ensuremath{{\mathbf{k}}}$ is
algebraically closed. Denote by $\phi_k$ the epimorphism $
\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V)\rTo A^{{\mathrm{min}}}(V) $. By
Proposition~\ref{proposition3.3},
$$
\bar L_p^k(V)=P_{p^k}(\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V))
$$
for each $k\geq 0$. By equation~(\ref{equation3.1}), $A^{{\mathrm{min}}}(V)$
is functorial coalgebra retract of $\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V)$. It
suffices to show by induction that the retraction map
$$
\phi_k|_{\bar L_p^k(V)}\colon \bar
L_p^k(V)=P_{p^k}(\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V))\to P_{p^k}A^{{\mathrm{min}}}(V)
$$
is an isomorphism for each $k\geq0$. The statement holds for $k=0$.
Suppose that the statement holds for $s\leq k$ with $k>0$ and
consider
$$
\phi_{k+1}|_{\bar L_p^{k+1}(V)}\colon \bar
L_p^{k+1}(V)=P_{p^{k+1}}(\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V))\to
P_{p^{k+1}}A^{{\mathrm{min}}}(V).
$$
By choosing $V$ to a module over the Steenrod algebra with a basis
given by $\{v,P^1_*v,P^2_*v,\ldots,P^{p-2}_*v\}$, then $V$ is an
indecomposable module over the Steenrod algebra. By
Lemma~\ref{lemma3.2}, $\bar L_p^{k+1}(V)$ is a suspension of $V$ as
a module over the Steenrod algebra. Thus $\bar L_p^{k+1}(V)$ is an
indecomposable module over the Steenrod algebra. It follows that the
retraction
$$
\phi_{k+1}|_{\bar L_p^{k+1}(V)}\colon \bar
L_p^{k+1}(V)=P_{p^{k+1}}(\ensuremath{{\mathbf{k}}}\otimes_{B(V)}T(V))\to
P_{p^{k+1}}A^{{\mathrm{min}}}(V)
$$
is either identically zero or an isomorphism.
Suppose that $\phi_{k+1}|_{\bar L_p^{k+1}(V)}=0$. Then
$$
P_{p^{k+1}}A^{{\mathrm{min}}}(V)=0
$$
and so
$$
L_{p^{k+1}}(V)\subseteq B^{\mathrm{max}}(V).
$$
By Lemma~\ref{lemma3.2}, $B^{[k+1]}(V)$ is the sub Hopf algebra of
$T(V)$ generated by $\bar L_p^{k+1}(V)$ and $L_n(V)$ for $n$ not a
power of $p$. It follows that
\begin{equation}\label{equation3.7}
B^{[k+1]}(V)\subseteq B^{\mathrm{max}}(V)
\end{equation}
and so
$$
PA^{{\mathrm{min}}}(V)\subseteq \bigoplus_{s=1}^k \bar L_p^s(V).
$$
From the induction. $\bar L^s_p(V)\subseteq A^{{\mathrm{min}}}(V)$ for each
$s\leq k$. Thus
$$
PA^{{\mathrm{min}}}(V)= \bigoplus_{s=1}^k \bar L_p^s(V).
$$
It follows that there is a (\textit{non-functorial}) isomorphism of
coalgebra
$$
A^{{\mathrm{min}}}(V)\cong \bigotimes_{s=0}^k E(\bar L_p^s (V)).
$$
By computing Poincar\'e series, the inequality~(\ref{equation3.7})
becomes the equality $$B^{[k+1]}(V)=B^{\mathrm{max}}(V).$$ Thus
$$
Q^{\mathrm{max}}_{p^{k+1}}(V)=Q_{p^{k+1}}B^{\mathrm{max}}(V)=Q_{p^{k+1}}B^{[k+1]}(V)\cong
\bar L_p^{k+1}(V).
$$
Observe that the inclusion
$$
\bar L_p^k(V)=P_{p^k}A^{{\mathrm{min}}}(V)\rInto A^{{\mathrm{min}}}_{p^k}(V)
$$
is an isomorphism and $\bar L_p^{k+1}(V)=\bar L_p(\bar L_p^k(V))$.
The composite of natural transformations
\begin{equation}\label{equation3.8}
Q^{\mathrm{max}}_{p^{k+1}}\rInto L_{p^{k+1}}\rTo L_p(T_{p^k}) \rTo^{L_p(r)}
L_p(A^{{\mathrm{min}}}_{p^k}) \rOnto \bar L_p(A^{{\mathrm{min}}}_{p^k})
\end{equation}
becomes an isomorphism by evaluating on the graded module $V$. Write
$\bar V$ for $ \bar L_p^k(V)=P_{p^k}A^{{\mathrm{min}}}(V)= A^{{\mathrm{min}}}_{p^k}(V).$
By Lemma~\ref{lemma3.4}, the composite of the first three natural
transformations in equation~(\ref{equation3.8}) admits a lifting of
natural transformation into $T_p(A^{{\mathrm{min}}}_{p^k})$ via the
epimorphism $\beta_p\colon T_p(A^{{\mathrm{min}}}_{p^k})\rOnto
L_p(A^{{\mathrm{min}}}_{p^k})$ and so the composite
$$
q\colon T_p(\bar V)=\bar V^{\otimes p}\rOnto^{\beta_p} L_p(\bar
V)\rOnto \bar L_p(\bar V)\rTo^{\cong} \bar V\otimes E_{p-1}(\bar V)
$$
admits a cross-section over the general linear group algebra
$\ensuremath{{\mathbf{k}}}(\mathrm{G L}(\bar V))$ by forgetting the grading of $\bar V$. This is
impossible according to Lemma~\ref{lemma3.6}.
Thus $\phi_{k+1}|_{\bar L_p^{k+1}(V)}\not=0$ and so
$\phi_{k+1}|_{\bar L_p^{k+1}(V)}$ must be an isomorphism. The
induction is finished and hence the result.
\end{proof}
\section{The Geometry of Natural Coalgebra Decompositions}\label{section4}
\subsection{Geometric Realizations}
Denote by $\mathbf{CoH}$ the category of $p$-local simply connected
co-$H$-spaces of finite type and co-$H$-maps. Let
$[\Omega,\Omega]_{\mathbf{CoH}}$ be the group of natural transformations of
the functor $\Omega$ from $\mathbf{CoH}$ to the homotopy category of spaces.
\begin{thm}[Geometric Realization Theorem]~\cite{STW2}\label{theorem4.1}
Let $Y$ be any simply connected co-$H$-space of finite type and let
$$
T(V)\cong A(V)\otimes B(V)
$$
any natural coalgebra decomposition for ungraded modules over
$\mathbb{Z}/p$. Then there exist homotopy functors $\bar A$ and $\bar B$
from $\mathbf{CoH}$ to spaces such that
\begin{enumerate}
\item there is a functorial decomposition
$$
\Omega Y\simeq \bar A(Y)\times \bar B(Y)
$$
\item On mod $p$ homology the decomposition $$H_*(\Omega Y)\cong H_*(\bar
A(Y))\otimes H_*(\bar B(Y))$$ is with respect to the augmentation
ideal filtration
\item On mod $p$ homology
$$
E^0H_*(\bar A(Y))=A(\Sigma^{-1}\bar H_*(Y)) \textrm{ and }
E^0_*(\bar B(Y))=B(\Sigma^{-1}\bar H_*(Y)),
$$
where $A$ and $B$ are the canonical extensions of the functors $A$
and $B$ for graded modules\hfill $\Box$
\end{enumerate}
\end{thm}
For the functors $A^{{\mathrm{min}}}$ and $B^{\mathrm{max}}$, we have the following geometric realization theorem.
\begin{thm}~\cite{STW2}\label{theorem4.2}
There exist homotopy functors $\bar A^{{\mathrm{min}}}$, $\bar Q^{\mathrm{max}}_n$, $n\geq2$,
from $\mathbf{CoH}$ to spaces such that for any $p$-local simply connected
co-$H$ space $Y$ of finite type the following hold:
\begin{enumerate}
\item[1)] $\bar B^{\mathrm{max}}(Y)=\Omega\left(\bigvee_{n=2}^\infty \bar Q^{\mathrm{max}}_n(Y)\right)$.
\item[2)] There is a functorial decomposition
$$
\Omega Y\simeq \bar A^{{\mathrm{min}}}(Y)\times \bar B^{\mathrm{max}}(Y).
$$
\item[3)] On mod $p$ homology the decomposition
$$ H_*(\Omega Y)\cong H_*(\bar A^{{\mathrm{min}}}(Y))\otimes H_*(\bar B^{\mathrm{max}}(Y))$$ is with
respect to the augmentation ideal filtration.
\item[4)]On mod $p$ homology
$$
\begin{array}{c}
E^0H_*(\bar A^{{\mathrm{min}}}(Y))=A^{{\mathrm{min}}}(\Sigma^{-1}\bar H_*(Y)),\\
E^0H_*(\bar B^{\mathrm{max}}(Y))=B^{\mathrm{max}}(\Sigma^{-1}\bar H_*(Y)),\\
E^0\bar H_*(\bar Q^{\mathrm{max}}(Y))=Q^{\mathrm{max}}(\Sigma^{-1}\bar H_*(Y)).\\
\end{array}
$$\hfill $\Box$
\end{enumerate}
\end{thm}
\subsection{Suspension Splitting Theorems}
In this subsection, we review the suspension splitting
theorems.
A \textit{graded space} means a space $W$ with a homotopy
decomposition
$$
\phi_W\colon W\rTo^{\simeq}\bigvee_{n=1}^\infty W_n.
$$
For any graded space $W$, the homology $\bar H_*(W)$ is filtered by
$$
I^t\bar H_*(W)=\phi_*^{-1}(\bar H_*(\bigvee_{n=t}^\infty W_n))
$$
for $t\geq 1$. A \textit{graded co-$H$ space} means a graded space
$W$ such that $W$ is a co-$H$ space. Thus each factor $W_n$ is also
a co-$H$ space. The following lemma gives a general criterion for
decomposing the retracts of graded co-$H$ spaces in term of grading
factors.
\begin{lem}~\cite{GTW}\label{lemma4.4}
Let $W$ be a simply connected $p$-local graded co-$H$ space of
finite type. Let $f\colon W\to W$ be a self-map such that on mod $p$
homology
\begin{enumerate}
\item[1)] $ f_*\colon \bar H_*(W)\to \bar H_*(W)$
preserves the filtration.
\item[2)] The induced bigraded map $E^0f_*$ is an idempotent:
$$E^0f_*\circ E^0f_*=E^0f_*\colon E^0H_*(W)\to E^0H_*(W).$$
\end{enumerate}
Let $A(f)=\operatorname{ho co l i m}_f W$ be the homotopy colimit and let
$A_n(f)=\operatorname{ho co l i m}_{g_n}W_n$, where $g_n$ is the composite
$$
g_n\colon W_n\rInto \bigvee_{k=1}^\infty W_k\rTo^{\phi_W^{-1}}
W\rTo^f W\rTo^{\phi_W} \bigvee_{k=1}^\infty W_k \rOnto W_n.
$$
Then there is a canonical homotopy decomposition of the homotopy
colimit
$$
A(f)\simeq \bigvee_{n=1}^\infty A_n(f)
$$
such that
$$
\bar H_*(A_n(f))\cong \mathrm{I m}(E^0_nf_*\colon E^0H_*(W)\cong \bar
H_*(W_n)\to E^0H_*(W)\cong \bar H_*(W_n)).
$$
\hfill $\Box$
\end{lem}
Let $X$ be any path-connected space. Let $H_*(\Omega \Sigma X)$ be
filtered by the powers of the augmentation ideal filtration. From
the classical suspension splitting Theorem~\cite{James1}
$$
\phi\colon \Sigma \Omega\Sigma X\rTo^{\simeq}
\bigvee_{n=1}^\infty\Sigma X^{(n)},
$$
$\Sigma\Omega\Sigma X$ is a simply graded co-$H$ space and the
filtration
$$
I^t\bar H_*(\Sigma \Omega\Sigma X)=\phi_*^{-1}(\bar
H_*(\bigvee_{n=t}^\infty \Sigma X^{(n)}))
$$
coincide with (the suspension) of the augmentation ideal filtration
of $H_*(\Omega\Sigma X)$.
Let $W$ and $W'$ be graded spaces. Then $W\wedge W'$ is a graded
space with homotopy equivalence
$$
W\wedge W'\rTo^{\phi_W\wedge
\phi_{W'}}_{\simeq}(\bigvee_{n=1}^\infty W_n)\wedge
(\bigvee_{n=1}^\infty W'_n)=\bigvee_{n=1}^\infty \bigvee_{i=1}^{n-1}
W_i\wedge W'_{n-i}
$$
and the homology $\bar H_*(W\wedge W')=\bar H_*(W)\otimes \bar
H_*(W')$ is the tensor product of filtered modules. Moreover if $W$
is a graded co-$H$ space, then $W\wedge W'$ is a graded co-$H$
space.
Recall that~\cite{SW1} any natural coalgebra retract $A(V)$ of $T(V)$ for
ungraded modules admits the tensor length decomposition
$$
A(W)=\bigoplus_{n=0}^\infty A_n(W)
$$
for any graded or ungraded module $W$.
\begin{thm}[Suspension Splitting Theorem~\cite{GTW}]\label{theorem4.8}
Let $A(V)$ be any natural coalgebra retract of $T(V)$ for any
ungraded modules $V$ and let $\bar A$ be the geometric realization of
$A$.Then for any $p$-local simply connected co-$H$ space $Y$ of
finite type and any $p$-local path-connected co-$H$-space $Z$, there
is a functorial splitting
$$
Z\wedge \bar A(Y)\simeq \bigvee_{n=1}^\infty [Z\wedge \bar A (Y)]_n
$$
such that
$$
\bar H_*([Z\wedge \bar A(Y)]_n)\cong \bar H_*(Z)\otimes
A_{n}(\Sigma^{-1}\bar H_*(Y))
$$
for each $n\geq 1$. In particular, for a $p$-local simply connected
co-$H$-space $Y$ there is a functorial suspension splitting
$$
\Sigma \bar A(Y)\simeq \bigvee_{n=1}^\infty \bar A_n(Y)
$$
such that
$$
\Sigma^{-1}\bar H_*(\bar A_n(Y))\cong A_n(\Sigma^{-1}\bar H_*(Y))
$$
for each $n\geq 1$.\hfill $\Box$
\end{thm}
\subsection{Hopf Invariants}
In this section, we review the results in Hopf invariants from~\cite{GTW}.
Let $A(V)$ be a natural coalgebra retract of $T(V)$ for ungraded
modules. From the suspension splitting theorem
(Theorem~\ref{theorem4.8}), there is a decomposition
$$
\Sigma \bar A(Y)\simeq \bigvee_{n=1}^\infty \bar A_n(Y)
$$
and so it induces Hopf invariants
$$
H_n\colon \bar A(Y)\rTo \Omega (\bar A_n(Y)).
$$
For computational purpose on homology, it is useful to make a
particular choice of Hopf invariants $H_n$. Let
$$
\ensuremath{\mathcal{H}}\colon \Sigma\Omega\Sigma X\simeq \Sigma J(X)\rTo^{\simeq}
\bigvee_{n=1}^\infty \Sigma X^{(n)}
$$
be the fat James-Hopf invariants. Let $H_*(\Omega\Sigma X)$ be
filtered by the products of the augmentation ideal and let
$H_*(\bigvee_{n=1}^\infty X^{(n)})$ be filtered by
$$
\bigoplus_{t\geq n} H_*(X^{(t)}).
$$
By~\cite[Proposition 3.7]{SW2}, the isomorphism
$$
\ensuremath{\mathcal{H}}_*\colon H_*(\Sigma\Omega\Sigma X)\rTo H_*(\bigvee_{n=1}^\infty \Sigma X^{(n)})
$$
preserves the filtration. Note that the composite
$$
\Sigma\Omega\Sigma X\rTo^{\ensuremath{\mathcal{H}}} \bigvee_{n=1}^\infty \Sigma
X^{(n)}\rTo^{\mathrm{proj}} \Sigma X^{(n)}
$$
is the James-Hopf invariant $H_n$. Let $Y\in\mathbf{CoH}$ be a simply
connected $p$-local co-$H$-space of finite type with the
cross-section map $s_{\mu'}\colon Y\rTo \Sigma \Omega Y$. Let $f$ be
the composite
$$
\Sigma \Omega \Sigma \Omega Y\rTo^{\Sigma\Omega \sigma} \Sigma\Omega
Y\rTo^{\Sigma \Omega s_{\mu'}}\Sigma \Omega \Sigma \Omega Y.
$$
By Lemma~\ref{lemma4.4}, the space $[\Sigma\Omega Y]_n$ is defined
to be the homotopy colimit of the self map $g_n(Y)$ given by the
composite
$$
\begin{array}{ccccc}
\Sigma (\Omega Y)^{(n)} &\rInto& \bigvee_{k=1}^\infty \Sigma (\Omega
Y)^{(k)}&\rTo^{\ensuremath{\mathcal{H}}^{-1}}&\Sigma \Omega \Sigma \Omega
Y\\
&\rTo^{\Sigma\Omega (s_{\mu'}\circ \sigma)} &\Sigma \Omega \Sigma
\Omega Y &\rTo^{\ensuremath{\mathcal{H}}}& \bigvee_{k=1}^\infty \Sigma (\Omega
Y)^{(k)}\\
&\rOnto^{\mathrm{proj}}& \Sigma (\Omega Y)^{(n)}.&&\\
\end{array}
$$
\begin{prop}\label{GTW}
For any $Y\in \mathbf{CoH}$, the composite
$$
\begin{array}{ccccc}
\ensuremath{\mathcal{H}}^Y\colon \Sigma\Omega Y &\rTo^{\Sigma\Omega s_{\mu'}}
&\Sigma\Omega \Sigma\Omega Y &\rTo^{\ensuremath{\mathcal{H}}}_{\cong}&
\bigvee_{n=1}^\infty \Sigma (\Omega
Y)^{(n)}\\
&\rTo& \bigvee_{n=1}^\infty \operatorname{ho co l i m}_{g_n}\Sigma (\Omega Y)^{(n)}\\
\end{array}
$$
is a homotopy equivalence.\hfill $\Box$
\end{prop}
Now define the $n\,$th Hopf invariant $H_n^Y$ to be the adjoint map
to the composite
\begin{equation}\label{equation4.21}
\Sigma \Omega Y \rTo^{\ensuremath{\mathcal{H}}^Y} \bigvee_{n=1}^\infty
\operatorname{ho co l i m}_{g_n}\Sigma (\Omega Y)^{(n)}\rTo^{\mathrm{proj}}
\operatorname{ho co l i m}_{g_n}\Sigma (\Omega Y)^{(n)}=[\Sigma\Omega Y]_n.
\end{equation}
\begin{thm}~\cite{GTW}\label{theorem4.11}
For any $Y\in\mathbf{CoH}$, there is a commutative diagram
\begin{diagram}
E^0H_*(\Omega Y)&\rTo^{E^0H^Y_{n\ast}}& E^0H_*(\Omega [\Sigma\Omega
Y]_n)\\
\uTo>{\cong}&&\uTo>{\cong}\\
T(\Sigma^{-1}\bar H_*(Y))&\rTo^{H_n}& T((\Sigma^{-1}\bar H_*(Y))^{\otimes n}),\\
\end{diagram}
where $H_n\colon T(V)\to T(V^{\otimes n})$ is the algebraic
James-Hopf map.\hfill $\Box$
\end{thm}
\section{Proof of Theorem~\ref{theorem1.5}}\label{section5}
\subsection{Proof of Theorem~\ref{theorem1.5}}
Now we give the proof of Theorem~\ref{theorem1.5}. All spaces are
localized at $p>2$. For a natural coalgebra retract $A$ of $T$, let
$\bar A$ be its geometric realization. According to the suspension
splitting theorem~\ref{theorem4.8}, there is a decomposition
$$
\Sigma \bar A(Y)\simeq \bigvee_{n=1}^\infty \bar A_n(Y).
$$
In particular, we have notations $\bar A^{{\mathrm{min}}}$ and $\bar
A^{{\mathrm{min}}}_n$. Let $X$ be a path-connected finite complex. Define $$
b_X=\sum\limits_{q=1}^\infty q\mathrm{dim} \bar H_q(X;\mathbb{Z}/p).$$
\begin{lem}\label{lemma5.1}
Let $Y$ be a simply connected co-$H$-space such that $\bar
H_{\mathrm{odd}}(Y)=0$ and $\mathrm{dim} \bar H_*(Y)=p-1$. Then
$$\bar
A^{{\mathrm{min}}}_{p^k-1}(Y)\simeq S^{(b_Y-p+1)\frac{p^k-1}{p-1}+1}.$$
\end{lem}
\begin{proof}
Let $V=\Sigma^{-1}\bar H_*(Y)$. Then $V_{\mathrm{even}}=0$ and $\mathrm{dim} V=p-1$.
By Theorem~\ref{theorem1.1},
$$
A^{{\mathrm{min}}}(V)=\bigotimes_{k=0}^\infty E(\bar L_p^k(V)).
$$
By considering tensor length $p^k-1$, we have
$$
A^{{\mathrm{min}}}_{p^k-1}(V)=E_{p-1}(V)\otimes E_{p-1}(\bar L_p(V))\otimes
\cdots\otimes E_{p-1}(\bar L_p^{k-1}(V)),
$$
which is a one-dimensional module. The assertion follows from
Theorem~\ref{theorem4.8}.
\end{proof}
Let $\bar Q^{\mathrm{max}}_n$ the geometric
realization of the indecomposables of $B^{\mathrm{max}}(V)$ as given in Theorem~\ref{theorem4.2}. Note that $Y$ is a retract of $\Sigma \bar
A^{{\mathrm{min}}}(Y)$ by Theorem~\ref{theorem4.8}.
\begin{lem}\label{lemma5.2}
Let $Y$ be any $p$-local simply connected co-$H$ space of finite
type. Then
\begin{enumerate}
\item there is a splitting cofibre sequence
$$
\bar Q^{\mathrm{max}}(Y)\longrightarrow \bar A^{{\mathrm{min}}}(Y)\wedge Y
\longrightarrow (\Sigma \bar A^{{\mathrm{min}}}(Y))/Y.
$$
Thus there is a decomposition $ Y\wedge \bar A^{{\mathrm{min}}}(Y) \simeq \bar
Q^{\mathrm{max}}(Y)\vee (\Sigma \bar A^{{\mathrm{min}}}(Y))/Y. $
\item There is a decomposition
$$
[Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n\simeq \tilde Q^{\mathrm{max}}_n(Y)\vee \tilde
A^{{\mathrm{min}}}_n(Y).
$$
\item If $Y=\Sigma X$, then
$$
[Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n\simeq X\wedge \bar A^{{\mathrm{min}}}_{n-1}(Y).
$$
\item If $\bar A^{{\mathrm{min}}}_{n-1}(Y)\simeq \Sigma Z$ for some $Z$,
then
$$
[Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n\simeq Y\wedge Z.
$$
\item $\bar A^{{\mathrm{min}}}_p(Y)\simeq [Y\wedge \bar
A^{{\mathrm{min}}}(Y)]_p$.
\end{enumerate}
\end{lem}
\begin{proof}
(1) and (2). Since $\bar B^{\mathrm{max}}(Y)\simeq \Omega\bar Q^{\mathrm{max}}(Y),$
there is a fibre sequence
$$
\Omega Y\rTo^{\partial} \bar A^{{\mathrm{min}}}(Y)\rTo^{\simeq\ast} \bar Q^{\mathrm{max}}(Y)\rTo Y.
$$
There is a (right) action of $\Omega Y$ on $\bar A^{{\mathrm{min}}}(Y)$ such
that the diagram
\begin{diagram}
\Omega Y\times \Omega Y&\rTo&\Omega Y\\
\dTo>{\partial\times \mathrm{id}_{\Omega Y}}&&\dTo>{\partial}\\
\bar A^{{\mathrm{min}}}(Y)\times \Omega Y&\rTo& \bar A^{{\mathrm{min}}}(Y)\\
\end{diagram}
commutes. Consider the commutative diagram
\begin{diagram}
& & & &\Sigma \Omega \bar Q^{\mathrm{max}}(Y) &\lInto& \bar
Q^{\mathrm{max}}(Y)\\
&&&&\dTo &&\\
(\Omega Y) \wedge Y &\rTo^{\mathrm{id}_{\Omega Y}\wedge s}& \Sigma\Omega
Y\wedge \Omega
Y&\rTo^{H}&\Sigma\Omega Y&\rTo& (\Sigma\Omega Y)/Y\\
\dTo&& \dTo &&\dTo&&\dTo\\
\bar A^{{\mathrm{min}}}(Y)\wedge Y&\rTo^{\mathrm{id}_{\bar A^{{\mathrm{min}}}(Y)}\wedge s}&
\bar A^{{\mathrm{min}}}(Y)\wedge \Sigma \Omega Y &\rTo^{H}& \Sigma \bar
A^{{\mathrm{min}}}(Y)&\rTo& (\Sigma
\bar A^{{\mathrm{min}}}(Y))/Y,\\
\end{diagram}
where $H$ is the Hopf construction. Observe that the composite of
the middle row is a homotopy equivalence. It follows that there is a
sequence
\begin{equation}\label{equation5.1}
\bar Q^{\mathrm{max}}(Y)\rTo^f \bar A^{{\mathrm{min}}}(Y)\wedge Y\rTo^g (\Sigma \bar
A^{{\mathrm{min}}}(Y))/Y
\end{equation}
such that $g\circ f\simeq\ast$, where $g$ is the composite of the
bottom rows and $f$ is the composite of the maps from top row down
to the middle right composing with the homotopy inverse from
$(\Sigma\Omega Y)/Y$ to $\Omega Y \wedge Y$ and the left column map.
Let $C_f$ be the homotopy cofibre of the map $f$. By
Lemma~\ref{lemma3.1}, the resulting map
$$
C_f\longrightarrow (\Sigma \bar A^{{\mathrm{min}}}(Y))/Y
$$
is a homotopy equivalence. Thus Sequence~(\ref{equation5.1}) is a
cofibre sequence. Let $r$ be the composite
\begin{equation}\label{equation5.2}
r\colon \bar A^{{\mathrm{min}}}(Y)\wedge Y\rInto^{s^{{\mathrm{min}}}\wedge \mathrm{id}_Y}
(\Omega Y)\wedge Y \rTo^{H\circ (\mathrm{id}_{\Omega Y}\wedge s)} \Sigma
\Omega Y \rOnto \Sigma\Omega \tilde Q^{\mathrm{max}}(Y)\rTo^{\sigma} \bar
Q^{\mathrm{max}}(Y).
\end{equation}
By the proof of ~\cite[Lemma 2.37]{Wu}, on homology
$$
(r\circ f)_*\colon \bar H_*(\tilde Q^{\mathrm{max}}(Y))\rTo \bar H_*(\bar
Q^{\mathrm{max}}(Y))
$$
is an isomorphism. Thus $f\colon \bar Q^{\mathrm{max}}(Y)\to \bar
A^{{\mathrm{min}}}(Y)\wedge Y$ admits a retraction and hence assertion (1).
Let $H_*(\Omega Y)$ be filtered by the products of the augmentation
ideal. Then $H_*(\bar A^{{\mathrm{min}}}(Y))$ and $H_*(\bar Q^{\mathrm{max}}(Y)$ have
the induced filtration. Observe that the maps $f,g$ and $r$ induces
filtration preserving maps on homology. Thus the decomposition
$$
Y\wedge \bar A^{{\mathrm{min}}}(Y) \simeq \bar Q^{\mathrm{max}}(Y)\vee (\Sigma \bar
A^{{\mathrm{min}}}(Y))/Y
$$
induces a graded decomposition and hence assertion (2).
(3) By Theorem~\ref{theorem4.8}, $$\Sigma \bar A^{{\mathrm{min}}}(Y)\simeq
\bigvee_{k=1}^\infty \bar A^{{\mathrm{min}}}_k(Y).$$ Let $s\colon \bar
A^{{\mathrm{min}}}_{n-1}(Y)\rInto \Sigma\bar A^{{\mathrm{min}}}(Y)$ be the inclusion.
Then composite
$$
\begin{array}{lcl}
X\wedge \bar A^{{\mathrm{min}}}_{n-1}(Y)&\rInto^{\mathrm{id}_X\wedge s}& X\wedge
\Sigma \bar A^{{\mathrm{min}}}(Y)\\
&\simeq &Y\wedge \bar A^{{\mathrm{min}}}(Y)\\
&\simeq &\bigvee\limits_{k=2}^\infty[Y\wedge \bar
A^{{\mathrm{min}}}(Y)]_k\\
&\rOnto^{\mathrm{pro j}.}& [Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n\\
\end{array}
$$
induces an isomorphism on homology and hence assertion (3).
(4) Let $s\colon Y\to \Sigma\Omega Y$ be a cross-section to the
evaluation map $\sigma \colon \Sigma\Omega Y\to Y$. Let $f$ be the
idempotent
$$
\Sigma \Omega Y\wedge \bar A^{{\mathrm{min}}}(Y)\rTo^{\sigma\wedge \mathrm{id}_{\bar
A^{{\mathrm{min}}}(Y)}} Y\wedge \bar A^{{\mathrm{min}}}(Y)\rInto^{s\wedge \mathrm{id}_{\bar
A^{{\mathrm{min}}}(Y)}} \Sigma\Omega Y\wedge \bar A^{{\mathrm{min}}}(Y)
$$
and let $g_n$ be the composite
$$
\begin{array}{lcl}
\Omega Y \wedge \bar A^{{\mathrm{min}}}_{n-1}(Y)&\simeq &[\Sigma \Omega Y
\wedge \bar A^{{\mathrm{min}}}(Y)]_n\\
&\rInto& \Sigma \Omega Y\wedge \bar A^{{\mathrm{min}}}(Y)\\
&\rTo^f &\Sigma \Omega Y\wedge \bar A^{{\mathrm{min}}}(Y)\\
&\simeq &\Omega Y\wedge \Sigma \bar A^{{\mathrm{min}}}(Y)\\
&\rOnto^{\mathrm{pro j}.}& \Omega Y \wedge \bar A^{{\mathrm{min}}}_{n-1}(Y),\\
\end{array}
$$
where the top homotopy equivalence follows from assertion (3) with
$X=\Omega Y$. According to Lemma~\ref{lemma4.4},
$$ [Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n\simeq \operatorname{ho co l i m}_{g_n}\Omega Y \wedge
\tilde A^{{\mathrm{min}}}_{n-1}(Y)$$ with the canonical retraction
$$
r\colon \Omega Y\wedge \bar A^{{\mathrm{min}}}_{n-1}(Y)\longrightarrow
[Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n.
$$
Now the composite
$$
\begin{array}{lcllcllcl}
Y\wedge Z&\rTo^{s\wedge\mathrm{id}_Z}&\Sigma\Omega Y \wedge Z &\simeq
&\Omega Y\wedge \bar A^{{\mathrm{min}}}_{n-1}(Y)
&\rTo^r &[Y\wedge \bar A^{{\mathrm{min}}}(Y)]_n\\
\end{array}
$$
induces an isomorphism on homology and so it is a homotopy
equivalence. Assertion (4) follows.
(5). According to~\cite[Section 11.2, p.97]{SW1}, $Q^{\mathrm{max}}_p(V)=0$.
Thus $$\bar H_*(\bar Q^{\mathrm{max}}_p(Y))=0$$ and so $\bar
Q^{\mathrm{max}}_p(Y)\simeq\ast$. By assertion (2),
$$\bar A^{{\mathrm{min}}}_p(Y)\simeq [Y\wedge\bar A^{{\mathrm{min}}}(Y)]_p$$ and hence
the result.
\end{proof}
\begin{lem}\label{lemma5.3}
Let $Y$ be a simply connected co-$H$-space such that $\bar
H_{\mathrm{odd}}(Y)=0$ and $\mathrm{dim} \bar H_*(Y)=p-1$. Then
$$
\bar A^{{\mathrm{min}}}_p(Y)\simeq \Sigma^{b_Y-p+1} Y.
$$
\end{lem}
\begin{proof}
By Lemma~\ref{lemma5.1}, $\bar A^{{\mathrm{min}}}_{p-1}=S^{b_Y-p+2}$. By
Lemma~\ref{lemma5.2} (4),
$$
[Y\wedge \bar A^{{\mathrm{min}}}(Y)]_p\simeq \Sigma^{b_Y-p+1}Y.
$$
The result then follows from Lemma~\ref{lemma5.2} (5).
\end{proof}
\begin{proof}[Proof of Theorem~\ref{theorem1.5}]
(1) and (2). Let
$$
H^Y_p\colon \Omega Y\rTo \Omega [\Sigma\Omega Y]_p
$$
be the James-Hopf invariant. By Lemma~\ref{lemma5.3},
$$\bar
A^{{\mathrm{min}}}_p(Y)\simeq \Sigma^{b_Y-p+1}Y.
$$
Define the map $H_p$ by
the composite
$$
\begin{array}{ccccc}
\bar A^{{\mathrm{min}}}(Y)&\rInto& \Omega Y&\rTo^{H^Y_p}& \Omega( [\Sigma
\Omega Y]_p)\\
&\rOnto& \Omega (\bar A^{{\mathrm{min}}}_p(Y))&\rOnto& \bar A^{{\mathrm{min}}}(\bar
A^{{\mathrm{min}}}_p(Y))=\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y).\\
\end{array}
$$
Let $\bar E(Y)$ be the homotopy fibre of the map $H_p$. Then there
is a fibre sequence
$$
\bar E(Y)\rTo \bar A^{{\mathrm{min}}}(Y)\rTo^{H_p} \bar A^{{\mathrm{min}}}(\bar
A^{{\mathrm{min}}}_p(Y)).
$$
We compute the homology of $\bar E(Y)$. Let $V=\Sigma^{-1}\bar
H_*(Y)$. According to Theorem~\ref{theorem4.11},
$$
E^0H_{p\ast}=H_p\colon T(V)=E^0H_*(\Omega Y)\rTo T(V^{\otimes
p})=E^0H_*(\Omega([\Sigma\Omega Y]_p))
$$
is the usual algebraic James-Hopf map. Thus there is a commutative
diagram
\begin{diagram}
T(V)&\rTo^{H_p}&T(V^{\otimes p})\\
\dOnto&&\dOnto\\
E^0H_*(\bar A^{{\mathrm{min}}}(Y))=A^{{\mathrm{min}}}(V)&\rTo^{E^0H_{p\ast}}&
E^0H_*(\bar
A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y)))\\
\end{diagram}
where $H_p$ is the algebraic James-Hopf map. By
Theorem~\ref{theorem1.1} and Lemma~\ref{lemma3.2}, the primitives of
$A^{{\mathrm{min}}}(V)$ is given by
$$
PA^{{\mathrm{min}}}(V)=\bigoplus_{k=0}^\infty \bar L^k_p(V).
$$
Let $\bar L$ be the sub Lie algebra of $L(V)$ generated by $L_p(V)$.
Note that
$$
\bar L_p^k(V)\subseteq \bar L
$$
for $k\geq 1$. By~\cite[Theorem 1.1]{Wu1}, the James-Hopf map
$$H_p|_{\bar L}\colon \bar L\rTo PT(V^{\otimes p})$$
is a morphism of Lie algebras. Since
$$H_{p}|_{L_p}\colon \bar L_p\to
V^{\otimes p}$$ is the canonical inclusion,
$$
H_p(\bar L_p^{k+1}(V))=\bar L_p^k(\bar L_p(V))
$$
for each $k\geq 0$. Note that $\bar L_p(V)\subseteq A^{{\mathrm{min}}}_p(V)$
and, since $\mathrm{dim} V=p-1$,
$$
\mathrm{dim} \bar L_p(V)=\mathrm{dim} A^{{\mathrm{min}}}_p(V) =p-1.
$$
Thus $\bar L_p(V)=A^{{\mathrm{min}}}_p(V)$. It follows that
\begin{equation}\label{equation5.3}
PE^0H_{p\ast}\colon PE^0H_*(\bar A^{{\mathrm{min}}}(Y))\rTo PE^0H_*(\bar
A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y)))
\end{equation}
is onto with the kernel $L^0_p(V)=V$.
For any graded Hopf algebra $A$, let $A^{\mathrm{a b}}$ be the abelianization
of $A$ and let
$$
\phi\colon A\rTo A^{\mathrm{a b}}
$$
be the quotient map. Define $\bar\phi$ to be the composite
$$
H_*(\bar A^{{\mathrm{min}}}(Y))\rInto H_*(\Omega Y)\rTo^{\phi} H_*(\Omega
Y)^{\mathrm{a b}}.
$$
Since $H_*(\Omega Y)$ is generated by odd dimensional elements,
$H_*(\Omega Y)^{\mathrm{a b}}$ is an exterior algebra. Moreover
$E^0H_*(\Omega Y)^{\mathrm{a b}}$ is primitively generated exterior algebra.
Thus
\begin{equation}\label{equation5.4}
E^0H_*(\Omega Y)^{\mathrm{a b}}=E(V).
\end{equation}
Consider the map
\begin{equation}\label{equation5.5}
\begin{array}{ccc}
\theta\colon H_*(\bar A^{{\mathrm{min}}}(Y))&\rTo^{\psi}& H_*(\bar
A^{{\mathrm{min}}}(Y))\otimes H_*(\bar A^{{\mathrm{min}}}(Y))\\
&\rTo^{\bar\phi\otimes H_{p\ast}}& H_*(\Omega Y)^{\mathrm{a b}}\otimes
H_*(\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))).\\
\end{array}
\end{equation}
From equations~(\ref{equation5.3}) and ~(\ref{equation5.4}),
$$
E^0P\theta\colon PE^0H_*(\bar A^{{\mathrm{min}}}(Y))\rTo PE^0(H_*(\Omega
Y)^{\mathrm{a b}}\otimes H_*(\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))))
$$
is an isomorphism. Thus
$$
E^0\theta\colon E^0H_*(\bar A^{{\mathrm{min}}}(Y))\rTo E^0(H_*(\Omega
Y)^{\mathrm{a b}}\otimes H_*(\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))))
$$
is a monomorphism. Since both sides have the same Poincar\'e series,
$E^0\theta$ is an isomorphism and so $\theta$ is an isomorphism.
Thus there is an algebra isomorphism on cohomology
$$
H^*(\bar A^{{\mathrm{min}}}(Y))\cong H^*(\bar A^{{\mathrm{min}}}(\bar
A^{{\mathrm{min}}}_p(Y)))\otimes (H_*(\Omega Y)^{\mathrm{a b}})^*.
$$
By the Eilenberg-Moore spectral sequence, there is an epimorphism
$$
H_*(\Omega Y)^{\mathrm{a b}}\rOnto H_*(\bar E(Y))
$$
and so the Poincar\'{e} series
$$
\chi(H_*(\bar E(Y)))\leq \frac{\chi(H_*(\bar
A^{{\mathrm{min}}}(Y)))}{\chi(H_*(\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))))}.
$$
By applying the Serre spectral sequence to the fibre sequence
$$
\bar E(Y)\rTo \bar A^{{\mathrm{min}}}(Y)\rTo \bar A^{{\mathrm{min}}}(\bar
A^{{\mathrm{min}}}_p(Y)),
$$
we have the $E^2$-terms given by $$H_*(\bar E(Y))\otimes H_*(\bar
A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y)))$$ and so
$$
\chi(H_*(\bar A^{{\mathrm{min}}}(Y)))\leq \chi(H_*(\bar E(Y))\chi(H_*(\bar
A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y)))).
$$
It follows that
$$
\chi(H_*(\bar E(Y)))=\frac{\chi(H_*(\bar
A^{{\mathrm{min}}}(Y)))}{\chi(H_*(\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))))}
$$
and so
$$
H_*(\Omega Y)^{\mathrm{a b}}\cong H_*(\bar E(Y)).
$$
This finishes the proof of assertion (1) and (2).
(3). Let $\bar Y$ be the homotopy cofibre of the map $f\colon S^n\to
Y$. Since $f$ is a co-$H$-map, $\bar Y$ is a co-$H$-space. By the
naturality of $\bar A^{{\mathrm{min}}}$, the pinch map $q\colon Y\to \bar Y$
induces a map
$$
\bar A^{{\mathrm{min}}}(Y)\rTo \bar A^{{\mathrm{min}}}(\bar Y).
$$
Let $V'=\Sigma^{-1}\mathrm{I m}(f_*\colon \bar H_*(S^n)\to \bar H_*(Y))$ and
let $\bar V=\Sigma^{-1}\bar H_*(\bar Y)\cong V/V'$. Since $\mathrm{dim} \bar
V=p-2$,
\begin{equation}\label{equation5.6}
E^0H_*(\bar A^{{\mathrm{min}}}(\bar Y))=E(\bar V)
\end{equation}
by~\cite[Corollary 11.6]{SW1}. From the proof of assertions (1) and
(2), the composite
\begin{equation}\label{equation5.7}
H_*(\bar E(Y))\rInto H_*(\bar A^{{\mathrm{min}}}(Y))\rInto H_*(\Omega Y)\rOnto
H_*(\Omega Y)^{\mathrm{a b}}
\end{equation}
is an isomorphism. Let $g$ be the composite
$$
\bar E(Y)\rTo \bar A^{{\mathrm{min}}}(Y)\rTo \bar A^{{\mathrm{min}}}(\bar Y).
$$
Consider the commutative diagram
\begin{equation}\label{equation5.8}
\begin{diagram}
H_*(\bar E(Y))&\rInto& H_*(\bar A^{{\mathrm{min}}}(Y))& \rInto &H_*(\Omega Y)&
\rOnto& H_*(\Omega Y)^{\mathrm{a b}}\\
&\rdTo>{g_*}&\dTo&&\dOnto&&\dOnto\\
& & H_*(\bar A^{{\mathrm{min}}}(\bar Y))&\rInto& H_*(\Omega \bar Y)&\rOnto & H_*(\Omega\bar Y)^{\mathrm{a b}},\\
\end{diagram}
\end{equation}
where the composites in the top and bottom rows are isomorphisms.
Thus $$g_*\colon H_*(\bar E(Y))\to H_*(\bar A^{{\mathrm{min}}}(Y))$$ is onto.
By~\cite[Proposition 4.20]{MM}, both $H_*(\Omega Y)^{\mathrm{a b}}$ and
$H_*(\Omega \bar Y)^{\mathrm{a b}}$ are primitively generated because they
are commutative with trivial restricted maps. Thus
$$
H_*(\Omega \bar Y)^{\mathrm{a b}}\cong E(\bar V) \textrm{ and } H_*(\Omega
Y)^{\mathrm{a b}}\cong E(V)\cong E(\bar V)\otimes E(V')
$$
as Hopf algebras. From diagram~\ref{equation5.8} the map $g_*$
induces a coalgebra decomposition
$$
H_*(\bar E(Y))\cong H_*(\bar A^{{\mathrm{min}}}(\bar Y))\otimes E(V').
$$
It follows that the homotopy fibre of the map $g\colon \bar E(Y)\to
\bar A^{{\mathrm{min}}}(\bar Y)$ is $S^{n-1}$ by using the Eilenberg-Moore
spectral sequence. Thus there is a homotopy commutative diagram of
fibre sequences
\begin{diagram}
\Omega\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))&\rTo^{P_f}& S^{n-1}&\rTo&
B_f&
\rTo&\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))\\
\dEq&&\dTo&&\dTo&&\dEq\\
\Omega\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))&\rTo^{P}& \bar
E(Y)&\rTo^{E}& \bar A^{{\mathrm{min}}}(Y)&
\rTo^{H_p} &\bar A^{{\mathrm{min}}}(\bar A^{{\mathrm{min}}}_p(Y))\\
& &\dTo&&\dTo&&\\
& &\bar A^{{\mathrm{min}}}(\bar Y)&\rEq&\bar A^{{\mathrm{min}}}(\bar Y)& &\\
\end{diagram}
and hence the result.
\end{proof}
\subsection{Self-maps of $\bar A^{{\mathrm{min}}}(Y)$}
Let $Y$ be a simply connected co-$H$-space of $(p-1)$-cell complex
with the cells in even dimensions. In this case, the integral
homology of $Y$ is torsion free and so we can take homology $H_*(Y)$
over $p$-local integers. Let $f$ be any self-map of $\bar
A^{{\mathrm{min}}}(Y)$. Let $V=\Sigma^{-1}\bar H_*(Y)$ with a basis
$\{x_{l_1},\ldots,x_{l_{p-1}}\}$, where the degree $|x_{l_i}|=l_i$
with $l_1\leq\cdots \leq l_{p-1}$. By Theorem~\ref{theorem1.5},
there is a coalgebra isomorphism
$$
H_*(\bar A^{{\mathrm{min}}}(Y))\cong \bigotimes_{k=0}^\infty H_*(\bar E(\bar
A^{{\mathrm{min}}}_{p^k}(Y)))\cong \bigotimes_{k=0}^\infty
E(A^{{\mathrm{min}}}_{p^k}(V))\cong A^{{\mathrm{min}}}(V).
$$
Thus the primitives
$$
PH_*(\bar A^{{\mathrm{min}}}(Y))=\bigoplus_{k=0}^\infty A^{{\mathrm{min}}}_{p^k}(V)
$$
and so on cohomology
$$
QH^*(\bar A^{{\mathrm{min}}}(Y))=\bigoplus_{k=0}^\infty A^{{\mathrm{min}}}_{p^k}(V)^*.
$$
Note that $A^{{\mathrm{min}}}_{p^k}(V)$ has a basis in the form
$$
w_{p^k,i}=[(x_{l_1}\wedge\cdots\wedge
x_{l_{p-1}})^{\frac{p^k-1}{p-1}}x_{l_i}]
$$
for $1\leq i\leq p-1$. Let $f^*_{k,l}$ be the composite
$$
A^{{\mathrm{min}}}_{p^k}(V)^*\rInto QH^*(\bar A^{{\mathrm{min}}}(Y))\rTo^{f^*} QH^*(\bar
A^{{\mathrm{min}}}(Y))\rTo^{\mathrm{proj.}} A^{{\mathrm{min}}}_{p^l}(V)^*.
$$
Note that $|w_{p^k,p-1}|<|w_{p^{k'},1}|$ for $k<k'$. Thus
\begin{equation}\label{equation5.9}
f^*_{k,l}=0 \textrm{ for } k\not=l
\end{equation}
and so
\begin{equation}\label{equation5.10}
f^*(A^{{\mathrm{min}}}_{p^k}(V)^*)\subseteq A^{{\mathrm{min}}}_{p^k}(V)^*.
\end{equation}
It follows that
$$
f^*\colon A^{{\mathrm{min}}}(V)^*=H^*(\bar A^{{\mathrm{min}}}(Y))\rTo H^*(\bar
A^{{\mathrm{min}}}(Y))=A^{{\mathrm{min}}}(V)^*
$$
is a graded map with respect to the tensor length. Since $\mathrm{dim}
A^{{\mathrm{min}}}_{p^k-1}(V)=1$, let $\deg^k(f)$ be the degree of the map
$$
f^*\colon A^{{\mathrm{min}}}_{p^k-1}(V)^*\rTo A^{{\mathrm{min}}}_{p^k-1}(V)^*.
$$
\begin{prop}
Let $f\colon \bar A^{{\mathrm{min}}}(Y)\to \bar A^{{\mathrm{min}}}(Y)$ be any map. Then
$\deg^k(f)$ over $p$-local integers is invertible if and only if
$$
f_*\colon H_j(\bar A^{{\mathrm{min}}}(Y);\mathbb{Z}/p)\rTo H_j(\bar A^{{\mathrm{min}}}(Y);\mathbb{Z}/p)
$$
is an isomorphism for $j< (l_1+\cdots +l_{p-1})\cdot
\frac{p^k-1}{p-1}+l_1$.
\end{prop}
\begin{proof}
The assertion follows from that
$$
\deg^k(f)=\prod_{j=0}^{k-1} \det\left(f^*\colon
A^{{\mathrm{min}}}_{p^j}(V)^*\rTo A^{{\mathrm{min}}}_{p^j}(V)^*\right)
$$
for cohomology with coefficients over $p$-local integers.
\end{proof}
Since $\bar A^{{\mathrm{min}}}(Y)$ is an $H$-space, there is a product $f\ast
g$ for any self maps $f$ and $g$ of $\bar A^{{\mathrm{min}}}(Y)$.
\begin{prop}
Let $f$ and $g$ be any self maps of $\bar A^{{\mathrm{min}}}(Y)$. Then
$$
\deg^k(f\ast g)=\prod_{j=0}^{k-1}\det\left(f_*+g_*\colon
A^{{\mathrm{min}}}_{p^j}(V)\rTo A^{{\mathrm{min}}}_{p^j}(V)\right).
$$
\end{prop}
\begin{proof}
Note that
$$
(f\ast g)_*(w)=f_*(w)+g_*(w)
$$
for any primitive element $w$. The assertion follows.
\end{proof}
\section{Proof of Theorem~\ref{theorem1.6} }\label{section6}
\begin{lem}\label{lemma6.1}
Let $Y$ be a $p$-local simply connected co-$H$-space such that
$H_{\mathrm{odd}}(Y)=0$ and $\mathrm{dim} \bar H_*(Y)=p-1$.
\begin{enumerate}
\item If $Y$ admits a nontrivial decomposition, then the $\mathrm{EHP}$
fibration splits off.
\item If $Y$ is atomic, then $\bar E(Y)$ is also atomic.
\end{enumerate}
\end{lem}
\begin{proof}
(1). Suppose that $Y\simeq Y_1\vee Y_2$ be a nontrivial
decomposition. Since $\mathrm{dim} \bar H_*(Y_i)<p-1$, $H_*(\bar
A^{{\mathrm{min}}}(Y_i))=E(\Sigma^{-1}\bar H_*(Y_i))$. Let $p_i$ be the
composite
$$
\bar A^{{\mathrm{min}}}(Y)\rInto \Omega Y\rOnto^{\mathrm{proj.}}\Omega
Y_i\rOnto \bar A^{{\mathrm{min}}}(Y_i).
$$
Then the composite
$$
\bar E(Y)\rTo^E \bar A^{{\mathrm{min}}}(Y)\rTo^{(p_1,p_2)} \bar
A^{{\mathrm{min}}}(Y_1)\times \bar A^{{\mathrm{min}}}(Y_2)
$$
is a homotopy equivalence as it induces an isomorphism on homology.
Assertion (1) follows.
(2). Let $f\colon \bar E(Y)\to \bar E(Y)$ be any self map inducing an
isomorphism on the bottom cell. Let $V=\Sigma^{-1}\bar H_*(Y)$. By
Theorem~\ref{theorem1.5}, $H_*(\bar E(Y))=E(V)$ as coalgebras. Then
$$
Pf_*\colon V=PE(V)\rTo V=PE(V)
$$
is an isomorphism on the bottom cell. From the suspension splitting
theorem (Theorem~\ref{theorem4.8}),
$$
\Sigma \bar E(Y)\simeq \bigvee_{n=1}^{p-1}\bar A^{{\mathrm{min}}}_n(Y)
$$
with $\Sigma^{-1} H_*(\bar A^{{\mathrm{min}}}_n(Y)=A^{{\mathrm{min}}}(V)$. Now the
composite
$$
g\colon Y=\bar A^{{\mathrm{min}}}_1(Y)\rInto \Sigma\bar E(Y)\rTo^{\Sigma f}
\Sigma \bar E(Y)\simeq \bigvee_{n=1}^{p-1}\bar A^{{\mathrm{min}}}_n(Y)\rOnto
\bar A^{{\mathrm{min}}}_1(Y)=Y
$$
is a homotopy equivalence because on homology
$\Sigma^{-1}g_*=Pf_*\colon V\to V$ is an isomorphism on the bottom
cell. It follows that $Pf_*\colon V\to V$ is an isomorphism and so
$f_*\colon E(V)\to E(V)$ is an isomorphism. Thus $f$ is a homotopy
equivalence and hence the result.
\end{proof}
Let $H_*(\Omega Y)$ be filtered by the products of the augmentation
ideal. For spaces $X$ and $Z$ such that $H_*(X)$ and $H_*(Z)$ are
filtered, a map $f\colon X\to Z$ is called \textit{filtered} of
$f_*\colon H_*(X)\to H_*(Z)$ is a filtered map. A retract $X$ of
$\Omega Y$ is called \textit{filtered} if there is a filtered self
map $f$ of $\Omega Y$ such that $X\simeq \operatorname{ho co l i m}_f\Omega Y$. Note
that for each filtered retract $X$ of $\Omega Y$ there is a
filtration on $H_*(X)$ induced from the filtration of $H_*(\Omega
Y)$. Consider all possible filtered retracts of $\Omega Y$. Divide
them in two types: a filtered retract $X$ of $\Omega Y$ is called of
\textit{type $A$} if $X$ contains the bottom cells of $\Omega Y$,
otherwise it is called of \textit{type B}.
\begin{lem}\label{lemma6.2}
For any $p$-local simply connected co-$H$-space $Y$, there exists a
minimal filtered retract $X^{{\mathrm{min}}}(Y)$ of $\Omega Y$ of type $A$
such that: (1) $X^{{\mathrm{min}}}(Y)$ is a filtered retract of $\Omega Y$ of
type $A$ and (2) any filtered retract of $\Omega Y$ of type $A$ is a
filtered retract of $X(Y)$. Moreover there is a filtered
decomposition
$$
\Omega Y\simeq X^{{\mathrm{min}}}(Y)\times \Omega Q.
$$
\end{lem}
\begin{proof}
Let $\ensuremath{\mathcal{S}}$ be the set of all filtered retract of $\Omega Y$ of type
$A$. Define a partial order on $\ensuremath{\mathcal{S}}$ by setting $X\leq X'$ if $X$
is a filtered retract of $X'$. From $X_1,X_2\in \ensuremath{\mathcal{S}}$, let $f$ be
the composite
$$
\Omega Y\rOnto X_1\rInto \Omega Y\rOnto X_2\rInto \Omega Y.
$$
Let $X_3=\operatorname{ho co l i m}_f\Omega Y$. Then $X_3$ is common filtered retract of
$X_1$ and $X_2$. Namely $X_3\leq X_1$ and $X_3\leq X_2$. By Zorn
Lemma, $\ensuremath{\mathcal{S}}$ has the minimal element $X^{{\mathrm{min}}}(Y)$. Consider the
commutative diagram
\begin{diagram}
\Omega\Sigma X^{{\mathrm{min}}}(Y)&\rTo& X^{{\mathrm{min}}}(Y)&\rTo &\Sigma
X^{{\mathrm{min}}}(Y)\wedge X^{{\mathrm{min}}}(Y)&\rTo^{H}&\Sigma X^{{\mathrm{min}}}(Y)\\
\uTo>{\Omega j}&&\uEq&&\uTo&\textrm{pull}& \uTo>{j}\\
\Omega Y&\rTo^{\partial}& X^{{\mathrm{min}}}(Y)&\rTo& Q&\rTo& Y\\
\end{diagram}
where $j$ is the adjoint map of the retraction $\Omega Y\to
X^{{\mathrm{min}}}(Y)$. Then the composite
$$
X^{{\mathrm{min}}}(Y)\rTo \Omega Y\rTo^{\partial} X^{{\mathrm{min}}}(Y)
$$
is a homotopy equivalence by the minimal assumption of
$X^{{\mathrm{min}}}(Y)$. It follows that there is a filtered decomposition
$$
\Omega Y\simeq X^{{\mathrm{min}}}(Y)\times \Omega Q
$$
and hence the result.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{theorem1.6}]
\noindent $(1)\Longrightarrow (2).$ If $\mathrm{EHP}$ fibration splits off,
then $\bar E(Y)$ is a retract of $\bar A^{{\mathrm{min}}}(Y)$. Note that
$\bar A^{{\mathrm{min}}}(Y)$ is an $H$-space as it is a retract of $\Omega Y$.
Thus $\bar E(Y)$ is an $H$-space.
$(2)\Longrightarrow (3)$. From the suspension splitting theorem
(Theorem~\ref{theorem4.8}),
$$
\Sigma \bar A^{{\mathrm{min}}}(Y)\simeq \bigvee_{n=1}^\infty \bar
A^{{\mathrm{min}}}_n(Y)
$$
with $\bar A^{{\mathrm{min}}}_1(Y)=Y$ and $\Sigma^{-1}\bar H_*(\bar
A^{{\mathrm{min}}}_n(Y))=A^{{\mathrm{min}}}(\Sigma^{-1}\bar H_*(Y))$. The composite
$$
\Sigma \bar E(Y)\rTo \Sigma \bar A^{{\mathrm{min}}}(Y)\rTo^{\mathrm{proj.}}
\bigvee_{n=1}^{p-1}\bar A^{{\mathrm{min}}}_n(Y)
$$
is a homotopy equivalence because $H_*(\bar E(Y))\to H_*(\bar
A^{{\mathrm{min}}}(Y))$ is a monomorphism and the connectivity of
$\bigvee_{n=p}^\infty \bar A^{{\mathrm{min}}}_n(Y)$ is greater than the
dimension of $\Sigma \bar E(Y)$. Thus $Y$ is a retract of
$\Sigma\bar E(Y)$. By Theorem~\ref{theorem1.5}, $H^*(\bar
E(Y))=E(\Sigma^{-1}\bar H^*(Y))$ and so $\bar E(Y)$ is an $H$-space
with $Y$ as a retractile generating complex.
$(3)\Longrightarrow (1)$.
Let $X$ be any $H$-space having $Y$ as a retractile generating
complex. Since $Y$ is a retract of $X$. Let $s\colon Y\to \Sigma X$
be an inclusion and let $r\colon \Sigma X\to Y$ be the retraction
such that $H^*(X)$ is generated by $M=r^*(\Sigma^{-1}\bar H_*(Y))$
with $M\cong QH^*(X)$. Let $V=\Sigma^{-1}\bar H_*(Y)$. Since
$H_{\mathrm{odd}}(Y)=0$, $V_{\mathrm{even}}=0$. Note that $H_*(X)$ is a quasi-Hopf
algebra as $X$ is an $H$-space. Thus $H^*(X)$ is the exterior
algebra generated by $V^*$ by the Borel Theorem~\cite[Theorem
7.11]{MM}. There is a homotopy commutative diagram of fibre
sequences
\begin{equation}\label{equation6.1}
\begin{diagram}
\Omega\Sigma X&\rTo^{\tilde\partial}& X&\rTo&\Sigma X\wedge X&\rTo^{H}& \Sigma X\\
\uTo>{\Omega s}&&\uEq&&\uTo&\textrm{pull}&\uTo>{s}\\
\Omega Y&\rTo^{\partial}& X&\rTo& B&\rTo& Y,\\
\end{diagram}
\end{equation}
where $H\colon \Sigma X\wedge X\to \Sigma X$ is the Hopf fibration
for the $H$-space $X$. Let $r'\colon X\to \Omega Y$ be the adjoint
map of $r$. Since the dimension of $X$ less than the connectivity of
$\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}(Y))$, the composite
$$
X\rTo^{r'}\Omega Y\rOnto \bar A^{{\mathrm{min}}}(Y)
$$
lifts to the fibre $\bar E(Y)$ of the $\mathrm{EHP}$-fibration. Let
$r''\colon X\to \bar E(Y)$ be a lifting of the above composite.
Consider the composite
\begin{equation}\label{equation6.2}
\begin{diagram}
\phi\colon \Omega Y&\rTo^{\partial}& X&\rTo^{r''}&\bar E(Y)&\rTo^E&\bar A^{{\mathrm{min}}}(Y)&\rInto& \Omega Y\\
\uInto>{j}&&\uInto>{j}&&\uInto>{j}& &\uInto>{j}&&\uInto>{j}\\
\bigvee_{\alpha}S^n&\rEq&\bigvee_{\alpha}S^n&\rEq&
\bigvee_{\alpha}S^n&\rEq&\bigvee_{\alpha}S^n&\rEq&\bigvee_{\alpha}S^n,\\
\end{diagram}
\end{equation}
where $j$ is the inclusion of the bottom cells. Let
$$
\bar X=\operatorname{ho co l i m}_{\phi}\Omega Y.
$$
Then $\bar X$ is a common retract of $X$, $\bar E(Y)$, $\bar
A^{{\mathrm{min}}}(Y)$ and $\Omega Y$, which contains the bottom cell. Since
$Y$ is atomic, $\bar E(Y)$ is also atomic by Lemma~\ref{lemma6.1}.
Thus the retraction
$$
\bar E(Y)\rTo^E \bar A^{{\mathrm{min}}}(Y)\rInto \Omega Y\rTo
\operatorname{ho co l i m}_{\phi}\Omega Y=\bar X
$$
is a homotopy equivalence. It follows that the $\mathrm{EHP}$-sequence
$$
\bar E(Y)\rTo^E \bar A^{{\mathrm{min}}}(Y)\rTo^{H_p} \bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)
$$
admits a retraction $\bar A^{{\mathrm{min}}}(Y)\to \bar E(Y)$ and so the
$\mathrm{EHP}$ fibration splits off.
Now $(1)\Longrightarrow (4)$ and $(4)\Longrightarrow (5)$ are
obvious. $(5)\Longrightarrow (6)$ follows from the homotopy
commutative diagram
\begin{diagram}
\Omega \bar A^{min}(\Sigma^{b_Y-p+1}Y)&\rTo^P& \bar E(Y)&\rTo^E&
\bar A^{{\mathrm{min}}}(Y)&\rTo^{H_p}& \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)\\
\uInto&&\uTo&&\uTo>{g}&&\uInto\\
\Sigma^{b_Y-p-1}Y&\rTo& \ast&\rTo&\Sigma^{b_Y-p}Y
&\rEq&\Sigma^{b_Y-p}Y,\\
\end{diagram}
where the bottom row is the cofibre sequence.
$(6)\Longrightarrow (3)$. Consider the commutative diagram
\begin{diagram}
\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)&\rInto& \Omega \Sigma^{b_Y-p+1}Y
&\rTo^{\tilde g}&\Omega Y&\rOnto& \bar
A^{{\mathrm{min}}}(Y)\\
& \luInto&\uInto& & \uInto&\ruEq &\\
& &\Sigma^{b_Y-p}Y&\rTo^{g}&\bar A^{{\mathrm{min}}}(Y),& &\\
\end{diagram}
where $\tilde g$ is the $H$-map induced by $g$. Let
$\bigvee_{\alpha}S^q$ be the bottom cells of $\Sigma^{b_Y-p}Y$.
Observe that
\begin{equation}\label{equation6.3}
H_{p\ast}\colon
H_q(\bar A^{{\mathrm{min}}}(Y))\rTo^{\cong} H_q(\bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)).
\end{equation}
Thus there is a commutative diagram
\begin{equation}\label{equation6.4}
\begin{diagram}
\theta\colon \Omega Y & \rOnto & \bar A^{{\mathrm{min}}}(Y)&\rTo^{H_p}& \bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)&\rInto& \Omega \Sigma^{b_Y-p+1}Y
&\rTo^{\tilde g}&\Omega Y\\
\uTo& &\uTo>{g|_{\vee_{\alpha}S^q}}& &\uInto>{j} & &\uInto>{j}& & \uTo\\
\bigvee_{\alpha}S^q & \rEq &\bigvee_{\alpha}S^q&\rEq &
\bigvee_{\alpha}S^q& \rEq& \bigvee_{\alpha}S^q&\rEq& \bigvee_{\alpha}S^q,\\
\end{diagram}
\end{equation}
where $j$ is the inclusion of the bottom cells.
Let $H_*(\Omega Y)$ be filtered by the products of the augmentation
ideal. Then $H_*(\bar A(Y))$ has the induced filtration. Observe
that
$$
g_*(\bar H_*(\Sigma^{b_Y-p}Y))\subseteq I^pH_*(\bar
A^{{\mathrm{min}}}(Y))\subseteq I^pH_*(\Omega Y)
$$
because the top dimension of $H_*(\bar
E(Y))=\oplus_{n=1}^{p-1}A^{{\mathrm{min}}}(\Sigma^{-1}Y)$ is less than the
connectivity of $\Sigma^{b_Y-p}Y$. It follows that
$$
\tilde g_*\colon
H_*(\Omega\Sigma^{b_Y-p+1}Y)=H_*(J(\Sigma^{b_Y-p}Y))\rTo H_*(\Omega
Y)
$$
preserves the filtration. Thus $\theta$ is a
filtered map because the other factors of $\theta$ are
filtration-preserving maps on homology. Let
$$
Z=\operatorname{ho co l i m}_{\theta} \Omega Y
$$
be the homotopy colimit. Then $Z$ is a common filtered retract of
$\Omega Y$, $\bar A^{{\mathrm{min}}}(Y)$, $\bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)$,
and $\Omega \Sigma^{b_Y-p+1}Y$. By Lemma~\ref{lemma4.4}, $\Sigma Z$
is a common graded retract of $\Sigma \Omega Y$, $\Sigma \bar
A^{{\mathrm{min}}}(Y)$, $\Sigma \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)$, and
$\Sigma\Omega \Sigma^{b_Y-p+1}Y$. Thus $[\Sigma Z]_p$ is a retract
of
$$
[\Sigma \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)]_p=\Sigma^{b_Y-p+1}Y
$$
containing the bottom cells by diagram~(\ref{equation6.4}). It
follows that
\begin{equation}\label{equation6.5}
[\Sigma Z]_p=[\Sigma \bar A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)]_p=\Sigma^{b_Y-p+1}Y
\end{equation}
because
$\Sigma^{b_Y-p+1}Y$ is atomic. Thus $Z$ contains $\Sigma^{b_Y-p}Y$
as the bottom piece.
Consider the filtered decomposition
$$
\Omega Y\simeq \bar B^{\mathrm{max}}(Y)\times \bar A^{{\mathrm{min}}}(Y).
$$
Since $Z$ is a filtered retract of $\bar A^{{\mathrm{min}}}(Y)$, there is a
further decomposition
$$
\Omega Y\simeq \bar B^{\mathrm{max}}(Y)\times Z\times A',
$$
where $A'$ is a filtered retract of $\Omega Y$ of type $A$. By
Lemma~\ref{lemma6.2}, $A'$ admits a further filtered decomposition
$$
A'\simeq B'\times X^{{\mathrm{min}}}(Y).
$$
This gives a filtered decomposition
\begin{equation}\label{equation6.6}
\Omega Y\simeq \bar B^{\mathrm{max}}(Y)\times Z\times B'\times X^{{\mathrm{min}}}(Y)
\end{equation}
and so the space $Q$ given in Lemma~\ref{lemma6.2} has the property that
\begin{equation}\label{equation6.7}
\Omega Q\simeq \bar B^{\mathrm{max}}(Y)\times Z\times B'.
\end{equation}
On homology
\begin{equation}\label{equation6.8}
PE^0H_*(Z)\oplus PE^0H_*(\bar B^{\mathrm{max}}(Y))\subseteq PE^0H_*(\Omega
Q).
\end{equation}
Let $V=\Sigma^{-1}\bar H_*(Y)$. From equation~(\ref{equation6.5}),
\begin{equation}\label{equation6.9}
\bar L_p(V)=A^{{\mathrm{min}}}_p(V)=\Sigma^{-1}\bar H_*([\Sigma \bar
A^{{\mathrm{min}}}(\Sigma^{b_Y-p+1}Y)]_p)\subseteq PE^0(\Omega Q).
\end{equation}
By equation~(\ref{equation6.8}),
$$
\bigoplus_{n=2}^{p-1}Q^{\mathrm{max}}_n(V)\oplus \bar L_p(V)\subseteq
PH_*(\Omega Q).
$$
Since $H_*(\Omega Q)$ is a Hopf algebra, the sub Hopf algebra
\begin{equation}\label{equation6.10}
B=\la \bigoplus_{n=2}^{p-1}Q^{\mathrm{max}}_n(V)\oplus \bar L_p(V)\ra
\subseteq H_*(\Omega Q).
\end{equation}
By Lemma~\ref{lemma3.2}, $B=B^{[1]}(V)$ with a short exact sequence
of Hopf algebras
$$
B\rInto T(V)\rOnto E(V).
$$
It follows that the Poincar\'e series
$$
\chi(H_*(\Omega Q))\geq \chi(B)=\frac{\chi(T(V))}{\chi(E(V))}.
$$
From the decomposition
$$
\Omega Y\simeq \Omega Q\times X^{{\mathrm{min}}}(Y),
$$
the Poincar\'e series
$$
\chi(H_*(X^{{\mathrm{min}}}(Y)))\leq \chi(E(V)).
$$
Since $X^{{\mathrm{min}}}(Y)$ is filtered retract of $\Omega Y$, $[\Sigma
X^{{\mathrm{min}}}(Y)]_1$ is a retract of $[\Sigma \Omega Y]_1=Y$. Because $Y$
is atomic, $[\Sigma X^{{\mathrm{min}}}(Y)]_1\simeq Y$. It follows that
$$
V\subseteq E^0H_*(X^{{\mathrm{min}}}(Y)).
$$
Since $X^{{\mathrm{min}}}(Y)$ is an $H$-space, $E^0H_*(X^{{\mathrm{min}}}(Y))$ is a Hopf
algebra and so
$$
\chi(H_*(X^{{\mathrm{min}}}(Y)))=\chi(E^0H_*(X^{{\mathrm{min}}}(Y)))\geq \chi(E(V)).
$$
Thus $\chi(H_*(X^{{\mathrm{min}}}(Y)))=\chi(E(V))$ and so
$H_*(X^{{\mathrm{min}}}(Y))=E(V)$ as coalgebras. It follows that $X^{{\mathrm{min}}}(Y)$
is an $H$-space having $Y$ as a retractile generating complex and
hence statement (3). The proof is finished.
\end{proof}
|
\section{Introduction, General Idea and Outline}
\label{sec: intro}
\subsection{Introduction and outline}
The AdS/CFT conjecture originally proposed by Maldacena \cite{Maldacena:1997re},
refined in \cite{Gubser:1998bc, Witten:1998qj}
and reviewed in \cite{Aharony:1999ti} has been one of the most interesting
developments in theoretical physics of the last decades.
It has become one of the most
powerful analytic tools to deal with strong coupling effects of some
particular
gauge theories in the planar limit $N_c \rightarrow \infty$.
The most studied and best understood case corresponds to $SU(N_c)$
${\cal N}=4$ SYM
which is a highly supersymmetric conformal theory and which only contains
matter in the adjoint representation of the gauge group.
Certainly, there are many interesting field theories which do not share these
properties and this fact has lead to an enormous amount of effort devoted to
extending the duality along different paths. Consequently, people have constructed
gravity duals of non-supersymmetric, non-conformal gauge theories,
in different vacua and
with diverse
matter contents.
One can mention the attempt of building a dual as close as possible to QCD as an
aim for these generalizations. However, one should keep in mind that this
is just one
among many desirable motivations, since understanding gauge theories at strong coupling
(or using gauge theories to understand gravity) is a very relevant problem {\it per se},
with both theoretical and phenomenological possible implications.
An important development of AdS/CFT has been to generalise the matter content of the gauge theories
under consideration and, in particular, to include fields which transform in the
fundamental representation of the gauge group, as the QCD quarks
do\footnote{With an abuse of language, we will use
throughout this review
the words {\it quark} or {\it flavor}
to refer to any field, fermionic or bosonic,
transforming in the fundamental representation of the gauge group. Accordingly, by {\it mesons} we will mean bound states of {\it quarks}. }.
A first possibility is to add the flavors in the {\it quenched} approximation.
The word {\it quenched} comes from the lattice literature
and, in that context, it amounts to setting
the quark fermion determinant to one. In more physical terms, {\it quenching} corresponds
to discarding the quark dynamical effects. This means that quantum effects produced by
the fundamentals are neglected, the quarks are considered as external non-dynamical objects
in the sense that they do not run in the loops\footnote{In the lattice, usually, quenching is
thought to be a good approximation for heavy quarks whereas for the gauge-gravity examples
the relevance of the quenched approximation comes from having parametrically less fundamental than
adjoint fields $N_f \ll N_c$.}. From the string side, adding quenched quarks to a given
gauge theory
corresponds to incorporating a set of brane probes in the dual background, which is not
modified with respect to the quark-less case. By analysing the
worldvolume physics of these flavor branes (typically using the Dirac-Born-Infeld +
Wess-Zumino action) a lot of physically interesting questions can be understood.
For instance and just to mention a few,
chiral symmetry breaking can be neatly described, phase diagrams can be
constructed and meson spectra can be exactly computed.
It is hard to do justice to the huge literature in the subject so let us just
mention the seminal papers \cite{Karch:2002sh}, \cite{Kruczenski:2003be} and a recent review
\cite{Erdmenger:2007cm}.
Thus, it is fair to say that the study of quenched flavor within the gauge-gravity
correspondence
has been a very fruitful program. Nonetheless, there are physical features which
are intimately related to the quantum effects of the quarks.
Examples are the consequences of the presence of fundamentals
on the running couplings, which may ultimately lead to conformal points, conformal windows
\cite{Banks:1981nn} or
Seiberg-like dualities \cite{Seiberg:1994pq}. More phenomenologically, multihadron production,
the screening of the color charge or the large mass of the $\eta'$ meson are spin-offs of
these quantum effects. Let us also mention that the most successful
application of string duals towards phenomenology has been the construction of solutions that
can be used as toy models for the experimental quark-gluon plasma.
Thus, a very interesting program is to build black hole solutions with unquenched flavor which
are really dual to {\it quark}-gluon plasmas, {\it i.e.} such that the effect of
the dynamical
quarks affects the plasma physics, as is expected to be the case in the real world.
These observations largely motivate the study
of theories with unquenched quarks
from the string theory dual point of view.
Unquenching the flavors of the gauge theory has a very precise implication for the
dual theory: the gravity background has to be modified by the inclusion of the quarks,
namely, one needs to take into account the back-reaction on the geometry produced by the
flavor branes. The main goal in the following will be to present methods to compute such
back-reacted solutions. This will be done by presenting different examples
that, hopefully, will help the reader to gain insight in both the physical questions and
the technical tools used to address them.
In this review, we will focus on a specific family of unquenched constructions.
Namely, we will discuss at length just
solutions of type IIA or type IIB string theory in which the fundamentals come from a
{\it smeared} set of flavor branes. In section \ref{heuristiczzz}, we will try to provide
a general understanding of this notion of smearing the flavor and argue why we find it
a case of particular interest.
As we will see, by considering the case of smeared D-branes we can build a systematic approach
applicable to different situations and which typically results in large simplifications as compared
to other kind of flavor D-brane distributions.
This smearing procedure referring to flavors
was first introduced in \cite{Bigazzi:2005md} in a non-critical string framework and
in \cite{Casero:2006pt} in a well-controlled ten-dimensional context.
It is important to remark that this smearing is by no means the only possibility to introduce
unquenched fundamentals in gauge-gravity duals. Many important works have
followed alternative
paths to construct different models. We are not able to review them here, but we provide a survey
of the literature in section \ref{sec:localized}.
\subsubsection*{Outline}
We will devote the rest of section \ref{sec: intro} to
further clarifying the kind of physical problems
we want to address and to give the general methods and notions which are common
to all the constructions we will present later.
Then, sections \ref{AdS5X5}-\ref{models2}
will analyse different models that
are ordered in increasing order of complexity. Each section can be
read mostly independently
from the rest.
The discussion of each model can always be regarded as a two step process.
First, one has to solve the equations for finding
the back-reacted solutions of type II supergravity
coupled to a set of D-brane sources. Second, one can use these solutions to extract
the physics of the conjectured gauge theory duals with unquenched flavors.
Readers interested in different aspects of the problem can consult the
different parts independently. We would like to stress that, even without making any
reference to the gauge-gravity correspondence, the string theory solutions and methods developed
to find them are interesting by themselves.
Section \ref{AdS5X5} deals with the backreaction of D7-branes on $AdS_5 \times X^5$
spaces, where $X^5$ stands for a Sasaki-Einstein space. As a matter of fact,
a large part of the discussion can
be carried out without specifying the $X^5$. Notwithstanding, the two most interesting
cases correspond to $X^5=S^5$ and $X^5=T^{1,1}$. At different points during section
\ref{AdS5X5}, we will refer to these particular examples in order to explain concrete
features. We will present supersymmetric solutions for massless and massive quarks, and
also non-supersymmetric black hole
solutions which are dual to theories at finite temperature,
in a deconfined plasma phase. All these solutions share the
property of having a singularity, associated with a UV Landau pole in the field theory
(when quarks are massless and the temperature is zero, there is also a naked IR singularity).
We will show how to make well-defined IR predictions from the geometry, even
in the presence of the UV singularity
(in much the same spirit as in field theory renormalization).
In section \ref{sec:D5D5}, we will discuss a model in which both the color and flavor
branes are D5's. It is dual to a (3+1)-dimensional ${\cal N}=1$ theory
with a UV completion.
Among several nice features of the model that will be presented, we would like to
remark here that it incorporates a geometrical description of a Seiberg-like duality.
Section \ref{sec:2+1} is also built from a D5-D5 intersection and in fact shares
several similarities with the previous model. The construction corresponds to color D5's
wrapping a compact 3-cycle and therefore the dual field theory is (2+1)-dimensional.
In section \ref{KS} we examine the addition of D7-branes to the
conspicuous Klebanov-Strassler model \cite{Klebanov:2000hb}. From the physical point of view, how
the unquenched flavors affect a duality cascade is particularly interesting.
From a technical point of view, the system is slightly more involved than the
previous ones because different RR and NSNS forms are turned on.
However, despite this complication, it is remarkable that almost all functions of the ansatz
can be integrated in a closed form.
Section \ref{models2} reviews a different class of models. The dual gauge theories
are built on wrapped supersymmetric D-branes
with the peculiarity that some of the adjoint scalars
remain massless. As we will explain, it is not sensible to smear the branes in
all the transverse directions.
The associated main technical difficulty will be
the fact that one has to solve
partial differential equations to find the background.
Profiting from the experience gained by discussing these examples one by one,
in section \ref{mathviewpointzz}, we will give a more mathematical viewpoint
of the constructions. In particular, we will take some tools of differential
geometry to describe in a concise and compact way the distributions of mass and
charge due to the presence of the flavor branes.
Finally, in section \ref{outlook} we will conclude by summarizing the whole topic
trying to give a general perspective of the results obtained and by
also providing an outlook of the subject.
\subsection{Presentation of the problem}\label{section1}
As anticipated in the introduction, we will discuss the addition of
flavors to field theories (mostly focusing on SUSY examples, but this is
not mandatory) using AdS/CFT or more generally, gauge-strings duality.
We hope it is clear to the reader that the addition of flavors (fields
transforming in the fundamental representation) is a very interesting
exercise from a dynamical point of view. Indeed, in a theory with adjoint
fields (let us for the sake of this discussion, consider the case of a
confining field theory) the presence
of fields transforming in the fundamental will produce the breaking of the
``QCD-string'' or screening. Of course, the fundamentals will add a new
symmetry, that can be $SU(N_f)$ or, like in massless QCD, $SU(N_f)\times
SU(N_f)$; a baryonic $U(1)_B$ symmetry should also appear. Obviously the
presence of global symmetries (and their possible spontaneous or explicit
breaking)
will directly reflect in
the spectrum. Apart from this, it will happen that the states before the
addition of fundamentals, that is the glueballs, will interact and mix
with the mesons, giving place to new diagonal combinations that will
be the observed states.
Moreover, anomalies will be modified, as fermions that transform in
the fundamental will run in the triangles. Also gauge couplings will run
differently and finally new dualities (Seiberg-like \cite{Seiberg:1994pq})
may appear.
In the rest of this article, we will discuss how all of the above
mentioned features are encoded in string backgrounds.
It is clear that we need to add new objects to our string background.
These new objects are D-branes, on which a gauge field propagates,
encoding
the
presence of a $U(N_f)$ gauge symmetry (in the bulk), dual to the global
$U(N_f)$ in the dual QFT. Also, it is on these D-branes that the meson
fields, represented by excitations of the branes, propagate and interact.
Following a nomenclature that by now became standard, we will call these
D-branes ``flavor branes''.
It is then clear that to add flavors to a field theory whose dual we know,
we should consider the original (unflavored) string background and add
flavor-branes. Now, the point is how to proceed technically to add these
new branes.
It may be useful to consider two developments of the 1970's, that will
turn to illuminate on the answer to this question.
In the papers \cite{'t Hooft:1973jz} and \cite{Veneziano:1976wm}
't Hooft and Veneziano respectively considered the influence of
fundamentals when the following scaling is taken
\begin{equation}
N_c \rightarrow \infty, \;\;\;\lambda=g_{YM}^2 N_c=fixed,
\end{equation}
and considered the two possible cases ('t Hooft and Veneziano
respectively)
\begin{eqnarray}
& & N_f=fixed,\;\;\; x=\frac{N_f}{N_c}\rightarrow 0,\nonumber\\
& & N_f \rightarrow \infty, \;\;\; x=\frac{N_f}{N_c}=fixed\,\,.\nonumber
\end{eqnarray}
It is very illuminating to see how different diagrams contributing to the
same physical process (to fix ideas, an n-point correlator of mesonic
currents) scale in these two cases.
In this respect,
a formula for the kinematical factor of the scattering of $n$ mesons was
produced in \cite{Capella:1992yb},
considering diagrams with $w$ internal fermion loops (windows), $h$
non-planar handles and $b$ boundaries,
\begin{equation}
<B_1....B_n> \sim \Big(\,\frac{N_f}{N_c}\,\Big)^w \, N_c^{(2-\frac{n}{2} -2h -b)}.
\label{scatk}
\end{equation}
Consider the case of scattering of two mesons $n=2$. We see that diagrams
like the first one in the figure ($w=h=0, b=1,n=2$) scale like a constant
$N_c^0\sim 1$, the
second diagram (with $w=1,h=0,b=1,n=2$) scales like $\frac{N_f}{N_c}$,
while the third one (with $w=0, h=0,b=2,n=2$, that is non-planar) goes
like $N_c^{-1}$.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.5]{Nscale.eps}
\caption{\small Diagrams for a meson propagator, with two
insertions of the meson operator ($n=2$) shown as thick
points on the boundaries. The dashed lines are gluons
that fill the diagram in the large $N_c$ limit and the
thick lines are quarks. A) Planar diagram with no internal quark loops
($h=0$, $w=0$), the scaling is $\sim 1$.
B) Planar diagram with an internal quark loop $h=0$, $w=1$,
the scaling is $\sim N_f/N_c$.
C) Non-planar diagram with no internal quark loops $h=0$, $w=0$, $b=2$,
the scaling is $\sim 1/N_c $.}
\label{htb!}\end{figure}
So, we see that from this view point, the Veneziano scaling captures more
physics, represented here by diagrams like $B$ in Fig. \ref{htb!}.
Nevertheless, there may be some particular problems for which studying
things in the 't Hooft scaling may be enough.
From the view point of a lattice theorist,
working in the 't Hooft scaling, hence neglecting the effects of fundamentals running inside loops,
is the same as working in what they would call the
``quenched approximation''.
We can think of the field theory as being quenched when the fundamental
fields do not propagate inside the loops. One natural way of doing this is
to consider the case of very massive quarks.
Indeed, when quenching,
one considers an expansion of the fermionic determinant (for massive quarks) of the form:
\begin{equation}
\log \det [\gamma^\mu D_\mu +m]= Tr \log(\gamma^\mu D_\mu +m)= Tr \log (m[1 + \frac{\gamma^\mu D_\mu }{m}]) \sim
Tr \log(m) + O(1/m)\,\,.
\end{equation}
Keeping only the constant term (or considering a very large mass) is equivalent to saying
that fundamentals are very difficult to pair-produce, hence
their presence inside loops will be very suppressed. Another way to quench
in the field theory is to consider the case in which the quotient
$x=\frac{N_f}{N_c}$ is very small. Notice that
the quenched theory is not equivalent to a theory with only adjoints, as
fundamentals can occur
in external lines, like in a correlator of two mesonic currents as
exemplified in the diagram A of Fig.\ref{htb!}.
Needless to say, lattice theorists developed techniques to quench
fundamental fields with arbitrary mass.
Also, while at first sight the quenching as described above is not a good operation as it breaks unitarity (not including all possible diagrams)
this kind of troubles
will be avoided when working in the 't Hooft scaling,
where unitarity problems will be suppressed in $1/N_c$ (but of course
will be present in a lattice version of theories with finite $N_c$).
The interesting point to be taken from this, by a physicist working in
gauge-string duality is
that both scalings ('t Hooft's and Veneziano's) can be realized with D-branes.
Indeed, in both cases we must add D-branes
(to realize symmetries and new states as discussed above), but we can add these flavor branes in two ways:
\begin{itemize}
\item{We can add $N_f$ flavor branes in such a way that we will only
{\it probe} the geometry produced by the $N_c$ color branes. In this case the
dynamics of the probe-flavor branes (the mesons) will be influenced by the
dynamics of the color branes (the glueballs) but not vice-versa. This is a
good
approximation if $x=\frac{N_f}{N_c}\rightarrow 0$, which immediately sets us in the 't Hooft scaling limit.
Notice however, that when the $N_c$ contribution to some
particular quantity vanishes, the flavor effects may
be the leading ones even when $N_f \ll N_c$.}
\item{We can add $N_f$ flavor branes, in such a way that they will deform the already existing geometry, in other words {\it backreacting} on the
original ``color'' geometry. In field theory language, we would say that the dynamics of the glueballs and that of the mesons influence each other,
leading to new states that will be a mixture of mesons and glueballs. This is surely what we need to do if $x=\frac{N_f}{N_c}=fixed$ and doing this
will set us in the Veneziano scaling limit. }
\end{itemize}
More technically, in the 't Hooft scaling
limit we are studying the Born-Infeld-Wess-Zumino dynamics for a D$k$ flavor
brane in a background created by
$N_c$ color D$p$-branes (we will always work in Einstein frame in the
following):
\begin{equation}
S_{BIWZ}= -T_{k}\int d^{k+1} x \, e^{\frac{k-3}{4}\phi}
\sqrt{\det[\hat g_{ab} + 2\pi \alpha'e^{-\frac{\phi}{2}} {\cal F}_{ab}]}+ T_{k}\int C \wedge e^{{\cal F}}\,\,,
\label{BIWZaction}
\end{equation}
where $\hat g_{ab}, {\cal F}_{ab}$
are fields {\it induced} by the color branes background on the (few) flavor branes.
The `shape' of the flavor branes (induced metric)
will then influence the mass spectrum and interactions of the fluctuations of the flavor branes (the mesons), explicitly realizing the picture
advocated above. This line of research was initiated
by Karch and Katz \cite{Karch:2002sh}
and substantially clarified in subsequent papers
\cite{Kruczenski:2003be}, \cite{Kruczenski:2003uq}-\cite{Sakai:2005yt}. This
line kept on growing in the
last few years, finding numerous
applications. See the paper \cite{Erdmenger:2007cm}
for a comprehensive review.
On the other hand, working in the Veneziano scaling limit
implies that we will need to study the action
\begin{equation}
S= S_{IIA/IIB} + S_{BIWZ}.
\label{iibbiwz}
\end{equation}
There will be new equations of motion,
encoding explicitly the numbers $N_f, N_c$.
As discussed above, it is now very clear that proceeding like this will be
the only
possibility when the number of flavors is comparable with the number of
colors.
Also, it makes manifest the fact that the
dynamics of glueballs (represented on the string side by $S_{IIA/IIB}$
which is of order $g_s^{-2}\sim N_c^2$)
is influenced and influences back on
the dynamics of fundamentals
(represented by $S_{BIWZ}$, of order $g_s^{-1} N_f \sim N_c N_f$).
The rest of this review will focus on this second scaling (Veneziano).
Notice that (in both scalings) we are making an explicit difference
between the color $SU(N_c)$ {\it gauged}
symmetry and the flavor $SU(N_f)$ {\it global} symmetry on the field
theory side.
From the string theory construction, this qualitative difference is connected to the
fact that the volume of
the flavor branes is infinite, as compared to the volume of the color branes.
Indeed, in the bulk, we only need to realize the field theory
global symmetry, and we do it with the gauge field present in the Born-Infeld-Wess-Zumino action. Searching for solutions of $D_p$ color branes in
interaction with $D_k$ flavor branes in pure IIA/IIB supergravity is an interesting problem, but will not
represent the physical system we are after, as only flavor singlet states would be included in the dynamics.
Before we proceed studying the
formalism and examples to clarify the details, some comments are in order.
\subsection{The string action and the scaling limit in $N_c$ and $N_f$}
Let us study a bit more the expression of eq. (\ref{iibbiwz}), being
careful about coefficients. We will consider the case of a set of $N_c$
``color'' D$p$-branes and $N_f$ ``flavor'' D$k$-branes. The action for
this system will be, in Einstein frame:
\begin{eqnarray}
& & S= \frac{1}{2\kappa_{10}^2}\int d^{10} x \sqrt{g_{10}}
\Big[R-\frac{1}{2}(\partial\phi)^2 -\frac{e^{-\phi}}{12}H_3^2 -\sum_l
\frac{e^{ \frac{5-l}{2} \phi}}{2 \times l!} F_{l}^2 \Big] + \int CS-terms
\nonumber\\
& &\,\,\,
- N_f T_{k} \int d^{k+1}x \,e^{\frac{k-3}{4}\phi}\sqrt{\det[\hat g_{ab} + 2\pi\alpha' e^{-\frac{\phi}{2}}
{\cal F}_{ab}]} + N_f T_{k} \int_{k+1} C\wedge e^{\cal F}\,\,,\nonumber\\\nonumber\\
& & S= \frac{1}{2\kappa_{10}^2} \Big[ \int L(IIA/IIB) - 2\kappa_{10}^2
N_f T_{k} \int L_{BIWZ} \Big],
\label{lagrangianzz}
\end{eqnarray}
where by $F_l$ we have denoted the various RR fields and with $CS-terms$ the
possible Chern-Simons terms. We have taken the simplification of writing the action for
the set of flavor branes as $N_f$ times that of a single D-brane, which is enough for
the large $N$ counting we want to undertake here.
The gravitational constant and D-brane tension are:
\begin{equation}
2\kappa_{10}^2= (2\pi)^7 g_s^2 \alpha'^{\,4},\;\;\;\qquad T_{k}=\frac{1}{(2\pi)^k g_s
(\alpha')^{\frac{k+1}{2}}} \,\,.
\label{stringthconsts}
\end{equation}
The typical quantization condition for the color branes reads:
\begin{equation}
\frac{1}{2\kappa_{10}^2} \int F_{8-p} = N_c T_p\,\,.
\label{quantDp}
\end{equation}
As a
consequence of (\ref{lagrangianzz}), we will have equations of
motion, that generically will read for the metric and dilaton:
\begin{eqnarray}
& & R_{\mu\nu}-\frac{1}{2}g_{\mu\nu}R= T_{\mu\nu}[IIA/B] -
2\kappa_{10}^2 N_f T_{k} T_{\mu\nu}[brane],\nonumber\\\nonumber\\
& & \nabla^2 \phi = \frac{\partial}{\partial \phi}\Big[L[IIA/B] -
2\kappa_{10}^2 N_f T_{k} L[BIWZ] \Big],
\label{verygeneraleom}
\end{eqnarray}
together with the modified (by the CS-terms) Maxwell equations and
importantly, the
Bianchi identity for the (magnetic) Ramond-Ramond field $F_{8-k}$ that couples to the
flavor D$k$-branes:
\begin{equation}
d F_{8-k}= 2\kappa_{10}^2 N_f T_{k} \delta^{9-k}(\vec{r})\,\,,
\end{equation}
indicating that the flavor branes are localized (all together) at the
position
$\vec{r}=0$. Similarly the $T_{\mu\nu}[brane]$ contains delta functions with support on
the position of the flavor branes. In principle, one will need to solve
second order, nonlinear, partial differential equations.
Instead of directly dealing with the above equations, we want to present
here an argument to understand which parameter controls the size of the flavor
effects on the action and, therefore, on the solution.
We remark that
the reasoning below is qualitative and in particular we
will just write a background for flat D$p$-branes
as considered for instance in \cite{Itzhaki:1998dd}.
This will be enough for understanding the
scaling with the parameters, at least in the
cases studied in this review.
In the following, we just focus on the behaviour with respect
to $N_f$, $N_c$, $g_{YM}^2$ and do not care about numerical prefactors.
We will use notation similar to \cite{Itzhaki:1998dd}.
Consider the
background associated to a stack of D$p$ color branes (in Einstein frame):
\begin{eqnarray}
ds^2&=& e^{-\frac{\phi}{2}}\alpha' \left[
\frac{(\sqrt{\alpha'}U)^{(7-p)/2} }{\alpha'\, c_p \sqrt{g_s N_c}}dx_{1,p}^2+
\frac{\alpha'\,c_p \sqrt{g_s N_c}}{(\sqrt{\alpha'}U)^{(7-p)/2}}dU^2
+ c_p \sqrt{g_s N_c} (\sqrt{\alpha'}U)^{(p-3)/2}d\Omega_{8-p}^2\right]\,\,,\nonumber\\
e^\phi &\sim& \left( \frac{g_s N_c}{(\sqrt{\alpha'}U)^{7-p}}\right)^{\frac{3-p}{4}}\,\,,
\label{itzhakimetric}
\end{eqnarray}
where $c_p$ is a known
numerical constant and $U$ is an energy scale.
On a background of this kind, we want to introduce
$N_f$ D$k$-flavor branes and to know which is the relative importance of the associated
terms in the action (\ref{lagrangianzz}) and equations of motion (\ref{verygeneraleom}).
With that aim, let us start by computing the coefficient in
front of the term coming from the RR-form sourced
by the color branes in (\ref{lagrangianzz}), namely
$(2\kappa_{10}^2)^{-1} \sqrt{g_{10}} e^{\frac{p-3}{2}\phi} F_{8-p}^2$.
Using (\ref{quantDp}) and (\ref{itzhakimetric}), we find that the lagrangian density associated
to the color branes goes as:
\begin{equation}
{\cal L}_{color} \sim (2\kappa_{10}^2)^{-1} \sqrt{g_{10}}\,\,
e^{\frac{5\phi}{2}} (\alpha'^{-1}) (g_s N_c)^{(p-4)/2} (\sqrt{\alpha'} U )^{(p-3)(p-8)/2}\,\,.
\label{LLcolor}
\end{equation}
Let us now look at the DBI term. We assume that the flavor D$k$-branes are extended
along the Minkowski directions, the radial direction $U$ and $k-p-1$ directions within the
sphere. We find:
\begin{equation}
{\cal L}_{flavor,DBI}
\sim N_f T_k e^{\frac{k-3}{4}\phi}
\sqrt{\hat g_{k+1}} \sim \,\, \frac{N_f}{N_c} \,\,g_{eff}^{\frac{k-p}{2}}\,\, {\cal L}_{color}\,\,,
\end{equation}
where in order to get the last expression we have used
(\ref{stringthconsts}),
(\ref{itzhakimetric}),
(\ref{LLcolor}) and defined a dimensionless effective coupling as in \cite{Itzhaki:1998dd}:
\begin{equation}
g_{eff}^2 \sim g_{YM}^2 N_c U^{p-3}\sim g_s N_c (\sqrt{\alpha'} U)^{p-3}\,\,.
\label{geff}
\end{equation}
Thus, parametrically, the action from the flavor branes as compared with
that from the color background
is weighed by $\frac{N_f}{N_c}\,\, g_{eff}^{\frac{k-p}{2}}$.
We now want to take a low energy decoupling limit as in \cite{Itzhaki:1998dd}
(see also \cite{Boonstra:1998mp}),
namely the dimensionless effective coupling
$g_{eff}$ and $U$ are fixed as $\alpha'\rightarrow 0$.
Thus, the Veneziano scaling limit in this framework amounts to:
\begin{equation}
N_c,N_f\rightarrow \infty\,\,,\qquad\quad g_{eff} \ \ \textrm{fixed},
\qquad\quad \frac{N_f}{N_c}\,\, g_{eff}^{\frac{k-p}{2}} \ \ \textrm{fixed},
\label{ncnfscaling}
\end{equation}
where the last relation comes from demanding that the flavor effects
are also fixed.
Staying in the supergravity limit requires:
\begin{equation}
1 \ll g_{eff}^2 \ll N_c^{\frac{4}{7-p}}\,\,,
\label{validitz}
\end{equation}
a constraint that limits the range of energy scales $U$ for which the supergravity
description is valid \cite{Itzhaki:1998dd}. Notice that if we further require that the flavor
terms do not parametrically
dominate over the color ones, this can further restrict $U$, depending on $p$ and $k$.
The probe limit, in which the flavor action is negligible as compared to the
gravity action, comes from making the last quantity
in (\ref{ncnfscaling})
vanishingly small\footnote{In the literature, it is
usually written that the probe limit is good when $N_f \ll N_c$. That is not strictly
correct. For instance in the D3-D7 case, the probe approximation is valid when,
parametrically, $\lambda\, N_f \ll N_c$.}.
As expected, that term is strictly zero in the 't Hooft limit.
We now comment on the values of $p$, $k$ that will appear in the following sections:
For the D3-D7 case of section
\ref{AdS5X5}, the parameter weighing the flavor effects
is $(g_{YM}^2 N_c)\frac{N_f}{N_c} \sim \lambda\frac{N_f}{N_c}$.
For the cascading case of section \ref{KS}, the result is similar but one has to replace
$N_c$ by the number of fractional branes $M$.
Getting ahead of the discussion of upcoming sections we notice that,
in these cases,
it is not enough to take this parameter
fixed, but it should also be small. This will be
due to positive beta functions, as will be thoroughly discussed.
From (\ref{ncnfscaling}), we see that $k=p$ (sections
\ref{sec:D5D5},\ref{sec:2+1} and \ref{models2})
is particularly interesting\footnote{
Even if in all these sections we
will deal with wrapped branes and therefore the backgrounds are not that similar
to (\ref{itzhakimetric}), the argument above still yields the correct result.}, since it is really $\frac{N_f}{N_c}$
what has to be taken fixed.
For this reason, only in these cases one
can hope to describe - within gravity - phenomena as Seiberg-like dualities
(see section \ref{sec:Seib}). Loosely speaking, the Klebanov-Strassler duality
cascade \cite{Klebanov:2000hb} lies in this class of $k=p$ theories, since it
can be understood as the interplay between two
sets of D5-branes wrapping vanishing two-cycles.
We close this section by mentioning other brane intersections
that will not be discussed further in later sections. In a D2-D6 system,
the effective coupling (\ref{geff}) decreases at large $U$ and therefore the
flavor backreaction on the glue fades away in the UV - see
(\ref{ncnfscaling}) - as expected in
a superrenormalizable theory. This was observed in \cite{cherkishash} when studying a solution
with localized D6-branes. In a D4-D8 intersection, the opposite happens. The probe brane approach can be
valid in an intermediate regime $1 \ll g_{eff}^2 \ll \frac{N_c}{N_f}$ but at a given $U$
the fundamentals eventually take over and dominate. Notice that the value of $U$ for which
the D4-D8 theory loses its validity is parametrically smaller than that for which the
unflavored D4-brane theory becomes pathological, which is set by (\ref{validitz}).
\subsection{The method}
Looking back at eq. (\ref{lagrangianzz}), one
can appreciate that in general finding the solution describing
the
backreaction between the type II closed strings and the open strings described
by the Born-Infeld
action is quite a challenging problem.
Indeed, the fact that the flavor branes (BIWZ) are localized in the
ten-dimensional space
implies that we will have to solve
second order, nonlinear, coupled and partial
differential equations with localized sources.
Basically what makes things so difficult
are the presence of delta function sources and the fact that the
differential equations describing the
dynamics are ``partial'' (in
principle depending on all the variables
describing the space transverse to the
flavor branes). In order to get some intuition of the answer, we may consider the case in which
we will ``erase'' the dependence on these
transverse coordinates (this is like considering the
``s-wave'' of the putative multipole decomposition of the full solution in
this transverse space) and delocalize the sources.
To achieve this, we will propose to {\it smear} the flavor branes
over their perpendicular space.
On the field theory side, this will amount to considering systems where the addition of the degrees
of freedom transforming in the fundamental does not break any of the global symmetries of the unflavored
QFT. Also, it may happen that the original $U(N_f)$ is explicitly broken to $U(1)^{N_f}$ as we are
separating the flavor branes - see the discussion in section 7 of
\cite{HoyosBadajoz:2008fw} and in section 2 of \cite{Bigazzi:2008zt}.
An intuitive
understanding of the smearing procedure will be discussed in
section \ref{heuristiczzz}, while a more formal approach will be
treated in section \ref{mathviewpointzz}.
For technical reasons, this procedure is cleaner in examples preserving some amount of SUSY, since the
force between flavor branes is cancelled and the smearing is at no cost of energy.
In the examples described in the following sections, we will proceed like this:
\begin{itemize}
\item{Consider an unflavored string background and find the embedding of flavor branes that will preserve
some SUSY, or (in non-SUSY examples) that will be stable and solve the equations of motion for the brane.
In the SUSY cases,
this can be achieved by
considering kappa-symmetric embeddings, that we review
generically below.}
\item{Consider now $N_f$ flavor branes in that particular
embedding and smear them, getting an action in ten dimensions, as will
be explained with all generality in section \ref{mathviewpointzz}.}
\item{Solve the equations derived from (\ref{lagrangianzz}),
that will now contain smeared branes and will
be ordinary differential
equations. In SUSY cases,
there will be a set of BPS equations to be solved. In non-SUSY examples
one might manage to get a fake superpotential and fake-BPS equations
\cite{Freedman:2003ax}. }
\end{itemize}
Moreover, one has to check that the flavor embeddings considered are still a solution
in the backreacted geometry.
Let us review briefly the main technical points collected above.
\subsubsection{BPS equations, Kappa symmetry (SUSY probes) and smearing}
\label{sec: kappa}
Let us consider the case of a supersymmetric background, namely a solution
of type II sugra
for which the supersymmetry variations of the gravitino and the dilatino
vanish $\delta \psi_\mu = \delta \lambda =0$.
We will not give here details on the form of these expressions, which can
be found elsewhere. For instance, the string frame
SUSY transformations of both type IIA and type IIB
are written down in appendix A of \cite{Martucci:2005rb}.
Given a background that preserves some amount of SUSY, the idea is to find
the hyper-surfaces in which to extend the flavor branes (in other words,
finding the embeddings for flavor branes) so that these will preserve all
(or a fraction) of the SUSY of the background.
One then writes an eigenvalue problem, imposing that the preserved spinors
of the background are eigenspinors of the kappa-symmetry matrix
\begin{equation}
\Gamma_\kappa\,\, \epsilon=\epsilon\,\,.
\end{equation}
See the papers \cite{Cederwall:1996ri}
for the definition of $\Gamma_\kappa$.
Once we have the kappa-symmetric embeddings as described above, we now
proceed to write an action describing the dynamics of the closed and open
strings, as in eq. (\ref{lagrangianzz}). We then realize that the problem
will lead (unless we are adding D9-branes) to a system of partial
differential equations. As discussed above, we then proceed to smear these
flavor branes. For this we propose an ansatz for the metric, where the
embedding of the flavor branes is clear and distribute them along the
directions of their transverse space. This distribution of the flavor
branes can be done in a uniform way. In some sense, we are `deconstructing'
the transverse space to the flavor branes by adding at each point one of
the many $N_f$ flavor branes.
The key point is that once the BPS equations and kappa-symmetry conditions
are simultaneously satisfied,
the problem is solved. In fact, it is a general result \cite{Koerber:2007hd} that
the SUSY equations $\delta \psi_\mu = \delta \lambda =0$, together with the
Bianchi identities
- and equations of motion -
for the different forms modified by calibrated (namely, kappa-symmetric) sources
imply the full set of equations of motion.
In the following, we will discuss first an intuitive way of
understanding this smearing. Then
we will apply this to different examples in sections \ref{AdS5X5}-\ref{models2}.
Finally, in section \ref{mathviewpointzz}, we will present
a formal way of implementing the backreaction from smeared sources.
\subsection{A heuristic viewpoint}
\label{heuristiczzz}
In the following sections, we will introduce the necessary mathematical machinery to consistently
compute solutions of string theory in which smeared backreacting flavor branes are present.
Before that, it is worth to make a digression in order to explain the
general set-up in
simple, heuristic terms.
Let us make an analogy with electrostatics. Suppose we want to compute the electric field
generated by a point-like charge and a couple of lines of charge, as depicted on the left
of Figure \ref{massless_plot}. In order to depict the situation, we show
dimension 1 lines
of charge in a total space of dimension 2, but clearly the situation can be generalized by changing
such dimensions. Since in the left plot there is no particular symmetry in the configuration, the
resulting electric field will have a not so simple expression. But let us imagine that we consider a huge
number of lines of charge as in the plot of the right, and homogeneously distribute them in the angle
they form with the horizontal axis.
In the limit of many lines, radial symmetry is recovered, the charge density
is ``smeared" and
will be just given by a single (monotonically decreasing) function
$\rho(r)$. The electric field, accordingly, will also be radially symmetric.
Notice that this process of describing a large number of discrete objects by a continuous distribution
is ubiquitous in physics: for instance a ``homogeneous" gas is a collection of atoms or the
``homogeneous" Universe considered in cosmological models contains a collection of galaxy clusters.
Also, solutions with different kinds of
smeared sources have been considered many times
in string theory contexts not necessarily related to gauge-gravity duality, see for
instance \cite{Grana:2006kf}, \cite{DeWolfe:2008zy}.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=.3\textwidth]{few_massless.eps}
\qquad\qquad\qquad
\includegraphics[width=.3\textwidth]{smear_massless.eps}}
\caption{On the left, a point-like charge (in red) and two lines of charge at different angles.
On the right, a configuration with many lines of charge. In the
asymptotic limit of an infinite
number of lines, they just correspond to a radial charge density. This picture depicts an
analogous situation to the case of smeared flavored branes, when the fundamental fields are
massless.}
\label{massless_plot}
\end{figure}
When comparing to the string theory set-up, the point-like
charge in the center corresponds to the color branes and the lines to the flavor branes
(which have to extend to infinity). The limiting, radially symmetric configuration, corresponds to
the kind of smeared situations that we will discuss in this review.\footnote{More precisely, it corresponds
to the situations analysed in sections \ref{AdS5X5}-\ref{KS}. For the
cohomogeneity 2 cases analysed in section
\ref{models2}, the different functions depend on two radial variables. A heuristic picture for such
situations is presented in section \ref{models2}.}
All functions of the ansatz can then be considered to depend on a single radial variable.
For flavor branes, the different
``angles" correspond to adding fundamental matter which couples differently to the rest of the fields.
In some of the cases discussed in the following, we will see how this is reflected in the field theory
superpotential (section \ref{masslessKW}).
We can still get further intuition from this simple analogy. In the right plot of figure
\ref{massless_plot}, we see that all lines intersect at the center. From the string point of view,
that means that the flavor branes are stretched down to the bottom of the geometry and the quarks
are massless.
In this situation, the charge density
$\rho(r)$ is highly peaked at $r=0$. Essentially, that is the reason why for the
solutions with massless
quarks described in the following sections there is a curvature singularity at the origin, where
all flavor branes meet.
Then, a simple way of getting rid of such a singularity is to displace the lines of charge from the origin,
while still keeping the radial symmetry. This is depicted in figure \ref{massive_plot}.
If we dub the distance from any of the lines to the center as $r_q$, the density of charge
$\rho(r)$ will vanish for $0<r<r_q$, while it will asymptote to the ''massless" $r_q=0$ one as $r\gg r_q$.
From the brane
construction, this displacement
typically corresponds to giving a mass to the fundamentals (or, in particular cases, it could
correspond to a non-trivial vacuum expectation value).
The solution of section \ref{massiveKW}, which indeed is regular in the IR, is a neat example of this notion.
Another possibility is to add temperature and to hide the singularity behind a horizon,
see section \ref{D3D7plasma}.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=.3\textwidth]{few_massive.eps}
\qquad\qquad\qquad
\includegraphics[width=.3\textwidth]{smear_massive.eps}}
\caption{These pictures depict
analogous situations to the case of flavored branes, when the fundamental fields are
massive. On the right, again, we add a large number of lines, such that in the limit radial
symmetry is recovered.}
\label{massive_plot}
\end{figure}
Going back to electrostatics for the right plot of figure \ref{massive_plot}, we know from Gauss' law
that the charge density outside does not affect the central region. The corresponding statement in the
field theory is that the massive fields decouple from the IR physics below the scale given by their mass.
We find it interesting that, through this heuristic reasoning, Gauss' law is
connected to the decoupling of heavy
particles (or holomorphic decoupling in the SUSY cases).
Even if the example of electrostatics is useful to qualitatively
picture what we will do in the following, the analogy is
by no means perfect. We note two differences:
first, we will be working with gravity, which is non-linear and, thus, one cannot find the final
solution by superposing the fields generated by different sources (which in the case of electrostatics
would make it rather trivial to find the electric field for the configurations depicted on the left
of figures \ref{massless_plot} and \ref{massive_plot}). Second, our ``lines of charge" (the flavor
branes) are dynamical. This means that is it not enough to compute the background fields generated from the
sources but one also has to check that the sources are stably embedded in the geometry.
We end this section by summarizing the pros and cons of looking
for duals of unquenched theories for which the string solutions include smeared flavor
branes, many of which can be inferred from the heuristic discussion above. On the positive side:
\begin{itemize}
\item The smearing simplifies the situation allowing us to write ans\"atze depending on a single radial coordinate,
and therefore the problem is eventually reduced to a set of ODEs. (For the cases of section \ref{models2}, they
depend on two radial coordinates and thus one finds PDEs in terms of two
independent variables, but again, without delta-function localized
sources.)
\item Possible issues related to singularities and strong coupling are ameliorated in the same sense as
they are washed out in electrostatics when
considering a smooth charge density rather than a sum of delta-functions over a large number of electrons.
\item It allows a simple application of the powerful mathematical tools of calibrated geometry \cite{Gaillard:2008wt},
see section \ref{mathviewpointzz}.
\end{itemize}
On the negative side:
\begin{itemize}
\item Obviously, if we require the flavor branes to be smeared, we are limiting ourselves to considering a very
particular subset of all the possible flavored theories. In particular, we require the superpotentials to
effectively recover (some of) the global symmetries of the theory without flavors.
\item Related to the previous point,
one cannot realize, in general, theories with $U(N_f)$ flavor groups. Since the flavor branes are required to
sit at different points in the internal space, the typical string connecting different flavor branes is heavy
and the flavor symmetry is typically broken to $U(1)^{N_f}$ (one may also interpret the
solutions as having flavor symmetry $U(k)^{N_f/k}$ for some $k\ll N_f$).
From the point of view of the field theory, this amounts to having a Veneziano expansion
with ``one window graphs'', as pointed out in \cite{Bigazzi:2008zt}.
In principle, this fact can hinder the realization of some interesting physical features in the dual
set-ups considered.
\end{itemize}
\subsection{Localized sources and other approaches}
\label{sec:localized}
As already remarked, this review focuses on solutions of string theory for which
there are D-brane sources homogeneously smeared over a given family of possible
embeddings, and that can be interpreted as duals of strongly coupled gauge
theories in the Veneziano limit. As stressed above, this is a very particular
subset of all the possible duals of theories with flavor. In a generic
case, one
should consider the sources to be localized at certain positions, such that the
density of charge is given by a sum over Dirac delta functions. Such generic case
is technically more challenging.
However, remarkable works along these lines have appeared, pursuing solutions
with the flavor branes localized at a single point of space (notice this is not
the most general case either).
We will not review them
in any detail here, but the goal of this section is to provide a brief guide
to the literature on the subject.
The main ingredient of this approach consists of finding solutions of
supergravity which can be interpreted as intersections of branes of different
dimension, with each stack of branes localized at a fixed position of space-time.
In the context of gauge-gravity duality, the search for such solutions was initiated
in \cite{Kehagias:1998gn}, \cite{Aharony:1998xz}. These papers discussed D3-D7 intersections,
which have been the most studied in the literature (see below for
different set-ups). A lot of progress was reported in subsequent years
\cite{Grana:2001xn}, \cite{Bertolini:2001qa}, \cite{Bertolini:2002xu}.
Among other aspects, these papers presented a better understanding of the
D3-D7 solutions,
the inclusion of fractional branes and clear matchings with field theory issues
such as the running of couplings and anomalies.
Further work on the D3-D7 localized system was performed in \cite{Burrington:2004id}
(where the conifold was also addressed),
\cite{Liu:2004ru} (where D7 brane backreaction on bubbling geometries was considered) and
\cite{Kirsch:2005uy}, where the solution corresponding to D3-D7 in flat space
was completed by providing an expression for
the warp factor in closed form.
It is also worth mentioning \cite{Ouyang:2003df} where a flavor D7-brane
in a cascading theory was considered and its backreaction introduced
as a perturbation. The finite temperature generalization of
the background of \cite{Ouyang:2003df} was discussed in \cite{Mia:2009wj}.
Let us now outline the literature on D2-D6 localized intersections, which can be interpreted
as duals of 2+1 supersymmetric gauge theories coupled to fundamentals introduced
by the D6-branes. The construction of the type IIA solutions (and their relation
to M-theory) was carried out in \cite{Itzhaki:1998uz}, \cite{Cherkis:2002ir}, \cite{cherkishash}.
In \cite{Erdmenger:2004dk}, meson excitations of this background were discussed and,
in particular,
the holographic dictionary relating meson-like operators
to certain (closed string) supergravity modes was presented.
On the other hand, the authors of \cite{GomezReino:2004pw} found a finite temperature
version of the solution, which was used to discuss the thermodynamics of the system.
Very recent progress in the D2-D6 systems, their M-theory uplifts and the detailed
relation to Chern-Simons theories with flavor has been reported in
\cite{gaiottoetal}.
Regarding D4-D8 intersections,
localized solutions in that
set-up were constructed in \cite{Nastase:2003dd} in an early attempt to build a QCD dual.
In the context of the Sakai-Sugimoto model \cite{Sakai:2004cn}, backreaction from
localized D8-$\bar{\rm {D}8}$ branes was analysed in \cite{Burrington:2007qd}.
It is also worth mentioning recent solutions in heterotic string theory which
were argued to be related to flavored theories \cite{Carlevaro:2009jx}.
\vskip.25cm
Interestingly, there are a few papers in which
similar situations were considered in subcritical string theory
and therefore defined in dimensions lower than ten.
In many of these cases,
each flavor brane fills the
whole space-time (therefore they
are not localized, neither
smeared).
Some physics can then be extracted by using exact string theory methods
but what these models have in common is that it is not possible to handle
them
within a well-controlled gravity description:
gravity-like actions with just two derivatives suffer curvature corrections which cannot
be neglected, nor consistently
computed. However,
there is the hope that the two-derivative actions
can nevertheless provide additional non-trivial insights in the
physics of the system.
This idea was put forward by Klebanov and Maldacena in \cite{Klebanov:2004ya},
who considered a D3-D5 system in a six dimensional background (the cigar).
Such a system is dual to 4D ${\cal N}=1$ SQCD as was shown using exact worldsheet methods
in \cite{Fotopoulos:2005cn,Murthy:2006xt}.
For a recent discussion on the dual to the flavor singlet sector of
${\cal N}=2$ superconformal QCD in a subcritical string framework, see
\cite{Gadde:2009dj}.
The set-up of \cite{Klebanov:2004ya} was generalized to different situations in \cite{Alishahiha:2004yv},
\cite{Bigazzi:2005md},
\cite{Casero:2005se}.
The finite temperature physics of a model in \cite{Bigazzi:2005md} was
analysed in \cite{Bertoldi:2007sf}.
Bottom-up approaches
(in which a high-dimensional gravity theory is proposed to describe
some specific features of QCD) with space-time filling flavor branes
have been discussed in \cite{Gursoy:2007cb}, \cite{Sin:2007ze}.
Recently, a bottom-up approach to the
conformal window along these lines has
appeared
\cite{Jarvinen:2009fe}.
\vskip.25cm
Finally, let us mention a recent contribution by Armoni
\cite{Armoni:2008jy}, in which a way of departing from the quenched approximation
was proposed. The fermion determinant is expanded in terms of Wilson loops. It then
turns out that a sum of correlators of an observable with the Wilson loops boils
down to an expansion in $\frac{N_f}{N_c}$, which can in principle be computed.
It would be nice to further develop possible implications of this
observation in holographic set-ups.
\setcounter{equation}{0}
\section{Flavor deformations of $AdS_5\times X^5$}
\label{AdS5X5}
Our first concrete application of the procedure described
will be the flavor deformation of $AdS_5\times S^5$.
This is the simplest possible case and,
hopefully, it will neatly illustrate the comments of section \ref{section1}.
In fact, for most of the discussion,
the formalism applies to any $AdS_5\times X^5$ geometry, $X^5$ being a five-dimensional
compact Sasaki-Einstein (SE) space, so we will refer to this more general case during this
whole section. At some points, we will use the two notable examples $X^5 = S^5$ or $T_{1,1}$
to clarify particular issues.
Let us start with a general comment. Since the $AdS_5\times S^5$
theories without flavor are conformal, we expect that once we include extra matter, a positive
beta function is generated. This is in fact the case and leads to the appearance of a Landau
pole. Nevertheless, as in QED,
the theory renders meaningful IR physics
even if the UV is ill-defined, as long as the IR and UV are well separated scales.
However, this separation does not allow to have $N_f$ and $N_c$ of the same order.
As we will see, one can define a parameter $\epsilon \sim \lambda N_f/N_c$
which weighs the internal
flavor
loops and that has to be kept small. The effect of the unquenched quarks can then be computed as
an expansion in $\epsilon$.
After introducing the framework in sections \ref{sec21}, \ref{sec22},
we present the unquenched supersymmetric
(${\cal N}=1$ in 4d) solutions in \ref{sec23}. In section \ref{sec:screening},
we present an instance of the effects of the unquenched flavors on a physical quantity,
namely on the mass of a particular meson tower.
Then, in section \ref{D3D7plasma}, we break supersymmetry by turning on temperature and analyse
the physics of the dual quark-gluon plasma. We end in section \ref{sec:rangeval} by discussing the range
of validity for the solutions and approximations used.
\subsection{The geometries and field theories without flavors}
\label{sec21}
The models we discuss here
are obtained by placing a stack of $N_c$ D3-branes at the origin of the six-dimensional cone over $X^5$. The corresponding type IIB background reads:
\begin{eqnarray}
&&ds^2\,=\,\big[\,h(r)\,\big]^{-1/2}\,dx^2_{1,3}\,+\,\big[\,h(r)\,\big]^{1/2}\,\big[\,dr^2\,+\,r^2\,ds^2_{X^5}\,\big]\,\,,\nonumber\\\rc
&&F_5\,=\,dh^{-1}\,dx^0\wedge\cdots\wedge dx^3\,+\,{\rm Hodge\,\, dual}\,\,,\nonumber\\\rc
&&h(r)\,=\,\frac{Q_c}{ 4r^4}\,\,,\qquad\qquad Q_c\,\equiv\,\frac{(2\pi)^4\,g_s\,\alpha'^2\,N_c}{{\rm Vol}(X^5)}\,\,,
\label{AdS5X5back}
\end{eqnarray}
where we have taken the near horizon limit. The dilaton is
constant and all the other fields of type IIB supergravity vanish.
In general the metric of the SE space $X^5$ can be written as a Hopf fibration over a four-dimensional K\"ahler-Einstein (KE) manifold:
\begin{equation}
ds^2_{X^5}\,=\,ds^2_{KE}\,+\,\big(\,d\tau\,+\,A_{KE}\,\big)^2\,\,,
\label{SEmetric}
\end{equation}
where $\tau$ is the fiber and $A_{KE}$ is the connection one-form whose exterior derivative gives the K\"ahler form $J_{KE}$ of the KE base:
\begin{equation}
dA_{KE}\,=\,2J_{KE}\,\,.
\label{A-J-KE}
\end{equation}
Let us first consider the particular case in which $X^5$ is the five-sphere $S^5$. In this case the KE base is the manifold $CP^2$ (with the Fubini-Study metric) and the space transverse to the color branes, with metric $dr^2+r^2\,ds^2_{S^5}$, is just
$\mathbb{R}^6$. When $X^5=S^5$ the coefficient $Q_c$ appearing in (\ref{AdS5X5back}) is just $Q_c\,=\,16\pi g_s\alpha'^2\,N_c$. Moreover, as is well-known, the field theory dual to the $AdS_5\times S^5$ background is
${\cal N}=4$ SYM in 4d, which, in ${\cal N}=1$ language, can be written in terms of a vector multiplet and of three chiral superfields $\Phi_i$ ($i=1,2,3$) transforming in the adjoint representation of the gauge group and interacting by means of the cubic superpotential:
\begin{equation}
W_{{\cal N}=4}\,=\,{\rm Tr}\,\Big[\,\Phi_1\,[\Phi_2,\Phi_3]\Big]\,\,.
\end{equation}
If we represent the transverse $\mathbb{R}^6$ of the $AdS_5\times S^5$ solution in terms of three complex variables $Z_i$ ($i=1,2,3$), one can regard the $Z_i$'s as the geometric realization of the adjoint superfields $\Phi_i$.
The second prominent example which we will analyze in detail is the one in which $X^5$ is the $T^{1,1}$ space with metric:
\begin{equation}
ds^2_{T^{1,1}} = \frac16\sum_{i=1}^2[d\theta_i^2+ \sin^2\theta_id\varphi_i^2] + \frac19 [d\psi+\sum_{i=1}^2\cos\theta_id\varphi_i]^2\,,
\label{T11metric}
\end{equation}
where the range of the angles is $\psi \in [0,4\pi)$,
$\varphi_i \in [0,2\pi)$, $\theta_i \in [0,\pi]$. Since ${\rm Vol}(T^{1,1})=16\pi^3/27$, the coefficient $Q_c$ for this solution is $Q_c=27\pi g_s\alpha'^2\,N_c$. In this case the space transverse to the color branes is the conifold, which is a 6d Calabi-Yau cone which can also be described as the locus of the solutions of the algebraic equation:
\begin{equation}
z_1z_2=z_3z_4\,,
\label{conifold-rel}
\end{equation}
where the $z_i$ are four complex coordinates. The relation between these variables and the coordinates used in (\ref{T11metric}) is the following:
\begin{eqnarray}
z_1 = r^\frac32 e^{\frac{i}{2}(\psi - \varphi_1 - \varphi_2)}
\sin\frac{\theta_1}{2}\sin\frac{\theta_2}{2}\,\,,\qquad
z_2 = r^\frac32 e^{\frac{i}{2}(\psi + \varphi_1 + \varphi_2)}
\cos\frac{\theta_1}{2}\cos\frac{\theta_2}{2}\,\,,\nonumber\\
z_3 = r^\frac32 e^{\frac{i}{2}(\psi + \varphi_1 - \varphi_2)}
\cos\frac{\theta_1}{2}\sin\frac{\theta_2}{2}\,\,,\qquad
z_4 = r^\frac32 e^{\frac{i}{2}(\psi - \varphi_1 + \varphi_2)}
\sin\frac{\theta_1}{2}\cos\frac{\theta_2}{2}\,\,.
\label{zetas}
\end{eqnarray}
Notice also that the metric written in (\ref{T11metric}) is of the form (\ref{SEmetric}) where the KE base is just the $S^2\times S^2$ space parameterized by the angles $(\theta_i, \varphi_i)$ and one should make the following identifications:
\begin{eqnarray}
&&\tau\,=\,\psi/3\,\,,\qquad\qquad
A_{T^{1,1}}\,=\,\frac13\,\big(\cos\theta_1\,d\varphi_1\,+\,\cos\theta_2\,d\varphi_2\,\big)\,\,,\nonumber\\\rc
&&J_{T^{1,1}}\,=\,\frac{dA_{T^{1,1}}}{ 2}\,=\,-\frac16\,
\big(\,\sin\theta_1\,d\theta_1\wedge d\varphi_1\,+
\,\sin\theta_2\,d\theta_2\wedge d\varphi_2\,\big)\,\,.
\label{Kahler-conifold}
\end{eqnarray}
The field theory dual to the $AdS_5\times T^{1,1}$ background is
the ${\cal N}=1$ superconformal quiver gauge theory with gauge group $SU(N_c)\times SU(N_c)$ and bifundamental matter fields $A_1, A_2$ and $B_1, B_2$ transforming respectively in the ({\bf $N_c,\bar N_c$}) and in the ({\bf$\bar N_c, N_c$}) representations of the gauge group \cite{kw}, \hbox{\it i.e.}\ the so-called Klebanov-Witten (KW) model. The matter fields form two $SU(2)$ doublets and interact through a quartic superpotential
\begin{equation}
W_{KW} = \hat h\,\epsilon^{ij}\epsilon^{kl}\,{\rm Tr}\, [A_iB_kA_jB_l]\,.
\label{KW-superpotential}
\end{equation}
The fields $A_i$ and $B_i$ can be related to the coordinates $z_i$ by means of the following relations:
\begin{equation}
z_1 = A_1 B_1, \quad z_2 = A_2 B_2, \quad z_3 = A_1B_2, \quad z_4 = A_2B_1 \,,
\label{maps}
\end{equation}
which automatically solve the defining conifold equation (\ref{conifold-rel}).
\subsection{Flavor branes and smeared charge distribution}
\label{sec22}
The flavor branes for the $AdS_5\times X^5$ backgrounds just described are D7-branes extended along the four Minkowski directions as well as along a non-compact submanifold of the cone over $X^5$. The type of flavor that the D7-branes add depends both on the space $X^5$ and on the submanifold they wrap in the transverse space. We first illustrate the situation
with the two examples of $X^5 = S^5$ and $T_{1,1}$ and at the end display the general expressions.
The first instance is
the case in which $X^5=S^5$. In this case a simple kappa symmetry analysis shows that, in order to preserve eight supersymmetries, the D7-branes must be extended along a codimension two hyperplane in $\mathbb{R}^6$ which, in terms of the complex coordinates $Z^i$ can be written as:
\begin{equation}
a_1\,Z^1\,+\,a_2\,Z^2\,+\,a_3\,Z^3\,=\,\mu\,\,,
\label{hol-embedding-S5}
\end{equation}
with the $a_i$ and $\mu$ being complex constants satisfying $\sum_1^3 |a_i|^2 = 1$.
On the field theory side these flavor branes introduce ${\cal N}=2$ fundamental hypermultiplets $(Q^r, \tilde Q_r)$ ($r=1,\cdots N_f$) - nonetheless,
a generic collection of branes
within the family (\ref{hol-embedding-S5}) retains just ${\cal N}=1$ susy.
The corresponding superpotential for an embedding such as the one in (\ref{hol-embedding-S5}) can be written as:
\begin{equation}
W\,=\,W_{{\cal N}=4}\,+\,\tilde Q_r\,\big[\,\sum_j a_j\,\Phi_j\,+\,m\,\big]\,Q^r\,\,,
\end{equation}
where the mass $m$ is related to the constant $\mu$ in (\ref{hol-embedding-S5}).
Notice that since the embeddings are holomorphic, it is not possible to smear them in a way
in which the full $SO(6)$ isometry is realised.
After smearing over the embeddings (\ref{hol-embedding-S5}), one can recover, at most, $SU(3)\times U(1)$,
as will be seen directly from the dual solution.
In the case of the $AdS_5\times T^{1,1}$ background there are two classes of holomorphic embeddings which correspond to different types of flavors in the KW theory. In terms of the $z_i$ coordinates of (\ref{zetas}) the representative embedding of the first class is given by the equation $z_1=\mu$. This is the so-called Ouyang embedding
\cite{Ouyang:2003df}, which has two branches in the massless limit $\mu=0$.
In each of these branches the D7-brane adds fundamental matter to one of the two nodes of the KW quiver and antifundamental matter to the second. The corresponding superpotential contains cubic couplings between the quark fields $q_i$ and $\tilde q_i$ ($i=1,2$) and the bifundamental fields $A_i$ and $B_i$. For example, for the massless embedding $z_1=0$ the superpotential (\ref{KW-superpotential}) is modified as:
\begin{equation}
W_{z_1=0} = W_{KW} + h_1{\tilde q_1}A_1q_2 + h_2 {\tilde q_2} B_1 q_1\,,
\end{equation}
where, here and in the following, traces over color indices and sums over the $N_f$ flavor indices are implied. The second class of D7-brane embeddings is the one giving rise to non-chiral flavors, whose representative element is given by the equation
$z_1-z_2=\mu$. In this case every D7-brane adds fundamental and antifundamental flavor to one node of the KW quiver and the flavor mass terms do not break the classical symmetry of the massless theory. The corresponding superpotential contains only
mass terms and
quartic couplings, namely:
\begin{equation}
W = W_{KW} +\hat h_1\, \tilde q_1 [A_1B_1-A_2B_2 ]q_1 + \hat h_2\,\tilde q_2 [B_1A_1-B_2A_2]q_2 + k_i\,(\tilde q_i q_i)^2 + m\,(\tilde q_i q_i)\,.
\label{KWsuperpquark}
\end{equation}
In order to develop our program and construct backreacted gravity solutions for smeared distributions of flavor branes following ref. \cite{Benini:2006hh}, we should be able to find a family of equivalent embeddings for each type of configuration described above. In the case of the $AdS_5\times S^5$ background eq. (\ref{hol-embedding-S5}) provides such a family.
Notice that, even if each individual embedding of the form (\ref{hol-embedding-S5}) preserves ${\cal N}=2$,
which supersymmetries are preserved depends on the $a_i$'s.
Nevertheless, one can check that all the holomorphic embeddings
of the type (\ref{hol-embedding-S5}) are mutually supersymmetric and,
due to the holomorphic nature of the linear equation (\ref{hol-embedding-S5}), they preserve the same common four supersymmetries (${\cal N}=1$)
for all values of the constants $a_i$. Thus, we can use these constants to parameterize the family of different planes that constitute our continuous distribution of flavor branes.
In the case of the $AdS_5\times T^{1,1}$ background one can generalize the chiral embedding $z_1=\mu$ by acting with the $SU(2) \times SU(2)$ symmetry of the conifold. The corresponding family of embeddings takes the form:
\begin{equation}
\sum_{i=1}^4 \alpha_i z_i = \mu\,,
\label{general_chiral}
\end{equation}
with the complex constants $\alpha_i$ spanning a conifold (up to overall complex rescalings)
\begin{equation}
\alpha_1 \alpha_2 - \alpha_3 \alpha_4 =0\,.
\label{ouyang_eq}
\end{equation}
Notice that embeddings like $z_1-z_2=\mu$ are not in this family. Indeed, the non-chiral embeddings $z_1-z_2=\mu$ can be generalized as:
\begin{equation}
\bar p z_1 - p z_2 + \bar q z_3 + q z_4 = \mu\,,
\label{genembeddingks}
\end{equation}
where $p, q$ span a unit 3-sphere, \hbox{\it i.e.}\ they satisfy $|p|^2+|q|^2\,=1$.
In spite of the differences among the cases presented above, the charge distribution generated by these families of embeddings can be written in a common form. The reason for this universality is the underlying Sasaki-Einstein structure. In order to illustrate this fact, let us consider the chiral embeddings (\ref{general_chiral}) in the case in which the mass parameter $\mu$ is zero. Without loss of generality we can rescale the $\alpha_i$ coefficients and fix $\alpha_1=1$. Then, (\ref{ouyang_eq}) fixes $\alpha_2=\alpha_3\alpha_4$ and, after using (\ref{zetas}), the massless embedding equation
\begin{equation}
z_1 + \alpha_3 \alpha_4 z_2 + \alpha_3 z_3 + \alpha_4 z_4 =0\,,
\label{genmless}
\end{equation}
nicely factorizes as:
\begin{equation}
\left( \sin\frac{\theta_1}{2}
+ \alpha_3 e^{i \varphi_1} \cos \frac{\theta_1}{2}\right)
\left( \sin\frac{\theta_2}{2}
+ \alpha_4 e^{i \varphi_2} \cos \frac{\theta_2}{2}\right)=0\,.
\label{genmlessex}
\end{equation}
Notice that the vanishing of each of the factors in (\ref{genmlessex}) determines a branch in which the branes sit at a fixed point of one of the two two-spheres parameterized by the angles $(\theta_i, \varphi_i)$. The constants $\alpha_3$ and $\alpha_4$ determine the particular point at which each brane is sitting in each $S^2$. Indeed, if $\xi^\alpha_1$ and $\xi^\alpha_2$ are systems of worldvolume coordinates for the D7-branes, these two branches can be written as:
\begin{eqnarray}
&&\xi^\alpha_1 = \{x^0,x^1,x^2,x^3,r,\theta_2,\varphi_2,\psi\}\,\,, \qquad
\theta_1=\theta_1^*\,=\,\text{const.} \,\,,\qquad \varphi_1=\varphi_1^*=\text{const.}, \qquad\nonumber\\\rc
&&\xi^\alpha_2 = \{x^0,x^1,x^2,x^3,r,\theta_1,\varphi_1,\psi\} \,\,,\qquad
\theta_2=\theta_2^*=\text{const.}\,\,, \qquad \varphi_2=\varphi_2^*=\text{const.}
\label{branches}
\end{eqnarray}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.95\textwidth]{smearing.eps}
\end{center}
\caption[smearingfig]{We see on the left side the two stacks of $N_f$ flavor-branes localized on
each of their respective $S^2$'s (they wrap the other $S^2$). The flavor
group is clearly $U(N_f) \times U(N_f)$. After the smearing on the right
side of the figure, this global symmetry is broken to $U(1)^{N_f - 1}\times
U(1)^{N_f -1} \times U(1)_B \times U(1)_A$. \label{smearingfig}}
\end{figure}
In figure \ref{smearingfig} we have represented the two branches
for the embedding (\ref{genmless}).
From the field theory side, which particular embedding we choose determines the coupling
between the associated
quarks and the bifundamentals. Roughly speaking, the contribution to the superpotential
of an embedding determined by some $\alpha_3$, $\alpha_4$ is
$h_1{\tilde q_1}(A_1 + \alpha_4 A_2) q_2 + h_2 {\tilde q_2}
(B_1 + \alpha_3 B_2) q_1$. Thus, when we {\it smear} and sum over all the possible $\alpha_3$ and $\alpha_4$,
both $SU(2)$'s (the one rotating the $A_i$'s and the one rotating the $B_i$'s) is effectively
recovered. Figure \ref{smearingfig} is the geometric interpretation of this effect.
It is straightforward to compute the charge density produced by this localized D7-brane configuration. Indeed, taking into account the contribution of the two branches, one gets:
\begin{equation}
\Omega^{loc}\,=\,\delta^{(2)}\,(\theta_1-\theta_1^{*}\,,\,\varphi_1-\varphi_1^{*})\,
d\theta_1\wedge d\varphi_1\,+\,
\delta^{(2)}\,(\theta_2-\theta_2^{*}\,,\,\varphi_2-\varphi_2^{*})\,
d\theta_2\wedge d\varphi_2\,\,.
\end{equation}
To produce an homogeneous configuration of $N_f$ D7-branes we should distribute in every branch the branes homogeneously along their transverse two-sphere. In the continuum limit $N_f\rightarrow\infty$ this procedure amounts to performing
an integration over each $S^2$ with the corresponding volume element, namely:
\begin{eqnarray}
&&\Omega\,=\, \Big[\,\int{N_f\over 4\pi}\,\sin\theta_1^*\,\,
\delta^{(2)}\,(\theta_1-\theta_1^{*}\,,\,\varphi_1-\varphi_1^{*})\,
d\theta_1^{*}\, d\varphi_1^{*}\,\Big] d\theta_1\wedge d\varphi_1\,+\,\nonumber\\\rc
&&\qquad
+\,\Big[\,\int{N_f\over 4\pi}\,\sin\theta_2^*\,\,
\delta^{(2)}\,(\theta_2-\theta_2^{*}\,,\,\varphi_2-\varphi_2^{*})\,
d\theta_2^{*}\, d\varphi_2^{*}\,\Big] d\theta_2\wedge d\varphi_2\,\,.
\label{Omega-micro}
\end{eqnarray}
The integrations over $\theta_i^{*}$ and $\varphi_i^{*}$ in (\ref{Omega-micro}) can be immediately performed, yielding the following expression for the smeared charge distribution of D7-branes:
\begin{equation}
\Omega\,=\,{N_f\over 4\pi}\,\,\Big(\,\sin\theta_1\,d\theta_1\wedge d\varphi_1\,+\,
\sin\theta_2\,d\theta_2\wedge d\varphi_2\,\Big)\,\,.
\label{Omega-massless-conifold}
\end{equation}
Notice that in (\ref{Omega-micro}) we have included the normalization factor $N_f/ 4\pi$ in such a way that the resulting distribution densities $\sin\theta_i^*\,N_f/ 4\pi\,$ are normalized to $N_f$ when they are integrated over $S^2$. Notice that, as already pointed out above, the flavor symmetry of the smeared configuration is $U(1)^{N_f}$ rather than $U(N_f)$, since the branes are not placed on top of each other. Interestingly, a similar calculation for the embeddings (\ref{genembeddingks}) in the massless case $\mu=0$ gives rise to the same charge density for the smeared configuration as in (\ref{Omega-massless-conifold}) \cite{Benini:2007gx}.
This is because the form of $\Omega$ in (\ref{Omega-massless-conifold}) is determined by the
$SU(2)\times SU(2) \times Z_2$ global symmetry which we want to recover after smearing.
Actually, one can rewrite (\ref{Omega-massless-conifold}) in a form which can be easily generalized to any continuous family of equivalent D7-brane massless embeddings in an arbitrary Sasaki-Einstein manifold. Indeed, by using (\ref{Kahler-conifold}) one can rewrite the right-hand side of (\ref{Omega-massless-conifold}) in terms of the K\"ahler form of $T^{1,1}$ as:
\begin{equation}
\Omega\,=\,-{3N_f\over 2\pi}\,\,J_{T^{1,1}}\,\,.
\label{Omega-T11}
\end{equation}
For an arbitrary Sasaki-Einstein space $X^5$, the expression (\ref{Omega-T11}) generalizes to:
\begin{equation}
g_s\,\Omega\,=\,-2\,Q_f\,J_{KE}\,\,,
\label{Omega-X5}
\end{equation}
where $Q_f$ is the following constant coefficient:
\begin{equation}
Q_f\,=\,{{\rm Vol}(X^3)\,g_s\,N_f\over 4\,{\rm Vol}(X^5)}\,\,.
\label{Qf}
\end{equation}
In (\ref{Qf}) $X^3$ is the compact submanifold of $X^5$ wrapped by the D7-brane in a massless embedding . Notice that in this case the D7-brane worldvolume along the space transverse to the color branes is always of the form $I\times X^3$, where $I$ is a non-compact interval along the holographic radial direction. It is worth noticing that the factor ${\rm Vol}(X^5)/{\rm Vol}(X^3)$ appearing on the right-hand side of (\ref{Qf}) is just the volume transverse to any individual flavor brane, over which we are distributing the D7-branes. For the massless chiral embeddings in the conifold one can readily check, after taking into account the contribution of both branches in (\ref{branches}), that
${\rm Vol}(X^3)=16\pi^2/9$. Since ${\rm Vol}(T^{1,1})=16\pi^3/27$, one can easily prove that (\ref{Omega-X5}) reduces to (\ref{Omega-T11}). In the case $X^5=S^5$ the three-manifold $X^3$ is just a unit $S^3$ and ${\rm Vol}(X^3)=2\pi^2$. Therefore, we obtain the following values of $Q_f$ for $X^5=S^5, T^{1,1}$:
\begin{equation}
Q_f\,=
\begin{cases}
{g_s\,N_f\over 2\pi}\,\, &{\rm for} \,\,\,\,X^5=S^5\,\,,\cr\cr
{3g_s\,N_f\over 4\pi}\,\,&{\rm for} \,\,\,\,X^5=T^{1,1}\,\,.
\end{cases}
\label{Qfs-cases}
\end{equation}
The charge density $\Omega$ determines the ansatz of $F_1$ in the backreacted geometry. Indeed, the WZ part of the D7-brane action contains a term in which the RR eight-form potential $C_8$ is coupled to the D7-brane worldvolume. The continuous limit for this term amounts to performing the following substitution:
\begin{equation}
S_{WZ} = T_7 \sum_{N_f} \int_{{\cal M}_8} \hat C_8\, \rightarrow\, T_7 \int_{{\cal M}_{10}} \Omega\wedge C_8\,\,,
\end{equation}
which leads to the following violation of the Bianchi identity for $F_1$:
\begin{equation}
dF_1 = - 2\kappa_{(10)}^2\,T_7 \Omega = - g_s \Omega\,\,.
\label{gen_bianchi}
\end{equation}
Taking into account the general expression of $\Omega$ for a massless embedding written in (\ref{Omega-X5}), as well as the relation (\ref{A-J-KE}) between the one-form $A_{KE}$ and the Sasaki-Einstein K\"ahler form $J_{KE}$, one is led to adopt \cite{Benini:2006hh} the following ansatz for $F_1$:
\begin{equation}
F_1\,=\,Q_f\,(d\tau+A_{KE})\,\,.
\label{F1-massless}
\end{equation}
A simple modification of (\ref{F1-massless}) for $F_1$ allows us to extend the ansatz to the case in which the quarks are massive \cite{Benini:2006hh}. This modification corresponds to introducing a function $p(\rho)$ of the holographic coordinate $\rho$ and performing the substitution $Q_f\rightarrow Q_f\,p(\rho)$ in (\ref{F1-massless}). This function $p(\rho)$ encodes the effects of the non-trivial profile of the D7-branes. Indeed, when the quarks are massive the brane does not extend along the full range of the radial coordinate $\rho$ and, accordingly, $p(\rho)$ must vanish for $\rho<\rho_q$, where $\rho=\rho_q$ is the radial location of the tip of the D7-brane. Moreover, the function $p(\rho)$ should approach the value $p=1$
when
$\rho \gg \rho_q$ since in this region the quarks are
effectively massless. The form of the function $p(\rho)$ is not universal and depends on the particular embedding of the D7-brane. For the three embeddings in the cases $X^{5}=S^{5}$ and $T^{1,1}$ discussed above, the expressions for $p(\rho)$ are given below. At this point let us simply notice that the charge density $\Omega$ is modified with respect to the massless case as:
\begin{equation}
g_s\,\Omega\,=\,-2\,p(\rho)\,Q_f\,J_{KE}\,-\,Q_f\,\dot p(\rho)\,d\rho\wedge (d\tau + A_{KE})\,\,,
\label{gsomega}
\end{equation}
where the dot denotes derivative with respect to the radial variable $\rho$.
\subsection{Backreacted ansatz and solution}
\label{sec23}
Let us now write an ansatz for the backreacted D3-D7 background for a generic Sasaki-Einstein space $X^5$ \cite{Benini:2006hh}. It is clear from the discussion of the previous subsection that, after performing the smearing, the resulting RR one-form $F_1$ introduces a distinction between the directions of the $U(1)$ fiber and of the KE base of $X^5$. Therefore, it seems clear that the effect of the smeared flavor branes on the metric should be an internal deformation of the $X^5$ in the form of a relative squashing between the KE space and the Hopf
fiber\footnote{Just in the case when $X^5$ is the sphere $S^5$, this squashing breaks part of the isometry
$SO(6) \rightarrow SU(3) \times U(1)$, where $SU(3)$ is the isometry of the K\"ahler-Einstein base $CP^2$.
}.
Accordingly, let us adopt the following ansatz for the metric in Einstein frame:
\begin{equation}
ds^2\,=\,\Big[\,h(\rho)\,\Big]^{-{1\over 2}}\,dx^2_{1,3}\,+\,
\Big[\,h(\rho)\,\Big]^{{1\over 2}}\,\Big[\,
e^{2f(\rho)}d\rho^2\,+\,e^{2g(\rho)}\,ds^2_{KE}\,+\,e^{2f(\rho)}\,
\big(\,d\tau+A_{KE})^2\,\Big]\,,
\label{metrictzero}
\end{equation}
where $g(\rho)$ and $f(\rho)$ are the functions that implement the squashing mentioned above and the function multiplying $d\rho^2$ amounts to choosing a particular radial variable $\rho$ which is convenient for our purposes. Moreover, the dilaton will depend on $\rho$ and the RR forms $F_5$ and $F_1$ have the form:
\begin{equation}
\phi=\phi(\rho)\,, \qquad
F_{5} = Q_c\,(1\,+\,*)\varepsilon(X^5)\,\,,\qquad F_1\,= Q_f\,p(\rho)\,
(d\tau+A_{KE})\,,
\label{dilaton-forms}
\end{equation}
where $\varepsilon(X^5)$ is the volume element of $X^5$ and $Q_c$ and $Q_f$ are written in (\ref{AdS5X5back}) and (\ref{Qf}) respectively. The function $p(\rho)$, whose form depends on the D7-brane embedding, takes into account the effects of massive quarks, as explained above.
Given the ansatz (\ref{metrictzero})-(\ref{dilaton-forms}) one can easily study the supersymmetric variations of the dilatino and gravitino in type IIB supergravity and find the corresponding first-order BPS equations, which ensure the preservation of four supersymmetries. The resulting equations are \cite{Benini:2006hh}:
\begin{eqnarray}
&&\partial_\rho g= e^{2f-2g} \,\,,\qquad
\qquad\qquad\quad\
\qquad\partial_\rho f = 3-2 e^{2f-2g} - \frac{Q_f}{2}p(\rho)
e^\phi\,\,,\nonumber\\\rc
&&\partial_\rho \phi = Q_f p(\rho) e^\phi \,\,,\qquad
\qquad\qquad\qquad
\partial_\rho h=-Q_c e^{-4g}\,\,.
\label{BSPsysKW}
\end{eqnarray}
Remarkably, the system (\ref{BSPsysKW}) can be integrated analytically for any function
$p(\rho)$. In order to present this solution, let us define the function $\eta(\rho)$ as follows:
\begin{equation}
\eta(\rho) = Q_f e^\phi \int_{\rho_q}^\rho e^{6\xi} p(\xi) d\xi\,\,,
\label{eta}
\end{equation}
where $\rho_q$ is the value of the radial coordinate at the tip of the flavor brane ($p(\rho<\rho_q)=0$). Then, we can write down quite simple expressions for $f,g,\phi$, namely:
\begin{eqnarray}
e^{-\phi} &=& e^{-\phi_*} -Q_f \int_{\rho_*}^\rho p(\xi) d\xi\,\,,\nonumber\\\rc
e^g &=& c_2 e^\rho e^{-\frac{\phi}{6}} \left(1+e^{-6\rho} (c_1 e^\phi + \eta) \right)^{\frac16}\,\,,\nonumber\\\rc
e^f &=& c_2 e^\rho e^{-\frac{\phi}{6}} \left(1+e^{-6\rho} (c_1 e^\phi + \eta) \right)^{-\frac13}\,\,,
\label{general-sol}
\end{eqnarray}
where we have introduced a reference scale $\rho_*$ and we have defined $\phi_*=\phi(\rho=\rho_*)$. Notice that the warp factor $h$ can be obtained as the integral of $e^{-4g}$, as follows from the last equation in the BPS system (\ref{BSPsysKW}). In (\ref{general-sol}) $c_1$ and $c_2$ are integration constants that we now fix.
First, if we demand
IR regularity of the solution, we need $g=f$ when $\rho \leq \rho_q$.
Since $\eta$ vanishes at
$\rho= \rho_q$, we need $c_1=0$. Moreover, the constant $c_2$ is just some overall scale and has no physical meaning. It is natural
to fix it to $\alpha'^{\frac12} e^{\frac{\phi_*}{6}}$ in order to give appropriate dimensions and to recover the usual expression for the metric when $Q_f=0$. Therefore, we find:
\begin{eqnarray}
e^{\phi-\phi_*} &=& \frac{1}{1 -e^{\phi_*} Q_f \int_{\rho_*}^\rho p(\xi) d\xi }\,\,,\nonumber\\\rc
e^g &=& \sqrt{\alpha'}\,\, e^\rho e^{-\frac{\phi-\phi_*}{6}} \left(1+e^{-6\rho} \eta \right)^{\frac16}\,\,,\nonumber\\\rc
e^f &=&\sqrt{ \alpha'} \,\,e^\rho e^{-\frac{\phi-\phi_*}{6}} \left(1+e^{-6\rho} \eta \right)^{-\frac13}\,\,.
\label{general-sol-fixed}
\end{eqnarray}
Notice that, when $Q_f=0$, we recover the unflavored $AdS_5\times X^5$ background. Indeed, in this case $\phi=\phi_*$ and $\eta=0$ and, after performing the change of the radial variable $r=\sqrt{\alpha'}\,e^{\rho}$, we get that $e^g=e^f=r$ and the background (\ref{metrictzero})-(\ref{dilaton-forms}) coincides with the one written in (\ref{AdS5X5back}).
Let us now introduce the following parameter
\begin{equation}
\epsilon_*\,\equiv\,Q_f\,e^{\phi_*}\,\,,
\label{epsilon_*}
\end{equation}
which, as we will see in a while, controls the effects of quark loops in the backreacted supergravity solution. Indeed, the gauge/gravity dictionary for the type of theories we are studying relates the exponential of the dilaton to the Yang-Mills coupling constant. For example, for the (flavored) ${\cal N}=4$ $SU(N_c)$ theory, dual to the deformed $AdS_5\times S^5$ background, the gauge coupling is
$g_{YM}^2=4\pi\,g_s e^\phi$ and, thus, the 't Hooft coupling at the scale $\rho_*$ is given by:
\begin{equation}
\lambda_*=4\pi\,g_s N_c e^{\phi_*}\,\,.
\label{lambda_*}
\end{equation}
For the quiver theories that correspond to different $X^5$ geometries, the gauge groups are of the form $SU(N_c)^n$. Let us generalize a relation from the orbifold constructions
$\sum_i^n 4\pi g_{YM,i}^{-2}= (g_s e^\phi)^{-1}$ \cite{kw,Lawrence:1998ja},
and consider all the gauge couplings
$g_{YM,i}$ to be equal. Then $4\pi\,g_s N_c e^{\phi}$, strictly speaking, gives the 't Hooft coupling at each node of the quiver, divided by $n$. However, with an abuse of language we will simply refer to it as the 't Hooft coupling. Therefore, by using (\ref{lambda_*}) and the definition of $Q_f$ in
(\ref{Qf}) in (\ref{epsilon_*}), we get:
\begin{equation}
\epsilon_* = \frac{Vol(X^3)}{16\pi\,Vol(X^5)}\lambda_* \frac{N_f}{N_c}\,\,.
\label{epsstar}
\end{equation}
In particular, when $X^5=S^5$ this relation becomes:
\begin{equation}
\epsilon_{*\,(X^5=S^5)} = \frac{1}{8\pi^2}\lambda_* \frac{N_f}{N_c}\,\,.
\end{equation}
Notice that the fact that $\phi$ is not constant in the backreacted solution is simply a reflection, in the gauge theory dual, of the running of the Yang-Mills coupling constant when matter is added to a conformal theory.
In terms of $\epsilon_*$ the dilaton and the function $\eta$ of (\ref{eta}) take the form:
\begin{equation}
e^{\phi-\phi_*}\,=\,{1\over 1-\epsilon_*\,\int_{\rho_*}^{\rho}\,
p(\xi)\,d\xi}\,\,,
\qquad\qquad
\eta\,=\,\epsilon_*\,e^{\phi-\phi_*}\,\,
\int_{\rho_q}^{\rho}\,\,e^{6\xi}\,p(\xi)\,d\xi\,\,.
\label{dilaton-eta-epsilon*}
\end{equation}
One of the prominent features of our solution is the fact that, for $N_f\not=0$, the dilaton blows up at some UV scale $\rho=\rho_{LP}$, determined by the condition:
\begin{equation}
\int_{\rho_*}^{\rho_{LP}}\,p(\xi)\,d\xi\,=\,\epsilon_*^{-1}\,\,.
\end{equation}
Clearly, in order to have a well-defined solution, we should restrict the value of the radial coordinate $\rho$ to the range $\rho<\rho_{LP}$. In view of the relation between the Yang-Mills coupling $g_{YM}$ and the dilaton ($g_{YM}^2\sim e^{\phi}$), the divergence of $\phi$ implies that $g_{YM}$ blows up at some UV scale, \hbox{\it i.e.}\ that the gauge theory develops a Landau pole. This UV pathology of our solution was expected on physical grounds since the flavored gauge theory has positive beta function. Indeed, we will check below in some particular case that our solution reproduces the running of the coupling constant of the dual field theory.
\subsubsection{The supersymmetric solution with massless quarks}
\label{masslessKW}
We now consider the particular case of massless quarks, which corresponds to taking the charge distribution given by (\ref{Omega-X5}) or simply $p(\rho)=1$. In this case (\ref{dilaton-eta-epsilon*}) simply gives:
\begin{equation}
e^{\phi-\phi_*}\,=\,{1\over 1+\epsilon_*\,(\rho_*-\rho)}\,\,,\qquad\qquad
e^{-6\rho}\,\eta\,=\,{\epsilon_*\over 6}\,e^{\phi-\phi_*}\,\,,
\label{masslessSUSY-phi-eta}
\end{equation}
and the solution written in (\ref{general-sol-fixed}) reduces to:
\begin{eqnarray}
e^{g}&=&\sqrt{\alpha'}\,\, e^\rho\,\left(1+\epsilon_* \Big(\frac16 +\rho_*-\rho\Big)\right)^\frac16\,\,,\nonumber\\\rc
e^{f}&=&\sqrt{\alpha'}\,\, e^\rho\,\left(1+\epsilon_* (\rho_*-\rho)\right)^\frac12
\left(1+\epsilon_* \Big(\frac16 +\rho_*-\rho \Big)\right)^{-\frac13}\,\,,\nonumber\\\rc
\frac{dh}{d\rho} &=& -{Q_c \over\alpha'^{2}}\,\,\, e^{-4\rho} \left(1+\epsilon_* \Big(\frac16 +\rho_*-\rho \Big)\right)^{-\frac23}\,\,.
\label{susysol}
\end{eqnarray}
Notice that the location of the Landau pole in this case is just $\rho_{LP}=\rho_*+\epsilon_*^{-1}$ and that the range of $\rho$ for which the solution (\ref{susysol}) makes sense is $\rho\in (-\infty, \rho_{LP})$. Moreover, by using the definition of $\epsilon_*$ in (\ref{epsilon_*}) one can immediately show that the dilaton can be written as:
\begin{equation}
e^{\phi(\rho)}\,=\,{1\over Q_f\,(\,\rho_{LP}-\rho\,)}\,\,.
\label{dilaton-rhoLP}
\end{equation}
Let us now verify that the dependence on $\rho$ of $\phi$ in (\ref{dilaton-rhoLP}) matches the expectations from field theory. For concreteness we will consider the case of ${\cal N}=4$ SYM with matter. Similar checks can be done in other cases (see \cite{Benini:2006hh} for the case of the Klebanov-Witten theory). By using the relation between the Yang-Mills coupling and the dilaton discussed above, as well as the value of $Q_f$ for $X^5=S^5$ written in (\ref{Qfs-cases}), one gets:
\begin{equation}
{8\pi^2\over g^2_{YM}}\,=\,N_f\,(\rho_{LP}-\rho)\,\,.
\label{g-rho}
\end{equation}
In order to read the running of the coupling constant from (\ref{g-rho}) we must convert the dependence on the coordinate $\rho$ in (\ref{g-rho}) into a dependence on the energy scale of the corresponding dual field theory. At an energy scale $\mu$ much lower that the Landau pole scale $\Lambda_{UV}$ (\hbox{\it i.e.}\ for $\rho<<\rho_{LP}$) the scaling dimensions of the adjoints and fundamentals take their canonical values and the natural radius/energy relation is:
\begin{equation}
\rho_{LP}\,-\,\rho\,=\,\log{\Lambda_{UV}\over \mu}\,\,.
\end{equation}
Plugging this relation in (\ref{g-rho}) we get:
\begin{equation}
{8\pi^2\over g^2_{YM}}\,=\, N_f\,\log{\Lambda_{UV}\over \mu}\,.
\end{equation}
Therefore, we get a logarithmic scaling of the coupling of the type
$8\pi^2/ g^2_{YM}=b\log E$, with $b=-N_f$, which matches the one-loop field theory
result\footnote{In principle, one could object that, being strongly coupled, the matter fields
could get large anomalous dimensions making this result suspicious. However,
since we
are performing a small perturbative (in $\epsilon_*$) deformation of the unflavored
backgrounds, the anomalous dimensions for the fundamental multiplets cannot differ
much from their quenched values. For the $X^5=S^5$ case, those anomalous dimensions
vanish.
We thank F. Bigazzi for stressing this point to us.}
in which one has that $b=3N_c-3N_c-N_f$.
In order to have a clearer understanding of the deformation of the $AdS_5\times S^5$ metric introduced by the flavor, it is very convenient to change to a new radial variable $r$, which is defined by requiring that the warp factor takes the same form as in the unflavored case (see (\ref{AdS5X5back})):
\begin{equation}
h\,=\,{R^4\over r^4}\,\,,\qquad\qquad
R^4\,=\,{Q_c\over 4}\,\,.
\label{hR}
\end{equation}
By integrating the last equation in (\ref{susysol}) we can get $h(\rho)$ and thus $r(\rho)$. We will perform this integration order by order in a series expansion in powers of $\epsilon_*$. The additive integration constant will be fixed by requiring that $r(\rho_*)\equiv r_*=\sqrt{\alpha'}\,e^{\rho_*}$. One gets:
\begin{eqnarray}
r&=&\alpha'^{\frac12} e^\rho\Big[1+\frac{\epsilon_*}{72} \Big(
e^{4\rho-4\rho_*}-1 +12(\rho_*-\rho)\Big)
+\frac{5\epsilon_*^2}{10368}\Bigl(e^{8\rho-8\rho_*} + 6 e^{4\rho-4\rho_*}(3+4(\rho_*-\rho))\nonumber\\
&&- (19 -24 (\rho_*-\rho) +144 (\rho_*-\rho)^2) \Bigr) + O(\epsilon_*^3)
\Big]\,\,.
\end{eqnarray}
It is now straightforward to obtain the functions
$f(r)$, $g(r)$ and the dilaton $\phi(r)$ as expansions in powers of $\epsilon_*$. Up to second order we have:
\begin{eqnarray}
e^f&=&r\Big[1-\frac{\epsilon_*}{24}(1+\frac13 \frac{r^4}{r_*^4})+
\frac{\epsilon_*^2}{1152}\left(17-\frac{94}{9}\frac{r^4}{r_*^4}+\frac59\frac{r^8}{r_*^8}
-48 \log(\frac{r}{r_*})\right)+ O(\epsilon_*^3)\Big]\,\,,\nonumber\\
e^g&=&r\Big[1+\frac{\epsilon_*}{24}(1-\frac13 \frac{r^4}{r_*^4})+
\frac{\epsilon_*^2}{1152}\left(9-\frac{106}{9}\frac{r^4}{r_*^4}+\frac{5}{9}\frac{r^8}{r_*^8}
+48 \log(\frac{r}{r_*})\right)+ O(\epsilon_*^3)\Big]\,\,,\nonumber\\
\phi&=&\phi_*+ \epsilon_* \log\frac{r}{r_*} + \frac{\epsilon_*^2}{72}\left(1-\frac{r^4}{r_*^4}
+12 \log\frac{r}{r_*} + 36 \log^2\frac{r}{r_*}\right)+ O(\epsilon_*^3)\,\,.
\label{susymasslesssol}
\end{eqnarray}
Eq. (\ref{susymasslesssol}) neatly displays the effects of quark loops in the deformation of the geometry and in the running of the dilaton (the latter is related to the running of the gauge coupling, as argued above). It is important to point out that the deformed geometry has a curvature singularity at the origin $r=0$ (or $\rho=-\infty$)
(this singularity is similar to the one that appears at $r=0$ in a 2-dimensional manifold with
metric $ds^2 = dr^2 + r^2 (1+r) d\varphi^2$). In the same IR limit, $e^\phi$ runs to $0$.
As argued in section \ref{heuristiczzz}, the appearance of this singularity can
be intuitively understood as due to the fact that, in this massless case, all branes of our smeared distribution pass through the origin and the charge density is highly peaked at that point.
From the field theory side, one can think of the singularity as appearing because the
theory becomes IR free, as first pointed out in \cite{Aharony:1998xz}.
Consistently with these interpretations and with the heuristic picture of
section \ref{heuristiczzz}, the IR singularity can be easily cured by giving a mass to the quarks
(it is a ``good" singularity according to the criteria of \cite{Maldacena:2000mw},\cite{Gubser:2000nd}).
We will explicitly verify this fact in the next subsection.
\subsubsection{The supersymmetric solution with massive quarks}
\label{massiveKW}
Let us now find the backreacted supergravity solution for massive quarks. As mentioned above, the function $p(\rho)$ entering the ansatz for $F_1$ in this case is not universal and depends on the particular
Sasaki-Einstein space $X^5$ and on the family of D7-brane embeddings chosen.
For concreteness we first concentrate in discussing the case in which $X^5=S^5$. The calculation of the function $p(\rho)$ in this case was performed in appendix C of \cite{Bigazzi:2009bk}. If
$|\mu|=e^{\rho_q}$, one has:
\begin{equation}
p(\rho)\,=\,\Big[\,1\,-\,e^{2(\rho_q-\rho)}\,\Big]^2\,\Theta(\rho-\rho_q)\,\,.
\end{equation}
When $\rho\ge \rho_q$ the function $p(\rho)$ is non-vanishing and one has to perform the integrals appearing in (\ref{dilaton-eta-epsilon*}) . These integrals can be straightforwardly done in analytic form and yield the result:
\begin{eqnarray}
e^g&=&\sqrt{\alpha'}\,\, e^\rho\,\left(1+\epsilon_* \Big(\frac16 +\rho_*-\rho-\frac16 e^{6\rho_q-6\rho}
-\frac32 e^{2\rho_q-2\rho} + \frac34 e^{4\rho_q-4\rho}-\frac14 e^{4\rho_q-4\rho_*}+ e^{2\rho_q-2\rho_*}
\Big)\right)^\frac16\,\,,
\nonumber\\\rc
e^f&=&\sqrt{\alpha'}\,\, e^\rho\,\frac{\left(1+\epsilon_* (\rho_*-\rho-e^{2\rho_q-2\rho}+
\frac14 e^{4\rho_q-4\rho}
+e^{2\rho_q-2\rho_*}-
\frac14 e^{4\rho_q-4\rho_*}
)\right)^\frac12}
{\left(1+\epsilon_* (\frac16 +\rho_*-\rho-\frac16 e^{6\rho_q-6\rho}
-\frac32 e^{2\rho_q-2\rho} + \frac34 e^{4\rho_q-4\rho}-\frac14 e^{4\rho_q-4\rho_*}+ e^{2\rho_q-2\rho_*}
)\right)^{\frac13}}
\,\,,
\nonumber\\\rc
\phi&=& \phi_* -\log\Big[1+\epsilon_*\, \Big(\rho_*-\rho-e^{2\rho_q-2\rho}+
\frac14 e^{4\rho_q-4\rho}
+e^{2\rho_q-2\rho_*}-
\frac14 e^{4\rho_q-4\rho_*}\Big)\Big]\,\,.
\label{susysolmassive}
\end{eqnarray}
As a check, notice that setting $\rho_q \rightarrow -\infty$ one recovers the massless
solution of (\ref{masslessSUSY-phi-eta}) and (\ref{susysol}).
We still have to write the solution for $\rho<\rho_q$. In this case $p(\rho)$ vanishes and
the dilaton is constant and, by continuity, it has the value that can be read from
(\ref{susysolmassive}) inserting $\rho=\rho_q$:
\begin{equation}
\phi_{IR}=\phi_q = \phi_* -\log(1+\epsilon_*\, (\rho_*-\rho_q-
\frac34
+e^{2\rho_q-2\rho_*}-
\frac14 e^{4\rho_q-4\rho_*}))\,\,.
\end{equation}
The functions $f$ and $g$ are equal and given by:
\begin{equation}
e^{f}=e^{g}=\alpha'^{\frac12}\, e^\rho e^{-\frac16(\Phi_{IR}-\Phi_*)}\,\,,\qquad\qquad
(\rho<\rho_q)\,\,.
\end{equation}
It follows straightforwardly from these results that the IR singularity at $\rho=-\infty$ of the massless case disappears when $\mu\not=0$ since the background reduces to $AdS_5\times S^5$ for $\rho<\rho_q$. Moreover, one can verify that the metric is also regular at $\rho=\rho_q$. Thus, as stressed in section \ref{heuristiczzz}, the smearing of massive flavors allows one to smooth out IR singularities.
Similar calculations can be done for the conifold theories. In this case we redefine the parameter $\mu$ of the embedding equations (\ref{general_chiral}) and (\ref{genembeddingks}) as $|\mu|=e^{3\rho_q\over 2}$. The charge distribution for the family
(\ref{general_chiral}) of chiral embeddings was obtained in ref. \cite{Bigazzi:2008zt}, with the result:
\begin{equation}
p(\rho)\,=\,\Big[\,1\,-\,e^{3(\rho_q-\rho)}\,\big(1+3\rho-3\rho_q\big)\,\Big]\,
\Theta(\rho-\rho_q)\,\,.
\end{equation}
Similarly,
for the non-chiral embeddings (\ref{genembeddingks}) the function $p(\rho)$ is given by \cite{fkw2}:
\begin{equation}
p(\rho)\,=\,\Big[\,1\,-\,e^{3(\rho_q-\rho)}\,\Big]\,\Theta(\rho-\rho_q)\,\,.
\end{equation}
The corresponding supergravity solutions have been written down in refs. \cite{Bigazzi:2008zt,fkw2}. They are regular in the IR, much in the same way as in the $X^5=S^5$ case detailed above.
\subsection{Screening effects on the meson spectrum}
\label{sec:screening}
The holographic theories with flavors present mesonic excitations, meaning
that there exists a spectrum of colorless physical states created by operators
which are bilinears in the fundamental fields.
They are associated to normalizable excitations of the flavor branes as was neatly explained
in the seminal paper
\cite{Kruczenski:2003be}. For a review of this broad subject, see \cite{Erdmenger:2007cm}.
Notice that the notion of ``meson" we use here generalizes that used
in QCD. For instance, the ``mesons" of \cite{Kruczenski:2003be} are excitations of a non-confining
theory and in this case the dimensionful quantity that sets the meson masses is just the
quark mass (divided by a power of the 't Hooft coupling), not a dynamically generated scale.
In the present section, we review how the presence of unquenched flavors can affect the discrete
mesonic spectrum. Again, we will restrict ourselves to the smeared set-up and follow
\cite{unquenchedmesons}. For discussions about screening effects on the spectrum
in cases with localized rather than
smeared backreacting flavor branes, we refer the reader to \cite{Kirsch:2005uy,Erdmenger:2007cm}.
The effect of the smeared flavors on the hydrodynamical transport coefficients (in a finite temperature
setting) was studied in \cite{Bigazzi:2009bk,BCT}.
It is also worth mentioning that,
within the model we will introduce in section \ref{sec:D5D5},
the screening effects on the glueball spectrum have been recently analyzed in
\cite{Elander:2009bm}.
For the sake of briefness, we will just focus on an example and discuss a
particular
mesonic excitation in the
backreacted Klebanov-Witten model.
The analysis and conclusions for different modes and/or different models should
be similar, see \cite{unquenchedmesons} for some other examples.
In particular, we will consider oscillations of a D7-brane which introduces massive
non-chiral flavor \cite{Kuperstein:2004hy}, and just look at the oscillation of
the gauge field that gives rise to a vector mode in the dual field theory.
Thus, we discuss the physics of a meson whose ``constituent quarks" are massive in the
presence of many dynamical massless flavors.
We write the gauge field along the Minkowski directions as
$A_\mu=a_v(\rho) \,\xi_\mu \, e^{ikx}$, where $\xi_\mu$ is a constant transverse vector.
The equation describing this oscillation was found in \cite{unquenchedmesons},
building on the formalism introduced in \cite{Kuperstein:2004hy}. It reads:
\begin{equation}
0=\partial_\rho\left(e^{2g-3\rho}(e^{3\rho}-e^{3\rho_Q})\partial_\rho a_v \right)
+ M_v^2 h\, e^{2g+2f}
\left(1+ e^{3\rho_Q - 3\rho} \Big(\frac34 e^{2g-2f} -1\Big) \right) a_v\,\,,
\label{eqforav}
\end{equation}
where $M_v^2 = - k^2$, the constant $\rho_Q$ is the minimal value of $\rho$ reached by the D7-brane
(related to the quark mass) and
$f,g,h$ are given in (\ref{susysol}).
Notice that for the meson excitation, we just use a D-brane probe, namely, we consider the oscillation
of a single brane in a fixed background. At first sight, this could look contradictory, since our aim is
always to take into account the effect of the flavor branes on the geometry.
Then, one may think about considering coupled fluctuations of brane and background fields.
Nevertheless, this is not necessary: there are $N_f \gg 1$ flavor branes which are affecting the background
but when we consider a meson, only one (or two) out of this $N_f$ is fluctuating. Therefore, the effect of this
oscillation on the background is suppressed by $N_f^{-1}$ with respect to the contribution of the whole set
of branes and therefore is consistently negligible. On the other hand, the existence of the rest of flavors
- and the associated quantum effects on the spectrum - are taken into account through the deformation they have produced in
the background geometry.
Following the standard procedure \cite{Kruczenski:2003be,Erdmenger:2007cm}, a discrete tower of values for
$M_v$ should be found when selecting solutions of (\ref{eqforav}) which are regular and normalizable.
Since the background has a Landau pole, some prescription is needed for dealing with the UV limit
(large $\rho$). Technically, we will just require that the fluctuation $a_v$ vanishes at $\rho_*$.
Physically, one can check that this is a consistent procedure if $\rho_Q << \rho_*$: we are interested
in some IR physics which should be independent of the UV completion of the theory at $\rho > \rho_*$, up
to corrections suppressed by powers of the UV scale. Namely, we neglect contributions of order
$e^{\rho_Q - \rho_*}\sim \frac{\Lambda_{IR}}{\Lambda_{UV}}$ and check that the spectrum can be written
in terms of IR quantities. The value $\rho_*$ disappears from the final result, apart from the quoted
negligible corrections. See \cite{unquenchedmesons} for further discussions on the issue.
In section \ref{D3D7plasma}, we will see similar examples of how to deal with
the Landau pole. In that case, the IR scale, which has to be much smaller that the arbitrary UV scale
at $\rho_*$, is set by the temperature rather than by the quark mass.
In order to estimate the spectrum from (\ref{eqforav}), we can use a WKB approximation.
In \cite{unquenchedmesons}, using a formalism developed in \cite{Russo:1998by}, an
expression for the mass tower in terms of the principal quantum number $n$ was found.
Adapting notation to the one we are using here:
\begin{equation}
M_v^{(n)} \approx \frac{\pi}{\Sigma_v}\,n\,\,,\qquad\quad
\Sigma_v \equiv \int_{\rho_Q}^{\rho_*} h^{\frac12} e^f \sqrt{\frac
{1+ e^{3\rho_Q - 3\rho} (\frac34 e^{2g-2f} -1)}{1 - e^{3\rho_Q - 3\rho}}}d\rho\,\,.
\end{equation}
Let us evaluate this integral at first order in $\epsilon_*$, by inserting
(\ref{susysol}). We still have to fix the additive constant for $h$, which we can
do by requiring $h(\rho_*)=0$ (in \cite{unquenchedmesons} $h(\rho_{LP})=0$ was used.
It is crucial that both prescriptions give the same result, up to quantities
in $e^{\rho_Q - \rho_*}\sim \frac{\Lambda_{IR}}{\Lambda_{UV}}$).
We shift
to a coordinate $u$ such that $u\equiv e^{\rho-\rho_Q},\ u_* \equiv e^{\rho_*-\rho_Q}$.
Defining $\lambda_Q$ as the 't Hooft coupling (\ref{lambda_*}) at the quark mass scale,
inserting the value of $Q_c$ in (\ref{AdS5X5back})
and defining
$T_Q\equiv \left(e^{\phi/2} \sqrt{-g_{tt} g_{xx}}\right)|_{\rho=\rho_Q}$ as the tension of a hypothetical
fundamental string stretched at constant $\rho=\rho_Q$, we can write the estimate for
the meson masses as:
\begin{equation}
M_v^{(n)} \approx \frac{T_Q^{\frac12}}{\lambda_Q^{\frac14}}\frac{\pi \, n}
{\frac{3^{\frac34}}{4\sqrt{2\pi}} \int_1^{u_*}\left( \frac{\sqrt{4u^3-1}}{u^2 \sqrt{u^3-1}}
+ \epsilon_Q \frac{7-4u^3 + 4 (4u^3-1)\log u}{24u^2 \sqrt{4u^3-1} \sqrt{u^3-1}} \right) du }\,\,.
\label{WKBest}
\end{equation}
It is important to stress once again that this expression is written only in terms of
IR quantities, once we discard terms of order $u_*^{-1} = e^{\rho_Q - \rho_*}$,
namely contributions like $\log u_*$ have cancelled out. Notice that the upper limit of the
integrals can be taken to infinity if we again insist in discarding $O(u_*^{-1})$
contributions. The expression (\ref{WKBest}) is a neat example of how, even having a Landau pole,
the holographic set-up is able to consistently obtain IR predictions, in exactly the same spirit as
in field theory. We can perform numerically the integration in (\ref{WKBest}), and
we get \cite{unquenchedmesons}:
\begin{equation}
M_v^{(n)} \approx \frac{T_Q^{\frac12}}{\lambda_Q^{\frac14}} n\, (5.2 - 6 \times 10^{-3} \frac{N_f\lambda_Q}{N_c}
+ \dots )\,\,,
\label{nummesons}
\end{equation}
where in order to substitute $\epsilon_Q$ as in (\ref{epsstar}) we have used
$Vol(X^3)= \frac{16}{9}\pi^2$, $Vol(X^5)= \frac{16}{27}\pi^3$. The expression (\ref{nummesons}) is the
result quoted in \cite{unquenchedmesons}, apart from a different factor of 2 in the definition
of $\lambda_Q$.
The lesson we want to take from this
section is that there is a well defined method to obtain the shift produced by the flavor quantum effects
on the meson spectrum (or, eventually, on any physical observable) as an expansion in the parameter
$\epsilon \sim \lambda \,N_f / N_c$
which weighs the flavor loops. Heuristically, it may be useful to think
of the computation leading to (\ref{nummesons}) as (partially)
a strong coupling analogue of the Lamb shift corrections of QED.
\subsection{Black hole solutions: D3-D7 quark-gluon plasmas}
\label{D3D7plasma}
In this subsection we will review the results in \cite{Bigazzi:2009bk}.
We start by
showing how one can find a black hole
solution which includes the backreaction effects due to massless quarks. To perform this analysis
it is more convenient to work with a new radial variable $\sigma$ such that the metric takes the form:
\begin{equation}
ds^2 = h^{-\frac12}\left[-b\ dt^2 + d\vec{x}_3^{\,2}\right] + h^\frac12
\left[\, b\,e^{8g+2f}\, d\sigma^2 + e^{2g}\, ds_{KE}^2 + e^{2f}\, (d\tau + A_{KE})^2\,
\right]\,\,.
\label{black10dmetric}
\end{equation}
Notice that we have introduced a new function $b$ which parameterizes the breaking of Lorentz invariance induced by the non-zero temperature $T$. All functions appearing in the metric (\ref{black10dmetric}), as well as the dilaton $\phi$, depend on $\sigma$. Moreover, the RR field strengths $F_5$ and $F_1$ are given by the ansatz (\ref{dilaton-forms}) with the function $p=1$. We remind the reader that fixing $p=1$ corresponds to taking massless quarks\footnote{In
reference \cite{Bigazzi:2009bk},
the more involved case of massive quarks $p\neq 1$
was also discussed. An extra complication is the necessity of finding the non-trivial
D7-brane embeddings in the backreacted geometry.}.
In this non-supersymmetric case we will not have the first-order BPS equations at our disposal and we will have to deal directly with the second-order equations of motion. Actually, since all the functions we need to compute depend only on the radial coordinate $\sigma$, it is possible to describe the system in terms of a one-dimensional effective action. One can find this effective action by directly substituting the ansatz in the gravity plus branes action (\ref{lagrangianzz}). One gets:
\begin{eqnarray}
S_{eff}&=&\frac{Vol(X^5)V_{1,3}}{2\kappa_{10}^2}\int d\sigma \left(
-\frac12\frac{(\partial_\sigma h)^2}{h^2} +12\,(\partial_\sigma g)^2
+ 8 \,\,\partial_\sigma g\,\,\partial_\sigma f\,-\,
\frac12 (\partial_\sigma \phi)^2+\right.\nonumber\\\rc
&+&\left.
\frac{(\partial_\sigma b)}{2b}
\left( \frac{\partial_\sigma h}{h}+
8 \,\partial_{\sigma}\,g
+ 2 \,\partial_{\sigma}\,f
\right)+
\right.
\nonumber\\\rc
&+&\left.
24\, b\,e^{2f+6g} - 4 b\,e^{4f+4g}\,-\,{Q_c^2\over 2} \frac{b}{h^2}\, -\,
{Q_f^2\over 2}\,b\, e^{2\phi+8g}
\, -\,
4Q_f\,b\,
e^{\phi+6g+2f}\right)\,\,.
\label{effeclagr}
\end{eqnarray}
In (\ref{effeclagr}) $V_{1,3}$ denotes the (infinite) integral over the Minkowski coordinates.
The second derivatives coming from the Ricci scalar have been integrated by parts and, as is customary,
only the angular part of $F_{5}$ is inserted in the $F_{5}^2$ term (otherwise
the $Q_c$ would not enter the effective action since, on-shell,
$F_{5}^2=0$ due to the self-duality condition).
The last term in (\ref{effeclagr}), proportional to $Q_f$, comes from the DBI contribution in (\ref{BIWZaction}). Notice also that the WZ term does not enter
(\ref{effeclagr}) because it does not depend on the metric or the dilaton.
The equations of motion stemming from the effective action (\ref{effeclagr}) are:
\begin{eqnarray}
\partial_\sigma^2(\log b)&=&0\,\,,\nonumber\\\rc
\partial_\sigma^2(\log h)&=&-Q_c^2 \frac{b}{h^2}\,\,,\nonumber\\\rc
\partial_\sigma^2\,g\,&=& -2 b \,e^{4g+4f} + 6 b\,e^{6g+2f} - Q_f \,b\, \,e^{\phi+6g+2f}\,\,,
\nonumber\\
\partial_\sigma^2\,f\,&=& 4b\,e^{4g+4f} - \frac{Q_f^2}{2}\,b\,e^{2\phi+8g}\,\,,
\nonumber\\\rc
\partial_\sigma^2\phi&=& Q_f^2\,b\,e^{2\phi+8g}\, +
4 Q_f\,b\, e^{\phi +6g+2f}\,\,.
\label{bhSeqsmassless}
\end{eqnarray}
It is straightforward to check that these equations solve the full set of Einstein equations
provided the following ``zero-energy'' constraint is also satisfied:
\begin{eqnarray}
0&=&
-\frac12\frac{(\partial_\sigma h)^2}{h^2} +12\,(\partial_\sigma g)^2
+ 8 \,\,\partial_\sigma g\,\,\partial_\sigma f\,-\,
\frac12 (\partial_\sigma \phi)^2+\nonumber\\\rc
&+&
\frac{(\partial_\sigma b)}{2b}
\left( \frac{\partial_\sigma h}{h}+
8 \,\partial_{\sigma}\,g
+ 2 \,\partial_{\sigma}\,f
\right)\,-\,24\, b\,e^{2f+6g} +4 b\,e^{4f+4g}\,+\,
\nonumber\\\rc
&+&
{Q_c^2\over 2} \frac{b}{h^2}\, +\,
{Q_f^2\over 2}\,b\, e^{2\phi+8g}
\, +\,
4Q_f\,b\,
e^{\phi+6g+2f}\,\,.
\label{constraintmassless}
\end{eqnarray}
This constraint can be thought of as the $\sigma\sigma$ component of the Einstein equations or, alternatively,
as the Gauss law from the gauge fixing of $g_{\sigma\sigma}$ in the ansatz
(\ref{black10dmetric}).
Let us now find a solution of the system of equations (\ref{bhSeqsmassless}) and of the ``zero-energy'' constraint (\ref{constraintmassless}) that corresponds to a black hole for the backreacted D3-D7 system. We will require that such a solution is regular at the horizon and tends to the supersymmetric one at energy scales much higher than the black hole temperature $T$. Actually, the biggest advantage of the radial variable $\sigma$ introduced above is that the equations of motion of $b$ and $h$ in (\ref{bhSeqsmassless}) are decoupled from the ones corresponding to the other functions of the ansatz. These decoupled equations can be easily integrated in terms of an integration constant $r_h$ as follows:
\begin{equation}
b=e^{4r_h^4 \,\sigma}
\,\,,\qquad\qquad
h=\frac{Q_c}{4r_h^4}(1-e^{4r_h^4 \,\sigma})\,\,.
\end{equation}
where $\sigma\in (-\infty,0)$. We now define a new radial coordinate $r$ by means of the relation:
\begin{equation}
e^{4r_h^4\,\sigma}=1-\frac{r_h^4}{r^4}\,\,,\qquad\qquad
r\in (r_h, +\infty).
\label{rcoorddef}
\end{equation}
Then, $b$ and $h$ take the form:
\begin{equation}
b\,=\,1\,-\,{r_h^4\over r^4}\,\,,\qquad\qquad
h\,=\,{R^4\over r^4}\,\,,
\label{bh-r}
\end{equation}
with $R^4=Q_c/4$. Notice that $h$ is given by the same expression as in (\ref{hR}). Moreover, it is clear from (\ref{bh-r}) that $r=r_h$ is the position of the horizon and, thus, the extremal limit is attained by sending $r_h$ to zero. In terms of $r$ the metric takes the form:
\begin{equation}
ds^2\,=\,{r^2\over R^2}\,\,\Big[\,\,
\Big(1-{r_h^4\over r^4}\Big)\,dt^2\,+\,d\vec{x}_3^{\,2}\,\Big]\,+\,
{R^2\over r^2}\,\,{e^{8\hat g+2\hat f}\over 1-{r_h^4\over r^4}}\,(dr)^2\,+\,
R^2\,\big[\,e^{2\hat g}\,ds_{KE}^2\,+\,e^{2\hat f}\,
(d\tau+A_{KE})^2\,\big]\,\,,
\end{equation}
where we have defined the functions $\hat f$ and $\hat g$ as follows:
\begin{equation}
e^{\hat f}\,\equiv\,{e^{f}\over r}\,\,,\qquad\qquad
e^{\hat g}\,\equiv\,{e^{g}\over r}\,\,.
\end{equation}
In order to determine completely the background we still have to solve eqs. (\ref{bhSeqsmassless}) and (\ref{constraintmassless}) for $ f$, $g$ and the dilaton $\phi$. We will find this solution by introducing a reference UV scale $r_*$ and by expanding the functions in terms of the parameter $\epsilon_*$ defined in (\ref{epsilon_*}). We will impose that the functions $f$, $g$ and $\phi$ are equal to the SUSY ones of (\ref{susymasslesssol}) when the extremality parameter $r_h$ vanishes. Moreover, we will also require that these functions coincide with those in (\ref{susymasslesssol}) at the UV scale $r_*$. These conditions fix uniquely a solution of (\ref{bhSeqsmassless}) and (\ref{constraintmassless}). Up to second order in $\epsilon_*$ this solution is given by:
\begin{eqnarray}
e^{\hat f}&=&1-\frac{\epsilon_*}{24}\Big(1+ \frac{2r^4-r_h^4}{6r_*^4-3r_h^4}\Big)+
\frac{\epsilon_*^2}{1152}\left(17-\frac{94}{9}\frac{2r^4-r_h^4}{2r_*^4-r_h^4}+
\frac59\frac{(2r^4-r_h^4)^2}{(2r_*^4-r_h^4)^2}+\right.\nonumber\\
&&\left.-\frac89 \frac{r_h^8 (r_*^4-r^4)}{(2r_*^4-r_h^4)^3}
-48 \log\Big(\frac{r}{r_*}\Big)\right)+ O(\epsilon_*^3)\,\,,\nonumber\\
e^{\hat g}&=&1+\frac{\epsilon_*}{24}\Big(1- \frac{2r^4-r_h^4}{6r_*^4-3r_h^4}\Big)+
\frac{\epsilon_*^2}{1152}\left(9-\frac{106}{9}\frac{2r^4-r_h^4}{2r_*^4-r_h^4}+
\frac59\frac{(2r^4-r_h^4)^2}{(2r_*^4-r_h^4)^2}+\right.\nonumber\\
&&\left.-\frac89 \frac{r_h^8 (r_*^4-r^4)}{(2r_*^4-r_h^4)^3}
+48 \log\Big(\frac{r}{r_*}\Big)\right)+ O(\epsilon_*^3)\,\,,\nonumber\\
\phi&=& \phi_*+\epsilon_* \log\frac{r}{r_*} + \frac{\epsilon_*^2}{72}\left(1-\frac{2r^4-r_h^4}{2r_*^4-r_h^4}
+12 \log\frac{r}{r_*} + 36 \log^2\frac{r}{r_*}+\right.\nonumber\\
&&\left.
+\frac92
\left(Li_2\Big(1-\frac{r_h^4}{r^4}\Big)-Li_2\Big(1-\frac{r_h^4}{r_*^4}\Big)\right)
\right)+ O(\epsilon_*^3)\,\,,
\label{finiteTsol}
\end{eqnarray}
where $Li_2(u)\equiv \sum_{n=1}^\infty \frac{u^n}{n^2}$ is a polylogarithmic function. The functions written in (\ref{finiteTsol}) determine a geometry that is regular at the horizon $r=r_h$. In the next subsection we will study its thermodynamics and we will extract some consequences for the dual field theory with dynamical quarks at non-zero temperature.
Let us conclude this section with some comments on the stability of our perturbative non-extremal solutions.
A possible way to check for the latter is to consider worldvolume fluctuations of a D7-brane in the setup. If, as in our cases, the brane corresponds to massless flavors, the related quasi-normal modes on the unflavored background all have frequencies with a negative imaginary part of the order of the temperature,
signaling stability \cite{mesonmelting}, \cite{myers}.
This result cannot be changed in the flavored case when a perturbative
expansion in $\epsilon_*$ is done. Thus, in our regime of approximations, stability with respect to those fluctuations is guaranteed.
\subsubsection{Thermodynamics of the solution}
\label{sec:thermo}
In the previous subsections we have defined the backreacted background in terms of an arbitrary UV scale $r_*$ as an expansion in powers of the parameter $\epsilon_*$ written in (\ref{epsstar}). This scale $r_*$ should be well separated from the Landau pole scale in order to avoid having the pathologies of the latter. Moreover, we are interested in analyzing the physical consequences of this background at energies much lower than the UV scale $r_*$. In a black hole background dual to a quark gluon plasma the natural IR scale is the location $r_h$ of the horizon, which should be related to temperature $T$ of the plasma. Accordingly, we define $\epsilon_h$ as:
\begin{equation}
\epsilon_h = \frac{\lambda_h\,Vol(X^3)}{16\pi\,Vol(X^5)}\frac{N_f}{N_c}\,\,,
\end{equation}
where, in what follows, the subscript $h$ means that the quantities are evaluated at the horizon $r=r_h$.
Thus, $\lambda_h$ is
naturally identified with the 't Hooft coupling at the scale of the plasma temperature.
We therefore have:
\begin{equation}
\epsilon_h = \epsilon_* \frac{e^{\Phi_h}}{e^{\Phi_*}} =
\epsilon_* + \epsilon_*^2 \log\frac{r_h}{r_*} + O(\epsilon_*^3)\,\,.
\label{epsirelation}
\end{equation}
We will use this relation to recast the expansions in powers of $\epsilon_*$ as series in $\epsilon_h$. We will assume in what follows that $r_h$ is well below the reference UV scale $r_*$ to ensure that the IR physics does not depend on the UV completion of the theory. In a Wilsonian sense of the renormalization group flow, the UV details of the theory should not affect the IR physics. Moreover, since, as we will see below, $r_h$ is proportional to the temperature
(at leading order), we have:
\begin{equation}
\frac{d\epsilon_h}{dT}=\frac{\epsilon_h^2}{T}+ O(\epsilon_h^3)\,\,,
\label{epsilonrun}
\end{equation}
and $T(d\lambda_h/dT)=\epsilon_h\lambda_h$ at leading order. These relations reflect the running of the gauge coupling induced by the dynamical flavors.
The thermodynamic properties of the black hole solution are determined by the metric functions at the horizon. After neglecting terms suppressed in powers of $r_h^4/r_*^4$, the values of the functions $\hat f$ and $\hat g$ at $r=r_h$ can be obtained from (\ref{finiteTsol}). One gets:
\begin{equation}
e^{\hat f_h}\,=\, 1 - \frac{\epsilon_h}{24} + \frac{17}{1152} \epsilon_h^2 + O(\epsilon_h^3)\,\,,\qquad
e^{\hat g_h}\,= \,1 + \frac{\epsilon_h}{24} + \frac{1}{128} \epsilon_h^2+ O(\epsilon_h^3)\,\,.
\label{FSphi}
\end{equation}
The black hole temperature can be obtained by requiring regularity of the euclideanized metric and by
identifying the temperature with the
inverse of the period of the euclideanized time.
A simple computation yields:
\begin{equation}
T=\frac{2r_h}{2\pi R^2 \,e^{4\hat g_h+\hat f_h}}
=\frac{r_h}{\pi R^2}\left[1-\frac18 \epsilon_h
-\frac{13}{384}\epsilon_h^2 + O(\epsilon_h^3) \right]\,\,,
\label{temperature}
\end{equation}
where in the last step we have used the values of $\hat f_h$ and $\hat g_h$ written in (\ref{FSphi}).
The entropy density $s$ is proportional to $A_8$, the volume at the horizon of the eight dimensional part of the space orthogonal to the $\hat t,r$
plane (where $\hat t$ is the Euclidean time), divided by the infinite constant volume of the 3d space directions
$V_3$. From the general form of the metric we get that:
\begin{equation}
s=\frac{2\pi\,\,A_8}{\kappa_{10}^2\,V_3} =
\frac{r_h^3 \,R^2 \,e^{4\hat g_h+\hat f_h}\,
Vol(X^5)}{2^5 \pi^6 g_s^2 \alpha'^4}
=\frac{\pi^5}{2Vol(X^5)} N_c^2 \frac{r_h^3}{\pi^3 R^6}\left[1+\frac18 \epsilon_h
+\frac{19}{384}\epsilon_h^2 + O(\epsilon_h^3) \right]\,\,,
\end{equation}
which in terms of the temperature reads:
\begin{equation}
s=
\frac{\pi^5}{2Vol(X^5)} N_c^2 T^3 \left[1+\frac12 \epsilon_h
+\frac{7}{24}\epsilon_h^2 + O(\epsilon_h^3) \right]\,\,.
\label{entropy}
\end{equation}
As for the other thermodynamic quantities which will follow, the leading term of this formula is the well-known unflavored result.
The $O(\epsilon_h)$ term was already calculated in \cite{myers} with the probe brane technique,
in the $X^5=S^5$ case. Here we have re-obtained this result in a quite standard way,
by computing the increase of the horizon area produced by the flavor branes. This can be considered
as a crosscheck of the validity of the whole construction.
Finally, the order $\epsilon_h^2$ was first obtained in \cite{Bigazzi:2009bk}.
The ADM energy of the solution can be computed as an integral of the extrinsic curvature of the eight-dimensional hypersurface of constant time and radius. This calculation is straightforward and has been done in the appendix B of \cite{Bigazzi:2009bk}, with the result:
\begin{equation}
\varepsilon= \frac{E_{ADM}}{V_3}=\frac38 \frac{\pi^5}{Vol(X^5)} N_c^2 T^4
\left[1+\frac12 \epsilon_h(T) + \frac13 \epsilon_h(T)^2 +O(\epsilon_h(T)^3)\right]\,\,.
\label{ADMresult}
\end{equation}
Again, terms suppressed as powers of $\frac{r_h}{r_*}$ have been neglected. Moreover,
since in the following derivatives with respect to $T$ are going to be taken, we find
it convenient to make explicit that $\epsilon_h$ depends on $T$ (see (\ref{epsilonrun})).
Eq. (\ref{ADMresult}) yields the energy density of the plasma and, thus, it
allows us to study the full thermodynamics. Indeed,
from (\ref{epsilonrun}) and (\ref{ADMresult}) we get immediately the heat capacity (density):
\begin{equation}
{c_V} = \partial_T \varepsilon = \frac32 \frac{\pi^5}{Vol(X^5)}
N_c^2 T^3 \left[1+\frac12 \epsilon_h(T) +\frac{11}{24} \epsilon_h(T)^2+O(\epsilon_h(T)^3)\right] ~.
\label{heatcapacity}
\end{equation}
The free energy density, and so (minus) the pressure, reads:
\begin{equation}\label{freeen}
\frac{F}{V_3} =-p = \varepsilon - T s= -\frac18 \frac{\pi^5}{Vol(X^5)}
N_c^2 T^4 \left[1+\frac12 \epsilon_h(T) +\frac16 \epsilon_h(T)^2+O(\epsilon_h(T)^3)\right] ~.
\end{equation}
Notice that, consistently, this satisfies the relation $s=\partial_T p$ (where it is crucial
to take (\ref{epsilonrun}) into account).
This result is confirmed by the direct computation of $F$ from the renormalized euclidean action (see, again, the appendix B of \cite{Bigazzi:2009bk}) and also by the calculation in \cite{BCT} of the correlator of the tensorial mode in the hydrodynamical approximation.
The speed of sound $v_s$ is obtained by combining (\ref{entropy}) and (\ref{heatcapacity}), namely:
\begin{equation}
v_s^2 = \frac{s}{c_V} = \frac13 \left[1-\frac{1}{6} \epsilon_h(T)^2+O(\epsilon_h(T)^3)\right]~.
\label{vs2}
\end{equation}
Note that the correction to the speed of sound, which measures the deviation from conformality, only appears at second order and that the sign of the correction is consistent with the bound $v_s^2\le \frac13$ conjectured in \cite{Cherman:2009tw}.
It is also interesting to point out that the solution
provides a direct measure of the breaking of conformality at second order, namely the so-called interaction measure, given by:
\begin{equation}
\frac{\varepsilon-3p}{T^4}=\frac{\pi^5 N_c^2}{16Vol(X^5)}\epsilon_h(T)^2\,\,.
\end{equation}
Let us now analyze the viscosity of the plasma predicted by the flavored black hole. Since we are not introducing higher derivatives of the metric in the action, the usual theorems apply and the shear viscosity $\eta$ saturates the Kovtun-Son-Starinets bound
\cite{Kovtun:2004de}, \hbox{\it i.e.}\ $\eta/s=1/4\pi$. Therefore, the shear viscosity $\eta$ can be obtained by dividing by $4\pi$ the entropy density written in (\ref{entropy}). Again, the first-order term coincides with the one calculated in the probe approach in \cite{Mateos:2006yd} while the second-order result was first computed in \cite{Bigazzi:2009bk}. On the other hand, one can also compute the bulk viscosity $\zeta$ for this model, with the result \cite{BCT}:
\begin{equation}
\zeta=\frac{\pi^4}{72 Vol(X^5)}N_c^2T^3 \left[\epsilon_h(T)^2+O(\epsilon_h(T)^3)\right]\, .
\label{bulk-vis}
\end{equation}
Interestingly, the value of $\zeta$ written in (\ref{bulk-vis}) saturates the bound proposed in \cite{buchel}:
\begin{equation}
\frac{\zeta}{\eta}\geq 2\left(\frac13 - v_s^2\right)\, .
\label{bbound}
\end{equation}
For the computation of other transport coefficients, we refer the reader to \cite{BCT}.
\subsubsection{Energy loss of partons}
One of the main phenomenological applications of holography is the analysis of the energy loss of a parton that moves through a quark-gluon plasma. One of the measures of this energy degradation is the so-called jet quenching parameter $\hat q$, which is a transport coefficient that measures the bremsstrahlung experienced by a parton probe due to its interactions with the quarks and gluons of the plasma \cite{Baier}. At very high energy, and using the eikonal approximation, the authors of \cite{liu} found a non-perturbative prescription for calculating $\hat q$
as the coefficient of $L^2$ in an almost light-like Wilson loop with dimensions $L^{-}\gg L$. By using
the generic formula in \cite{aredmas} (and cutting the integral at
$r_*$), we can write:
\begin{equation}
\hat q^{-1}= \pi\, \alpha' \int_{r_h}^{r_*} e^{-\frac{\phi}{2}}
\frac{\sqrt{g_{rr}}}{g_{xx}\sqrt{g_{xx}+g_{tt}}}dr
= \frac{\pi\, \alpha'\,R^4}{r_h^2} e^{-\frac{\phi_h}{2}}\int_{r_h}^{r_*} e^{-\frac{(\phi-\phi_h)}{2}}
\frac{e^{4\hat g+\hat f}}{\sqrt{r^4-r_h^4}}\,\, dr
\,\,.
\label{qhat-general}
\end{equation}
The dilaton enters the formula because we are considering the Einstein frame metric. By plugging in (\ref{qhat-general}) the expressions of $\hat f$, $\hat g$ and $\phi$ written in (\ref{finiteTsol}), and by performing the corresponding integrals in $r$, one gets $\hat q$ as a power series expansion in $\epsilon_*$. In the course of this calculation we will neglect terms that are suppressed by powers of $r_h/r_*$ and we will write the result in a series in $\epsilon_h$ rather than in $\epsilon_*$.
In terms of gauge theory quantities one gets \cite{Bigazzi:2009bk}:
\begin{equation}
\hat q=\frac{\pi^3\sqrt{\lambda_h}\Gamma(\frac34)}{\sqrt{Vol(X^5)}\,\Gamma(\frac54)}T^3
\left[1 + \frac18(2+\pi) \epsilon_h + \gamma\, \epsilon_h^2+ O(\epsilon_h^3)
\right]\,\,,
\label{jetq}
\end{equation}
where we have introduced a constant $\gamma$:
\begin{equation}
\gamma= \frac{11}{96}+\frac{\pi}{48} + \frac{3\pi^2}{128} + \frac18 {\cal C}+
\frac{1}{48}\ {}_4 F_3\left(
1,1,1,\frac32;\frac74,2,2;1\right)\approx 0.5565\,\,,
\end{equation}
with ${\cal C}\sim 0.91597$ being the Catalan constant. Notice that the flavor correction to $\hat q$ is positive, \hbox{\it i.e.}\ fundamentals enhance the jet quenching. Actually, eq. (\ref{jetq}) can be used to estimate this enhancement in the
extrapolation to the realistic RHIC regime. Let us take
$X^5=S^5$, $N_c=N_f=3$ and $\alpha_s\,=\,g^2_{YM}/4\pi\sim 1/2$. Then, $\lambda_{h}\sim 6\pi$ and $\epsilon_h\sim {N_f\over 4 \pi}\sim 0.24$. Using this value in (\ref{jetq}) we have that $\hat q$ is increased by $20\%$. For example, at
$T=300$ MeV we obtain $\hat q \sim 5.3$ (Gev)$^2$/fm, to be compared with the value \cite{liu} $\hat q \sim 4.5$ (Gev)$^2$/fm of the unflavored plasma (the RHIC values are
$\hat q \sim 5-15$ (Gev)$^2$/fm). It is also interesting to rewrite (\ref{jetq}) in terms of the entropy density $s$. One gets:
\begin{equation}
\hat q = c \,\sqrt{\lambda_h} \sqrt{\frac{s}{N_c^2}}T^\frac32
\left[1+\frac{\pi}{8}\epsilon_h+(\gamma-\frac{11}{96}-\frac{\pi}{32})\epsilon_h^2+
O(\epsilon_h^3) \right]\,\,,
\qquad
c=\sqrt{2\pi}\frac{\Gamma(\frac34)}{\Gamma(\frac54)}\,\,,
\end{equation}
which shows a deviation (driven by $\epsilon_h$) from the general expression
put forward in \cite{lrw2}. In this setting, the presence of fundamentals and the breaking of
conformality are inevitably mingled. It would be interesting to have the dual of a conformal theory
with fundamentals to check whether the conjecture \cite{lrw2} holds in such situation.
That was analyzed in \cite{Bertoldi:2007sf} in a non-critical string framework
and, interestingly, the result differs from \cite{lrw2}. The caveat is that
the model studied in \cite{Bertoldi:2007sf} suffers from the usual
problem of gravity-like approaches to non-critical strings, namely
there are uncontrolled approximations.
Another way of characterizing the energy loss of a parton probe in the plasma is by modeling it as a macroscopic string attached to a probe flavor brane. The string is dragged by a constant force $f$ which keeps its velocity $v$ fixed and transfers to the parton energy and momentum, which is then lost in the plasma at a constant rate. This energy loss is measured by the drag coefficient $\mu$, which relates the force $f$ and the parton momentum $p$: $f=\mu p$. To compute this drag force one can apply the general procedure of refs. \cite{herzog1,gubserdrag,herzog2}. By using the Nambu-Goto action for a string in the black hole background one gets that the rate of momentum transferred to the medium is given by \cite{herzog2}:
\begin{equation}
\frac{dp}{dt} = -\frac{1}{2\pi\alpha'}C = -
\frac{r_h^2}{2\pi\alpha'\,R^2}e^\frac{\Phi(r_c)}{2}\frac{v}{\sqrt{1-v^2}}=-
\mu\,M_{kin}\,\frac{v}{\sqrt{1-v^2}}\,\,,
\label{drag1}
\end{equation}
where $C$ is the constant
determined from the equation $g_{xx}(r_c)g_{tt}(r_c)+C^2=0$ with the point $r_c$ given by $g_{tt}(r_c)+g_{xx}(r_c)v^2=0$, namely $r_c=r_h(1-v^2)^{-\frac14}$.
In (\ref{drag1}) we have introduced, following ref. \cite{herzog1},
the kinematical mass $M_{kin}$ such that $p=M_{kin}\frac{v}{\sqrt{1-v^2}}$.
From (\ref{drag1}), using (\ref{finiteTsol}), (\ref{epsirelation}), (\ref{temperature}),
we find:
\begin{eqnarray}\label{muemme}
\mu\,M_{kin} &=& \frac{\pi^{5/2}}{2} \frac{\sqrt{\lambda_h}}{\sqrt{Vol(X^5)}}\, T^2 \left[
1+\frac18 (2-\log(1-v^2)) \epsilon_h +
\right. \\
&& \left.+
\frac{1}{384}\left[44-20 \log(1-v^2)
+9 \log ^2(1-v^2)+12 Li_2(v^2) \right] \epsilon_h^2 + O(\epsilon_h^3)\right]\,\,.\nonumber
\end{eqnarray}
As happens with the jet quenching, the energy loss (at fixed $v$) is enhanced by the presence of fundamental matter. The quantity $\mu\,M_{kin}$ grows when increasing the velocity. From (\ref{muemme}), formally,
it would diverge as $v\rightarrow 1$. However,
(\ref{muemme}) is not applicable in that limit since
we have to require $\epsilon_h \log(1-v^2) \ll 1$ for the expansions to be valid.
\subsection{A discussion on the range of validity}
\label{sec:rangeval}
We now discuss, following \cite{Bigazzi:2009bk}, the restriction on the physical
parameters needed for the deconfined flavored plasma
solution to be physically meaningful. Before we go on, two comments
are in order: first, notice that, even if we will use here the plasma
temperature as the IR scale at
which the relevant physics takes place, this can be substituted by any other IR
scale, depending on what one wants to study. Thus, for instance, when computing meson
masses at zero temperature as in section \ref{sec:screening}, the discussion below holds,
just taking into account the IR scale there is set by the quark mass.
Second, notice that the restriction
of small $\epsilon_*$ (which leads to $N_f \ll N_c$)
of the D3-D7 case at hand comes from the existence of a Landau pole.
In holographic theories in which
there is no Landau pole in the geometry (sections \ref{sec:D5D5}, \ref{sec:2+1}),
there is in principle no
restriction to $N_f$. In particular, it is possible to consider in those theories
$N_f$ to be of the same order as $N_c$.
As we have already remarked,
having a pathological UV means that
there must exist a separation of scales between IR and UV.
Concretely, there must exist a hierarchy, which
in terms of the $r$ radial coordinate reads:
\begin{equation}
r_h \ll r_* \ll r_a < r_{LP}\,\,.
\end{equation}
The quantity $r_h$ sets the scale of the plasma temperature $\frac{r_h}{R^2}\sim \Lambda_{IR}\sim T$,
which is the scale at which we want to analyze the physics.
The point $r_{LP}$ is where the dilaton diverges, signaling a Landau pole in the dual theory.
At a scale $r_a$ the string solution starts presenting subtler
pathologies, whose
discussion we delay until the end of this section.
Finally, $r_*$ sets an (arbitrary) UV cutoff scale $\frac{r_*}{R^2}\sim \Lambda_{UV}$.
The solution (\ref{finiteTsol}) will only be used for $r<r_*$.
In a Wilsonian sense of a renormalization group flow, the UV details
should not affect the IR physical predictions.
This feature is reflected in the fact that physical quantities do not depend (up to suppressed contributions)
on $r_*$ or functions evaluated at that point, but only on IR parameters.
Even if the precise value of $r_*$ is arbitrary, we have to make sure that it is possible
to choose it such that it is well above the IR scale (so that the UV
completion
only has negligible effects on the IR physics) and well below the pathological $r_a, r_{LP}$ scales
(so that the solution we use is meaningful and the expansions do not break down).
To this we turn now.
Let us start by computing the hierarchy between $r_*$ and $r_{LP}$. Since at
$r_*$ we can approximate the solution by the supersymmetric one, we can read the
position of the Landau pole from (\ref{susysol}). If we insert
the approximate relation between radial coordinates
$r \approx \sqrt{\alpha'} e^{\rho}$, we find:
\begin{equation}
\frac{r_*}{r_{LP}}\approx e^{-\frac {1} {\epsilon_*}} \ll 1\,\,,
\end{equation}
as long as $\epsilon_* \ll 1$.
Moreover, one has to make sure that the Taylor expansions (\ref{finiteTsol})
are valid in the region $r_h<r<r_*$. This of course requires $\epsilon_* \ll 1$,
but also that $\epsilon_* \left|\log\frac{r_h}{r_*}\right|\ll 1$ (notice that the absolute value of the
logarithm can be big because $r_h \ll r_*$). This means that
$\frac{r_h}{r_*} \gg e^{-\frac {1} {\epsilon_*}}$.
On the other hand, when computing physical quantities in the previous sections, we always neglect
quantities suppressed as powers of $\frac{r_h}{r_*} \sim \frac{T}{\Lambda_{UV}}$.
This is the order of magnitude
of the corrections due to the eventual UV completion of the theory at $r_*$. One has to make sure that
the corrections in $\epsilon_*$ we are keeping are much larger than the neglected ones,
namely $\epsilon_* \gg \frac{r_h}{r_*}$. In summary, we have the
following hierarchy of
parameters (in the following, in order to avoid overly messy expressions, we insert the value of $\epsilon_*$ for the
$X^5=S^5$ case, remembering that for a generic $X^5$, its value is given by
(\ref{epsstar})):
\begin{equation}
e^{-\frac {1} {\epsilon_*}}\sim
e^{-\frac {8\pi^2\,N_c} {\lambda_*\,N_f}}
\ll \frac{r_h}{r_*} \sim \frac{T}{\Lambda_{UV}}
\ll \epsilon_* \sim \frac{\lambda_*\,N_f}{8\pi^2\,N_c} \ll 1\,\,.
\end{equation}
As long as $\epsilon_* \sim \frac{\lambda_*\,N_f}{8\pi^2\,N_c} \ll 1$, there always
exists a range of $r_*$ such that this inequality is satisfied. Since we focus on the IR physics of the plasmas, at the scale set by their temperature, the actual physical constraint on the parameters will be
$
\frac{\lambda_h}{8\pi^2}\frac{N_f}{N_c} \ll 1\,\,,
$
which we have written in terms of the coupling at the scale of the horizon,
$\lambda_h = \lambda_* (1 + O(\epsilon_*))$.
On top of this, we have to make sure that the SUGRA+DBI+WZ action we are using is
valid. As usual, the suppression of
closed string loops requires
$N_c \gg 1$ whereas the suppression of $\alpha'$-corrections
is guaranteed by $\lambda_h \gg 1$.
We have written the D7-brane worldvolume contribution to the action as a sum of $N_f$ single brane contributions.
This is justified if the typical energy of a string connecting two different branes is large
(in $\alpha'$ units). Since the branes are distributed on a space whose size is controlled
by $R \sim \lambda_h^\frac14 \sqrt{\alpha'}$, we again need $\lambda_h \gg 1$.
The smearing approximation will be good if
the distribution of D7-branes on the transverse space is dense, i.e. $N_f\gg 1$.
The discussion up to now is summarized in the following validity regime:
\begin{equation}\label{validity}
N_c \gg 1\,\,,\qquad \lambda_h \gg 1\,\,,\qquad N_f \gg 1 \,\,,\qquad
\epsilon_h= \frac{\lambda_h}{8\pi^2}\frac{N_f}{N_c} \ll 1\,\, .
\end{equation}
Finally, we want to find the regime of parameters in which the flavor corrections are not
only {\it valid} but are also the {\it leading} ones. With this aim,
we ought to demand that the leading $\alpha'$-corrections to the supergravity action
(which typically scale as $\lambda_h^{-\frac32}$
due to terms of the type $\alpha'^3 {\cal R}^4$) are smaller than the flavor
ones, controlled by $\epsilon_h$, namely:
\begin{equation}
\lambda_h^{-\frac32} \ll \epsilon_h\,\,.
\label{leadingcond}
\end{equation}
Demanding that corrections to the D7-branes contributions
(for instance curvature corrections to the worldvolume action itself or corrections produced by possible
modifications of the brane embeddings due to curvature
corrections to the background metric)
are subleading does not impose any further restriction. The reason is that their contribution
is typically of order $\epsilon_h \lambda_h^{-c}$ for some $c>0$ which is always subleading
with respect to $\epsilon_h$ as long as (\ref{validity}) is satisfied.
\subsubsection*{The holographic a-function}
As discussed in \cite{unquenchedmesons} and mentioned above, the string solution starts presenting
pathologies at a scale $r_a < r_{LP}$, where the
holographic $a$-function is singular. The utility of the solution for $r>r_a$ is doubtful, but
since we have only used the solutions up to $r_* \ll r_a$ in order to derive the IR physics, this
subtlety does not affect the physical results. We now briefly review the argument in
\cite{unquenchedmesons}, which used the backreacted Klebanov-Witten solution at zero temperature.
The qualitative picture holds for the rest of the cases addressed in the present section and for the case of section \ref{KS} too.
Let us start by considering the metric of a generic dimensional reduction to
five dimensions, giving a 5d Einstein frame metric of the form (the $u$ here is, obviously, a redefined
holographic coordinate, namely $u=u(r)$):
\begin{equation}
ds_5^2 = H(u)^{1/3}\left[dx_{1,3}^2 + \beta(u) du^2\right]\,.
\end{equation}
In standard set-ups, the function $H(u)^{1/6}$, which can be roughly identified with the dual field theory energy scale, monotonically varies with the radial coordinate. This is also required in order for the ``holographic $a$-function'' \cite{holaf}:
\begin{equation}
a(u)\sim \beta(u)^{3/2} H(u)^{7/2} [H'(u)]^{-3}\,,
\label{holadef}
\end{equation}
to be finite.\footnote{The monotonicity of $H(u)$ also plays a crucial role in holographic computations of the entanglement entropy, see \cite{kkm}. The notations of that paper are used in the equations above.}
Instead, the function $H(u)$ is not monotonic here: it increases with $u$ from zero up to a maximum
at a point $u_a$ and then it decreases back to zero where $h$ vanishes\footnote{
For the present discussion and in particular for figure \ref{Hflav}
we will choose the additive integration constant of $h$ such that
$h$ is zero at the Landau pole. The specific point $u_a$ (namely $r_a$) at which this UV pathology
sets in depends on this choice. Again, we stress that the important point is the IR results do not
depend on this choice (modulo suppressed contributions) as long as $r_a \gg r_*$. What we show here is
that the integration constant can be naturally chosen such that this condition is satisfied.
}. In the flavored supersymmetric KW case, the $H$ and $a$ functions
read:
\begin{equation}
H(\rho)\,\sim\,h\,e^{2f+8g}\,\,,\qquad\quad
a(\rho) \,\sim\, h^{3/2}\,e^{3f}\, H^{7/2} [\partial_\rho H]^{-3}\,,
\label{holafkw}
\end{equation}
where we have not written unimportant overall factors.
A representative plot is given in figure \ref{Hflav}.
\begin{figure}
\centering
\includegraphics[width=.4\textwidth]{Hflavoredkwp.eps}
\caption{ The function $H$ in the massless-flavored KW model at zero temperature.}
\label{Hflav}
\end{figure}
The non-monotonic behavior of $H$
implies that the holographic $a$-function is singular and discontinuous at the ``$a$-scale".
From the plot, we see that $r_a \sim e^{\rho_a}$ is below, but not parametrically separated from $r_{LP}$.
\setcounter{equation}{0}
\section{A dual to ${\cal N}=1 $ SQCD-like theories}
\label{sec:D5D5}
In the following section, we will study a system that in some sense is qualitatively different from those of the previous sections, though the procedure to deal with the addition of flavors is identical. The main qualitative difference will be that there need not be a hierarchical difference between the number of flavors and the number of colors. The case treated here will represent the addition of fundamental matter to
a field theory that is originally confining and four dimensional at low energies, but that gets some higher dimensional completion in the UV (in principle this allows one to extend the range of the radial coordinate to arbitrarily large values). Some of the qualitative changes that observables of a confining theory undergo when fundamentals are added will be discussed. The developments described in the present section were
applied to model possible aspects that could appear in
physics beyond the standard model, as we will briefly mention below.
More concretely, in this section, we will study a dual
to a version of ${\cal N}=1$ SQCD. The model is
based
on D5-branes wrapped on two-cycles inside the resolved conifold-leading
to a
geometry related to the deformed conifold. We will first briefly present
the model without flavors, then study the addition of flavors following
the
ideas described in the first sections of this paper: kappa symmetric
embeddings, smearing, backreaction, system of BPS equations, particular
solutions to this system and finally present a set of checks that the
correspondence we are proposing is valid and robust, we also explain some
predictions about the field theory obtained with the string background.
\subsection{The model without flavors}
The proposal is to construct a dual to a field theory with minimal SUSY in
four dimensions using wrapped branes. Ideas of this kind were first
explored by Witten in the early days of AdS/CFT. In the paper
\cite{Witten:1998zw}, Witten presented a model dual to a version of
Yang-Mills theory
(with an extra massless scalar and UV-completed by an infinite tower of
massive vectors, scalars and fermions), by wrapping
a set of $N_c$ D4 branes on a
circle with SUSY-breaking boundary
conditions.
The idea here is very similar, only that we will work with D5-branes and
we will preserve some amount of SUSY. We will compactify the five branes
in a very subtle way (involving a twisting of the 6-d
theory) so that only four supercharges will be preserved in the
compactified theory for all energies \cite{Maldacena:2000yy}\footnote{The
fact that a
twisting procedure (see
\cite{Witten:1994ev} for a very nice presentation of this idea) is at
work, implies that even in the far UV, the theory is still preserving only
four supercharges .}, in other words, the partial SUSY breaking is {\it
not} due
to the presence of relevant operators, like mass terms.
This kind of compactification of the six dimensional theory living on a
stack of D5-branes (when the D5's wrap a two-cycle inside the resolved
conifold) was well studied in various papers, see \cite{Bertolini:2003iv}
for various reviews. We will follow mostly the detailed study of
\cite{Andrews:2005cv}.
One can show that a very generic string background describing a stack of
$N_c$ D5-branes wrapping a two cycle and preserving four supercharges
includes a metric, RR-three form $F_3=dC_2$, a dilaton $\phi(\rho)$ and is
given by,
\begin{eqnarray}
ds^2 &=& \alpha' g_s e^{ \frac{\phi(\rho)}{2}} \Big[dx_{1,3}^2 +
e^{2k(\rho)}d\rho^2
+ e^{2 h(\rho)}
(d\theta^2 + \sin^2\theta d\varphi^2) +\nonumber\\
&+&\frac{e^{2 g(\rho)}}{4}
\left((\tilde{\omega}_1+a(\rho)d\theta)^2
+ (\tilde{\omega}_2-a(\rho)\sin\theta d\varphi)^2\right)
+ \frac{e^{2 k(\rho)}}{4}
(\tilde{\omega}_3 + \cos\theta d\varphi)^2\Big], \nonumber\\
F_{(3)} &=&\frac{g_s\alpha' N_c}{4}\Bigg[-(\tilde{\omega}_1+b(\rho)
d\theta)\wedge
(\tilde{\omega}_2-b(\rho) \sin\theta d\varphi)\wedge
(\tilde{\omega}_3 + \cos\theta d\varphi)+\nonumber\\
& & b'd\rho \wedge (-d\theta \wedge \tilde{\omega}_1 +
\sin\theta d\varphi
\wedge
\tilde{\omega}_2) + (1-b(\rho)^2) \sin\theta d\theta\wedge d\varphi \wedge
\tilde{\omega}_3\Bigg],
\label{nonabmetric424}
\end{eqnarray}
where $\tilde\omega_i$ are the left-invariant forms of $SU(2)$
\begin{eqnarray}\lab{su2}
&&\tilde{\omega}_1= \cos\psi d\tilde\theta\,+\,\sin\psi\sin\tilde\theta
d\tilde\varphi\,\,,\nonumber\\
&&\tilde{\omega}_2=-\sin\psi d\tilde\theta\,+\,\cos\psi\sin\tilde\theta
d\tilde\varphi\,\,,\nonumber\\
&&\tilde{\omega}_3=d\psi\,+\,\cos\tilde\theta d\tilde\varphi.
\end{eqnarray}
For convenience, below we will set the parameters $g_s=\alpha'=1$. The
presence of the $N_c$ color D5-branes is indicated in $F_3$ that
satisfies
the quantization condition:
\begin{equation}
\frac{1}{2\kappa_{(10)}^2}\int_{S^3} F_{(3)} = N_c T_5\,\,.
\label{quantcond}
\end{equation}
The $S^3$ on which we integrate is parameterized by $\tilde\theta,
\tilde\varphi,\psi$.
We then impose that a fraction of SUSY is preserved, hence we need to
impose some projections on the Type IIB spinors and a set of BPS
equations reflecting this arise (see the appendix B in the paper
\cite{Casero:2006pt} for generous details). The BPS equations are
non-linear, first order and coupled for the functions of the background
in eq. (\ref{nonabmetric424})-for details see appendix B in
\cite{Casero:2006pt}. Certainly, solving first order equations is simpler
than solving the second order Einstein equations; nevertheless the BPS
equations for the functions $(\phi,h,g,k,a,b)$
are non-linear and coupled, rendering the problem
complicated.
It is technically convenient to make a `change of basis' to another set of
functions, so that the BPS equations become first order and nonlinear (of
course) but can be decoupled, and then solved independently. A change of
variables that does the job partially was obtained in
\cite{HoyosBadajoz:2008fw}. The change of basis is from the set of
functions $[\phi,h,g,k,a,b]$ into the functions
$[P,Q,\tau,\Phi,Y,\sigma]$. The map reads \cite{HoyosBadajoz:2008fw},
\begin{equation}
e^{2h}=\frac14\left(\frac{P^2-Q^2}{P\cosh\tau-Q}\right),
\quad e^{2g}=P\cosh\tau-Q,\quad e^{2k}=4Y,\quad
a=\frac{P\sinh\tau}{P\cosh\tau-Q},\quad b=\frac{\sigma}{N_c}.
\label{changevariables}
\end{equation}
and
\begin{equation}
\Phi= (P^2-Q^2)\sqrt{Y} e^{2\phi}.
\end{equation}
As explained in detail in \cite{HoyosBadajoz:2008fw} (see section 3 of
that paper), the BPS equations can be solved one by one for these new
functions, obtaining:
\begin{eqnarray}
& & Q(\rho)=(Q_o+ N_c)\cosh\tau + N_c (2\rho \cosh\tau -1),\nonumber\\
& & \sinh\tau(\rho)=\frac{1}{\sinh(2\rho-2\rho_o)},\quad
\cosh\tau(\rho)=\coth(2\rho-2\rho_o),\nonumber\\
& & Y(\rho)=\frac{P'}{8},\nonumber\\
& & e^{4\phi}=\frac{e^{4\phi_o} \cosh(2\rho_o)^2}{(P^2-Q^2) Y
\sinh^2\tau},\nonumber\\
& & \sigma=\tanh\tau (Q+N_c)= \frac{(2N_c\rho + Q_o + N_c)}{\sinh(2\rho
-2\rho_o)}.
\label{BPSeqs}
\end{eqnarray}
Note that both $Q$ and the dilaton are given {\em algebraically} in terms
of the rest of the functions
parametrizing the backgrounds.
Here $Q_o$ and $\phi_o$ are constants of
integration and we have chosen the integration constant in the dilaton
field equation $\phi_0$
such that it admits a smooth
limit as $\rho_o\rightarrow -\infty$ (this limit gives $\tau=\sigma=0$ and so corresponds
to what in the paper \cite{Casero:2007jj}
were called type {\bf A} backgrounds).
The function $P$ satisfies the following second order equation:
\begin{equation}
P'' + P'\Big(\frac{P'+Q'}{P-Q} +\frac{P'-Q'}{P+Q} - 4 \coth(2\rho-2\rho_o)
\Big)=0.
\label{master}
\end{equation}
We will refer to this equation as the `master' equation, since once we
have a solution of (\ref{master}) all other functions
are determined via (\ref{BPSeqs}).
\subsection{Some solutions}
There are many solutions to the master equation
(\ref{master}).
A very simple one is given by
\begin{equation}
P=2 N_c \rho ,\;\;\; Q_o=-N_c.
\label{cv}
\end{equation}
Once processed back, one computes the functions in the original
background of
eq. (\ref{nonabmetric424}) and one recovers an old solution
\cite{Chamseddine:1997nm}. To avoid nasty singular behaviors, in the
following, we will choose the value
of the integration constant $Q_o = - N_c$, so that the first term in the
expression for $Q(\rho)$, namely $(Q_o+N_c)$, vanishes\footnote{If we do not make
this choice, the space ends before $\rho=\rho_0$, since $Q>P$ possibly giving
place to
geodesically incomplete spaces and a divergent dilaton. Hence, we will
choose the term
proportional to $\cosh\tau(\rho)$ in $Q(\rho)$ to vanish.}.
Aside from the simple
solution presented above, there are a variety of very
interesting solutions. For example when the function $P(\rho)$ is given near
$\rho=0$ by the following Taylor series:
\begin{equation}
P= h_1 \rho+ \frac{4 h_1}{15}\left(1-\frac{4 N_c^2}{h_1^2}\right)\rho^3
+\frac{16 h_1}{525}\left(1-\frac{4N_c^2}{3h_1^2}
-\frac{32N_c^4}{3h_1^4}\right)\rho^5+{\cal O}(\rho^7),
\label{solutionnear0xx}\end{equation}
where $h_1$ is again an arbitrary constant (notice that for $h_1=2N_c$
we get back to the solution in eq. (\ref{cv}); we will also assume that
$h_1>2N_c$). It is
interesting that
this solution can be numerically connected in a smooth way with a solution
for large values of the radial coordinate ($\rho\rightarrow \infty$) that
differs greatly from the linear behavior of the solution in eq. (\ref{cv}). In
this case,
it is given by
\begin{equation}
P\sim e^{\frac{4}{3}\rho}\Big[c (1- \frac{8}{3} \rho e^{-4\rho}) +\frac{1}{64
c}( 256 \rho^2 +256 Q_o\rho +144 N_c^2 + 64 Q_o^2
)e^{-\frac{8}{3}\rho}+O(e^{-4\rho}) \Big].
\label{Pinfinity}
\end{equation}
These solutions were studied explicitly in section 8 of the paper
\cite{Casero:2006pt} and have a variety of interesting applications that
we will briefly mention in the following sections.
\subsubsection{An exact recursive solution}
There is one recursive way of obtaining solutions, described in
\cite{Nunez:2008wi}, that basically uses the fact that the master equation
(\ref{master}) can be written as (we choose here and in the following
$\rho_0=0$)
\begin{equation}\label{master-c+}
\pa_\rho\left(s(P^2-Q^2)P'\right)+4sP'QQ'=0,\;\;\;\;
s(\rho)=\sinh^2\tau=\frac{1}{\sinh^2(2\rho)}.
\end{equation}
Integrating eq. (\ref{master-c+}) twice we obtain
\begin{equation}\label{master-c+-integrated}
P^3-3Q^2P+6\int_{\rho_2}^\rho d\rho'QQ'P+12\int_{\rho_2}^\rho d\rho'
s^{-1}\int^{\rho'}_{\rho_1} d\rho'' s P'QQ'=c^3 R(\rho)^3,
\end{equation}
where
\begin{equation}
R(\rho)\equiv\left(\cos^3\alpha+\sin^3\alpha(\sinh(4\rho)-4\rho)\right)^{1/3},
\end{equation}
being $(c,\alpha)$ the two integration constants of the master
equation.
Following
\cite{HoyosBadajoz:2008fw} we write $P$ in a formal expansion in
inverse powers of $c$ - the integration constant encountered above - as:
\begin{equation}\label{solution}
P=\sum_{n=0}^\infty c^{1-n} P_{1-n}.
\end{equation}
Inserting this expansion in eq. (\ref{master-c+-integrated}) we obtain
recursively
\begin{eqnarray}
&&P_1=R,\nonumber\\%\nonumber\\
&&P_0=0,\nonumber\\%\nonumber\\
&&P_{-1}=-\frac13 P_1^{-2}\left(-3Q^2P_1+6\int_{\rho_2}^\rho d\rho' QQ'
P_1+12\int_{\rho_2}^\rho d\rho' s^{-1}\int^{\rho'}_{\rho_1} d\rho''
sQQ'P_1'\right),\nonumber\\\nonumber\\
&&P_{-2}=0,\nonumber\\%\nonumber\\
&&P_{-n-2}=-\frac13P_1^{-2}\left\{\sum_{m=1}^{n+2}\left(2P_1P_{1-m}P_{m-n-2}+\sum_{k=1}^{n-m+3}P_{1-m}P_{1-k}
P_{m+k-n-2}\right)-3Q^2P_{-n}\right.\nonumber\\
&&\left.+6\int_{\rho_2}^\rho d\rho' QQ' P_{-n}+12\int_{\rho_2}^\rho d\rho'
s^{-1}\int^{\rho'}_{\rho_1} d\rho'' sQQ'P_{-n}'\right\},\quad n\geq 1.
\end{eqnarray}
It follows by induction that $P_k=0$ for all even $k$. The large $\rho$
expansion of these solutions coincide with that described in
eq. (\ref{Pinfinity}) Once again, solutions written in this form have
interesting applications to the physics of cascading quivers on the
baryonic branch \cite{Maldacena:2009mw}, \cite{gmnp}.
We will not study the physics encoded in the solutions
described above, suggesting the interested reader to consult the papers
\cite{Bertolini:2003iv},
\cite{Casero:2006pt}, \cite{Nunez:2008wi}.
There is another set of solutions, proposed in \cite{Nunez:2008wi}
and whose physics content was developed in \cite{Nunez:2009da},\cite{Elander:2009pk}
that correspond to what are called ``walking solutions''. The idea here
is to construct string backgrounds such that the dual QFT has a gauge
coupling with very slow running
(or ``walking'' coupling). See the papers
\cite{Nunez:2008wi},\cite{Nunez:2009da},\cite{Elander:2009pk} for detailed explanations on
the physical implications of these solutions.
\subsubsection{A comment about the dual field theory}\label{qftwof}
The system of wrapped D5-branes has a field theory realized on its
worldvolume, whose dual background and various
solutions were described above. The field theory is a version of
minimally SUSY Yang-Mills. Again, some UV completion takes over at
high energies \footnote{We are not saying that Super-Yang-Mills needs a
UV completion, just that the system of D5-branes realizes a theory with
these characteristics.}. The field theory is minimally SUSY (${\cal N}=1$) and its
{\it perturbative} spectrum, aside from a massless vector multiplet contains a
tower of massive vector and chiral multiplets. A careful study of the perturbative dual
field theory obtained by compactification and twisting of the
six-dimensional theory living on (unwrapped) D5-branes was done in
\cite{Andrews:2005cv}. In that paper, the degeneracies and masses of the
(perturbative) states in the tower mentioned above are given. More
interestingly, the authors of \cite{Andrews:2005cv} showed that the theory
is equivalent to ${\cal N}=1^*$ Yang-Mills in a particular Higgs vacuum, where
the
extra dimensions appear by deconstruction. In this sense, we will think of
the theory without flavors either as a six-dimensional theory compactified
or as a four-dimensional theory with an infinite set of fields.
For our purposes, it will be enough to use the fact that the lagrangian of the
field theory reads
\begin{equation}
L= Tr[-\frac{1}{4}F_{\mu\nu}^2 -i \bar{\lambda}\gamma^\mu D_\mu \lambda +
L(\Phi_k, W_k,W)]\,\,,
\end{equation}
where $\Phi_k$ and $W_k$ represent the infinite number of
massive chiral and vector multiplets and $W$ denotes the massless vector
multiplet. The term $L(\Phi_k, W_k,W)$, represents all the kinetic terms
and interactions that can be deduced from \cite{Andrews:2005cv}. More
comments about this field theory can be found in the appendix A of
\cite{HoyosBadajoz:2008fw}.
In what follows, we will summarize the procedure of adding flavors to this
field theory. The flavor branes, in this particular case are D5-branes.
\subsection{Addition of flavors}
The study of supersymmetric embeddings in backgrounds of the form of
eq. (\ref{nonabmetric424}), more precisely for the solution given in
eq. (\ref{cv}) was initiated in the paper \cite{Nunez:2003cf}. There the
eigenspinors of the kappa symmetry matrix were found to be the spinors
preserved by the background for a variety of D5-brane embeddings. For the
purposes of this review, we will focus on the ``cylinder embeddings''
described in section 6.3 of \cite{Nunez:2003cf} and in more detail in
section 6.5.3 of the third paper in reference \cite{Bertolini:2003iv}.
In this case the flavor D5-branes are extended along the $R^{1,3}$
Minkowski directions, on the radial direction $\rho$ and also wrap the
R-symmetry direction $\psi$. Intuitively, the flavor branes are localized
in the directions $(\theta,\tilde{\theta},\varphi,\tilde{\varphi})$, but
interestingly enough, any constant value of these coordinates ensures that
we have a kappa symmetric configuration. This is a very important fact, as
we can put one flavor brane ``at each point'' of the four manifold
$\Sigma[\theta,\tilde{\theta},\varphi,\tilde{\varphi}]$ and still have a
SUSY configuration.
This is precisely what we will take advantage of when smearing. Let us see
this in more detail: if, as discussed in the first section, we
write the action describing the closed strings (IIB) and the open strings
(BIWZ), we will have
\begin{eqnarray}
& & S=\frac{1}{2\kappa_{(10)}^2}
\int d^{10}x \sqrt{-g} \left[R-\frac12 (\partial_\mu \phi)
(\partial^\mu \phi)-\frac{1}{12}e^{\phi}F_{(3)}^2\right]\qquad\qquad
\nonumber\\\nonumber\\
&& \qquad\qquad-T_5 \sum^{N_f}
\int_{{\cal M}_6} d^6x e^{\frac{\phi}{2}}\sqrt {-\hat g_{(6)}}\,+\, T_5 \sum^{N_f}
\int_{{\cal M}_6} P[C_{6}]\,\,,
\label{gravaction}
\end{eqnarray}
where the integrals are taken over the six-dimensional
worldvolume of the flavor
branes ${\cal M}_6$,
and $\hat g_{(6)}$ stands for the determinant of the pull-back
of the metric in such a worldvolume.
As discussed in previous sections, we then think of the $N_f \rightarrow \infty$ branes as being
homogeneously
smeared along the four transverse directions parameterized by the coordinates
$\theta, \varphi$ and $\tilde \theta, \tilde \varphi$.
The smearing erases
the dependence on the angular coordinates and makes it
possible to consider an ansatz with functions only depending on $r$,
enormously simplifying computations. One has:
\begin{eqnarray}
-T_5 \sum^{N_f}
\int_{{\cal M}_6} d^6x e^{\frac{\phi}{2}} \sqrt {-\hat g_{(6)}} &\rightarrow&
-\frac{T_5 N_f}{(4\pi)^2} \int d^{10}x
\sin\theta \sin \tilde \theta e^{\frac{\phi}{2}} \sqrt {-\hat g_{(6)}}\,\,,
\label{BIterm}\\
T_5 \sum^{N_f} \int_{{\cal M}_6} P[C_{6}] &\rightarrow& \frac{T_5 N_f}{(4\pi)^2}
\int Vol({\cal Y}_4) \wedge C_{(6)}\,\,,
\label{WZterm}
\end{eqnarray}
where we have defined $Vol({\cal Y}_4)=\sin \theta \sin \tilde \theta
d\theta \wedge d\varphi \wedge d\tilde\theta \wedge d\tilde\varphi$
and the new integrals span the full space-time.
We will need the expressions (with the choice explained
above $\alpha'=g_s=1$).
\begin{equation}
T_5=\frac{1}{(2\pi)^5}\,\, ,\qquad\qquad
{2\kappa_{(10)}^2} ={ (2\pi)^7 }.
\label{T5value}
\end{equation}
From here, we will have a set of BPS equations describing the dynamics of
this open-closed string system.
The same change of basis with the purposes described around
eq. (\ref{changevariables}) can be performed - see
\cite{HoyosBadajoz:2008fw} for details. The solution
in this case is dependent on the number of flavor branes $N_f$ and reads
(reinstating momentarily the integration constant $\rho_o$):
\begin{equation}\label{tau}
\sinh\tau=\frac{1}{\sinh(2(\rho-\rho_o))},
\end{equation}
for the function $\tau$, while for $Q,\Phi$ we have,
\begin{equation}\label{Q}
Q=\left(Q_o+\frac{2N_c-N_f}{2}\right)\cosh\tau+\frac{2N_c-N_f}{2}\left(2\rho\cosh\tau-1\right),
\end{equation}
\begin{equation}\label{dilaton}
e^{4(\phi-\phi_o)}=\frac{\cosh^2(2\rho_o)}{(P^2-Q^2)Y\sinh^2\tau}.
\end{equation}
In the case with flavors, like in the unflavored case previously discussed,
both $Q$ and the dilaton are given {\em algebraically} in terms
of the rest of the functions
parametrizing the backgrounds.
As before, $\rho_o$, $Q_o$ and $\phi_o$ are constants of
integration and we have chosen the integration constant in (\ref{dilaton})
such that it admits a smooth
limit as $\rho_o\rightarrow -\infty$ (this limit gives $\tau=\sigma=0$ and so corresponds
to the type {\bf A} backgrounds).
The function $Y$ is determined in terms of $P$ as
\begin{equation}\label{Y}
Y=\frac18(P'+N_f),
\end{equation}
while the only remaining unknown, the function $P$, then satisfies the
{\it new} decoupled second order master equation
\begin{equation}\label{masterf}
P''+(P'+N_f)\left(\frac{P'+Q'+2N_f}{P-Q}+\frac{P'-Q'+2N_f}{P+Q}-4\coth(2\rho-2\rho_o)\right)=0.
\end{equation}
One can redefine $P(\rho)=N_c p(\rho)$ and factor out $N_c $ from the master
equation.
We will mention some solutions to eq. (\ref{masterf}), that explicitly
include the quotient
$x=\frac{N_f}{N_c}$, hence the solutions will capture the nontrivial
physics of the fields transforming in the fundamental representation of
the gauge group.
\subsection{Study of solutions}
We now describe various solutions to the ``flavored'' master equation (\ref{masterf}).
Some solutions were found exactly, for the particular relation $N_f=2N_c$ while some other
are known as asymptotic expansions, near the UV (large $\rho$) and the IR (small $\rho$). In
these latter cases, a smooth numerical interpolation can be found.
\subsubsection{Exact solutions for $N_f=2N_c$}
One can find some {\it exact} solutions for the
case $N_f=2N_c$ or $x=2$.
They were first discussed in the papers
\cite{Casero:2006pt},\cite{HoyosBadajoz:2008fw}.
For $N_f=2N_c$ an exact type {\bf A} ($\rho_o\rightarrow -\infty$) solution of eq. (\ref{masterf}) is:
\begin{equation}\label{self-dual-1}
P=N_c+\sqrt{N_c^2+Q_o^2}, \quad Q=Q_o\equiv
4N_c\frac{(2-\xi)}{\xi(4-\xi)},\quad 0<\xi<4\,\,.
\end{equation}
Another solution with a qualitatively different UV behavior is:
\begin{equation}\label{self-dual-2}
P=\frac{9 N_c}{4}+c e^{4\rho/3},
\quad c>0,\quad Q=\pm\frac{3N_c}{4}\,\,.
\end{equation}
One can check that in these solutions the radial coordinate moves all over
the real axis and that for $\rho\rightarrow -\infty$ the solutions take the same
form, but as anticipated above, differ substantially in the far UV, for
$\rho\rightarrow \infty$. Also, for the case $N_f=2N_c$ the papers
\cite{Caceres:2007mu} discuss some extra solutions
apart from the ones
mentioned, including,
interestingly, the generalization to near-extremal solutions\footnote{
The metric for the simplest non-extremal solution can be written in
terms of a constant $\xi$ and a function ${\cal F}=1- (z_h/z)^4$ as:
\begin{eqnarray}
ds^2&=&e^\frac{\phi_0}{2}\,z
\Big[-{\cal F}dt^2 + dx_1^2+ dx_2^2+ dx_3^2+ N_c \Big(
\frac{4}{z^2}{\cal F}^{-1}
dz^2 +\frac{1}{\xi} ( d\theta^2+\sin^2\theta d\varphi^2)+\nonumber\\
&&+\frac{1}{4-\xi}(d\tilde\theta^2+\sin^2\tilde\theta d\tilde\varphi^2)
+\frac14 (d\psi + \cos \theta d\varphi + \cos \tilde \theta d\tilde\varphi)^2
\Big)\Big]\,\,.
\label{bhflav}
\end{eqnarray}
The solution also contains non-trivial RR $F_{(3)}$ and dilaton, see
\cite{Casero:2006pt} for details.
Different features of this black hole solution have been analysed in \cite{Bertoldi:2007sf},
\cite{talaveraetal}. An important remark is that the
theory is in a Hagedorn phase and, indeed, the temperature coincides with the
Hagedorn temperature of Little String Theory.
For this reason, this solution is a bit problematic for studying the effect of
quarks in a field theory plasma,
unlike the finite temperature
solution of section \ref{D3D7plasma}.
}.
\subsubsection{Asymptotic expansions of generic solutions}
Other solutions of interest have been discussed in the papers
\cite{Casero:2006pt},
\cite{Casero:2007jj},\cite{HoyosBadajoz:2008fw}. We will summarize the
results, but suggest to the interested reader to go over those papers for
details of all the metric functions.
In the UV (for $\rho\rightarrow\infty$), two possible asymptotics were found, that
were called Class I and Class II in \cite{HoyosBadajoz:2008fw}.
Table \ref{UV-classes} summarizes the situation.
\begin{table}
\begin{center}
\begin{tabular}{|c||l|l|}
\hline
$N_f$ & I & II \\
\hline\hline &&\\
$<2N_c$ & $P\sim Q\sim|2N_c-N_f|\rho$ \mbox{}& \\
& $e^{2h}\sim\frac12\left(2N_c-N_f\right) \rho $
&\\
& $e^{2g}\sim N_c$ &\\
& $Y\sim \frac{N_c}{4}$ &\\
& $e^{4(\phi-\phi_o)}\sim \frac{e^{4 \left(\rho -\rho
_o\right)}\sinh^2\left(2 \rho _o\right)}
{2N_c^2\left(2 N_c- N_f\right) \rho }$ &\\
& $a\sim \frac{2}{N_c}\left(2N_c- N_f\right) e^{-2 \left(\rho
-\rho _o\right)}\rho $ &\\
& &\\ \cline{1-2} && \\
$>2N_c$ & $P\sim -Q\sim|2N_c-N_f|\rho$ & $P\sim c_+e^{4\rho/3}$ \\
& $e^{2h}\sim \frac{1}{4} \left(N_f-N_c\right)$ &
$e^{2h}\sim \frac14 c_+e^{4\rho/3}$ \\
& $e^{2g}\sim \frac{1}{2} \left(N_f-2 N_c\right) \rho$
& $e^{2g}\sim c_+e^{4\rho/3}$\\
& $Y\sim \frac{1}{4} \left(N_f-N_c\right)$ &
$Y\sim \frac16 c_+e^{4\rho/3}$\\
& $e^{4(\phi-\phi_o)}\sim \frac{e^{4 \left(\rho -\rho
_o\right)}\sinh^2\left(2 \rho _o\right)}
{2 \left(N_c-N_f\right){}^2 \left(N_f-2 N_c\right) \rho }$
& $e^{4(\phi-\phi_o)}\sim 1 $ \\
& $a\sim e^{-2 \left(\rho -\rho _o\right)}\rho $ & $a\sim2
e^{-2(\rho-\rho_o)} $\\
&&\\ \cline{1-2} && \\
$=2N_c$ & $P\sim N_c+\sqrt{N_c^2+Q_o^2}\sim \frac{8 N_c}{(4-\xi ) \xi }$ &
\\
& $e^{2h}\sim \frac{N_c}{\xi }$ &\\
& $e^{2g}\sim \frac{4 N_c}{4-\xi }$ &\\
& $Y\sim \frac{N_c}{4}$ &\\
& $e^{4(\phi-\phi_o)}\sim e^{4 \left(\rho -\rho
_o\right)}\sinh^2\left(2 \rho _o\right)\frac{(4-\xi ) \xi }
{16 N_c^3}$ &\\
& $a\sim \frac{4}{\xi }e^{-2 \left(\rho -\rho _o\right)}$
&\\
\hline
\end{tabular}
\caption{The two classes of leading UV behaviors.}
\label{UV-classes}
\end{center}
\end{table}
In the IR ($\rho \rightarrow 0$), three
types of
solutions were found, called Type I, II and III (there exist other,
qualitatively different solutions reported in \cite{gmnp} ). The function
$P(\rho)$ in these cases is,
\begin{eqnarray}
P=-N_f\rho+P_o+\frac43c_+^3P_o^2\rho^3-2c_+^3N_fP_o\rho^4+\frac{4}{5}c_+^3\left(\frac{4}{3}P_o^2+N_f^2\right)\rho^5+{\cal O}\left(\rho^6\right),
\end{eqnarray}
for Type I.
For the Type II asymptotics, we assume that this behavior occurs when the
IR
is located at $\rho_{IR}>\rho_o$. Without loss of generality we can choose
$\rho_{IR}=0$. With this choice we then
necessarily have $\rho_o<0$. Expanding $Q$ in (\ref{Q}) around $\rho=0$ we
obtain
\begin{equation}\label{Q-exp}
Q=b_0+b_1\rho+{\cal O}(\rho^2),
\end{equation}
where
\begin{eqnarray}
b_0&=&-\coth(2\rho_o)\left(Q_o+\frac{2N_c-N_f}{2}\right)-\frac{2N_c-N_f}{2},\nonumber\\
b_1&=&-\frac{2}{\sinh^2(2\rho_o)}\left(Q_o+\frac{2N_c-N_f}{2}\right)-(2N_c-N_f)\coth(2\rho_o)\,\,.
\end{eqnarray}
Looking for IR solutions of (\ref{masterf}) we
find that we must require that $b_0>0$. The
corresponding asymptotic solution then takes the form
\begin{eqnarray}\label{G-II-1}
P&=&Q+h_1\rho^{1/2}-\frac{1}{6b_0}\left(h_1^2+12b_0(b_1+N_f)\right)\rho\nonumber\\
&&+\frac{h_1}{72b_0^2}\left(5h_1^2+6(5b_1+2N_f)b_0-72b_0^2\coth(2\rho_o)\right)\rho^{3/2}+{\cal O}(\rho^2),
\end{eqnarray}
where $h_1$ is an arbitrary constant. Note that this expansion for $P$
admits a smooth limit when $\rho_o\rightarrow-\infty$ and so it
is valid for both solutions of type {\bf A } ($\rho_o\rightarrow-\infty$).
Finally, for Type III asymptotics we consider
$\rho_{IR}>\rho_o$ and we take $\rho_{IR}=0$. In terms of the expansion
(\ref{Q-exp}) this requires that $b_0=0$. We then find,
\begin{eqnarray}\label{III}
P=h_1\rho^{1/3}-\frac{9N_f}{5}\rho-\frac{2h_1}{3}\coth(2\rho_o)\rho^{4/3}-\frac{1}{175h_1}\left(50b_1^2-18N_f^2\right)\rho^{5/3}
+{\cal O}(\rho^2),
\end{eqnarray}
where $h_1\neq 0$ is an arbitrary constant.
To leading order the solutions for large $\rho$ - UV solutions - are quoted in table
\ref{UV-classes}. It is the presence of subleading terms that allow the smooth numerical
interpolation with three possible IR behaviors discussed.
The physics of the dual field theory encoded in these solutions was
discussed in detail in the papers \cite{Casero:2006pt},\cite{HoyosBadajoz:2008fw},\cite{Casero:2007jj}
by computing various
observables using the string solution of eq. (\ref{nonabmetric424})
evaluated on the solutions above\footnote{Finding a numerical
interpolation between the IR solutions and the solutions of Class
I in the far UV is (numerically) delicate. One can see some plots in
section V of the paper \cite{Nunez:2009da} .}. We move now to discuss
general features of the dual field theory.
\subsection{The dual field theory}
The proposal here is the following: without the addition of the flavor branes, the field
theory is
known to be a twisted version of six-dimensional Yang-Mills,
or as we discussed above, a four
dimensional QFT with an infinite number of massive fields.
See section \ref{qftwof}.
To get an {\it intuitive} understanding of the modifications of the dynamics produced by the
``quark'' fields (that feature below), we will consider that all the infinite massive fields
are chiral multiplets
and then argue that the dynamics is ruled by a lagrangian of the form:
\begin{equation}
L= Tr[-\frac{1}{4}F_{\mu\nu}^2 -i \bar{\lambda}\gamma^\mu D_\mu \lambda +
L(\Phi_k,W)]\sim \int d^2\theta W_\alpha W^\alpha +\sum_k \int d^4
\theta\Phi^{\dagger}_k e^{V}\Phi_k +\int d^2\theta \mu_k |\Phi_k|^2+\cdots\,\,.
\end{equation}
When the
flavor branes come into play, we are adding ``quark superfields'' that are
realized as the
open
strings going from the non-compact flavor branes to the compact (or wrapped) color branes (as
usual, the open strings that begin and end on a
flavor brane decouple and do not contribute to the
four dimensional dynamics). More concretely, we add the quark and anti-quark superfields
($Q,\tilde{Q}$) and propose that we have a lagrangian
for the massive fields interacting with the quark-antiquark superfields $Q, \tilde{Q}$
schematically of the form:\footnote{For more
details, see the appendix A of the paper \cite{HoyosBadajoz:2008fw} }
\begin{equation}
L(\Phi_k,W, Q,\tilde{Q})= \sum_k \int d^4\theta \Phi_k^{\dagger} e^{V}\Phi_k + \kappa_k\int
d^2\theta
\tilde{Q}\Phi_k
Q
+\mu_k |\Phi_k|^2+\cdots\,\,,
\end{equation}
and canonical kinetic terms for $(Q,\tilde{Q})$. In this system, the $SU(N_f)_L\times
SU(N_f)_R$ symmetry is explicitly broken to
the diagonal $SU(N_f)_D$ by the presence of the coupling $\tilde{Q}\Phi_k Q$. In this
respect,
the theory is qualitatively different from ${\cal N}=1$ SQCD.
One may be interested in the theory at low energies and hence integrate out the massive
fields (either
massive vectors or massive chirals) and after some algebra end with a theory of the form (again
schematically):
\begin{equation}
L= \int d^4\theta\Big( Q^\dagger e^{V} Q + \tilde{Q}^\dagger e^{-V}\tilde{Q}\Big) + \int
d^2\theta W_\alpha
W^\alpha + \kappa \int d^2\theta (\tilde{Q}Q)^2\,\,,
\label{flavorslagzz}\end{equation}
where we have a (naively irrelevant) deformation of ${\cal N}=1$ SQCD.
We emphasize that this is an {\it intuitive} way of understanding
the field theory dual to the
flavored system described above.
As we will summarize below there are various observables that can be
computed that match the predicted (or expected) result. So, the {\it precise} dual QFT should be
something similar to what we described above, or at least with the same qualitative physics.
\subsection{Checks and predictions}
This subsection summarizes results developed in the papers
\cite{Casero:2006pt},
\cite{HoyosBadajoz:2008fw},
\cite{Casero:2007jj}. There is a point that should be emphasized here. All the solutions to the BPS
equations or the master equation (\ref{masterf})
that have been found up to the time of writing this review, present a singularity in the IR (typically at
$\rho=0$). In spite of this being a ``good singularity'' according to some criteria developed in
the literature \cite{Maldacena:2000mw};
the presence of the singularity makes the interpretation of IR observables a bit
unclear. In other words, though one gets the ``correct or expected'' result one should perhaps handle
those particular computations with care.
Let us then concentrate on various quantities computed in the UV and then we will specify some that
are mostly influenced by the IR of the geometry.
\subsubsection{Beta function and anomalies}
The gauge coupling and the theta angle of the dual QFT can be defined as explained in various places,
see for example section 4.1 of the paper \cite{Casero:2007jj} or section 5 of the paper
\cite{HoyosBadajoz:2008fw}. One gets, after some algebra,
that the gauge coupling is related to the functions of the background as:
\begin{equation}
\frac{8\pi^2}{g^2}= e^{-\tau}P\,\,.
\end{equation}
Choosing a particular radius-energy relation, that was discussed in
\cite{Apreda:2001qb},
\cite{DiVecchia:2002ks},
one can compute the variation of the coupling with
respect to energy. Using the solutions where the
dilaton asymptotes to a linear function $(e^{4\phi}\sim\frac{e^{4\rho}}{\rho})$ and working to
leading order in an
expansion in inverse powers of the radial coordinate, we get
\begin{equation}
\beta_{\frac{8\pi^2}{g^2}}=\frac{3}{2}(2N_c-N_f)\,\,,
\end{equation}
that coincides with the result predicted by the NSVZ result, once we assign
anomalous dimensions to the quark superfields $\gamma_Q=\gamma_{\tilde{Q}}=-\frac{1}{2}$.
Similarly, one can define a geometrical quantity that can be associated with the quartic coupling.
See section 4.2 of
the paper \cite{Casero:2007jj}. The beta function can be computed using the anomalous dimensions
discussed above and again get matching with the interesting fact that for $N_f>2N_c$ the quartic
coupling is irrelevant, for $N_f<2N_c$ the coupling is relevant while for $N_f=2N_c$ the coupling is
not running. See \cite{Strassler:2005qs} for a nice explanation of this fact.
One can also assign
a value of the R-charge to the quark superfields to get the correct R-symmetry
transformation properties of the quartic superpotential of
eq. (\ref{flavorslagzz}), that is $R[Q]= R[\tilde{Q}]=\frac{1}{2}$.
This predicts that the R-symmetry
anomaly, the triangle with one R-current and two gauge currents is proportional to the
quantity ($2N_c-N_f$) times the phase by which we are rotating the fermions. This is the
precise result that
the string background gives. Indeed, if we compute the $\Theta$-angle as explained in section 4.1 of
the paper \cite{Casero:2007jj} or in section 5 of the paper
\cite{HoyosBadajoz:2008fw}, we will get
\begin{equation}
\Theta=\frac{\psi-\psi_0}{2}(2N_c-N_f)\,\,,
\end{equation}
where we associated $\frac{\psi-\psi_0}{2}$ with the change in phase of the fermions in the
quark multiplet
and the gauge multiplet to get perfect matching.
In the same vein, it is possible to attempt a 't Hooft matching of anomalies, that is of triangles
involving three global currents. The reader will find it quite instructive to go over section
4.7 of the
paper \cite{Casero:2007jj}. There, a detailed study of the matching of the
correlator of three global
currents -some of them corresponding to discrete symmetries,
some of them being continuous symmetries-
is presented. The treatment is performed in the case of the Type {\bf A} backgrounds, that
are
characterized by the fact that the functions $a=b=0$ in eq. (\ref{nonabmetric424}).
This
translates to the fact that the R-symmetry is broken to ${\mathbb Z}_{2N_c-N_f}$ without the further
(spontaneous)
breaking
to ${\mathbb Z}_2$.
\subsubsection{Seiberg duality}
\label{sec:Seib}
It is known that Seiberg duality manifests beautifully in a QFT like the one
of eq. (\ref{flavorslagzz}). This is explained in section 1.10 of \cite{Strassler:2005qs}. The
backgrounds discussed here show this in a very nice way. Indeed, as discussed for example in the
paper \cite{HoyosBadajoz:2008fw}, we can see that the master equation (and the whole system) is
invariant under the change,
\begin{equation}
P\rightarrow P,\; Q\rightarrow -Q,\;\; \sigma\rightarrow - \sigma,\;\;\; N_c\rightarrow N_f-N_c,\;\; N_f\rightarrow N_f\,\,,
\label{changeszzz}
\end{equation}
while all other functions are invariant.
Geometrically, this change amounts to swapping the two $S^2$ in the background,
namely those parameterized by $\theta,\varphi$ and $\tilde \theta, \tilde \varphi$ in
(\ref{nonabmetric424}).
This should be interpreted as follows:
suppose that we are presented with
a background, representing the dynamics of a field theory with $N_c$ colors and $N_f$
flavors. This
implies that we have a particular solution to the master equation (\ref{masterf}) for the
function $P$.
With this solution, and applying the changes of eq. (\ref{changeszzz}), we can construct another
solution, that will be related to the first one by a differomorphism and that will describe the
physics of a field theory with $N_f-N_c$ colors and $N_f$ flavors. Various aspects of this
interesting duality have been discussed in the papers \cite{Casero:2007jj},
\cite{HoyosBadajoz:2008fw}, \cite{Elander:2009bm}
and probably elsewhere.
The implementation of Seiberg duality in ${\cal N}=1$ subcritical string models
was discussed in
\cite{Murthy:2006xt}, \cite{Ashok:2007sf}. One can think of the sphere exchange mentioned above as being the
geometrical version of the mechanism described in these papers.
Interestingly, similar methods were used to propose a non-supersymmetric Seiberg duality in
\cite{Armoni:2008gg}.
\subsection{IR Physics: Domain Walls and some comments on Wilson/'t Hooft Loops}
One observable that can be computed and that strongly depends on the $\rho\rightarrow 0$ region of the solution
(the IR) is the tension of domain walls. Indeed, domain walls can be thought of as D5-branes that
wrap a three cycle inside the internal six dimensional manifold and that extend on two of the
Minkowski directions (and time, of course). We can compute the tension of a wall by considering a
probe D5-brane that sits on the manifold
$\Sigma_6=[t,x_1,x_2,\tilde{\theta},\tilde{\varphi},\psi]$, at
constant $\rho=0$, constant $\theta,\varphi$. The Born-Infeld action for this probe can be computed and
one reads that the effective tension is given by (see for example section 5.6 of the paper
\cite{Casero:2006pt}):
\begin{equation}
T_{DW}= 4 \pi^2 e^{2\phi+ 2g +k}\,\, T_{D5}\,\,,
\end{equation}
which, when evaluated at $\rho=0$, gives a constant proportional to $2N_c-N_f$. This is a good example
for an observable, since even when a singularity is present (this typically reflects in some of the
functions of the background being divergent), the combination above is finite. This is
typical of ``good
singularities''.
There are other observables that can be computed using D-branes or fundamental strings, examples of
these are Wilson or 't Hooft loops. A similar
conspiracy of
functions that avoids an infinite result occurs here.
Nevertheless, one should be quite careful with these
quantities as noted above. Indeed, it was found in section V of the paper \cite{Nunez:2009da}
that for the particular case of the backgrounds studied in this section, the Nambu-Goto action for
the fundamental string might cease to be a good approximation as the string develops a cusp when
approaching the singularity.
In other words, we believe that the solutions presented in this section surely capture correctly
many UV aspects of the field theory, together with some IR observables. Probably, we could think of
the presence of the singularity in the same way as we think about the singularity in the
Klebanov-Tseytlin background\footnote{In contrast to what happens for the solutions
discussed here, the Klebanov-Tseytlin background presents a {\it bad} singularity and the
IR physics computed with that solution is not trustable.}, that captures some of
the physics, but some is lost and the singularity
must be resolved. There are different ways of attempting a resolution of the singularity; for
example, considering massive quarks. This is under present study.
\subsubsection{Wilson loops and first order phase transitions}
\label{CNPwilson}
One can consider the situation (of course, this is an idealized situation) in which all the flavors are
massive and with a fixed sharp mass $M_0$ (corresponding to a given value of the
radial coordinate, that we call $\rho=m_0$). The way to model this in a first approximation is
to
consider $N_f(\rho)=N_f \Theta(\rho-m_0)$, where $\Theta(\rho)$ is a Heaviside
step function. Once again, there is
some
dynamics that is being lost in doing this, for example, the matching of the derivatives of the
solutions is not smooth at the point $\rho=m_0$ and the curvature of the background is
not well defined at that point.
Nevertheless, it is possible (within this approximately correct way of proceeding) to find a solution
that for energies below the scale set by $m_0$ corresponds to the theory {\it without} flavors, say
those discussed around eq. (\ref{solutionnear0xx}) and far in the UV corresponds to the flavored
theory, as represented by solutions of the Class I in table \ref{UV-classes}. One can then compute
the Wilson loop following the well-known prescription \cite{Maldacena:1998im}. This was done explicitly in
\cite{Bigazzi:2008gd}. The qualitative result is the following: for a range of ratios between
the
mass of the quarks $M_0$ and the
value of the gaugino condensate set by the function $a(\rho)$ one observes that the
relation between the quark-antiquark potential $V_{QQ}$ and their separation $L_{QQ} $ presents a
first order
transition\footnote{Notice that we should talk of a
``Quantum'' phase transition, as the system is at zero temperature.}.
In other words, a point where $\frac{dV_{QQ}}{d L_{QQ}}$ is discontinuous. The
same kind of behavior was observed in systems
where a more careful study is possible. Indeed, in the backreacted
Klebanov-Witten (see section \ref{AdS5X5}) and Klebanov-Strassler
(see section \ref{KS}) models,
it was possible to
find the precise form for the function $p(\rho)$ - that shows that the ``Heaviside
approximation'' described above is not
a bad one.
The same qualitative behavior for the first order phase
transition was found in
the papers
\cite{Bigazzi:2008zt}, \cite{Bigazzi:2008qq}, \cite{Ramallo:2008ew}.
This kind of first order transitions for the Wilson loop string configurations
are by no means particular of systems with dynamical fundamental fields.
In fact, they were first found in a different context in \cite{sfetsos},
where a nice connection between these Wilson loops computations and the Van der Waals
gas
(paradigm of the first order transition) was put forward
(further discussions can be found for instance in
\cite{Bigazzi:2008qq},
\cite{Nunez:2009da}).
Different examples of such phase transitions in systems without flavors
have been worked out in
\cite{vandertrans}.
The ``morale'' seems to be the following: when we have a physical system that has two {\it
independent}
scales (that is two scales that can be tuned independently, in the present example, the mass
of the quarks $M_0$ and the gaugino condensate $\Lambda^3$) the first order phase transition
for the
quantity $V_{QQ}(L_{QQ})$ will be present.
Of course, like in any other
first order transition, it
will happen that the discontinuity in the derivative will disappear for some ratio between
the scales
mentioned above.
\setcounter{equation}{0}
\section{A dual to a (2+1)-dimensional ${\cal N}=1$ SQCD-like model}
\label{sec:2+1}
In this section we will study gravity duals to minimal supersymmetric theories in 2+1 dimensions. These backgrounds can be obtained by wrapping D5-branes along three-cycles of manifolds with $G_2$ holonomy \cite{Chamseddine:2001hk,Maldacena:2001pb,Schvellinger:2001ib}. The corresponding field theory dual is a (2+1)-dimensional ${\cal N}=1$ supersymmetric $U(N_c)$ Yang-Mills theory with a level $k$ Chern-Simons interaction. Such a theory coupled to an adjoint massive scalar field should arise on the domain walls separating the different vacua of pure ${\cal N}=1$ super-Yang-Mills in 3+1 dimensions. The corresponding unflavored background was studied in ref. \cite{Maldacena:2001pb}, where it was argued to be dual to a $U(N_c)$ gauge theory with Chern-Simons level $k=N_c/2$. In what follows we will review a generalization of these results, following closely ref. \cite{Canoura:2008at}. We will present the deformation of the background of
\cite{Chamseddine:2001hk,Maldacena:2001pb,Schvellinger:2001ib} induced by a smeared distribution of massless flavors. In order to formulate these generalized backgrounds, let $\sigma^i$ and $\omega^i$ $(i=1,2,3)$ be two sets of SU(2) left-invariant one forms, obeying:
\begin{equation}
d\sigma^i=-{1\over 2}\,\epsilon_{ijk}\,\sigma^j\wedge \sigma^k\,\,,
\qquad\qquad
d\omega^i=-{1\over 2}\,\epsilon_{ijk}\,\omega^j\wedge \omega^k\,\,.
\label{sigma-w}
\end{equation}
The forms $\sigma^i$ and $\omega^i$ parameterize two three-spheres. In the geometries we will be dealing with, these spheres are fibered by a one-form $A^i$. The corresponding ten-dimensional metric of the type IIB theory in the Einstein frame is given by:
\begin{equation}
ds^2\,=\,e^{{\phi\over 2}}\left[ dx_{1,2}^2+dr^2+\frac{e^{2h}}{4}(\sigma^i)^2+\frac{e^{2g}}{4}(\omega^i-A^i)^2\right]\,\,,
\label{2+1ansatz}
\end{equation}
where $\phi(r)$ is the dilaton of type IIB supergravity and $g$ and $h$ are functions of the radial variable $r$. In addition, the one-form $A^i$ will be taken as:
\begin{equation}
A^i\,=\,{1+w(r)\over 2}\,\,\sigma^i\,\,,
\label{Ai-w}
\end{equation}
with $w(r)$ being a new function of $r$. For convenience in this section we will take $g_s=\alpha'=1$, as we did in section \ref{sec:D5D5}. The backgrounds considered here are also endowed with a RR three-form $F_{3}$. We will represent $F_{3}$ as the sum of two contributions:
\begin{equation}
F_3\,=\,{\cal F}_3\,+\,f_3\,\,,
\label{F3f3}
\end{equation}
where $d{\cal F}_3=0$ and $f_3$ is the part of the RR three-form which is responsible for the violation of the Bianchi identity ($df_3\not=0$) and which is sourced by the flavor D5-branes. Let us first parametrize the component ${\cal F}_3$ as:
\begin{equation}
\frac{{\cal F}_3}{N_c}=-\frac{1}{4}(\omega^1-B^1)\wedge(\omega^2-B^2)\wedge(\omega^3-B^3)+\frac{1}{4}F^i\wedge(\omega^i-B^i)+H\,\,,
\label{F3ansatz}
\end{equation}
where $B^i$ is a new one-form and $F^i$ are the components of its field strength, given by:
\begin{equation}
F^i\,=\,dB^i\,+\,{1\over 2}\,\epsilon_{ijk}\,B^j\wedge B^k\,\,.
\label{Fi}
\end{equation}
In (\ref{F3ansatz}) $H$ is a three-form that is determined by imposing the Bianchi identity for ${\cal F}_3$, namely:
\begin{equation}
d{\cal F}_3\,=\,0\,\,.
\label{Bianchi-id}
\end{equation}
By using (\ref{sigma-w}) one can easily check from the explicit expression written in
(\ref{F3ansatz}) that, in order to fulfill (\ref{Bianchi-id}), the three-form $H$ must satisfy the equation:
\begin{equation}
dH\,=\,{1\over 4}\,F^i\wedge F^i\,\,.
\label{dH}
\end{equation}
In what follows we shall adopt the following ansatz for $B^i$:
\begin{equation}
B^i\,=\,{1+\gamma(r)\over 2}\,\,\sigma^i\,\,,
\label{Bi}
\end{equation}
where $\gamma(r)$ is a new function. After plugging the ansatz of $B^i$ written in (\ref{Bi}) into (\ref{Fi}), one gets the expression for $F^i$ in terms of $\gamma(r)$:
\begin{equation}
F^i\,=\,{\gamma'\over 2}\,\,dr\wedge \sigma^i\,+\,
{\gamma^2-1\over 8}\,\,\epsilon_{ijk}\,\sigma^j\wedge \sigma^k\,\,,
\label{Fi-gamma}
\end{equation}
where the prime denotes the derivative with respect to the radial variable $r$. Using this result for $F^i$ in (\ref{dH}) one can easily determine the three-form $H$ in terms of $\gamma$. Let us parameterize $H$ as:
\begin{equation}
H\,=\,{1\over 32}\,\,{1\over 3!}\,\,{\cal H}(r)\,\,\epsilon_{ijk}\,\,
\sigma^i\wedge \sigma^j\wedge\sigma^k\,\,.
\end{equation}
Then, by solving (\ref{dH}) for $H$, one can verify that
${\cal H}(r)$ is the following function of the radial variable:
\begin{equation}
{\cal H}\,=\,2\gamma^3-6\gamma\,+\,8\kappa\,\,,
\label{calH}
\end{equation}
with $\kappa$ being an integration constant.
Let us now consider the contribution $f_3$ to the RR three-form $F_3$. As explained above, this contribution violates the Bianchi identity and is non-zero when flavor branes are present. Indeed, let us write the WZ term of the action of a system of flavor D5-branes as:
\begin{equation}
S_{flavor}^{WZ}\,=\,T_5\,\int_{{\cal M}_{10}}\,\Omega\wedge C_6\,\,,
\label{C6-Omega}
\end{equation}
with $\Omega$ being a four-form with components along the space transverse to the worldvolume of the branes. Then, the coupling to the RR potential $C_6$ written in (\ref{C6-Omega}) gives rise to the following modified Bianchi identity:
\begin{equation}
dF_3\,=\,df_3\,=\,{4\pi^2\Omega}\,\,.
\label{Bianchi2+1}
\end{equation}
To write an specific ansatz for $\Omega$ and $f_3$ we have to select some family of supersymmetric embeddings for the flavor branes. As explained above, this can be done by using kappa symmetry. In the simplest case one looks for massless embeddings, which extend along the full range of the radial coordinate $r$. Those are the configurations considered in \cite{Canoura:2008at}, in which the D5-brane is extended along the three Minkowski directions $x^{\mu}$, as well as along a three-dimensional cylinder spanned by $r$ and two other angular directions. Actually, it was shown in \cite{Canoura:2008at} that these two angular directions could be the ones corresponding to $\sigma^{3}$ and $\omega^3$. The corresponding transverse volume for this configuration is just:
\begin{equation}
{\rm Vol}(\,{\cal Y}_4^{1,2}\,)\,=\,
\sigma^1\wedge \sigma^2\wedge \omega^1\wedge \omega^2\,\,.
\label{Vol12}
\end{equation}
However, there is nothing special in our background about these directions. Indeed, both in the metric and in the RR three-form ${\cal F}_3$, we are adopting a round ansatz which does not distinguish among the directions of the two three-spheres. Thus we could as well consider supersymmetric cylinder embeddings that span the $1, \hat 1$ or $2, \hat 2$ directions. The volume forms of the spaces transverse to these embeddings are clearly:
\begin{equation}
{\rm Vol}(\,{\cal Y}_4^{2,3}\,)\,=\,\sigma^2\wedge \sigma^3\wedge \omega^2\wedge \omega^3\,\,,
\qquad\qquad
{\rm Vol}(\,{\cal Y}_4^{1,3}\,)\,=\,\sigma^1\wedge \sigma^3\wedge \omega^1
\wedge \omega^3\,\,.
\label{Vol23-Vol13}
\end{equation}
To construct a backreacted supergravity solution with the same type of ansatz as in (\ref{2+1ansatz}) we should consider a brane configuration that combines these three possible types of embeddings in an isotropic way. The corresponding transverse volume form ${\rm Vol}(\,{\cal Y}_4)$ of this three-branch brane system would be just the sum of the three four-forms written in eqs. (\ref{Vol12}) and (\ref{Vol23-Vol13}). The corresponding smearing form $\Omega$ is obtained by multiplying by the suitable normalization factor, namely :
\begin{equation}
\Omega\,=\,-{N_f\over 16\pi^2}\,\,{\rm Vol}(\,{\cal Y}_4\,)\,=\,-{N_f\over 64\pi^2}\,
\epsilon_{ijk}\,\epsilon_{ilm}\,
\sigma^j\wedge \sigma^k\wedge \omega^l\wedge \omega^m\,\,,
\label{Omega2+1}
\end{equation}
where the minus sign has its origin in the different orientation (required in the kappa symmetry analysis of \cite{Canoura:2008at}) of the D5-brane worldvolume with respect to the ten-dimensional space. It is now straightforward to use the $\Omega$ written in (\ref{Omega2+1}) and get an expression of $f_3$ whose modified Bianchi identity is the one of (\ref{Bianchi2+1}). One has:
\begin{equation}
f_3=\frac{N_f}{8}\epsilon_{ijk}(\omega^i-\frac{\sigma^i}{2})\wedge\sigma^j\wedge\sigma^k\,.
\label{f3}
\end{equation}
Eq. (\ref{f3}) completes our ansatz for the general flavored case. Using these expressions of the metric and RR three-form in the supersymmetry variations of the dilatino and gravitino of type IIB supergravity, after imposing that the background preserves two supersymmetries, we arrive at a system of first-order BPS equations. These equations, which are rather involved, have been derived and analyzed in detail in \cite{Canoura:2008at}. They admit several consistent truncations which lead to simpler solutions. One can, for example, first consider the unflavored case $N_f=0$. If, in addition, we require that the function $g$ is constant and that the fibering functions $w$ and $\gamma$ are equal, our ansatz reduces to the one considered in \cite{Chamseddine:2001hk,Maldacena:2001pb,Schvellinger:2001ib}. Actually, in this case the BPS equations fix the value of $g$ to be $e^{2g}=N_c$ and, in order to have a regular solution one should take the constant $\kappa$ of (\ref{calH}) to be equal to $1/2$ ($\kappa$ is related to the Chern-Simons level $k$ of the dual field theory). Other unflavored solutions exist and have been studied in detail in ref. \cite{Canoura:2008at}. Here we will concentrate on reviewing the case in which $N_f\not=0$, starting from a particular truncation of the BPS system which is very interesting and serves to classify the different more involved solutions in the UV.
\subsection{The truncated system}
\label{Mnas-truncated}
In this section we will analyze the truncation of the general system of BPS equations that corresponds to taking $w=\gamma= \kappa=0$. In this case the BPS equations of
\cite{Canoura:2008at} for the remaining functions $h$ and $g$ of the metric and for the dilaton $\phi$ consistently reduce to the following simple system of differential equations:
\begin{eqnarray}
&&\phi'\,=\,N_c\,e^{-3g}\,\,-\,{3\over 4}\,(\,N_c-4N_f\,)\,e^{-g-2h}\,\,,\nonumber\\\rc
&&h'\,=\,{1\over 2}\,e^{g-2h}\,+\,{N_c-4N_f\over 2}\,\,e^{-g-2h}\,\,,\nonumber\\\rc
&&g'\,=\,e^{-g}\,-\,{1\over 4}\,e^{g-2h}\,-\,N_c\,e^{-3g}\,+\,{N_c-4N_f \over 4}\,
e^{-g-2h}\,\,.
\label{ab-system-flavor}
\end{eqnarray}
By inspecting the system (\ref{ab-system-flavor}) one readily realizes that there is a special solution for which the metric functions $h$ and $g$ are constant. Actually this solution only exists when $N_c<2N_f$ and the corresponding expressions for $g$ and $h$ are the following:
\begin{equation}
e^{2g}\,=\,4N_f-N_c\,\,,
\qquad\qquad
e^{2h}\,=\,{1\over 4}\,\,{(4N_f-N_c)^2\over 2N_f-N_c}\,\,,
\qquad\qquad (\,N_c\, <\,2N_f\,)\,\,,
\end{equation}
while the dilaton grows linearly with the holographic coordinate $r$, namely:
\begin{equation}
\phi\,=\,{2(3N_f-N_c)\over \big[\,4N_f-N_c]^{{3\over 2}}}\,\,r\,\,+\,\,\phi_0\,\,.
\end{equation}
Let us next consider solutions for which the function $h$ is not constant. In this case we can use $\rho=e^{2h}$ as a radial variable and one can define a new function $F(\rho)$ as $F(\rho)=e^{2g}$. It follows from (\ref{ab-system-flavor}) that the BPS equation for $F(\rho)$ is now:
\begin{equation}
{dF\over d\rho}\,=\,{(F-N_c)\,\Big(\,2-{F\over 2\rho}\,\Big)\,-\,{2N_f\over \rho}\,\,F
\over F+N_c-4N_f}\,\,,
\label{F-ab-flavor}
\end{equation}
while the equation for the dilaton as a function of $\rho$ can be written as:
\begin{equation}
{d\phi\over d\rho}\,=\,{N_c\over F(F+N_c-4N_f)}\,\Big[\,1\,-\,{3\over 4\rho}\,
\Big(\,1\,-\,{4N_f\over N_c}\,\Big)\,F\,\Big]\,\,.
\label{dilaton-ab-flavor}
\end{equation}
Moreover, from the second equation in (\ref{ab-system-flavor}) we can obtain the relation between the two radial variables $r$ and $\rho$, namely:
\begin{equation}
{dr\over d\rho}\,=\,{\sqrt{F(\rho)}\over F(\rho)+N_c-4N_f}\,\,.
\label{r-rho-flavored-ab}
\end{equation}
Notice that the sign of the right-hand side of (\ref{r-rho-flavored-ab}) could be negative when $N_f\not=0$. This means that we have to be careful in identifying the UV and IR domains in terms of the new radial variable $\rho$. We can use the result of integrating eqs. (\ref{F-ab-flavor})-(\ref{r-rho-flavored-ab}) to obtain the metric in terms of the new variable $\rho$, which takes the form:
\begin{equation}
ds^2\,=\,e^{{\phi\over 2}}\,\Bigg[\,dx^2_{1,2}\,+\,
\Big(\,{dr\over d\rho}\,\Big)^2\,(d\rho)^2\,+\,
{\rho\over 4}\,(\sigma^i)^2\,+\,{F\over 4}\,\Big(\,\omega^i\,-\,A^i\,\Big)^2
\,\Bigg]\,\,.
\label{metric-in-rho}
\end{equation}
Let us now study the different solutions of eqs. (\ref{F-ab-flavor})-(\ref{dilaton-ab-flavor}).
\subsubsection{Linear dilaton backgrounds}
\label{FMnas-abe-subsection}
When $N_f=0$, eq. (\ref{F-ab-flavor}) can be simply solved by taking $F=N_c$. However,
it is clear from (\ref{F-ab-flavor}) that in the flavored case $F=N_c$ is no longer a solution of the equations. Nevertheless, there are solutions for which this constant value of $F$ is reached asymptotically when $\rho\rightarrow\infty$. Indeed, one can check this fact by solving (\ref{F-ab-flavor}) as an expansion in powers of $1/\rho$. One gets:
\begin{equation}
F\,=\,N_c\,+\, N_c\, N_f\,{1\over \rho}\,-\,
{3\over 4}\, N_c\, N_f\,(N_c-4N_f)\,{1\over \rho^2}\,+\,\cdots\,\,,
\qquad\qquad (\rho\rightarrow\infty)\,\,.
\label{UVF}
\end{equation}
By plugging the expansion (\ref{UVF}) into (\ref{dilaton-ab-flavor}) one can prove that, when $N_c\not=2N_f$, these solutions have a dilaton that depends linearly on $\rho$ in the UV and, actually, one can verify that:
\begin{equation}
{d\phi\over d\rho}\,=\,{1\over 2(N_c-2N_f)}\,-\,
{3N_c^2\,-\,12 N_c N_f\,+\,16N_f^2\over 8(N_c-2N_f)^2}\,\,\,{1\over \rho}
\,+\,\cdots\,\,,
\qquad\qquad (\rho\rightarrow\infty)\,\,.
\label{UVdilaton}
\end{equation}
Notice the different large $\rho$ behavior of the dilaton in the two cases $N_c>2N_f$ and $N_c<2N_f$. Indeed, when $N_c>2N_f$ the dilaton grows linearly with the holographic coordinate $\rho$ (the behavior expected for a confining theory in the UV), while for $N_c<2N_f$ the field $\phi$ decreases linearly with $\rho$. This seems to suggest that the sign of the beta function of the dual gauge theory depends on $N_c$ and $N_f$ through the combination $N_c-2N_f$. Actually one can verify by means of a probe calculation in the complete system that the beta function is positive for $N_c>2N_f$ and changes its sign when $N_c<2N_f$ \cite{Canoura:2008at}.
Eq. (\ref{F-ab-flavor}) can be solved numerically by imposing the behavior (\ref{UVF}) for large $\rho$. Once $F(\rho)$ is known one can obtain the dilaton $\phi(\rho)$ by direct integration of the right-hand side of (\ref{dilaton-ab-flavor}). The result of this numerical calculation was analyzed in detail in \cite{Canoura:2008at}. Let us only mention here that, in the most interesting case $N_c>2N_f$, the function $F$ diverges for $\rho\rightarrow 0$, while the dilaton $\phi$ remains finite for small $\rho$. This bad IR behavior of $F$ is cured in the untruncated solution with the same leading UV form of $F$ and $\phi$ but with $w, \gamma\not=0$ (see below).
\subsubsection{Flavored $G_2$ cone}
\label{G2cone}
Let us now consider the solution of the equations (\ref{F-ab-flavor}) and (\ref{dilaton-ab-flavor}) that leads to a metric which is asymptotically a $G_2$-cone with constant dilaton in the UV. It can be checked that there exists a solution of (\ref{F-ab-flavor}) which can be expanded for large values of $\rho$ as:
\begin{equation}
F\,=\,{4\over 3}\,\,\rho\,+\,4(N_f-N_c)\,+\,
{15 N_c^2-39N_c N_f\,+\,24 N_f^2\over \rho}\,+\,\cdots\,\,.
\label{G2-F-flavored}
\end{equation}
The corresponding expansion for $\phi(\rho)$ is:
\begin{equation}
\phi\,=\,\phi_{*}\,-\,{9N_f\over 4}\,{1\over \rho}\,-\,{27\over 32}\,N_c\,(N_c+2N_f)\,
{1\over \rho^2}\,+\,\cdots\,\,,
\label{G2-int-phi-flavored}
\end{equation}
where $\phi_{*}$ is the constant limiting value of $\phi$ in the UV. In order to explore the asymptotic form of the metric for large $\rho$, it is convenient to perform a change in the radial variable, namely:
\begin{equation}
\rho\,=\,{1\over 3}\,\tau^2\,\,,
\end{equation}
in terms of which the metric asymptotically becomes the one corresponding to the direct product of (2+1)-dimensional Minkowski space and a seven-dimensional cone with $G_2$ holonomy, namely:
\begin{equation}
ds^2\,\approx\,e^{{\phi_{*}\over 2}}\,\Big[\,dx^2_{1,2}\,+\,
(d\tau)^2
\,+\,
{\tau^2\over 12}\,(\sigma^i)^2\,+\,{\tau^2\over 9}\,\,
\Big(\,\omega^i\,-\,{\sigma^i\over 2}\,\Big)^2
\,\Big]\,\,.
\label{UV-asymp-metric}
\end{equation}
To find the solution in the whole range of the radial coordinate one can numerically integrate the system (\ref{F-ab-flavor})-(\ref{dilaton-ab-flavor}) by imposing the asymptotic behavior (\ref{G2-F-flavored}) to the function $F(\rho)$. For $N_c\ge 2 N_f$ one can show that $F(\rho)$ is well-defined for $\rho>0$, while it diverges for $\rho\rightarrow 0$
(see \cite{Canoura:2008at} for further details). Notice that, at least in the unflavored case $N_f=0$, it is natural to regard these solutions with finite dilaton in the UV as corresponding to D5-branes wrapped on a three-cycle of a $G_2$ cone, in which the near horizon limit has not been taken and, thus, as we move towards the large $\rho$ region the effect of the branes on the metric becomes asymptotically negligible and we recover the geometry of the $G_2$ cone where the branes are wrapped.
\subsection{The complete system}
Let us now consider the solutions of the BPS equations for our general ansatz. These complete BPS equations have been derived in the appendix A of \cite{Canoura:2008at}. Here we will restrict ourselves from now on to the cases with $N_c> 2 N_f$, which are the ones that lead to more sensible solutions. As in the truncated case of section \ref{Mnas-truncated}, we will use $\rho=e^{2h}$ as radial variable and $F=e^{2h}$ as a function of $\rho$. In order to solve the general BPS equations we must impose initial conditions to the functions $w(\rho)$ and $\gamma(\rho)$ introduced in (\ref{Ai-w}) and (\ref{Bi}), and we must fix the value of the constant $\kappa$ of (\ref{calH}). These initial conditions are determined by imposing some regularity requirements at $\rho=0$ that we now review (see \cite{Canoura:2008at} for additional details). First of all, we will demand that the function $F$ approaches a constant finite value when $\rho\rightarrow 0$ (\hbox{\it i.e.}\ $F\sim F_0$ for $\rho\rightarrow 0$). In order to fix the value of the function $w(\rho)$ at $\rho=0$ let us recall (see (\ref{Ai-w})) that $w$ parameterizes the one-form $A^i$ which, in turn, determines the mixing of the two three-spheres in the ten-dimensional fibered geometry. The curvature of the gauge connection $A^i$ (defined as in (\ref{Fi}) with $B^i\rightarrow A^i$) determines the non-triviality of this mixing. When this curvature vanishes one can choose a new set of three one-forms in which the two three-spheres are disentangled in a manifest way and one can factorize the directions parallel and orthogonal to the color brane worldvolume in a well-defined way. From the wrapped brane origin of our solutions, one naturally expects such an un-mixing of the two $S^3$'s to occur in the IR limit $\rho=0$ of the metric. Moreover, by a direct calculation using (\ref{sigma-w}) it is easy to verify that for $w=1$ the curvature of the one-form $A^i$ vanishes . Thus, it follows that the natural initial condition for $w(\rho)$ is:
\begin{equation}
w(\rho=0)=1\,\,.
\label{initialw}
\end{equation}
Actually, the three-cycle that the color branes wrap can be identified with the one that shrinks when $\rho\rightarrow 0$, which is the one given by:
\begin{equation}
\Sigma\equiv\{\omega^i=\sigma^i\}.
\label{mixc}
\end{equation}
In order to have a non-singular flux at the origin, the RR three-form $F_3$ should vanish on $\Sigma$ when $\rho\rightarrow 0$. It is easy to check that this occurs if the constant $\kappa$ takes the value:
\begin{equation}
\kappa\,=\,{1\over 2}\,-\,{3N_f\over 2 N_c}\,\,.
\label{kappa-flavors}
\end{equation}
Actually, (\ref{kappa-flavors}) is also a necessary condition to have a finite dilaton at $\rho=0$. Indeed, it was shown in \cite{Canoura:2008at} that, in addition to (\ref{kappa-flavors}), the dilaton remains finite in the IR if the function $\gamma(\rho)$ takes the following value for $\rho=0$:
\begin{equation}
\gamma(\rho=0)\,=\,1\,-\,{2N_f\over N_c}\,\,.
\label{gamma-initial-flavored}
\end{equation}
Eqs. (\ref{initialw}) and (\ref{gamma-initial-flavored}) provide the initial conditions for the functions $w$ and $\gamma$ we were looking for.
\subsubsection{Asymptotic linear dilaton}
\label{FlaMnas-subsection}
As explained above,
we are interested in solutions of the BPS equations such that asymptotically $F$ is constant. Actually, by solving the BPS system in powers of $1/\rho$,
one can check that there are solutions in which $F$ has the following asymptotic
behavior:
\begin{equation}
F\,=\,N_c\,+\,{a_1\over \rho}\,+\,{a_2\over \rho^2}\,+\,{a_3\over \rho^3}\,+\,\cdots\,\,,
\label{F-inseries-untruncated}
\end{equation}
where the coefficients $a_1$, $a_2$ and $a_3$ are given by:
\begin{eqnarray}
&&a_1\,=\,N_f\, N_c\,\,,\nonumber\\\rc
&&a_2\,=\,-{3\over 4}\, N_c N_f(\,N_c\,-\,4N_f\,)\,\,,\nonumber\\\rc
&&a_3\,=\,{N_f\,N_c\over 16}\,
\Big[\,21N_c^2\,-\,148\,N_f\, N_c\,+\,240\,N_f^2\,\Big]\,\,.
\label{a-coefficients-untruncated}
\end{eqnarray}
Notice that the first two terms in (\ref{F-inseries-untruncated}) and
(\ref{a-coefficients-untruncated}) coincide with the one written in (\ref{UVF}) for the truncated system. Similarly, the functions $w$ and $\gamma$ can be represented as:
\begin{eqnarray}
&&w\,=\,{b_1\over \rho}\,+\,{b_2\over \rho^2}\,+\,{b_3\over \rho^3}\,+\,\cdots\,\,,\nonumber\\\rc
&&\gamma\,=\,{c_1\over \rho}\,+\,{c_2\over \rho^2}\,+\,{c_3\over \rho^3}\,+\,\cdots\,\,,
\label{wgamma-inseries-untruncated}
\end{eqnarray}
where the coefficients $b_i$ and $c_i$ are the following:
\begin{eqnarray}
&&b_1\,=\,c_1\,=\,{1\over 2}\,\,(N_c\,-3N_f)\,\,,\nonumber\\\rc
&&b_2\,=\,c_2\,=\,{5\over 8}\,\,(N_c\,-3N_f)\,(N_c-2N_f)\,\,,\nonumber\\\rc
&&b_3\,=\,{1\over 32}\,\,(N_c\,-3N_f)\,
\Big[\,49 N_c^2\,-\,184\, N_c\,N_f\,+\,204\,N_f^2\,\Big]\,\,,\nonumber\\\rc
&&c_3\,=\,{1\over 32}\,\,(N_c\,-3N_f)\,
\Big[\,49 N_c^2\,-\,208\,N_f N_c\,+\,252\,N_f^2\,\Big]\,\,.
\label{bc-coefficients-untruncated}
\end{eqnarray}
Moreover, for $N_c>2N_f$ the dilaton grows linearly with $\rho$ as in (\ref{UVdilaton}), \hbox{\it i.e.}\
$\phi\sim \rho/[2(N_c-2N_f)]$ for large $\rho$.
The solution for the full range of the holographic coordinate can be found by numerical integration of the BPS system with the IR regularity conditions (\ref{initialw}), (\ref{kappa-flavors}) and (\ref{gamma-initial-flavored}) and with $F(\rho=0)=F_0$ finite. One has to perform an interpolation between the $\rho\rightarrow 0$ and $\rho\rightarrow\infty$ behaviors by means of a shooting technique in which the only free parameter $F_0$ is varied until a solution with $F(\rho)\approx N_c$ for large $\rho$ is obtained (which only occurs when $F_0$ is fine tuned to a very precise value).
After obtaining this solution of the equations of motion of the gravity plus brane system, we can see if it incorporates some of the features that the supergravity dual of 2+1 dimensional gauge theory plus flavors should exhibit. In particular, we can study the evolution of the gauge coupling constant with the holographic coordinate. In order to do that, let us consider a D5-brane probe extended along the three Minkowski directions and wrapping the internal three-cycle $\Sigma$ defined in (\ref{mixc}) at a fixed value of the holographic coordinate $\rho$. By looking at the ${\cal F}^2$ terms in the DBI action of this probe, we get the value of the Yang-Mills coupling constant of the dual (2+1)-dimensional gauge theory, namely:
\begin{equation}
{1\over g^2_{YM}}\,\sim\,e^{-{3\over 4}\,\phi}\,\,
\int_{\Sigma}\,\,\sqrt{- \det \big(\,\hat G_3\,\big)}\,\, \,d^3\,\xi\,\,\sim\,
\Big[\,\rho+\,{F\over 4}\,(1-w)^2\,\Big]^{{3\over 2}}\,\,,
\label{gYM}
\end{equation}
where $\hat G_3$ is the induced metric on the three-cycle $\Sigma$ and we have neglected all constant numerical factors. Due to our initial condition (\ref{initialw}), the right-hand side of (\ref{gYM}) vanishes for $\rho=0$, which corresponds to having $g^2_{YM}\rightarrow\infty$ in the IR, as expected in a confining theory. Moreover, $1/g^2_{YM}$ grows as we move towards the UV region $\rho\rightarrow\infty$, in agreement with the expected property of asymptotic freedom. Other gauge theory observables for these backgrounds, such as the Wilson loops, can be also analyzed (see ref. \cite{Canoura:2008at}). Notice that, despite the regularity conditions we have imposed, in the flavored case $N_f\not=0$ the explicit calculation of the scalar curvature for the linear dilaton solutions shows that the metric is singular at the origin of the radial coordinate. Notice that, as argued for other backgrounds, it is physically reasonable to expect that massless flavors drastically alter the backreacted geometry in the deep IR. However, as our initial conditions are such that the dilaton is finite at the origin, the value of the $g_{tt}$ component of the metric is also bounded and then, according to the criterium of \cite{Maldacena:2000mw}, the singularity is ``good" and the background can be used to extract non-perturbative information of the dual gauge theory.
\subsubsection{Asymptotic $G_2$ cones}
\label{G2cones-untruncated-subsection}
When $F(\rho=0)$ takes values in a certain range, the solutions of the BPS equations lead to the metric (\ref{UV-asymp-metric}) at the UV, which is the direct product of 2+1 dimensional Minkowski space and a $G_2$ cone. The solutions in this case are very similar in the UV to the ones discussed in subsection \ref{G2cone} (with better IR behavior) and we will not discuss them further here. Let us only mention that the asymptotic values of $F$, $w$ and $\gamma$ for $\rho\rightarrow\infty$ can be determined analytically and are given by:
\begin{eqnarray}
&&F\,\approx\,{4\over 3}\,\rho\,+\,4\,(\,N_f-N_c)\,+\,\cdots\,\,,\nonumber\\\rc
&&w\,\approx\,{3(N_c-3N_f)\over 2\rho}\,+\,\cdots\,\,,
\qquad\qquad (\rho\rightarrow\infty)\,\,,
\nonumber\\\rc
&&\gamma\,\approx\,{1\over 3}\,-\,{N_f\over N_c}\,+\,\cdots\,\,\,.
\end{eqnarray}
\setcounter{equation}{0}
\section{Flavors in the Klebanov-Strassler Model}
\label{KS}
The so-called Klebanov-Strassler (KS) solution \cite{Klebanov:2000hb} is dual
to a cascading, confining theory, and has been a popular and successful laboratory in which
to study numerous issues related to gauge-gravity duality and to cosmology.
The gauge theory lives on a stack of regular and fractional D3-branes at the tip of the
deformed conifold, as we now briefly review.
The deformed conifold is a regular, six dimensional, non-compact manifold defined by the equation $z_1\,z_2-z_3\,z_4 =\hat\mu^2$ in ${\mathbb C}^4$. When the complex deformation parameter $\hat\mu$ is turned off, it reduces to the singular conifold, which is invariant under complex rescaling of the $z_i$.
The base of the conifold
has $SU(2)\times SU(2)\times U(1)$ isometry and $S^2\times S^3$ topology. The deformation parameter breaks the scale invariance, produces a blown-up $S^3$ at the apex of the conifold and breaks the $U(1)$ isometry to ${\mathbb Z}_2$.
The low energy dynamics of $N$ regular and $M$ fractional D3-branes on the deformed conifold is described by a cascading ${\cal N}=1$ 4d gauge theory with gauge group $SU(N+M)\times SU(N)$ and bifundamental matter fields $A, B$ transforming as $SU(2)\times SU(2)$ doublets and interacting with a quartic superpotential $W_{KW}=\epsilon^{ij}\epsilon^{kl}A_iB_kA_jB_l$.
The dual to this theory is the
KS solution \cite{Klebanov:2000hb}, that
is relevant for the $N=n\,M$ case, where $n$ is an integer. The related theory develops a Seiberg duality cascade which stops after $n-1$ steps when the gauge group is reduced to $SU(2M)\times SU(M)$. The regular KS solution precisely accounts for the physics of an $A\leftrightarrow B$-symmetric point in the baryonic branch of the latter theory, which exhibits confinement and $U(1)_R\rightarrow {\mathbb Z}_{2N}\rightarrow {\mathbb Z}_2$
where the second
breaking is
due to the formation of a gluino condensate $\langle\lambda\lambda\rangle\sim\Lambda^3_{IR}$. The complex parameter $\epsilon$ is the geometric counterpart of this condensate.
In this section, we will discuss how the solution is modified when a smeared distribution
of D7-branes is introduced. In the dual theory, they correspond to fundamental fields, but
the precise way in which they couple to the rest of fields depends on the D7-brane embeddings,
as we will discuss below.
In what follows, we only discuss cases in which the flavor D7-branes do not break any supersymmetry,
such that the four-dimensional ${\cal N}=1$ of the KS solution is preserved.
The material we summarize in this section was developed in \cite{Benini:2007gx},
\cite{Benini:2007kg}, \cite{Bigazzi:2008qq}.
\subsection{Backreaction with non-chiral flavors}
\label{nonchiralKS}
\subsubsection{Brane embeddings}
Let us start by choosing an appropriate family of supersymmetric D7-brane embeddings.
A particularly interesting
example is given by D7-branes wrapping the holomorphic 4-cycle defined by an equation of the form \cite{Kuperstein:2004hy}:
\begin{equation}
z_1-z_2=\mu\,,
\label{kuperemb}
\end{equation}
where $\mu$ is a constant.
It was shown in \cite{Kuperstein:2004hy} that this embedding is $\kappa$-symmetric and hence preserves the four supercharges of the deformed conifold theory.
A D7-brane wrapping the 4-cycle defined above is conjectured to add a massless (if $\mu=0$) or massive (anti) fundamental flavor to a node of the KS model. The resulting gauge theory is said to be
``non-chiral'' because the flavor mass terms do not break the classical flavor symmetry of the massless theory. The related perturbative superpotential is, just as in the singular conifold case \cite{Ouyang:2003df}, which
we wrote in (\ref{KWsuperpquark}).
The complex mass parameter $m$ in $W$ is mapped to the geometrical parameter $\mu$. The different fields are summarized in the quiver diagram of figure \ref{quiverN}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.5\textwidth]{quiverN.eps}
\end{center}
\caption{The quiver diagram of the gauge theory. Circles are gauge groups, squares are flavor groups, and arrows are bifundamental chiral superfields. $N_{f1}$ and $N_{f2}$ sum up to $N_f$. }
\label{quiverN}
\end{figure}
By acting on this fiducial embedding (\ref{kuperemb}) with the generators of the broken symmetries,
we can build the family of embeddings over which we want to smear.
This is the obvious generalization to the deformed conifold case
of the discussion in section \ref{AdS5X5} and, in fact, a generic non-chiral embedding
is still given by (\ref{genembeddingks}).
\subsubsection{The ansatz}
We now write the ansatz for the metric and forms. It is similar to the ansatz for the
KS solution, but, due to the presence of D7-branes, the RR one-form $F_{(1)}$ is non-trivial
and the dilaton runs.
It is useful to introduce the $g_i$ one-forms
used in \cite{Klebanov:2000hb}:
\begin{eqnarray}
g^1 &=& \frac{-\sin\theta_1\,d\varphi_1 -\cos\psi\sin\theta_2\,d\varphi_2 +\sin\psi\,d\theta_2}{\sqrt{2}}\,,\quad
g^2 = \frac{d\theta_1-\sin\psi\sin\theta_2\,d\varphi_2 -\cos\psi\,d\theta_2}{\sqrt{2}}\,,\nonumber\\
g^3 &=& \frac{-\sin\theta_1\,d\varphi_1 +\cos\psi\sin\theta_2\,d\varphi_2 -\sin\psi\,d\theta_2}{\sqrt{2}}\,,\quad
g^4 = \frac{d\theta_1+\sin\psi\sin\theta_2\,d\varphi_2 +\cos\psi\,d\theta_2}{\sqrt{2}}\,,\nonumber\\
g^5&=& d\psi +\cos\theta_1\,d\varphi_1 + \cos\theta_2\,d\varphi_2\,.
\label{gis}
\end{eqnarray}
The Einstein frame metric ansatz is\footnote{A more generic form of the ansatz was
used in \cite{Benini:2007gx},
\cite{Bigazzi:2008qq}. By requiring supersymmetry and performing some algebra, one
ends up with (\ref{metric}). We will skip those intermediate steps here for the sake of briefness.}:
\begin{eqnarray}
ds^2 &=& h^{-1/2}(\tau)\,dx^2_{1,3} + h^{1/2}(\tau)
{1\over 2}\,\,\hat\mu^{{4\over 3}}\,\,e^{-{\phi(\tau)\over 3}}\,{\cal K}(\tau)\,\,
\Bigg[\,{1\over
3{\cal K}^3(\tau)}\,\,\big(\,d\tau^2\,+\,(g^5)^2\,\big)\,+\nonumber\\\,
&+& \cosh^2\Big({\tau\over 2}\Big)\,\Big(\,(g^3)^2\,+\,
(g^4)^2\,\Big)\,
+\,\sinh^2\Big({\tau\over 2}\Big)\,
\Big(\,(g^1)^2\,+\, (g^2)^2\,\Big)\,
\,\Bigg]\,\,,
\label{metric}
\end{eqnarray}
where $\hat \mu$ is the complex deformation parameter of the conifold, $dx^2_{1,3}$ denotes the four-dimensional Minkowski metric and
${\cal K}(\tau)$, $h(\tau)$ and the dilaton $\phi$ are unknown
functions of the radial variable to be determined.\footnote{
The relation of the $z_i$ complex variables as used above to the $\tau, \theta_1,\varphi_1,
\theta_2,\varphi_2,\psi$ coordinates can be found, for instance in \cite{Bigazzi:2008qq}.
The embedding equation (\ref{genembeddingks})
expressed in terms of the ``deformed conifold $\tau$ variable'' looks the same in terms of the ``backreacted ansatz $\tau$ variable''. See \cite{Bigazzi:2008qq} for details. }
For the forms we will adopt the following ansatz:
\begin{eqnarray}
F_5&=&d h^{-1}(\tau)\wedge dx^0\wedge\cdots\wedge dx^3\,-\,\dot h \frac{\hat\mu^{\frac83}}{16}
e^{-\frac{2\phi}{3}}\sinh^2 \tau\, {\cal K}^4 g_1\wedge g_2\wedge g_3\wedge g_4\wedge g_5\,,\nonumber\\
B_2 &=& g_s\alpha'\frac{M}{2} \Bigl[ f\, g^1 \wedge g^2\,+\,k\, g^3 \wedge
g^4 \Bigr]\,,\nonumber\\
H_3&=& g_s\alpha'\frac{M}{2} \, \Bigl[ d\tau \wedge (\dot f \,g^1 \wedge g^2\,+\,
\dot k\,g^3 \wedge g^4)\,+\,{1 \over 2}(k-f)\, g^5 \wedge (g^1
\wedge g^3\,+\,g^2 \wedge g^4) \Bigr]\,,\nonumber\\
F_1&=&g_s{N_f\,p(\tau)\over 4\pi}\,\,g^5\,,\nonumber\\
F_3&=& g_s\alpha'\frac{M}{2} \Big\{ g^5\wedge \Big[ \big( F+\frac{g_sN_f\,p(\tau)}{4\pi}f\big)g^1\wedge g^2 + \big(1- F+\frac{g_sN_f\,p(\tau)}{4\pi}k\big)g^3\wedge g^4 \Big] +\nonumber\\
&&+\dot F d\tau \wedge \big(g^1\wedge g^3 + g^2\wedge g^4 \big)\Big\} \,,
\label{theansatz}
\end{eqnarray}
where $f=f(\tau)$, $k=k(\tau)$, $F=F(\tau)$ are functions of the radial coordinate (and
where the dot denotes derivatives with respect to $\tau$). We have implemented the self-duality
condition for $F_5$.
Notice that, consistently, $dF_1=-g_s \Omega$, where $\Omega$ is the
symmetry preserving
D7-brane density distribution form analogous to (\ref{gsomega}):
\begin{equation}
\Omega =\frac{N_f}{4\pi}\Big(p(\tau)(\sin\theta_1 d\theta_1 \wedge d\varphi_1
+ \sin\theta_2 d\theta_2 \wedge d\varphi_2)
-\dot p(\tau) d\tau\wedge g_5\Big)\,.
\label{massive_2form}
\end{equation}
When quarks are massless \cite{Benini:2007gx}, one just has $p(\tau)=1$, whereas $p(\tau)$
becomes non-trivial when quarks are massive.
In what follows, we will keep $p(\tau)$ generic. We refer the reader to \cite{Bigazzi:2008qq} for the computation of $p(\tau)$ from the massive non-chiral brane embeddings
(\ref{genembeddingks}).
The source contributions to the modified Bianchi identities for $F_3$ and $F_5$
\begin{eqnarray}
&&dF_3\,=\,H_3\wedge F_1\,-\,g_s\Omega\wedge B_2\,\,,\nonumber\\
&&dF_5\,=\,H_3\wedge F_3\,-\,{1\over 2}\,g_s\Omega\wedge
B_2\wedge B_2\,,
\label{newBianchi}
\end{eqnarray}
follow from the WZ term of the smeared D7-brane action \cite{Bigazzi:2008qq}.
Given (\ref{theansatz}) and (\ref{massive_2form}), the equations (\ref{newBianchi}) are
satisfied provided:
\begin{equation}
\dot h \frac{\hat \mu^{\frac83}}{16}
e^{-\frac{2\phi}{3}}\sinh^2 \tau\, {\cal K}^4 = \text{const}
- \frac14 (g_s \alpha' M)^2 \left[ f - (f-k)F + \frac{g_s N_f }{4\pi}p(\tau) f\,k\right]\,\,.
\label{hdot}
\end{equation}
\subsubsection{The BPS equations}\label{kssolution}
By requiring the vanishing of the bulk fermionic supersymmetry variations, one
finds a set of first order BPS equations. The computation is lengthy but straightforward
and was carried out in \cite{Benini:2007gx}
(since in that paper $p(\tau)=1$, the substitution $N_f \rightarrow N_f\, p(\tau)$
has to be implemented in the equations of \cite{Benini:2007gx}).
In the present notation, the differential equations are:
\begin{eqnarray}
\dot\phi &=& \frac{g_s N_f p(\tau)}{4\pi}e^{\phi}\,\,,\nonumber\\
\dot k &=& e^{\phi}\left(F+\frac{g_s N_f p(\tau)}{4\pi}\,f\right)\coth^2\frac{\tau}{2}\,,\nonumber \\
\dot f &=& e^{\phi}\left(1-F+\frac{g_s N_f p(\tau)}{4\pi}\,k\right)\tanh^2\frac{\tau}{2}\,,\nonumber \\
\dot F &=& \frac{1}{2}e^{-\phi}(k-f)\,,\nonumber\\
\frac{\dot {\cal K}}{\cal K} &=& \frac{2}{3{\cal K}^3 \sinh \tau } + \frac{\dot \phi}{3} -\coth \tau\,\,,
\label{BPSKS}
\end{eqnarray}
supplemented by the algebraic constraint
\begin{equation}
e^{-\phi}(k-f)=\tanh\frac{\tau}{2}- 2F\,\coth\tau + \frac{g_s N_f p(\tau)}{4\pi}\left[k\,\tanh\frac{\tau}{2} -f\,\coth\frac{\tau}{2}\right]\,.
\label{fluxconstr}
\end{equation}
Quite remarkably, the equations (\ref{BPSKS})-(\ref{fluxconstr})
can be (almost) explicitly integrated. In the following,
we will use notations similar to those employed in section \ref{AdS5X5}.
We introduce an arbitrary value of the radial coordinate $\tau_*$ at which the
dilaton is $\phi_*$. Then, we can write the dilaton as:
\begin{equation}
e^{\phi-\phi_*}=\frac{1}{1+\epsilon_* \int_{\tau}^{\tau_*} p(\xi) d\xi}\,\,,
\label{dilKS}
\end{equation}
where we have introduced the deformation parameter which
weighs the flavor loops as:
\begin{equation}
\epsilon_* = \frac{N_f}{16\pi^2 M}\lambda_* \qquad {\rm with}\qquad
\lambda_* \equiv 4\pi g_s M\, e^{\phi_*}\,\,.
\label{KSdefs}
\end{equation}
Let us also introduce a function:
\begin{equation}
\eta(\tau) = \epsilon_* e^{\phi-\phi_*} \int_0^\tau (\sinh 2 \xi -2\xi) p(\xi) d\xi\,\,.
\end{equation}
Then, we can integrate for the rest of the functions of the ansatz:
\begin{eqnarray}
{\cal K}\,&=&\,{\big[\sinh 2\tau-2\tau\,+\,\eta(\tau)\big]^{{1\over 3}}\over
2^{{1\over 3}}\,\sinh\tau}\,\,,\qquad\quad F= \frac{\sinh \tau - \tau}{2 \sinh \tau}\,, \nonumber\\
f&=& e^{\phi}\,\frac{\tau \coth \tau -1}{2 \sinh \tau} (\cosh \tau -1)\,\,,\qquad
k= e^{\phi}\,\frac{\tau \coth \tau -1}{2 \sinh \tau} (\cosh \tau +1)\,\,.
\label{fkFexplicit}
\end{eqnarray}
Finally, the function $h$ can be obtained by integrating (\ref{hdot}).
The KS solution without flavors \cite{Klebanov:2000hb} is obtained by taking $\epsilon_*=0$, such that
the dilaton is constant and $\eta(\tau)=0$. For $p(\tau)=1$, we find the solution
backreacted with massless flavors \cite{Benini:2007gx}.
In this case, the integrals for the dilaton and $\eta(\tau)$ can be explicitly performed:
\begin{eqnarray}
\eta(\tau) &=& \epsilon_* e^{\phi-\phi_*} (\sinh^2 \tau - \tau^2) \,\,,\nonumber\\
e^{\phi-\phi_*} &=& \frac{1}{1+\epsilon_* ( \tau_* - \tau)}\,\,. \qquad\qquad\qquad
(\text{for}\, p(\tau)=1)
\end{eqnarray}
In this massless case, the solution has a curvature singularity in the IR $\tau=0$.
Some cases where $p(\tau)$ is non-trivial
were discussed in \cite{Bigazzi:2008qq}.
\subsubsection{Some physical features}
The solution presented in the preceding sections has been used to extract some
of the physics encoded in the unquenched background. In \cite{Benini:2007gx},
the running of the couplings and anomalies were discussed. As anticipated above, in
\cite{Bigazzi:2008qq}, the solution with massive flavors was found.
Quark masses erase the IR singularity in the same way as explained in section
\ref{heuristiczzz} or in section \ref{massiveKW}.
Quark-antiquark potentials, screening lengths and associated quantum phase transitions
were discussed in the same paper. Finally, in \cite{unquenchedmesons}, it was computed
how the screening effects due to unquenched fundamental matter affect the mass spectra of
the KS model, with results similar to section \ref{sec:screening}.
Due to space constraints, we cannot go explicitly through all of these features and we refer
the interested reader to the original papers. Here, we will just briefly discuss how the
solution captures the phenomenon of a duality wall \cite{Benini:2007gx} and how gauge
groups ranks
change upon Seiberg duality.
We will make use of the following holographic formulae,
which can be derived in the $\mathcal{N}=2$ orbifold case by looking at the lagrangian of the low energy field theory living on probe (fractional) D3-branes:%
\begin{equation}
\label{holographic relations}
\frac{4\pi^2}{g_{YM}(l)^2} + \frac{4\pi^2}{g_{YM}(s)^2} = \frac{\pi\, e^{-\phi}}{g_s} \,\,,\qquad
\frac{4\pi^2}{g_{YM}(l)^2} - \frac{4\pi^2}{g_{YM}(s)^2} =
\frac{2 \pi\,e^{-\phi}}{g_s} \Bigl[\frac{1}{4\pi^2 \alpha'} \int_{S^2} B_2 - \frac12 \;
\Bigr]\,\,.
\end{equation}
The labels $(l),(s)$ in (\ref{holographic relations}) refer to the gauge group with the larger or smaller rank. Strictly speaking, these formulae need to be corrected for small values of the gauge couplings and are only valid in the large 't Hooft coupling regime (see \cite{Benvenuti:2005wi,Strassler:2005qs,Benini:2006hh}), which is the case under consideration. Moreover, they are also expected to be precise just in the UV region, where the cascade takes place
and the region on which we will focus below.
The expressions (\ref{holographic relations})
give positive squared couplings only if the expression inside the
square bracket is in the range $[-\frac12,\frac12]$. Define:
\begin{equation}
b_0(\tau) \equiv \frac{1}{4\pi^2\alpha'}\int_{S^2} B_2 = \frac{g_s M}{\pi}f
= \frac{\lambda_*}{8\pi^2} \frac{\tau -1}{1-\epsilon_*(\tau-
\tau_*)}\,,\qquad
\tilde b_0 \equiv b_0 - [b_0] \in [0,1] \,,
\label{b0val}
\end{equation}
where $[b_0]$ denotes the integer part of $b_0$.
In order to get the explicit expression for $b_0$ we have integrated over the $S^2$ parameterized by
$\theta_1=\theta_2$, $\varphi_1 = 2\pi - \varphi_2$, $\psi = const$ \cite{Benini:2007gx}, considered the UV limit of (\ref{dilKS}), (\ref{fkFexplicit}) such that
$p(\tau)\approx 1$ and $f\approx k \approx e^\phi (\tau -1)/2$ and inserted the
definitions (\ref{KSdefs}).
Now we see that what we
have to insert in (\ref{holographic relations}) is indeed $\tilde b_0$.
This is the physical content of the cascade: at a given energy scale we must perform a large gauge transformation on $B_2$ in supergravity to shift $\int B_2$ by a multiple of $4\pi^2\alpha'$
to get a field theory description with positive squared couplings.
Let us restrict our attention to an energy range, between two subsequent Seiberg dualities,
where a field theory description in terms of specific ranks holds.
When flowing towards the IR, $\tilde b_0$ decreases from 1 to 0. From
(\ref{holographic relations}) and inserting the solution, we can find an expression
for each of the gauge couplings:
\begin{equation}
\frac{1}{\lambda_l}=\frac{1}{\lambda_*}(1-\epsilon_* (\tau-\tau_*))\, \tilde b_0\,\,,\qquad
\frac{1}{\lambda_s}=\frac{1}{\lambda_*}(1-\epsilon_* (\tau-\tau_*))\, (1-\tilde b_0)\,\,.
\end{equation}
In this energy range, the coupling $\lambda_l$
starts different from zero and flows to $\infty$ at the end of this range,
where a Seiberg duality on its gauge group is needed. The coupling $\lambda_s$ of
the gauge group with smaller rank is the one which starts very large (actually
divergent) after the previous Seiberg duality on its gauge group, and then flows toward weak coupling.
The qualitative picture of the RG flow in the UV can be extracted from our supergravity solution even without discussing
the precise radius-energy relation, simply recalling that the radius must be a monotonic function of the energy scale.
First, notice that at a finite $\tau$
(and therefore at a finite energy scale $E_{UV}$), the dilaton diverges making both gauge couplings diverge.
This happens at:
\begin{equation}
\tau_{dw}= \tau_* + \frac{1}{\epsilon_*}\,\,.
\end{equation}
From (\ref{b0val}), we see that the derivative $\frac{db_0}{d\tau}$ grows unbounded near
$\tau_{dw}$, meaning that the interval (in $\tau$) between Seiberg dualities becomes shorter and
shorter.
The Seiberg dualities pile up the more we approach the UV cut-off $E_{UV}$.
The picture which stems from the flavored Klebanov-Tseytlin/Strassler solution is that
$\tau_{dw}$ is a so-called ``Duality Wall", namely an accumulation point of energy scales at which a Seiberg duality is required in order to have a weakly coupled description of the gauge theory \cite{Strassler:1996ua}. Above the duality wall, Seiberg duality does not proceed and a weakly coupled dual description of the field theory is not known.
See Figure \ref{wall}. Nevertheless, in full analogy with the discussion of section \ref{AdS5X5}, the
derivative of the holographic $a$-function changes sign at a finite distance in $\tau$ below $\tau_{dw}$
so one should not trust the solution all the way up to the singular point $\tau_{dw}$.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.45\textwidth]{wall.eps}
\end{center}
\caption[wall]{Qualitative plot of the running gauge couplings as functions of the logarithm of the energy scale in the cascading gauge theory. The blue lines are the inverse squared gauge couplings, while the red line is their sum. \label{wall}}
\end{figure}
Duality walls were studied in the context of quiver gauge theories first by Fiol \cite{Fiol:2002ah} and later in a series of papers by Hanany and collaborators \cite{Hanany}.
To our knowledge, the solution above
is the only explicit realization of this exotic ultraviolet phenomenon on the supergravity side of the gauge/gravity correspondence.
To end this section, we discuss how the effective number of regular and fractional
D3-branes change when undergoing a step of the cascade of Seiberg dualities.
We will not compute the explicit shift in $\tau$ but rather the shift in the function $f$ ($\approx k$).
From (\ref{b0val}), we have:
\begin{equation}
\begin{split}
b_0 (\tau) \rightarrow b_0(\tau') = b_0 (\tau) - 1
\end{split}
\quad
\Longrightarrow \quad
\begin{split}
f(\tau) &\rightarrow f(\tau') = f(\tau) - \frac{\pi}{g_s M} \,\,.
\end{split} \label{shift f k}
\end{equation}
On the other hand, we compute the effective number of branes at a given
energy scale by integrating the
appropriate RR-forms:
\begin{eqnarray}
N_{eff} (\tau) &\equiv& \frac{1}{(2\pi)^4 g_s \alpha'^2} \int_{{\cal M}_5} F_5 \,=N_0 +
\,{g_s M^2\over \pi}\,\Big[\,
f +{g_s N_f\over 4\pi}f^2 \Big]\,\,,\nonumber\\
M_{eff}(\tau) &\equiv& \frac{1}{4\pi^2g_s \alpha'} \int_{S^3} F_3 = M \Bigl[ 1 + \frac{g_s N_f}{2\pi} f
\Bigr]\;.
\label{neffmeff}
\end{eqnarray}
In these expressions we have substituted (\ref{theansatz}), (\ref{hdot}) and
already taken the UV limit $f = k$ and $p(\tau)=1$.
The $S^3$ for the second integral is the one parameterized by
$\theta_2=\text{constant}$, $\varphi_2=\text{constant}$. Notice that $N_{eff}$ and $M_{eff}$ are not quantized. This is because they are {\it Maxwell} charges, as opposed to
{\it Page} charges. See \cite{Benini:2007gx} for thorough explanations.
We can compute how $N_{eff}$, $M_{eff}$ vary in a Seiberg duality step
(\ref{shift f k}).
A bit of algebra shows that:
\begin{eqnarray}
M_{eff} (\tau) &\rightarrow& M_{eff} (\tau') = M_{eff} (\tau) - \frac{N_f}{2} \label{scaling sugra Meff}\,\,,
\nonumber\\
N_{eff} (\tau) &\rightarrow& N_{eff} (\tau') = N_{eff} (\tau) - M_{eff} (\tau) + \frac{N_f}{4} \,\,,
\nonumber
\label{scaling sugra 2}
\end{eqnarray}
whereas $N_f$ remains unchanged. A careful analysis in
\cite{Benini:2007gx} showed that this is in full agreement with field theory expectations.
\subsection{Backreaction with chiral flavors}
In a remarkable paper \cite{Benini:2007kg}, Benini discussed the solution dual to
having smeared chiral flavors on the conifold. In the probe approximation, the D7-brane embeddings
that correspond to chiral flavors were discussed in \cite{Ouyang:2003df}, \cite{Levi:2005hh}.
How these flavors transform under the gauge groups is shown in the quiver diagram
\ref{quiverBenini}. In this case,
the quiver theory is not self-similar under the duality cascade; in each step
of the cascade a meson field is generated. Its couplings to the
rest of the fields are, however, irrelevant \cite{Benini:2007kg}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.3\textwidth]{quiver_el1bis.eps}
\caption{Quiver diagram of the KS theory with chiral flavors. }
\end{center}
\label{quiverBenini}
\end{figure}
The backreacted solution of \cite{Benini:2007kg} uses the singular conifold and therefore it
can be considered as the deformation of the Klebanov-Tseytlin solution \cite{Klebanov:2000nc}
due to smeared chiral flavors.
From the gravity point of view, the extra complication with respect to section \ref{nonchiralKS}
is that the worldvolume gauge field on the D7s has to be turned on.
In fact, this is crucial when matching the shifts in the ranks of the gauge groups upon Seiberg
dualities to the supergravity background (there are subtle differences with respect to the
non-chiral case). We will not report further on this solution here, but
refer the reader to \cite{Benini:2007kg}.
\setcounter{equation}{0}
\section{Models with cohomogeneity 2}\label{models2}
In this section we present some situations in which, even smearing the flavor
branes, the system cannot be reduced to a one-dimensional problem. In fact,
the different fields will depend on two different radial coordinates and,
accordingly, one has to solve partial differential equations rather than
ordinary differential equations.
In order to provide a heuristic picture, the situation is depicted in
figure \ref{smearingfig2}. Concretely, we will refer here to the model of section
\ref{secN2}, but the situation is
very similar
for all the cases discussed in this section.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.45\textwidth]{smearing_cod2.eps}
\end{center}
\caption{A qualitative plot of the situation with
cohomogeneity 2 models. The red dot in the center represents the color branes and each
vertical line is a flavor brane. Taking a lot of them smeared along the
$\phi_2$ angle, the rotational symmetry associated to this angle is
effectively recovered. Any function of the ansatz depends on the radial coordinates
$\rho$ and $\sigma$.}
\label{smearingfig2}
\end{figure}
In figure \ref{smearingfig2}, the color branes are placed at the tip of a Calabi-Yau
($\sigma$ is a radial coordinate along the CY, the rest of the directions of the CY
are omitted from the plot).
The $\rho-\phi_2$ plane is transverse both to the color branes and to the CY. Each flavor
brane lies at a point in this plane and is extended along $\sigma$.
Distributing the flavor branes along $\phi_2$, it is possible to recover
(in the smeared limit)
the associated $U(1)$ isometry. On the contrary, as is apparent from
figure \ref{smearingfig2}, there is no way in which one can place the flavor brane to
recover the full radial symmetry. Hence, the solution associated to this brane configuration
must be cohomogeneity two, meaning that all functions of the eventual ansatz
will depend on $\rho$ and $\sigma$.
As a matter of fact, if one wishes to construct a deformation
of $AdS_5\times S^5$ with smeared flavor such that the supersymmetry
preserved is ${\cal N}=2$ (rather than ${\cal N}=1$ as
in section \ref{AdS5X5}), the solution would have cohomogeneity two
and, presumably, would share some similarities with the examples presented in
this section.
This is an interesting open problem for the future.
\subsection{A dual to (3+1)-dimensional ${\cal N}=2$ SQCD-like theory}
\label{secN2}
In this section, we study the dual solution to the brane intersection summarized in
table \ref{table: N2}. The gauge theory lives on $N_c$
D5-branes wrapping a two-sphere with the appropriate twisting
to preserve eight supercharges, {\it i.e.} ${\cal N}=2$ in the
effective four-dimensional low energy theory. Geometrically, it corresponds
to wrapping the branes along a compact SLag two-cycle inside a
non-compact Calabi-Yau two-fold.
This leaves two flat transverse dimensions which are identified with
the moduli space corresponding to giving vevs to the complex scalar
inside the ${\cal N}=2$ vector multiplet.
The $N_f$ flavor D5-branes do not further break supersymmetry and
provide
fundamental hypermultiplets in order to
build
${\cal N}=2$ SQCD.
They are extended in the non-compact $\sigma$ direction and, thus, their volume
is infinite, making exactly zero the
effective four-dimensional gauge coupling living on them.
They would provide a global symmetry group $U(N_f)$ if they were placed on top of each other,
but due to the smearing, only $U(1)^{N_f}$ is left.
The dual solution without flavors was found in \cite{Gauntlett:2001ps}, and the flavored
case was discussed in \cite{Paredes:2006wb}.
\begin{table}[ht!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\multicolumn{1}{c}{ }
&
\multicolumn{1}{c}{ }
&
\multicolumn{4}{c}{$\overbrace{\phantom{\qquad\ \qquad
\qquad
\ \ \,}}^{CY_2}$}
&
\multicolumn{2}{c}{$\overbrace{\phantom{\ \quad
\ \ \,}}^{\mathbb{R}_2}$}
\\
\hline
\multicolumn{1}{|c|}{ }
&
\multicolumn{1}{|c|}{$x_{1,3}$}
&\multicolumn{1}{|c|}{$\sigma$}
&\multicolumn{1}{|c|}{$\phi_1$}
&\multicolumn{1}{|c|}{$\tilde{\theta}$}
&\multicolumn{1}{|c|}{$\tilde{\varphi}$}
&\multicolumn{1}{|c|}{$\rho$}
&\multicolumn{1}{|c|}{$\phi_2$}
\\
\hline
$N_c$ D5 &$-$&$\cdot$&$\cdot$&$\bigcirc$&$\bigcirc$&$\cdot$&$ \sim $\\
\hline
$N_f$ D5 &$-$&$-$&$\bigcirc$&$ \sim $&$ \sim $&$\cdot$&$ \sim $\\
\hline
\end{tabular}
\caption{A scheme of the setup:
for the brane configuration,
a line $-$ means that the brane spans a non-compact dimension, a point
$\cdot$ that it is point-like in that direction, a circle
$\bigcirc$ that it wraps a compact cycle and $\sim$ indicates smearing in the
direction.
Above, it is shown which directions spanned the Calabi-Yau
and which the transverse plane before backreaction.
\label{table: N2}}
\end{center}
\end{table}
We start by writing an ansatz for the metric consistent with the symmetries of the problem.
In Einstein frame:
\begin{eqnarray}
ds_{10}^2 &=& g_s N_c \alpha' e^\frac{\Phi}{2} \left[
\frac{1}{g_s N_c \alpha'} dx_{1,3}^2 + z (d\tilde{\theta}^2 +
\sin^2 \tilde{\theta} d\tilde{\varphi}^2) + \right.\nonumber\\
&&
\left.
+ e^{-2\Phi} (d\rho^2 + \rho^2 d\phi_2^2)
+ \frac{e^{-2\Phi}}{z} \left( d\sigma^2 + \sigma^2
(d\phi_1 + \cos \tilde{\theta} d\tilde{\varphi})^2
\right) \right]\,\,,
\label{metricN2}
\end{eqnarray}
where $z$ and $\Phi$ depend on both radial coordinates $\rho$, $\sigma$.
The Calabi-Yau twofold directions are $0 \leq \tilde{\theta} \leq \pi$, $0\leq \tilde{\varphi} < 2\pi$, $0\leq\sigma<\infty$,
$0\leq \phi_1 < 2\pi$
(of course, in this solution with fluxes there is not a
Calabi-Yau any more, but it can be
thought of as a deformation of the Calabi-Yau that was present
before backreaction).
The coordinates
$0\leq \rho < \infty$, $0 \leq \phi_2 < 2 \pi$
span the transverse two-dimensional plane, so they should be identified with the moduli space,
and therefore rotations in $\phi_2$ are related to the $U(1)_R$ symmetry of the field theory.
Out of the $SU(2)_R$ symmetry, only its diagonal $U(1)_J$ is manifest in the geometry, as
rotations in $\phi_1$. The extra $SO(3)$ isometry which acts on $\tilde{\theta},\tilde{\varphi},\phi_1$
does not play
a role in the low energy ${\cal N}=2$ SQCD theory \cite{Gauntlett:2001ps}.
As anticipated in table \ref{table: N2}, we want to consider a set of $N_f$ D5-branes extended
in $x_{1,3}$, $\sigma$ and wrapped in $\phi_1$. They lie at fixed $\rho=\rho_Q$, where $\rho_Q$ is
proportional to the modulus of the mass of the fundamental hypermultiplets. These D5-branes are
homogeneously smeared over the $S^2$ parameterized by $\tilde \theta$, $\tilde \varphi$ and on the
angle $\phi_2$, which corresponds to the phase of the mass of the hypers.
This distribution is described by the four-form:
\begin{equation}
\Omega=\frac{N_f}{8\pi^2}
\delta(\rho-\rho_Q)\ \sin \tilde{\theta} d\rho \wedge d\phi_2
\wedge d\tilde{\theta} \wedge d\tilde{\varphi} \,\,,
\label{y4}
\end{equation}
such that the source-modified Bianchi identity for $F_{(3)}$ reads:
\begin{equation}
dF_{(3)}= 2\kappa_{(10)}^2 T_5 \Omega=
g_s \alpha' \ \frac{N_f}{2}
\delta(\rho-\rho_Q)\ \sin \tilde{\theta} d\rho \wedge d\phi_2
\wedge d\tilde{\theta} \wedge d\tilde{\varphi} \,\,.
\label{newdF}
\end{equation}
We can write an ansatz for $F_{(3)}$ consistent with this expression:
\begin{eqnarray}
F_{(3)}&=& N_c g_s \alpha' \left[ -g' d\phi_2 \wedge d\rho \wedge
(d\phi_1 + \cos \tilde{\theta} d\tilde{\varphi}) - \dot g d\phi_2 \wedge d\sigma \wedge
(d\phi_1 + \cos \tilde{\theta} d\tilde{\varphi})+\right. \nonumber\\
&& \left. + (g + \frac{N_f}{2N_c}\Theta(\rho-\rho_Q))
\sin \tilde{\theta} d\phi_2\wedge d\tilde{\theta} \wedge
d\tilde{\varphi} \right]\,.
\label{F3N2}
\end{eqnarray}
where $\Theta$ is the Heaviside step function\footnote{Notice that,
as opposed to section \ref{CNPwilson} where a Heaviside function was introduced as an approximation to
the effect of the massive flavors,
the $\Theta$ here is exactly what comes from the family of D-brane embeddings considered, since
they all lie at fixed $\rho=\rho_Q$.},
$g$ a new function of $\rho$ and $\sigma$ that
needs to be determined, and we have introduced the following notation for the partial
derivatives:
\begin{equation}
' \equiv \partial_\rho \,,\qquad \dot{} \equiv \partial_\sigma
\label{prime&dot}
\end{equation}
The next step is to insert the ansatz (\ref{metricN2}), (\ref{F3N2}) into the type
IIB supersymmetry transformations $\delta \psi_\mu = \delta \lambda =0$,
as outlined in section \ref{sec: kappa}.
This procedure was carefully performed in
\cite{Paredes:2006wb},
whereas here we just quote the resulting system of first order equations:
\begin{eqnarray}
&&g + \frac{N_f}{2N_c}\Theta(\rho-\rho_Q) = - \rho z' \,,\qquad\quad\qquad\qquad
g' = -2 e^{-2\Phi} \rho\sigma \dot \Phi
\,,\qquad\qquad\nonumber\\
&&e^{2\Phi}= \frac{\sigma}{z \dot z}
\,, \qquad\quad\quad\qquad
\dot g = - z^{-2} e^{-2\Phi} \sigma (g+\frac{N_f}{2N_c}\Theta(\rho-\rho_Q)) +
2 z^{-1} \rho\sigma e^{-2\Phi} \Phi'\,.\qquad
\label{neweq4}
\end{eqnarray}
It is easy to check that the last equation is not independent of the previous ones and
that equations (\ref{neweq4}) ensure the equation of
motion for the 3-form
$d(e^\Phi\ {}^*F_{(3)}) = 0$. This system of equations can be recast as a single, non-linear,
second order PDE for $z(\rho,\sigma)$:
\begin{equation}
\sigma \frac{N_f}{2N_c}\delta(\rho-\rho_Q) +
\rho z (\dot z - \sigma \ddot z)=\sigma (\rho \dot z^2 + z' + \rho z'')\,\,.
\label{neweqinz}
\end{equation}
Once $z(\rho,\sigma)$ is computed, $g$ and $\Phi$ are read from (\ref{neweq4}).
In general, the equation (\ref{neweqinz}) cannot be solved explicitly.
In the unflavored case $N_f=0$, there is in fact an exact solution
\cite{Gauntlett:2001ps} (see \cite{Paredes:2006wb} and the first paper
of \cite{DiVecchia:2002ks} for the adaptation of the solution \cite{Gauntlett:2001ps} to
the present coordinate system).
Equation (\ref{neweqinz}), however, can be studied numerically \cite{Paredes:2006wb}.
We will not pursue that here, but we will verify using (\ref{neweq4}) that the
expected beta-function for the gauge coupling stems from the differential equations.
In order to read the effective four-dimensional gauge coupling from the geometry, we consider
a ``color" D5-brane probing the Coulomb branch of the theory, namely, a D5 wrapping
the $S^2$ parameterized by $\tilde \theta$, $\tilde \varphi$, sitting at $\sigma=0$
\cite{DiVecchia:2002ks}. After integrating the volume of the $S^2$, we find:
\begin{equation}
\frac{1}{g_{YM}^2}=\frac{N_c}{4\pi^2}(z|_{\sigma=0})\,\,.
\end{equation}
Thus, in order to understand the running of the coupling it is not necessary to know the
geometry everywhere, but just at $\sigma=0$. From the second equation of (\ref{neweq4}),
we see that $g$ is a constant at $\sigma=0$, which then results in the fact that the first
equation of (\ref{neweq4}) can be trivially integrated.
But before doing that, let us find out which is the value of $g|_{\sigma=0}$.
With that purpose, let us consider the normalization condition:
\begin{equation}
\frac{1}{2\kappa_{(10)}^2}\int F_{(3)} = N_c T_5\,\,,
\label{quantt}
\end{equation}
where we have to integrate along $\phi_1$, $\phi_2$ and an angle built in the
``plane" of the two radial directions $\rho,\sigma$
(heuristically, think of introducing some polar coordinates
$r,\theta$ such that $\rho=r \sin\theta$ and $\sigma=r \cos\theta$. Then
we want to integrate in $\theta$ from 0 to $\frac{\pi}{2}$ at fixed and large
$r$). Inserting (\ref{F3N2}),
we find
\begin{equation}
g|_{(\sigma=\infty ,\rho=0 )} - g|_{(\sigma=0,\rho=\infty )}=1\,\,.
\end{equation}
But from the first equation in (\ref{neweq4}) we read that
$g|_{\rho=0 }=0$ and, thus, $g|_{\sigma=0}=-1$.
We are now ready to integrate the first equation in (\ref{neweq4})
at $\sigma=0$:
\begin{equation}
\frac{N_c}{4\pi^2}(z|_{\sigma=0})=\frac{1}{g_{YM}^2}=\frac{1}{4\pi^2}\Big[
\Big(N_c - \frac{N_f}{2} \Theta(\rho-\rho_Q)\Big) \log \rho
+ \frac{N_f}{2} \Theta(\rho-\rho_Q) \log \rho_Q+const\Big]\,\,,
\label{flavgym}
\end{equation}
where the next to last term comes from requiring continuity of the metric
at $\rho=\rho_Q$.
Making use of the radius-energy relation
$\rho=\frac{\mu}{\Lambda}$ found in \cite{DiVecchia:2002ks},
we get:
\begin{equation}
\beta (g_{YM})(\mu) = -\frac{g_{YM}^3}{8\pi^2} \Big(N_c - \frac{N_f(\mu)}{2}\Big)\,\,,
\label{bgym}
\end{equation}
where $N_f (\mu)$ is defined as the number of flavors for which the
modulus of their masses is
smaller than the scale. Matter fields with bigger mass are holomorphically decoupled
at lower scales, as expected.
The expression (\ref{bgym}) fits field theory expectations and is a non-trivial check
of the described unquenched set-up.
For further discussion of this model, see \cite{Paredes:2006wb}.
\subsection{Flavors in lower-dimensional SQCD Models}
The approach described in subsection \ref{secN2} can be also applied to construct supergravity duals of SQCD-like models in two and three dimensions by considering lower dimensional branes wrapping different cycles of Calabi-Yau manifolds. In this subsection we will review two of such constructions. First of all, following refs. \cite{Maldacena:2000mw, Arean:2008az}, we will consider the case of D3-branes wrapping a two-cycle of a Calabi-Yau twofold, which is dual to a two-dimensional gauge theory with ${\cal N}=(4,4)$ supersymmetry. Secondly, we will review the similar construction of refs. \cite{Maldacena:2000mw, Divecchia, Ramallo:2008ew} of the gravity dual of three-dimensional ${\cal N}=4$ gauge theories from D4-branes wrapping two-cycles in a $CY_2$. Backgrounds dual to 2d and 3d flavored theories with reduced supersymmetry have been also constructed \cite{Gaillard:2008wt,Arean:2009gc,D3D4D5}, and they will be also very briefly reviewed.
\subsubsection{Two-dimensional theories}
Let us consider the following setup for two sets of D3-branes in a Calabi-Yau cone of complex dimension two:
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\multicolumn{3}{c}{ }&
\multicolumn{4}{c}{$\overbrace{\phantom{\qquad\qquad\qquad}}^{\text{CY}_2}$}\\
\hline
&\multicolumn{2}{|c|}{$\mathbb{R}^{1,1}$}
&\multicolumn{2}{|c|}{$S^2$}
&\multicolumn{2}{|c|}{$N_2$}
&\multicolumn{4}{|c|}{$\mathbb{R}^{4}$}\\
\hline
$N_c$\,\,\,D$3$ (color) &$-$&$-$&$\bigcirc$&$\bigcirc$&$\cdot$&$\cdot$&$\cdot$&$\cdot$&$\cdot$&$\cdot$\\
\hline
$N_f$\,\,\,D$3$ (flavor) &$-$&$-$&$\cdot$&$\cdot$&$-$&$-$&$\cdot$&$\cdot$&$\cdot$&$\cdot$\\
\hline
\end{tabular}
\label{44flavored-array}
\end{center}
where $S^2$ represents the directions of a compact two-cycle and $N_2$ are the directions of the corresponding normal bundle. Notice also that the symbols ``$-$" and ``$\cdot"$ represent respectively unwrapped worldvolume directions and transverse directions, while a circle denotes wrapped directions. Let us parameterize the cycle by means of two angular coordinates $(\theta, \phi)$ and let $\sigma$ be the radial coordinate of the CY cone. The ansatz for the string frame metric which we will adopt is the following:
\begin{eqnarray}
&&ds_{st}^2\,=\,H^{-{1\over 2}}\,\,\Big[\,dx_{1,1}^2\,+\,{z\over m^2}\,
\Big(\,d\theta^2\,+\,\sin^2\theta\,d\phi^2\,\,\Big)\,\Big]\,+\,\nonumber\\\rc
&&\qquad\qquad+\,
H^{{1\over 2}}\,\,\,\Big[\,{1\over z}\,\,\Big(\,
d\sigma^2\,+\,\sigma^2\,
\Big(\,d\psi+\cos\theta d\phi\,\Big)^2\,\Big)
\,+\,
d\rho^2\,+\,\rho^2\,d\Omega_3^2\,\Big]\,\,,
\qquad\qquad
\label{D3metric}
\end{eqnarray}
where $m$ is a constant with units of mass which, for convenience, we will take as:
\begin{equation}
{1\over m^2}\,=\,\sqrt{4\pi g_s N_c}\,\,\alpha'\,\,.
\label{m}
\end{equation}
Notice that in this setup there is another radial coordinate $\rho$, which represents the distance along $\mathbb{R}^{4}$, the directions
orthogonal to both the D3-brane worldvolume and the CY cone. Moreover, $d\Omega_3^2$ is the metric of a unit three-sphere. Furthermore, the function $z$ (which controls the size of the cycle) and the warp factor $H$ should be considered as functions of the two radial variables $(\rho, \sigma)$: $H\,=\,H(\rho,\sigma)$, $z\,=\,z(\rho,\sigma)$.
As in any background created by D3-branes, our solution should be endowed with
a self-dual RR five-form $F_5$, that we write as:
\begin{equation}
F_5\,=\,{\cal F}_5\,+\,{}^*{\cal F}_5\,\,.
\label{F5}
\end{equation}
The presence of $N_f$ flavor D3-branes induces a violation of the Bianchi identity of $F_5$. Indeed, the WZ term of the flavor brane action contains the term $\sum_{N_f}\,\int_{{\cal M}_4}\,\hat C_4$ that acts as a source for this violation. Actually, the smearing procedure amounts to performing the following substitution in this term:
\begin{equation}
\sum_{N_f}\,\int_{{\cal M}_4}\,\hat C_4\,\rightarrow\,\int_{{\cal M}_{10}}\,
\Omega\wedge C_4\,\,,
\end{equation}
where $\Omega$ is a six-form proportional to the volume form of the space transverse to the worldvolume of the flavor brane. The modified Bianchi identity takes the form
$dF_5=2\kappa_{10}^2\,T_3\,\Omega$. As in the four-dimensional example discussed in subsection \ref{secN2}, we shall locate the flavor branes at a particular value $\rho=\rho_Q$ of the $\rho$ coordinate (the mass of the matter fields is just $m_Q=\rho_Q/(2\pi\alpha')$). Moreover, we will smear the $N_f$ D3-branes along the angular directions $(\theta, \phi)$ of the cycle as well as along the external three-sphere. The corresponding smearing form is:
\begin{equation}
\Omega\,=\,-{N_f\over 8\pi^3}\,\,\,\delta(\rho-\rho_Q)\,d\rho\wedge \omega_3\wedge\omega_2\,\,,
\label{Omega}
\end{equation}
with $\omega_2=\sin\theta d\theta\wedge d\phi$ and $\omega_3$ is the volume element of the external $S^3$ with line element $d\Omega_3^2$ (the minus sign in (\ref{Omega}) is due to the orientation of the worldvolume required by supersymmetry). It is clear that the modified Bianchi identity in this case is:
\begin{equation}
dF_5\,=\,-2\pi\,g_s\,(\alpha')^2\,N_f\,\delta(\rho-\rho_Q)\,d\rho\wedge \omega_3\wedge\omega_2\,\,.
\label{newBianchi2d}
\end{equation}
Accordingly, let us represent $ F_5$ as in (\ref{F5}) with ${\cal F}_5$ being given by:
\begin{equation}
{\cal F}_5\,=\,f_5\,-\,2\pi\,g_s\,(\alpha')^2\,N_f\,\Theta(\rho-\rho_Q)\,\omega_3\wedge \omega_2\,\,,
\label{calF5-flavored}
\end{equation}
with $f_5$ such that $df_5=0$. We shall represent $f_5$ in terms of a potential ${\cal C}_4$ as $f_5=d{\cal C}_4$, where ${\cal C}_4$ is given by the ansatz:
\begin{equation}
{\cal C}_4\,=\,g\,\omega_3\,\wedge\,(d\psi+\cos\theta \,d\phi)\,\,,
\qquad g=g(\rho,\sigma)\,\,.
\label{CalC4}
\end{equation}
Proceeding as in subsection \ref{secN2}, one gets in this case the following set of BPS equations:
\begin{eqnarray}
&&m^2\,\big[\,g\,-\,2\pi\,g_s\,(\alpha')^2\,N_f\,\Theta(\rho-\rho_Q)\,\big]\,=\,\rho^3\,z'\,\,,\nonumber\\\rc
&&m^2\,H\,=\,{z\dot z\over \sigma}\,\,,
\qquad\qquad\qquad
g'\,=\,-\sigma\,\rho^3\,\dot H\,\,,\nonumber\\\rc
&&\dot g\,=\,{\sigma \rho^3\over z}\,\,H'\,-\,{\sigma\over z^2}\,H\,m^2\,
\big[\,g\,-\,2\pi\,g_s\,(\alpha')^2\,N_f\,\Theta(\rho-\rho_Q)\,\big]
\,\,,
\label{flavored-BPSsystem}
\end{eqnarray}
where the prime and the dot have the same meaning as in (\ref{prime&dot}). The fulfillment of (\ref{flavored-BPSsystem}) ensures the preservation of eight supersymmetries by the background, which corresponds to ${\cal N}=(4,4)$ SUSY of the dual gauge theory. Moreover, one can prove that $z(\rho,\sigma)$ satisfies the following PDE:
\begin{equation}
\rho\,z\,(\dot z\,-\,\sigma\,\ddot z\,)\,=\,\sigma\,(\, \rho\dot z^2\,+\,\rho z''\,+\,3 z'\,)\,+\,
{N_f\over 2 N_c}\,{\sigma\over m^2\,\rho^2}\,\,\delta (\rho-\rho_Q)\,\,.
\label{flavoredPDE}
\end{equation}
In the unflavored case $N_f=0$, the BPS system (\ref{flavored-BPSsystem}) (and the PDE equation (\ref{flavoredPDE})) can be solved analytically \cite{Arean:2008az} by constructing the solution in five-dimensional gauged supergravity and by uplifting it to ten dimensions \cite{Maldacena:2000mw}. After a suitable change of variables one can show \cite{Arean:2008az} that the metric and RR five-form of this gauged supergravity solution can be written as in our ansatz. In the general flavored case one has to apply numerical techniques. However, as in the four-dimensional case, one only needs to know the solution for $\sigma=0$ in order to get the behavior of the gauge coupling. Indeed, by means of a probe calculation one can check \cite{Arean:2008az} that the supersymmetric locus of a color D3-brane occurs precisely at $\sigma=0$ and that the gauge coupling is related to $z(\rho, \sigma=0)$ by means of the relation:
\begin{equation}
{1\over g^2_{YM}(\rho)}\,=\,{z(\rho, \sigma=0)\over m^2 g_s}\,\,.
\label{gYM-gravity}
\end{equation}
It follows from the system (\ref{flavored-BPSsystem}) that $g(\rho, \sigma=0)$ is constant. Actually, by using a flux quantization condition similar to the one employed for the 4d case, one can verify that $g(\rho, \sigma=0)=1/m^4$, where $m$ is the constant defined in (\ref{m}). By using this result in the first equation in (\ref{flavored-BPSsystem}) one readily integrates $z(\rho, \sigma=0)$. By imposing continuity of the solution at $\rho=\rho_Q$, one gets:
\begin{equation}
z(\rho, 0)\,=\,z_{*}\,-\,{\pi m^2 g_s\,(\alpha')^2\over \rho_Q^2}\,N_f\,
\Theta(\rho-\rho_Q)\,-\,
{2\pi\, m^2\,g_s\,(\alpha')^2\over \rho^2}
\,\Big[\,N_c\,-\,{N_f\over 2}\, \Theta(\rho-\rho_Q)\,\Big]\,\,,
\label{z-sigma0-flavored}
\end{equation}
where $z_{*}$ is a constant of integration. Plugging this result in (\ref{z-sigma0-flavored}), and assuming that the energy scale $\mu$ is related to the holographic coordinate $\rho$ as $\rho=2\pi\alpha'\,\mu$, one gets:
\begin{equation}
{1\over g_{YM}^2(\mu)}\,=\,
{1\over g^2_{YM}}\,\Big(\,1\,-\,{g^2_{YM}\over 2\pi\mu^2}\,\big(\,N_c\,-\,{N_f(\mu)\over 2}\,\big)\Big)\,\,,
\label{44gymrunning-gravity}
\end{equation}
where $N_f(\mu)$ is again the number of flavors with mass smaller that the scale $\mu$ and $g_{YM}$ is the bare UV Yang-Mills coupling. The dependence on the scale $\mu$ of the Yang-Mills coupling displayed in (\ref{44gymrunning-gravity}) matches precisely the one in field theory, which constitutes a non-trivial test of the gravity result.
Backgrounds dual to 2d theories with ${\cal N}=(2,2)$ SUSY can be obtained by wrapping D5-branes along a four-cycle of a Calabi-Yau threefold \cite{Arean:2009gc}. An alternative construction, which improves the UV behavior of the solution, involves D3-branes wrapping a two-cycle of a $CY_3$ \cite{Maldacena:2000mw, D3D4D5}. One can further reduce the amount of supersymmetry by considering a D5-brane wrapping a four-cycle of a manifold of $G_2$ holonomy, which leads to a dual of an ${\cal N}=(1,1)$ supersymmetric gauge theory. In all these cases the flavor branes are extented along some of the non-compact normal directions of the cycle wrapped by the color branes and the corresponding backreacted solutions can be obtained numerically and are similar to the one reviewed here.
\subsubsection{Three-dimensional theories}
A similar analysis can be carried out to obtain the gravity dual of ${\cal N}=4$ three-dimensional gauge theories. In this case one must consider flavor and color D4-branes wrapping two-cycles according to the array:
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\multicolumn{4}{c}{ }&
\multicolumn{4}{c}{$\overbrace{\phantom{\qquad\qquad\qquad}}^{\text{CY}_2}$}\\
\hline
&\multicolumn{3}{|c|}{$\mathbb{R}^{1,2}$}
&\multicolumn{2}{|c|}{$S^2$}
&\multicolumn{2}{|c|}{$N_2$}
&\multicolumn{3}{|c|}{$\mathbb{R}^{3}$}\\
\hline
$N_c$\,\,\,D$4$ (color) &$-$&$-$&$-$&$\bigcirc$&$\bigcirc$&$\cdot$&$\cdot$&$\cdot$&$\cdot$&$\cdot$\\
\hline
$N_f$\,\,\,D$4$ (flavor) &$-$&$-$&$-$&$\cdot$&$\cdot$&$-$&$-$&$\cdot$&$\cdot$&$\cdot$\\
\hline
\end{tabular}
\label{44flavored-arrayD4}
\end{center}
The concrete ansatz for the ten-dimensional string frame metric we will adopt in this case is very similar to the 2d and 4d cases studied above, namely:
\begin{eqnarray}
&&ds^2=e^{2\Phi}\left[dx^2_{1,2}+{z \over m^2}\,\left(d\tilde{\theta}^2+\sin^2\tilde{\theta}d\tilde{\phi}^2\right)\right]\,+\,\nonumber\\
&&+e^{-2\Phi}\left[\frac{1}{z}\left(d\sigma^2+\sigma^2\left(d\psi+\cos\tilde{\theta}d\tilde{\phi}\right)^2\right)+
d\rho^2+\rho^2\left(d\theta^2+\sin^2\theta d\phi^2\right)
\right] \ ,
\label{metric-ansatz}
\end{eqnarray}
where $\Phi=\Phi(\rho,\sigma)$ is the dilaton and the constant $m$ is now given by:
\begin{equation}
{1\over m^3}\,=\,8\pi g_s\,N_c\,(\alpha')^{{3\over 2}}\,\,.
\label{R}
\end{equation}
As before, $z=z(\rho,\sigma)$ and the background should include an RR form whose Bianchi identity is violated due to the presence of flavor branes. For D4-branes the appropriate RR form is a four-form $F_4$. If we locate the flavor branes at a fixed distance $\rho=\rho_Q$ in the transverse $\mathbb{R}^{3}$ and we smear them along their orthogonal angular directions, the modified Bianchi identity is:
\begin{equation}
dF_4\,=\,2\,\kappa_{10}^2 \, T_4\,\Omega\,=\,\frac{N_f}{2 N_c}\,\,\frac{1}{8 m^3}\,\delta(\rho-\rho_Q)
\,d\rho\wedge\, \omega_2 \, \wedge \, \tilde{\omega}_2 \, ,
\label{newBianchi3d}
\end{equation}
where $\omega_2$ and $\tilde{\omega}_2$ are the volume forms of the unit $(\theta, \phi)$ and $(\tilde\theta, \tilde\phi)$ two-spheres. Let us solve (\ref{newBianchi3d}) by means of the following ansatz:
\begin{equation} \label{F4-flavor}
F_4\,=\, dC_3\,+\,
\frac{N_f}{2 N_c}\,\,\frac{1}{8\,m^3}\,\,\Theta(\rho-\rho_Q)
\,\,\tilde{\omega}_2 \, \wedge \, \omega_2 \, ,
\end{equation}
where $C_3$ is the following potential depending on the function $g(\rho,\sigma)$:
\begin{equation}
C_3\,=\, - \, g\,\omega_2 \, \wedge \, (d\psi+\cos\tilde{\theta} \,d\tilde{\phi}) \, .
\label{C3}
\end{equation}
By imposing that the system preserves eight supersymmetries, we arrive at the following system of BPS equations:
\begin{eqnarray}
\label{flavored-BPSsystemD4}
&&g\,+\,{N_f \over 2\,N_c}\,{1\over
8\,m^3}\,\Theta(\rho-\rho_Q)\,=-\ \,{\rho^2\,{z}' \over m^2} \ , \quad \quad\qquad
e^{-4\Phi}\,\sigma\,=\,{1\over m^2}\,\, z\dot{z} \ , \nonumber\\
\\
&& g'\,=\,-4\,\sigma\,\rho^2\,e^{-4\phi}\dot{\Phi} \ , \nonumber\\\rc
&&\dot g\,=\,\,-m^{2}\,\sigma\,{
z}^{-2}\,e^{-4\Phi}\,\left[\,g\,+\,{N_f \over 2\,N_c}\,{1
\over 8\,m^3}\,\Theta(\rho-\rho_Q)\,\right]\,+\,4\,\sigma\,\rho^2\,{
z}^{-1}\,e^{-4\Phi}\,\Phi' \, . \nonumber
\end{eqnarray}
Again, one can combine the different equations in (\ref{flavored-BPSsystemD4}) and get a single second-order PDE for $z(\rho,\sigma)$, namely:
\begin{equation}
\rho^2\,{z}\,\left(\,\dot{{z}}\,-\,\sigma\, \ddot{{
z}}\,\right)\,=\,\rho\,\sigma\,\left(\,\rho\,\dot{{z}}^2\,+\, \rho\,{
z}''\,+\, 2 {z}'\,\right)\,+\sigma\,{N_f \over 2\,N_c}\,{1
\over 8\,m^3}\,\delta(\rho-\rho_Q).
\label{PDE-z-flavored}
\end{equation}
As in the 4d and 2d cases, eqs. (\ref{flavored-BPSsystemD4}) and (\ref{PDE-z-flavored}) can be solved analytically when $N_f=0$ by using gauged supergravity \cite{Maldacena:2000mw,Divecchia}. In the general flavored case one can get analytically the form of the solution for $\sigma=0$ \cite{Ramallo:2008ew}. Indeed, one can verify from (\ref{flavored-BPSsystemD4}) and the corresponding flux quantization condition
that $g(\rho,\sigma=0)=-1/(8m^3)$ and that $ z(\rho, 0)$ is:
\begin{equation}
z(\rho, 0)\,=\, z_{*}\,- \,{1\over 8\,\rho_Q m }
\,{N_f\over 2\,N_c}\,\Theta(\rho-\rho_Q)\,-\,{1\over 8m\,\rho}
\left[\,1\,-\,{N_f\over 2\,N_c}\,\Theta(\rho-\rho_Q)\,\right]
\, ,\label{z-sigma0-flavorD4}
\end{equation}
with $ z_{*}$ being a constant. Moreover, by means of a probe calculation one readily verifies that $\sigma=0$ is the SUSY locus of the color D4-branes and that the relation between the YM coupling and $ z(\rho, 0)$ is
$ g^{-2}_{YM}(\rho)=z(\rho, 0)/(2\pi g_s \sqrt{\alpha'}\,m^2)$. Using this result and the radius-energy relation $\rho=2\pi\alpha'\mu$, one can convert (\ref{z-sigma0-flavorD4}) into the following equation for the running of the YM coupling of the 3d theories:
\begin{equation}
{1\over g_{YM}^2(\mu)}\,=\, {1\over
g^2_{YM}}\,\Big[\,1\,-\,{g^2_{YM} \over
4\pi\mu}\,\Big(\,N_c-{N_f(\mu)\over 2}\,\Big) \,\,\Big] \ ,
\label{gYM-QFT-D4}
\end{equation}
which again matches the corresponding field theory result.
A gravity dual of ${\cal N}=2$ three-dimensional gauge theory based on D5-branes wrapping a three-cycle was found in \cite{Gomis:2001aa,Gauntlett:2001ur}. The addition of flavor to this background along the lines discussed here is carried out in \cite{Gaillard:2008wt,D3D4D5}. Alternatively, for this same amount of supersymmetry one can consider D4-branes wrapping a two-cycle of a Calabi-Yau threefold \cite{Maldacena:2000mw,D3D4D5}.
\setcounter{equation}{0}
\section{A Mathematical Viewpoint}
\label{mathviewpointzz}
In the approach we have followed up to now in this review on how to add
unquenched flavor, we considered a family of equivalent embeddings of the flavor branes. This family can be generated by acting with the isometries of the background on a
fiducial representative embedding. When the number $N_f$ of flavor branes is large, considering the set of branes as a continuous distribution is a good approximation. We then computed the RR charge density generated by the branes, \hbox{\it i.e.}\ the smearing form $\Omega$, by explicitly performing the average over the set of embeddings and, subsequently, we have studied the deformation induced on the metric and forms due to the backreaction.
The outcome of this microscopic approach is a system of supergravity plus delocalized sources preserving some amount of supersymmetry. It turns out that, in this process, very interesting mathematical structures emerge. The reason for this is the fact that the supersymmetric sources that we are using satisfy a calibration condition. As a consequence, one can use the methods of modern geometry to find backgrounds with smeared flavors in a systematic way \cite{Gaillard:2008wt}. In this procedure one does not deal with the set of embeddings and, for this reason, we will refer to it as the macroscopic approach, as opposed to the microscopic approach reviewed in previous sections. The goal is computing (or at least constraining) the smearing form $\Omega$ by using the same type of technology as the one employed in the analysis of flux compactifications of string theory (see refs. \cite{Gutowski:1999tu}-\cite{Koerber:2006hh}, \cite{Koerber:2007hd}).
The central object in this geometric approach is the
so-called `calibration form' ${\cal K}$. For D$p$-branes ${\cal K}$ is a
$(p+1)$-form, which can be represented in a vielbein basis as:
\begin{equation}
{\cal K}\,=\,{1\over (p+1)!}\,{\cal K}_{a_1\cdots a_{p+1}}\,e^{a_1\,\cdots a_{p+1}}\,\,,
\end{equation}
with $e^{a_1\,\cdots a_{p+1}}=e^{a_1}\wedge\cdots \wedge e^{a_{p+1}}$. The
different components ${\cal K}_{a_1\cdots a_{p+1}}$ are given by fermionic bilinears of the type:
\begin{equation}
{\cal K}_{a_1\cdots a_{p+1}}\,=\,\,\epsilon^{\dagger}\,\tau\,\Gamma_{a_1\,\cdots a_{p+1}}\,\epsilon\,\,,
\label{K-bilinear}
\end{equation}
where $\epsilon$ are Killing spinors of the background, conveniently normalized, and
$\tau$ is a constant matrix which (in the type IIB theory) is $\tau=\tau_3 ^{{p-3\over 2}}\,i\tau_2$, where $\tau_2$ and $\tau_3$ are Pauli matrices and the spinor $\epsilon$ is represented as a two-dimensional vector of Majorana-Weyl spinors ($\tau$ is the same matrix that appears in the expression of the kappa symmetry matrix $\Gamma_{\kappa}$ of a D$p$-brane when all worldvolume fluxes are switched off). The form ${\cal K}$ can be used to characterize $(p+1)$-dimensional surfaces in the ten-dimensional geometry. A $(p+1)$-dimensional surface ${\cal M}_{p+1}$ is said to be calibrated by ${\cal K}$ if its pullback to ${\cal M}_{p+1}$ is equal to the induced volume form on ${\cal M}_{p+1}$, namely:
\begin{equation}
\hat {\cal K}\,=\,\sqrt{-\det \hat g}\,\,\,\,d^{p+1}\,\xi\,\,,
\label{cal-con}
\end{equation}
where the $\xi$'s are a set of local coordinates of ${\cal M}_{p+1}$.
When there are no NSNS fluxes or worldvolume gauge fields the calibration
condition (\ref{cal-con}) characterizes the supersymmetric embeddings of D$p$-branes (this can be easily established by using kappa symmetry). Actually, a D$p$-brane whose worldvolume ${\cal M}_{p+1}$ is calibrated by ${\cal K}$ is electrically charged with respect to an RR (p+2)-form field strength $F_{p+2}$ and, in the Einstein frame, $F_{p+2}$ is related to ${\cal K}$ as:
\begin{equation}
F_{p+2}\,=\,d\Big(\,e^{{p-3\over 4}\phi}\,\,{\cal K}\,\Big)\,\,.
\label{F-K}
\end{equation}
Eq. (\ref{F-K}) is a consequence of supersymmetry \cite{Gutowski:1999tu} and, actually, in our backreacted backgrounds it follows from the system of BPS equations. Moreover, as a consequence of (\ref{cal-con}), the action of a localized embedding of a D$p$-brane (without NSNS flux and with worldvolume gauge fields switched off) can be written as:
\begin{equation}
S_{Dp}^{loc}\,=\,-T_p\,\int_{{\cal M}_{p+1}}\,\,
\Big[\,e^{{p-3\over 4}\phi}\,\hat {\cal K}\,-\,\hat C_{p+1}\,\Big]\,\,.
\label{loc-actionDp}
\end{equation}
Following our prescription, the smeared version of the brane action is obtained by performing the wedge product with $\Omega$ of the $(p+1)$-form inside the brackets in (\ref{loc-actionDp}) and by integrating the result over the full ten-dimensional spacetime:
\begin{equation}
S_{Dp}^{smeared}\,=\,-T_p\,\int_{{\cal M}_{10}}\,\,
\Big[\,e^{{p-3\over 4}\phi}\, {\cal K}\,-\, C_{p+1}\,\Big]\wedge \Omega\,\,.
\label{smeared-Dp}
\end{equation}
Let us now define the $(8-p)$-form $F_{8-p}$, under which the D$p$-brane is magnetically charged, as:
\begin{equation}
F_{8-p}\,=\,\pm e^{{3-p\over 2}\phi}\,\,{}^*F_{p+2}\,\,,
\label{F(8-p)-F(p+2)}
\end{equation}
where the sign depends on the particular value of $p$ and on the conventions used. As in the examples studied in previous sections, the D$p$-brane modifies the Bianchi identity of $F_{8-p}$, namely:
\begin{equation}
dF_{8-p}\,=\,\pm 2\kappa_{10}^2\,T_p\,\Omega\,\,.
\label{Bianchi-Dp}
\end{equation}
Eq. (\ref{Bianchi-Dp}) establishes a crucial
connection between the smearing form $\Omega$ and
the calibration form ${\cal K}$. Indeed, by using (\ref{F-K}) and (\ref{F(8-p)-F(p+2)}), the right-hand side of (\ref{Bianchi-Dp}) can be written in terms of ${\cal K}$ and its exterior derivative. Moreover, from the inspection of the smeared brane action (\ref{smeared-Dp}), one concludes that $\Omega$ can be regarded as a kind of orthogonal complement
(the Poincare dual) of ${\cal K}$ in ${\cal M}_{10}$. Interestingly,
the possible calibration forms ${\cal K}$ in a
manifold are known and are related to
its supersymmetric cycles and G-structures.
In the case of a manifold preserving minimal SUSY in 4d, ${\cal K}$ can
be written in terms of powers
of the K\"ahler form and of the holomorphic volume form. Thus, geometry and topology constrain the form of the charge density distribution of supersymmetric configurations and, actually, one could adopt the point of view in which the expression of $\Omega$ is determined from these constraints without explicitly performing the average over the family of embeddings. This macroscopic approach was followed in refs. \cite{Gaillard:2008wt,Gaillard:2009kz,Arean:2009gc, D3D4D5} for some particular brane setups.
To finish this section let us detail the implementation of these mathematical concepts in the case discussed in section \ref{AdS5X5}, namely the D3-D7 system. From now on we will assume that the metric, dilaton and forms are given by the expressions written in (\ref{metrictzero}) and (\ref{dilaton-forms}). It is convenient to define the following two-form:
\begin{equation}
{\cal J}\,=\,h^{{1\over 2}}\,\,\big[\,e^{2g}\,J_{KE}\,+\,e^{2f}\,d\rho\wedge (d\tau+A_{KE})\,\big]\,\,,
\end{equation}
which is such that $h^{-{1\over 2}}\,{\cal J}$ is the K\"ahler form of the transverse 6d space. Actually, one can immediately verify that $d\big[\,h^{-{1\over 2}}\,{\cal J}\,\big]=0$ as a consequence of the BPS equation for $g$ in (\ref{BSPsysKW}). By explicitly computing the fermion bilinear in (\ref{K-bilinear}) and by using the projections satisfied by the Killing spinor of the flavored $AdS_5\times S^5$ background, one gets that the calibration form ${\cal K}$ in this case is given by:
\begin{equation}
{\cal K}\,=\,{1\over 2}\,{\rm Vol}(M_{1,3})\wedge {\cal J}\wedge {\cal J}\,\,,
\end{equation}
with ${\rm Vol}(M_{1,3})=h^{-1}\,d^4x$ being the volume form of the Minkowski part of the space.
Using the fact that $dA_{KE}=2J_{KE}$, one gets:
\begin{equation}
d\Big( e^{\phi}\, {\cal K}\,\Big)\,=\,{1\over 2}\,\,e^{2g+\phi}\,\,
\Big[\,\big(4g'+\phi')\,e^{2g}\,-\, 4 e^{2f}\,\Big]\,d^4x\,\wedge
J_{KE}\wedge J_{KE}\wedge d\rho\,=\,F_9\,\,,
\label{F9-K}
\end{equation}
where, in the last step, we have used the condition (\ref{F-K}) for $p=7$. Let us now verify that the value of $F_9$ obtained in (\ref{F9-K}) is consistent with the expression for $F_1$ written in our ansatz (\ref{dilaton-forms}) and, thus, with the $\Omega$ displayed in eq. (\ref{gsomega}). Taking into account that the volume form of the KE space is ${1\over 2} J_{KE}\wedge J_{KE}$, one can easily compute the Hodge dual of $F_1$ and get the following result for $F_9$:
\begin{equation}
F_9\,=\,-e^{2\phi}\,{}^*F_1\,=\,{Q_f\over 2}\,p(\rho)\,e^{4g+2\phi}\,
d^4x\,\wedge J_{KE}\wedge J_{KE}\wedge d\rho\,\,.
\label{F9-F1}
\end{equation}
The expressions (\ref{F9-K}) and (\ref{F9-F1}) for $F_9$ coincide if the following relation holds:
\begin{equation}
Q_f\,p(\rho)\,=\,e^{-\phi}\,\big[\,4g'+\phi'\,-\,4\,e^{2f-2g}\,\big]\,\,.
\label{Qf-relation}
\end{equation}
One can easily check that (\ref{Qf-relation}) is a consequence of the BPS system (\ref{BSPsysKW}).
\section{Discussion}
\label{outlook}
In hindsight, we can say that the program of finding
solutions dual to theories with unquenched fundamentals
with smeared flavor branes has been quite successful.
As expected, it simplifies matters both when looking for the background solution
and when discussing the physics they encode.
We have presented a series of example of solutions of ten-dimensional type
IIA or type IIB supergravity coupled to D-brane sources. The philosophy and
methods used in the different cases are quite similar. Finding a consistent
solution requires solving at the same time the closed string degrees of freedom
(namely, finding solutions of the generalised Einstein equations in the presence
of sources) and the open string degrees of freedom (namely, checking that the D-brane
embeddings which generate the mass and charge source density are indeed solutions of
the background).
Supersymmetric solutions are easier to deal with and indeed, preserving SUSY
simplifies enormously the technical work.
It is rather remarkable that sometimes such complicated coupled systems can be
(at least almost) completely integrated and the solutions given in a simple closed
form (in particular in sections \ref{AdS5X5}, \ref{KS};
for the other sections,
profuse numerical integration was necessary).
However, supersymmetry is not mandatory for the construction and we
have presented non-supersymmetric black hole solutions.
The solutions are conjectured to be dual to theories with
unquenched quarks. Since we have always dealt with the particular case of smearing
the flavor branes over the transverse directions, we have built duals of a very
particular class of such unquenched theories.
We have used the solutions to discuss many physical features of the different set-ups.
Many crosschecks of field theory expectations have been discussed. Just to mention
a few instances, the
running of the gauge coupling in different theories, the behaviour of the cascade in
section \ref{KS} or the direct computation of the first flavor contribution
to the entropy of the
D3-D7 plasma (section \ref{D3D7plasma}) which was previously known from an indirect method
(namely, from first computing the free energy) \cite{myers}.
All this asserts that the dualities discussed in this review are on firm ground.
Due to obvious space constraints, we have not been able to include all the material
that may deserve to be reviewed, but we hope that we have given enough references to the
original literature.
\vskip.5cm
It is worth recapitulating
about the presence of singularities in the different solutions
discussed. First, in all the cases presented
there are IR curvature singularities when all of the flavor branes reach the bottom of
the geometry, see the heuristic picture of section \ref{heuristiczzz}.
They pertain to the kind of singularities usually called {\it good}.
In fact, we have shown explicitly in the examples of sections \ref{AdS5X5} and
\ref{KS} how adding (even an infinitesimal) quark mass leads to regular backgrounds
(the analogous generalization for the set-ups of sections \ref{sec:D5D5}, \ref{sec:2+1}
remains an interesting
open question). Moreover, heating up the theories can result in the formation of a black
hole horizon behind which the IR singularity is hidden, see section \ref{D3D7plasma}
and eq. (\ref{bhflav}) for examples.
On the other hand, the solutions of sections \ref{AdS5X5} and \ref{KS} are singular in the UV
(an effect connected to having flavor D7-branes)
since the dilaton diverges at a finite position of the radial variable.
This is expected, since
it is the consequence of the Landau pole of the dual theory (more precisely,
in the case of
section \ref{KS} it is a duality wall). Despite
the singularity, we have shown that it is possible to consistently compute IR observables
as long as the IR scale is well separated from the pathological UV.
Clear examples are the meson spectrum (section
\ref{sec:screening}) and the black hole thermodynamical properties
(section \ref{sec:thermo}).
In the D5D5 setups of sections \ref{sec:D5D5}, \ref{sec:2+1}, the dilaton diverges linearly
in the UV, signalling a little string theory-like UV completion of the dual field theory.
We want to stress here that this already happens in the unflavored solutions and thus
is not problem associated to the backreaction.
Finally, all the solutions in section \ref{models2} have a singularity in the IR.
This singularity is not associated to the flavors as it is already present in the
flavorless solutions and, at least in some cases, can be resolved by the
worldsheet CFT \cite{Hori:2002cd}. On top of that,
for the same models, typically,
when $N_f$ is sufficiently large, a Landau pole is generated and, jointly, a UV
singularity appears in the geometry.
\vskip.5cm
We end this discussion with two clarifications.
\begin{itemize}
\item
We have repeatedly
stressed that our main goal is to build duals
to theories in which $N_c$ and $N_f$ are of the same order. Nevertheless, for the set-ups
discussed in section \ref{AdS5X5}, which include the
specially interesting
flavored $AdS_5 \times S^5$ case,
$N_f \ll N_c$ is needed, see section \ref{sec:rangeval} (similar comments
apply to section \ref{KS}). This is because, starting with a conformal theory, the
introduction of extra matter generates an UV pathology, namely a Landau pole.
Then, roughly speaking, in order to have a meaningful IR, it has to be well separated
from the pathological region, enforcing the number of flavors not to be too large.
However, backreaction effects and, accordingly, the effect of unquenched quarks,
can still be computed as an expansion in
$N_f/N_c$.
On the other hand, for the models in sections \ref{sec:D5D5}, \ref{sec:2+1},
\ref{models2}, this restriction is not present and, indeed, it makes sense
to talk about solutions with $N_f \sim N_c$. In fact, this is imperative for
instance when discussing Seiberg-like dualities as in section \ref{sec:Seib}.
\item
Since the (DBI) action is used to model the D-brane
sources, one could be wary for the following reason: the effective string coupling
on a stack of $N_f$ D-branes is $g_s N_f \sim N_f / N_c$ and this should be small
for the DBI to be a good approximation \cite{Callan:1986bc}, whereas there is not a good effective description
for strong string coupling. However, this caveat is
circumvented because we do not deal with
stacks of localized flavor branes: due to the smearing, the
typical distance between any pair of flavor branes is of the order of the
size of the transverse space, which is typically large. As a result, the flavor symmetry
is usually broken to $U(1)^{N_f}$ and the effective open string coupling
remains small.
As already pointed out, this amounts to keeping
``one window graphs'' in the Veneziano expansion \cite{Bigazzi:2008zt}.
\end{itemize}
\subsubsection*{Outlook}
There are still many open questions that deserve to be addressed within the framework presented
in this review.
We briefly mention a few examples of
possible future projects.
They comprise both making further progress in studying the models here
presented and building new solutions that could be useful in exploring the consequences
of the formalism for different physical points.
Along the first of these lines, it would be nice to generalise the solutions of section \ref{sec:D5D5}
to the massive quark case in order to remove the IR singularity. Also, we expect the black hole
solution of section \ref{D3D7plasma} to encode interesting physical information. For instance, one
could consider massive embeddings in the search of a first order phase
transition similar
to those in \cite{Kruczenski:2003uq},\cite{Kirsch:2004km},\cite{myers}.
The peculiarity of the back-reacted setting
would be, conceivably, that the area of the horizon would undergo a finite jump
at the transition.
Along the second line, a back-reacted D4-D6 solution building on the model of
\cite{Kruczenski:2003uq} could be useful in discussing QCD-like properties.
Another conceivable program is to look for a solution, which, as in
\cite{Chen:2009kx}, may correspond to a color-flavor locking phase. The
study of fluctuations in these backgrounds, that will also contain
fluctuations of the fields in the flavor branes, with a view on
understanding holographic renormalization would be a highly interesting
result.
Aside from this, it would be nice to find solutions (with backreacted
flavor branes) that contain an $AdS_5$ factor. The study of conformal
anomalies there may give interesting results.
As stressed in the introduction, finding the kind of solutions discussed
here, including the
D-brane backreaction,
has an interest on their own, independently of AdS/CFT.
It would be nice to understand whether they
may turn out to be useful for different physical applications.
For instance, for models of inflation built with D3-D7 systems on the conifold
(see \cite{Baumann:2010sx} for recent progress in this direction), the analysis of section \ref{KS}
could have some relevance.
\section*{Acknowledgments}
This review is based on work done with several
collaborators in the last few years.
We thank all of them for their insights and the many discussions
during the course of those collaborations.
Aside from them, we are specially grateful to D. Are\'an,
A. Armoni, F. Bigazzi, E. Conde, A.
Cotrone, S.
Cremonesi, J.
Gaillard, D. Martelli, I. Papadimitriou, J. Shock,
J. Tarr\'\iota o
and D. Zoakos for a critical
reading of the manuscript and the many useful comments, discussions and
remarks they made.
The research of A.P is
supported by grants FPA2007-66665C02-02 and DURSI
2009 SGR 168, and by the CPAN CSD2007-00042 project of the
Consolider-Ingenio
2010 program. The work of AVR was funded in part by MEC and FEDER under grant
FPA2008-01838, by the Spanish Consolider-Ingenio 2010 Programme CPAN (CSD2007-00042) and by Xunta de Galicia (Conseller\'\iota a de Educaci\'on and grant INCITE09 206 121 PR).
|
\section{Introduction}
The classical topic of separation of variables, often realized by a Howe dual
pair acting on the representation of interest, is one of the cornerstones in harmonic
analysis.
Some well-known examples are the Howe dual pair for the orthogonal Lie group $O(n)$ acting on the space of polynomials on $\mathbb{R}^{n}$ given by $O(n)\times \mathfrak{sl}_{2}$ and the dual pair for the representation on spinor valued polynomials given by $O(n) \times \mathfrak{osp}(1|2)$. Recently, there has been a lot of interest in deformations of these dual pairs where the orthogonal group is replaced by a finite reflection group $\mathcal{G} < O(n)$. This leads to the introduction of the so-called Dunkl operators (see e.g. \cite{MR951883, MR1827871}). They allow to construct a Dunkl Laplacian and it was proven in \cite{He} that the relevant operators again generate $\mathfrak{sl}_{2}$ (see also \cite{BO}). In \cite{Orsted}, a Dunkl Dirac operator has been introduced, leading to a dual pair $\mathcal{G} \times \mathfrak{osp}(1|2)$.
The Dunkl Laplacian can be seen as a spherical deformation of the classical Laplacian. Very recently, also radial deformations have been obtained in this context. We refer the reader to \cite{Orsted2} for radial deformations of $\mathfrak{sl}_{2}$ and to \cite{H12} for radial deformations of $\mathfrak{osp}(1|2)$.
A natural question to be posed now is whether similar things are also possible for the symplectic group $Sp(2n)$, namely do there exist dual pairs $Sp(2n) \times \mathfrak{sl}_{2}$ or $Sp(2n) \times \mathfrak{osp}(1|2)$ and do there exist Dunkl analogs of these dual pairs. In other words, can the operators realizing these dual pairs be deformed to be only invariant under a (finite) subgroup of the symplectic group?
The answer is partly positive. In e.g. \cite{H4, HDBThesis} a theory of harmonic analysis has been established for operators acting on Grassmann algebras (or superspaces), leading to dual pairs $Sp(2n) \times \mathfrak{sl}_{2}$ and $Sp(2n) \times \mathfrak{osp}(1|2)$. However, although this yields a satisfying framework, it is unclear how the operators obtained there can be deformed to be only invariant under a subgroup of the symplectic group. The construction leading to the classical Dunkl operators (see \cite{MR951883}) cannot immediately be transferred to this new framework, due to the different algebraic structure of Grassmann algebras.
Fortunately, there seems to be an alternative way of obtaining symplectic analogs of the orthogonal constructions.
Many years ago B. Kostant introduced a symplectic analog of the Dirac operator called symplectic Dirac operator. This symplectic Dirac operator was mainly studied from the geometrical point of view, see \cite{MR2252919} and references
therein, and also as an invariant differential operator in \cite{KAD}, but its spectral properties are difficult to obtain.
In this article we aim to study the symplectic Dirac operator from the point of view of special functions and harmonic analysis. We will establish the Howe duality and the corresponding Fischer decomposition. As a consequence, we determine its kernel on $\mathbb{R}^{2n}$. It is important to note that the symplectic Dirac operator that we will study does not have a Laplace counterpart.
The paper is organized as follows. In section 2 and 3 we repeat some well-known facts on symplectic Lie algebras and their finite dimensional representations. In section 4 we define the symplectic Dirac operator and show how it leads to a realization of $\mathfrak{sl}_{2}$. In section 5 we obtain the Fischer decomposition and construct explicit projection operators on all summands. We end with some conclusions and an outlook for further research.
\section{Symplectic Lie algebra, symplectic Clifford algebra and simple
highest weight modules for $sp(2n)$}
Let us consider a symplectic vector space $(\mathbb{R}^{2n},\omega )$ and a
symplectic basis
$$
e_1,\dots ,e_n,f_1,\dots ,f_n
$$
with respect to the non-degenerate two-form $\omega$ on $\mathbb{R}^{2n}$.
Let $E_{i,j}$ be the $2n\times 2n$ matrix with $1$ on the intersection
of the $i$-th row and $j$-th column, and zero otherwise.
The symplectic Lie algebra $sp(2n)$ is generated by (see e.g. \cite{MR1153249})
\begin{eqnarray*}
X_{ij}=E_{i,j}-E_{n+i,n+j}, Y_{ij}=E_{i,n+j}+E_{j,n+i}, Z_{ij}=E_{n+i,j}+E_{n+j,i}
\end{eqnarray*}
for $i,j\in\{1,\dots ,n\}$, and can be realized by the first order differential operators
\begin{eqnarray*}
X_{ij}=x_j\partial_i-x_{n+j}\partial_{n+i}, Y_{ij}=x_{n+j}\partial_i+x_{n+i}\partial_j,
Z_{ij}=x_j\partial_{n+i}+x_i\partial_{n+j}.
\end{eqnarray*}
For example, in the case $n=1$ the three vector fields
$\{2x_2\partial_1,2x_1\partial_2,x_1\partial_1-x_2\partial_2\}$
generate $sp(2)$.
Next, we introduce the symplectic Clifford algebra over $(\mathbb{R}^{2n}, \omega)$ (see \cite{MR1044769}). This is the associative algebra over $\mathbb{R}$ with unit element, which is multiplicatively generated by the elements $e_{1}, \ldots, e_{n}, f_{1}, \ldots, f_{n}$ under the relations
\begin{align*}
&e_{i} e_{j} = e_{j} e_{i}\\
&f_{i} f_{j} = f_{j} f_{i}\\
&e_{i} f_{j} - f_{j} e_{i} = \omega(e_{i}, f_{j}) = \delta_{ij}
\end{align*}
for all $i, j \in \{1, \ldots, n\}$.
The metaplectic Lie algebra $mp(2n,\mathbb{R})$ is a Lie algebra attached to the twofold covering
$\rho : Mp(2n,\mathbb{R})\to Sp(2n,\mathbb{R})$ of the symplectic Lie group $Sp(2n)$. It can be realized
by homogeneity two elements in
the symplectic Clifford algebra, where the homomorphism $\rho_\star : mp(2n,\mathbb{R})\to sp(2n,\mathbb{R})$
is given by
\begin{eqnarray}
& & \rho_\star (e_ie_j)=-Y_{ij},
\nonumber \\
& & \rho_\star (f_{i}f_{j})=Z_{ij},
\nonumber \\
& & \rho_\star (e_if_{j}+f_{j}e_i)=2X_{ij}
\end{eqnarray}
for $i,j\in\{1,\dots ,n\}$.
There is another useful realization of the symplectic Lie algebra as a subalgebra
of the Weyl algebra of rank $n$. Let $x_i$, $i\in\{1,\dots ,n\}$ be the generators of
polynomial algebra. The Weyl algebra is the associative algebra generated by
$\{x_i,\partial_i\}$, $i\in\{1,\dots ,n\}$, partial differentiation with respect to
$x_i$ and multiplication operators $x_i$, acting on polynomials on $\mathbb{R}^n$. The
root spaces of $sp(2n)$ corresponding to positive simple roots $\alpha_i$ are
spanned by $x_{i+1}\partial_{i}$, $i\in\{1,\dots ,n-1\}$ and $\alpha_n$ is
spanned by $-\frac{1}{2}\partial_n^2$.
The Segal-Shale-Weil representation $L$ is the minimal highest weight representation
of $Sp(2n,\mathbb{R})$ on the vector space $L^2(\mathbb{R}^{n},d\mu)$, where
$d\mu =\exp^{-||x||^2} dx_{\mathbb{R}^{n}}$ with $dx_{\mathbb{R}^{n}}$ the Lebesgue measure
on $\mathbb{R}^{n}$. We take for the basis of this vector
space polynomials on the isotropic subspace $\mathbb{R}^{n}\subset\mathbb{R}^{2n}$.
The differential $L_\star :mp(2n,\mathbb{R})\to End(Pol(\mathbb{R}^n)) $ of the Segal-Shale-Weil representation is
\begin{eqnarray}
& & L_\star (e_ie_j)=ix_ix_j,
\nonumber \\
& & L_\star (f_{i}f_{j})=-i\partial_i\partial_j,
\nonumber \\
& & L_\star (e_if_{j}+f_{j}e_i)=x_i\partial_j+x_j\partial_i
\end{eqnarray}
for $i,j\in\{1,\dots ,n\}$.
\section{Decomposition of tensor products of finite dimensional representations with the Segal-Shale-Weil
representation}
In this section, we will make explicit several
results in \cite{BL} on the decomposition of the tensor product of completely
pointed highest weight modules (specifically, we will consider two irreducible components of
the Segal-Shale-Weil representation) with a suitable class of
finite dimensional representations (specifically, symmetric powers
of the fundamental vector representation $\mathbb{R}^{2n}$) of $sp(2n,\mathbb{R})$. The property of a module
being completely pointed means that all of its weight spaces
are uniformly bounded by a constant, see \cite{BL}, Lemma 2.1.
Throughout the article, $V(\mu)$ denotes the Verma module of highest weight $\mu$ and
$L(\mu)$ denotes the simple module of highest weight $\mu$, i.e. the quotient of $V(\mu)$
by its unique maximal submodule $I(\mu)\subset V(\mu)$.
Let us introduce the set
\[
\tau^{i}=\{\sum_{j=1}^n d_j L_j \,|\, d_j+\delta_{1,i}\delta_{n,j}\in\mathbb{N}, \sum_{j=1}^n d_{j}=0 \, mod\, 2\}.
\]
Here $\mathbb{N}$ means the set of natural numbers including $0$, $d_j\in\mathbb{N}$ and
$i=0$ resp. $i=1$ for $V(-\frac{1}{2}\omega_n)$ resp. $V(\omega_{n-1}-\frac{3}{2}\omega_n)$.
These sets are bijective with the set of weights of two irreducible parts of the
Segal-Shale-Weil representation.
Let $\lambda=\sum_{i=1}^{n}\lambda_i\omega_i$ be a dominant integral weight, i.e.
the highest weight of a finite dimensional irreducible representation. Define the
set of weights
\begin{eqnarray*}
\tau_\lambda^{i}&=&\{\mu|\lambda-\mu=\sum_{j=1}^nd_jL_j\in\tau^i,\\
&& 0\leq d_j\leq \lambda_j (j=0,\dots ,n-1),
0\leq d_n+\delta_{1,i}\leq 2\lambda_n+1\}.
\end{eqnarray*}
The central result of \cite{BL} is
\begin{thm}
Let $L(-\frac{1}{2}\omega_n)$ resp. $L(\omega_{n-1}-\frac{3}{2}\omega_n)$ denote the
Verma modules corresponding to irreducible subrepresentations of the Segal-Shale-Weil
representations. Then for any finite dimensional irreducible representation $F(\lambda)$
with highest weight $\lambda$ we have
\begin{eqnarray}
L(-\frac{1}{2}\omega_n)\otimes F(\lambda)\simeq \bigoplus_{\mu\in \tau_\lambda^{0}}L(-\frac{1}{2}\omega_n+\mu)
\end{eqnarray}
(the same is true for $L(\omega_{n-1}-\frac{3}{2}\omega_n)$ with $\tau_\lambda^{1}$ instead of
$\tau_\lambda^{0}$.) In particular, the decomposition is direct and so the tensor product is
completely reducible.
\end{thm}
Recall that the Verma modules $V(\mu)$ appearing in the previous decomposition are
irreducible, i.e. equal to $L(\mu)$. In particular, they have no other
singular vector than the highest weight one.
The consequence of this result is the decomposition of the tensor product
of $L(-\frac{1}{2}\omega_n)$ resp. $L(\omega_{n-1}-\frac{3}{2}\omega_n)$ with
symmetric powers $S^k(\mathbb{C}^{2n})$ ($k\in\mathbb{N}$) of the fundamental vector
representation $\mathbb{C}^{2n}$ of $sp(2n)$. Note that these are irreducible
representations (see \cite{MR1153249}).
\begin{cor}\label{decomposition}
We have for $L(-\frac{1}{2}\omega_n)$
\begin{enumerate}
\item
In the even case $k=2l$ ($2l+1$ terms on the right-hand side):
\begin{eqnarray*}
L(-\frac{1}{2}\omega_n)\otimes S^k(\mathbb{C}^{2n}) & \simeq &
L(-\frac{1}{2}\omega_n)\oplus L(\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n) \nonumber\\&&
\oplus L(2\omega_1-\frac{1}{2}\omega_n)\oplus L(3\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)
\oplus\dots
\nonumber \\
& & \oplus L((2l-1)\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)\oplus
L(2l\omega_1-\frac{1}{2}\omega_n),
\end{eqnarray*}
\item
In the odd case $k=2l+1$ ($2l+2$ terms on the right-hand side):
\begin{eqnarray*}
L(-\frac{1}{2}\omega_n)\otimes S^k(\mathbb{C}^{2n}) & \simeq &
L(\omega_{n-1}-\frac{3}{2}\omega_n)\oplus L(\omega_1-\frac{1}{2}\omega_n)
\\
&&\oplus L(2\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n) \oplus
L(3\omega_1-\frac{1}{2}\omega_n)
\oplus\dots
\nonumber \\
& & \oplus L(2l\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)\oplus
L((2l+1)\omega_1-\frac{1}{2}\omega_n).
\end{eqnarray*}
\end{enumerate}
We have for $L(\omega_{n-1}-\frac{3}{2}\omega_n)$
\begin{enumerate}
\item
In the even case $k=2l$ ($2l+1$ terms on the right-hand side):
\begin{eqnarray*}
L(\omega_{n-1}-\frac{3}{2}\omega_n)\otimes S^k(\mathbb{C}^{2n}) & \simeq &
L(\omega_{n-1}-\frac{3}{2}\omega_n)\oplus L(\omega_1-\frac{1}{2}\omega_n)
\\ && \oplus L(2\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)
\oplus\dots
\nonumber \\
& & \oplus L((2l-1)\omega_1-\frac{1}{2}\omega_n)\oplus
L(2l\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n),
\end{eqnarray*}
\item
In the odd case $k=2l+1$ ($2l+2$ terms on the right-hand side):
\begin{eqnarray*}
L(\omega_{n-1}-\frac{3}{2}\omega_n)\otimes S^k(\mathbb{C}^{2n}) & \simeq &
L(-\frac{1}{2}\omega_n)\oplus L(\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)
\oplus\dots
\nonumber \\
& & \oplus L(2l\omega_1-\frac{1}{2}\omega_n)\oplus
L((2l+1)\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n).
\end{eqnarray*}
\end{enumerate}
\end{cor}
Reformulated in the language of differentials operators, this Corollary leads to Theorem \ref{SymplecticFischer}. We also introduce the notation $\mathcal{C} = L(-\frac{1}{2}\omega_n)\oplus L(\omega_{n-1}-\frac{3}{2}\omega_n)$.
\section{Symplectic structure and its Lie algebra $\mathfrak{sl}_2$}
\label{sl2}
Let $(\mathbb{R}^{2n},\omega)$ be the symplectic vector space with coordinates $x_1,\dots ,x_{2n}$,
coordinate vector fields $\partial_1,\dots ,\partial_{2n}$ and
symplectic frame $e_1,f_1,\dots ,e_n,f_n$, i.e. $\omega(e_i,e_j)=0$, $\omega(f_i,f_j)=0$
and $\omega(e_i,f_j)=\delta_{ij}$ for all $i,j=1,\dots ,n$. It follows from the action of
$sp(2n)$ on these vectors that
\begin{eqnarray}
& & X_s:=\sum_{j=1}^{n} (x_{2j-1} f_{j} + x_{2j}e_{j}),
\nonumber \\
& & D_s:=\sum_{j=1}^{n} (\partial_{x_{2j-1}} e_{j} - \partial_{x_{2j}}f_{j}),
\nonumber \\
& & \mathbb{E}:=\sum_{j=1}^{2n} x_{j}\partial_{x_{j}}
\end{eqnarray}
are invariant and so will be used as linear maps intertwining the $sp(2n)$ action on
the space $\mathcal{P} \otimes \mathcal{C}$ of symplectic spinor valued polynomials on $\mathbb{R}^{2n}$, with
$\mathcal{P} = Pol(\mathbb{R}^{2n})$. The space of homogeneous polynomials of degree $k$ will be denoted by $\mathcal{P}_{k}$.
It is easy to verify that they fulfill $\mathfrak{sl}_2$ commutation relations:
\begin{eqnarray}
& & [\mathbb{E} +n,D_s]=-D_s,
\nonumber \\
\label{slRels}
& & [\mathbb{E} +n,X_s]=X_s,\\
& & [D_s,X_s]=\mathbb{E} +n.\nonumber
\end{eqnarray}
The action of $\mathfrak{sl}_2 \times sp(2n)$ will generate the multiplicity free decomposition
of the representation of interest.
Further we introduce the operator
\begin{equation}
\Gamma_{s} = X_{s} D_{s} - \frac{1}{2} \mathbb{E}(2n-1 + \mathbb{E}),
\label{gammaop}
\end{equation}
which is the Casimir operator in $\mathfrak{sl}_2$. Using the formulas (\ref{slRels}) it is easy to check that $\Gamma_{s}$ commutes with both $X_{s}$ and $D_{s}$.
\section{Fischer decomposition and homomorphisms of $sp(2n,\mathbb{R})$-modules appearing
in the decomposition of polynomials valued in the Segal-Shale-Weil representation}
Before introducing the scheme in full generality, we start with a few explicit remarks
concerning homogeneity zero and one parts in the decomposition.
The tensor product $L(-\frac{1}{2}\omega_n)\otimes \mathbb{C}^{2n}$
(analogously, one can consider $L(\omega_{n-1}-\frac{3}{2}\omega_n)\otimes \mathbb{C}^{2n}$)
decomposes as the direct sum $V_{1} \oplus V_{2}$ of two invariant subspaces, given by
\begin{eqnarray}
& & V_1:=\{\sum_{i=1}^{n}e_i s\otimes e_i +\sum_{i=1}^{n}f_i s\otimes e_i|s\in L(-\frac{1}{2}\omega_n)\},
\nonumber \\
\nonumber
& & V_2:=\{\sum_{i=1}^{n}s_i\otimes e_i+\sum_{j=1}^{n}s_j\otimes f_j;s_i,s_j\in L(-\frac{1}{2}\omega_n)|\sum_{i=1}^{n}e_is_i
+\sum_{j=1}^{n}f_js_j =0\}.
\end{eqnarray}
The map $i:L(\omega_{n-1}-\frac{3}{2}\omega_n)\to L(-\frac{1}{2}\omega_n)\otimes \mathbb{C}^{2n}$
(resp. $L(-\frac{1}{2}\omega_n)\to L(\omega_{n-1}-\frac{3}{2}\omega_n)\otimes \mathbb{C}^{2n}$) is
injective and onto $V_1$. The reason is that injectivity $i(s)=0$ is equivalent to $e_is=0$
resp. $f_is=0$ for all
$i\in\{1,\dots ,n\}$, and so $(e_if_{j}-f_{j}e_i)s=0$. The symplectic Clifford algebra
relation $e_if_{j}-f_{j}e_i=\delta_{ij}$ implies $s=0$ and the result follows. In other words,
the action of $X_s$ induces an isomorphism between two irreducible submodules in homogeneity
zero and one.
Another remark is an application of tools in representation theory, in particular
of infinitesimal character.
The sum of fundamental weights (or half of the sum of positive roots) for $sp(2n)$
is $\delta =(n,n-1,\dots ,2,1)$. The highest weights of irreducible simple $sp(n)$-modules,
coming from the decomposition of the tensor product, were determined
for each homogeneity $k\in\mathbb{N}$ in Corollary \ref{decomposition}. The
multiplication by $X_s$ gives an intertwining map between neighboring columns, say the
$k$-th and $(k+1)$-th. Let us determine possible target modules when restricting the
action of $X_s$ to a given simple irreducible $sp(2n)$-module $L(a\omega_1-\frac{1}{2}\omega_{n})$
with highest weight $a\omega_1-\frac{1}{2}\omega_{n}$ for some $a\leq k$ (the case of
$L(b\omega_1+\omega_{n-1}-\frac{3}{2}\omega_{n})$ being analogous.) The comparison of
infinitesimal characters of the collection of weights
$\{\mu_a=a\omega_1-\frac{1}{2}\omega_{n},\nu_b=b\omega_1+\omega_{n-1}-\frac{3}{2}\omega_{n}\}$
($a,b\in\mathbb{N}$) yields $||\mu_a+\delta||^2=||\nu_b+\delta||^2$ if and only if either
\begin{enumerate}
\item
$2a+n-\frac{1}{2}=2b+n-\frac{1}{2}$, which implies $a=b$, or
\item
$2a+n-\frac{1}{2}=-(2b+n-\frac{1}{2})$, i.e. $a+b=-n+\frac{1}{2}$
and there is no solution in this case.
\end{enumerate}
It remains to prove that the image of $X_s$, when restricted to an irreducible
simple module in the $k$-th column, is nonzero (or, as follows from irreducibility,
is the irreducible simple module in the $(k+1)$-th column with the same infinitesimal
character.)
To complete this line of reasoning, we employ the Lie algebra $\mathfrak{sl}_{2}$ from
Section \ref{sl2}. To illustrate it explicitly, we start in homogeneity
zero with only one simple module $L(-\frac{1}{2}\omega_n)$ and
assume that it is $Ker(X_s)$. Because it is in $Ker(D_s)$, it is in the
kernel of the commutator $[D_s,X_s]=\mathbb{E} +n$. However, $\mathbb{E} +n$ acts in homogeneity
zero by $n$, which is the required contradiction and so $X_s$ acts as an
isomorphism $L(-\frac{1}{2}\omega_n)\to L(\omega_{n-1}-\frac{3}{2}\omega_n)$.
Let us now consider the action of $X_s$ on $L(\omega_{n-1}-\frac{3}{2}\omega_n)$
sitting in the homogeneity one part and assume it acts trivially. Then due
to the previous isomorphism, this kernel is $Ker(X_s^2)$ when $X_s^2$ is acting on
$L(-\frac{1}{2}\omega_n)$ in homogeneity zero. As before, the commutator
$[X_s^2,D_s]$ acts by zero. Because it is equal to $-X_s (\mathbb{E}+n)-(\mathbb{E}+n) X_s$, it acts
on homogeneity zero elements by $-(2n+1)X_s$ and due to the fact that $X_s$
is an isomorphism, it is nonzero and so yields the contradiction. In conclusion,
$X_s:L(\omega_{n-1}-\frac{3}{2}\omega_n)\to L(-\frac{1}{2}\omega_n)$ acting
between homogeneity one and two is an isomorphism. Clearly, one can iterate
the procedure further using the subsequent Lemmas \ref{ActSympl} and \ref{calcSympDirac} on $sp(2n)$-invariant
intertwining operators acting on the direct sum of simple highest weight
$sp(2n)$-modules. An analogous induction procedure can be used to prove the
isomorphic action of
$D_s$.
In what follows, we formulate the previous qualitative statements more quantitatively.
Denote by ${\fam2 M}^+_l$ resp. ${\fam2 M}^-_l$ the (irreducible) simple $sp(2n)$-module
with highest weight $L(l\omega_1-\frac{1}{2}\omega_n)$ resp. $L(l\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n))$,
and call them {\bf symplectic monogenics} of degree $l$ (or $l$-homogeneous symplectic monogenics). We put ${\fam2 M}_l:={\fam2 M}^+_l\oplus {\fam2 M}^-_l$ and we also have
\[
\mathcal{M}_{l} = \ker{D_{s}} \cap (\mathcal{P}_{l} \otimes \mathcal{C}).
\]
We then obtain two auxiliary lemmas.
\begin{lem}
Suppose $M_{\ell} \in \mathcal{M}_{\ell}$ is a symplectic monogenic of degree $\ell$. Then
\[
D_{s} (X_{s}^{k}M_{\ell}) = \frac{1}{2}k(2n+2l+k-1)X_{s}^{k-1}M_{\ell}.
\]
\label{ActSympl}
\end{lem}
\begin{proof}
By induction.
\end{proof}
\begin{lem}
Suppose $M_{\ell} \in \mathcal{M}_{\ell}$ is a symplectic monogenic of degree $\ell$. Then
\[
D_{s}^{j} (X_{s}^{k}M_{\ell}) = c_{j,k,\ell} X_{s}^{k-j}M_{\ell}
\]
with
\[
c_{j,k,l} = \left\{ \begin{array}{ll}\dfrac{1}{2^{j}} \dfrac{k!}{(k-j)!} \dfrac{(2n+2l+k-1)!}{(2n+2l+k-j-1)!}&j \leq k\\0&j>k. \end{array}\right.
\]
\label{calcSympDirac}
\end{lem}
\begin{proof}
The lemma follows from $j$ iterations of Lemma \ref{ActSympl}.
\end{proof}
The previous considerations can be summarized in the symplectic analog of the classical theorem
on separation of variables in the orthogonal case, see e.g. \cite{Orsted} and
the references therein.
\begin{thm}
\label{SymplecticFischer}
The space $\mathcal{P}\otimes \mathcal{C}$ decomposes under the action of $\mathfrak{sl}_{2}$ into the direct sum of simple highest weight $sp(2n)$-modules
\begin{eqnarray}
\bigoplus_{l=0}^\infty\bigoplus_{j=0}^\infty X_s^j{\fam2 M}_l,
\end{eqnarray}
where we used the notation ${\fam2 M}_l:={\fam2 M}^+_l\oplus {\fam2 M}^-_l$. The decomposition takes
the form of an infinite triangle \[
\xymatrix@=11pt{\mathcal{P}_0 \otimes \mathcal{C} & \mathcal{P}_1 \otimes \mathcal{C} & \mathcal{P}_2 \otimes \mathcal{C} & \mathcal{P}_3 \otimes \mathcal{C} \ar@{=}[d] & \mathcal{P}_4 \otimes \mathcal{C} & \mathcal{P}_5 \otimes \mathcal{C} &\ldots \\
\mathcal{M}_0 \ar[r] & X_s \mathcal{M}_0 \ar[r] & X_s^2 \mathcal{M}_0 \ar[r] & X_s^3 \mathcal{M}_0 \ar @{} [d] |{\oplus}
\ar[r] & X_s^4 \mathcal{M}_0 \ar[r] & X_s^5 \mathcal{M}_0 &\ldots\\
&\mathcal{M}_1 \ar[r] & X_s \mathcal{M}_1 \ar[r] & X_s^2 \mathcal{M}_1 \ar @{} [d] |{\oplus}
\ar[r] & X_s^3 \mathcal{M}_1 \ar[r] & X_s^4 \mathcal{M}_1 &\ldots\\
&&\mathcal{M}_2 \ar[r] & X_s \mathcal{M}_2 \ar @{} [d] |{\oplus}
\ar[r] & X_s^2 \mathcal{M}_2 \ar[r] & X_s^3 \mathcal{M}_2 &\ldots\\
&&&\mathcal{M}_3 \ar[r] & X_s \mathcal{M}_3 \ar[r] & X_s^2 \mathcal{M}_3 &\ldots\\
&&&&\mathcal{M}_4 \ar[r] & X_s \mathcal{M}_4 &\ldots\\
&&&&&\mathcal{M}_5&\ldots
}
\]
where all summands are simple highest weight $sp(2n)$-modules. The $k$-th column gives the decomposition of homogeneous polynomials of degree $k$ taking values in $\mathcal{C} = L(-\frac{1}{2}\omega_n)\oplus L(\omega_{n-1}-\frac{3}{2}\omega_n)$. The $l$-th row forms a highest weight
$\mathfrak{sl}_{2}$-module $\oplus_{j=0}^\infty X_s^j{\fam2 M}_l$ generated by the space of symplectic monogenics ${\fam2 M}_l$.
\end{thm}
One immediate Corollary is the structure of polynomial solutions of the symplectic Dirac operator
on $\mathbb{R}^{2n}$. The statement is given for both symplectic spin modules
$L(-\frac{1}{2}\omega_n)$ and $L(\omega_{n-1}-\frac{3}{2}\omega_n)$ separately.
\begin{cor}
\label{symkernel}
The kernel of (half of) the symplectic Dirac operator $D_s$ acting
on $L(-\frac{1}{2}\omega_n)$-valued polynomials is
\begin{eqnarray*}
Ker^+(D_s)\simeq \bigoplus_{l\in\mathbb{N}}
\left(L(2l\omega_1-\frac{1}{2}\omega_n)\oplus L((2l+1)\omega_1-\frac{1}{2}\omega_n)\right).
\end{eqnarray*}
The kernel of (half of) the symplectic Dirac operator $D_s$ acting
on $L(\omega_{n-1}-\frac{3}{2}\omega_n)$-valued polynomials is
\begin{eqnarray*}
Ker^-(D_s)\simeq \bigoplus_{l\in\mathbb{N}} \left(
L(2l\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)\oplus L((2l+1)\omega_1+\omega_{n-1}-\frac{3}{2}\omega_n)\right).
\end{eqnarray*}
\end{cor}
Every homogeneous polynomial of degree $k$, taking values in $\mathcal{C}$, can now be decomposed in monogenic components as follows.
\begin{thm}
Let $p\in \mathcal{P}_{k} \otimes \mathcal{C}$. Then there exists a unique representation of $p$ as
\begin{eqnarray}
p=\sum_{i=0}^k p_i,
\end{eqnarray}
where $p_i=X_s^{k-i}m_i$ and $m_i\in {\fam2 M}_i$.
\label{thmReprHomK}
\end{thm}
We now proceed to construct projection operators that allow to explicitly compute the representation given in Theorem \ref{thmReprHomK}. They are given in the following theorem.
\begin{thm}
The operators
\begin{equation}
\pi^{k}_{i} = \sum_{j=0}^{k-i} a^{i,k}_{j} X^{i+j}_{s} D^{i+j}_{s}
\label{projversion1}
\end{equation}
with
\[
a^{i,k}_{j} = (-1)^{j} (2n+2k-2i-1) \frac{2^{i+j}}{i! j!} \frac{(2n+2k-2i-j-2)!}{(2n+2k-i-1)}
\]
and $i=0, \ldots, k$ satisfy
\[
\pi^{k}_{i} (X^{j}_{s} \mathcal{M}_{k-j}) = \delta_{ij} X^{i}_{s} \mathcal{M}_{k-i}.
\]
\end{thm}
\begin{proof}
Using lemma \ref{calcSympDirac} it is easy to see that $\pi^{k}_{i} (X^{j}_{s} \mathcal{M}_{k-j}) =0$ for all $j<i$. The coefficients $a^{i,k}_{j}$, for fixed $i$ and $k$, can now be determined iteratively. First of all, expressing $\pi^{k}_{i} (X^{i}_{s} \mathcal{M}_{k-i}) = X^{i}_{s} \mathcal{M}_{k-i}$ yields
\[
a^{i,k}_{0} = \frac{1}{c_{i,i,k-i}} = \frac{2^{i}}{i!} \frac{(2n+2k-2i-1)!}{(2n+2k-i-1)!}.
\]
Similarly, expressing $\pi^{k}_{i} (X^{i+1}_{s} \mathcal{M}_{k-i-1}) =0$ then yields
\[
a^{i,k}_{1} = - \frac{c_{i,i+1,k-i-1}}{c_{i+1,i+1,k-i-1}} a^{i,k}_{0} = - \frac{1}{n+k-i-1} a^{i,k}_{0}.
\]
Thus continuing we arrive at the hypothesis
\[
a^{i,k}_{j} = (-1)^{j} \frac{2^{j}}{j!} \frac{(2n+2k-2i-j-2)!}{(2n+2k-2i-2)!} a^{i,k}_{0}
\]
which can be proven using induction. Indeed, suppose that the statement holds for $a_j^{i,k}$, $j \leq l$, then we prove that it also holds for $a_{l+1}^{i,k}$. This last coefficient has to satisfy
\[
\sum_{j=0}^{l+1} a_j^{i,k} \, c_{i+j,i+l+1,k-i-l-1}=0.
\]
Substituting the known expressions we obtain
\begin{align*}
a_{l+1}^{i,k} &= -\sum_{j=0}^l a_j^{i,k} \frac{c_{i+j,i+l+1,k-i-l-1}}{c_{i+l+1,i+l+1,k-i-l-1}}\\
&= -\sum_{j=0}^l a_j^{i,k} \frac{2^{l+1-j}}{(l+1-j)!} \frac{(\alpha-2l-1)!}{(\alpha-l-j)}\\
&= -\frac{2^{l+1}}{(l+1)!} \frac{(\alpha-2l-1)!}{\alpha!} a^{i,k}_{0} \sum_{j=0}^{l} (-1)^{j} \binom{l+1}{j} \frac{(\alpha-j)!}{(\alpha-l-j)!}
\end{align*}
where we have put $\alpha = 2n+2k-2i-2$. The proof is now complete by remarking that
\[
\sum_{j=0}^{l+1} (-1)^{j} \binom{l+1}{j} \frac{(\alpha-j)!}{(\alpha-l-j)!}=0.
\]
This can either be obtained directly (see e.g. Lemma 5 in \cite{HDBThesis}) or as a consequence of Gauss's hypergeometric theorem, expressing $\,_2F_1(a,b;c;1)$ in terms of a product of Gamma functions.
\end{proof}
Note that there exists another way of computing the projection operators
on irreducible summands, namely using the Casimir operator of $\mathfrak{sl}_{2}$. First observe that
\[
\Gamma_{s} \mathcal{M}_{k} = -\frac{1}{2} k (2n-1+k) \mathcal{M}_{k}.
\]
It is then clear that the operators
\[
\mathbb{P}_{i}^{k} = \prod_{j=0, j \neq i}^{k} \dfrac{2\Gamma_{s} + j (2n-1+j)}{j(2n-1+j)-i(2n-1+i)}, \quad i = 0, \ldots, k
\]
defined on the space $\mathcal{P}_{k} \otimes \mathcal{C}$ satisfy
\[
\mathbb{P}_{i}^{k} (X_{s}^{k-j} \mathcal{M}_{j}) = \delta_{i j}X_{s}^{k-i} \mathcal{M}_{i}.
\]
\section{Open questions and unresolved problems}
In \cite{KAD}, the symplectic Dirac operator $D_s$ on $\mathbb{R}^{2n}$ was studied as an
$sp(2n+2)$-invariant differential operator in the context of contact parabolic geometry.
As a consequence, the kernel of $D_s$ has the structure of an $sp(2n+2)$-module. Clearly,
$Ker(D_s)$ is, as a vector space, isomorphic to $Pol(\mathbb{R}^{2n+2})$
(see Corollary \ref{symkernel}) and we leave the question of its representation theoretic
content open.
In \cite{Orsted}, the authors studied the specific deformation of Howe duality and Fischer
decomposition for the Dirac operator acting on spinor valued polynomials, coming from the Dunkl
deformation of the Dirac operator. It is an interesting question to develop the Dunkl version
of the symplectic Dirac operator in the context of symplectic reflection algebras (see \cite{MR1881922}).
Another interesting question is whether the reproducing kernel of the space of sympletic monogenics $\mathcal{M}_{l}$ can again be expressed in terms of Gegenbauer polynomials, as in the orthogonal and Dunkl case.
|
\section{Introduction}
Corresponding to any sequence of real numbers $\ds \{\gamma_k\}_{k=0}^{\infty}$ one can define a linear operator $T$ on $\mathbb{R}[x]$ by declaring $T[x^n] = \gamma_n x^n$ for all $n$. If the linear operator $T$ has the property that $T[p]$ has only real zeros whenever $p$ has only real zeros, then $\{\gamma_k\}_{k=0}^{\infty}$ is called a \textit{multiplier sequence}. Examples of such sequences were demonstrated in the late 1800's by Jensen and Laguerre and, in the early 1900's, all such sequences were completely characterized by P\'olya and Schur.
\begin{thm}\label{PS}
\emph{(P\'olya-Schur \cite{PS})} Let $\ds \{\gamma_k\}_{k=0}^{\infty}$ be a sequence of nonnegative real numbers. The following are equivalent.
\begin{enumerate}
\item{$\ds \{\gamma_k\}_{k=0}^{\infty}$ is a multiplier sequence}
\item{For each $n$, the polynomial $\displaystyle T[(1+x)^n]:= \sum_{k=0}^{n}\binom{n}{k} \gamma_k x^k$ has only real zeros}
\item{The series $\displaystyle \varphi(z) = \sum_{k=0}^{\infty} \frac{\gamma_k}{k!}z^k$ converges in the whole plane and either $\varphi(z)$ or $\varphi(-z)$ is of the form $\displaystyle c e^{\sigma z}z^m \prod_{k=1}^{\omega}\left(1+ w_k{z}\right)$ where $c\in\mathbb{R}$, $\sigma\geq 0$, $m$ is a nonnegative integer, $0\leq \omega\leq \infty$, $z_k>0$, and $\displaystyle \sum_{k=1}^{\omega} {w^k}<\infty$}.
\end{enumerate}
\end{thm}
Similarly, corresponding to any sequence of real numbers $\ds \{\gamma_k\}_{k=0}^{\infty}$ one can define a linear operator $T_H$ on $\mathbb{R}[x]$ by declaring $T_H[H_n(x)] = \gamma_n H_n(x)$ for all $n$, where $H_n(x)$ denotes the $n^{th}$ Hermite polynomial $\displaystyle H_n(x) = (-1)^n e^{x^2} D^n e^{-x^2}$. If the linear operator $T_H$ has the property that $T_H[p]$ has only real zeros whenever $p$ has only real zeros, then $\{\gamma_k\}_{k=0}^{\infty}$ is called an \textit{Hermite multiplier sequence}. Examples of such sequences were demonstrated in the mid 1900's by Tur\'an \cite{Turan}, and also in 2001 by Bleecker and Csordas \cite{BC}. In 2007, all such sequences were completely characterized by Piotrowski.
\begin{thm}\label{PS}
\emph{(Piotrowski, Theorem 152 in \cite{andrzej})} Let $\ds \{\gamma_k\}_{k=0}^{\infty}$ be a sequence of nonnegative real numbers. The following are equivalent.
\begin{enumerate}
\item{$\ds \{\gamma_k\}_{k=0}^{\infty}$ is a non-trivial Hermite multiplier sequence}
\item{$\ds \{\gamma_k\}_{k=0}^{\infty}$ is an nondecreasing multiplier sequence}
\item{The series $\displaystyle \varphi(z) = \sum_{k=0}^{\infty} \frac{\gamma_k}{k!}z^k$ converges in the whole plane and either $\varphi(z)$ or $\varphi(-z)$ is of the form $\displaystyle c e^{\sigma z}z^m \prod_{k=1}^{\omega}\left(1+\frac{z}{z_k}\right)$ where $c\in\mathbb{R}$, $\sigma\geq 1$, $m$ is a nonnegative integer, $0\leq \omega\leq \infty$, $z_k>0$, and $\displaystyle \sum_{k=1}^{\omega}\frac{1}{z^k}<\infty$}.
\end{enumerate}
\end{thm}
In this paper we investigate a related problem, where we use the generalized Laguerre polynomials in place of the Hermite polynomials. To any sequence of real numbers $\ds \{\gamma_k\}_{k=0}^{\infty}$, one can define a linear operator $T_L$ on $\mathbb{R}[x]$ by declaring $T_L\left[L_n^{(\alpha)}(x)\right] = \gamma_n L_n^{(\alpha)}(x)$ for all $n$, where $L_n^{(\alpha)}(x)$ denotes the $n^{th}$ Laguerre polynomial $ \displaystyle L_n^{(\alpha)}(x) = \sum_{k=0}^{n}\binom{n+\alpha}{n-k}\frac{(-x)^k}{k!}$ and $\alpha>-1$. If the linear operator $T_L$ has the property that $T_L[p]$ has only real zeros whenever $p$ has only real zeros, then $\{\gamma_k\}_{k=0}^{\infty}$ is called an \textit{$L^{(\alpha)}$-multiplier sequence}. We opted not use the terminology ``Laguerre multiplier sequence," as this phrase has been used by other authors with a different meaning (See, for example, \cite{CCsurvey}).
In a similar way, one can define $Q$-multiplier sequences, where $\displaystyle Q=\{q_k\}_{k=0}^{\infty}$ is any simple polynomial set (i.e., $\deg(q_k)=k$ for each $k$). Remarkably, every $Q$-multiplier sequence must be a (classical) multiplier sequence, regardless of the choice of $Q$. In particular, the following result guarantees that every $L^{(\alpha)}$-multiplier sequence must also be a multiplier sequence.
\begin{thm}\label{thmqmsms}\emph{(Piotrowski, Theorem 158 in \cite{andrzej})}
Let $Q = \{q_k\}_{k=0}^{\infty}$ be a simple set of polynomials. If the sequence $\{ \gamma_k\}_{k=0}^{\infty}$ is a $Q$-multiplier sequence, then the sequence $\{ \gamma_k\}_{k=0}^{\infty}$ is a multiplier sequence.
\end{thm}
In general, we will say that an operator $T$ \textit{preserves reality of zeros} if it has the property that $T[p]$ has only real zeros whenever $p$ has only real zeros. Thus, a sequence is a multiplier sequence if its corresponding operator preserves reality of zeros. Very recently, Borcea and Br\"and\'en gave a complete characterization of stability preserving operators. A special case of their result is a characterization of linear operators which preserve reality of zeros.
\begin{thm}\label{BBthm}\emph{(Borcea-Br\"and\'en, Theorem 5 in \cite{BB})}
A linear operator $T:{\mathbb R}[x]\to {\mathbb R}[x]$ preserves reality of zeros if and only if either
\begin{enumerate}
\item{$T$ has range of dimension at most two and is of the form $T[f] = \alpha(f)P+ \beta(f)Q$ where $\alpha$ and $\beta$ are linear functionals on ${\mathbb R}[x]$ and $P$ and $Q$ are polynomials with only real interlacing zeros.}
\item{$\displaystyle T[\exp(-xw)] = \sum_{n=0}^{\infty} \frac{(-w)^n T[x^n]}{n!}\in \overline{A}$, or }
\item{$\displaystyle T[\exp(xw)] = \sum_{n=0}^{\infty} \frac{w^n T[x^n]}{n!}\in \overline{A}$,}
\end{enumerate}
where $\overline{A}$ denotes the set of entire functions in 2 variables that are limits, uniformly on compact subsets, of polynomials in the set
$$
A = \{f\in{\mathbb R}[x,w] \,\big|\, f(x,w) \neq 0 \emph{ whenever } \emph{Im }x>0 \emph{ and } \emph{Im }w >0 \}.
$$
\end{thm}
With this characterization at hand, the crux of our problem is to find necessary and sufficient conditions on a sequence of real numbers under which the corresponding operator $T_L$ satisfies one of the conditions (1)-(3) above. This task is quite difficult for a generic sequence, and as such we have not yet arrived at a complete characterization of
$L^{(\alpha)}$-multiplier sequences.
\newline \indent Finally, we note that throughout the paper we adopt the following convention: to avoid trivialities, we consider the identically zero function $f \equiv 0$ to have only real zeros, although this is clearly not the case.
\section{Trivial, Geometric, and Linear Sequences}
It is well known that the generalized Laguerre polynomials form an orthogonal set over the positive real axis with respect to the weight function $x^{\alpha}e^{-x}$ (recall that in this paper we are only considering $\alpha>-1$). Orthogonal polynomials have only simple real zeros. Furthermore, the zeros of consecutive polynomials in the sequence are interlacing. As a consequence of this, it is easy to verify that any linear combination $aL_n^{(\alpha)}(x) + bL_{n+1}^{(\alpha)}(x)$ has only real zeros (one only needs to consider the sign of the individual terms and count zeros with the aid of the Mean Value Theorem). We thus have the following fact:
\begin{prop}\label{trivial} Given $\gamma_{n}, \gamma_{n+1} \in \mathbb{R}$, any sequence of the form $( 0,0,\ldots, 0,0, \gamma_{n}, \gamma_{n+1}, 0,0,\ldots )$ is an $L^{(\alpha)}$-multiplier sequence.
\end{prop}
We will call sequences of the above form {\it trivial $L^{(\alpha)}$-multiplier sequences}. Unless stated otherwise, in what follows we only consider nontrivial $L^{(\alpha)}$-multiplier sequences.
\subsection{Geometric $L^{(\alpha)}$-Multiplier Sequences}
We now consider the geometric sequences $\{ r^k \}_{k=0}^{\infty}$, $r \in \mathbb{R}$. These sequences are (classical) multiplier sequences for all nonzero $r$ and are Hermite multiplier sequences if and only if $|r|\geq 1$. In contrast with these results, the only geometric sequence which is an $L^{(\alpha)}$-multiplier sequence is the constant sequence $\{1\}_{k=0}^{\infty}$.
\begin{prop} \label{r^k} The sequence $\displaystyle \{ r^k \}_{k=0}^{\infty}$ is an $L^{(\alpha)}$-multiplier sequence if and only if $r=1$.
\end{prop}
\begin{proof} Consider the polynomial $p(x)=(x+b)^2$ for $b \in {\mathbb R}$. We can write $p(x)$ as
\[
p(x)=2 L_2^{(\alpha)}(x)-2(\alpha+2+b) L_1^{(\alpha)}(x)+(\alpha+b)^2+3\alpha+2b+2.
\]
Applying the sequence $\{ r^k\}_{k=0}^{\infty}$ and then expanding in terms of the standard basis we obtain the polynomial
\begin{eqnarray*}
\bar p(x)&=&r^2 x^2+(2(\alpha+2+b)r-(2\alpha+4)r^2)x\\
&+&2+\alpha^2+2b+b^2+\alpha(3+2b)-2(2+\alpha+b)(1+\alpha)r+r^2(2+3\alpha+\alpha^2),
\end{eqnarray*}
with discriminant
\[
\Delta=-4 r^2 (r-1)((2+\alpha)(1-r)+2b).
\]
From this representation we immediately see that $(i)$ if $r=1$ the discriminant is equal to zero and $(ii)$ large positive values (if $r >1$) or large negative values (if $r<1$) of $b$ result in a negative discriminant. This establishes the claim.
\end{proof}
\subsection{Linear $L^{(\alpha)}$-Multiplier Sequences}
In \cite{andrzej} it is shown that for the simple Laguerre polynomials ($\alpha=0$) the sequence $\{a+k \}_{k=0}^{\infty}$ is not an $L^{(0)}$-multiplier sequence for $a >1 $ and $a < 0$ but it is an $L^{(0)}$-multiplier sequence for $a=1$ and $a=0$. The question whether $\{a+k\}_{k=0}^{\infty}$ is an $L^{(0)}$-multiplier sequence for $0< a < 1$ is left open. In this section we answer this question and completely characterize linear $L^{(\alpha)}$-multiplier sequences.
\begin{lem} \label{a_alpha_1} $\{k+a\}_{k=0}^{\infty}$ is not an $L^{(\alpha)}$-multiplier sequence for any $\alpha$ if $a<0$.
\end{lem}
\begin{proof} The set $\left\{ L_k^{(\alpha)}(x) \right\}_{k=0}^{\infty}$ is a simple set of polynomials. Thus, by Theorem \ref{thmqmsms}, any sequence of real numbers $\{ \gamma_k \}_{k=0}^{\infty}$ that is an $L^{(\alpha)}$-multiplier sequence is a (classical) multiplier sequence. Since $\{ a+k\}_{k=0}^{\infty}$ is not a (classical) multiplier sequence for $a<0$ the result follows.
\end{proof}
\begin{lem} \label{a_alpha_2} $\{k+a\}_{k=0}^{\infty}$ is not an $L^{(\alpha)}$-multiplier sequence if $a > \alpha+1$.
\end{lem}
\begin{proof} We recall that the polynomials $L_n^{(\alpha)}(x)$ satisfy the following ordinary differential equation:
\begin{equation} \label{LaguerreODE}
nL_n^{(\alpha)}(x)=(x-\alpha-1)L_n^{(\alpha)'}(x)-xL_n^{(\alpha)''}(x).
\end{equation}
It follows that
\[
(a+k)L_k^{(\alpha)}(x)=aL_k^{(\alpha)}(x)+(x-\alpha-1)L_k^{(\alpha)'}(x)-xL_k^{(\alpha)''}(x).
\]
Thus the action of the sequence $\{ a+k\}_{k=0}^{\infty}$ on a polynomial is represented by the operator
\begin{equation}
\label{T}
T:=a+(x-\alpha-1)D-xD^2.
\end{equation}
Consider now the polynomial $(x+n)^n$, which clearly has only real zeros. We have
\begin{eqnarray*}
T[(x+n)^n]&=&a(x+n)^n+(x-\alpha-1)n(x+n)^{n-1}-xn(n-1)(x+n)^{n-2}\\
&=&(x+n)^{n-2}[ a(x+n)^2+(x-(\alpha+1))n(x+n)-x(n^2-n)]\\
&=&(x+n)^{n-2}[x^2(a+n)+x(2an-n\alpha)+an^2-n^2(\alpha+1)].
\end{eqnarray*}
Calculating the discriminant of the polynomial in the square brackets we get
\begin{eqnarray*}
\Delta(n)&=&n^2[4a^2-4a\alpha+\alpha^2-4(a+n)(a-(\alpha+1))]\\
&=&n^2[\alpha^2+4a-4n(a-(\alpha+1))].
\end{eqnarray*}
It follows that if $a>(\alpha+1)$ then $\Delta(n) < 0$ for $n$ sufficiently large. Therefore $T[(x+n)^n]$ will have non-real zeros for large enough $n$. This completes the proof.
\end{proof}
\begin{lem} \label{a_alpha_3}
If $0\leq a\leq \alpha+1$, then $\{k+a\}_{k=0}^{\infty}$ is an $L^{(\alpha)}$-multiplier sequence. In particular, if $0 \leq a \leq 1$, then $\{k+a\}_{k=0}^{\infty}$ is an $L$-multiplier sequence.
\end{lem}
\begin{proof}
Consider the differential operator representation of the sequence at hand.
$$
T = (x-\alpha-1)D - x D^2 + a
$$
By the result of Borcea and Br\"and\'en (Theorem $\ref{BBthm}$) this operator preserves reality of zeros provided the polynomial
$$
a+(z-\alpha - 1)(-w) - z (-w)^2 = a-w(w+1)z+ w(\alpha+1)
$$
does not vanish whenever Im $z>0$ and Im $w>0$. Setting the above equation equal to zero and solving for $z$ we obtain
$$
z = \frac{w(\alpha+1)+a}{w(w+1)}= (\alpha+1)\frac{w+w_0}{w(w+1)}, \qquad \left(w_0 = \frac{a}{\alpha+1}\right).
$$
Suppose Im $w>0$ and that $0\leq a \leq \alpha+1$. Then $w_0<1$, and we have
$$
0< \arg(w)\leq \arg(w+w_0)\leq \arg(w+1)< \pi,
$$
from which we obtain
$$
-\pi < -\arg(w+1) \leq \arg(w+w_0) - \arg(w) - \arg(w+1) \leq -\arg(w)< 0.
$$
Thus Im $z<0$ whenever Im $w>0$ and $0 \leq a \leq \alpha+1$. The proof is complete.
\end{proof}
Combining lemmas $\ref{a_alpha_1}$, $\ref{a_alpha_2}$, and $\ref{a_alpha_3}$ we obtain the following theorem.
\begin{thm}\label{linchar}
$\{k+a\}_{k=0}^{\infty}$ is an $L^{(\alpha)}$-multiplier sequence if and only if $0\leq a\leq \alpha+1$.
\end{thm}
\section{The sequence $\{k(k-1)(k-2) \cdots (k-(n-1)) \}_{k=0}^{\infty}$} \label{arbdegree}
The purpose of this section is to prove that the above sequence is an $L^{(\alpha)}$-multiplier sequence for $\alpha >-1$ and $n \geq 1$. To establish this fact we need several auxiliary results. We begin with the following lemma.
\begin{lem} \label{commutator} Let $\delta$ be the operator defined by $\delta:=(x-(\alpha+1))D-xD^2$. Then for $k \geq 0$ we have
\[
[\delta, D^k]:=\delta D^k-D^k\delta=-k(1-D)D^k.
\]
\end{lem}
\begin{proof} If $k=0$ the result is trivial. Supposing the result holds for all integers up to $k$ we calculate
\begin{eqnarray*}
[\delta,D^{k+1}]&=&\delta D^{k+1}-D^{k+1} \delta=(\delta D^k)D-D(D^k \delta)\\
&=&(\delta D^k)D-D(\delta D^k+k(1-D)D^k)\\
&=&\delta D^{k+1}-(\delta D+(1-D)D)D^k-k(1-D)D^{k+1}\\
&=&-(k+1)(1-D)D^{k+1},
\end{eqnarray*}
establishing the desired equality.
\end{proof}
\begin{prop} \label{BIGlemma} Let $\delta$ be the operator defined by $\delta:=(x-(\alpha+1))D-xD^2$ and let $L_n^{(\alpha)}(x)$ be the $n^{th}$ generalized Laguerre polynomial. If
\begin{equation}\label{deltas}
\delta (\delta-1)(\delta-2)\cdots(\delta-(n-1))=\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x) D^k
\end{equation}
then
\begin{equation}\label{LaguerreRep}
\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x)z^k=n! (-1)^{n} z^n L_n^{(\alpha)}(x-xz).
\end{equation}
\end{prop}
\begin{proof} We proceed by induction on $n$. If $n=1$ the left hand side of $(\ref{deltas})$ is just $(x-(\alpha+1))D-xD^2$ which, after replacing $D^k$ by $z^k$ gives $(x-(\alpha+1))z-xz^2=-z L_1^{(\alpha)}(x-xz)$. Thus the statement of the proposition holds in case $n=1$. Next we calculate
\begin{eqnarray*}
\delta(\delta-1)(\delta-2) \cdots (\delta-(n-1))&=&\sum_{k=n}^{2n}{}_{2n}q_{k,\alpha}(x)D^k \\
&=&\left(\sum_{k=n}^{2n-2} {}_{2n-2}q_{k,\alpha}(x)D^k \right)(\delta-(n-1))\\
&=&\sum_{k=n}^{2n-2} {}_{2n-2}q_{k,\alpha}(x)(x-(\alpha+1))D^{k+1}+\sum_{k=n}^{2n-2} {}_{2n-2}q_{k,\alpha}(x)kD^k\\
&-&\sum_{k=n}^{2n-2} {}_{2n-2}q_{k,\alpha}(x)kD^{k+1}-x\sum_{k=n}^{2n-2} {}_{2n-2}q_{k,\alpha}(x)D^{k+2}\\
&-&(n-1)\sum_{k=n}^{2n-2} {}_{2n-2}q_{k,\alpha}(x)D^k.
\end{eqnarray*}
Going from the second to the third line in this calculation we made use of Lemma \ref{commutator}. Replacing $D^k$ by $z^k$ in this expression along with the inductive hypothesis gives
\begin{eqnarray*}
\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x)z^k&=&z(x-(\alpha+1))(n-1)!(-1)^{n-1}z^{n-1}L_{n-1}^{(\alpha)}(x-xz)\\
&+&(z-z^2)D_z\left[ (n-1)!(-1)^{n-1}z^{n-1}L_{n-1}^{(\alpha)}(x-xz) \right]\\
&-&xz^2(n-1)!(-1)^{n-1}z^{n-1}L_{n-1}^{(\alpha)}(x-xz)\\
&-&(n-1)(n-1)!(-1)^{n-1}z^{n-1}L_{n-1}^{(\alpha)}(x-xz)\\
&=&(n-1)!(-1)^{n-1}z^{n-1}\left\{z(x-(\alpha+1))L_{n-1}^{(\alpha)}(x-xz) \right.\\
&+& (1-z)\left[(n-1)L_{n-1}^{(\alpha)}(x-xz)-zx\frac{d}{dw}\left[L_{n-1}^{(\alpha)}(w) \right]_{w=x-xz} \right]\\
&-&\left. xz^2L_{n-1}^{(\alpha)}(x-xz)-(n-1)L_{n-1}^{(\alpha)}(x-xz)\right\}\\
&=&(n-1)!(-1)^{n-1}z^{n-1}\left\{-z(\alpha+n)L_{n-1}^{(\alpha)}(x-xz) \right.\\
&+& \left. z(x-xz)L_{n-1}^{(\alpha)}(x-xz)-z(x-xz)\frac{d}{dw}\left[L_{n-1}^{(\alpha)}(w) \right]_{w=x-xz} \right\}.
\end{eqnarray*}
Since the generalized Laguerre polynomials satisfy the relations
\begin{eqnarray}
xDL_n^{(\alpha)}(x)&=&nL_n^{(\alpha)}(x)-(\alpha+n)L_{n-1}^{(\alpha)}(x) \label{recrel1} \\
DL_n^{(\alpha)}(x)&=&DL_{n-1}^{(\alpha)}(x)-L_{n-1}^{(\alpha)}(x) \label{recrel2}
\end{eqnarray}
(see for example Ch. 12 in \cite{Rainville}), it follows that
\begin{eqnarray*}
&&(n-1)!(-1)^{n-1}z^{n-1}\left\{-z(\alpha+n)L_{n-1}^{(\alpha)}(x-xz) \right.\\
&+& \left. z(x-xz)L_{n-1}^{(\alpha)}(x-xz)-z(x-xz)\frac{d}{dw}\left[L_{n-1}^{(\alpha)}(w) \right]_{w=x-xz}\right\}\\
&=&(n-1)!(-1)^{n-1}z^{n-1}\left\{-znL_{n}^{(\alpha)}(x-xz) \right.\\
&+& z(x-xz)DL_{n-1}^{(\alpha)}(x-xz)-z(x-xz)L_{n-1}^{(\alpha)}(x-xz) \\
&+& \left. z(x-xz)L_{n-1}^{(\alpha)}(x-xz)-z(x-xz)DL_{n-1}^{(\alpha)}(x-xz)\right\}\\
&=&(n-1)!(-1)^{n-1}z^{n-1}(-znL_n^{(\alpha)}(x-xz))\\
&=&n!(-1)^nz^nL_n^{(\alpha)}(x-xz).
\end{eqnarray*}
The proof of Proposition \ref{BIGlemma} is complete.
\end{proof}
\begin{thm} The sequence $\set{k(k-1)(k-2)\cdots(k-(n-1))}_{k=0}^{\infty}$ is an $L^{(\alpha)}$-multiplier sequence for $n \in {\mathbb N}, \ n \geq 1$.
\end{thm}
\begin{proof}
Let $T$ be the linear operator defined by $T[L_k^{(\alpha)}(x)] = k(k-1)\cdots(k-n+1)L_k^{(\alpha)}(x)$. Then $T = \delta (\delta-1)(\delta-2)\cdots(\delta-(n-1))$, where $\delta:=(x-(\alpha+1))D-xD^2$ and $D$ denotes differentiation with respect to $x$. Using the definition of the generalized Laguerre polynomials, we have
\begin{eqnarray*}
n!(-1)^n z^n L_n^{(\alpha)}(x-xz) &=& n!(-1)^n z^n \sum_{k=0}^{n}\left(\begin{array}{c} n+\alpha \\ n-k \end{array} \right)(-1)^k \frac{(x-xz)^k}{k!}\\
&=& n!(-1)^n \sum_{k=0}^{n}\left(\begin{array}{c} n+\alpha \\ n-k \end{array} \right)(-1)^k \frac{x^k}{k!} z^n (1-z)^k.
\end{eqnarray*}
Thus, by Proposition \ref{BIGlemma},
$$
T = n!(-1)^n \sum_{k=0}^{n}\left(\begin{array}{c} n+\alpha \\ n-k \end{array} \right)(-1)^k \frac{x^k}{k!} D^n (1-D)^k,
$$
and we have
\begin{eqnarray*}
T[\exp(-xw)] &=& n!(-1)^n \sum_{k=0}^{n}\left(\begin{array}{c} n+\alpha \\ n-k \end{array} \right)(-1)^k \frac{x^k}{k!} D^n (1-D)^k [\exp(-xw)]\\
&=& n!(-1)^n \sum_{k=0}^{n}\left(\begin{array}{c} n+\alpha \\ n-k \end{array} \right)(-1)^k \frac{x^k}{k!} (-w)^n (1+w)^k \exp(-xw)\\
&=& n!(-1)^n (-w)^n L_n^{(\alpha)}(x+xw) \exp(-xw).
\end{eqnarray*}
Note that
$$
f_m(x,w) = n!(-1)^n (-w)^n L_n^{(\alpha)}(x+xw) \left(1-\frac{xw}{m}\right)^m \qquad (m\in{\mathbb N})
$$
converges uniformly on compact subsets to $T[\exp(-xw)]$ as $m\to\infty$. Let $0<x_1<x_2<\cdots<x_n$ be the zeros of $L_n^{(\alpha)}(x)$ (recall that the generalized Laguerre polynomials have only real simple positive zeros). Then $f_m(x,w)=0$ if and only if either $w=0$, $x(1+w)=x_k$ or $xw=m$, none of which occur when Im $x>0$ and Im $w>0$. Therefore, by Theorem \ref{BBthm}, $T$ preserves reality of zeros.
\end{proof}
We conclude this section with a corollary to this theorem. Although the corollary does not have a direct application to the development of $L^{(\alpha)}$-multiplier sequences, it is a quick result so we include it here.
\begin{cor} Let $\delta$ be as in Proposition \ref{BIGlemma} and let
\[
\delta (\delta-1) \cdots (\delta-(n-1))=\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x)D^k.
\]
Then
\[
\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x)=(-1)^n \prod_{k=1}^n (\alpha+k).
\]
\end{cor}
\begin{proof} By Proposition \ref{BIGlemma} we have
\[
\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x)=\sum_{k=n}^{2n} {}_{2n}q_{k,\alpha}(x)z^k \Big|_{z=1}=n!(-1)^nL_n^{(\alpha)}(0).
\]
On the other hand, using the generating function
\[
\frac{1}{(1-t)^{1+\alpha}} e^{\frac{-xt}{1-t}}=\sum_{n=0}^{\infty} L_n^{(\alpha)}(x)t^n
\]
we see that
\[
n!(-1)^nL_n^{(\alpha)}(0)=(-1)^n \prod_{k=1}^n (\alpha+k).
\]
\end{proof}
\section{Properties of $L^{(\alpha)}$-multiplier sequences}
\subsection{Classical Properties}
There are a number of properties of the classical multiplier sequences which are easily verified. Here we list those that carry over to $L^{(\alpha)}$-multiplier sequences.
\begin{lem} \label{properties} Let $\set{\gamma_k}_{k=0}^{\infty}$ be an $L^{(\alpha)}$-multiplier sequence. Then:
\begin{itemize}
\item[(i)] If there exists an integers $n> m \geq 0$ such that $\gamma_m \neq 0$ and $\gamma_n=0$, then $\gamma_k=0$ for all $k \geq n$.
\item[(ii)] The terms of $\set{\gamma_k}_{k=0}^{\infty}$ are either all of the same sign, or they alternate in sign.
\item[(iii)] For any $r \in {\mathbb R}$, the sequence $\set{r \gamma_k}_{k=0}^{\infty}$ is also an $L^{(\alpha)}$-multiplier sequence.
\item[(iv)] The terms of $\set{\gamma_k}_{k=0}^{\infty}$ satisfy Tur\'an's inequality
\[
\gamma_k^2-\gamma_{k-1}\gamma_{k+1} \geq 0, \quad \quad k=1,2,3, \ldots
\]
\end{itemize}
\end{lem}
\begin{proof} These claims follow immediately from Theorem \ref{thmqmsms} and the fact the generalized Laguerre polynomials form a simple set of polynomials. Properties $(i)-(iv)$ for classical multiplier sequences have been established in \cite{Levin}.
\end{proof}
\begin{rmk} To draw further contrast between $L^{(\alpha)}$-multiplier sequences, Hermite multiplier sequences, and classical multiplier sequences, we demonstrate that the following two properties, which hold for multiplier sequences and Hermite multiplier sequences, do \emph{not} hold for $L^{(\alpha)}$-multiplier sequences.
\begin{itemize}
\item[(a)] If $\set{\gamma_k}_{k=0}^{\infty}$ is a multiplier sequence, then $\set{\gamma_k}_{k=m}^{\infty}$ is a multiplier sequence for any $m\in\mathbb{N}$.
\item[(b)] If $\set{\gamma_k}_{k=0}^{\infty}$ is a multiplier sequence, then $\set{(-1)^k\gamma_k}_{k=0}^{\infty}$ is a multiplier sequence.
\end{itemize}
For property (a), we note that for the simple Laguerre polynomials ($\alpha=0$), the sequence $\{k+1\}_{k=0}^{\infty}$ is an $L^{(0)}$-multiplier sequence, but $\{k+1\}_{k=1}^{\infty} = \{k+2\}_{k=0}^{\infty}$ is not (see Theorem \ref{linchar}).
For property (b), we note again that $\{k+1\}_{k=0}^{\infty}$ is an $L^{(0)}$-multiplier sequence. We now show that $\{(-1)^k(k+1)\}_{k=0}^{\infty}$ is not. The polynomial
$$
p(x) = (x-10)^2 = 82L_0^{(0)}(x)+16 L_1^{(0)}(x)+2 L_2^{(0)}(x)
$$
has only real zeros, while
$$
3 \cdot 82L_0^{(0)}(x)-2\cdot16 L_1^{(0)}(x)+1\cdot2 L_2^{(0)}(x) = 3x^2+20x+56
$$
has two non-real zeros.
\end{rmk}
\begin{prop} Suppose that $\set{\gamma_n}$ is a non-trivial $L^{(\alpha)}$-multiplier sequence for some $\alpha >-1$. Then there exists an $m \in \mathbb{Z}$ such that $\gamma_k=0$ for all $k<m$ and $\gamma_k \neq 0$ for all $k \geq m$.
\end{prop}
\begin{proof} Since $\set{\gamma_n}$ is a non-trivial multiplier sequence, there is at least one $k \in \mathbb{Z}$ such that $\gamma_k \neq 0$. Let $m$ be the minimal index such that $\gamma_m \neq 0$. It is easy to see that $\gamma_{m+1}$ and $\gamma_{m+2}$ are non-zero, for if either of them were zero, in light of Lemma \ref{properties} we would have to conclude that $\set{\gamma_k}_{k=0}^{\infty}$ is a trivial multiplier sequence. Suppose now that there exists a $n>m+2$ such that $\gamma_n=0$. By Lemma $\ref{bmax}$ (see below) there are constants $a_m, a_{m+2}$ such that the polynomial
\[
\widetilde{q(x)}=a_m\gamma_mL^{(\alpha)}_m(x)+a_{m+2}\gamma_{m+2} L^{(\alpha)}_{m+2}(x)
\]
has some non-real zeros. On the other hand, by Lemma $\ref{open}$ then there exists $a_n$ such that
\[
q(x)=a_mL^{(\alpha)}_m(x)+a_{m+2}L^{(\alpha)}_{m+2}(x)+a_nL^{(\alpha)}_n(x) = a_n\left(L^{(\alpha)}_n(x) +\frac{a_m}{a_n}L^{(\alpha)}_m(x)+\frac{a_{m+2}}{a_n}L^{(\alpha)}_{m+2}(x)\right)
\]
has only real zeros. Applying the $L^{(\alpha)}$-multiplier sequence $\set{\gamma_k}_{k=0}^{\infty}$ to $q(x)$ we obtain the polynomial $\widetilde{q(x)}$, a contradiction. Hence $\gamma_k \neq 0$ for all $k \geq m$ and the proof is complete.
\end{proof}
\subsection{Monotonicity of $L^{(\alpha)}$-multiplier sequences}
The main result in this section is that if a classical multiplier sequence is an $L^{(\alpha)}$-multiplier sequence, then it must be non-decreasing. We note that an analogous statement is true for the Hermite multiplier sequences. The converse is also true for Hermite multiplier sequences, but not for $L^{(\alpha)}$-multiplier sequences (recall the sequences $\{r^k\}$ for $r>1$!). We next lay the necessary groundwork to establish the stated monotonicity result for $L^{(\alpha)}$-multiplier sequences.
\newline \indent We begin with two simple, but very useful lemmas. The first one essentially says that if a polynomial has only simple real zeros and one makes a small perturbation of the coefficients, then the resulting polynomial also has only real zeros.
\begin{lem}\label{open}
Let $p$ and $q$ be real polynomials and suppose $\deg(q)<\deg(p)$. If $p$ has only simple real zeros then there exists $\epsilon>0$ such that $p(x)+ b q(x)$ has only real zeros whenever $|b|<\epsilon$.
\end{lem}
\begin{proof}
Suppose no such $\epsilon$ exists. Then we can obtain a sequence of real numbers $\{b_n\}_{n=1}^{\infty}$ converging to zero such that, for each $n$, the real polynomial $p_n(x) = p(x)+ b_n q(x)$ has some non-real zeros. The polynomials $p_n$ converge uniformly on compact subsets of $\mathbb{C}$ to $p$. By Hurwitz' Theorem, the zeros of $p$ must be limits of the zeros of $p_n$, contradicting the fact that the zeros of $p$ are all real and simple.
\end{proof}
The next result is similar in nature. If we begin with a polynomial which has some non-real zeros then any small perturbation of the coefficients will result in another polynomial which has some non-real zeros.
\begin{lem}\label{closed}
Let $p$ and $q$ be real polynomials and suppose $\deg(q)<\deg(p)$. If $p$ has some non-real zeros then there exists $\epsilon>0$ such that $p(x)+ b q(x)$ has some non-real zeros whenever $|b|<\epsilon$.
\end{lem}
\begin{proof}
We appeal to Hurwitz' Theorem once again. If no such $\epsilon$ exists, then we can obtain a sequence of real numbers $\{b_n\}_{n=1}^{\infty}$ converging to zero such that, for each $n$, the real polynomial $p_n(x) = p(x)+ b_n q(x)$ has only real zeros. The polynomials $p_n$ converge uniformly on compact subsets of $\mathbb{C}$ to $p$. By Hurwitz' Theorem, the zeros of $p$ must be limits of the zeros of $p_n$, but non-real numbers are never the limit of a sequence of real numbers, a contradiction.
\end{proof}
\begin{lem}\label{bmax}
For $n\geq 2$ and $b\in\mathbb{R}$, define
\begin{eqnarray*}
f_{n, b, \alpha} (x)&:=&L_n^{(\alpha)}(x)+b L_{n-2}^{(\alpha)}(x), \quad \textrm{and}\\
E_n&:=&\set{ b \in {\mathbb R} \ | \ f_{n, b, \alpha}(x) \quad \textrm{has only real zeros}}.
\end{eqnarray*}
Then $\max(E_n)$ exists, and is a positive real number.
\end{lem}
\begin{proof}
By Lemma $\ref{open}$, there exists $\epsilon>0$ such that $(-\epsilon, \epsilon)\subseteq E_n$. In particular, $E_n$ is nonempty and $\max(E_n)$, if it exists, is positive. It now suffices to show that $E_n$ is closed and bounded above.
Suppose $t \in (\mathbb{R}\setminus E_n)$. Then, by Lemma $\ref{closed}$, there exists $\delta>0$ such that
$$
f_{n, t, \alpha}(x)+ b L_{n-2}^{(\alpha)}(x) = L_n^{(\alpha)}(x)+(t+b) L_{n-2}^{(\alpha)}(x).
$$
has non-real zeros whenever $|b|<\delta$. That is to say, $(b-\delta, b+\delta)\subseteq (\mathbb{R}\setminus E_n)$. Whence, $\mathbb{R}\setminus E_n$ is open and, therefore, $E_n$ is closed.
To show that $E_n$ is bounded above, we consider the $(n-2)^{nd}$ derivative of $f_{n,b,\alpha}$. A calculation shows
$$
\frac{d^{n-2}}{dx^{n-2}} f_{n, b, \alpha} (x) = \frac{1}{2}x^2 - (n+\alpha) x + \frac{(n+\alpha)(n+\alpha-1)}{2} + b.
$$
Thus $\displaystyle \frac{d^{n-2}}{dx^{n-2}} f_{n, b, \alpha} (x)$, and therefore $ f_{n, b, \alpha} (x)$, have some non-real zeros whenever $b$ is sufficiently large.
\end{proof}
\begin{thm} If the sequence of positive real numbers $\displaystyle{\set{\gamma_k}_{k=0}^{\infty}}$ is a non-trivial $L^{(\alpha)}$-multiplier sequence, then $\gamma_k \leq \gamma_{k+1}$ for all $k \geq 0$.
\end{thm}
\begin{proof} Let $T_L$ denote the operator associated to the $\displaystyle{L^{(\alpha)}}$-multiplier sequence $\set{\gamma_k}_{k=0}^{\infty}$. With the notation of Lemma \ref{bmax}, for each $n \geq 2$ the function
\[
f_{n,\beta_n^*,\alpha}(x)=L_n^{(\alpha)}(x)+\beta^*_n L_{n-2}^{(\alpha)}(x) \qquad (\beta_n^* = \max(E_n))
\]
has only real zeros. It follows that
\[
T_L[f_{n,\beta_n^*,\alpha}(x)]=\gamma_n L_n^{(\alpha)}(x)+\gamma_{n-2}\beta^*_n L_{n-2}^{(\alpha)}(x)=\gamma_{n} \left(L_n^{(\alpha)}(x)+\frac{\gamma_{n-2}}{\gamma_n}\beta^*_n L_{n-2}^{(\alpha)}(x) \right)
\]
also only has real zeros. By Lemma \ref{bmax} we must have $\displaystyle{\frac{\gamma_{n-2}}{\gamma_n}\beta^*_n \leq \beta_n^*}$ which gives $\displaystyle{0 < \frac{\gamma_{n-2}}{\gamma_n} \leq 1}$. On the other hand, by Lemma \ref{properties}, we have
\[
\gamma_{n-1}^2-\gamma_{n}\gamma_{n-2} \geq 0, \quad \quad (n \geq 2)
\]
which means $\displaystyle{\left( \frac{\gamma_{n-1}}{\gamma_{n-2}}\right)^2 \geq \frac{\gamma_{n}}{\gamma_{n-2}}\geq1}$. In other words $\gamma_{n-1} \geq \gamma_{n-2}$ and the proof is complete.
\end{proof}
\section{Open questions}
Contrary to the linear sequences, quadratic (and higher degree) multiplier sequences for generalized Laguerre bases are not well understood and are far from being completely characterized. Recall from Section \ref{arbdegree} that $L^{(\alpha)}$-multiplier sequences of arbitrary degrees exist. As a result, investigations into quadratic, cubic, and higher degree $L^{(\alpha)}$-multiplier sequences are not vacuous, and rather challenging. One of the reasons for this is that although one can naturally get higher order $L^{(\alpha)}$-multiplier sequences from lower order ones, one can not get them all this way. There are for example quadratic $L^{(\alpha)}$-multiplier sequences that do not factor as a product of the linear ones (in the differential operator sense). In this section we present some partial results in the characterization of sequences of the form $\{k^2+ak+b \}_{k=0}^{\infty}$ for the simple Laguerre polynomials ($\alpha=0$) and pose some open questions.
\newline
Based on some partial results, we believe that the following conjecture is true:
\begin{conj} The sequence $\{k^2+ak+b\}_{k=0}^{\infty}$ is an $L^{(0)}$-multiplier sequence if and only if
$$
-1\leq a \leq 3 \quad \text{ and }\quad \max \{ 0, a-1 \} \leq b \leq \displaystyle{\frac{1}{8}(1+a)^2}.
$$
\end{conj}
It is easy to show that if $\{k^2+ak+b\}_{k=0}^{\infty}$ is an $L^{(0)}$-multiplier sequence then then $a\geq -1$ and $\displaystyle 0\leq b \leq \frac{1}{4} (a+1)^2$. It is a bit more involved to improve the upper bound on $b$ to $\frac{1}{8} (a+1)^2$ but it can be done by using the result of Borcea and Br\"and\'en. The proof involves the verification of stability of a certain polynomial in two complex variables, and is a bit technical. We believe that for the characterization of polynomial type $L^{(\alpha)}$-multiplier sequences of arbitrary (fixed) degree, additional techniques will be needed.
Using a theorem due to Newton, we can easily establish the bounds $a \leq 4$ and $a-1\leq b$ and another application of Borcea Br\"and\'en gives that if $1 \leq a \leq 3$ then $b=a-1$ is allowed, in other words $\{k^2+ak+a-1 \}$ is an $L^{(0)}$-multiplier sequence. Though these results pointed to the formulation of the above conjecture, we were unable to prove the result so far and the question remains open. The situation is similar for all polynomial sequences of degree 3 or higher.
|
\section{Introduction}
Hindman's Theorem states that, in any finite coloring of the integers, some color contains an infinite set and all its finite sums. Hindman's original proof \cite{hindman74} is quite complicated; fortunately, there are both simpler combinatorial arguments \cite{baumgartner74,towsner09:hindman_simple} and an elegant proof based on the topology of ultrafilters (see, for instance, \cite{comfort77}).
Strikingly, the ultrafilter argument gives, sometimes with little additional work, various strengthenings of the theorem for which combinatorial proofs are iether much harder, or not known to exist. (\cite{hindman98} gives a thorough exploration of many uses of ultrafilters in this context.) One such strengthening promises, not a single monochromatic infinite set and its finite sums, but a tree of such sets in which there are ``many'' possible first elements, for each such first element ``many'' possible second elements, and so on.
In this paper we give the first combinatorial proof of this theorem, modeled on Baumgartner's proof of the ordinary Hindman's Theorem. The key idea is the use of approximate ultrafilters, as introduced by Hirst \cite{hirst04}---countable collections of sets of integers which nonetheless contain enough information to complete the proof. The proof here is modeled on our related proof of the ordinary Hindman's Theorem \cite{towsner09:hindman_unwind}\footnote{Indeed, we originally found a proof quite similar to that one, and only subsequently found proof in the style of Baumgartner which we present here. This proof is slightly more elegant, and we hope that it will shed some light on the relationship between Baumgartner's proof of the ordinary Hindman's Theorem and the ultrafilter proof.}.
The reverse mathematical strength of even the ordinary Hindman's Theorem is open; bounds are given in \cite{blass87}, and the gap between the lower and upper bounds on reverse mathematical strength there have not been improved. The proof given here is entirely within the bounds of second order arithmetic, but well above their upper bounds; no lower bound for the Dense Hindman's Theorem is known besides the obvious one, that any lower bound for the ordinary Hindman's Theorem must also bound the Dense Hindman's Theorem.
We thank Mathias Beiglb\"ock for bringing this question to our attention, and for many discussions about the mathematics around Hindman's Theorem.
\section{General Definitions}
Throughout this paper, variables denoted by lowercase letters will typically be natural numbers, except for $b$, which will always belong to $\{-1,1\}$. Variables denoted by upppercase letters will be sets of natural numbers, and variables denoted by caligraphic letters ($\mathcal{U}$, $\mathcal{F}$, etc.) will be sets of sets of natural numbers. In fact, it would cause no harm to assume that all sets of sets of natural numbers appearing in this paper are countable, and therefore to code them using sets of natural numbers. The one exception is the property $\mathfrak{P}$, which represents the set of sets of natural numbers satisfying some shift-invariant divisible'' property, such as the infinite sets, the piecewise syndetic sets, or the sets of positive upper Banach density. (The notion of divisibility is introduced in \cite{glasner80}.
We adopt the notation that if $X$ is a set of integers, $1\cdot X=X$ and $-1\cdot X=\mathbb{N}\setminus X$. We write $X-n$ for $\{m\mid m+n\in X\}$.
\begin{definition}
Let $\mathfrak{P}$ be a collection of sets of integers such that\footnote{To keep our promise that the proof go through in second order arithmetic, we should insist that $\mathfrak{P}$ be given by some arithmetic formula; this includes all the examples given.}:
\begin{itemize}
\item $\mathbb{N}\in\mathfrak{P}$
\item $\emptyset\not\in\mathfrak{P}$
\item If $X\subseteq Y$ and $X\in\mathfrak{P}$ then $Y\in\mathfrak{P}$
\item If $X_0\cup X_1=X$ and $X\in\mathfrak{P}$ then either $X_0\in\mathfrak{P}$ or $X_1\in\mathfrak{P}$
\item For any $X$ and any $n$, $X\in\mathfrak{P}$ iff $X-n\in\mathfrak{P}$
\end{itemize}
We say a collection $\mathcal{U}$ of sets of integers has the $\mathfrak{P}$-finite intersection property ($\mathfrak{P}$-fip) if for every finite collection $\mathcal{F}\subseteq\mathcal{U}$,
\[\bigcap_{S\in \mathcal{F}}S\in\mathfrak{P}.\]
\end{definition}
Natural examples of such properties $\mathfrak{P}$ include:
\begin{itemize}
\item $\mathfrak{P}$ is the collection of infinite sets
\item $\mathfrak{P}$ is the collection of sets with positive upper Banach density
\item $\mathfrak{P}$ is the collection of piecewise syndetic sets
\item $\mathfrak{P}$ is the collection of sets $X$ such that $\sum_{x\in X}1/x=\infty$
\end{itemize}
Properties satisfying all but the final condition, that $X\in\mathfrak{P}$ iff $X-n\in\mathfrak{P}$, are called \emph{divisible} \cite{glasner80}.
\begin{definition}
Let $\mathcal{U}$ be a countable collection of sets of integers\footnote{None of our arguments would change if uncountable collections---say, true ultrafilters---are allowed. However we wish to emphasize that none of our arguments will require more than countable collections}. We write $X\tilde\in\mathcal{U}$ if there is a finite $\mathcal{F}\subseteq\mathcal{U}$ such that $\bigcap_{S\in\mathcal{F}}S\subseteq X$.
We say $\mathcal{U}$ is a $\mathfrak{P}$-semigroup if $\mathcal{U}$ satisfies $\mathfrak{P}$-fip, $\mathcal{U}$ is closed under finite intersections, and whenever $X\in\mathcal{U}$, there is a $Y\tilde\in\mathcal{U}$ such that $X-n\tilde\in\mathcal{U}$ for each $n\in Y$.
\end{definition}
There are two useful ways to view $\mathfrak{P}$-semigroups. The first is to observe that every $\mathfrak{P}$-semigroup represents a closed semigroup in the space of ultrafilters on the integers (namely, the collection of ultrafilters extending $\mathcal{U}$).
The second is to recall that an IP set is a set $S$ such that there is an infinite $T\subseteq S$ all of whose finite sums also belong to $S$. Then the collection $\{S-n\mid n\in FS(T)\}$ (where $FS(T)$ is the finite sums from $T$) is a canonical example of a $\mathfrak{P}$-semigroup where $\mathfrak{P}$ is the collection of infinite sets. The notion of a $\mathfrak{P}$-semigroup generalizes an IP set in two directions: first, it allows for arbitrary $\mathfrak{P}$. (Note that the appropriate requirement is \textit{not} that $T\in\mathfrak{P}$; rather, the requirement is that there are many possible choices for the infinite set $T$---indeed, the an infinite tree of such sets, with the number of possible branches at each level belonging to $\mathfrak{P}$. This discussion will be made precise when we state the main theorem below.) Second, if we have an infinite descending sequence of IP sets $S_1\supseteq S_2\supseteq\cdots S_n\supseteq$, their intersection may well be $0$. However the union of the corresponding $\mathfrak{P}$-semigroups is still a $\mathfrak{P}$-semigroup. So $\mathfrak{P}$-semigroups also generalize IP sets by accomodating the result of infinitely many successive refinements of an IP set.
The argument here is very similar to a proof using ultrafilters, but we emphasize that the $\mathfrak{P}$-semigroups appearing in the proof are much simpler objects: they are countable collections, built with no use of the axiom of choice.
The following lemma is our essential building block:
\begin{lemma}
Let $\mathcal{U}$ satisfy $\mathfrak{P}$-fip, and let $A$ be a set of integers. Then either $\mathcal{U}\cup\{A\}$ or $\mathcal{U}\cup\{-1\cdot A\}$ satisfies $\mathfrak{P}$-fip.
\end{lemma}
\begin{proof}
Suppose neither collection satisfies $\mathfrak{P}$-fip. Then choose finite sets $\mathcal{F},\mathcal{F}'\subseteq \mathcal{U}$ such that
$\bigcup_{S\in\mathcal{F}}S\cap A\not\in\mathfrak{P}$
and
$\bigcup_{S\in\mathcal{F}'}S\cap (-1\cdot X)\not\in\mathfrak{P}$.
Then
\[\bigcap_{S\in\mathcal{F}\cup\mathcal{F'}}S\subseteq (\bigcup_{S\in\mathcal{F}}S\cap X)\cup(\bigcup_{S\in\mathcal{F}'}S\cap (-1\cdot X)).\]
But this is impossible, since $\bigcap_{S\in\mathcal{F}\cup\mathcal{F'}}S\in\mathfrak{P}$ must hold.
\end{proof}
\section{Dense Hindman's Theorem}
\begin{lemma}
If $\mathcal{U}$ is a $\mathfrak{P}$-semigroup, $A$ is a set, $S$ is a set with the property that
\[\mathcal{U}\cup\{-1\cdot(A-n)\mid n\not\in S\}\]
satisfies $\mathfrak{P}$-fip, and $\mathcal{U}\cup\{S-n\mid n\in S\}$ does not satisfy $\mathfrak{P}$-fip, then there is a finite $F\subseteq S$ and a $Y\tilde\in\mathcal{U}$ such that $0\in Y$ and $\mathcal{U}\cup\{\bigcup_{n\in F}-1\cdot(A-n-m)\mid m\in Y\}$ satisfies $\mathfrak{P}$-fip.
\label{3.1}
\end{lemma}
\begin{proof}
Since $\mathcal{U}\cup\{S-n\mid n\in S\}$ does not satisfy $\mathfrak{P}$-fip, let $X\tilde\in\mathcal{U}$ and $F\subseteq S$ be such that $X\cap\bigcap_{n\in F}S-n\not\in\mathfrak{P}$. Let $Y$ be the set of $n$ such that $X-n\tilde\in\mathcal{U}$; clearly $0\in Y$ since $X\tilde\in\mathcal{U}$.
We claim that $\mathcal{U}\cup\{\bigcup_{n\in F}-1\cdot(A-n-m)\mid m\in Y\}$ satisfies $\mathfrak{P}$-fip. Let $Z\tilde\in\mathcal{U}$ and $G\subseteq Y$ be finite. Let $Z'\tilde\in\mathcal{U}$ be such that $Z-n\tilde\in\mathcal{U}$ for each $n\in Z'$; then $Z'\cap\bigcap_{m\in G}\bigcup_{n\in F}-1\cdot(S-n-m)\tilde\in\mathcal{U}$. If we choose any $k$ in this set, we have $Z-k\cap\bigcap_{m\in G}\bigcup_{n\in F}-1\cdot (A-k-n-m)\in\mathfrak{P}$, and therefore $Z\cap\bigcap_{m\in G}\bigcup_{n\in F}-1\cdot (A-n-m)\in\mathfrak{P}$.
\end{proof}
In particular, observe that if
\[\mathcal{U}\cup\{A-n\mid n\in S\}\]
satisfies $\mathfrak{P}$-fip but whenever $m\not\in S$,
\[\mathcal{U}\cup\{A-n\mid n\in S\}\cup\{A-m\}\]
does not satisfy $\mathfrak{P}$-fip then $S$ satisfies the condition of this lemma.
\begin{definition}
We say $A$ is \emph{large} relative to a $\mathfrak{P}$-semigroup $\mathcal{U}$ if whenever $\mathcal{V}$ is a $\mathfrak{P}$-semigroup extending $\mathcal{U}$, $\mathcal{V}\cup\{A\}$ satisfies $\mathfrak{P}$-fip.
\end{definition}
\begin{lemma}
If $C=C_1\cup\cdots\cup C_n$ and $C$ is large relative to $\mathcal{U}$, there is a $\mathcal{V}$ extending $\mathcal{U}$ and an $i$ such that $C_i$ is large for $\mathcal{V}$.
\end{lemma}
\begin{proof}
Applying induction, it suffices to consider the case $n=2$. If $C_1$ is large relative to $\mathcal{U}$ then $1$ and $\mathcal{U}$ suffice. Otherwise there is a $\mathcal{V}$ extending $\mathcal{U}$ such that $\mathcal{V}\cup\{C_1\}$ fails to satisfy $\mathfrak{P}$-fip, and therefore $C_2$ is large relative to $\mathcal{V}$.
\end{proof}
\begin{theorem}
Let $\mathcal{U}$ be a $\mathfrak{P}$-semigroup and $A$ a set. If $A$ is large for $\mathcal{U}$ then there is a $\mathcal{V}$ extending $\mathcal{U}$ such that the set of $n\in A$ for which $A\cap A-n$ is large for $\mathcal{V}$ belongs to $\mathfrak{P}$.
\end{theorem}
\begin{proof}
Given $\mathcal{U}$, expand to some $\mathfrak{P}$-semigroup $\mathcal{V}$ such that for every finite set $F$, either $\bigcap_{n\in F}A-n$ is large for $\mathcal{V}$ or $\bigcup_{n\in F}-1\cdot(A-n)\tilde\in\mathcal{V}$.
Let $S$ be a maximal set of $n$ so that $0\in S$, $F\subseteq S$ implies $\bigcap_{n\in F}A-n$ is large for $\mathcal{V}$, and $m\not\in S$ implies that there is an $F\subseteq S$ so that $\bigcap_{n\in F\cup\{m\}}A-n$ is not large for $\mathcal{V}$.
Then if $F\subseteq -1\cdot S$ is finite, we may choose a $G\subseteq S$ so that $\bigcap_{n\in G\cup\{m\}}A-n$ is not large for $\mathcal{V}$ for any $m\in F$. Then $\bigcup_{n\in G\cup\{m\}}-1\cdot(A-n)\tilde\in\mathcal{V}$ for each $m\in F$, and since $\mathcal{V}\cup\{\bigcap_{n\in G}A-n\}$ satisfies $\mathfrak{P}$-fip, it follows that $\mathcal{V}\cup\{-1\cdot (A-m)\mid m\in F\}$ satisfies $\mathfrak{P}$-fip. This holds for every finite $F\subseteq -1\cdot S$, so $\mathcal{V}\cup\{-1\cdot(A-m)\mid m\not\in S\}$ satisfies $\mathfrak{P}$-fip.
Therefore, by Lemma \ref{3.1}, if $\mathcal{V}\cup\{S-n\mid n\in S\}$ does not satisfy $\mathfrak{P}$-fip then there is a finite set $F\subseteq S$ and a $\mathcal{W}$ extending $\mathcal{V}$ such that $\bigcup_{n\in F}-1\cdot (A-n)\tilde\in\mathcal{W}$. But this would contradict the fact that $\bigcap_{n\in F}A-n$ is large for $\mathcal{V}$. So $\mathcal{V}\cup\{S-n\mid n\in S\}$ satisfies $\mathfrak{P}$-fip, and is therefore a $\mathfrak{P}$-semigroup.
Since $A$ is large for $\mathcal{V}$, also $\mathcal{V}\cup\{S-n\mid n\in S\}\cup\{A\}$ satisfies $\mathfrak{P}$-fip, so in particular, $S\cap A\in\mathfrak{P}$. Since for each $n\in S$, $A\cap A-n$ is large for $\mathcal{V}$, the claim is proven.
\end{proof}
\begin{definition}
For any set of integers $S$, let $FS(S)$ be the collection of finite sums from $S$:
\[FS(S)=\{\sum_{i\in T}i\mid T\subseteq S, T\text{ finite}\}.\]
\end{definition}
\begin{theorem}
Let $\mathbb{N}=A_1\cup\cdots\cup A_r$. There is some $i\leq r$ and a tree $\mathcal{T}$ of finite sets of integers such that:
\begin{itemize}
\item $\emptyset\in\mathcal{T}$
\item If $F\in\mathcal{T}$, $\{n\mid F\cup\{n\}\in\mathcal{T}\}$ belongs to $\mathfrak{P}$
\item If $F\in\mathcal{T}$ then $FS(F)\setminus\{0\}\subseteq A_i$
\end{itemize}
\end{theorem}
\begin{proof}
Since $\mathbb{N}$ is large for the trivial $\mathfrak{P}$-semigroup $\{\mathbb{N}\}$, we may choose a $\mathfrak{P}$-semigroup $\mathcal{U}$ and an $i$ so that $A_i$ is large for $\mathcal{U}$. By the preceeding theorem, we obtain a $\mathfrak{P}$-semigroup $\mathcal{U}_1$ extending $\mathcal{U}$ so that $\{n\in A_i\mid A_i\cap A_i-n\text{ is large for }\mathcal{U}'\}$ belongs to $\mathfrak{P}$. We place $\{n\}$ in $\mathcal{T}$ if $A_i\cap A_i-n$ is large for $\mathcal{U}'$.
Suppose we have $F\subseteq\mathcal{T}$ and have a $\mathcal{V}$ so that $A':=\bigcap_{n\in FS(F)\setminus\{0\}}A_i-n$ is large for $\mathcal{V}$. Again by the preceeding theorem, we may find a $\mathcal{V}'$ such that $\{m\in A'\mid A'\cap A'-m\text{ is large for }\mathcal{V}'\}$ belongs to $\mathfrak{P}$; then we place $F\cup\{m\}$ in $\mathcal{T}$ for each such $m$, thereby maintaining the inductive assumption.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
Star formation is a ubiquitous phenomenon in spiral galaxies such as our very own Milky Way. However the detailed process by which gas is converted into stars in such galaxies is still relatively poorly understood. One of the key open questions is why star formation is so remarkably inefficient, with only a few percent of the mass of gas in a molecular cloud ending up in stars. Recent observational results for nearby molecular clouds in the Spitzer cores-to-discs survey \citep{evansetal09} find the mass of stars in a star-forming clouds typically around 3-6\% of the mass of the parental molecular cloud, the latter estimated by multiplying the column density inferred from interstellar dust extinction maps by the area of the cloud (defined by an extinction threshold). Equivalently the observational result can be restated as saying that only a few percent of molecular cloud gas is converted into stars per gravitational free-fall time.
Thus star formation appears to be both \emph{inefficient} --- in the sense that not much gas has been converted into stars --- and \emph{slow}, in the sense that, over the timescales necessary for gravity to act on the global cloud, not many stars have formed. It is also important to note that the inefficiencies found by \citet{evansetal09} refer to nearby molecular clouds that are actively forming stars. It is a further challenge to explain molecular clouds where relatively little star formation occurs at all, such as the Pipe Nebula (efficiency $\sim 0.06\%$, \citet{forbrichetal09}) that lies in stark contrast to the profusion of star formation occurring in the nearby $\rho$-Ophiuchus cloud.
The fact that star formation appears to occur on a slower timescale than the gravitational one indicates that the answer must lie in physics beyond gravity, or at least beyond the \emph{self-gravity} of the cloud. To achieve inefficiency of star formation over time scales much greater than the dynamical time must further involve involve unbinding the cloud in some way -- for example by internal driving of turbulence by jets and outflows \citep{nl07} or invoking tidal forces from the Galactic potential \citep{bpetal09}.
Magnetic fields are, observationally, a good candidate for explaining why clouds in otherwise similar environments can have vastly different star formation efficiencies. For example, recent optical polarisation maps of the Pipe Nebula \citep{alvesetal08} reveal a remarkable degree of uniformity in the magnetic field (as inferred from the polarisation angles), in contrast to the wide dispersion in polarisation angles (on large scales) seen in active star forming regions like Orion \citep{lietal09}. The nearby Taurus molecular cloud, recently surveyed by \citet{goldsmithetal08}, forms an intermediate case, with relatively inefficient star formation (but more efficient than the Pipe Nebula) and also a well-ordered large scale magnetic field (better ordered than Orion, though less well ordered than the Pipe), together with compelling evidence for magnetic fields strong enough to control the flow of gas in (relatively) low density outer regions \citep{heyeretal08}.
The effect of a magnetic field on the self-gravitating collapse of gas to form stars can be quantified in terms of the ratio of mass within a volume to the magnetic flux threading the surface of that volume. At a critical value magnetic fields are able to prevent collapse entirely, unless some decoupling of the magnetic field from the gas occurs (i.e., ambipolar diffusion). Indeed, this theoretical understanding led to the so-called `standard model' of star formation as a quasi-static diffusion process mediated by magnetic fields \citep{sal87}. However for all star formation to occur in this manner requires that all molecular cloud cores are sub-critical (magnetic fields able to prevent collapse), whereas Zeeman measurements of field strengths in cores indicate that they are generally marginally supercritical (mass-to-flux ratios of ~few times critical, see \citet{crutcher99}). A further difficulty is the problem of how to maintain the observed turbulent motions in the cloud, since supersonic turbulence decays rapidly with or without magnetic fields in the absence of a continual driving mechanism \citep[e.g.,][]{osg01}.
This latter consideration in particular has led many to consider so-called `rapid' or `turbulent' models of star formation, where the main controlling ingredient is turbulence rather than magnetic fields, with clouds that assemble, form stars and disperse within roughly one crossing time \citep{elmegreen00}. Indeed, simulations based on simply the interaction of turbulent gas dynamics and gravity alone \citep[e.g.][]{bbb03,bb05,bate09a} do a remarkably good job of reproducing many observed statistical properties of star formation, including the gross characteristics of the initial mass function, multiplicity as a function of mass and the frequency of low mass binary stars \citep{bate09a}. However, the star formation efficiency in these calculations is much higher than observed, since the fraction of gas that is initially bound and will remain so in the absence of feedback processes, giving a SFE of order 50\% (the typical fraction of gas that is bound at the end of the calculations). Improved statistics in more recent calculations of larger clouds \citep{bate09a} also suggest that there is also a problem in terms of an over-production of very low mass stars and brown dwarfs.
Whilst observations indicate that magnetic fields are not strong enough to prevent global collapse in typical clouds, they can nevertheless play a determining role in the internal dynamics by acting as a source of pressure within the cloud (quantified by the ratio of gas-to-magnetic pressure: The plasma $\beta$) and by the magnetic braking of rotating cores. Indeed, observationally typical values for $\beta$ in molecular cloud cores are of order $0.3$ (that is, magnetic pressure dominant over gas pressure by a factor of 3) \citep{crutcher99,bourkeetal01}, similar to the value found in the cold neutral medium thought to be the precursor of molecular clouds \citep{ht04}.
In a recent series of papers \citep{pb07,pb08,pb09} we have studied the effect of magnetic fields on the formation of stars in precisely this regime: where the magnetic field is too weak to prevent global collapse but sufficiently strong to play an important role as a source of additional pressure and in magnetic braking of rotating cores. In addition we have combined this with an improved treatment of the thermodynamics of the gas on small scales by incorporating a full treatment of radiative transfer in the flux-limited diffusion approximation. We propose that slow and inefficient star formation can be explained by the combined effects of magnetic fields and radiative feedback on the star formation process \citep{pb09}. We give an overview of our findings below.
\section{Numerical methods}
Modelling star formation is made difficult by the tremendous range of length and time scales involved. For example, to follow the collapse of a giant molecular cloud of size $\sim$a few pc ($\sim 10^{11}$ km) and containing up to $10^{4} M_{\odot}$ of material to a star the size and mass of the Sun ($R_{\odot}\sim 10^{5}$ km) requires resolution over 6 orders of magnitude in length, around 14 orders of magnitude in density [$10^{4}M_{\odot} / (10^{6} R_{\odot})^{3} \to M_{\odot}/R_{\odot}^{3}$] and roughly 11 orders of magnitude in timescale (from the dynamical time of a GMC, $\sim$1 million years, to the timescale for sound-waves in the Sun of order a few minutes). Thus the computational challenge is extreme, and one cannot hope to model star formation, using for example, uniform grid techniques, though the development of such methods are well advanced. Furthermore the physics of star formation is far from simple, involving self-gravitating, turbulent gas dynamics over a huge range of length and timescales with important contributions from magnetic fields --- ultimately requiring non-ideal magnetohydrodynamics ---, radiation transport in both optically thick and thin regimes, dust and molecular chemistry and many other physical effects which should be incorporated into a realistic model.
Smoothed Particle Hydrodynamics \citep[SPH, for recent reviews see][]{price04,monaghan05} is a method very well suited to star formation studies because the resolution follows the mass by discretising the fluid equations onto Lagrangian particles that follow the fluid motion. The Lagrangian formulation, giving an exact treatment of mass advection, means that important conservation properties such as that of angular momentum are automatically satisfied, which requires very high resolution in grid techniques, even using adaptive mesh refinement (AMR). Techniques for solving the equations of self-gravitating hydrodynamics over large ranges of length and time in SPH are well established, based on the tree code algorithms used in $N-$body codes. By contrast, the solution of the equations of magnetohydrodynamics (MHD) in SPH has proved more challenging, in part due to the early discovery \citep{pm85} of numerical instabilities associated with particular formulations of the MHD equations. Our goal over the last few years has been to develop the techniques for MHD in SPH sufficiently to be able to study the role of magnetic fields in star formation problems. This has involved dealing carefully with many of the numerical issues including the aforementioned instabilities \citep{pm05}, the treatment of MHD shocks \citep{pm04a,pm04b} and perhaps most importantly (and the main difficulty), exploring methods for enforcing the $\nabla \cdot {\bf B} = 0$ ``no monopoles'' constraint with sufficient accuracy to perform calculations which evolve beyond the actual point of star formation \citep{pm05,pb07,price10}.
The method we have found for enforcing the divergence constraint that is sufficiently robust for star formation studies has been to use the `Euler Potentials' or `Clebsch' formulation, whereby the magnetic field is written in terms of two scalar potentials in the form ${\bf B} = \nabla\alpha \times \nabla\beta$. The corresponding induction equation for the magnetic field takes the form
\begin{equation}
\frac{d\alpha}{dt} = 0; \hspace{1cm} \frac{d\beta}{dt} = 0,
\label{eq:ind}
\end{equation}
corresponding to the advection of magnetic field lines by Lagrangian particles. The Euler potentials are thus very naturally suited to a Lagrangian description, but there are important limitations to their use. The main one is that fields with complicated topologies (such as a poloidal field wrapped by a toroidal one) cannot be represented by Euler potentials because they would become double-valued. A corollary to this is that such fields also cannot be generated during the calculation and thus important dynamo processes are not captured. Another way to understand this is to appreciate that evolution of a field using (\ref{eq:ind}) is, in effect, a mapping of the field from the initial to final positions of the SPH particles, and requires a one-to-one mapping, after which the field winding will no longer be captured. A further issue is that it is difficult to formulate non-ideal MHD terms for the Euler potentials --- although we add artificial dissipative terms to capture shocks it is clear that these do not and cannot be used to represent a correct physical dissipation \citep[see][]{brandenburg10}.
Nevertheless, with the above caveats in mind, we have been able to study the effect of magnetic fields on the star formation process, mainly studying the influence of the magnetic field in supporting the cloud in the initial stages of collapse, and the effect of this on the subsequent star formation sequence. Rather than starting with global turbulent-cloud star cluster formation calculations, we first studied the effect of magnetic fields on the formation of individual stars at small scales, from which we have proceeded to study star cluster formation on larger scales (see following sections).
Alongside the development of the MHD algorithms, we have also developed an algorithm for incorporating the effect of radiation using the flux-limited diffusion approximation. This is an approximation in radiation is assumed to be transported by diffusion through both optically thick and thin regions, but with the diffusion speed limited to the speed-of-light in optically thin regimes. The key challenge for adapting grid-based flux-limited diffusion techniques into an SPH context was to develop an implicit integration method that enables the radiative transport to (which is much faster than the gas dynamics, particularly in optically thin regions) to be computed on a timescale similar to the hydrodynamics \cite[for details see][]{wb04,wb05}.
\begin{figure}
\centering
\includegraphics[angle=90,width=\textwidth]{effect_on_discs.ps}
\caption{Effect of magnetic fields on the formation of circumstellar discs around young stars: Results of a simulation following the collapse of a rotating $1M_{\odot}$ spherical cloud core without (top) and with (bottom) a uniform magnetic field threading the initial cloud. Field strengths are given in terms of the mass-to-magnetic flux ratio divided by the critical value at which magnetic fields prevent collapse altogether. Despite the relatively weak field with respect to gravity the magnetic field is able to almost completely prevent disc formation due to a combination of magnetic braking and magnetic pressure in the collapsing core. Time is shown in units of the gravitational free-fall time ($t_{ff}$).}
\label{fig:discs}
\end{figure}
\section{Single and binary star formation}
Our first application of our MHD-SPH algorithm was to the collapse of a $1M_{\odot}$, $R=4\times 10^{16}$cm molecular cloud core to form single and binary stars. As the initial condition we assumed a dense, spherical, cold ($T\sim 10K$) core in pressure-equilibrium with a warm, low density medium with an initially uniform magnetic field threading the core and the medium. The sphere was given an initial solid body rotation, of $\Omega = 1.77 \times 10^{-13}$ rad s$^{-1}$ for the case of a single star, and $\Omega = 10^{-12}$ rad s$^{-1}$ for a binary, in the latter case also imposing an initial $m=2$ perturbation in density to seed the binary formation. We considered a range of field strengths from zero up to the observed mass-to-flux ratios of a few. The efficient cooling of the molecular gas was modelled using a barotropic (pressure-dependent-on-density) equation of state that assumes isothermality whilst the density is below a threshold physical value ($\rho_{c}=10^{-14}$ g/cm$^{3}$) and becomes polytropic with $\gamma=1.4$ above the critical density in order to approximate the effect of the gas becoming optically thick to radiation. Above a particular (much higher) density and provided the gas is gravitationally bound and collapsing, a `sink' particle is inserted to replace the densest, bound gas in the calculation, such that the calculations can be followed beyond the point of actual star formation.
The results of a typical set of calculations are shown in Fig.~\ref{fig:discs}, showing the projected column density as it evolves beyond one gravitational free-fall time (left-to-right, times given in units of $t/t_{ff}$), for a calculation with no magnetic fields (top) and with a mass-to-flux ratio of 4 times the critical value --- that is magnetic fields that are too weak to prevent collapse by a factor of 4). Despite the relatively weak field the effect on the formation of the circumstellar disc is catastrophic. In the absence of a magnetic field a disc is formed that is sufficiently massive so as to become unstable to gravitational instability in the form of large scale spiral arms, yet with a magnetic field only the faintest trace of a disc is visible even at the end of the calculations.
Since our initial calculations by several other groups have found similar results based on numerical simulations \citep[e.g.][]{ht08,hc09} and also semi-analytic calculations by \citet{gallietal06} (see Galli, this volume). In fact \cite{ht08} somewhat alarmingly discuss a `fragmentation crisis' and speculate further that, given the paucity of observational evidence for discs in the earliest (class 0) phase of star formation, perhaps they do not exist (instead forming later). More likely the solution lies in the fact that we have assumed ideal MHD in a regime where it is clear that non-ideal MHD effects are known to be important. Indeed later analysis by \citet{shuetal06} suggests that Ohmic resistivity can provide a solution, though nonetheless requiring a diffusion parameter considerably higher than the microscopic value. We intend to explore non-ideal MHD effects in the near future, though it requires a shift away from the Euler potentials formulation \cite[for recent progress on this, see][]{price10}.
A similarly dramatic effect of magnetic fields on binary formation was also observed, though for the case of binaries the effect depended more strongly on the magnetic field configuration, since in certain circumstances the field configuration could assist binary formation by forming a ``magnetic cushion'' between two overdense, collapsing regions. It was also found that a sufficiently large perturbation would produce a binary regardless of the magnetic field strength. Nevertheless it is clear that the presence of even a relatively weak magnetic field in a molecular cloud core can drastically change the star formation picture.
\section{Effect of magnetic fields on cluster formation}
We have also considered the effect of magnetic fields on larger scales, important to the formation of whole star clusters \citep{pb08,pb09}. Our initial study was to evaluate the influence of magnetic fields in star cluster formation calculations similar to those performed by \citet{bbb03}. The initial conditions consist of a cold ($T=10K$), 50$M_{\odot}$, uniform density cloud of radius $\sim 0.2$ pc, with an imposed turbulent velocity field with a power spectrum and Mach number $\mathcal{M} = 6.7$ consistent with observed motions in molecular clouds. As previously we initially adopted a barotropic equation of state (see above) to approximate the effect of the gas becoming optically thick, and thus heating and halting the collapse, beyond a certain critical density. The initially turbulent cloud was threaded with a uniform magnetic field, though with no external medium. Instead, SPH particles in the initially expanding outer layers of the cloud carry the magnetic field outwards and form a low density medium into which the field is anchored.
Despite the relatively weak field strengths with respect to gravity, magnetic fields were found to have a dramatic effect on the large scale structure of the clouds, as can be seen from the column density projection shown at $t/t_{ff} = 1.23$ in Fig.~\ref{fig:global}. This is because the fields are \emph{not} weak with respect to gas pressure, so the magnetic field is able to act as the dominant source of pressure within the cloud, producing large-scale magnetically-supported voids (middle and bottom rows) that are completely absent from purely hydrodynamical calculations (top row).
\begin{figure}
\centering
\includegraphics[width=\textwidth]{mclusterRT_all_global.ps}
\caption{Effect of magnetic fields and radiation on the large scale structure of star-forming 50 $M_{\odot}$ molecular cloud cores. Showing calculations with no magnetic fields (top row), a mass-to-flux ratio of 5 (middle row) and 3 times the critical value (bottom row). In the regime where magnetic pressure exceeds gas pressure the magnetic fields there is a dramatic influence on the global cloud structure, with the appearance of large-scale, magnetically supported voids. The large scale evolution of the cloud with radiative transfer explicitly calculated (right panels) is identical to that using an approximate, barotropic equation of state (left panels), at least for low mass star formation. }
\label{fig:global}
\end{figure}
The means by which magnetic fields are able to act as a source of pressure on large scales is relatively simple to understand, since in ideal MHD the gas motions are tied to the magnetic field lines. For a relatively strong field, this means that gas is channelled along field lines as it collapses (rather than the gas dragging the field lines around in the weak field case). Since in ideal MHD the mass-to-flux ratio is conserved along any given flux tube, any gas collapsing to form dense structures inevitably leaves behind a region evacuated of gas pressure but with the magnetic field strength (and magnetic pressure) unchanged. Thus the ratio of gas-to-magnetic pressure decreases substantially away from the densest gas. New material is prevented from re-entering the evacuated region because of the inability to cross magnetic field lines. Thus the region, once evacuated, remains as a magnetic-pressure supported void. At a recent meeting the above mechanism was paraphrased by Carl Heiles as ``magnetic fields abhor a vacuum'', since it is easy to remove gas from a region of space along the magnetic field lines, but the magnetic fields themselves will remain.
The effect of the support provided to the large scale regions of the cloud is a dramatic slow-down in the star formation rate with increasing magnetic field strength (discussed below, see Fig.~\ref{fig:sfr}), most effective in the regime where magnetic pressure exceeds gas pressure ($\beta < 1$) and independent of the fact that the field may be weak relative to gravity and/or turbulence. An unexpected finding from \citep{pb08} was the resultant change to the initial mass function of stars formed in the calculations, in the form of a reduction in the number of sub-stellar objects (i.e., brown dwarfs) relative to higher mass objects (i.e., stars). This occurs not because of some complicated influence of the magnetic fields on the fragmentation --- we do not resolve the magnetic fields structure on the smallest scales in these calculations --- but simply because of the overall slowdown in the star formation rate and a consequent reduction in the importance of dynamical interactions and the associated ejection of low mass objects from multiple systems. Given the low number of objects formed overall in the strong magnetic field calculations, it is not possible to state whether or not this effect is sufficient to resolve the statistical disagreement in the number of low mass objects and the observed IMF found by \citet{bate09a}, but the trend is certainly in the right direction.
In the strongest field calculation, we also found that the expanding outer regions of the cloud started to show a `striped' appearance as the gas was channelled along the magnetic field lines. This is strongly reminiscent of the `magnetically aligned striations' observed in the outer regions of the Taurus molecular cloud in the recent $^{12}$CO and $^{13}$CO molecular line emission maps by \citet{goldsmithetal08}, co-located with measurable velocity anisotropy aligned with the global magnetic field \citep{heyeretal08}. This is a good indication that Taurus lies in a regime where the magnetic field is able to exert considerable influence on the star formation process.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{mclusterRT_all.ps}
\caption{Combined effect of magnetic fields and radiative feedback on star cluster formation. The plots show a zoomed-in subsection of the clouds shown in Fig.~\ref{fig:global} at 1.23 initial gravitational free-fall times for the calculations of three different magnetic field strengths (top to bottom), without (left) and with (right) a full modelling of radiative transport in the gas. The effect on the small scale fragmentation is dramatic: Once the gas becomes optically thick to radiation the heating effect provided to neighbouring material completely inhibits any subsequent fragmentation within a radius of several AU. The result is a trend towards fewer but more massive stars and a further reduction in the overall star formation rate on top of the large-scale effect provided by the magnetic field.}
\label{fig:zoomin}
\end{figure}
\section{Influence of radiative feedback on star cluster formation}
A key limitation to all of the calculations discussed above was the approximate treatment of the thermodynamics of the gas via the use of a barotropic equation of state where gas pressure is a function of density alone, rather than being a function of density and temperature. Naturally this assumption simplifies the calculations considerably, but it misses important feedback processes, especially once the gas enters the optically thick regime. In particular, using the barotropic approximation the temperature is assumed to rise strictly with density, but this neglects the fact that radiation in actual fact should diffuse from the hot, dense, compressed gas into the less-dense surrounds, thus heating it and preventing it from fragmenting further.
In the initial phases of the cloud evolution and on the largest scales the radiation has very little influence, evident from Fig.~\ref{fig:global} which compares the cloud structures using the barotropic equation of state (left panels) with calculations incorporating the transport of radiation within the gas via the flux-limited diffusion approximation (right panels). This is partly the case because we form only low-mass stars in the calculations, but also because we have neglected both the accretion luminosity within the sink radius and the luminosity of the protostars themselves. Thus the effect of radiation that we consider is very much a lower limit to the true feedback effect. However, to capture as much of the radiative feedback as possible we have reduced the accretion radius on the sink particles to a mere $0.5$ AU in size, compared to $5$ AU sink radii in previous calculations \citep[e.g.][]{bbb03,pb08}. Other authors \citep{offneretal09} adopt a prescription for providing radiative feedback from the protostars themselves, but this brings a number of assumptions about the protostellar evolution process that at this stage we have preferred to avoid since it ultimately requires a detailed sub-grid stellar evolution model in order to constrain the parameters.
The effect provided by the transport of radiation from the hot, collapsing, compressed gas into neighbouring material does however have a dramatic impact on the small-scale fragmentation in the cloud. This is illustrated in Fig.~\ref{fig:zoomin} that shows a close up view of the star formation occurring in the 6 model clouds shown in Fig.~\ref{fig:global}. Whilst the effect of the magnetic field on the large scales is to reduce the overall rate at which the global cloud collapses, radiative feedback completely inhibits any secondary fragmentation in the material immediately surrounding the protostars (comparing left to right panels). The result is a reduction in the formation of low mass objects (visibly being ejected from the multiple systems resulting in the top left panel), particularly those that initially resulted from fragmentation in circumstellar discs which with the radiative feedback effect included become sufficiently heated such that no further sub-fragmentation occurs. Similarly in the middle panels at moderate field strength it may be observed that an object that initially fragmented into a binary system using the barotropic equation of state no longer fragments when radiative transport is accounted for, instead producing a single star with a circumstellar disc.
The reason for the dramatic reduction in small scale fragmentation that occurs when radiative feedback is easily understood from our above discussion, and may be readily illustrated by a plot of the integrated temperature $\int \rho T {\rm dz}/ \int \rho {\rm dz}$ shown in Fig.~\ref{fig:temp} for the hydrodynamic calculations using the barotropic equation of state (left panel) and with radiative feedback included (right panel). With the barotropic equation of state (left) the temperature is high only at several discrete points corresponding to where the gas density exceeds the threshold for the polytropic index to change from $\gamma = 1$ to $\gamma = 1.4$. Once the transport of radiation from hot to cold regions is modelled (right panel), a spatially extended region of high temperature gas is produced in the region surrounding each collapsing protostar, producing the effect on the fragmentation seen in Fig.~\ref{fig:zoomin}.
Our results regarding the effect on small-scale fragmentation produced by the transport of radiation have been confirmed by calculations employing similar physics performed by \citet{offneretal09} using an adaptive mesh refinement code.
The resultant effect on the initial mass function strengthens the trend already produced by the magnetic field, namely towards producing fewer and more massive objects. As previously stated, this is in the right direction to resolve the discrepancy with the observed IMF found by \citet{bate09a} in barotropic calculations, but given the very low number statistics --- particularly with the overall reduction in star formation rate produced by the combined influence of the magnetic fields and radiation --- we are reluctant to draw firm conclusions in this regard.
\begin{figure}
\centering
\includegraphics[angle=270,width=\textwidth]{hyd_temp.ps}
\caption{Effect of radiative feedback on star cluster formation: The plot shows the distribution of average temperature $\int \rho T {\rm dz}/ \int \rho {\rm dz}$ from the hydrodynamic calculations shown in the top row of Figs.~\ref{fig:global} and \ref{fig:zoomin} employing a barotropic (pressure-proportional to density) equation of state (left panel) compared to the calculation (right) where the radiation is explicitly modelled and thus the transport of radiation from hot to cold regions is captured. Whereas using the barotropic approximation only the material above the critical density becomes hot, in the radiation hydrodynamics calculation an extended region surrounding each protostar is heated and thus fragments no further (see Fig.~\ref{fig:zoomin}).}
\label{fig:temp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{sinkmass.ps}
\caption{Combined influence of magnetic fields and radiative feedback on the star formation rate in our $50 M_{\odot}$ model clouds. Line styles correspond to the four different magnetic field strengths employed: no magnetic fields (solid, black), and mass-to-flux ratios of 10, 5 and 3 in units of the critical value for preventing collapse altogether (dotted red, dashed blue and dot-dashed magenta lines respectively), whilst the line width shows whether (thick lines) or not (thin lines) radiative feedback was modelled (if not, a barotropic equation of state was employed). The magnetic field strength has the dominant influence on the star formation rate, with a secondary effect due to radiative feedback occurring at later times.}
\label{fig:sfr}
\end{figure}
\section{Combined influence of magnetic field and radiative feedback on the star formation rate and efficiency}
Having assessed the effect of both magnetic fields and radiative feedback on the star cluster formation, we may return to our original question: namely, are these two pieces of missing physics the necessary and sufficient ingredients required to explain the kind of slow and inefficient star formation observed in real molecular clouds?
The combined effect on the star formation rate is shown in Fig.~\ref{fig:sfr}, showing the total mass accreted onto the sink particles for the full suite of calculations as a function of time. The primary influence on the star formation rate is the strength of the initial magnetic field, since it affects the large scale structure of the cloud and thus the amount of material that is able to later collapse and form stars. Radiative feedback enters as a secondary effect, reducing the star formation rate further, particularly at later times as the radiation diffuses further from the protostars into the surrounding medium.
It is notable that only the calculations employing the strongest magnetic fields (mass-to-flux ratio of 3 in units of the critical value) produce a star formation rate that is even remotely close to the observed rate of 3-6\% per gravitational free-fall time found by \citet{evansetal09}: The rate in the strongest field calculation with radiative feedback is $0.18 M_{\odot} / 0.34 t_{ff} /50 M_{\odot} \approx 10\%$ per free-fall time. This is not unreasonable since molecular cloud cores are indeed observed to have mass-to-flux ratios of a few times the critical value, and radiative feedback is clearly an important effect. Relative differences in the star formation rate across the Galaxy can also be explained as being due to variations in the global flux threading individual star forming clouds.
We can speculate that the remaining discrepancy between our results and observational estimates of the star formation rate is due to our neglect of additional feedback processes, namely the intrinsic and accretion luminosity from the protostars themselves as well as mechanical feedback from jets and outflows.
The question of the overall star formation \emph{efficiency} in the presence of magnetic fields and radiation transport is more difficult to answer given the limited time for which we have been able to evolve the calculations beyond one free-fall time. Ideally one would continue the calculations over several global dynamical times until star formation activity has ceased, however this is currently prohibitively expensive in terms of CPU time. Observational estimates are limited in a similar manner because a star forming molecular cloud is defined as one in which star formation has initiated but not completed, and once completed one has little insight as to the initial mass of the parental cloud. If we assume that star formation continues at the rate observed in Fig.~\ref{fig:sfr} and that the molecular cloud survives for 2-3 free-fall times beyond star formation, then the overall star formation efficiency in the strongest field case would be of order 20-30\%. By contrast, for the calculations without magnetic fields the efficiency would be close to 100\% on a similar timescale. Since at supercritical mass-to-flux ratios the field is relatively weak compared to gravity, the fraction of bound gas at the end of the calculation remains relatively high even for the highest field strengths, of order 85\% for the mass-to-flux ratio of 3 (times critical) calculation with radiative feedback, so the main requirement for a low overall efficiency is that the cloud should be dispersed after several dynamical times and that the star formation rate should not accelerate considerably with time.
\section*{Acknowledgements}
DJP is supported by a Monash Fellowship, though much of this work was completed whilst funded by a UK Royal Society University Research Fellowship at the University of Exeter. We thank the organisers for their hospitality in both Milan and Como, the opportunity to attend and present at the conference.
|
\section{Introduction}
Recently, progress has been made on many-user approximations to the sum-capacity
$\Csum$ of random Gaussian interference networks.
In particular, in a 2009 paper, Jafar \cite{jafar} proved a result on the
asymptotic sum-capacity of a particular random Gaussian interference network:
\begin{theorem}[\cite{jafar}, Theorem 5]
Suppose direct $\SNR$s are fixed and identical, so $\SNR_i = \snr$ for all $i$,
and suppose that all $\INR$s are IID random and supported on some neighbourhood
of $\snr$. Then the average per-user capacity $\Csum/n$ tends in probability
to $\frac12 \log(1+2\snr)$ as $n\to\infty$.
\end{theorem}
(Here and elsewhere, we use $\Csum$ to denote the sum-capacity of the network,
and interpret $\Csum/n$ as the average per-user capacity.)
A subsequent result by the current authors \cite{johnson} concerned a more
physically realistic model:
\begin{theorem}[\cite{johnson}, Theorem 1.5]
Suppose receivers and transmitters are placed IID uniformly at random on the unit square $[0,1]^2$,
and suppose that signal power attenuates like a polynomial in $1/\mathrm{distance}$.
Then the average per-user capacity $\Csum/n$ tends in probability
to $\frac12 \EE \log(1+2\SNR)$ as $n\to\infty$.
\end{theorem}
In this paper, we prove a similar -- but more general -- result to Theorem 2, with a neater proof, using ideas
from Jafar's proof of Theorem 1. We assume transmitters and receivers are situated independently at random
in space (not necessarily uniformly), and that the power of signals depends in a natural way on the distance they travel.
Specifically our result is the following (full definitions of non-italicised technical terms are in Section II):
\begin{theorem} \label{thm:main}
Consider a Gaussian interference network formed by $n$ pairs of
nodes placed in an \emph{spatially-separated IID network} with \emph{power law attenuation}.
Then the average per-user capacity $\Csum/n$ converges in probability to
$\frac12 \EE \log(1 + 2 \SNR)$, in that for all $\epsilon > 0$
\[ \PP \left( \left| \frac{\Csum}{n} - \frac12 \EE \log(1 + 2 \SNR) \right| > \epsilon \right)
\to 0 \quad \text{as $n \to \infty$.} \]
\end{theorem}
The direct part of the proof uses interference alignment. Interference alignment is a new
way of dealing with interference in networks, particularly when that interference is of a similar
strength to the desired signal. Interference alignment allows communication at faster rates
than traditional resource division strategies such as time-division or frequency-divibysion
multiple-access. Two early papers on interference alignment are those by Maddah-Ali, Motahari
and Khandani \cite{motahari} and Cadambe and Jafar \cite{cadambejafar}.
Specifically, we take advantage of so-called ergodic interference alignment, developed by
Nazer, Gastpar, Jafar and Vishwanath \cite{nazer}.
The converse part of the proof uses the idea of `bottleneck links' developed by Jafar
\cite{jafar}. An information theoretic argument gives a capacity bound on such bottleneck
links, and probabilistic counting arguments show there are sufficiently many such links
to tightly bound the sum-capacity of the whole network.
A different approach towards finding the capacity of large communications
networks is given by the deterministic approach of Avestimehr, Diggavi and Tse \cite{deterministic}. This paper
shows how capacities can be calculated up to a gap determined by the
number of users $n$, across all values of $\snr$. However, we identify a sharp
limit as the number of users $n$ tends to infinity.
A wider literature review is available in our previous paper \cite{johnson}.
The plan of this paper is as follows: In Section II we define our network model. In
Section III we prove the direct part of Theorem 3. Our main contribution comes in Section
IV where we use new ideas to prove the converse part of Theorem 3. We conclude in Section V.
\section{Model}
\subsection{Node position model}
We believe that our techniques should work in a variety of models for the node positions. We outline
one very natural
scenario here.
These ideas were introduced in our earlier paper \cite{johnson}, but were not
fully exploited, due to that paper's concentration on the uniform case.
\begin{definition} \label{def:nodeplace}
Consider two probability distributions $\PP_T$
and $\PP_R$ defined on $d$-dimensional space $\RR^d$. Given an integer $n$, we sample the
$n$ transmitter node positions
$\vec T_1, \ldots, \vec T_n$ independently from the distribution $\PP_T$. Similarly, we sample
the $n$ receiver node $\vec R_1, \ldots, \vec R_n$ positions independently from distribution $\PP_R$. We
refer to such a model of node placement as an `IID network'.
\end{definition}
Equivalently, we could state that transmitter and receiver positions are distributed according to two independent
(non-homogeneous) Poisson processes, conditioned such that there are $n$ points
of each type.
We pair the transmitter and receiver nodes up so that transmitter $i$ at $\vec T_i$ wishes to communicate
with receiver $i$ at $\vec R_i$ for each $i$.
We make the following definition:
\begin{definition} \label{def:spatsep}
Let $\vec T \sim \PP_T$ and $\vec R \sim \PP_R$ be placed independently in $\RR^d$.
We say the IID network is spatially separated if there exists constants $\beta$
and $\csep$ such that for all $\rho$
\[ \PP( \| \vec T - \vec R \| \leq \rho) \leq \csep \rho^{\beta} . \]
\end{definition}
In particular, it can be shown \cite[proof of Lemma 2.2i]{johnson} that the standard dense network
is spatially separated.
(The $d$-dimensional standard dense network is defined by $\PP_T$ and $\PP_R$ being
independent uniform measures on $[0,1]^d$.)
The standard dense network has been the subject of much research (see for example
the review paper of Xue and Kumar \cite{xuekumar} and references therein). However, we emphasise
that our result holds for a wider range of models.
\subsection{Transmission model}
Our results are in the context of so-called `line of sight'
communication models, without multipath interference. That is,
we consider a model where signal strengths attenuate
deterministically with distance according to
some function $a$.
The definitions in this section are adapted from our previous paper \cite{johnson}.
\begin{definition}
\label{def:transprot}
Fix transmitter node positions $\{ \vec T_1, \ldots, \vec T_n \}
\in \RR^d$ and receiver node positions
$\{\vec R_1, \ldots, \vec R_n \} \in \RR^d$, and consider Euclidean distance
$\| \quad \|$ and an attenuation function $a$.
We define $\SNR_i = a(\|\vec T_i-\vec R_i\|)$, and for all pairs with $i \neq j$, define
$\INR_{ij} = a(\|\vec T_i-\vec R_j\|)$.
\end{definition}
We consider the $n$-user Gaussian interference network defined so that
transmitter $i$ sends a message encoded as a string of $T$ complex
numbers $\vec{x}_i = (x_i[1], \ldots, x_i[T])$ to receiver $i$, under a power constraint
$\frac{1}{T} \sum_{t=1}^T |x_i[t]|^2
\leq 1$ for each $i$.
Our result requires that the fading random variables be circularly symmetric.
For definiteness, we hold the modulus constant and choose the argument uniformly
at random. (We discuss Rayleigh fading in Section V.) So the $t$th symbol received at receiver $j$ is given as
\begin{multline} \label{eq:transmod} Y_j[t] = \exp(\ii \Theta_{jj}[t])
\sqrt{\SNR_{j}} x_j[t] \\
+ \sum_{i\neq j} \exp(\ii \Theta_{ij}[t])
\sqrt{\INR_{ij}} x_i[t] + Z_j[t],
\end{multline}
where noise terms $Z_j[t]$ are independent standard complex Gaussian random variables,
and the phases $\Theta_{ij}[t]$ are
independent $U[0,2\pi]$ random variables independent of all other terms.
The $\INR_{ij}$ and $\SNR_i$ remain fixed over time, since the node positions themselves are fixed,
but the phases are fast-fading, in that they are renewed for each $t$.
\begin{definition} \label{def:decay}
We say an attenuation function $a$ has power law attenuation if there exist constants
$\alpha$ and $\cdec$ such that for all $\rho$
$ a(\rho) \leq \cdec \rho^{-\beta} $
\end{definition}
(Tse and Viswanath \cite[Section 2.1]{tsevis} discuss a variety of models under which
power law attenuation is an appropriate model for different exponents $\alpha$.)
For brevity, we write $S_{ij}$ for the random variables $\frac12 \log(1 + 2 \INR_{ij})$ (when $i \neq j$),
and $S_{ii}$ for $\frac12 \log(1 + 2 \SNR_{i})$
which are functions of the distance
between the transmitters and receivers.
In particular, since the nodes are positioned
independently, under this model
the random variables $S_{ij}$ are identically distributed, and $S_{ij}$ and
$S_{kl}$ are IID when $\{i,j\}$ and $\{k,l\}$ are disjoint.
We will also write
$E = \EE S_{ii} = \frac12 \EE \log(1 + 2 \SNR)$,
noting that this is independent of $i$. (It is also true that
$E = \EE S_{ij}$ for all $i$ and $j$.)
\section{Proof: direct part}
We can now prove our main theorem, Theorem 3, by breaking the probability into two terms which we deal
with separately. So
\begin{multline} \label{eq:unionbd}
\PP \left( \left| \frac{\Csum}{n} - E \right| > \epsilon \right) \\
= \PP \left( \frac{\Csum}{n} - E < -\epsilon \right) +
\PP \left( \frac{\Csum}{n} - E > \epsilon \right).\end{multline}
Bounding the first term of (\ref{eq:unionbd}) corresponds to the direct part of the proof.
Bounding the second term of (\ref{eq:unionbd}) corresponds to the converse part, and
represents our major contribution.
We prove the direct part as previously \cite{johnson}.
\begin{IEEEproof}
The first term of (\ref{eq:unionbd})
can be bounded relatively simply, using an achievability
argument based on an interference alignment scheme presented
by Nazer, Gastpar, Jafar and Vishwanath \cite{nazer}.
Their theorem \cite[Theorem 3]{nazer} implies that the rates $R_i =
1/2 \log(1 + 2 \SNR_i) = S_{ii}$ are simultaneously achievable. This implies that
$\Csum \geq \sum_{i=1}^n R_i = \sum_{i=1}^n S_{ii}$. This allows us to bound the first term in
(\ref{eq:unionbd}) as
\[ \PP \left( \frac{\Csum}{n} - E < -\epsilon \right)
\leq \PP \left( \frac{\sum_{i=1}^n S_{ii}}{n} < E-\epsilon \right).
\label{eq:lln}
\]
But $E = \EE S_{ii}$, so this probability tends to $0$ by the weak law of large numbers.
\end{IEEEproof}
\section{Proof: converse part}
We now need to show that the second term of (4) tends to $0$ too.
Specifically, we must prove the following: for all $\epsilon > 0$
\begin{equation} \label{converse}
\Prob \left( \frac{C_\Sigma}{n} \geq E + \epsilon \right)
\to 0
\end{equation}
as $n \to \infty$.
The proof of the converse part is the major new part of this paper.
First, bottleneck links are introduced, and we prove a tight information-theoretic
bound on the capacity of such links. Second, a probabilistic counting argument
ensures there are (with high probability) sufficiently many bottleneck links to bound
the sum-capacity of the entire network.
\subsection{Bottleneck links}
The important concept is that of the bottleneck link, an idea first used
by Jafar \cite{jafar} and later adapted \cite{johnson} in the following form:
\begin{definition}
We say the link $i\link j$, $i\neq j$, is a \emph{$\epsilon$-bottleneck
link}, if the the following three conditions hold:
\begin{description}
\item[B1:] $S_{ii} \leq E + \epsilon/2$,
\item[B2:] $S_{ij} \leq E + \epsilon/2$,
\item[B3:] $S_{jj} \leq S_{ji}$.
\end{description}
We let $B_{ij}$ be the indicator function that the crosslink $i\link j$ is a $\epsilon$-bottleneck
link. We also define the \emph{bottleneck probability} $\beta :=
\Ex B_{ij}$ to be the probability that a given link is an $\epsilon$-bottleneck which is
independent of $i$ and $j$ for an IID network. (We suppress the $\epsilon$ dependence for simplicity.)
\end{definition}
The crucial point about bottleneck links is the constraints they place
on achievable rates in a network.
\begin{lemma} \label{links}
Consider a crosslink $i\link j$ in a $n$-user Gaussian interference
network. If $i\link j$ is a $\epsilon$-bottleneck link, then the
sum of their achievable transmission rates is bounded by
$r_i + r_j \leq 2E + \epsilon$.
\end{lemma}
\begin{IEEEproof}
First, note that we make things no worse by considering the two-user
subnetwork:
\begin{align*}
Y_i &= \exp (\ii\Theta_{ii}) \sqrt{\SNR_i}X_i
+ \exp (\ii\Theta_{ji}) \sqrt{\INR_{ji}}X_j + Z_i \\
Y_j &= \exp (\ii\Theta_{ij}) \sqrt{\INR_{ij}}X_i
+ \exp (\ii\Theta_{jj}) \sqrt{\SNR_j}X_j + Z_j
\end{align*}
where receiver $i$ needs to determine signal $X_i$, and receiver $j$
signal $X_j$. (The time index is ommited for clarity.)
From bottleneck conditions B1 and B2 we have
\[ 1 + 2\SNR_i \leq \exp (2E + \epsilon) , \quad
1 + 2\INR_{ji} \leq \exp (2E + \epsilon) . \]
Summing and taking logs gives
\begin{equation} \label{logs}
\log (1+\SNR_i+\INR_{ji}) \leq 2E + \epsilon .
\end{equation}
We combine this with the argument given by Jafar \cite{jafar}.
Let $r_i$ and $r_j$ be jointly achievable rates for the subnetwork.
In particular, receiver $i$ can determine signal $X_i$ with an arbitrarily
low probability of error.
We certainly do no worse if a genie presents signal $X_i$ to receiver
$j$ -- so assume $j$ can indeed recover $X_i$. But condition B3
ensures that it is easier for receiver $i$ to determine $X_j$ than
it is for receiver $j$ (since the weighting is larger in the first
case). So since receiver $j$ can recover $X_j$ (as $r_j$ is achievable),
receiver $i$ can recover $X_j$ also.
Because receiver $i$ can determine both $X_i$ and $X_j$, these
two signals must have been transmitted at a sum-rate no higher
than the sum-capacity of the Gaussian multiple-access channel
(see, for example, Cover and Thomas \cite[Section 14.3.6]{cover}).
Hence,
\[ r_i + r_j \leq \log(1+\SNR_i+\INR_{ji}) \leq 2E + \epsilon, \]
where the second inequality comes from \eqref{logs}.
\end{IEEEproof}
\subsection{Three technical lemmas}
A few technical lemmas are required in order to prove
\eqref{converse}.
First, we need to ensure that very high $\SNR$s are very rare (Lemma \ref{SNRs}).
Second, we need to show that bottleneck links will actually occur (Lemma \ref{beta}).
Last, we must show that the number of bottleneck links cannot vary
too much (Lemma \ref{Var}).
Under any network model where these three lemmas are true, our theorem will hold.
We emphasise that our model of IID networks with power law attenuation is one such
model; we believe the result holds more widely.
\begin{lemma} \label{SNRs}
Consider a spatially-separated IID network, with power law attenuation. Then for any $\eta > 0$,
\[ \Prob \bigg(\max_{1\leq i\leq n} S_{ii} > n^{\eta/2}\bigg) = O(n^{-1})
\quad \text{as $n \to \infty$.} \]
\end{lemma}
In fact, in our case the convergence to $0$ is considerably quicker than
$O(n^{-1})$, but this is sufficient.
It is worth noting that this fast decay in the tails of $S_{ii}$ ensures that the
expectation $E = \EE S_{ii}$ does indeed exist and is finite.
\begin{IEEEproof}
First, we have by the union bound
\[ \Prob (\max S_{ii} > n^{\eta/2}) \leq n \Prob ( S_{11} > n^{\eta/2}) . \]
But by the definition of $S_{11}$
\begin{align*}
\Prob ( S_{11} > n^{\eta/2})
&= \Prob \left( \SNR_{11} > \frac12 (2^{2n^{\eta/2}}-1) \right) \\
&= \Prob \left( a(\| \vec T_1 - \vec R_1 \|) > \frac12 (2^{2n^{\eta/2}}-1) \right)
\end{align*}
and the proof follows by applying the definitions of $\SNR$, spatial separation and
power law attenuation.
\end{IEEEproof}
We will often condition off this event; that is, condition on the complementary event
$\{ \max S_{ii} \leq n^{\eta/2} \}$. We use $\PP_n$, $\EE_n$ and $\Var_n$ to
denote such conditionality, and write $\beta_n=\EE_nB_{ij}$ for the
conditional bottleneck probability.
The next two lemmas concern showing that conditional probabilities are
nonzero. However, we have for any event $A$,
\begin{align*}
\Prob(A) &= \Prob(A \mid \max S_{ii} \leq n^{\eta/2}) \Prob(\max S_{ii} \leq n^{\eta/2}) \\
& \qquad\qquad {}+ \Prob(A \mid \max S_{ii} > n^{\eta/2}) \Prob(\max S_{ii} > n^{\eta/2}) .
\end{align*}
and hence by Lemma \ref{SNRs} we have the bounds
\begin{align*}
\Prob(A) &\leq \Prob(A \mid \max S_{ii} \leq n^{\eta/2}) + \Prob(\max S_{ii} > n^{\eta/2}) \\
&= \Prob_n(A) + O(n^{-1})
\end{align*}
and
\begin{align*}
\Prob(A) &\geq\Prob(A \mid \max S_{ii} \leq n^{\eta/2}) \Prob(\max S_{ii} \leq n^{\eta/2}) \\
&= \Prob_n(A) \big(1-O(n^{-1})\big),
\end{align*}
and so $\Prob(A) = \Prob_n(A) + O(n^{-1})$. This will be useful in the next two proofs.
\begin{lemma} \label{beta}
Consider a spatially-separated IID network, with power law attenuation. Then the conditional
bottleneck probability $\beta_n$ is bounded away from $0$ for
all $n$ sufficiently large.
\end{lemma}
\begin{IEEEproof}
First note that by the comment above, we need only show that the
unconditional bottleneck probability $\beta$ is nonzero.
Second, note that by the exchangeability of $\vec R_i$ and $\vec R_j$,
we have
$ \Prob(\text{B1 and B2 and B3}) \geq \frac12 \Prob(\text{B1 and B2}) $.
It is left to show that $\Prob(\text{B1 and B2})$ is non-zero.
Note that B1 requires $S_{ii}$ to be less than its expectation plus $\epsilon$.
So $\vec R_i$ must be situated such that this has nonzero probability. So $\vec T_i$
has a nonzero probability of being positioned such that B1 occurs. But
$\vec T_i$ and $\vec T_j$ are also exchangeable, so we are done.
\end{IEEEproof}
\begin{lemma} \label{Var}
Consider a spatially-separated IID network, with power law attenuation.
Then, conditional on $\{\max_i S_{ii} < n^{\eta/2} \}$,
\[ \Var_n (\# \text{\emph{ bottleneck links}})
= \Var_n \left( \sum_{i\neq j} B_{ij} \right)
= O(n^3) , \]
where the sum is over all crosslink pairs $(i,j)$, $i\neq j$.
\end{lemma}
In general, one might assume that $\Var_n (\# \text{ bottleneck links})$
would be proportional to the total number of links, and thus be $O(n^4)$.
However, because of the independences in the IID network, the variance is in fact
much lower.
\begin{IEEEproof}
First consider the unconditional version. We have
\[ \Var \left( \sum_{i\neq j} B_{ij} \right)
= \sum_{i\neq j} \sum_{k\neq l} \Cov(B_{ij}, B_{kl}) . \]
The important observation is that for $i,j,k,l$ all distinct,
$B_{ij}$ and $B_{kl}$ are independent giving $\Cov(B_{ij}, B_{kl}) = 0$.
(This is because they depend only on the position of distinct and
independently-positioned nodes.) Hence there are only $O(n^3)$ non-zero
terms in the sum, each of which is trivially bounded by
$\frac12 (1-\frac12) = \frac14$.
But by the comment above, if $\Cov(B_{ij}, B_{kl}) = 0$, then the
conditional covariance is $\Cov_n(B_{ij}, B_{kl}) = O(n^{-1})$.
Hence,
\[ \Var_n \left( \sum_{i\neq j} B_{ij} \right) \leq O(n^3)\frac14 + O(n^4)O(n^{-1}) = O(n^3), \]
as desired.
\end{IEEEproof}
\subsection{Completing the proof of Theorem 3}
We are now in a position to prove \eqref{converse}, and hence prove
Theorem 3.
\begin{IEEEproof}
We need to show
\[
\forall\, \epsilon > 0 \ \forall\, \delta>0 \ \exists\, N
\ \forall\, n\geq N \quad
\Prob\left( \frac{C_\Sigma}{n} \geq E + \epsilon \right)
\leq \delta.
\]
So choose $\epsilon>0$, $\delta>0$, fix $n \geq N$ (where $N$ will
be determined later), and pick a rate vector $\mathbf{r}
\in \bR_+^n$ with sum-rate
\begin{equation} \label{sumrate}
\frac{r_\Sigma}{n} > E + \epsilon ;
\end{equation}
we need to show that $\Prob(\mathbf r\text{ is achievable}) \leq \delta$.
(Here, we are writing $r_\Sigma := \sum_{i=1}^n r_i$ for the sum-rate.)
We divide into two cases: when there is a very high $\SNR$, which is unlikely
to happen; and when there is not, in which case $\mathbf r$ is unlikely to be
achievable. Formally,
\begin{align}
&\Prob(\mathbf r \text{ achievable}) \notag \\
&\ = \Prob(\mathbf r\text{ achievable} \given \max S_{ii} \leq n^{\eta/2})
\Prob(\max S_{ii} \leq n^{\eta/2}) \notag \\
&\ \quad\ {}+ \Prob(\mathbf r\text{ achievable} \given \max S_{ii} > n^{\eta/2})
\Prob(\max S_{ii} > n^{\eta/2}) \notag \\
&\ \leq \Prob(\mathbf r\text{ achievable} \given \max S_{ii} \leq n^{\eta/2})
+ \Prob(\max S_{ii} > n^{\eta/2}) \notag \\
&\ \leq \Prob_n (\mathbf r\text{ achievable} )
+ \frac{\delta}{2} , \label{bound}
\end{align}
for $N$ sufficiently large, by Lemma \ref{SNRs}. We need to bound the first term in
\eqref{bound}.
First, note that our assumption on $\max_i S_{ii}$ means that if $r_i > 2n^{\eta/2}$,
than we break the single-user capacity bound, since we would have
\begin{multline*}
r_i > 2n^{\eta/2} \geq 2 \max_j S_{jj} \\
\geq 2 S_{ii} = \log (1+2\SNR_i) > \log (1+\SNR_i)
\end{multline*}
meaning $\mathbf r$ is not achievable, and we are done. Thus we assume this does
not hold; that
\begin{equation} \label{boundR}
r_i \leq 2n^{\eta/2} \quad \text{for all $i$}.
\end{equation}
(The rest of our argument closely follows Jafar \cite{jafar}.)
Now, if $\mathbf r$ is achievable, it must at least satisfy the constraints on the
$\epsilon$-bottleneck links $i\link j$ from Lemma \ref{links}, and hence also the sum of those constraints. So
\begin{align}
\Prob_n(\mathbf r &\text{ achievable}) \notag \\
&\quad \leq \Prob_n ( r_i + r_j \leq 2E + \epsilon
\text{ on bottleneck links $i\link j$} ) \notag \\
&\quad \leq \Prob_n \left( \sum_{i\neq j} B_{ij} (r_i + r_j)
\leq \bigg( \sum_{i\neq j} B_{ij} \bigg) (2E + \epsilon) \right) \notag \\
&\quad = \Prob_n (U \leq V) , \label{UV}
\end{align}
where we have defined
\begin{align*}
U &:= \frac{1}{n(n-1)} \sum_{i\neq j} B_{ij} (r_i + r_j) , \\
V &:= \frac{1}{n(n-1)} \bigg( \sum_{i\neq j} B_{ij} \bigg) (2E + \epsilon) .
\end{align*}
The conditional expectations of $U$ and $V$ are
\[ \Ex_n U = 2\beta_n \frac{r_\Sigma}{n}, \quad
\Ex_n V = \beta_n (2E + \epsilon)
= 2\beta_n \left(E + \frac{\epsilon}{2} \right). \]
Note that since $\beta_n > 0$ by Lemma 6, we can rewrite
\eqref{sumrate} as
$ \Ex_n U > \Ex_n V + \beta_n\epsilon $ ,
or equivalently,
\[ \Ex_n U - \frac{\beta_n\epsilon}{2} > \Ex_n V + \frac{\beta_n\epsilon}{2} . \]
The proof is completed by formalising the following idea: since
the expectations are ordered $\Ex_n U > \Ex_n V$, we can only rarely
have the opposite ordering $U < V$. Hence the expression in \eqref{UV} is small.
Formally, by (the conditional version of) Chebyshev's inequality and the union bound, we have
\begin{align}
&\Prob_n ( U \leq V) \notag \\
&\quad \leq \Prob_n \bigg( U \leq \Ex_n U - \frac{\beta_n\epsilon}{2}
\text{ or } V' \geq \Ex_n V + \frac{\beta_n\epsilon}{2} \bigg) \notag \\
&\quad\leq \Prob_n \bigg( |U - \Ex_n U| \geq \frac{\beta_n\epsilon}{2} \bigg)
+ \Prob_n \bigg( |U - \Ex_n U| \geq \frac{\beta_n\epsilon}{2} \bigg) \notag \\
&\quad\leq \bigg(\frac{2}{\beta_n\epsilon}\bigg)^2 \Var_n U + \bigg(\frac{2}{\beta_n\epsilon}\bigg)^2 \Var_n V
\notag \\
&\quad = \frac{4}{\beta_n^2\epsilon^2} ( \Var_n U + \Var_n V ) . \label{UV2}
\end{align}
Using Lemma \ref{Var} we can bound these variances as
\begin{align*}
\Var_n U &= \frac{1}{n^2(n-1)^2} \Var_n \left( \sum_{i\neq j} B_{ij} (r_i + r_j) \right) \\
&\leq \frac{1}{n^2(n-1)^2} O(n^3) 16 n^\eta = O(n^{-(1-\eta)}). \\
\Var_n V &= \frac{1}{n^2(n-1)^2} O(n^3) (2E+\epsilon)^2 = O(n^{-1}) ,
\end{align*}
where we used \eqref{boundR} to bound $\Var_n U$. Choosing $\eta$ to be
less than $1$, we can ensure $N$ is sufficiently large that for all
$n \geq N$
\[ \Var_n U + \Var_n V \leq \frac{\beta_n^2 \delta \epsilon^2}{8} . \]
\noindent This makes \eqref{UV2} $ \Prob_n (U \leq V) < \delta/2$.
Together with \eqref{UV} and \eqref{bound}, this yields the result.
\end{IEEEproof}
\section{Conclusion}
In this paper we have defined IID interference networks with power law attenuation. We have shown
that this setup fulfils necessary properties for the average per-user capacity
$\Csum/n$ to tend in probability to $\frac12 \EE \log(1 + 2 \SNR)$. We have
also noted that this result is not unique to our setup.
\newpage
We briefly mention one more example.
Suppose Rayleigh fading is added to our model. That is, now let
$\SNR_i := |H_{ii}|^2 a(\|\vec T_i-\vec R_i\|)$ and
$\INR_{ij} := |H_{ij}|^2 a(\|\vec T_i-\vec R_j\|)$, where the $H_{ij}$
are IID standard complex Gaussian random variables.
Because ergodic interference alignment still works with Rayleigh fading \cite{nazer},
the direct part of the theorem still holds. But also, because the fading coefficients
are IID, the independence structure from the non-fading case remains, ensuring Lemmas
5--7 hold. Hence, the theorem is still true.
Characterising all networks for which such a limit for average per-user capacity
exists is an open problem.
At the moment, Theorem 3 should perhaps be regarded as being of
theoretical interest. That is, our major contribution is to provide a sharp
upper bound on the performance of interference networks. However, the
lower bound relies on an ergodic interference alignment \cite{nazer} which, while
rigorously proved, may not be feasible to implement in practice for large
number of users. Examination of the proof of the effectivenes of ergodic
interference alignment \cite[Theorem 1]{nazer} shows that,
even for a model with alphabet size $q$, the channel needs to be used
$O( (q-1)^{K^2})$ times. Even for $K \approx 10$, this is a prohibitive
requirement.
However, recent work by the current authors \cite{delay} characterises the delay--rate tradeoff
for ergodic interference alignment.
Also, note that for $K=3$, El Ayach, Peters, and Heath \cite{ayach}
have shown that the interference alignment scheme of Cadambe and Jafar \cite{cadambejafar} can perform
close to the theoretical bounds.
\section*{Acknowledgments}
M. Aldridge and R.~Piechocki thank Toshiba Telecommunications Research
Laboratory and its directors for supporting this work. The authors thank
Justin Coon and Magnus Sandell of Toshiba for their advice and support with this
research.
|
\section{Introduction}
The main numerical techniques for solving problems related to elliptic linear
partial differential equations with variable coefficients in one way or
another involve a discretization of a domain and solution of systems of
thousands of algebraic equations. Seldom the method of separation of variables
is applied due to its natural limitations related to the requirements of a
complete agreement between the geometry of the domain and the symmetry of the
coefficients. Moreover, the method of separation of variables implies solution
of Sturm-Liouville spectral problems which is not an easy task itself.
In the present paper we propose a different method based on some old and new
results from pseudoanalytic function theory \cite{Berskniga}, \cite{APFT}. Its
applicability is not so universal as the applicability of the finite
difference method or the finite element method. First of all, it is applicable
to problems in bounded domains in the plane and up to now only for the
operator $\operatorname{div}p\operatorname{grad}+q$. Moreover, at present we
can apply the method only when the equation
\begin{equation}
\left( \operatorname{div}p\operatorname{grad}+q\right)
u(x,y)=0\label{maineqintro}%
\end{equation}
possesses a particular solution $u_{0}$ such that the function $f=p^{1/2}%
u_{0}$ is sufficiently smooth, nonvanishing in the domain of interest and
representable in the form $f=S(s)T(t)$ where $s$ and $t$ represent an
orthogonal coordinate system. This separable form of $f$ may cause
associations with the method of separation of variables. Nevertheless this is
a completely different technique, based on different ideas and free of the
mentioned above limitations of the method of separation of variables.
The heart of the method is the construction of a complete system of solutions
for (\ref{maineqintro}) in the domain of interest, complete in the sense
explained below (see \cite[Sect. 1.3]{Colton} for related ideas and additional
details). The system of solutions is used for approximating the solution of a
boundary value problem. Due to the linearity of equation (\ref{maineqintro})
after the construction of the system of solutions the problem reduces to
approximation of boundary conditions, for which a variety of methods can be
used. Here we apply the collocation method.
The complete system of solutions is constructed in the following way. The
knowledge of a particular solution of (\ref{maineqintro}) allows us to propose
a corresponding Vekua equation \cite{Vekua} closely related to
(\ref{maineqintro}) in the sense that the real part of any of its solutions
has the form $p^{1/2}u$ where $u$ is a solution of (\ref{maineqintro}), and
vice versa given $u$ one can easily construct a corresponding solution of the
Vekua equation \cite{Krpseudoan}, \cite{APFT}. The relation between
(\ref{maineqintro}) and the Vekua equation is similar to the relation between
the Laplace equation and the Cauchy-Riemann system. L. Bers developed
\cite{Berskniga}, \cite{BersFormalPowers} a theory of so-called pseudoanalytic
formal powers. They are generalizations of the analytic powers $(z-z_{0})^{n}$
in the sense that they are solutions of the corresponding Vekua equation and
behave locally like the analytic powers. The theory of Bers includes
generalizations of Taylor series, Runge's theorem and other basic facts from
analytic function theory. Thus, under certain quite natural conditions the
system of pseudoanalytic formal powers is complete in the space of all
pseudoanalytic functions (solutions of the Vekua equation) in the same sense
as the system of powers $(z-z_{0})^{n}$ is complete in the space of analytic
functions. To construct the pseudoanalytic formal powers the knowledge of a
corresponding generating sequence is required. Recently
\cite{KrRecentDevelopments}, \cite{APFT} an algorithm for construction of
generating sequences under additional conditions on the coefficients in the
Vekua equation was proposed. This implies that when $f=p^{1/2}u_{0}$ is
representable in a separable form the complete system of formal powers for the
Vekua equation associated with (\ref{maineqintro}) can be constructed
explicitly following Bers' recursive procedure.
We investigate the efficiency of the proposed method which we call MPFP, the
Method of Pseudoanalytic Formal Powers. We show its fast convergence and
compare its accuracy with that of the finite element method. In general, we
show that in problems addmitting the explicit construction of formal powers
and hence the application of the MPFP its use is advantageous compared to
other computational techniques based on discretization of the problem.
It is worth mentioning that the MPFP is a direct generalization of the method
of harmonic polynomials for solving boundary value problems for the Laplace
equation which has been considered in dozens of works (see, e.g.,
\cite{Cannon}, \cite{Genev1984}, \cite{Hozejovski}, \cite{Suetin}). Indeed, in
a special case when $p\equiv1$, $q\equiv0$ and $u_{0}\equiv1$ the
corresponding complete system of solutions constructed by means of the MPFP
coincides with the system of harmonic polynomials $\left\{ \operatorname{Re}%
(z-z_{0})^{n},\quad\operatorname{Im}(z-z_{0})^{n}\right\} _{n=0}^{\infty}$.
The knowledge of a complete system of solutions for an equation corresponding
to any value of a spectral parameter allows one to use it for solving
eigenvalue problems. We consider this possibility in section \ref{SectEigen}.
The numerical results are highly promising, and it is clear that in the case
of eigenvalue problems as well as for boundary value problems further work
should be done in investigation of optimal ways of application of the MPFP.
For example, for solving eigenvalue problems by means of the MPFP we used the
simplest possible idea reducing the problem to calculation of zeros of a
certain determinant obtained by evaluating the first $N$ solutions from the
constructed complete system in $N$ points on the boundary of the domain under
consideration. Meanwhile, in principle, this natural approach works there
exist other techniques offering different ways of using the available exact
solution systems (see \cite[Sect. 1.13]{Alexidze}, where similar questions are discussed).
\section{Factorization of the operator $\operatorname{div}p\operatorname{grad}%
+q$.\label{SectFactorization}}
Let $\Omega$ be a domain in $\mathbf{R}^{2}$. Throughout the whole paper we
suppose that $\Omega$ is a simply connected domain. Denote $\partial
_{\overline{z}}=\frac{1}{2}\left( \frac{\partial}{\partial x}+i\frac
{\partial}{\partial y}\right) $ and $\partial_{z}=\frac{1}{2}\left(
\frac{\partial}{\partial x}-i\frac{\partial}{\partial y}\right) $. By $C$ we
denote the operator of complex conjugation.
Note that the operator $\partial_{\overline{z}}$ applied to a real valued
function $\varphi$ can be regarded as a kind of gradient, and if we know that
$\partial_{\overline{z}}\varphi=\Phi$ in a whole complex plane or in a convex
domain, where $\Phi=\Phi_{1}+i\Phi_{2}$ is a given complex valued function
such that its real part $\Phi_{1}$ and imaginary part $\Phi_{2}$ satisfy the
equation
\begin{equation}
\partial_{y}\Phi_{1}-\partial_{x}\Phi_{2}=0, \label{casirot}%
\end{equation}
then we can reconstruct $\varphi$ up to an arbitrary real constant $c$ in the
following way%
\[
\varphi(x,y)=2\left( \int_{x_{0}}^{x}\Phi_{1}(\eta,y)d\eta+\int_{y_{0}}%
^{y}\Phi_{2}(x_{0},\xi)d\xi\right) +c
\]
where $(x_{0},y_{0})$ is an arbitrary fixed point in the domain of interest.
Note that this formula can be easily extended to any simply connected domain
by considering the integral along an arbitrary rectifiable curve $\Gamma$
leading from $(x_{0},y_{0})$ to $(x,y)$%
\begin{equation}
\varphi(x,y)=2\left( \int_{\Gamma}\Phi_{1}dx+\Phi_{2}dy\right) +c.
\label{Antigr}%
\end{equation}
By $\overline{A}$ we denote the integral operator in (\ref{Antigr}):%
\[
\overline{A}[\Phi](x,y)=2\left( \int_{x_{0}}^{x}\Phi_{1}(\eta,y)d\eta
+\int_{y_{0}}^{y}\Phi_{2}(x_{0},\xi)d\xi\right) +c.
\]
Thus if $\Phi$ satisfies (\ref{casirot}), there exists a family of real valued
functions $\varphi$ such that $\partial_{\overline{z}}\varphi=\Phi$, given by
the formula $\varphi=\overline{A}[\Phi]$.
The following result is in the core of the method proposed in the present work.
\begin{theorem}
\cite{KrJPhys06} \label{ThFactGenSchr}Let\ $p$ and $q$ be real valued
functions, $p\in C^{2}(\Omega)$ and $p\neq0$ in $\Omega$, $u_{0}$ be a
positive particular solution of the equation
\begin{equation}
(\operatorname{div}p\operatorname{grad}+q)u=0\text{\qquad in }\Omega\text{.}
\label{maineq}%
\end{equation}
Then for any real valued continuously twice differentiable function $\varphi$
the following equality holds%
\begin{equation}
\frac{1}{4}(\operatorname{div}p\operatorname{grad}+q)\varphi=p^{1/2}\left(
\partial_{z}+\frac{f_{\overline{z}}}{f}C\right) \left( \partial
_{\overline{z}}-\frac{f_{\overline{z}}}{f}C\right) p^{1/2}\varphi,
\label{mainfact}%
\end{equation}
where
\begin{equation}
f=p^{1/2}u_{0}. \label{fandu}%
\end{equation}
\end{theorem}
\begin{remark}
Let $q\equiv0$. Then $u_{0}$ can be chosen as $u_{0}\equiv1$. Hence
(\ref{mainfact}) gives us the equality
\[
\frac{1}{4}\operatorname{div}(p\operatorname{grad}\varphi)=p^{1/2}\left(
\partial_{z}+\frac{\partial_{\overline{z}}p^{1/2}}{p^{1/2}}C\right) \left(
\partial_{\overline{z}}-\frac{\partial_{\overline{z}}p^{1/2}}{p^{1/2}%
}C\right) (p^{1/2}\varphi).
\]
\end{remark}
Let $f$ be a real function of $x$ and $y$. Consider the Vekua equation
\begin{equation}
W_{\overline{z}}=\frac{f_{\overline{z}}}{f}\overline{W}\text{\qquad in }%
\Omega\text{.}\label{Vekuamain}%
\end{equation}
This equation plays a crucial role in all that follows and hence we will call
it the \textbf{main Vekua equation}. We notice that the operator of this
equation is precisely the second factor in (\ref{mainfact}).
Denote $W_{1}=\operatorname*{Re}W$ and $W_{2}=\operatorname{Im}W$.
\begin{theorem}
\cite{KrJPhys06}\label{ThConjugate2} Let $W=W_{1}+iW_{2}$ be a solution of
(\ref{Vekuamain}). Assume that $f=p^{1/2}u_{0}$, where $u_{0}$ is a positive
solution of (\ref{maineq}) in $\Omega$. Then $u=p^{-1/2}W_{1}$ is a solution
of (\ref{maineq}) in $\Omega$, and $v=p^{1/2}W_{2}$ is a solution of the
equation
\begin{equation}
(\operatorname*{div}\frac{1}{p}\operatorname*{grad}+q_{1})v=0\qquad\text{in
}\Omega, \label{assocmaineq}%
\end{equation}
where
\begin{equation}
q_{1}=-\frac{1}{p}\left( \frac{q}{p}+2\left\langle \frac{\nabla p}{p}%
,\frac{\nabla u_{0}}{u_{0}}\right\rangle +2\left( \frac{\nabla u_{0}}{u_{0}%
}\right) ^{2}\right) . \label{q1}%
\end{equation}
\end{theorem}
Theorem \ref{ThConjugate2} shows us that as much as real and imaginary parts
of a complex analytic function are harmonic functions, the real and imaginary
parts of a solution of the main Vekua equation (\ref{Vekuamain}) multiplied by
$p^{-1/2}$ and $p^{1/2}$ respectively are solutions of the associated elliptic
equations (\ref{maineq}) and (\ref{assocmaineq}). The following natural
question arises then. We know that given an arbitrary real valued harmonic
function in a simply connected domain, a conjugate harmonic function can be
constructed explicitly such that the obtained couple of harmonic functions
represent the real and imaginary parts of a complex analytic function. What is
the corresponding more general fact for solutions of associated elliptic
equations (\ref{maineq}) and (\ref{assocmaineq}) (which we slightly
generalizing the definition of I. N. Vekua call metaharmonic functions). The
precise result is given in the following theorem.
\begin{theorem}
\cite{KrJPhys06} \label{CorConjugate}Let $f=p^{1/2}u_{0}$, where $u_{0}$ is a
positive solution of (\ref{maineq}) in a simply connected domain $\Omega$ and
$u$ be a solution of (\ref{maineq}). Then a solution $v$ of (\ref{assocmaineq}%
) with $q_{1}$ defined by (\ref{q1}) such that $W=p^{1/2}u+ip^{-1/2}v$ is a
solution of (\ref{Vekuamain}), is constructed according to the formula%
\begin{equation}
v=u_{0}^{-1}\overline{A}(ipu_{0}^{2}\partial_{\overline{z}}(u_{0}^{-1}u)).
\label{transfDarboux}%
\end{equation}
Let $v$ be a solution of (\ref{assocmaineq}), then the corresponding solution
$u$ of (\ref{maineq}) such that $W=p^{1/2}u+ip^{-1/2}v$ is a solution of
(\ref{Vekuamain}), is constructed according to the formula%
\begin{equation}
u=-u_{0}\overline{A}(ip^{-1}u_{0}^{-2}\partial_{\overline{z}}(u_{0}v)).
\label{transfDarbouxinv}%
\end{equation}
\end{theorem}
\begin{remark}
When $p\equiv1$, $q\equiv0$ and $u_{0}\equiv1$, equalities
(\ref{transfDarboux}) and (\ref{transfDarbouxinv}) turn into the well known
formulas in complex analysis for constructing conjugate harmonic functions.
\end{remark}
\section{Formal powers}
Briefly speaking formal powers are solutions of a Vekua equation
\begin{equation}
W_{\overline{z}}=aW+b\overline{W} \label{VekuaGeneral}%
\end{equation}
(with $a$ and $b$ being complex valued functions) generalizing the usual
analytic powers $\left\{ (z-z_{0})^{n}\right\} _{n=0}^{\infty}$ in the sense
that locally when $z\rightarrow z_{0}$ they behave asymptotically like the
usual powers and under some additional conditions on the coefficients $a$ and
$b$ they form a complete system in the space of all solutions of the Vekua
equation in the same sense as the analytic powers $\left\{ (z-z_{0}%
)^{n}\right\} _{n=0}^{\infty}$ form a complete system in the space of
analytic functions. Generalizations of the extension theorem, the Runge
theorem and of other important results about the convergence of corresponding
series are valid. The construction of formal powers is one of the main
problems of pseudoanalytic function theory. Recently it was solved
\cite{KrRecentDevelopments}, \cite{APFT} for a wide class of Vekua equations
of the form (\ref{Vekuamain}) which as was shown in the preceding section are
of main interest for studying problems for second order equations of the form
(\ref{maineq}).
The main ingredient for obtaining the explicit form of formal powers for a
certain Vekua equation is the generating sequence, a concept introduced by
Bers. If one knows a generating sequence for a given Vekua equation then the
construction of formal powers reduces to a simple algorithm. Here we briefly
explain the main ideas and steps refering the reader to \cite{Berskniga} and
\cite{APFT} for further details.
\subsection{Generating pair and generating sequence}
\begin{definition}
A pair of solutions $F$ and $G$ of a Vekua equation (\ref{VekuaGeneral}) in
$\Omega$ possessing partial derivatives with respect to the real variables $x$
and $y$ is said to be a generating pair if it satisfies the inequality
\begin{equation}
\operatorname{Im}(\overline{F}G)>0\qquad\text{in }\Omega. \label{GenPairCond}%
\end{equation}
\end{definition}
Condition (\ref{GenPairCond}) implies that every complex function $W$ defined
in a subdomain of $\Omega$ admits the unique representation $W=\phi F+\psi G$
where the functions $\phi$ and $\psi$ are real valued. Thus, the pair $(F,G)$
generalizes the pair $(1,i)$ which corresponds to usual complex analytic
function theory. The following expressions are known as characteristic
coefficients of the pair $(F,G)$%
\begin{align*}
a_{(F,G)} & =-\frac{\overline{F}G_{\overline{z}}-F_{\overline{z}}%
\overline{G}}{F\overline{G}-\overline{F}G},\qquad b_{(F,G)}=\frac
{FG_{\overline{z}}-F_{\overline{z}}G}{F\overline{G}-\overline{F}G},\\
A_{(F,G)} & =-\frac{\overline{F}G_{z}-F_{z}\overline{G}}{F\overline
{G}-\overline{F}G},\qquad B_{(F,G)}=\frac{FG_{z}-F_{z}G}{F\overline
{G}-\overline{F}G}.
\end{align*}
If $(F,G)$ is a generating pair of a Vekua equation (\ref{VekuaGeneral}) then
$a_{(F,G)}=a$ and $b_{(F,G)}=b$. \ The other two characteristic coefficients
are related to the concept of a derivative \cite{Berskniga}. The
$(F,G)$-derivative $\overset{\cdot}{W}=\frac{d_{(F,G)}W}{dz}$ of a
continuously differentiable function $W$ exists and has the form
\begin{equation}
\overset{\cdot}{W}=W_{z}-A_{(F,G)}W-B_{(F,G)}\overline{W} \label{FGder}%
\end{equation}
if and only if%
\[
W_{\overline{z}}=a_{(F,G)}W+b_{(F,G)}\overline{W}.
\]
Solutions of this equation are called $(F,G)$-pseudoanalytic functions.
\begin{definition}
\label{DefSuccessor}Let $(F,G)$ and $(F_{1},G_{1})$ - be two generating pairs
in $\Omega$. $(F_{1},G_{1})$ is called \ successor of $(F,G)$ and $(F,G)$ is
called predecessor of $(F_{1},G_{1})$ if%
\begin{equation}
a_{(F_{1},G_{1})}=a_{(F,G)}\qquad\text{and}\qquad b_{(F_{1},G_{1})}%
=-B_{(F,G)}\text{.} \label{charcoeffsuccessor}%
\end{equation}
\end{definition}
This definition arises naturally in relation to the notion of the
$(F,G)$-derivative due to the following fact.
\begin{theorem}
\label{ThBersDer}Let $W$ be an $(F,G)$-pseudoanalytic function and let
$(F_{1},G_{1})$ be a successor of $(F,G)$. Then $\overset{\cdot}{W}$ is an
$(F_{1},G_{1})$-pseudoanalytic function.
\end{theorem}
Thus, to the difference of analytic functions whose derivatives are again
analytic, the $(F,G)$-derivatives of pseudoanalytic functions are in general
solutions of another Vekua equation with the coefficients given by
(\ref{charcoeffsuccessor}). Obviously this process of construction of new
Vekua equations associated with the previous ones via relations
(\ref{charcoeffsuccessor}) can be continued and we arrive at the following definition.
\begin{definition}
\label{DefSeq}A sequence of generating pairs $\left\{ (F_{m},G_{m})\right\}
$, $m=0,\pm1,\pm2,\ldots$, is called a generating sequence if $(F_{m+1}%
,G_{m+1})$ is a successor of $(F_{m},G_{m})$. If $(F_{0},G_{0})=(F,G)$, we say
that $(F,G)$ is embedded in $\left\{ (F_{m},G_{m})\right\} $.
\end{definition}
\begin{definition}
A generating sequence $\left\{ (F_{m},G_{m})\right\} $ is said to have
period $\mu>0$ if $(F_{m+\mu},G_{m+\mu})$ is equivalent to $(F_{m},G_{m})$
that is their characteristic coefficients coincide.
\end{definition}
We will need the following notation introduced by Bers. The $(F,G)$-integral
is defined as follows
\[
\int_{\Gamma}Wd_{(F,G)}z=F(z_{1})\operatorname{Re}\int_{\Gamma}\frac
{2\overline{G}}{F\overline{G}-\overline{F}G}Wdz-G(z_{1})\operatorname{Re}%
\int_{\Gamma}\frac{2\overline{F}}{F\overline{G}-\overline{F}G}Wdz
\]
where $\Gamma$ is a rectifiable curve leading from $z_{0}$ to $z_{1}$.
Let $W$ be an $(F,G)$-pseudoanalytic function. Using a generating sequence in
which $(F,G)$ is embedded we can define the higher derivatives of $W$ by the
recursion formula%
\[
W^{[0]}=W;\qquad W^{[m+1]}=\frac{d_{(F_{m},G_{m})}W^{[m]}}{dz},\quad
m=0,1,\ldots\text{.}%
\]
A generating sequence defines an infinite sequence of Vekua equations. If for
a given (original) Vekua equation we know not only a corresponding generating
pair but the whole generating sequence, that is a couple of exact and
independent solutions for each of the Vekua equations from the infinite
sequence of equations corresponding to the original one, we are able to
construct an infinite system of solutions of the original Vekua equation as is
shown in the next definition.
\begin{definition}
\label{DefFormalPower}The formal power $Z_{m}^{(0)}(a,z_{0};z)$ with center at
$z_{0}\in\Omega$, coefficient $a$ and exponent $0$ is defined as the linear
combination of the generators $F_{m}$, $G_{m}$ with real constant coefficients
$\lambda$, $\mu$ chosen so that $\lambda F_{m}(z_{0})+\mu G_{m}(z_{0})=a$. The
formal powers with exponents $n=1,2,\ldots$ are defined by the recursion
formula%
\begin{equation}
Z_{m}^{(n)}(a,z_{0};z)=n\int_{z_{0}}^{z}Z_{m+1}^{(n-1)}(a,z_{0};\zeta
)d_{(F_{m},G_{m})}\zeta. \label{recformula}%
\end{equation}
\end{definition}
This definition implies the following properties.
\begin{enumerate}
\item $Z_{m}^{(n)}(a,z_{0};z)$ is an $(F_{m},G_{m})$-pseudoanalytic function
of $z$.
\item If $a^{\prime}$ and $a^{\prime\prime}$ are real constants, then
$Z_{m}^{(n)}(a^{\prime}+ia^{\prime\prime},z_{0};z)=a^{\prime}Z_{m}%
^{(n)}(1,z_{0};z)+a^{\prime\prime}Z_{m}^{(n)}(i,z_{0};z).$
\item The formal powers satisfy the differential relations%
\[
\frac{d_{(F_{m},G_{m})}Z_{m}^{(n)}(a,z_{0};z)}{dz}=nZ_{m+1}^{(n-1)}%
(a,z_{0};z).
\]
\item The asymptotic formulas
\begin{equation}
Z_{m}^{(n)}(a,z_{0};z)\sim a(z-z_{0})^{n},\quad z\rightarrow z_{0}
\label{asymptformulas}%
\end{equation}
hold.
\end{enumerate}
Assume now that
\begin{equation}
W(z)=\sum_{n=0}^{\infty}Z^{(n)}(a_{n},z_{0};z) \label{series}%
\end{equation}
where the absence of the subindex $m$ means that all the formal powers
correspond to the same generating pair $(F,G),$ and the series converges
uniformly in some neighborhood of $z_{0}$. It can be shown that the uniform
limit of pseudoanalytic functions is pseudoanalytic, and that a uniformly
convergent series of $(F,G)$-pseudoanalytic functions can be $(F,G)$%
-differentiated term by term. Hence the function $W$ in (\ref{series}) is
$(F,G)$-pseudoanalytic and its $r$th derivative admits the expansion
\[
W^{[r]}(z)=\sum_{n=r}^{\infty}n(n-1)\cdots(n-r+1)Z_{r}^{(n-r)}(a_{n}%
,z_{0};z).
\]
From this the Taylor formulas for the coefficients are obtained%
\begin{equation}
a_{n}=\frac{W^{[n]}(z_{0})}{n!}. \label{Taylorcoef}%
\end{equation}
\begin{definition}
Let $W(z)$ be a given $(F,G)$-pseudoanalytic function defined for small values
of $\left\vert z-z_{0}\right\vert $. The series%
\begin{equation}
\sum_{n=0}^{\infty}Z^{(n)}(a_{n},z_{0};z) \label{Taylorseries}%
\end{equation}
with the coefficients given by (\ref{Taylorcoef}) is called the Taylor series
of $W$ at $z_{0}$, formed with formal powers.
\end{definition}
The Taylor series always represents the function asymptotically:%
\begin{equation}
W(z)-\sum_{n=0}^{N}Z^{(n)}(a_{n},z_{0};z)=O\left( \left\vert z-z_{0}%
\right\vert ^{N+1}\right) ,\quad z\rightarrow z_{0}, \label{asympt}%
\end{equation}
for all $N$. This implies (since a pseudoanalytic function can not have a zero
of arbitrarily high order without vanishing identically) that the sequence of
derivatives $\left\{ W^{[n]}(z_{0})\right\} $ determines the function $W$ uniquely.
If the series (\ref{Taylorseries}) converges uniformly in a neighborhood of
$z_{0}$, it converges to the function $W$.
\subsection{Convergence theorems\label{SubsectConvergence}}
S. Agmon and L. Bers \cite{AgmonBers} and L. Bers developed a theory of
expansions in pseudoanalytic formal powers which in its generality is
presented in \cite{Berskniga}, \cite{BersFormalPowers}. We do need here the
general results concerning a general Vekua equation (\ref{VekuaGeneral}).
Fortunately the situation with the main Vekua equation (\ref{Vekuamain}) in a
bounded simply connected domain under quite natural conditions on the function
$f$ is much easier than in the general case, and we have the following
expansion theorem and Runge theorem \cite{KrRecentDevelopments}, \cite{APFT}.
\begin{theorem}
\label{ThConvPer} Let $D$ be a disk of a finite radius $R$ and center $z_{0}$,
and $f\in C^{1}(\overline{D})$ be positive in $\overline{D}$. Then any
solution $W$ of (\ref{Vekuamain}) \ in $D$ admits a unique normally convergent
expansion\footnote{Following \cite{Berskniga}, \cite{Dettman} we shall say
that a sequence of functions $W_{n}$ converges normally in a domain $\Omega$
if it converges uniformly on every bounded closed subdomain of $\Omega$.} of
the form $W(z)=\sum_{n=0}^{\infty}Z^{(n)}(a_{n},z_{0};z)$.
\end{theorem}
\begin{theorem}
\label{ThRunge} any solution $W$ of (\ref{Vekuamain}) defined in a simply
connected domain can be expanded into a normally convergent series of formal
polynomials (linear combinations of formal powers with positive exponents).
\end{theorem}
\begin{remark}
This theorem admits a direct generalization onto the case of a multiply
connected domain (see \cite{BersFormalPowers}).
\end{remark}
We mention here another important result obtained by Menke in \cite{Menke}
which gives a useful estimate for the rate of convergence of the series from
the preceding theorem in the case when $W$ is a H\"{o}lder continuous function
up to the boundary of the domain of interest.
\begin{theorem}
\label{ThMenke} Let $W$ be a pseudoanalytic function in a domain $\Omega$
bounded by a Jordan curve and satisfy the H\"{o}lder condition on
$\partial\Omega$ with the exponent $\alpha$ ($0<\alpha\leq1$). Then for any
$\varepsilon>0$ and any natural $n$ there exists a pseudopolynomial of order
$n$ satisfying the inequality
\[
\left\vert W(z)-P_{n}(z)\right\vert \leq\frac{\operatorname*{Const}}%
{n^{\alpha-\varepsilon}}\qquad\text{for any }z\in\overline{\Omega}%
\]
where the constant does not depend on $n$, but only on $\varepsilon$.
\end{theorem}
The following statements are direct corollaries of the relations established
in section \ref{SectFactorization} between pseudoanalytic functions (solutions
of (\ref{Vekuamain})) and solutions of second-order elliptic equations, and of
the convergence theorems formulated above. Here we assume the existence of a
positive solution $u_{0}$ of (\ref{maineq}) in the domain $\Omega$ and the
function $f$ in (\ref{Vekuamain}) to be defined by $f=p^{1/2}u_{0}$ and belong
to $C^{1}(\overline{\Omega})$.
\begin{definition}
Let $u(z)$ be a given solution of the equation (\ref{maineq}) defined for
small values of $\left\vert z-z_{0}\right\vert $, and let $W(z)$ be a solution
of (\ref{Vekuamain}) constructed according to theorem \ref{CorConjugate}, such
that $\operatorname*{Re}W=p^{1/2}u$. The series
\begin{equation}
p^{-1/2}(z)\sum_{n=0}^{\infty}\operatorname*{Re}Z^{(n)}(a_{n},z_{0};z)
\label{TaylorForU}%
\end{equation}
with the coefficients given by (\ref{Taylorcoef}) is called the Taylor series
of $u$ at $z_{0}$, formed with formal powers.
\end{definition}
\begin{theorem}
\cite{KrJPhys06}, \cite{APFT} \label{ThExpMainEq}Let $u(z)$ be a solution of
(\ref{maineq}) defined for $\left\vert z-z_{0}\right\vert <R$. Then it admits
a unique expansion of the form
\[
u(z)=p^{-1/2}(z)\sum_{n=0}^{\infty}\operatorname*{Re}Z^{(n)}(a_{n},z_{0};z)
\]
which converges normally for $\left\vert z-z_{0}\right\vert <R$.
\end{theorem}
\begin{theorem}
\label{ThRungeSchr}An arbitrary solution of (\ref{maineq}) defined in a simply
connected domain where there exists a positive particular solution $u_{0}$
such that $f=p^{1/2}u_{0}\in C^{1}(\overline{\Omega})$ can be expanded into a
normally convergent series of formal polynomials multiplied by $p^{-1/2}$.
\end{theorem}
More precisely the last theorem has the following meaning. Due to Property 2
of formal powers we have that $Z^{(n)}(a,z_{0};z)$ for any Taylor coefficient
$a$ can be expressed through $Z^{(n)}(1,z_{0};z)$ and $Z^{(n)}(i,z_{0};z)$.
Then due to theorem \ref{ThRunge} any solution $W$ of (\ref{Vekuamain}) can be
expanded into a normally convergent series of linear combinations of
$Z^{(n)}(1,z_{0};z)$ and $Z^{(n)}(i,z_{0};z)$. Consequently, any solution of
(\ref{maineq}) can be expanded into a normally convergent series of linear
combinations of real parts of $Z^{(n)}(1,z_{0};z)$ and $Z^{(n)}(i,z_{0};z)$
multiplied by $p^{-1/2}$.
Obviously, for solutions of (\ref{maineq}) the results on the interpolation
and on the degree of approximation like, e.g., theorem \ref{ThMenke} are also valid.
Let us stress that theorem \ref{ThRungeSchr} gives us the following result.
The functions
\begin{equation}
\left\{ p^{-1/2}(z)\operatorname*{Re}Z^{(n)}(1,z_{0};z),\quad p^{-1/2}%
(z)\operatorname*{Re}Z^{(n)}(i,z_{0};z)\right\} _{n=0}^{\infty}
\label{complsystem}%
\end{equation}
represent a complete system of solutions of (\ref{maineq}) in the sense that
any solution of (\ref{maineq}) can be represented by a normally convergent
series formed by functions (\ref{complsystem}) in any simply connected domain
$\Omega$ where a positive solution of (\ref{maineq}) exists, and the rate of
convergence of the series can be estimated with the aid of theorem
\ref{ThMenke}.
\subsection{Explicit construction of generating sequences and formal
powers\label{SectTheMainVekua}}
The results of section \ref{SectFactorization} show us that the theory of the
elliptic equation
\[
(\operatorname{div}p\operatorname{grad}+q)u=0
\]
is closely related to equation (\ref{Vekuamain}):
\begin{equation}
W_{\overline{z}}=\frac{f_{\overline{z}}}{f}\overline{W}. \label{Vekuamain1}%
\end{equation}
It is interesting that for this equation we always know a generating pair.
Namely, it is easy to see that the functions $F=f\quad$and\quad$G=\frac{i}{f}$
\ satisfy (\ref{Vekuamain1}) together with the condition (\ref{GenPairCond}).
Then the corresponding characteristic coefficients $A_{(F,G)}$ and $B_{(F,G)}$
have the form%
\[
A_{(F,G)}=0,\quad\text{\quad}B_{(F,G)}=\frac{f_{z}}{f},
\]
and the $(F,G)$-derivative according to (\ref{FGder}) is defined as follows%
\[
\overset{\cdot}{W}=W_{z}-\frac{f_{z}}{f}\overline{W}=\left( \partial
_{z}-\frac{f_{z}}{f}C\right) W.
\]
Due to Theorem \ref{ThBersDer} we obtain the following statement.
\begin{proposition}
\label{PrDer} Let $W$ be a solution of (\ref{Vekuamain1}). Then its
$(F,G)$-derivative, the function $w=\overset{\cdot}{W}$ is a solution of the
equation $\left( \partial_{\overline{z}}+\frac{f_{z}}{f}C\right) w=0$.
\end{proposition}
In spite of having given a generating pair for (\ref{Vekuamain1}) in general
it is not known how to construct a corresponding generating sequence necessary
for calculating the system of formal powers. Nevertheless a recent result from
\cite{KrRecentDevelopments}, \cite{APFT} which we formulate in the following
statement gives an answer to this question in a quite general situation.
\begin{theorem}
\label{ThGenSeq} Let $F=S(s)T(t)$ and $G=\frac{i}{S(s)T(t)}$ where $S$ and $T$
are arbitrary differentiable nonvanishing real valued functions, $\Phi=s+it$
is an analytic function of the variable $z=x+iy$ in $\Omega$ such that
$\Phi_{z}$ is bounded and has no zeros in $\Omega$. Then the generating pair
$(F,G)$ is embedded in the generating sequence $(F_{m},G_{m})$, $m=0,\pm
1,\pm2,\ldots$ in $\Omega$ defined as follows
\[
F_{m}=\left( \Phi_{z}\right) ^{m}F\quad\text{and\quad}G_{m}=\left( \Phi
_{z}\right) ^{m}G\quad\text{for even }m
\]
and%
\[
F_{m}=\frac{\left( \Phi_{z}\right) ^{m}}{S^{2}}F\quad\text{and\quad}%
G_{m}=\left( \Phi_{z}\right) ^{m}S^{2}G\quad\text{for odd }m.
\]
\end{theorem}
In order to appreciate the generality of this construction let us remind that
orthogonal coordinate systems in a plane are obtained (see \cite{Madelung})
from Cartesian coordinates $x$, $y$ by means of the relation%
\[
s+it=\Phi(x+iy)
\]
where $\Phi$ is an arbitrary analytic function. Quite often a transition to
more general coordinates is useful
\[
\xi=\xi(s),\quad\eta=\eta(t).
\]
$\xi$ and $\eta$ preserve the property of orthogonality. To illustrate the
point, besides the obvious example of Cartesian coordinates which are
generated by the analytic function $z$ we give some other examples taken from
\cite{Madelung}.
\begin{example}
\textbf{Polar coordinates }%
\[
s+it=\ln(x+iy),
\]%
\begin{equation}
s=\ln\sqrt{x^{2}+y^{2}},\quad t=\arctan\frac{y}{x}. \label{trueorthogcoord}%
\end{equation}
Usually the following new coordinates are introduced%
\[
r=e^{s}=\sqrt{x^{2}+y^{2}},\quad\varphi=t=\arctan\frac{y}{x}.
\]
\end{example}
\begin{example}
\textbf{Parabolic coordinates }%
\[
\frac{s+it}{\sqrt{2}}=\sqrt{x+iy},
\]%
\[
s=\sqrt{r+x},\quad t=\sqrt{r-x}.
\]
More frequently the parabolic coordinates are introduced as follows%
\[
\xi=s^{2},\quad\eta=t^{2}.
\]
\end{example}
\begin{example}
\textbf{Elliptic coordinates }%
\[
s+it=\arcsin\frac{x+iy}{\alpha},
\]%
\[
\sin s=\frac{s_{1}-s_{2}}{2\alpha},\quad\cosh t=\frac{s_{1}+s_{2}}{2\alpha}%
\]
where $s_{1}=\sqrt{(x+\alpha)^{2}+y^{2}}$, $s_{2}=\sqrt{(x-\alpha)^{2}+y^{2}}%
$. The substitution
\[
\xi=\sin s,\quad\eta=\cosh t
\]
is frequently used.
\end{example}
\begin{example}
\textbf{Bipolar coordinates}%
\[
s+it=\ln\frac{\alpha+x+iy}{\alpha-x-iy},
\]%
\[
\tanh s=\frac{2\alpha x}{\alpha^{2}+x^{2}+y^{2}},\quad\tan t=\frac{2\alpha
y}{\alpha^{2}-x^{2}-y^{2}}.
\]
The following substitution is frequently used%
\[
\xi=e^{-s},\quad\eta=\pi-t.
\]
\end{example}
The last theorem opens the way for explicit construction of formal powers
corresponding to the main Vekua equation (\ref{Vekuamain1}) in the case when
$f$ has the form
\begin{equation}
f=S(s)T(t) \label{fST}%
\end{equation}
and hence for explicit construction of complete systems of solutions for
corresponding second-order elliptic equations admitting a particular solution
of this form.
\section{Description of the method\label{SectDescription}}
We consider boundary value problems of Dirichlet, Neumann or mixed type for
the elliptic equation of the form (\ref{maineq}) in a bounded, simply
connected domain $\Omega\subset\mathbb{R}^{2}$. The main assumption required
for the applicability of the method of pseudoanalytic formal powers (MPFP) is
the existence in $\overline{\Omega}$ of a positive solution $u_{0}$ such that
the function $f=p^{1/2}u_{0}\in C^{1}(\overline{\Omega})$ be representable in
a separable form (\ref{fST}) in an orthogonal coordinate system. Let us stress
that very often such a particular solution $u_{0}$ is readily available. The
simplest example of such situation is when $q\equiv0$ and $p$ is of the form
(\ref{fST}). For example, the cases when $p(x,y)=X(x)Y(y)$ or $p=p(\sqrt
{x^{2}+y^{2}})$ frequently occur in practice \cite{Demidenko}.
When the equation of the form
\begin{equation}
(-\Delta+q(y))u(x,y)+\lambda^{2}u(x,y)=0 \label{Schr0}%
\end{equation}
is considered, it is sufficient to obtain a particular solution for the
ordinary differential equation
\begin{equation}
(-\frac{d^{2}}{dy^{2}}+q(y))h(y)=0. \label{ordinary}%
\end{equation}
Then a particular solution of (\ref{Schr0}) can be constructed as follows
\begin{equation}
u_{0}(x,y)=e^{\lambda x}h(y). \label{u0}%
\end{equation}
It has a convenient separable form. Notice that in this example we come to an
important open problem. It is related to the requirement that $u_{0}$ should
be different from zero in the domain of interest. Meanwhile in many
practically significant situations it is easy to guarantee that $h(y)\neq0$
when $(x,y)\in\overline{\Omega}$, sometimes this condition becomes a
considerable obstacle. Moreover, when $\lambda$ in (\ref{Schr0}) is purely
imaginary, the solution (\ref{u0}) is not acceptable because it is no longer
real valued. In this case one should\ take instead of $e^{ikx}$, where
$\lambda=ik$, a solution in the form of $\sin kx$ or $\cos kx$ but then for a
big domain or large $k$ one cannot avoid the appearance of zeros of the
resulting particular solution $u_{0}$ and in this case the proposed scheme in
general does not work.
One possibility to overcome this problem is to include under consideration a
complex valued particular solution $u_{0}$ but then we would need to consider
a corresponding bicomplex main Vekua equation (see \cite{KrJPhys06} and
\cite{APFT}). Then the whole algorithm for the construction of generating
sequences and formal powers would go through with no modification compared to
the complex case, but up to now there is no proof of the completeness of the
system of formal powers for a bicomplex Vekua equation. As a consequence there
is no guarantee that the infinite system of exact solutions obtained similar
to (\ref{complsystem}) will be complete in the space of solutions of
(\ref{maineq}) in $\Omega$. Our conjecture is that at least in the case when
$u_{0}(x,y)=g(x)h(y)$ where $g$ and $h$ are complex valued nonvanishing
functions the system of formal powers for the corresponding bicomplex main
Vekua equation is complete in the same sense as was established earlier for
the complex case. We continue this discussion in section \ref{SectEigen} where
we use complex valued particular solutions of the form (\ref{u0}) for solving
eigenvalue problems for operators of the form $-\Delta+q(y)$.
Turning back to equation (\ref{maineq}) we assume that it admits a positive
solution $u_{0}$ in the domain $\Omega$ such that $f=p^{1/2}u_{0}\in
C^{1}(\overline{\Omega})$ is representable in a separable form (\ref{fST}) and
that $\Phi=s+it$ is an analytic function of the variable $z=x+iy$ in $\Omega$
such that $\Phi_{z}$ is bounded and has no zeros in $\Omega$. Then applying
theorem \ref{ThGenSeq} one can construct a corresponding generating sequence.
Construction of formal powers $\left\{ Z^{(n)}(1,z_{0};z),\quad
Z^{(n)}(i,z_{0};z)\right\} _{n=0}^{\infty}$ reduces then to the recursive
algorithm described in Definition \ref{DefFormalPower}, and in this way one
obtains the complete system of solutions for (\ref{maineq}) in $\Omega$ given
by (\ref{complsystem}). By construction $\operatorname*{Re}Z^{(0)}%
(i,z_{0};z)\equiv0$. Taking this into account we introduce the notations
\begin{align*}
u_{1}(z) & =p^{-1/2}(z)\operatorname*{Re}Z^{(1)}(1,z_{0};z),\quad
u_{2}(z)=p^{-1/2}(z)\operatorname*{Re}Z^{(1)}(i,z_{0};z),\\
u_{3}(z) & =p^{-1/2}(z)\operatorname*{Re}Z^{(2)}(1,z_{0};z),\quad
u_{4}(z)=p^{-1/2}(z)\operatorname*{Re}Z^{(2)}(i,z_{0};z),\ldots
\end{align*}
and obtain the complete system of solutions for (\ref{maineq}) given by
$\left\{ u_{0},u_{1},u_{2},\ldots\right\} $. We look for an approximate
solution of a boundary value problem for (\ref{maineq}) in the form
\begin{equation}
u^{N}=\sum_{k=0}^{N}b_{k}u_{k}\label{uN}%
\end{equation}
where $b_{k}$ are real coefficients which should be found from boundary
conditions. To obtain $N+1$ equations for finding $\left\{ b_{k}\right\}
_{k=0}^{N}$ one can use, e.g., the collocation method. Chosing $N+1$ points
$\zeta_{j}\in\partial\Omega$ we obtain $N+1$ equations
\[
\sum_{k=0}^{N}b_{k}B[u_{k}](\zeta_{j})=v(\zeta_{j}),\quad j=\overline{0,N}%
\]
where $v$ is a given function and $B$ is the linear operator of the boundary
condition. For the Dirichlet condition one has $B[u]=u$ and for the Neumann
condition, $B[u]=\frac{\partial u}{\partial\overrightarrow{n}}$ -- the normal
derivative of $u$. Finding $\left\{ b_{k}\right\} _{k=0}^{N}$ we have the
approximate solution $u^{N}$.
Thus, the proposed here MPFP belongs to the class of boundary methods because
due to the linearity of the problem the function (\ref{uN}) is an exact
solution of (\ref{maineq}) in $\Omega$ and only boundary conditions should be
approximated. An estimate for the rate of convergence of the method is given
in theorem \ref{ThMenke}. In the next section we discuss the numerical
realization of MPFP and results of numerical tests.
\section{Approximate solution of boundary value problems}
As a first example we considered the Dirichlet problem for the equation
\begin{equation}
\left( -\Delta+c^{2}\right) u=0\label{Yukawa}%
\end{equation}
where $c$ is a real constant. The interest in this relatively simple equation
is not due to its numerical simplicity. In fact this is not the case,-
numerical solution of this equation is not less difficult than that of an
equation with $c$ being a reasonably good function with a range of values
comparable with $c$. The attractiveness of this example consists in the
possibility to calculate\ a large number of the functions $u_{k}$ (see the
preceding section) symbolically, using an appropriate software for symbolic
calculations like Mathematica (Wolfram), Maple or Matlab. In this work we used
Matlab 2006 and a PC of 2 GB in RAM and a processor of 1.73 GHz.
Implementation of the symbolically calculated base functions $u_{k}$ gives us
the possibility to estimate the accuracy of the MPFP itself without
considering the precision of recursive numerical integrations. We also compare
the results obtained using symbolically calculated $u_{k}$ with the results
obtained purely numerically.
For equation (\ref{Yukawa}) it is easy to propose a positive particular
solution. It can be chosen, e.g., as $f=e^{cy}$. Then the first functions
$u_{k}$ constructed as described in the preceding section taking as a center
of the formal powers the origin will have the form \cite{APFT}%
\begin{align}
u_{0}(x,y) & =e^{cy},\qquad u_{1}(x,y)=xe^{cy},\qquad u_{2}(x,y)=-\frac
{\sinh(cy)}{c},\nonumber\\
& \label{uk}\\
u_{3}(x,y) & =\left( x^{2}-\frac{y}{c}\right) e^{cy}+\frac{\sinh(cy)}%
{c^{2}},\qquad u_{4}(x,y)=-\frac{2x\sinh(cy)}{c},\ldots\text{.}\nonumber
\end{align}
It is interesting to mention that using Matlab we obtained the first 101
functions of this system calculated symbolically. According to theorem
\ref{ThRungeSchr} this system of solutions is complete in any bounded simply
connected domain containing the origin. First we show results obtained with
the help of the system of functions $u_{k}$ calculated symbolically.
\subsection{Numerical results obtained with symbolically calculated base
functions}
We begin with the unitary disk $D$ with center in the origin. As a test exact
solution we take the function
\begin{equation}
u=e^{cx}. \label{exact}%
\end{equation}
Thus, the problem we consider is to solve (\ref{Yukawa}) in $D$ with the
boundary condition $\left. u\right\vert _{\partial D}=e^{cx}$. We look for an
approximate solution $u^{N}$ in the form (\ref{uN}) with the base functions
(\ref{uk}). We use the collocation method for satisfying the boundary
condition, the collocation points are distributed uniformly on $\partial D$.
Their number is equal to the number of solutions $u_{k}$.
According to theory from subsection \ref{SubsectConvergence} the coefficients
$b_{k}$ in (\ref{uN}) are obtained in the case under consideration from the
Taylor coefficients which appear in (\ref{TaylorForU}). More precisely we have
that according to theorem \ref{ThExpMainEq} the solution $u$ can be
represented as follows
\[
u(z)=\sum_{n=0}^{\infty}\operatorname*{Re}Z^{(n)}(a_{n},0;z)=\sum
_{n=0}^{\infty}\left( a_{n}^{\prime}\operatorname*{Re}Z^{(n)}(1,0;z)+a_{n}%
^{\prime\prime}\operatorname*{Re}Z^{(n)}(i,0;z)\right)
\]
where $a_{n}=a_{n}^{\prime}+ia_{n}^{\prime\prime}$ are the Taylor coefficients
given by (\ref{Taylorcoef}). In the case of the exact solution (\ref{exact})
the Taylor coefficients have the form \cite[Sect. 7.3]{APFT}%
\[
a_{n}=\frac{c^{n}}{n!}(1+i).
\]
Thus, the exact values for the coefficients $b_{k}$ from (\ref{uN}) in our
example are as follows%
\[
b_{0}=1,\quad b_{1}=b_{2}=c,\quad b_{3}=b_{4}=\frac{c^{2}}{2},\ldots.
\]
Having compared the numerically calculated constants $b_{k}$ which we denote
by $\widetilde{b}_{k}$ for $N=34$ with their exact values in the case $c=1$ we
obtained their coincidence up to $10^{-14}$ for every $k=0,\ldots,34$. For
smaller values of $c$ the situation is the same. The difference between
$\widetilde{b}_{k}$ and $b_{k}$ tends to become larger for larger values of
$c$. In Table \ref{Table1} we show results for $c=5$ and $N=34$.
\begin{center}
Table 1.\label{Table1} Comparison of the values of $\widetilde{b}_{k}$ and
$b_{k}$ as $k$ increases%
\begin{tabular}
[c]{|c|c|c|c|}\hline
$k$ & The values of $\widetilde{b}_{k}$ & The values of $b_{k}$ & $\left\vert
\widetilde{b}_{k}-b_{k}\right\vert $\\\hline
$5$ & $20.83333333333382$ & $20.83333333333333$ & $0.00000000000048$\\\hline
$8$ & $26.04166666666448$ & $26.04166666666667$ & $0.00000000000219$\\\hline
$13$ & $15.50099206509864$ & $15.50099206349206$ & $0.00000000160657$\\\hline
$17$ & $5.38228885848900$ & $5.38228891093474$ & $0.00000005244574$\\\hline
$25$ & $0.19601580023149$ & $0.19603324996120$ & $0.00001744972971$\\\hline
$31$ & $0.00743227875004$ & $0.00729290364439$ & $0.00013937510565$\\\hline
$34$ & $0.00172611010091$ & $0.00214497166011$ & $0.00041886155920$\\\hline
\end{tabular}
\end{center}
The maximum number of functions $u_{k}$ that we used here is limited not by
the possibility of obtaining them symbolically but rather by the time required
for numerical calculations involving the corresponding quite long symbolic expressions.
In the following two tables, the convergence of MPFP is shown by comparison of
the maximum absolute error obtained for different values of $N$, for the case
$c=1$ and $c=5$.
\begin{center}
Tables 2 and 3\label{Table2} Maximum absolute error depending on $N$ for $c=1$
and $c=5$
\begin{tabular}
[c]{|c|c|}\hline
$N$ & Maximum absolute error\\\hline
$8$ & $0.00698626935341$\\\hline
$14$ & $2.534633673767495\times10^{-5}$\\\hline
$22$ & $1.432881036045330\times10^{-9}$\\\hline
$28$ & $4.276579090856103\times10^{-13}$\\\hline
$32$ & $1.776356839400251\times10^{-15}$\\\hline
$36$ & $8.881784197001252\times10^{-16}$\\\hline
$38$ & $1.110223024625157\times10^{-15}$\\\hline
\end{tabular}
\qquad%
\begin{tabular}
[c]{|c|c|}\hline
$N$ & Maximum absolute error\\\hline
$6$ & $3.59578971016677\times10^{2}$\\\hline
$14$ & $22.38029523897584$\\\hline
$22$ & $0.73431266884919$\\\hline
$32$ & $0.00194275813006$\\\hline
$44$ & $0.59167057031573\times10^{-7}$\\\hline
$54$ & $0.72795103278622\times10^{-11}$\\\hline
$60$ & $8.781864124784988\times10^{-14}$\\\hline
\end{tabular}
\end{center}
In Table 4 for a fixed number ($N=34$) of the base functions $u_{k}$ we show
the dependence of the maximum absolute error on the parameter $c$. Here we
also indicate the maximum absolute error obtained for the same problem using
the standard PDE tool of Matlab.
\begin{center}
Table 4.\label{Table4} Performance of MPFP compared to Matlab's PDE tool in
terms of the maximum absolute error for increasing values of $c$ as $N=34$%
\begin{tabular}
[c]{|c|c|c|}\hline
$c$ & Maximum absolute error of MPFP & PDE tool (2129 nodes)\\\hline
$0.1$ & $0.89\times10^{-15}$ & $1.5\times10^{-6}$\\\hline
$0.5$ & $0.26\times10^{-14}$ & $4.5\times10^{-6}$\\\hline
$1$ & $0.12\times10^{-14}$ & $1.6\times10^{-4}$\\\hline
$2$ & $0.14\times10^{-10}$ & $1.4\times10^{-3}$\\\hline
$5$ & $0.29\times10^{-3}$ & $3.0\times10^{-2}$\\\hline
$10$ & $4.06\times10^{2}$ & $8.0$\\\hline
\end{tabular}
\end{center}
As it can be observed in the last table the result of application of MPFP in
the case of $c=10$ is less satisfactory as that of PDE tool. This is due to
the fact that for larger values of $c$ one should consider a bigger $N$. In
Table 5 we show the absolute error of MPFP for $c=10$ and $N\geq42$.
\begin{center}
Table 5. Improvement in the maximum absolute error due to MPFP as the number
of functions $u_{k}$ keeps increasing%
\begin{tabular}
[c]{|c|c|}\hline
$N$ & Maximum absolute error of MPFP\\\hline
$42$ & $3.89$\\\hline
$44$ & $3.25$\\\hline
$46$ & $1.81$\\\hline
$48$ & $0.81$\\\hline
$50$ & $0.41$\\\hline
$52$ & $0.10$\\\hline
\end{tabular}
\end{center}
Thus, one can see that for $N\geq42$ the result obtained with the aid of MPFP
is more accurate than that given by Matlab. We stress that in the case of
using MPFP a system of $N+1$ linear algebraic equations is solved which means
solution of dozens of equations instead of thousands required by the finite
element method implemented in the PDE tool.
We experimented also with the shape of the domain. We considered the elliptic
form as well as a unitary disk with a triangle shaped deformation. In the
first case it is possible to see how the maximum absolute error increases with
the excentricity $e$, Table 6. Here in all cases the area of the considered
ellipses was kept constant, equal to $\pi$, while the excentricity was being increased.
\begin{center}
Table 6. Maximum absolute error for different values of the excentricity of
the elliptic domain with the area of the domain being equal to $\pi$. The case
$e=0$ corresponds to the unitary disc.%
\begin{tabular}
[c]{|c|c|c|c|c|c|c|}\hline
$N$ & $e=0$ & $e=0.5$ & $e=0.7$ & $e=0.9$ & $e=0.95$ & $e=0.99$\\\hline
$30$ & $2.2\times10^{-14}$ & $0.4\times10^{-13}$ & $0.5\times10^{-13}$ &
$0.3\times10^{-12}$ & $0.2\times10^{-11}$ & $1\times10^{-10}$\\\hline
\end{tabular}
\end{center}
For the case of the domain with a triangular deformation (see Fig. 1),
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
natheight=4.375100in,
natwidth=5.833200in,
height=3.0173in,
width=4.0171in
]%
{Fig1.jpg}%
\caption{The unitary disk with a triangular deformation.}%
\end{center}
\end{figure}
the errors were tested for different heights of the peak and different values
of $c$ and $N$ with satisfactory results. In Table 7 we present the maximum
absolute error of the approximate solution of the boundary value problem in
dependence on the height of the triangular peak over the unitary circunference.
\begin{center}
Table 7. Maximum absolute error for $N=31$, $c=1$ and different heights of the peak%
\begin{tabular}
[c]{|c|c|}\hline
Height of the peak over the unitary disk & Maximum absolute error\\\hline
$0.5$ & $0.92\times10^{-12}$\\\hline
$0.7$ & $0.62\times10^{-11}$\\\hline
$1.0$ & $0.72\times10^{-10}$\\\hline
\end{tabular}
\bigskip
\end{center}
\subsection{Results obtained with numerically calculated base functions}
The use of the numerically calculated base functions which we denote by
$\widetilde{u}_{k}$ poses the natural question about the accuracy of their
calculation. Consideration of equation (\ref{Yukawa}) gives us the possibility
to compare $\widetilde{u}_{k}$ with the symbolically calculated exact
solutions (\ref{uk}). In the following table we give the difference between
$u_{k}$ and $\widetilde{u}_{k}$ for $c=1$.
\begin{center}
Table 8. Maximum absolute error of calculation of the base functions for $c=1$%
\begin{tabular}
[c]{|c|c|}\hline
$k$ & $\left\vert u_{k}-\widetilde{u}_{k}\right\vert $\\\hline
$1$ & $0.00000667646050\times10^{-4}$\\\hline
$5$ & $0.00022726096338\times10^{-4}$\\\hline
$11$ & $0.01492959925020\times10^{-4}$\\\hline
$16$ & $0.07644765777970\times10^{-4}$\\\hline
$20$ & $0.22545767650040\times10^{-4}$\\\hline
\end{tabular}
\end{center}
Regarding the results given in the last table it is important to note that in
fact the weight of $u_{k}$ in the expansion of a solution decreases as
$\left( \frac{k+1}{2}\right) !$ or $\left( \frac{k}{2}\right) !$ for odd
or even $k$ respectively. This is due to the factor $1/n!$ in the definition
of the Taylor coefficients (\ref{Taylorcoef}). That is in fact the real
accuracy of calculation of the base functions would be given by $\left\vert
u_{k}-\widetilde{u}_{k}\right\vert /K!$ where $K=\left\{
\begin{array}
[c]{c}%
\left( \frac{k+1}{2}\right) !\text{ for }k\text{ odd}\\
\left( \frac{k}{2}\right) !\text{ for }k\text{ even}%
\end{array}
\right. $. It is easy to see that in the case of the results given in Table 7
one then obtains $\left\vert u_{20}-\widetilde{u}_{20}\right\vert
/10!\simeq6.213\times10^{-12}$ which is a remarkably good agreement.
For the numerical computation of $\widetilde{u}_{k}$ we implemented the
following procedure. Before integrating on each new step according to
(\ref{recformula}) along segments joining the center of the formal powers with
points on the boundary of the domain the integrand was represented as a cubic
spline which then was integrated using the standard Matlab routine for
integration of splines. This procedure is simple but clearly not optimal.
Nevertheless the approximate results presented in this work show that even
such integration procedure gives satisfactory agreement between the exact base
functions and those calculated numerically.
The accuracy of the approximate solution obtained with the aid of
$\widetilde{u}_{k}$ in our numerical tests did not differ significantly from
that of the solution obtained using the exactly calculated $u_{k}$. The order
of the maximum absolute error for a given $N$ and $c$ coincided in both cases.
Hence here we present results corresponding to another test problem for which
we did not have the exactly calculated base functions.
Consider the equation
\begin{equation}
\left( -\Delta+\frac{e^{y}}{4}\right) u(x,y)=0.\label{secondexample}%
\end{equation}
An exact solution for this equation can be found using the fact that the
change of variables $\xi=e^{\frac{y}{2}}\cos\frac{x}{2}$, $\eta=e^{\frac{y}%
{2}}\sin\frac{x}{2}$ leads to (\ref{Yukawa}) in the new variables. Thus, e.g.,
the function $u(x,y)=\exp(e^{\frac{y}{2}}\cos\frac{x}{2})$ is an exact
solution of (\ref{secondexample}). Consequently, as a test problem we can
consider the problem of finding a solution of (\ref{secondexample}) in
$\Omega$, satisfying the boundary condition
\begin{equation}
u(x,y)=\exp(e^{\frac{y}{2}}\cos\frac{x}{2}),\quad(x,y)\in\partial
\Omega.\label{test2}%
\end{equation}
In order to construct a particular solution $u_{0}$ in a separable form we
solve numerically the ordinary differential equation
\[
\left( -\frac{d^{2}}{dy^{2}}+\frac{e^{y}}{4}\right) u_{0}(y)=0.
\]
The obtained solution we then use for constructing the system of functions
$\left\{ u_{k}\right\} $. Some results on the accuracy of the approximate
solution are given in the following table.
\begin{center}
\bigskip Table 9. Maximum absolute error of the approximate solution of the
test problem (\ref{secondexample}), (\ref{test2}) considered in a unitary disk
in dependence on $N$%
\begin{tabular}
[c]{|c|c|}\hline
$N$ & Maximum absolute error\\\hline
$4$ & $0.079$\\\hline
$6$ & $0.021$\\\hline
$8$ & $0.005$\\\hline
$10$ & $0.001$\\\hline
$12$ & $0.0004$\\\hline
$14$ & $0.000099$\\\hline
$16$ & $0.000020$\\\hline
$18$ & $0.0000033$\\\hline
$20$ & $0.00000060$\\\hline
$28$ & $0.00000000072$\\\hline
$32$ & $0.00000000028$\\\hline
\end{tabular}
\end{center}
\section{Approximate solution of eigenvalue problems\label{SectEigen}}
In this section we consider the application of MPFP to solution of eigenvalue
problems for operators of the form $-\Delta+q(y)$. \ For simplicity we keep
working with the Dirichlet boundary conditions and suppose that $q$ is
continuous and $q(y)\geq0$ in $\overline{\Omega}$. Then the spectrum of the
operator is discrete and positive. As was explained in section
\ref{SectDescription} for the equation%
\begin{equation}
(-\Delta+q(y))u_{0}(x,y)=\lambda^{2}u_{0}(x,y)\label{Schrodwitheigen}%
\end{equation}
it is easy to propose a particular solution in a separable form for any value
of $\lambda$. We are interested here in positive values, and hence a natural
choice of a nonvanishing solution would be $u_{0}(x,y)=e^{i\lambda x}h(y)$
where $h(y)$ is a positive solution of (\ref{ordinary}). As was observed in
section \ref{SectDescription} the completeness of the system of solutions
$\left\{ u_{k}\right\} _{k=0}^{\infty}$ obtained in this case is up to now
an open problem due to the fact that $u_{0}$ is complex valued and one should
consider bicomplex pseudoanalytic formal powers for which the whole theory is
still underdeveloped. Nevertheless we used the constructed system of exact
solutions $\left\{ u_{k}\right\} _{k=0}^{\infty}$ for finding the
eigenvalues $\lambda^{2}$ in the following way. Assuming that $\left\{
u_{k}\right\} _{k=0}^{\infty}$ is complete in the same sense as was proved in
the case of the real-valued particular solution $u_{0}$ (subsection
\ref{SubsectConvergence}) we have then that if a nontrivial solution $u$ of
(\ref{Schrodwitheigen}) exists satisfying the boundary condition $\left.
u\right\vert _{\partial\Omega}=0$ then $u\simeq\sum_{k=0}^{N}b_{k}u_{k}$ and
the coefficients $b_{k}$ are such that the trivial boundary condition is
approximately fulfilled. This means that one can require that $\sum_{k=0}%
^{N}b_{k}u_{k}(z_{j})=0$ for $z_{j}\in\partial\Omega$ and $j=\overline{0,N}$.
This is possible iff the determinant of the matrix $U=\left( u_{jk}\right)
_{j,k=0}^{N}$ vanishes where $u_{jk}$ $=u_{k}(z_{j})$. The determinant of $U$
for a fixed $N$ is a function of $\lambda$. Thus, the problem of finding
eigenvalues reduces to the problem of finding zeros of the function $\det
U(\lambda)$.
As a test problem we considered the problem of calculating the eigenvalues of
the Dirichlet problem for the Helmholtz equation $\left( \Delta+\lambda
^{2}\right) u=0$. For every $\lambda$ \ a system of exact solutions $\left\{
u_{k}\right\} _{k=0}^{\infty}$ can be constructed using (\ref{uk}) where $c$
should be replaced by $i\lambda$. Then following the described scheme we
looked for zeros of $\det U(\lambda)$. As it is well known (see, e.g.,
\cite{Coleman}) the eigenvalues of the Dirichlet problem for the Helmholtz
equation in a unitary disk are squares of zeros of Bessel functions $J_{n}%
(x)$. Our numerical experiments showed that a relatively small value of $N=21$
was needed for computing the first five eigenvalues with the accuracy of four
decimals. With $N=23$ we obtained six first eigenvalues with the same
accuracy. Thus, indeed, the MPFP is clearly competitive in solving eigenvalue
problems for elliptic operators. In applying the MPFP we detected a similar
problem to that described by Alexidze in \cite[Sect. 1.13]{Alexidze} where the
method of fundamental solutions (or auxiliary sources) was applied to
eigenvalue problems. The considered determinant shows a very fast decrement
(in spite of this the method gives good numerical results). \cite{Alexidze}
contains references to other publications where different ways of using the
knowledge of a system of exact solutions for numerical solution of eigenvalue
problems were studied. In this direction further research is needed.
\section{Conclusions}
A new approach for solving boundary value and eigenvalue problems for elliptic
operators in bounded planar domains is proposed. It is based on some classical
and some new results from pseudoanalytic function theory which allow one to
construct complete systems of solutions of the elliptic equations. We showed
the practical applicability of the numerical method based on this
construction, studied the rate of its convergence, accuracy and other
parameters of its performance.
|
\section*{Introduction}
Consider the group~$\mathbb{G} = GL_n(\mathbb{F}_q)$ of invertible matrices over the finite field~$\mathbb{F}_q$. Denote by~$\mathbb{B}$ its subgroup of upper triangular matrices, and by~$\mathbb{T}$ its subgroup of diagonal matrices. Set~$\varepsilon = \frac{1}{|\mathbb{B}|}\sum_{b\in \mathbb{B}}b$ in~$\mathbb{C}[\mathbb{G}]$. The quotient group~$N_\mathbb{G}(\mathbb{T})/\mathbb{T}$ is isomorphic to the symmetric group~$S_n$. Moreover, the Iwohori-Hecke $\mathbb{C}$-algebra $\mathcal{H}(\mathbb{G},\mathbb{B}) = \varepsilon \mathbb{C}[\mathbb{G}]\varepsilon$ is isomorphic to $\oplus_{w\in S_n}\mathbb{C}w$ as a $\mathbb{C}$-vector space, and the structure constants in the multiplicative table lie in $\mathbb{Z}[q]$. More generally, if $G$ is a finite reductive group over~$\overline{\mathbb{F}}_q$, $B$ is a Borel subgroup of $G$, and $T$ a maximal torus included in~$B$, then~$N_G(T)/T$ is a Weyl group and the above results extend to the Hecke algebra $\mathcal{H}(G,B)$. Now, consider a \emph{finite reductive monoid}~$M$ over~$\overline{\mathbb{F}}_q$ as defined by Renner in~\cite{Ren2}. Such a monoid is a unit regular monoid and its unit group is a finite reductive group~$G$. Solomon introduced in~\cite{Sol} the notion of a \emph{Iwahori-Hecke algebra}~$\mathcal{H}(M,B)$ of a finite reductive monoid~$M$. Here,~$B$ is a Borel subgroup of~$G$. This $\mathbb{C}$-algebra is defined by $\mathcal{H}(M,B) = \varepsilon \mathbb{C}[M]\varepsilon$ where as before~$\varepsilon = \frac{1}{|B|}\sum_{b\in B}b$ in~$\mathbb{C}[M]$. In this framework, the Weyl group is replaced by an inverse monoid~$R$, which is called the \emph{Renner monoid} of~$M$. Its turns out that~$\mathcal{H}(M,B)$ is isomorphic to $\oplus_{r\in R}\mathbb{C}r$ as a $\mathbb{C}$-vector space. An isomorphism is given by $r\mapsto \tilde{T}_r = \sum_{x\in BrB}x$. Therefore, this is natural to address the question of the existence of a normalisation~$T_r = a_r\tilde{T}_r$ of the basis~$(\tilde{T}_r)_{r\in R}$ such that in this new basis~$(T_r)_{r\in R}$, the structure constants in the multiplicative table lie in $\mathbb{Z}[q]$ as in the case of finite reductive groups. Solomon considered this question in~\cite{Sol} and answered in the positive in the specific case where~$M = M_n(\mathbb{F}_q)$. In~\cite{Sol1}, he announced that in a forthcoming paper, he was going to extend his result and its proof to every finite reductive monoid that arises as the set of fixed points of a \emph{reductive monoid} over~$\overline{F}_q$ (see Section~\ref{sousect} for a definition) by the Frobenius map~$\sigma$ defined by $\sigma(x_{i,j}) = x_{i,j}^q$. But it seems that this result has never be published. In~\cite{Put} Putcha proves that for every finite reductive monoid, one can normalised the basis~$(\tilde{T}_r)_{r\in R}$ such that the structure constants become rational in~$q$. Howewer, the question remained open, and Renner concluded in~\cite[sec.~8.3]{Ren} that ``the delicate part here is obtaining integral structure constants''. The main object of this article is to answer Solomon's question in the positive for every finite reductive monoid. We prove:
\begin{The}
Let $M$ be a finite reductive monoid over~$\overline{\mathbb{F}}_q$. Denote by $R$ the associated Renner monoid. There exists a normalisation of the basis~$(\tilde{T}_r)_{r\in R}$ of the Iwahori-Hecke algebra~$\mathcal{H}(M,B)$ such that the structure constants in the multiplicative table lie in~$\mathbb{Z}[q]$. Moreover, the coefficients of the polynomials depend on~$R$ only. \label{THintro}
\end{The}
In Section~\ref{sectionprincipale}, we provide explicit formulae (see Theorem~\ref{Th:genThe}), which are related to the existence of a length function on~$R$. Moreover, we deduce a finite presentation of~$\mathcal{H}(M,B)$ in the spirit of the classical presentation of~$\mathcal{H}(G,B)$ (see Corollary~\ref{THconclu} in Section~\ref{sousectionrappelamt}).
Mokler, Renner and Putcha consider families of monoids that are closed to reductive monoids (see \cite{Mok2,Mok3,Mok,Put2,Put3,PutRen} for instance. They are called \emph{finite monoids of Lie type} and \emph{face monoids}. Indeed, finite reductive monoids are special cases of finite monoids of Lie type. To each of these groups can be associated a so-called \emph{Renner monoid}, whose properties are closed to Renner monoids of (finite) reductive monoids (See Examples~\ref{exe:ex2} and~\ref{exe:ex3} below). This explains why these monoids are still called Renner monoids in the latter references. However, there is some differences between these monoids (see Remark~\ref{Rem:diffrenner} for a discussion). We introduce here the notion of a \emph{generalised Renner monoid}. All Renner monoids are examples of generalised Renner monoids. One motivation for this definition is to introduce a notion that plays for these various Renner monoids the role of the notion of a Coxeter system for Weyl group. We prove that all the properties shared by the various Renner monoids hold for generalised Renner monoid. In particular, it is a factorisable monoid and its unit group~$G$ is a Coxeter group. The crucial point regarding Solomon's question is that we can associate with each such generalised Renner monoid~$R$ a \emph{generic Hecke algebra}~$\mathcal{H}(R)$ which is a ring on the free $\mathbb{Z}[q]$-module with basis~$R$. It turns out that Theorem~\ref{THintro} is a consequence of
\begin{The}
Let $M$ be a finite reductive monoid over~$\overline{\mathbb{F}}_q$ with Renner monoid~$R$. The Iwahori-Hecke algebra~$\mathcal{H}(M,B)$ is isomorphic to the $\mathbb{C}$-algebra~$\mathbb{C}\otimes_\mathbb{Z}\mathcal{H}_q(R)$, where $\mathcal{H}_q(R)$ is the specialisation of the generic Hecke algebra~$\mathcal{H}(R)$ at $q$. \label{THintro2}
\end{The}
The second main ingredient used in the proof of Theorem~\ref{THintro} is the existence of a length function~$\ell$ on every generalised Renner monoid~$R$. This length function is related to the canonical generating set~$S\cup\Lambda$, which equips every generalised Renner monoid. In the case of reductive monoids, we investigate the relation of this length function with the product of double classes. We prove in particular that
\begin{Prop} Let $M$ be a reductive monoid with unit group~$G$ and Renner monoid~$R$. Fix a maximal torus~$T$ and a Borel subgroup~$B$ that contains~$T$ in $G$.\\(i) Let~$r$ lie in~$R$ and~$s$ lie in~$S$, then $$B s B r B = \left\{\begin{array}{ll}BrB,&\textrm{if } \ell(sr) = \ell(r);\\BsrB,&\textrm{if } \ell(sr) = \ell(r)+1;\\BsrB\cup BrB,&\textrm{if } \ell(sr) = \ell(r)-1.\end{array}\right.$$ \label{Pr:lienlongBBintro}
\\(ii) Let~$r$ lie in~$R$ and~$e$ lie in~$\Lambda$, then then $$B e B r B = BerB \textrm{ and } BrBeB = BreB$$ \end{Prop}
This result extends results obtained in~\cite{God,God2}, and leads to a similar result for finite reductive monoids.
The paper is organised as follows. In Section~\ref{Sec:genheckalgr}, we introduce the notion of a \emph{generalised Renner monoid}, provide examples and investigate properties of such monoids. In particular, we define the length function~$\ell$ and prove that a \emph{generic Hecke algebra} can be associated with every generalised Renner monoid. In Section~\ref{sectionprincipale}, we first recall the notion of a reductive monoid and prove Proposition~\ref{Pr:lienlongBBintro}. Then we introduce the notion of a Iwahori-Hecke algebra in the context of monoid theory. We prove some motivating general results for such algebras. These results are probably well-known by semigroup experts, but we have not be able to find references for them. Finally, we turn to finite reductive monoids and conclude with the proof of Theorem~\ref{THintro} and~\ref{THintro2}.
\section{Generic Hecke algebra}
\label{Sec:genheckalgr}
The notion of a Coxeter group has been introduced in order to study Weyl groups. Our objective in this section is to develop a similar theory for Renner monoids. We need first to recall some standard notions and introduce useful notations.
\subsection{Basic notions and notations}
We refer to~\cite{How} for a general introduction on Semigroup Theory, and to~\cite{Fit} for a survey on factorisable inverse monoids. We refer to~\cite{Bou} for general theory and proofs on Coxeter systems.
\subsubsection{Background on Semigroup Theory}
If $M$ is a monoid, we let $E(M)$ and $G(M)$ its idempotent set and its unit group. We see a (\emph{lower}) \emph{semi-lattice} as a commutative idempotent semigroup where $a\leq b$ iff $ab = ba = a$. In particular, $a\land b = ab$. A semigroup is \emph{unit regular} if $M = E(M)G(M) = G(M)E(M)$, and it is \emph{factorisable} if it is unit regular and $E(M)$ is a semi-lattice. In this latter case $M$ is \emph{invertible}, that is for every $x$ in $M$ there exists a unique $y$ in $M$ such that $xyx= x$ (and therefore $yxy = y$).
\subsubsection{Background on Coxeter Group Theory}
\label{sousectionrappelcgtintro}
\begin{definition} Let $\Gamma$ be a finite simple labelled graph whose labels are positive integers greater or equal than~$3$. We let denote~$S$ the vertex set of $\Gamma$. We let $\mathcal{E}(\Gamma)$ denote the set of pairs $(\{s,t\},m)$ such that either $\{s,t\}$ is an edge of $\Gamma$ labelled by~$m$, or $\{s,t\}$ is not an edge of $\Gamma$ and $m=2$. When $(\{s,t\},m)$ belongs to~$\mathcal{E}(\Gamma)$, we let $|s,t\rangle^m$ denote the word~$sts\cdots$ of length~$m$. The \emph{Coxeter group}~$W(\Gamma)$ associated with~$\Gamma$ is defined by the following group presentation $$\left\langle S \left| \begin{array}{ll}s^2 = 1&s\in S\\ |s,t\rangle^m = |t,s\rangle^m &(\{s,t\},m)\in \mathcal{E}(\Gamma) \end{array}\right.\right\rangle$$
In this case, one says that the pair~$(W(\Gamma),S)$ is a \emph{Coxeter system}, and that $W$ is a Coxeter group. The Coxeter graph is uniquely defined by the Coxeter system. \end{definition}
\begin{definition} Let~$(W,S)$ be a \emph{Coxeter system}.\\ (i) Let $w$ belong to $W$. The \emph{length}~$\ell(w)$ of $w$ is the minimal integer $k$ such that $w$ has a word representative of length~$k$ on the alphabet~$S$. Such a word is called a \emph{minimal word representative} of~$w$.\\(ii) The subgroup~$W_I$ generated by a subset~$I$ of $S$ is called a \emph{standard parabolic subgroup} of $W$.\end{definition}
A key tool in what follows is the following classical result.
\begin{Prop}\cite{Bou} Let $(W,S)$ be a \emph{Coxeter system} with Coxeter graph~$\Gamma$.\\(i) For every~$I\subseteq S$, the pair $(W_I,I)$ is a Coxeter system. Its graph~$\Gamma_I$ is the full subgraph of $\Gamma$ spanned by~$I$. \\(ii) For every~$I,J\subseteq S$ and every element $w\in W$ there exists a unique element~$\hat{w}$ of minimal length in the double-class $W_JwW_I$. Furthermore there exists $w_1$ in $W_I$ and $w_2$ in $W_J$ such that $w = w_2\hat{w}w_1$ with $\ell(w) = \ell(w_1)+\ell(\hat{w})+\ell(w_2)$. \label{proppopa2}
\end{Prop}
Note that~$(ii)$ holds, in particular, when $I$ or $J$ are empty. The element~$\hat{w}$ is said to be \emph{$(I,J)$-reduced}. In the sequel, we let $\textrm{Red}(I,J)$ denote the set of \emph{$(I,J)$-reduced} elements. Note also that the pair~$(w_1,w_2)$ is not unique in general, but it becomes unique if we require that $w_2\hat{w}$ is $(\emptyset,J)$-reduced (or that $\hat{w}w_1$ is $(I,\emptyset)$-reduced).
\subsection{Generalised Renner monoids}
\subsubsection{Generalised Renner-Coxeter System}
If $R$ is a factorisable monoid and $e$ belongs to $E(R)$ we let $W(e)$ and $W_\star(e)$ denote the subgroups defined by $$W(e) = \{w\in G(R)\mid we = ew\}$$ $$W_\star(e) = \{w\in G(R)\mid we = ew = e\}.$$ The unit group~$G(R)$ acts on $E(R)$ by conjugacy.
\begin{definition} (i) An \emph{generalised Renner-Coxeter system} is a triple~$(R,\Lambda,S)$ such that
\begin{enumerate}
\item[(ECS1)] $R$ is a factorisable monoid;
\item[(ECS2)] $\Lambda$ is both a transversal of $E(R)$ for the action of $G(R)$ and a sub-semi-lattice;
\item[(ECS3)] $(G(R),S)$ is a Coxeter system;
\item[(ECS4)] for every pair $e_1\leq e_2$ in $E(R)$ there exists $w$ in $G(R)$ and $f_1\leq f_2$ in~$\Lambda$ such that $wf_iw^{-1} = e_i$ for~$i=1,2$;
\item[(ECS5)] for every $e$ in $\Lambda$, the subgroups~$W(e)$ and $W_\star(e)$ are standard Coxeter subgroups of $G(R)$;
\item[(ECS6)] the map $e\in\Lambda\mapsto \lambda^\star(e) = \{s\in S\mid se = es \neq e\}$ is not decreasing: $e\leq f \implies \lambda^\star(e)\subseteq \lambda^\star(f)$.
\end{enumerate}
In this case, we say that $R$ is a \emph{generalised Renner monoid}. Following the standard terminology for Renner monoids, we call the section~$\Lambda$ the \emph{cross section lattice} of $R$, and we define the~\emph{type map} of $R$ to be the map~$\lambda :\Lambda\to S$ defined by $W(e) = W_{\lambda(e)}$.
\end{definition}
\begin{Not}
for $e$ in $\Lambda$, we set $$\lambda_\star(e) = \{s\in S\mid se = es = e\}$$ $$W^\star(e) = W_{\lambda^\star(e)}$$
\end{Not}
\begin{Rem} Assume $(R,\Lambda,S)$ is a generalised Renner-Coxeter system.\\
(i) Since $W_\star(e)$ is a standard Coxeter subgroup of $W(e)$, we have $$W_\star(e) = W_{\lambda_\star(e)}.$$ Moreover, This is clear that $W_\star(e)$ is a normal subgroup of~$W(e)$. As a consequence, $$W(e) = W_\star(e)\times W^\star(e)\textrm{ and }\lambda(e) = \lambda_\star(e)\cup\lambda^\star(e).$$(ii) Below, several results can be proved without assuming Property~(ECS6). However this is a crucial tool in the proof of Theorem~\ref{Th:genThe} and Proposition~\ref{Prop:inverprese}.\\(iii) If $E(R)$ is finite and a lower semi-lattice, then it has to be a lattice. This is so for \emph{Renner monoids} associated with \emph{reductive monoids}.\\(iv) the map $\lambda_\star$ is not increasing: $$e\leq f \implies \lambda_\star(f)\subseteq \lambda_\star(e).$$(v) We can have $\lambda_\star(e) = \lambda_\star(f)$ and $\lambda^\star(e) = \lambda^\star(f)$ for $e\neq f$ (see \cite[Sec.~2.3]{God2}). \label{remdepar}
\end{Rem}
Now we provide some examples of generalised Renner monoids.
\begin{Exe} Let $M$ be a reductive monoid (see Section~\ref{sousect} for a definition, and Example~\ref{exerook}). The associated Renner\label{exe:ex1} monoid~$R(M)$ of~$M$ is a generalised Renner monoid by \cite{Ren}.
\end{Exe}
\begin{Exe}
Let $M$ be a \emph{abstract finite monoid of Lie type} (see~\cite{Put2},\cite{PutRen} or \cite{Ren} for a definition. Note that these groups are called \emph{regular split monoids} in~\cite{Put2}, and \emph{finite monoids of Lie type} in~\cite{PutRen}). The associated \emph{Renner monoid}~$R(M)$ of~$M$ is a generalised Renner monoid. \label{exe:ex2} Property~(ECS6) follows from~\cite[Cor.~3.5(i)]{Put2}. The other defining properties hold by~\cite[Sec.~2]{Put3}. The seminal examples of an abstract finite monoid of Lie type is a Renner monoid of a \emph{finite reductive monoid}~\cite{Ren2}. In Section~3 we focus on these monoids.
\end{Exe}
\begin{Exe} \label{exe:ex3} Let~$G$ be a Kac-Moody group over a field~$\mathbb{F}$ of characteristic zero whose derived group is the special Kac-moody group introduced in~\cite{KaPe,KaPe2}. Denote by $(W,S)$ the associated Coxeter system. The Coxeter group~$W$ is infinite. Let~$Fa(X)$ be the set of \emph{faces} of its associated Tits cone~$X$ (see~\cite{Mok2} for details). The action of $W$ on $X$ induces an action on the lattice $Fa(X)$. The \emph{Renner monoid}~$R$ is the monoid~$W \ltimes Fa(X) /\sim$ where $\sim$ is the congruence on~$W\ltimes Fa(X)$ defined by $(w,R)\sim (w',R')$ if $R = R'$ and $w'^{-1}w$ fixes $R$ pointwise~\cite{Mok2}. Then $R$ is a generalised Renner monoid. Properties (ECS1), (ECS2), (ECS3) and (ECS5) are proved in~\cite{Mok2} (see also~\cite{Mok}). The cross section lattice~$\Lambda$ can be identified with the set of infinite standard parabolic subgroups of $W$ that have no finite proper normal standard parabolic subgroups. The semi-lattice structure is given by~$W_I\leq W_J$ if $J\subseteq I$. If $\Theta$ belongs to $\Lambda$, then $\lambda_\star(\Theta) = \Theta$ and $\lambda^\star(\Theta) = \{s\in S\mid \forall t\in \Theta, st = ts \}$. The latter equality clearly implies (ECS6). Finally, Property~(ECS4) can be deduced from~\cite[Theorem 2 and 4]{Mok}.
\end{Exe}
\begin{Rem} \label{Rem:diffrenner} In Examples~\ref{exe:ex1},~\ref{exe:ex2} and~\ref{exe:ex3} we provide examples of generalised Renner monoids that are all called \emph{Renner monoid} in the literature. From our point of view, this is not a suitable terminology since there is crucial differences between these monoids. Therefore, using the same terminology may be misleading. For instance, for Renner monoids of reductive monoids one has~$\lambda_\star(e) = \bigcap_{f\leq e} \lambda(f)$ and~$\lambda^\star(e) = \bigcap_{f\geq e} \lambda(f)$. This is not true in general for Renner monoids associated with abstract finite monoids of Lie type (see~\cite{PutRen} for a details). In Renner monoids of reductive monoids and of abstract finite monoids of Lie type, all maximal chains of idempotents have the same size. This is not true for Renner monoids of example~\ref{exe:ex3}, as explained in~\cite{Mok2}.
\end{Rem}
\subsubsection{Presentation for generalised Renner monoids}
For all this section, we fix a generalised Renner-Coxeter system~$(R,\Lambda,S)$. We let~$W$ denote the unit group of~$R$. Our objective is to prove that important properties shared by Renner monoids of Examples~\ref{exe:ex1}, \ref{exe:ex2}, \ref{exe:ex3} can be deduced from their generalised Renner-Coxeter system structure. In particular, we extend to this context the results obtained in~\cite{God2}. By Proposition~\ref{proppopa2}, For every~$w$ in~$W$ and every~$e,f$ in~$\Lambda$, each of the sets~$w\sn(e)$, $\sn(e) w$, $w\sn_\star(e)$, $\sn_\star(e) w$ and $\sn(e) w \sn(f)$ has a unique element of minimal length. In order to simplify notation, we set $\Dre{e} = \textrm{Red}(\emptyset;\lambda(e))$, $\Gae{e} = \textrm{Red}(\lambda(e),\emptyset)$; $\DrTe{e} = \textrm{Red}(\emptyset,\lambda_\star(e))$; $\GaTe{e} = \textrm{Red}(\lambda_\star(e),\emptyset)$; $\textrm{Red}(e,f) = \textrm{Red}(\lambda(e),\lambda(f))$.
\begin{Prop}\label{fnrenner} For every~$r$ in~$R$,\\(i) there exists a unique triple~$(w_1,e,w_2)$ with~$e\in \Lambda$,~$w_1\in \DrTe{e}$ and~$w_2\in \Gae{e}$ such that~$r = w_1ew_2$;\\(ii) there exists a unique triple~$(v_1,e,v_2)$ with~$e\in \Lambda$,~$w_1\in \Dre{e}$ and~$w_2\in \GaTe{e}$ such that~$r = v_1ev_2$ \end{Prop}
Following~\cite{Ren}, we call the triple~$(w_1,e,w_2)$ the \emph{normal decomposition} of $r$.
\begin{proof} Let us prove~(i). The proof of~(ii) is similar. Let $r$ belong to the monoid~$R$. By Property~(ECS1), there exists $e$ in~$E(R)$ and $w$ in $W$ such that $r = ew$. By Property~(ECS2) there exists $e_1$ in $\Lambda$ and $v$ in~$W$ such that $e = ve_1v^{-1}$. Then $r = vew_1$ with $w_1 = v^{-1}w$. By Remark~\ref{remdepar}(i), we can write $v = v_1v'_1$ and $w_1 = w'_2w''_2w_2$ with $v_1$, $w_2$, $v'_1$, $w'_2$ and $w''_2$ in $\DrTe{e}$, $\Gae{e}$, $W_\star(e)$, $W^\star(e)$ and $W_\star(e)$, respectively. Then we have $r = v_1w'_2ew_2$, and $v_1w'_2$ belongs to $\DrTe{e}$, still by Remark~\ref{remdepar}(i). Now assume~$r = w_1ew_2 = v_1fv_2$ with $e,f$ in $\Lambda$, $w_1,v_1$ in~$\DrTe{e}$ and~$\DrTe{f}$, respectively, and $w_2,v_2$ in~$\Gae{e}$ and in~$\Gae{f}$, respectively. Then $(w_1w_2)w_2^{-1}ew_2 = (v_1v_2)v_2^{-1}fv_2$. This implies $w_2^{-1}ew_2 = v_2^{-1}fv_2$ by~\cite{Fit}. As a consequence, $e = f$ and $v_2w_2^{-1}$ lies in $W(e)$. Since $v_2$ and $w_2$ both belong to $\Gae{e}$, we must have $v_2 = w_2$. Now, it follows that $w_1e = v_1e$ and $w_1^{-1}v_1$ lies in~$W_\star(e)$. This implies $w_1 = v_1$ in $\DrTe{e}$.
\end{proof}
\begin{Lem} Let $e,f$ belong to $\Lambda$ and $w$ lie in $\textrm{Red}(e,f)$.\\\label{lemclef}
(i) There exists $h$ in $\Lambda$ such that $w$ belongs to~$W(h)$ and $ewf = wh$.\\
(ii) The element~$w$ lies in~$W_\star(h)$. Therefore, $wh = h$.
\end{Lem}
Note that in the above lemma we have $h\leq e\wedge f = ef$. In the sequel the element~$h$ is denoted by~$e\wedge_wf$.
\begin{proof} The proof is similar to~\cite[Prop~1.21]{God2}. (i) Consider the normal decomposition~$(w_1,h,w_2)$ of $ewf$. By definition~$w_1$ belongs to~$\DrTe{h}$ and $w_2$ belongs to~$\Gae{h}$. The element $w^{-1}ewf$ is equal to $w^{-1}w_1hw_2$ and belongs to $E(R)$. Since $w_2$ lies in~$\Gae{h}$, this implies that $w_3 = w_2w^{-1}w_1$ lies in $W_\star(h)$, and that $f\geq w_2^{-1}hw_2$. By Property~(ECS4), there exists $w_4$ in $W$ and $f_1,h_1$ in $\Lambda$, with $f_1\geq h_1$, such that $w_4^{-1}f_1w_4 = f$ and $w_4^{-1}h_1w_4 = w_2^{-1}hw_2$. Since $\Lambda$ is a cross section for the action of $W$, we have~$f_1 = f$ and~$h_1 = h$. In particular, $w_4$ belongs to~$W(f)$. Since $w_2$ belongs to $\Gae{h}$, we deduce that there exists $r$ in $W(h)$ such that $w_4 = rw_2$ with $\ell(w_4) = \ell(w_2)+\ell(r)$. Then $w_2$ lies in $W(f)$, too. Now, write $w_1 = w'_1w''_1$ where $w''_1$ lies in $W^\star(h)$ and $w'_1$ belongs to $\Dre{h}$. One has $ewf = w'_1hw''_1w_2$, and $w_1w''_2$ lies in~$\GaTe{h}$. By symmetry, we get that $w'_1$ belongs to $W(e)$. The element~${w'}^{-1}_1ww^{-1}_2$ is equal to $w''_1{w}_3^{-1}$ and belongs to $W(h)$. But, by hypothesis $w$ lies in~$\textrm{Red}(e,f)$. Then we must have $\ell(w''_1{w}_3^{-1}) = \ell({w'}^{-1}_1)+\ell(w)+\ell({w}^{-1}_2)$. Since $w''_1{w}_3^{-1}$ belongs to $W(h)$, it follows that $w'_1$ and $w_2$ belong to $W(h)$ too. This implies $w_2 = w'_1 = 1$ and $w = w''_1w_3^{-1}$. Therefore, $ewf = hw''_1 = hw = wh$.\\(ii) This is a direct consequence of the following fact: for $h,e$ in $\Lambda$ such that $h \leq e$, we have $W(h)\cap \Gae{e}\subseteq W_\star(h)$ and $W(h)\cap \Dre{e}\subseteq W_\star(h)$. Assume $w$ lies in~$W(h)\cap \Dre{e}$, then we can write $w = w_1w_2 = w_2w_1$ where $w_1$ lies in $W_\star(h)$ and $w_2$ lies in $W^\star(h)$. Since $h\leq e$, we have $\lambda^\star(h)\subseteq\lambda^\star(e)$ and $W^\star(h)\subseteq W^\star(e)$. Since $w$ belongs to~$\Dre{e}$, this implies~$w_2 = 1$. The proof of the second inclusion is similar.
\end{proof}
\begin{Cor}
(i) For every chain $e_1\leq e_2\leq\cdots\leq e_m$ in $E(R)$ there exists $w$ in $G(R)$ and a chain $f_1\leq f_2\leq\cdots\leq f_m$ in~$\Lambda$ such that $wf_iw^{-1} = e_i$ for every index~$i$.\\
(ii) If $\Lambda$ has an infimum~$e$, then $\lambda(e) = S$.\\
(iii) For all $e,f$ in~$\Lambda$ and $w$ in~$\textrm{Red}(e,f)$, one has $$ewf = \max\{h\in \Lambda\mid h\leq e,\ h\leq f,\ w\in W(h)\} = fw^{-1}e.$$
\end{Cor}
In the case of Renner monoids of reductive monoids, the lattice~$\Lambda$ has an infimum~$e$ and $\lambda(e) = \lambda_\star(e) = S$. In other words, $e$ is a zero element of $R$.
\begin{proof}
(i) Assume $w_1e_1w_1^{-1}\leq \cdots \leq w_me_mw_m^{-1}$. We prove the result by induction on $m$. For $m = 2$ this is true by Property (ECS4). Assume $m\geq 3$. By induction hypothesis, we can assume $w_2 = \cdots =w_m$. We can also also assume that $w_1$ belongs to~$\Dre{e_1}$. By hypothesis, we have $w_1e_1w_1^{-1}w_2e_2w_2^{-1} = w_1e_1w_1^{-1}$. We can write $w_1^{-1}w_2 = v_1v_3v^{-1}_2$ with $v_1$ in $W(e_1)$, $v_2$ in $W(e_2)$ and $v_3$ in $\textrm{Red}(e_1,e_2)$. Then $w_1e_1w_1^{-1}w_2e_2w_2^{-1} = w_1v_1e_1v_3e_2v^{-1}_2w_2^{-1}$. If $v_3\neq 1$, then we get a contradiction by Lemma~\ref{lemclef}(i) and Proposition~\ref{fnrenner}. Then $v_2 = 1$ and $e_1e_2=e_1$. It follows that $w_1v_1 = w_2v_2$. Write $v_1 = v_{1\star}v^\star_1$ and $v_2 = v_{2\star}v^\star_2$ with $v_{i\star}$ in $W_\star(e_i)$ and $v_i^\star$ in $W^\star(e_i)$. We have $w_1v_{1\star}v^{-1}_{2\star} = w_2v^\star_2{v_1^\star}^{-1}$. Since~$\lambda_\star(e_2)\subseteq \lambda_\star(e_1)$ and~$\lambda^\star(e_1)\subseteq \lambda^\star(e_2)$, we get that~$v_{1\star}v^{-1}_{2\star}$ and~$v^\star_2{v^\star_1}^{-1}$ lie in $W(e_1)$ and $W(e_2)$, respectively. Then $w_1e_1w_1^{-1} = we_1w^{-1}$ and $w_2e_2w_2^{-1} = we_2w^{-1}$ with $w = w_1v_{1\star}v^{-1}_{2\star}$. But $W(e_2)\subseteq W(e_j)$ for $j\in\{2,\cdots,m\}$. Therefore, $w_2e_jw_2^{-1} = we_jw^{-1}$ for every~$j\geq 2$.\\
(ii) if $s\in S$ does not belong to~$\lambda(e)$, then $ese< e$ in $\Lambda$.\\
(iii) This is clear that~$e\wedge_wf$ lies in~$\{h\in \Lambda\mid h\leq e,\ h\leq f,\ w\in W(h)\}$. Now, if $h\in \Lambda$ verifies $h\leq e$, $h\leq f$, and $w\in W(h)$, then $h (e\wedge_wf) = hw^{-1}(ewf) = w^{-1}hwf = hf = h$. Therefore, $h\leq ewf$. The last equality follows form the fact that $w^{-1}$ belongs to $\textrm{Red}(f,e)$.
\end{proof}
\begin{Prop} For every $w$ in $W$, we fix an arbitrary reduced word representative~$\underline{w}$. We set $\Lambda_{\scriptscriptstyle \circ} = \Lambda\setminus\{1\}$. The monoid~$R$ admits the monoid presentation whose generating set is~$S\cup \Lambda_{\scriptscriptstyle \circ}$ and whose defining relations are:
\begin{center}\begin{tabular}{lll}
(COX1)&$s^2 = 1$,&$s\in S$;\\
(COX2)&$|s,t\rangle^m = |t,s\rangle^m$,&$(\{s,t\},m)\in \mathcal{E}(\Gamma)$;\\
(REN1)&$se = es$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda^\star(e)$;\\
(REN2)&$se = es = e$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda_\star(e)$;\\
(REN3)&$e\underline{w}f = e\!\wedge_w\!f$,& $e,f\in \Lambda_{\scriptscriptstyle \circ}$, $w\in \textrm{Red}(e,f)$.
\end{tabular}\end{center}
\label{proppres2}
\end{Prop}
\begin{proof} This is clear that the relations stated in the proposition hold in $R$. Conversely, every element~$r$ in~$R$ has a unique representing word~$\underline{w}e\underline{v}$ such that $(w,e,v)$ is its normal decomposition, and this is immediate that every representing word of~$r$ on~$S\cup \Lambda_{\scriptscriptstyle \circ}$ can be transformed into~$\underline{w}e\underline{v}$ using the given relations only.
\end{proof}
\begin{Rem} (i) The above presentation is not minimal in general. Some of the relations of type~(REN3) can be removed (see the proof of~\cite[Theorem~0.1]{God2} and Remark~\ref{Rem:rem2} below).\\
(ii) The reader may verify that the result of Proposition~\ref{proppres2} and its proof still hold if we do not assume Property~(ECS6), except that Relation (REN3) must be replace by
\begin{center}\begin{tabular}{lll}(REN3')&$e\underline{w}f = \underline{w} (e\wedge_wf)$,& $e,f\in \Lambda_{\scriptscriptstyle \circ}$, $w\in \textrm{Red}(e,f)$.
\end{tabular}\end{center} Indeed, Lemma~\ref{lemclef}(i) still hold.
\end{Rem}
One may wonder whether every monoid defined by a monoid presentation like in Proposition~\ref{proppres2}. The answer is positive under some necessary assumptions:
\begin{definition} A \emph{generalised Renner-Coxeter data} is $4$-uple~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$ such that~$\Gamma$ is a Coxeter graph with vertex set $S$, $\Lambda_{\scriptscriptstyle \circ}$ is a lower semi-lattice and~$\lambda^\star$, $\lambda_\star$ are two maps from $\Lambda_{\scriptscriptstyle \circ}$ to $S$ that verifies \begin{enumerate}
\item[(a)] for every $e$ in~$\Lambda_{\scriptscriptstyle \circ}$, the graphs spanned by $\lambda_\star(e)$ and $\lambda^\star(e)$ in~$\Gamma$ are not connected, and $$e\leq f \Rightarrow \lambda_\star(f)\subseteq\lambda_\star(e)\textrm{ and }\lambda^\star(e)\subseteq\lambda^\star(f).$$
\item[(b)] for every $f,g$ in~$\Lambda_{\scriptscriptstyle \circ}$ and every $w\in Red(f,g)$ the set $$\left\{e\in \Lambda_{\scriptscriptstyle \circ} \mid e\leq f,\ e\leq g\textrm{ and }w\in W_{\lambda(e)}\right\}$$ has a greatest element, denoted by $f\wedge_w g$.
\end{enumerate} \label{Prop:inverpreseee}
with $\lambda(e) = \lambda_\star(e)\cup \lambda^\star(e)$ for $e\in\Lambda_{\scriptscriptstyle \circ}$ and $\textrm{Red}(e,f) = \textrm{Red}(\lambda(e),\lambda(f))$ in the Coxeter group~$W(\Gamma)$ associated with~$\Gamma$.
\end{definition}
Note that properties (a) and (b) hold in every generalised Renner-Coxeter system. Actually, if~$\Lambda_{\scriptscriptstyle \circ}$ is any lower semi-lattice such that all maximal chains are finite, then Assumption~(b) is necessarily verified.
\begin{The} Assume~$M$ is a monoid. There exists a generalised Renner-Coxeter system~$(M,\Lambda,S)$ if and only if there exists a generalised Renner-Coxeter data~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$, where~$S$ is the vertex set of~$\Gamma$, such that $M$ admits the following monoid presentation
\begin{center}\begin{tabular}{lll}
(COX1)&$s^2 = 1$,&$s\in S$;\\
(COX2)&$|s,t\rangle^m = |t,s\rangle^m$,&$(\{s,t\},m)\in \mathcal{E}(\Gamma)$;\\
(REN1)&$se = es$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda^\star(e)$;\\
(REN2)&$se = es = e$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda_\star(e)$;\\
(REN3)&$e\underline{w}f = e\!\wedge_w\!f$,& $e,f\in \Lambda_{\scriptscriptstyle \circ}$, $w\in \textrm{Red}(e,f)$.
\end{tabular}\end{center}
Where~$\underline{w}$ is an arbitrary fixed minimal representing word of~$w\in W(\Gamma)$.\\ In this case, $W(\Gamma)$ is canonically isomorphic to the unit group of $M$, and $\Lambda_{\scriptscriptstyle \circ}$ embeds in $M$ with $\Lambda = \Lambda_{\scriptscriptstyle \circ}\cup\{1\}$.
\label{Prop:inverprese}\end{The}
Note that given a generalised Renner-Coxeter data~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$, Relations (COX1) and (COX2) implies that the monoid~$M$ defined by the presentation stated in Theorem~\ref{Prop:inverprese} does not depend on the chosen representing words~$\underline{w}$. Theorem~\ref{Prop:inverprese} follows from the following lemmas.
\begin{Lem} Consider a generalised Renner-Coxeter data~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$ and the monoid~$M$ defined by the presentation stated in Theorem~\ref{Prop:inverprese}. Then for every $f,g$ in~$\Lambda_{\scriptscriptstyle \circ}$ and every $w\in \textrm{Red}(f,g)$,\label{lem:propreci}
\begin{enumerate}
\item[($b_1$)] $e\wedge_1 f = e\wedge f$ and $e\wedge_w f \leq e\wedge f$;
\item[($b_2$)] $e\wedge_wf = f\wedge_{w^{-1}}e$;
\item[($b_3$)] $w\in W_{\lambda_\star(e\wedge_w f)}$.
\end{enumerate}
\end{Lem}
\begin{proof} Properties~$(b_1)$ and $(b_2)$ are immediate consequences of Assumption~(b). Properties~$(b_3)$ follows from Assumption~(a). The main argument is like in the proof of Lemma~\ref{lemclef}(ii). If $w$ doesnot belong to $W_{\lambda_\star(e\wedge_w f)}$, then we can write $w = w_\star w^\star$ with $w_\star\in W_{\lambda_\star(e\wedge_w f)}$ and $w^\star\in W_{\lambda^\star(e\wedge_w f)}$. But $W_{\lambda^\star(e\wedge_w f)}\subseteq W_{\lambda^\star(f)}$ and $w$ lies in~$\textrm{Red}(e,f)$. Therefore, $w^\star = 1$.
\end{proof}
\begin{Lem} Consider a generalised Renner-Coxeter data~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$ and the monoid~$M$ defined by the presentation stated in Theorem~\ref{Prop:inverprese}. Let $\textrm{F\!M\!}(S\cup \Lambda_{\scriptscriptstyle \circ})$ be the free monoid on~$S\cup \Lambda_{\scriptscriptstyle \circ}$, and $\equiv$ be the congruence on $F\!M\!(S\cup \Lambda_{\scriptscriptstyle \circ})$ generated by the defining relations of~$M$. Hence by definition,~$M$ is equal to $\textrm{F\!M\!}(S\cup \Lambda_{\scriptscriptstyle \circ})/\!\equiv$.\\
(i) If $\omega_1$ and $\omega_2$ are two words on~$S$ such that $\omega_1\equiv \omega_2$, then they represent the same element in $W(\Gamma)$.\\
(ii) If $e$ lie in $\Lambda_{\scriptscriptstyle \circ}$ and $\omega$ lie in $\textrm{F\!M\!}(S\cup \Lambda_{\scriptscriptstyle \circ})$ with $e\equiv \omega$, then the word~$\omega$ is equal to~$\nu_1e_1\nu_2\cdots e_k\nu_{k+1}$ where for every~$i$ we have $e\leq e_i$ in $\Lambda_{\scriptscriptstyle \circ}$ and $\nu_i$ are words on $S$ whose images in~$W(\Gamma)$ belong to $W_{\lambda(e)}$. Furthermore, the image of the word~$\nu_1\nu_2\cdots \nu_{k+1}$ in $W_{\lambda^\star(e)} = W_{\lambda(e)}/W_{\lambda_\star(e)}$ is trivial.\label{leminterm}
\end{Lem}
\begin{proof} In this proof we write $\omega_1 \dot{=} \omega_2$ if the two words $\omega_1$, $\omega_2$ are equals. If the words $\omega_1$ $\omega_2$ represent the elements $w_1, w_2$ in $M$, respectively, then $\omega_1 \dot{=} \omega_2$ implies $\omega_1\equiv \omega_2$ and $w_1 = w_2$. Conversely, $w_1 = w_2$ if and only if $\omega_1\equiv \omega_2$. Point~(i) is clear: if $\omega_1\equiv \omega_2$ then one can transform $\omega_1$ into $\omega_2$ using relations (COX1) and (COX2) only, since the words in both sides of Relations (REN1-3) contain letters in~$\Lambda_{\scriptscriptstyle \circ}$. Let us prove~$(ii)$. Write $\omega_1\equiv_1 \omega_2$ if one can transform $\omega_1$ into $\omega_2$ by applying one defining relation of $M$ on $\omega_1$. If $e\equiv \omega$, then there exists $\omega_0 \dot{=} e, \omega_1, \cdots, \omega_r \dot{=} \omega$ such that $\omega_i\equiv_1 \omega_{i+1}$. We prove the result by induction on $r$. If $r = 0$ we have nothing to prove. Assume $r\geq 1$. By induction hypothesis, $\omega_{r-1} \dot{=} \mu_1f_1\mu_2\cdots \mu_{j}f_j\mu_{j+1}$ with $e\leq f_i$ in $\Lambda_{\scriptscriptstyle \circ}$ and $\mu_i$ is a word on $S$ whose image in~$W(\Gamma)$ belongs to $W_{\lambda(e)}$, and the image of the word~$\mu_1\mu_2\cdots \mu_{j+1}$ in $W_{\lambda^\star(e)} = W_{\lambda(e)}/W_{\lambda_\star(e)}$ is trivial. We deduce the result for $\omega \dot{=} \omega_r$ by considering case by case the type of the defining relation applied to $\omega_{r-1}$ to obtain~$\omega_r$. The cases where the relation is of one of the types (COX1), (COX2) or (REN1) are trivial. The case where the relation is of type (REN2) follows from Property~(a) in Definition~\ref{Prop:inverpreseee}: by induction hypothesis, one has $\lambda_\star(f_i)\subseteq \lambda_\star(e)\subseteq\lambda(e)$. Finally, the case where the relation is of type (REN3) follows from properties~(a) and (b) by Lemma~\ref{lem:propreci}. If the image~$u_i$ of~$\mu_i$ in $W(\Gamma)$ belongs to~$\textrm{Red}(f_{i-1},f_{i})$ with $\mu_i = \underline{u}_i$ and~$\omega \dot{=} \mu_1f_1\cdots \mu_{i-1} (f_{i}\!\wedge_{u_i}\!f_{i+1}) \mu_{i+1}f_{i+2}\cdots f_j\mu_{j+1}$ then $e\leq f_{i-1}\!\wedge_{u_i}\!f_i$. Conversely, if $\omega = \mu_1f_1\mu_2\cdots f_{i-1}\mu_ie_{i}\underline{u}_ie_{i+1}\mu_{i+1}\cdots \mu_{j}f_j\mu_{j+1}$ where $f_i = e_{i}\!\wedge_{u_i}\!e_{i+1}$ for $e_i,e_{i+1}$ in $\Lambda_{\scriptscriptstyle \circ}$ and some $u_i$ in $\textrm{Red}(e_i,e_{i+1})$, then $e\leq f_i\leq e_i$ and $e\leq f_i\leq e_{i+1}$; Moreover,~$u_i$ belongs to $W_{\lambda_\star(f_i)}$, which is included in~$W_{\lambda_\star(e)}$. In all these cases the words~$\nu_1\nu_2\cdots \nu_{k+1}$ and~$\mu_1\mu_2\cdots \mu_{j+1}$ represent the same element in $W_{\lambda(e)}/W_{\lambda_\star(e)}$, which is trivial by induction hypothesis.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{Prop:inverprese}]
Consider a generalised Renner-Coxeter system~$(M,\Lambda,S)$. Denote by $\Gamma$ the Coxeter graph with vertex set~$S$ of the unit group of $M$, and set $\Lambda_{\scriptscriptstyle \circ} = \Lambda\setminus \{1\}$. It follows from previous results that~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$ is a generalised Renner-Coxeter data, and by Proposition~\ref{proppres2} that $M$ has the required monoid presentation. Conversely, consider a generalised Renner-Coxeter data~$(\Gamma,\Lambda_{\scriptscriptstyle \circ},\lambda_\star,\lambda^\star)$ and let~$M$ denote the monoid defined by the presentation stated in Theorem~\ref{Prop:inverprese}. By Lemma~\ref{leminterm}(i), the subgroup of $M$ generated by $S$ can be identified with~$W(\Gamma)$. Lemma~\ref{leminterm}(ii) implies that $\Lambda_{\scriptscriptstyle \circ}$ injects in~$M$, as a set. Let $e,f$ be in~$\Lambda_{\scriptscriptstyle \circ}$. In $M$ one has $ef = fe = e\wedge_1 f = e\wedge f$. Assume furthermore that $w$ lies in~$W(\Gamma)$. Lemma~\ref{leminterm}(ii) implies also that $(wew^{-1})f = wew^{-1}$ if and only if $e\leq f$ in~$\Lambda_{\scriptscriptstyle \circ}$ and $w$ lie in $W_{\lambda(e)}$. Let $wew^{-1}$ and $vfv^{-1}$ be in $E(M)$ with $e,f$ in~$\Lambda_{\scriptscriptstyle \circ}$. Write $w^{-1}v = v_1v_2v_3$ with $v_2$ in $\textrm{Red}(e,f)$, $v_1$ in $W_{\lambda(e)}$ and $v_3$ in $W_{\lambda(f)}$. Then $ev_2f = e\wedge_{v_2}f$ and $v_2$ lies in~$W_{\lambda_\star(e\wedge_{v_2}f)}$. We get, $$\displaylines{wew^{-1}vfv^{-1} = wv_1e\wedge_{v_2}fv_3v^{-1} = wv_1f\wedge_{v_2}^{-1}ev_3v^{-1} = wv_1v_2fv^{-1}_2ev_2v_3v^{-1} =\hfill\cr\hfill wv_1v_2v_3fv_3^{-1}v^{-1}_2v_1^{-1}ev_1v_2v_3v^{-1} = vfv^{-1}wew^{-1}.}$$ It is easy to see that every representing word~$\omega$ on $S\cup\Lambda_{\scriptscriptstyle \circ}$ of en element~$w$ of $M$ can be transformed into a word $\omega_1e\omega_2 \equiv \omega_1e\omega_1^{-1}\omega_1\omega_2$ where $e$ belongs to~$\Lambda = \Lambda_{\scriptscriptstyle \circ}\cup\{1\}$ and $\omega_1,\omega_2$ represent words in~$W(\Gamma)$. Moreover, if $\omega$ contains some letter in~$\Lambda_{\scriptscriptstyle \circ}$, then~$e$ has to be in~$\Lambda_{\scriptscriptstyle \circ}$. Therefore, $M$ is unit regular and $G(M) = W(\Gamma)$. In particular Property~(ECS3) holds. Assume $w = w_1ew_2$ lies in~$E(M)$ with $w_1,w_2$ in $W(\Gamma)$ and $e$ in~$\Lambda$. If $e = 1$ then $w_1w_2$ has to be equal to $1$ in~$W(\Gamma)$. Assume $e\neq 1$. Then $w_1ew_2w_1ew_2 = w_1ew_2$, and $ew_2w_1e = e$. By Lemma~\ref{leminterm}(ii), $w_2w_1$ belongs to $W_{\lambda_\star(e)}$ and~$w=w_1ew_1^{-1}$. Thus $E(M) = \{wew^{-1}\mid e\in\Lambda, w\in W(\Gamma)\}$ is a semi-lattice and Property~(ECS1) holds. Let $w_1,w_2,v_1,v_2$ be in~$W(\Gamma)$ and $e,f$ be in $\Lambda$ such that $w_1ew_2 = v_1fv_2$ in $M$. Then $e = w_1^{-1}v_1fv_2w_2^{-1}$ and $e\leq f$. By symmetry, $e = f$ and the elements~$w_1^{-1}v_1$ and~$v_2w_2^{-1}$ belong to~$W_{\lambda(e)}$. This implies that $\Lambda$ is a transversal of $E(M)$ for the action of $W(\Gamma)$ and a sub-semi-lattice of~$E(M)$. Therefore, we get Property~(ECS2). Furthermore, if $w_2 = v_1 = 1$ and $v_2 = w_1$, then $w_1$ lies in $W_{\lambda(e)}$. If $w_2 = v_1 = v_2 = 1$, then~$w_1$ lies in $W_{\lambda_\star(e)}$ by Lemma~\ref{leminterm}(ii). Property~(ECS5) follows. If $wew^{-1}\leq vfv^{-1}$, then $wew^{-1} vfv^{-1} = wew^{-1}$ and $ew^{-1}vfv^{-1}w= e$ Then $w^{-1}v$ lies in $W_{\lambda_\star(e)}\times W_{\lambda^\star(e)}$, which is included in $W_{\lambda_\star(e)}\times W_{\lambda^\star(f)}$. As a consequence, Property~(ECS4) holds. Finally, Property~(ECS6) holds by hypothesis.
\end{proof}
\subsubsection{Length function for generalised Renner-Coxeter systems}
As explained in the introduction, to answer Solomon's question, we need to define a length function on finite reductive monoids. Here we introduce this length function in the general context of generalised Renner-Coxeter systems. This extends results obtained in~\cite{God} and~\cite{God2}. As before,~$(R,\Lambda,S)$ is a generalised Renner-Coxeter system. The unit group of~$R$ is denoted by~$W$, and we set~$\Lambda_{\scriptscriptstyle \circ} = \Lambda\setminus\{1\}$.\label{seclong}
\begin{definition}\label{deflenfun2} (i) We set~$\ell(s) = 1$ for~$s$ in~$S$ and~$\ell(e) = 0$ for~$e$ in~$\Lambda$. Let~$x_1,\ldots, x_k$ be in~$S\cup\Lambda_{\scriptscriptstyle \circ}$ and consider the word~$\omega = x_1\cdots x_k$. Then the \emph{length} of the word~$\omega $ is the integer~$\ell(\omega )$ defined by~$\ell(\omega ) = \sum_{i = 1}^k\ell(x_i)$.\\
(ii) The \emph{length} of an element~$w$ which belongs to $R$ is the integer~$\ell(w)$ defined by $$\ell(w) = \min\left\{\ell(\omega) \mid \omega \textrm{ is a word representative of }w \textrm{ over }S\cup\Lambda_{\scriptscriptstyle \circ}\right\}.$$ If $\omega$ is a word representative of~$\omega$ such that~$\ell(w) = \ell(\omega)$, we say that~$\omega$ is a \emph{minimal word representative} of $w$. \end{definition}
\begin{Prop} \label{propproplenght}Let~$r$ belong to~$R$.\\ (i) The length function~$\ell$ on~$R$ extends the length function~$\ell$ defined on~$W$.\\(ii)~$\ell(r) = 0$ iff~$r$ lies in~$\Lambda$.\\(iii) If~$s$ lies in~$S$ then~$|\ell(sr)-\ell(r)|\leq 1$.\\(iv) If~$r'$ belongs to~$R$, then~$\ell(rr')\leq \ell(r)+\ell(r')$.\end{Prop}
\begin{proof}This is direct consequences of the definition of the length function. \end{proof}
\begin{Prop}\label{propproplenght2}
Let~$r$ belong to~$R$. If $(w_1,e,w_2)$ is the normal decomposition of $r$, then $$\ell(r) = \ell(w_1)+\ell(w_2).$$
\end{Prop}
\begin{proof} Using the relations of the monoid presentation of~$R$ stated in Proposition~\ref{proppres2}, every representative word of $r$ can be transformed into~$\underline{w}_1e\underline{w}_2$ without increasing the length. Therefore~$\ell(r) = \ell(\underline{w}_1)+\ell(e)+\ell(\underline{w}_2) = \ell(w_1)+\ell(w_2)$. \end{proof}
From the proof of the above proposition, we also deduce that
\begin{Cor}\label{cor:pasaugleng}
Let~$r$ belong to~$R$ and $\omega_1,\omega_2$ be two minimal word representatives of $r$. Using the relations of the monoid presentation of~$R$ stated in Proposition~\ref{proppres2}, one can transform~$\omega_1$ into~$\omega_2$ without increasing the length.
\end{Cor}
\subsubsection{Matsumoto's Lemma for generalised Renner-Coxeter systems}
In this section we state and prove some technical results that play the role of \emph{Matsumoto's Lemma} in the context of generalised Renner-Coxeter systems. we need these results when proving Theorem~\ref{Th:genThe}.
As before,~$(R,\Lambda,S)$ is a generalised Renner-Coxeter system. Let us first recall Matsumoto's Lemma.
\begin{Lem}\cite[Sec.~7.2]{Hum}\label{lemcoxbienconnu}
Consider a Coxeter system~$(W,S)$. Let $w$ belong to~$W$ and $s,t$ belong to $S$. If $\ell(swt) = \ell(w)$ and $\ell(sw) = \ell(wt)$, then $sw = wt$.
\end{Lem}
\begin{Lem} \label{lem:ewwt1}Let $r$ belong to $R$ and $s,t$ belong to $S$. Let $(w_1,e,w_2)$ be the normal decomposition of $r$. Then\\(i) $\ell(sr) = \ell(r)\pm 1$ if and only if the normal decomposition of $sr$ is $(sw_1,e,w_2)$. In this case, $\ell(sr) - \ell(r) = \ell(sw_1) - \ell(w_1)$.\\(ii) $\ell(sr) = \ell(r)$ if and only if $sr = r$ if and only if $sw_1 = w_1u$ for some $u$ in~$\lambda_\star(e)$. In this case, $\ell(sw_1) = \ell(w_1) +1$. \\(iii) $\ell(rt) = \ell(r)\pm 1$ if and only if the normal decomposition of $rt$ is either $(w_1,e,w_2t)$ or $(w_1u,e,w_2)$ for some $u$ in~$\lambda^\star(e)$. Furthermore, in the former case $\ell(rt) - \ell(r) = \ell(w_2t) -\ell(w_2)$, and in the latter case $w_2t = uw_2$ with $\ell(w_2t) = \ell(w_2)+1$.\\(iv) $\ell(rt) = \ell(r)$ if and only if $r = rt$ if and only if $w_2t = uw_2$ for some $u$ in~$\lambda_\star(e)$.\\(v) If~$\ell(srt) = \ell(r)$ and $\ell(sr) = \ell(rt) \neq \ell(r)$, then there exists $u$ in~$\lambda^\star(e)$ such that $sw_1 = w_1u$ and $uw_2 = w_2t$. As a consequence, $sr = rt$. \end{Lem}
\begin{proof} Recall that $|\ell (sr) -\ell (r)|\leq 1$ and $|\ell (rt) -\ell (r)|\leq 1$. The normal decomposition of $sr$ is $(sw_1,e,w_2)$ if and only if $sw_1$ belongs to $\DrTe{e}$. Since $w_1$ belongs to $\DrTe{e}$, this is clearly the case if $\ell(sw_1) = \ell(w_1)-1$. Assume~$\ell(sw_1) = \ell(w_1)+1$ and $sw_1$ does not belong to $\DrTe{e}$. Then we can write $sw_1 = w'_1u$ for some $u$ in $\lambda_\star(e)$ such that $\ell(sw_1) = \ell(w'_1)+1$. In particular, $\ell(sw_1u) = \ell(w'_1) = \ell(w_1)$. On the other hand, $\ell(w_1u) = \ell(w_1)+1 = \ell(sw_1)$ because $w_1$ belongs to $\DrTe{e}$, and $u$ lies in~$\lambda_\star(e)$. By Lemma~\ref{lemcoxbienconnu}, we get $sw_1 = w_1u$ and $sr = sw_1ew_2 = w_1uew_2 = w_1ew_2 = r$. This proves $(i)$ and $(ii)$ since the other implications are obvious. The normal decomposition of $rt$ is $(w_1,e,w_2t)$ if and only if $w_2t$ belongs to $\Gae{e}$. Since $w_2$ belongs to~$\Gae{e}$, this is clearly the case if $\ell(w_2t) = \ell(w_2)-1$. Assume~$\ell(w_2t) = \ell(w_2)+1$ and $w_2t$ does not belong to $\Gae{e}$. Then we can write $w_2t = uw'_2$ for some $u$ in $\lambda(e)$ such that $\ell(w_2t) = \ell(w'_2)+1$. As before we can conclude that $w_2t = uw_2$. If $u$ lies in $\lambda_\star(e)$ then $rt = r$. Otherwise, $u$ belongs to $\lambda^\star(e)$ and $w_1u$ belongs to $\DrTe{e}$. This is true since $u$ belongs to~$\lambda^\star(e)$ and therefore commutes with each element of~$\lambda_\star(e)$. Then the normal decomposition of~$rt$ is $(w_1u,e,w_2)$. This proves $(iii)$ and $(iv)$. Now assume $\ell(srt) = \ell(r)$ and $\ell(sr) = \ell(rt) \neq \ell(r)$. We claim that $\ell(w_2t) = \ell(w_2)+1$ and there exists $u$ in $\lambda(e)$ such that $uw_2 = w_2t$. If it was not the case, by above arguments, the normal decomposition of $srt$ would be $(sw_1,e,w_2t)$ and $\ell(srt) = \ell(r) \pm 2$. Since we assume~$\ell(rt) \neq \ell(r)$, the element~$u$ has to belong to~$\lambda^\star(e)$. Finally, using that $\ell(sr) = \ell(rt)\neq \ell(r) = \ell(srt)$ we deduce that $\ell(sw_1) = \ell(w_1u)$ and $\ell(w_1) = \ell(sw_1u)$, which in turn implies~$sw_1 = w_1u$ by Lemma~\ref{lemcoxbienconnu}.
\end{proof}
\begin{Lem}\label{lem:ewwt2}Let $r$ belong to $R$,~$s$ belong to~$S$ and $f$ belong to $\Lambda$. Let $(w_1,e,w_2)$ be the normal decomposition of $r$.\\(i) If $\ell(rf) = \ell (r)$ then $w_2$ belongs to~$W(f)$.\\(ii) If $\ell(fr) = \ell (r)$ then $w_1 = w'_1w''_1$ where $w'_1$ lies in ~$W(f)$ and $w''_1$ lies in $W^\star(e)$.\\(iii) If $\ell(sr) = \ell(r)-1$, then $\ell(srf)\leq\ell(rf)$. If $\ell(sr) = \ell(r)+1$, then $\ell(srf) \geq \ell(rf)$. \\(iv) If $\ell(rs) = \ell(r)-1$, then $\ell(frs)\leq\ell(fr)$. If $\ell(rs) = \ell(r)+1$, then $\ell(frs) \geq \ell(fr)$.
\end{Lem}
\begin{proof} By definition of the normal decomposition, $w_2$ belongs to~$\Gae{e}$. Write $w_2 = w'_2w''_2$ with $w'_2,w''_2$ in the unit group~$W$ of~$R$ such that $\ell(w_2) = \ell(w'_2)+\ell(w''_2)$, $w''_2$ belongs to~$W(f)$ and $w'_2$ belongs to~$\Dre{f}$. Then $w'_2$ lies in $\textrm{Red}(e,f)$. By Relation~(REN3), we have~$rf = w_1(e\wedge_{w'_2}f)w''_2$. It follows that $\ell(w'_2) = 0$, and $w_2 = w''_2$. This proves~(i). The prove of (ii) is similar except that we need first to decompose $w_1$ in $w'_1w''_1$ where $w''_1$ lies in $W^\star(e)$ and $w'_1$ lies in~$\Dre{e}$.\\
(iii) Assume $\ell(sr) = \ell(r)-1$. Write $w_1 = sv_1$ with $\ell(w_1) = \ell(v_1)+1$, and write $w_2 = w'_2w''_2v'''_2$ with $w'_2,w''_2,w'''_2$ in $W$ such that $\ell(w_2) = \ell(w'_2)+\ell(w''_2)+\ell(w'''_2)$, where $w''_2$ belongs to~$W^\star(f)$, $w'''_2$ belongs to~$W_\star(f)$ and $w'_2$ belongs to~$\textrm{Red}(e,f)$. Then $(v_1,e,w_2)$ is the normal decomposition of~$sr$. One has $srf = v_1ew'_2fw''_2 = v_1e'w''_2$ where $e' = e\land_{w'_2}f$ belongs to~$\Lambda$. Write $w''_2 = v''_2v'_2v_2$ such that $\ell(w''_2) = \ell(v''_2)+\ell(v'_2)+\ell(v_2)$ with $v''_2\in W_\star(e')$, $v'_2\in W^\star(e')$ and $v_2\in \Gae{e'}$. We claim that $v''_2 = 1$. Indeed $w'_2$ belongs to~$W_\star(e')$ by Lemma~\ref{lemclef}(ii), and $w_2 = w'_2v''_2v'_2v_2w'''_2 = v'_2w'_2v''_2v_2w'''_2$ with $\ell(w_2) = \ell(v_2)+\ell(v'_2)+\ell(w'_2)+\ell(v''_2)+\ell(w'''_2)$. But $v'_2\in W^\star(e')\subseteq W^\star(e)$, since~$e'\leq e$ by Property~(ECS6), whereas $w_2$ belongs to~$\Gae{e}$ by definition of the normal decomposition. Hence, $v'_2 = 1$. Now, write~$v_1 = v'_1v''_1$ such that $\ell(v_1) = \ell(v'_1)+\ell(v''_1)$ with $v'_1\in \DrTe{e'}$ and $v''_1\in W_\star(e')$. Then $srf = v'_1e'v_2$ and $(v'_1,e',v_2)$ is the normal decomposition of $swf$. Since $\ell(ssr) = \ell(sr)+1$, we have $\ell(sv'_1v''_1) = \ell(sv_1) = \ell(v_1)+1$ by Lemma~\ref{lem:ewwt1}(i). This implies $\ell(sv'_1) = \ell(v'_1)+1$ and we cannot have $\ell(ssrf) = \ell(srf)-1$, still by Lemma~\ref{lem:ewwt1}(i). Assume $\ell(sr) = \ell(r)+1$. let $(v_1,e,w_2)$ be the normal decomposition of~$r$, and $(v'_1,e',v_2)$ be the normal decomposition of~$rf$. It follows from above arguments that $v'_1$ left divides $v_1$. We conclude using Lemma~\ref{lem:ewwt1}: $\ell(sr) = \ell(r)+1\Rightarrow \ell(sv_1) = \ell(v_1)+1 \Rightarrow \ell(sv'_1) = \ell(v'_1)+1 \Rightarrow \ell(srf) \geq \ell(rf)$. The proof of (iv) is similar.
\end{proof}
\subsection{Free module over $R$}
For all this section, we assume~$(R,\Lambda,S)$ is a generalised Renner-Coxeter system. We let $W$ denote the unit group of $R$, and set $\Lambda_{\scriptscriptstyle \circ} = \Lambda\setminus\{1\}$. We fix an arbitrary unitary associative ring~$A$. We let~$V$ denote the free $A$-module with basis elements $T_r$ for~$r\in R$.
\begin{The} Fix $q$ in $A$. There exists a unique structure of unitary associative $A$-algebra on~$V$ such that~$T_1$ is the unity element and the following conditions hold for every $x$ in $S\cup\Lambda_{\scriptscriptstyle \circ}$ and every $r$ in $R$:\label{Th:genThe}
\begin{center}\begin{tabular}{ll}$T_xT_r = T_{xr}$,&if $x\in S$ and $\ell(xr) = \ell(r) +1 $;\\
$T_xT_r = qT_{r}$,&if $x\in S$ and $\ell(xr) = \ell(r)$;\\$T_xT_r = (q-1)T_r+qT_{xr}$,&if $x\in S$ and $\ell(xr) = \ell(r) - 1;$\\ $T_xT_r = q^{\ell(r)-\ell(xr)}T_{xr}$,&if $x\in \Lambda_{\scriptscriptstyle \circ}$.
\\\end{tabular}\end{center}
\end{The}
We follow the method explained in~\cite[Sec.~7.1]{Hum1} for the Hecke algebra of Coxeter groups. Let~$\mathcal{E} = End_A(V)$ the $A$-algebra of endomorphisms of the $A$-module~$V$. For~$s$ in~$S$ and $r$ in~$R$, we define $\rho_s$ in~$\mathcal{E}$ by \begin{center}\begin{tabular}{ll}$\rho_s(T_r) = T_{sr}$, &if $\ell(sr) = \ell(r)+1$;\\$\rho_s(T_{r}) = qT_{r}$, &if $\ell(sr) = \ell(r)$;\\$\rho_s(T_{r}) = (q-1)T_r+qT_{sr}$, &if $\ell(sr) = \ell(r) -1$.\end{tabular}\end{center} For~$e$ in~$\Lambda$ and $r$ in~$R$, we define $\rho_e$ by $$\rho_e(T_r) = q^{\ell(r)-\ell(er)}T_{er}$$ Similarly, for~$s$ in~$S$ and $r$ in~$R$, we define~$\overline{\rho}_s$ in~$\mathcal{E}$ by \begin{center}\begin{tabular}{ll}$\overline{\rho}_s(T_r) = T_{rs}$,&if $\ell(sr) = \ell(r)+1$;\\$\overline{\rho}_s(T_{r}) = qT_{r}$,&if $\ell(r) = \ell(rs)$;\\$\overline{\rho}_s(T_{r}) = (q-1)T_r+qT_{rs}$,&if $\ell(sr) = \ell(r)-1$.\end{tabular}\end{center} For~$e$ in~$\Lambda$ and $r$ in~$R$, we define $\overline{\rho}_e$ by $$\overline{\rho}_e(T_r) = q^{\ell(r)-\ell(re)}T_{re}.$$
The key tool in the proof of Theorem~\ref{Th:genThe} is the following result.
\begin{Lem}\label{lem:commut} For every $x,y$ in $S\cup \Lambda$, $$\rho_x\overline{\rho}_y = \overline{\rho}_y\rho_x.$$
\end{Lem}
\begin{proof} Let $r$ belong to~$R$ and $x,y$ belong to $S\cup\Lambda$. We prove that~$\rho_x(\overline{\rho}_y (T_r))= \overline{\rho}_y(\rho_x(T_r))$. Clearly we can assume $x\neq 1$ and $y\neq 1$. By Proposition~\ref{propproplenght}, $\ell(xry)\leq \ell(x)+\ell(r)+\ell(y)\leq(r)+2$. We provide case by case as in~\cite{Hum}.\\
\underline{Case $1$: $\ell(xry) = \ell(r)+\ell(x)+\ell(y)$}. We must have $\ell(xr) = \ell(r)+\ell(x)$, $\ell(ry) = \ell(r)+\ell(y)$ and $\ell(xry) = \ell(ry)+\ell(x) = \ell(xr)+\ell(y)$. Therefore~$\rho_x(\overline{\rho}_y (T_r)) = \rho_x(T_{ry})= T_{xry} = \overline{\rho}_y(T_{xr}) = \overline{\rho}_y(\rho_x(T_r))$.\\
\noindent\underline{Case~$2$: $\ell(xry) = \ell(r)+1$}. We must have~$\ell(xr)\geq\ell(r)$,~$\ell(ry)\geq\ell(r)$, and~$x$ or~$y$, possibly both, belongs to $S$. If~$x$ or $y$ belongs to~$\Lambda_{\scriptscriptstyle \circ}$, we are in Case~$1$. So we assume $x$ and $y$ belong to~$S$.\\
{\it Subcase~1:} $\ell(xr) = \ell(r)$, that is $xr = r$. Then $\ell(ry) = \ell(xry) = \ell(r)+1$ and $\ell(xry) = \ell(xr)+1$. Therefore~$\rho_x(\overline{\rho}_y (T_r)) = \rho_x(T_{ry})= qT_{xry} = \overline{\rho}_y(qT_{xr}) = \overline{\rho}_y(\rho_x(T_r))$. The case~$\ell(ry) = \ell(r)$ is similar. \\
{\it Subcase~2:}~$\ell(ry) = \ell(xr)= \ell(r)+1$. Then~$\ell(ry) = \ell(xr)= \ell(xry)$. We deduce that~$\rho_x(\overline{\rho}_y (T_r)) = \rho_x(T_{ry})= qT_{xry} = \overline{\rho}_y(T_{xr}) = \overline{\rho}_y(\rho_x(T_r))$.\\
\noindent\underline{Case~$3$: $\ell(xry) = \ell(r)$}. If~$x$ and $y$ belong to~$\Lambda_{\scriptscriptstyle \circ}$, we are in Case~$1$. So we assume this is not the case.\\{\it Subcase~1:}~$x$ and $y$ belong to~$S$. Consider first the case $\ell(xr) = \ell(r)$. Then $xr = r$ and $\ell(xry) = \ell(ry) = \ell(r)$. Therefore,~$\rho_x(\overline{\rho}_y (T_r)) = \overline{\rho}_y(\rho_x(T_r)) = q^2T_r$. Assume now~$\ell(xr) \neq \ell(r)$. This implies $\ell(ry)\neq \ell(y)$ by symmetry. If $\ell(xr) = \ell(ry)$, by Lemma~\ref{lem:ewwt1}$(v)$ we have $xr = ry$. Hence, if $\ell(xr) = \ell(ry) = \ell(r) +1$, we have~$\overline{\rho}_y(\rho_x(T_r)) = \overline{\rho}_y(T_{xr}) = (q-1)T_{xr}+qT_{xry}$ and~$\rho_x(\overline{\rho}_y (T_r)) = \rho_x(T_{ry}) = (q-1)T_{ry}+qT_{xry}$. If $\ell(xr) = \ell(ry) = \ell(r) - 1$, we have~$\overline{\rho}_y(\rho_x(T_r)) = \overline{\rho}_y((q-1)T_{r}+qT_{xr}) = (q-1)T_{yr}+qT_{xry}$ and~$\rho_x(\overline{\rho}_y (T_r)) = \rho_x((q-1)T_{r}+qT_{ry}) = (q-1)T_{xr}+qT_{xry}$. Consider now the case $\ell(xr) = \ell(r)+1$ and $\ell(ry) = \ell(r)-1$. Then~$\overline{\rho}_y(\rho_x(T_r)) = \overline{\rho}_y(T_{xr}) = (q-1)T_{xr}+qT_{xry} = \rho_x((q-1)T_{r}+q T_{ry}) = \rho_x(\overline{\rho}_y (T_r))$. The case where $\ell(xr) = \ell(r)-1$ and $\ell(ry) = \ell(r)+1$ is similar. \\{\it Subcase~2:}~$x$ belongs to~$S$ and~$y$ belong to~$\Lambda_{\scriptscriptstyle \circ}$. We must have $\ell(xr)\geq \ell(r)$. Assume first~$\ell(xr) = \ell(r)$. We have $xr = r$ and $\ell(xry) = \ell(ry) = \ell(r)$. We get, $\overline{\rho}_y(\rho_x(T_r)) = \overline{\rho}_y(qT_r)) = q^{1+\ell(r)-\ell(ry)}T_{ry} =q^{\ell(r)-\ell(ry)}\rho_x(T_{ry}) = \rho_x(\overline{\rho}_y (T_r))$. Assume now~$\ell(xr) = \ell(r)+1$, then~$\overline{\rho}_y(\rho_x(T_r)) = \overline{\rho}_y(T_{xr}) = q^{\ell(xr)-\ell(xry)}T_{xry} = q T_{xry}$. If $\ell(ry) = \ell(r)$ then $\ell(xry) = \ell(ry)$ and~$\rho_x(\overline{\rho}_y (T_r)) = \rho_x(T_{ry}) = qT_{xry}$. If $\ell(ry) < \ell(r)$, then $\ell(xry) = \ell(r) = \ell(ry) +1$ and ~$\rho_x(\overline{\rho}_y (T_r)) = q\rho_x(T_{ry}) = qT_{xry}$. The case $x\in \Lambda_{\scriptscriptstyle \circ}$ and $y\in S$ is similar.\\
\noindent\underline{Case~$4$: $\ell(xry) < \ell(r)$}.\\
{\it Subcase~1:}~$x,y$ belong to~$\Lambda_{\scriptscriptstyle \circ}$. Clearly, $\rho_x(\overline{\rho}_y (T_r)) = \overline{\rho}_y(\rho_x(T_r)) = q^{\ell(r)-\ell(xry)}T_{xry}$.\\
{\it Subcase~2:}~$x$ belongs to~$S$, $y$ belongs to~$\Lambda_{\scriptscriptstyle \circ}$ and $\ell(xr) = \ell(r)$. Then $xr = r$ and $xry = ry$. This case is similar to the first case in Case~$3$ Subcase~$2$.\\
{\it Subcase~3:}~$x$ belongs to~$S$, $y$ belongs to~$\Lambda_{\scriptscriptstyle \circ}$ and~$\ell(xr) = \ell(r) - 1$. Applying Lemma~\ref{lem:ewwt2}, we get~$\ell(xry) \leq \ell(ry)$. We have $\overline{\rho}_y(\rho_x(T_r)) =$ $\overline{\rho}_y((q-1)T_r+qT_{xr}) =$ $(q-1)q^{\ell(r)-\ell(ry)}T_{ry}+q^{1+\ell(xr)-\ell(xry)}T_{xry}$ and $(\overline{\rho}_y (T_r)) = q^{\ell(r)-\ell(ry)} \rho_x(T_{ry})$.\\ Assume first $\ell(xry) = \ell(ry)-1$. Then $\ell(xr)-\ell(xry) = \ell(r)-\ell(ry)$ and $(\overline{\rho}_y (T_r)) = (q-1)q^{\ell(r)-\ell(ry)} T_{ry} + q^{1+\ell(r)-\ell(ry)} T_{xry}$.\\ Assume secondly that $\ell(xry) = \ell(ry)$, that is $xry = ry$. In this case, $(\overline{\rho}_y (T_r)) = q^{1+\ell(r)-\ell(ry)}T_{xry}$. But $1+\ell(xr)-\ell(xry) = \ell(r)-\ell(ry)$, therefore $\overline{\rho}_y(\rho_x(T_r)) = q^{1+\ell(r)-\ell(ry)} T_{ry}$.\\
{\it Subcase~4:}~$x$ belongs to~$S$, $y$ belongs to~$\Lambda_{\scriptscriptstyle \circ}$ and $\ell(xr) = \ell(r)+1$. By Lemma~\ref{lem:ewwt2}, we get $\ell(xry) \geq \ell(ry)$. We have $\overline{\rho}_y(\rho_x(T_r)) = \overline{\rho}_y(T_{xr}) = q^{\ell(xr)-\ell(xry)}T_{xry}$. If $\ell(xry) = \ell(ry) +1$, then $\rho_x(\overline{\rho}_y(T_r)) = \rho_x(q^{\ell(r)-\ell(ry)}T_{ry}) = q^{\ell(r)-\ell(ry)}T_{xry}$. If $\ell(xry) = \ell(ry)$, then $\rho_x(\overline{\rho}_y(T_r)) = \rho_x(q^{\ell(r)-\ell(ry)}T_{ry}) = q^{\ell(r)-\ell(ry)+1}T_{xry}$. Thus, in both case, $\overline{\rho}_y(\rho_x(T_r)) = \rho_x(\overline{\rho}_y(T_r))$.\\
{\it Subcase~5:}~$x,y$ belong to~${S}$. If $\ell(xry) = \ell(r)-2$, then $\ell(xr) = \ell(ry) = \ell(r)-1$ and a calculation similar to~\cite[page~148 case (b)]{Hum} lied to~$\overline{\rho}_y(\rho_x(T_r)) =\rho_x( \overline{\rho}_y(T_r)) = q^2T_{xry} +q(q-1)T_{xr} + q(q-1)T_{ry} + (q-1)^2T_{r}$. So, we consider the case~$\ell(xry) = \ell(r)-1$. If $\ell(xr) = \ell(r)$, then $xr = r$ and $xry = ry$. Therefore $\ell(ry)<\ell(r)$ and~$\overline{\rho}_y(\rho_x(T_r)) =\rho_x( \overline{\rho}_y(T_r)) = q(q-1)T_{xr}+q^2T_{xry}$. Now, consider the case $\ell(xr) = \ell(r)-1$. If $\ell(ry) = \ell(r)$, then~$\overline{\rho}_y(\rho_x(T_r)) =\rho_x( \overline{\rho}_y(T_r)) = q(q-1)T_{r}+q^2T_{xr}$; finally, if $\ell(ry) = \ell(r)-1$ then~$\overline{\rho}_y(\rho_x(T_r)) =\rho_x( \overline{\rho}_y(T_r)) = (q-1)^2T_{r}+q(q-1)T_{rt} + q^2T_{xry}$.
\end{proof}
Once we have Lemma~\ref{lem:commut}, we can almost repeat the argument of~\cite[Sec~7.3]{Hum} to prove Theorem~\ref{Th:genThe}.
\begin{Lem} Let $\mathcal{L}$ be the sub-algebra of~$\mathcal{E}$ generated the $\rho_x$ for $x$ in $R$. The map~$\varphi$ from~$\mathcal{L}$ to $V$ which sends $\rho$ to $\rho(T_1)$ is an isomorphism of $A$-modules. \label{lem:iso}
\end{Lem}
\begin{proof} This is clear that~$\varphi$ is a morphism of $A$-modules. Let $r$ belong to~$R$, and let~$x_1\cdots x_k$ be a minimal word representative. Then by definition of the maps~$\rho_{x_i}$, we have $T_r = \varphi (\rho_{x_1}\cdots \rho_{x_k})$. Therefore, $\varphi$ is surjective. Assume $\varphi(\rho) = 0$ for some $\rho$ in~$\mathcal{L}$. Consider~$r$ and $x_1\cdots x_k$ as before, such that $k$ is minimal. We prove by induction on $k$ that $\rho(T_r) = 0$. For~$k = 0$, that is $r = 1$, this is true by hypothesis. The word~$x_1\cdots x_{k-1}$ is a minimal word representative of some element~$r'$. By induction hypothesis, we have $\rho(T_{r'}) = 0$. It follows $\rho(T_r) = \rho(T_{r'x_k}) = \rho(\overline{\rho}_{x_k}(T_{r'})) = \overline{\rho}_{x_m}(\rho(T_{r'})) = \overline{\rho}_{x_m}(0) = 0$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{Th:genThe}] Consider the notation of Lemma~\ref{lem:iso}. Assume $r$ belongs to $R$ and $x_1\cdots x_k$ is a minimal word representative of~$r$. Iterating the first defining relation in Theorem~\ref{Th:genThe}, we get $T_r = T_{x_1}\cdots T_{x_k}$. The unicity follows. Since $\varphi$ is an isomorphism, the endomorphism~$\rho_{r} = \rho_{x_1}\cdots\rho_{x_k}$ does not depend on the minimal word representing~$x_1\cdots x_k$, and the set~$\{\rho_r \mid r\in R\}$ is a free $A$-basis for~$\mathcal{L}$ with $\varphi(\rho_r) = \rho_r(T_1) = T_r$. Moreover, we can transfer the $A$-algebra structure of~$\mathcal{L}$ to $V$ using the isomorphism~$\varphi$. It remains to verify that the structure constants of the obtained $A$-algebra are the one stated in the theorem. Let $x$ belongs to~$S\cup \Lambda_{\scriptscriptstyle \circ}$ and $r$ in~$R$. If $\ell(xr) = \ell(x)+\ell(r)$ and $\omega$ is a minimal word representative of $r$, then $x\omega$ is clearly a minimal word representative of~$xr$. Therefore~$\rho_x\rho_r(T_1) = \rho_x(T_r) = T_{xr} = \rho_{xr}(T_1)$. Therefore, $\rho_x\rho_r = \rho_{xr}$, and~$T_xT_r = T_{xr}$. Assume $x$ lies in $\Lambda_{\scriptscriptstyle \circ}$ and $\ell(xr) < \ell(r)$. Then~$\rho_x\rho_r(T_1) = \rho_x(T_r) = q^{\ell(r)-\ell(xr)}T_{xr} = q^{\ell(r)-\ell(xr)} \rho_{xr}(T_1)$. We get $\rho_x\rho_r = q^{\ell(r)-\ell(xr)}\rho_{xr}$ and~$T_xT_r = q^{\ell(r)-\ell(xr)}T_{xr}$. Assume $x$ lies in $S$. If $\ell(xr) = \ell(r)$, then~$\rho_x\rho_r(T_1) = \rho_x(T_r) = qT_{xr} = q\rho_{xr}(T_1)$ and $T_xT_r = qT_{rx}$. Finally, consider the case $\ell(xr) = \ell(r) -1$. One has ~$\rho_x\rho_r(T_1) = \rho_x(T_r) = (q-1)T_r + qT_{xr} = (q-1)\rho_r(T_1) + q\rho_{xr}(T_1) = ((q-1)\rho_r + q\rho_{xr})(T_1)$. Therefore,~$\rho_x\rho_r = (q-1)\rho_r + q\rho_{xr}$ and~$T_xT_r = (q-1)T_r + qT_{xr}$.
\end{proof}
\begin{definition} Let~$q$ be an indeterminate and set~$A = \mathbb{Z}[q]$. The \emph{generic Hecke algebra}~$\mathcal{H}(R)$ of the generalised Renner monoid~$R$ is the $A$-algebra described in Theorem~\ref{Th:genThe}. \label{def:genHechalg}
\end{definition}
\begin{Cor} The generic Hecke algebra~$\mathcal{H}(R)$ of~$R$ admits the following $\mathbb{Z}[q]$-algebra presentation: the generators are $T_x$ for $x$ in $S\cup\Lambda_{\scriptscriptstyle \circ}$; the defining relations are\begin{center}
\begin{tabular}{lll}
(HEC1)&$T^2_s = (q-1)T_1+qT_{s}$,&$s\in S$;\\
(HEC2)&$|T_s,T_t\rangle^m = |T_t,T_s\rangle^m$,&$(\{s,t\},m)\in \mathcal{E}(\Gamma)$;\\
(HEC3)&$T_sT_e = T_eT_s$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda^\star(e)$;\\
(HEC4)&$T_sT_e = T_eT_s = qT_e$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda_\star(e)$;\\
(HEC5)&$T_eT_wT_f = q^{\ell(w)}T_{e\wedge_wf}$,& $e,f\in \Lambda_{\scriptscriptstyle \circ}$, $w\in \textrm{Red}(e,f)$.
\end{tabular} \end{center}\label{cor:preshHR}
\end{Cor}
In the special case of the rook monoid (see Example~\ref{exerook} below), we recover the presentation obtained in~\cite{God}.
\begin{proof} Consider the presentation of~$\mathcal{H}(R)$ given in Theorem~\ref{Th:genThe}. Then Relations (HEC1)---(HEC5) clearly hold in~$\mathcal{H}(R)$. For instance~$|T_s,T_t\rangle^m = T_{|s,t\rangle^m} = T_{|t,s\rangle^m} = |T_t,T_s\rangle^m$. Conversely, consider the algebra~$\mathcal{H}$ defined by the presentation given in the corollary. We claim that for two minimal word representatives~$\omega_1 = x_1\cdots x_k$ and $\omega_2 = y_1\cdots y_k$ on~$S\cup\Lambda_{\scriptscriptstyle \circ}$ that represent the same element~$r$ in~$R$, we have $T_{x_1}\cdots T_{x_k} = T_{y_1}\cdots T_{y_k}$. Indeed, it follows from Corollary~\ref{cor:pasaugleng} that we can transform $T_{x_1}\cdots T_{x_k}$ into $T_{y_1}\cdots T_{y_k}$ by using $(HEC2)$, $(HEC3)$ and $(HEC5)$. So we set $T_r = T_{x_1}\cdots T_{x_k}$ in~$\mathcal{H}$.
If $(w_1,e,w_2)$ is the normal decomposition of $r$ we have $T_r = T_{w_1}T_eT_{w_2}$. Now, we deduce that the defining relations of~$\mathcal{H}(R)$ given in Theorem~\ref{Th:genThe} hold in~$\mathcal{H}$ using lemma~\ref{lem:ewwt1} and~\ref{lem:ewwt2}. If $\ell(xr) = \ell(x)+\ell(r)$ and $x_1\cdots x_k$ is a minimal word representative of $r$, then $xx_1\cdots x_k$ is a minimal word representative of $xr$ and $T_{xr} = T_xT_{x_1}\cdots T_{x_k} = T_xT_r$. If $x$ belong to $S$ and $\ell(xr) = \ell(r) - 1$, then $T_xT_r =T_xT_{w_1}T_eT_{w_2} = ((q-1)T_{w_1} + qT_{xw_1}) T_eT_{w_2} = (q-1)T_r +qT_{xw_1}$. Here we use that Relations $(HEC1)$ and $(HEC2)$ implies $T_{w} = (q-1)T_{w} + qT_{xw}$ when $w$ belongs to $W$ such that $\ell(xw) = \ell(w) - 1$ ({\it cf.} \cite[Sec.~7]{Hum}). If $x$ belongs to $S$ and $\ell(xr) = \ell(r)$, then by Lemma~\ref{lem:ewwt1}, there exists~$u$ in~$\lambda_\star(e)$ such that $xw_1 = w_1u$, and~$\ell(xw_1) = \ell(w_1)+1$. It follows that~$T_xT_r = T_xT_{w_1}T_eT_{w_2} = T_{xw_1}T_eT_{w_2} = T_{w_1}T_uT_eT_{w_2} = qT_{w_1}T_eT_{w_2} = qT_r$. Finally, assume $x$ belongs to~$\Lambda_{\scriptscriptstyle \circ}$ and $\ell(xr)<\ell(r)$. Write $w_1 = w'''_1w''_1w'_1$ such that $\ell(w_1) = \ell(w'''_1)+\ell(w''_1)+\ell(w'_1)$ with~$w'''_1$ in~$W_\star(x)$,~$w''_1$ in~$W^\star(x)$ and~$w'_1$ in $\textrm{Red}(x,e)$. We have~$T_xT_r = T_xT_{w_1}T_eT_{w_2} = T_xT_{w'''_1}T_{w''_1}T_{w'_1}T_eT_{w_2} = q^{\ell(w'''_1)}T_{x} T_{w''_1}T_{w'_1}T_eT_{w_2} = q^{\ell(w'''_1)} T_{w''_1}T_{x}T_{w'_1}T_eT_{w_2}$. We get~$T_xT_r = q^{\ell(w'''_1)+\ell(w'_1)} T_{w''_1}T_{x\wedge_{w'_1}e}T_{w_2}$. We can decompose~$w''_1$ and~$w_2$ such that $w''_1 = v'_1v''_1$ and $w_2 = v''_2v'_2$ where $v''_1,v''_2$ belong to~$W_\star(x\wedge_{w'_1}e)$, $v'_1$ belongs to $\DrTe{x\wedge_{w'_1}e}$ and $v'_2$ belongs to $\GaTe{x\wedge_{w'_1}e}$. We have~$\ell(xr) = \ell(v'_1)+\ell(v'_2)$ and $v'_1(x\wedge_{w'_1}e)v'_2$ is a minimal word representative of $xr$. Hence, $T_xT_r = q^{\ell(w'''_1)+\ell(w'_1)+\ell(v''_1)+\ell(v''_2)} T_{v'_1}T_{x\wedge_{w'_1}e}T_{v'_2} = q^{\ell(x)-\ell(xr)} T_{xr}$.
\end{proof}
\begin{Rem}\label{Rem:rem2}
(i) For $e,f$ in~$\Lambda_{\scriptscriptstyle \circ}$, we set $$\textrm{Red}_\star(e,f) = \textrm{Red}(e,f)\bigcap W_{\cap_{h> e}\lambda(h)} \bigcap W_{\cap_{h> f}\lambda(h)}.$$ It is not difficult to see that in Relations~$(HEC5)$ of the presentation stated in Corollary~\ref{cor:preshHR}, we can assume $w$ belongs to $\textrm{Red}_\star(e,f)$ ({\it cf.} the proof of ~\cite[Theorem~0.1]{God2}).\\(ii) In $\mathcal{H}(R)$ the following relations hold :
\begin{center}\begin{tabular}{ll}$T_rT_x = T_{xr}$,&if $x\in S$ and $\ell(rx) = \ell(r)+1$;\\
$T_rT_x = qT_{r}$,&if $x\in S$ and $\ell(rx) = \ell(r)$;\\$T_rT_x = (q-1)T_r+qT_{rx}$,&if $x\in S$ and $\ell(rx) = \ell(r) - 1;$\\ $T_rT_x = q^{\ell(r)-\ell(rx)}T_{rx}$,&if $x\in \Lambda_{\scriptscriptstyle \circ}$.
\\\end{tabular}\end{center}
This can be deduced directly from Theorem~\ref{Th:genThe}, but this is an immediate consequence of Corollary~\ref{cor:preshHR} since the defining relations~$(HEC1)-(HEC5)$ have a right-left symmetry.
\end{Rem}
\section{Iwahori-Hecke algebra of finite reductive monoids}
Here, we first recall basic results on Algebraic Monoid Theory, then we introduce the notion of an Iwahori-Hecke algebra in the general framework of Monoid Theory, we recall some basic properties and explain why this Iwahori-Hecke algebra is interesting. Finally, we turn to finite reductive monoids and prove that the Iwahori-Hecke algebra of such monoids is related to the generic Hecke algebra of the associated Renner monoid. As a consequence, we prove Theorems~\ref{THintro} and \ref{THintro2}.
\label{sectionprincipale}
\subsection{Regular monoids and reductive groups}
\label{sousect}
We introduce here the basic definitions and
notation on Algebraic Monoid Theory that we shall need in the sequel. We fix an algebraically closed field~$\mathbb{K}$. We let~$M_n$ denote the set of all~$n\times n$ matrices over~$\mathbb{K}$, and by~$GL_n$ the set of all invertible matrices in~$M_n$. We refer to~\cite{Put,Ren,Sol1} for the general theory and proofs involving linear algebraic monoids and Renner monoids; we refer to~\cite{Hum} for an introduction to Linear Algebraic Groups Theory. If $X$ is a subset of~$M_n$, we let $\overline{X}$ denote its closure for the Zariski topology. Recall that a semigroup~$M$ is said to have a \emph{zero element} if it contains an element~$0$ such that~$0\times x = x\times 0 = 0$ for every~$x$ in~$M$.
\begin{definition}[Algebraic monoid] An \emph{algebraic monoid} is a submonoid of~$M_n$, for some positive integer~$n$, that is closed for the Zariski topology. An algebraic monoid is \emph{irreducible} if it is irreducible as a variety.\end{definition}
It is very easy to construct algebraic monoids. Indeed, the Zariski closure~$M = \overline{G}$ of any submonoid~$G$ of~$M_n$ is an algebraic monoid. The main example occurs when for $G$ one considers an algebraic subgroup of~$GL_n$. It turns out that in this case, the group~$G$ is the unit group of~$M$. Conversely, if $M$ is an algebraic monoid, then its unit group~$G(M)$ is an algebraic group. The monoid~$M_n$ is the seminal example of an algebraic monoid, and its unit group~$GL_n$ is the seminal example of an algebraic group.
The next result, which is the starting point of the theory, was obtained independently by Putcha and Renner in 1982.
\begin{The} Let $M$ be an irreducible algebraic monoid with a zero element. Then $M$ is regular if and only if its unit group~$G(M)$ is reductive.
\end{The}
\begin{definition}[Reductive monoid] A \emph{reductive monoid} is an irreducible algebraic monoid whose unit group is a reductive group. \end{definition}
\begin{definition}[Renner monoid] Let $M$ be a reductive monoid. The normaliser of a maximal torus~$T$ of $G(M)$ is denoted by $N_{G(M)}(T)$. The \emph{Renner monoid}~$R(M)$ of~$M$ is the monoid~$\overline{N_{G(M)}(T)}/T$.
\end{definition}
It is clear that~$R(M)$ does not depend on the choice of the maximal torus of the algebraic group~$G(M)$.
\begin{Prop} Let $M$ be reductive monoid. Fix a maximal torus~$T$ of $G(M)$ and a Borel subgroup~$B$ of $G(M)$ that contains~$T$. The unit group of~$R(M)$ is the Weyl group~$W$ of $G(M)$. If $S$ is the standard generating set of~$W$ associated with the Borel $B$ and $\Lambda(B) = \{e\in E(\overline{T})\mid \forall b\in B,\ be = ebe \}$, then $(R(M),\Lambda(B),S)$ is a generalised Renner-Coxeter system such that~$R(M)$ is a generalised Renner monoid. Moreover, there is a canonical order preserving isomorphism of monoids between $E(R(M))$ and $E(\overline{T})$.\label{propRgenparEetW}
\end{Prop}
\begin{Exe} \label{exerook}Consider $M = M_n$. Choose the Borel subgroup~$\mathbb{B}$ of invertible upper triangular matrices and the maximal torus~$\mathbb{T}$ of invertible diagonal matrices. The Renner monoid is isomorphic to the monoid of matrices with at most one nonzero entry, that is equal to~$1$, in each row and each column. This monoid is called the rook monoid~$R_n$~\cite{Sol2}. Its unit group is the group of monomial matrices, which is isomorphic to the symmetric group~$S_n$. Denote by~$e_i$ the diagonal matrix~$\left(\begin{array}{cccccc}Id_i&0\\0&0\end{array}\right)$ of rank~$i$. Then the set~$\Lambda(\mathbb{B})$ is~$\{e_0,\ldots, e_n\}$. One has $e_i\leq e_{i+1}$ for every index~$i$. One has~$\lambda_\star(e_i) = \{s_j\mid j > i\}$ and $\lambda^\star(e_i) = \{s_j\mid j < i\}$.\begin{figure}[ht]
\begin{picture}(250,75)
\put(18,0){\includegraphics[scale = 0.6]{Bndiagethasse.eps}}
\put(148,15){$s_1$} \put(125,15){$s_2$}\put(40,15){$s_{n-2}$}\put(17,15){$s_{n-1}$}
\put(225,0){$e_0$} \put(225,20){$e_1$} \put(225,70){$e_n$}
\end{picture}
\caption{Coxeter graph~$\Gamma(S)$ and Hasse diagram~$\Lambda(\mathbb{B})$ for $M_n$.}\label{fig:hassecrlAn}
\end{figure}\\
Other examples can be found in~\cite{God2}. \end{Exe}
In the framework of algebraic monoids, Renner monoid plays the role of Weyl groups in Algebraic Group Theory. In particular we still have a \emph{Bruhat decomposition}: the monoid~$M$ is equal to the disjoint union~$\cup_{r\in R}BrB$. Moreover, the product of double classes~$BrB$ is related to the length function that we introduce in Section~\ref{seclong}:
\begin{Prop} Let $M$ be a reductive monoid. Fix a maximal torus~$T$ of $G(M)$ and a Borel subgroup~$B$ of $G(M)$ that contains~$T$. Consider the generalised Renner-Coxeter system~$(R(M),\Lambda,S)$ of~$R(M)$ defined in Proposition~\ref{propRgenparEetW}.\\(i) Let~$r$ lie in~$R(M)$ and~$s$ lie in~$S$, then $$B s B r B = \left\{\begin{array}{ll}BrB,&\textrm{if } \ell(sr) = \ell(r);\\BsrB,&\textrm{if } \ell(sr) = \ell(r)+1;\\BsrB\cup BrB,&\textrm{if } \ell(sr) = \ell(r)-1.\end{array}\right.$$ \label{Pr:lienlongBB}
(ii) Let~$r$ lie in~$R(M)$ and~$s$ lie in~$S$, then $$B r B s B = \left\{\begin{array}{ll}BrB,&\textrm{if } \ell(rs) = \ell(r);\\BrsB,&\textrm{if } \ell(rs) = \ell(r)+1;\\BrsB\cup BrB,&\textrm{if } \ell(rs) = \ell(r)-1.\end{array}\right.$$\\(iii) Let~$r$ lie in~$R(M)$ and~$e$ lie in~$\Lambda$, then $$B e B r B = BerB \textrm{ and } BrBeB = BreB$$ \end{Prop}
\begin{proof} (i) is proved in~\cite[Prop.~0.2]{God2} in the case of irreducible regular monoid~$M$ with a zero element. Same arguments can be applied for any reductive monoids; let us deduced~(ii): by the remark following~\cite[Prop.~8.6]{Ren} we know that~$$B r B s B\subseteq BrB\cup BrsB$$ and, clearly, $B r B s B$ is a union of double classes. Hence, $B r B s B$ has to be equal to $BrB$, $BrsB$ are $BrB\cup BrsB$. If~$\ell(rs) = \ell(r)$ then $rs = r$ and we are done. if $\ell(rs) = \ell(r)+1$ and $r = x_1\cdots x_k$ is a minimal word representative of $r$ then $BrBsB = Bx_1B\cdots Bx_{k-1}Bx_kBsB = Bx_1B\cdots Bx_{k-1}Bx_ksB = \cdots = BrsB$. Finally, if $\ell(rs) = \ell(rs)-1$, and $x_1\cdots x_{k-1}s$ is a minimal word representative of $r$, then $BrBsB = Bx_1B\cdots Bx_{k-1}BsBsB = Bx_1B\cdots Bx_{k-1}B (B\cup BsB) = BrsB \cup BrB$. Let us proof (iii). Since $e$ belongs to $\Lambda$, $Be\subseteq eB$ \cite{Ren}. Thus, $BrBeB\subseteq BreB$. The inclusion~$BreB\subseteq BrBeB$ is trivial. Let us prove that $BeBrB = BerB$.
If $r = s_{i_1}\cdots s_{i_{\ell(r)}}$ belongs to the Weyl group~$W$, the results follows from~(ii) since for $\ell(es_{i_1}\cdots s_{i_j})\geq \ell(es_{i_1}\cdots s_{i_{j-1}})$. Therefore, we may assume that $r = w_1fw_2$ where $f$ lies in~$\Lambda_{\scriptscriptstyle \circ}$ and $(w_1,f,w_2)$ is the normal decomposition of~$r$.
We can write $w_1 = v_1v_2v_3v_4$ with $v_1\in W_\star(e)$, $v_2\in W^\star(e)$, $v_3\in \textrm{Red}(e,f)$, $v_4$ in $W^\star(f)$ and $\ell(w_1) = \ell(v_1)+\ell(v_2)+\ell(v_3)+\ell(v_4)$. Then $$\displaylines{BeBrB = BeBw_1fw_2B = BeBv_1v_2v_3fv_4w_2B = BeBv_1Bv_2v_3BfBv_4w_2B=\hfill\cr\hfill Bev_2v_3BfBv_4w_2B = Bv_2ev_3fBv_4w_2B = Bv_2(e\wedge_{v_3}f)Bv_4w_2B.}$$ Write $v_4w_2 = v_5v_6v_7$ such that $\ell(v_4w_2) = \ell(v_5)+\ell(v_6)+\ell(v_7)$ and $v_5\in W_\star(e\wedge_{v_3}f)$, $v_6\in W^\star(e\wedge_{v_3}f)$, $v_7\in \Gae{e\wedge_{v_3}f}$. Then $BeBrB = Bv_2(e\wedge_{v_3}f)Bv_6v_7B$. We claim that~$\ell(er) = \ell(v_2(e\wedge_{v_3}f)v_6v_7) = \ell(v_2(e\wedge_{v_3}f)) + \ell(v_6v_7)$, which implies $BeBrB = Bv_2(e\wedge_{v_3}f)v_6v_7B = BerB$ by (ii). If it was not the case, By Lemma~\ref{lem:ewwt1}~(iii), $v_6v_7 = uv_8$ with $u\in \lambda^\star(e\wedge_{v_3}f)$, $\ell(v_6v_7) = \ell(v_8)+1$ and $\ell(v_2u) = \ell(v_2)-1$. But $\lambda^\star(e\wedge_{v_3}f)\subseteq \lambda^\star(f)$, $uv_5 = v_5u$ and $uv_2 = v_2u$ since $v_2$ lies in $W_\star(e\wedge_{v_3}f)$. Therefore, this leads to $r = w_1ew_2 = v_1v_2v_3v_4fw_2 = v_1v_2v_3fv_4w_2 = v_1v_2v_3fv_5uv_8 = v_1v_2uv_3fv_5v_8$. But this is impossible since $$\displaylines{\ell(r) = \ell(v_1v_2uv_3fv_5v_8)\leq \ell(v_1)+\ell(v_2u)+\ell(v_3)+\ell(v_5)+\ell(v_8)=\hfill\cr\hfill \ell(v_1)+\ell(v_2)-1+\ell(v_3)+\ell(v_5)+\ell(v_8)< \ell(w_1)+\ell(w_2) = \ell(r).}$$
\end{proof}
\subsection{Iwahori-Hecke algebra}
We introduce here the notion of a Iwahori-Hecke algebra in the general framework of Monoid Theory. The equivalent notion in the context of Group Theory is well-known~(\cite[Sec.~8.4]{Gec} for instance). There is no difficulty to translate the notion from Group Theory to Monoid Theory. The point is to verify that definitions and proofs can be written without using the existence of inverse elements. This is not the case for the whole theory (see Remarks~\ref{rem:1} and~\ref{rem:2} below) but the main results still hold as far as one considers the Iwahori-Hecke algebra associated with a subgroup. We have no find general references for Iwahori-Hecke Algebra of a monoid. This is why we start with an introduction to these notions with included proof.
\label{sect:monhecke}
For all this section, we assume $M$ is a finite monoid. We let $G$ denote its unit group and we fix a subgroup~$H$ of $G$.
We let~$\mathbb{C}[M]$ denote the monoid algebra of~$M$. An element of~$\mathbb{C}[M]$ has the form $\sum_{x\in M}\lambda_x x$ where the $\lambda_x$ belong to~$\mathbb{C}$. We set $$\varepsilon = \frac{1}{|H|}{\sum_{h\in H}}h$$ in~$\mathbb{C}[M]$. All the considered algebras are unit associative algebras, and all modules are left modules. We begin with two easy lemma whose proofs are left to the reader. \\
\begin{Lem}\label{lem:basic1}
Consider the $\mathbb{C}$-algebra~$\mathbb{C}^M$ of linear maps from $M$ to~$\mathbb{C}$ where the product is the convolution product~$\star$, defined by $$f\star g (x) = \sum_{y,z\in M, yz = x}f(y)g(z).$$ There is a canonical isomorphism of $\mathbb{C}$-algebra from~$\mathbb{C}[M]$ to $\mathbb{C}^M$ which sends~$X = \sum_{x\in M}\lambda_x x$ to the map~$\overline{X}: x\mapsto\lambda_x$.
\end{Lem}
The following lemma is immediate. We left the proof to the reader.
\begin{Lem}\label{lem:basic2}
(i) $\varepsilon^2 = \varepsilon$, and for every~$h$ in $H$ one has $h\varepsilon = \varepsilon h = h$.\\(ii) $\mathbb{C}[M]\varepsilon$ and $\mathbb{C}[M\slash H]$ are isomorphic as~$\mathbb{C}[M]$-modules and as~$\mathbb{C}$-vector spaces.
\end{Lem}
\begin{Rem}We remark that Lemma~\ref{lem:basic2} is no more true in general if we only assume~$H$ is a submonoid of $M$. Indeed, $\varepsilon$ is not necessarily an idempotent.\label{rem:1}\end{Rem}
\begin{Prop} \label{Prop:algisom}There is a canonical isomorphism between the following $\mathbb{C}$-algebras:\\(a) the subalgebra of $\mathbb{C}^M$ whose elements are the linear maps which are constant on the double-classes $H\backslash M\slash H$;\\
(b) the algebra~$\varepsilon \mathbb{C}[M] \varepsilon$;\\
(c) the algebra~$\left(End_{\mathbb{C}[M]}(\mathbb{C}[M\slash H])\right)^{op}$ of endomorphisms of $\mathbb{C}[M\slash H]$ considered as a $\mathbb{C}[M]$-module (for the opposite product).
\end{Prop}
\begin{proof}
The second and third algebras are isomorphic by \cite[Lemma~3.19]{CurRei}. This is clear that $\varepsilon X \varepsilon = X$ if and only if $X$ belongs to~$\varepsilon \mathbb{C}[M] \varepsilon$. Consider the notation of Lemma~\ref{lem:basic1}. Denote by $Hx_1,\ldots, Hx_k$ the left classes of $M$ modulo the subgroup~$H$. Let $X = \sum_{x\in M}\lambda_x x$ belong to~$\mathbb{C}[M]$. Then $$\varepsilon X = \frac{1}{|H|}\sum_{i = 1}^k\sum_{x\in H\!x_i}\sum_{h\in H} \lambda_{x} hx = \sum_{i = 1}^k\sum_{x\in H\!x_i}\left(\frac{1}{|H|}\sum_{y\in Hx_i} \alpha_{y,x}\lambda_{y} \right) x$$ where $\alpha_{y,x} = \#\{h\in H\mid hy = x\}$. If $M$ is a group, then $\alpha(y,x) = 1$ for every $y,x$ in $Hx_i$. In the general case one has $\alpha(y,x) = \frac{|H|}{|Hx_i|}$ because $H$ is a group. Therefore, $\varepsilon X = \sum_{x\in M}\left(\frac{1}{|Hx|}\sum_{y\in Hx}\lambda_{y} \right) x$, and $\varepsilon X = X$ if and only if $\overline{X}$ is constant on each left class. by a similar computation, $X \varepsilon = X$ if and only if $\overline{X}$ is constant on each right class. Therefore $\varepsilon X \varepsilon = X$ if and only if $\overline{X}$ is constant on each double class.
\end{proof}
\begin{Rem} The isomorphism between~$\left(End_{\mathbb{C}[M]}(\mathbb{C}[M\slash H])\right)^{op}$ and~$\varepsilon \mathbb{C}[M] \varepsilon$ is given by $f\mapsto \varepsilon f(\varepsilon)\varepsilon$ for every endomorphism~$f$.
\end{Rem}
Following Solomon~\cite{Sol} and Putcha~\cite{Put}, who consider the case of finite reductive monoids, we introduce the Iwahori-Hecke algebra~$\mathcal{H}(M,H)$:
\begin{definition}[Iwahori-Hecke algebra] Let $M$ be a finite monoid, and assume $H$ is a subgroup of $M$. Let~$\varepsilon = \frac{1}{|H|}\sum_{h\in H}h$ in~$\mathbb{C}[M]$. We define the~\emph{Iwahori-Hecke algebra~$\mathcal{H}(M,H)$ of $M$ relatively to $H$} to be the algebra~$\varepsilon \mathbb{C}[M] \varepsilon$.
\end{definition}
It is immediate that for every $\mathbb{C}[M]$-module~$N$, we get an induced structure of left $\mathcal{H}(M,H)$-module on~$\varepsilon N$. Proposition~\ref{Prop:algisom} explains why the Hecke algebra is interesting. Another motivation for such a definition is the following result.
\begin{Prop} Assume~$\mathbb{C}[M]$ is semisimple.\\(i) The Hecke algebra~$\mathcal{H}(M,H)$ is semisimple.\\ (ii) The map $N\mapsto \varepsilon N$ induced a one-to-one correspondence between the set of simple~$\mathbb{C}[M]$-modules in the induced~$\mathbb{C}[M]$-module $\mathbb{C}[M]\varepsilon = \mathbb{C}[M]\otimes_{\mathbb{C}[H]}\mathbb{C}[H]$ and the set of isomorphic classes of simple $\mathcal{H}(M,H)$-modules. Furthermore, the multiplicity of $N$ in $\mathbb{C}[M]\varepsilon$ is equal to the dimension of the $\mathcal{H}(M,H)$ module~$\varepsilon N$ considered as a $\mathbb{C}$-vector space.
\end{Prop}
Note that this is known by~\cite{OkPu} that~$\mathbb{C}[M]$ is semisimple for abstract finite monoids of Lie type ({\it cf.} Example~\ref{exe:ex2}), and therefore for finite reductive monoids.
\begin{proof}
Since~$\mathbb{C}[M]$ is semisimple, the algebra~$\varepsilon \mathbb{C}[M] \varepsilon$ is semisimple. Assume $N$ is a simple~$\mathbb{C}[M]$ module and let~$f$ belong to $Hom_{\mathbb{C}[M]}(\mathbb{C}[M]\varepsilon,N)$. For every $x$ in $\mathbb{C}[M]\varepsilon$ one has $f(x) = f(x\varepsilon) = xf(\varepsilon)$. If we consider $x = \varepsilon$, we get that $f(\varepsilon)$ belongs to~$\varepsilon N$. Moreover, it follows that the map~$f\mapsto f(\varepsilon)$ from $Hom_{\mathbb{C}[M]}(\mathbb{C}[M]\varepsilon,N)$ to $\varepsilon N$ is $\mathbb{C}$-linear and one-to-one. Thus $dim_\mathbb{C}(\varepsilon N)$ is equal to $dim(Hom_{\mathbb{C}[M]}(\mathbb{C}[M]\varepsilon,N))$, that is to the multiplicity of $N$ in~$\mathbb{C}[M]\varepsilon$. Now write $\mathbb{C}[M]\varepsilon = \oplus_i M_i$ where the $M_i$ are simple $\mathbb{C}[M]$-modules. Then $\varepsilon\mathbb{C}[M]\varepsilon = \oplus_i \varepsilon M_i$ and each~$\varepsilon M_i$ is a non-trivial simple $\mathcal{H}(M,H)$-modules: its $\mathbb{C}$-dimension is at least one, and for $m$ in $M_i$ such that $\varepsilon m\neq 0$ one has $\mathcal{H}(M,H)\varepsilon m = \varepsilon\mathbb{C}[M]\varepsilon m = \varepsilon M_i$ since $M_i$ is a simple $\mathbb{C}[M]$-module.
\end{proof}
By Proposition~\ref{Prop:algisom}, this is immediate to obtain a $\mathbb{C}$-basis of $\mathcal{H}(M,H)$:.
\begin{Prop} Let $\{D_1,\cdots, D_\ell\}$ be the set of double classes of $M$ modulo~$H$. We fix some arbitrary non-zero complex numbers~$a_1,\cdots a_\ell$, and we set $\displaystyle X_i = a_i\sum_{x\in D_i}x$ for $i$ in $\{1,\cdots,\ell\}$. Then the $X_i$ form a $\mathbb{C}$-basis for $\mathcal{H}(M,H)$. If we write $\displaystyle X_iX_j = \sum_{k = 1}^\ell \mu(i,j,k)X_k$, then $\mu(i,j,k) = \frac{a_ia_j}{a_k}\#\{(x,y)\in D_i\times D_j\mid xy = x_k\}$ where $x_k$ is an arbitrary fixed element of $D_k$. \end{Prop}
\begin{proof} The first part is clear. The second part come from the fact that $H$ is a group: we can write~$X_iX_j = \sum_{k = 1}^\ell \sum_{z\in D_k} \alpha(i,j,z) z$ where $\alpha(i,j,z) = \#\{(x,y)\in D_i\times D_j\mid xy = z\}$. But if $z$ belongs to $D_k$, then $\alpha(i,j,z) = \alpha(i,j,x_k)$. Indeed, if $z = h_1x_kh_2$ then the map $(x,y)\mapsto (h_1x,yh_2)$ is one-to-one from $\{(x,y)\in D_i\times D_j\mid xy = x_k\}$ onto $\{(x,y)\in D_i\times D_j\mid xy = z\}$.
\end{proof}
As explained in~\cite[Sec~4]{Sol} and in~\cite[Sec~2]{Put}, an important issue is to determined the structure constants~$\mu_{i,j,k}$ and, if possible, to suitably choose the $a_i$ so that the $\mathbb{Z}$-module generated by the~$a_i X_i$ becomes a $\mathbb{Z}$-subalgebra of $\mathcal{H}(M,B)$, in other words, so that the structure constants~$\mu_{i,j,k}$ belong to~$\mathbb{Z}$.
\begin{Rem}\label{rem:2} Let $\varphi$ belong to $End_{\mathbb{C}}(\mathbb{C}[M\slash H])$. Define $\dot{\varphi} : M\slash H\times M\slash H\to \mathbb{C}$ by $\varphi(xH) = \sum_{yH\in M\slash H}\dot{\varphi}(yH,xH)yH$. If $M$ is a group, it turns out that $\varphi$ belongs to~$End_{\mathbb{C}[M]}(\mathbb{C}[M\slash H])$, that is to $\mathcal{H}(M,H)$, if and only if $\dot{\varphi}$ is constant on the orbits of~$M$ on $M\slash H\times M\slash H$~\cite[Sec~8.4]{Gec}, which are naturally related to the double classes $HxH$ when $M$ is a group. This is no more true if we only assume $M$ is a monoid. One can verify that in the general case, $\varphi$ belongs to~$End_{\mathbb{C}[M]}(\mathbb{C}[M\slash H])$ if and only for every $xH$ and $yH$ in $M/H$ and every $g$ in $M$, one has $\dot{\varphi}(yH,gxH) = 0$ if $yH\cap gM$ is empty, and $$\dot{\varphi}(gyH,gxH) = \frac{1}{|\mathcal{C}_g(yH)|}\sum_{zH\in \mathcal{C}_g(yH)}\dot{\varphi}(zH,xH)$$ where $\mathcal{C}_g(yH) = \{zH\mid gzH = gyH\}$. If $M$ is a group then $yH\cap gM$ is never empty, and $\mathcal{C}_g(yH) = \{yH\}$.
\end{Rem}
\subsection{Finite reductive monoids}
\label{sousectionrappelamt}
We can now turn to the proof of Theorems~\ref{THintro} and~\ref{THintro2}. Let us recall the definition of finite reductive monoids~\cite{Ren2}, which is in the spirit of the definition of finite reductive groups~\cite{Ste}.
\begin{definition}[finite reductive monoid] Let~$\underline{M}$ be a reductive monoid defined over $\overline{\mathbb{F}}_q$. A finite submonoid~$M$ of~$\underline{M}$ is a \emph{finite reductive monoid} if there exists a surjective endomorphism of algebraic monoid~$\sigma : \underline{M}\to \underline{M}$ such that $$M = \{x\in \underline{M}\mid \sigma(x) = x\}.$$\label{def:finredmon} \end{definition}
\begin{Exe}
Consider a reductive monoid~$\underline{M}$ over $\overline{\mathbb{F}}_q$. The finite reductive monoid~$M$ associated with the map~$(x_{i,j})\mapsto (x_{i,j}^q)$ is~$M_n(\mathbb{F}_q)$. See~\cite{Sol} for more details.
\end{Exe}
Finite reductive monoids are special cases of abstract finite monoids of Lie type~\cite{Put3}, and their unit groups are finite groups of Lie type. Therefore, they are \emph{groups with a BN pair} and possess Borel subgroups and a generalised Renner monoid~$R$ ({\it cf.} Example~\ref{exe:ex2}). As a consequence, we can associate with $M$ a generic Hecke algebra~$\mathcal{H}(R)$ as defined in Section~2, and a Iwahori-Hecke algebra as defined in Section~\ref{sect:monhecke}. Our objective is to prove Theorem~\ref{THintro2}, which explains how these two notions are related.
\begin{Not} Assume~$M$ is a finite reductive monoid over $\overline{\mathbb{F}}_q$, and consider the notation of Definition~\ref{def:finredmon}. There exists a maximal torus~$\underline{T}$ of~$G(\underline{M})$ and a Borel subgroup~$\underline{B}$ of $\underline{G} = G(\underline{M})$ that contains $\underline{T}$ such that $\sigma(\underline{T}) = \underline{T}$ and $\sigma(\underline{B}) = \underline{B}$ \cite{Ste,Ren2}. Moreover, $\sigma(N_{\underline{G}}(\underline{T})) = N_{\underline{G}}(\underline{T})$. Let $\underline{R}$ be the Renner monoid associated with~$\underline{M}$, and $\underline{W}$ be its unit group. Then $\sigma$ induces an isomorphism~$\sigma : \underline{R} \to \underline{R}$. We set $$G = \{b\in \underline{G}\mid \sigma(g) = g\}$$ $$B = \{b\in \underline{B}\mid \sigma(b) = b\}$$ $$T = \{t\in \underline{T}\mid \sigma(t) = t\}$$ $$W = \{w\in \underline{W}\mid \sigma(w) = w\}$$ $$R = \{r\in \underline{R}\mid \sigma(r) = r\}$$ $$\Lambda = \{e\in \underline{\Lambda}\mid \sigma(e) = e\}$$\label{not:fintredmon}
\end{Not}
\begin{Prop}\cite{Ren2,Ste} Consider Notation~\ref{not:fintredmon}. The group~$G$ is the unit group of~$M$, and $B$ is a Borel subgroup of~$G$ with maximal torus~$T$. The Renner monoid of $M$ is~$R$. The unit group of $R$ is $W$, and $\Lambda$ is the cross section lattice of~$R$ associated with~$B$. Denote by $\underline{S}$ the canonical generating set of~$\underline{W}$ associated with~$\underline{T}$ and~$\underline{B}$. For a conjugated class~$X$ of elements of~$S$ under~$\sigma$, we let~$\Delta_X$ denote the greatest element of~$W_X$. Let $S$ be the set of all $\Delta_X$. Then~$(W,S)$ is a Coxeter system, and $(R,\Lambda,S)$ is a generalised Renner-Coxeter system. Moreover, we have a disjoint union \emph{Bruhat decomposition}~$M = \cup_{r\in R}BrB$. \label{Prop:redtofinred}
\end{Prop}
From the Bruhat decomposition of~$M$, we deduce for every $r$ in $R$ that $$BrB = \{x\in \underline{B}r\underline{B}\mid \sigma(x) = x\}.$$ It is immediate that for $e$ in~$\Lambda$ one has $\sigma(\underline{\lambda}(e)) = \underline{\lambda}(e)$ and $\sigma(\underline{\lambda}_\star(e)) = \underline{\lambda}_\star(e)$ in $\underline{R}$, with obvious notation. Therefore, $\omega_X$ belongs to $\lambda(e)$ in $R$ ({\it resp.} to~$\lambda_\star(e)$) if and only if $X$ is included in~$\underline{\lambda}_(e)$ ({\it resp.} to~$\underline{\lambda}_\star(e)$) in $\underline{R}$.
\begin{Lem} Consider Notation~\ref{not:fintredmon}. Denote by $\ell$ the length function on~$R$.\\ (i) Let~$r$ lie in~$R$ and~$s$ lie in~$S$. Then $$B s B r B = \left\{\begin{array}{ll}BrB,&\textrm{if } \ell(sr) = \ell(r);\\BsrB,&\textrm{if } \ell(sr) = \ell(r)+1;\\BsrB\cup BrB,&\textrm{if } \ell(sr) = \ell(r)-1.\end{array}\right.$$
(ii) Let~$r$ lie in~$R$ and~$s$ lie in~$S$. Then $$B r B s B = \left\{\begin{array}{ll}BrB,&\textrm{if } \ell(rs) = \ell(r);\\BrsB,&\textrm{if } \ell(rs) = \ell(r)+1;\\BrsB\cup BrB,&\textrm{if } \ell(rs) = \ell(r)-1.\end{array}\right.$$
(iii) Let $e$ lie in $\Lambda_{\scriptscriptstyle \circ}$ and $r$ lie in $R$. Then $$B e B r B = BerB.$$\end{Lem}
\begin{proof} The result follows from Proposition~\ref{Pr:lienlongBB}. (i) Denote by $\underline{\ell}$ the length function on~$\underline{R}$. Let $r$ lie in $R$ and $\Delta_X$ lie in $S$ ({\it cf.} Proposition~\ref{Prop:redtofinred}). Fix a minimal representative word~$x_1\cdots x_k$ on~$\underline{S}$ of $\Delta_X$. Using the map~$\sigma$, we deduce that there is three possibilities:\\ (a) $\forall x \in X$, $\underline{\ell}(xr) = \underline{\ell}(r)+1$.\\ In this case, $\underline{\ell}(\omega_Xr) = \underline{\ell}(r)+\underline{\ell}(\omega_X)$, $\underline{B} \omega_X \underline{B} r \underline{B} = \underline{B} \omega_Xr \underline{B}$ and $\ell(\omega_Xr) = \ell(r)+1$. Therefore, $B \omega_X B r B \subseteq \{x\in\underline{B} \omega_Xr \underline{B}\mid \sigma(x) = x\} = B \omega_Xr B$. But $B \omega_X B r B$ is an union of double classes~$ByB$. Then the latter inclusion has to be an equality.\\ (b) $\forall x \in X$, $\underline{\ell}(xr) = \underline{\ell}(r)$.\\
In this case $\omega_Xr = r $, and in particular $\underline{\ell}(\omega_Xr) = \underline{\ell}(r)$, $\underline{B} \omega_X \underline{B} r \underline{B} = \underline{B} r \underline{B}$ and $\ell(\omega_Xr) = \ell(r)$. It follows that~$B \omega_X B r B = B r B$ as in the previous case.\\
(c) $\forall x \in X$, $\underline{\ell}(xr) = \underline{\ell}(r)-1$.\\ In this case, $\underline{\ell}(\omega_Xr) = \underline{\ell}(r)-\underline{\ell}(\omega_X)$, $\ell(\omega_Xr) = \ell(r)-1$ and $\underline{B} \omega_X \underline{B} r \underline{B} = \bigcup_{v}\underline{B} vr \underline{B}$, where~$v$ ranges over all the elements $x_{i_1}\cdots x_{i_j}$ with $1\leq i_1<\cdots< i_j\leq k$ and $0\leq j\leq k$. But for such an element~$v$ of~$\underline{R}$, the set $\{x\in \underline{B} vr \underline{B}\mid \sigma(x) = x\}$ is empty, except if $vr$ belongs to~$R$, that is $v = 1$ or $v = \omega_X$. Therefore, $\{x\in \underline{B} \omega_X \underline{B} r \underline{B}\mid \sigma(x) = x\} = B \omega_Xr B\cup B r B$. But~$\underline{B} \omega_X \underline{B} r \underline{B} = \bigcup_{b\in \underline{B}}\underline{B}\omega_Xbr \underline{B}$. We deduce that~$$M\cap \underline{B} \omega_X \underline{B} r \underline{B} = \bigcup_{b\in \underline{B}}M\cap \underline{B}\omega_Xbr \underline{B} = \bigcup_{b\in B}B\omega_Xbr B = B\omega_XBr B.$$(ii) the proof is similar to~(ii). \\(iii) $BeBrB$ is included in $\{x\in \underline{B} er \underline{B}\mid \sigma(x) = x\} = BerB$. But $BeBrB$ is an union of double classes $ByB$. Therefore, $BeBrB = BerB$.
\end{proof}
We are now ready to prove Theorem~\ref{THintro2}.
\begin{proof}[Proof of Theorem~\ref{THintro2}] By Theorem~\ref{Th:genThe} and Definition~\ref{def:genHechalg}, $\mathbb{C}\otimes_{\mathbb{Z}}\mathcal{H}_q(R)$ is the unique $\mathbb{C}$-algebra such that the relations stated in Theorem~\ref{Th:genThe} hold. But, by Section~\ref{sect:monhecke}, $\mathcal{H}(M,B)$ is a $\mathbb{C}$-algebra over the free $\mathbb{C}$-module with basis $\sum_{x\in BrB}x$, for $r\in R$. We set $$T_r = \frac{q^{\ell(r)}}{|BrB|} \sum_{x\in BrB}x$$ in~$\mathcal{H}(M,B)$. We are going to prove that the relations stated in Theorem~\ref{Th:genThe} hold
in~$\mathcal{H}(M,B)$ for the basis $T_r$, $r\in R$. The main arguments are like in~\cite[Sec.~4]{Sol}. Denote by~$$\pi: \mathcal{H}(M,B)\to \mathbb{C}$$ the restriction of the one-dimensional representation from~$\mathbb{C}[M]\to\mathbb{C}$ that sends every $g$ in $M$ to $1$. We have $\pi (T_r) = \pi(\frac{q^{\ell(r)}}{|BrB|} \sum_{x\in BrB}x) = q^{\ell(r)}$. Let $r_1,r_2,r_3$ lie in $R$ such that $Br_1B Br_2B = Br_3B$. Applying the map~$\pi$, we get $$T_{r_1}T_{r_2} = q^{\ell(r_1)+\ell(r_2)-\ell(r_3)} T_{r_3}.$$
Therefore, it follows from Lemma~\ref{not:fintredmon} that
$$\begin{array}{ll}T_sT_r = T_{sr},&\textrm{if }s\in S \textrm{ and } \ell(r) = \ell(r)+1;\\
T_sT_r = qT_{r},&\textrm{if }s\in S\textrm{ and }\ell(sr) = \ell(r);\\ T_eT_r = q^{\ell(r)-\ell(er)}T_{er},&\textrm{if }e\in \Lambda_{\scriptscriptstyle \circ}.
\end{array}$$
Assume $s$ lies in $S$ and $r$ lies in~$R$ such that $\ell(sr) = \ell(r)-1$. Denote by $(w_1,e,w_2)$ the normal decomposition of~$r$. By Lemma~\ref{lem:ewwt1}(i), $\ell(sw_1) = \ell(w_1)-1$ and $\ell(sw_1ew_2) = \ell(sw_1)+\ell(w_2)$. Therefore, $T_sT_{w_1} = qT_{sw_1} + (1-q)T_{w_1}$, by~\cite[Theorem~8.4.6]{Gec}, and $$T_sT_r = T_sT_{w_1}T_{ew_2} = qT_{sw_1}T_{ew_2} + (1-q)T_{w_1}T_{ew_2} = qT_{sr} + (1-q)T_{r}.$$
\end{proof}
Now, using Theorem~\ref{Th:genThe}, Theorem~\ref{THintro} is a corollary of Theorem~\ref{THintro2}. More precisely,
gathering Corollary~\ref{cor:preshHR} and Theorem~\ref{THintro2}, we get the following result.
\begin{Cor} Let $M$ be a finite reductive monoid over~$\overline{\mathbb{F}}_q$. Consider Notation~\ref{not:fintredmon}. Then the Iwahori-Hecke algebra~$\mathcal{H}(M,B)$ admits the following $\mathbb{C}$-algebra presentation:
\begin{center}\label{THconclu}
\begin{tabular}{lll}
(HEC1)&$T^2_s = (q-1)T_1+qT_{s}$,&$s\in S$;\\
(HEC2)&$|T_s,T_t\rangle^m = |T_t,T_s\rangle^m$,&$(\{s,t\},m)\in \mathcal{E}(\Gamma)$;\\
(HEC3)&$T_sT_e = T_eT_s$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda^\star(e)$;\\
(HEC4)&$T_sT_e = T_eT_s = qT_e$,& $e\in \Lambda_{\scriptscriptstyle \circ}$, $s\in \lambda_\star(e)$;\\
(HEC5)&$T_eT_wT_f = q^{\ell(w)}T_{e\wedge_wf}$,& $e,f\in \Lambda_{\scriptscriptstyle \circ}$, $w\in \textrm{Red}(e,f)$.
\end{tabular}
\end{center}
\end{Cor}
\ \\\ \\\noindent {\bf Acknowledgements.} The author is in debt with F. Digne, B. Leclerc, for useful discussions and with C. Mokler, M. Putcha and L. Renner for helpful email correspondences.
|
\section{Introduction}
The Bayesian approach~\cite{Robert2007} to inference plays an increasingly
important role in particle physics research. This is due, in part, to a better
understanding of Bayesian reasoning within the field and the concomitant abating
of the frequentist/Bayesian debate. Moreover, the small but growing number of
successful applications provide concrete examples of how the Bayesian approach
fares in practice.
In spite of these successes the specification of priors in a principled way
remains a conceptual and practical hurdle. In the so-called subjective Bayesian
approach~\cite{OHagan1994}, one is invited to elicit the prior based on one's
actual beliefs about the unknown parameters in the problem. If one has
well-understood information, for example based on subsidiary measurements or
simulation studies, one can encode this partial information in an evidence-based
prior~\cite{Cox2007}. Such priors generally occasion little or no controversy.
On the other hand, if one knows little about a given parameter, or if one
prefers to act as if one knows little, then it is far from clear how one ought
to encode this minimal information in a prior probability.
Since there is, in fact, no unique way to model prior ignorance, a viewpoint
has evolved in which this lack of knowledge is represented by one's willingness
to {\em adopt} a standard prior for certain parameters~\cite{Kass1996}, just as
one has adopted a standard for quantities such as length and weight. In this
spirit, our field adopted as a convention a uniform (flat) prior for unknown
cross sections and other parameters (see for example Ref.~\cite{Bertram2000}),
mainly because this prescription is simple to implement and seems to embody
Laplace's principle of insufficient reason. Unfortunately, uniform priors are
both conceptually and practically flawed. The conceptual difficulty is with
their justification: lack of knowledge about a parameter $\sigma$ implies lack
of knowledge about any one-to-one transform $\sigma^{\prime}$ of $\sigma$, and
yet a prior distribution that is uniform in $\sigma$ will not be so in
$\sigma^{\prime}$ if the transform is non-linear. The practical problem is
that careless use of uniform priors can lead to improper posteriors, that is,
posteriors whose integrals are infinite and which can therefore not be used to
assign meaningful probabilities to subsets of parameter space. An example of
this pathology is found in a common method for reporting the exclusion of a new
physics signal, where one estimates an upper limit from a posterior distribution
for the signal's production cross section. When constructed from a Poisson
probability mass function for the observations, a flat prior for the signal
cross section, and a truncated Gaussian prior for the signal acceptance, this
posterior is actually improper. However, for small acceptance uncertainties
the divergence of the upper limit is often concealed by the inevitable
truncation of numerical computations~\cite{Demortier2002}.
The specification of priors that encode minimal information is of such
importance in practice that a large body of literature exists describing
attempts to construct priors that yield results with provably useful
characteristics. These priors are typically arrived at using formal rules.
In this paper, therefore, we refer to them as {\em formal
priors}~\cite{FormalPriors} to distinguish them from evidence-based priors.
Many such formal rules exist~\cite{Kass1996}. In this paper we study, and
then recommend, a rule which is arguably the most successful: that developed by
Bernardo~\cite{Bernardo1979} and Berger and Bernardo~\cite{Berger1989,
Berger1992a,Berger1992b}. Formal priors constructed according to the
Bernardo-Berger rule are called {\em reference priors}, a somewhat unfortunate
name given that the term reference prior is sometimes used as a synonym for
what we have called a formal prior.
Reference priors have been shown to yield results with several desirable
properties, all of which should appeal to particle physicists. Therefore, in
principle such priors could be a foundation for Bayesian inference in particle
physics research. However, reference priors and the associated methods
collectively referred to as reference analysis~\cite{Bernardo2005,Demortier2005}
have yet to enter the field in a significant way. The purpose of this paper is
to initiate this process by applying the Bernardo-Berger method to a familiar,
but important class of problems, namely that of calculating posterior densities
for signal cross sections.
In the next section we describe the general goals of reference prior
construction and show how these are implemented via the concept of
missing information. For simplicity we limit that discussion to one-parameter
problems. Section~\ref{InclPartInfo} then considers the treatment of nuisance
parameters about which prior information is available. Examples of such
parameters include detector calibration constants, background contaminations,
geometrical acceptances, and integrated luminosities. We describe two methods
for handling these parameters, depending on the type of information that is
available about them. These methods are then applied to counting experiments
with uncertain background contamination and effective luminosity. In the
simplest cases we have obtained analytical expressions for the marginal
posterior for the quantity of interest. For the general case we have developed
a numerical algorithm. Some appealing properties of these posteriors are
examined in Sec.~\ref{ValidationStudies}. In Sec.~\ref{SingleTop} the
reference prior methodology is applied to a recent measurement of the production
cross section for single top quarks at the Tevatron. Final comments are
presented in Sec.~\ref{FinalComments}.
\section{Reference Priors}
\label{sec:RefPriors}
In 1979, Bernardo~\cite{Bernardo1979} introduced a formal rule for constructing
what he called a reference prior. The goal was to construct a prior which, in a
sense to be made precise, contained as little information as possible relative
to the statistical model under consideration. By statistical model he meant a
representation of the entire experimental design, including the probability
distribution of the data, the sampling space, and the stopping rule. Hence, by
construction reference priors depend on all these aspects of a statistical
model, and so will inferences derived from data with the help of a reference
prior. This may seem to violate the so-called {\em likelihood
principle}~\cite{Birnbaum1962}, according to which all the information
about unknown model parameters obtainable from an experiment is contained in the
likelihood function, i.e. the probability distribution of the data, evaluated
at the observations and viewed as a function of the parameters. While this
is formally true, it should be kept in mind that the likelihood principle
applies after data have been observed, whereas reference priors are constructed
at the experimental design stage. Their purpose is to approximate a consensus
of opinions that is suitable for scientific communication. This is generally
unproblematic in large-sample situations, where posterior inferences are
dominated by the likelihood function. In small sample cases however, results
obtained with reference priors should be considered preliminary, and a careful
study should be conducted of the degree to which inferences about the physics
model underlying the observations can be trusted. This can be achieved by
examining the sensitivity of the results to changes in the prior, and
subsequently assessing the need for additional observations.
Reference priors have several desirable properties, including
\begin{enumerate}
\item {\em generality:} a well-defined algorithm exists to create a reference
prior for almost any type of estimation problem, and the resulting
posterior is proper;
\item {\em invariance:} given a one-to-one map from a parameter $\theta$ to
a parameter $\phi$, applying the reference prior construction separately
to $\theta$ and $\phi$ yields posteriors that are related by the correct
transformation law, $\pi(\phi\,|\,x) = \pi(\theta\,|\,x) \,
|\partial\theta/\partial\phi|$;
\item {\em sampling consistency:} the posterior densities from an ensemble of
experiments tend to cluster around the true values of the parameters; and
\item {\em coherence:} inferences derived from reference priors avoid
marginalization paradoxes.
\end{enumerate}
Marginalization paradoxes~\cite{Dawid1973} arise in multiparameter problems
when a posterior density can be calculated in different ways that ought to give
the same answer but do not (see Fig.~\ref{fig:mparadox}).
\begin{figure}
\includegraphics[width=15cm]{mparadox}
\caption{Let $\vec{x}$ be a dataset modeled by the probability density
$p(\vec{x}\,|\,\theta,\phi)$, where $\theta$ and $\phi$ are unknown parameters,
and consider the following two paths to a posterior density for $\theta$. In
path~1, we use a formal prior $\pi_{\scriptscriptstyle\rm F1}(\theta,\phi)$ to
construct the joint posterior for $\theta$ and $\phi$, and then integrate out
$\phi$. Suppose that the result of this operation only depends on the data
$\vec{x}$ through the statistic $t$; this gives us
$\pi_{\scriptscriptstyle 1}(\theta\,|\,t)$. For path~2, assume further that
the sampling distribution of $t$ only depends on $\theta$. We can then
directly construct a posterior for $\theta$, say
$\pi_{\scriptscriptstyle 2}(\theta\,|\,t)$. A marginalization paradox occurs if
$\pi_{\scriptscriptstyle 1}(\theta\,|\,t)\ne
\pi_{\scriptscriptstyle 2}(\theta\,|\,t)$ regardless of the choice of
prior $\pi_{\scriptscriptstyle 2}(\theta)$ in path~2.
\label{fig:mparadox}
}
\end{figure}
This incoherence does not happen with subjective or evidence-based priors because
these priors are always proper. With formal priors however, it can only be
avoided by allowing the joint prior for all the parameters in a given statistical
model to depend on the quantity of interest. This is in fact what the reference
prior construction does. For a simple illustration, consider $n$ measurements
$x_{i}$ from the normal model with unknown mean $\mu$ and standard deviation
$\sigma$. The likelihood function is:
\begin{equation}
p(\vec{x}\,|\,\mu,\sigma)
\;=\; \prod_{i=1}^{n} \frac{e^{-\tfrac{1}{2}
\bigl(\tfrac{x_{i}-\mu}{\sigma}\bigr)^{2}}}{\sqrt{2\pi}\,\sigma}
\;=\; \frac{e^{-\tfrac{n-1}{2}\bigl(\tfrac{s}{\sigma}\bigr)^{2}-
\tfrac{n}{2}\bigl(\tfrac{\bar{x}-\mu}{\sigma}\bigr)^{2}}}
{\bigl(\sqrt{2\pi}\,\sigma\bigr)^{n}},
\end{equation}
where $\bar{x}=\sum_{i}^{n}x_{i}/n$ and $s^{2}=\sum_{i}^{n}(x_{i}-\bar{x})^{2}
/(n-1)$. When $\mu$ is the quantity of interest, the reference prior derived
from this likelihood is $1/\sigma$. Restricting the remaining calculations to
the case $n=2$ for convenience, the joint reference posterior for $\mu$ and
$\sigma$ is then:
\begin{equation}
\pi(\mu,\sigma\,|\,\vec{x})\;=\;\frac{\sqrt{2}\,s}{\pi\,\sigma^{3}}\;
e^{-\tfrac{1}{2}\bigl(\tfrac{s}{\sigma}\bigr)^{2}-
\bigl(\tfrac{\mu-\bar{x}}{\sigma}\bigr)^{2}}.
\label{eq:MuSigmaPost}
\end{equation}
Integrating out $\sigma$ yields the marginal $\mu$-posterior, which is a Cauchy
distribution with location parameter $\bar{x}$ and scale parameter $s/\sqrt{2}$.
Suppose however that our interest lies in the standardized mean
$\theta=\mu/\sigma$. In a non-reference approach one would perform the
transformation $(\mu,\sigma)\rightarrow (\theta,\sigma)$ in
Eq.~\eqref{eq:MuSigmaPost} and integrate out $\sigma$ in order to obtain the
marginal $\theta$-posterior. The latter only depends on the data through the
statistic $t\equiv \sqrt{2}\,\bar{x}/s$:
\begin{equation}
\pi(\theta\,|\,\vec{x})
\;=\;\frac{e^{-\tfrac{\theta^{2}}{1+t^{2}}}}{\sqrt{\pi\,(1+t^{2})}}\,
\left[1+\textrm{erf}\left(\frac{t\theta}{\sqrt{1+t^{2}}}\right)\right]
\;=\;p(\theta\,|\,t),
\label{eq:phiPost}
\end{equation}
where $\textrm{erf}$ is the error function. Furthermore, the sampling
distribution of $t$ turns out to depend on $\theta$ only and is a noncentral
Student's $t$ distribution for one degree of freedom and with noncentrality
parameter $\theta$:
\begin{equation}
p(t\,|\,\theta)\;=\; \frac{e^{-\theta^{2}}}{\pi\,(1+t^{2})}\;+\;
\frac{\theta\,t\,e^{-\tfrac{\theta^{2}}{1+t^{2}}}}{\sqrt{\pi}\,(1+t^{2})^{3/2}}\;
\left[1 + \textrm{erf}\left(\frac{\theta\, t}{\sqrt{1+t^{2}}}\right)\right].
\label{eq:tSamplingDis}
\end{equation}
It is clear that there exists no prior (no function of $\theta$ only) that,
multiplied by the likelihood~\eqref{eq:tSamplingDis}, leads to the
posterior~\eqref{eq:phiPost}. Hence the marginalization paradox: someone who
is only given the data value of the statistic $t$ will be able to make
inferences about the parameter $\theta$, but these inferences are guaranteed
to disagree with those previously made by the Bayesian who had access to the
full dataset. Resolution of this paradox hinges on the realization that lack
of information about $\mu$ is not the same as lack of information about
$\theta$. Therefore, the choice of which quantity is of interest must be
done {\em before} calculating the prior. Since reference priors are derived
from the likelihood function, the latter must first be expressed in terms of
the relevant parameters, $\theta$ and $\sigma$:
\begin{equation}
p(\vec{x}\,|\,\theta,\sigma)
\;=\; \frac{e^{-\tfrac{n-1}{2}\bigl(\tfrac{s}{\sigma}\bigr)^{2}-
\tfrac{n}{2}\bigl(\tfrac{\bar{x}}{\sigma}-\theta\bigr)^{2}}}
{\bigl(\sqrt{2\pi}\,\sigma\bigr)^{n}}.
\end{equation}
Applying the reference algorithm to this likelihood while treating $\theta$ as
the quantity of interest yields the prior $\pi(\theta,\sigma) =
1/(\sigma\,\sqrt{1+\theta^{2}/2})$, which is very different from the prior
$1/\sigma$ obtained by treating $\mu$ as the quantity of interest. The
resulting reference posterior suffers no marginalization problems. Further
details about this example can be found in Refs.~\cite{Bernardo1979,Bernardo2005}.
Our discussion of marginalization also helps to clarify the behavior of
reference posteriors under transformations in the multiparameter setting.
Reference posteriors are invariant under one-to-one transformations of the
parameter of interest, but not under transformations that redefine the parameter
of interest by mixing in one or more nuisance parameters. However, redefining
the nuisance parameters is permitted. Suppose for example that $\varphi$ is the
parameter of interest, $\nu$ the nuisance parameter(s), and consider an
invertible transformation of the form $(\varphi,\nu)\rightarrow
(\varphi,\lambda)$, where $\lambda$ is a function of both $\varphi$ and $\nu$.
Then the reference posterior for $\varphi$ is unchanged by the transformation.
Reference priors on unbounded parameter spaces are usually improper, which
invalidates the application of Bayes' theorem. To circumvent this problem one
introduces a nested sequence of compact subsets $\Theta_{1}\subset\Theta_{2}
\subset\ldots$ of the parameter space $\Theta$, such that $\Theta_{\ell}
\rightarrow\Theta$ as $\ell\rightarrow\infty$. Given an improper prior
$\pi(\theta)$, its restriction to $\Theta_{\ell}$ will be proper, so that Bayes'
theorem can be applied to construct the corresponding restricted posterior
$\pi_{\ell}(\theta|x)$. The unrestricted posterior for the entire parameter
space is then defined by the limit of the $\pi_{\ell}(\theta|x)$ as
$\ell\rightarrow\infty$. The practical justification for this procedure is
that one often knows the shape, but not the size, of the physical region of
parameter space where the prior has nonzero weight. As this size is typically
very large, the limiting posterior can be viewed as an approximation to the
posterior on the physical region.
Interestingly, the limiting posterior can also be obtained by direct, formal
application of Bayes' theorem to the improper prior $\pi(\theta)$, provided the
marginal distribution of the data,
\begin{equation}
m(x)\;\equiv\; \int p(x|\theta)\, \pi(\theta) \,d\theta,
\end{equation}
is finite. It can then be shown that the restricted posteriors
$\pi_{\ell}(\theta|x)$ converge {\em logarithmically} to their limit
$\pi(\theta|x)$:
\begin{equation}
\lim_{\ell\rightarrow\infty} D[\pi_{\ell}(\theta|x),\,\pi(\theta|x)]\;=\;0,
\end{equation}
where
\begin{equation}
D[p(\theta),\,q(\theta)]\;\equiv\;
\int\!q(\theta)\,\log\frac{q(\theta)}{p(\theta)}\,d\theta
\end{equation}
is the Kullback-Leibler divergence between $p(\theta)$ and $q(\theta)$. This
divergence is a parametrization-independent, non-negative measure of the
separation between two densities; it is zero if and only if the densities are
identical. Unfortunately, pointwise logarithmic convergence is not enough to
avoid inferential inconsistency in some special cases~\cite{Berger2009}, so
that a stronger form of convergence is needed, {\em expected} logarithmic
convergence:
\begin{equation}
\lim_{\ell\rightarrow\infty}
\mathbb{E}_{m}\Bigl\{D[\pi_{\ell}(\theta|x),\,\pi(\theta|x)]\Bigr\}\;=\;0,
\end{equation}
where the expectation is taken with respect to the marginal density $m(x)$.
The above discussion motivates the following terminology~\cite{Berger2009}.
Given a statistical model on a parameter space $\Theta$, a {\em standard prior}
is a strictly positive and continuous function on $\Theta$ that yields a proper
posterior. A {\em permissible prior} is a standard prior for which the
posterior is the expected logarithmic limit of a sequence of posteriors defined
by restriction to compact sets.
\subsection{The Concept of Missing Information}
Reference priors make use of the notion of expected intrinsic information.
For one observation from a model $p(x|\theta)$, the expected intrinsic
information about the value of $\theta$ when the prior is $\pi(\theta)$ is
given by the functional
\begin{equation}
I\{\pi\} \;=\; \mathbb{E}_{m}\Bigl\{ D[\pi(\theta|x),\;\pi(\theta)] \Bigr\}.
\label{eq:xii}
\end{equation}
The more informative the observation, the greater the expected separation
between the posterior and the prior. The larger this separation, the greater
the expected intrinsic information $I\{\pi\}$. Thus, $I\{\pi\}$ measures the
amount of information about the value of $\theta$ that might be expected from
one observation when the prior is $\pi(\theta)$.
Suppose next that we make $k$ independent observations $x_{(k)}=\{x_{1}, x_{2},
\ldots, x_{k}\}$ from the model $p(x|\theta)$. The definition of expected
intrinsic information can be generalized to include all $k$ observations:
\begin{equation}
I_{k}\{\pi\} \;=\;
\mathbb{E}_{m}\Bigl\{ D[\pi(\theta|x_{(k)}),\;\pi(\theta)] \Bigr\},
\label{eq:xii2}
\end{equation}
where the expectation is a $k$-dimensional integral over $x_{(k)}$ weighted by:
\begin{equation}
m(x_{(k)})\;\equiv\; \int p(x_{(k)}|\theta)\, \pi(\theta) \,d\theta
\;=\; \int \left[\prod_{i=1}^{k}p(x_{i}|\theta)\right]\;\pi(\theta)\,d\theta.
\end{equation}
As the sample size $k$ grows larger, one expects the amount of information about
$\theta$ to increase, and in the limit $k\rightarrow\infty$, the true value of
$\theta$ would become exactly known. In this sense, the limit $I_{\infty}\{\pi\}
\equiv\lim_{k\rightarrow\infty}I_{k}\{\pi\}$ represents the {\em missing} information
about $\theta$ when $\pi(\theta)$ is the prior. This concept of missing
information is central to the construction of reference priors.
\subsection{Reference Priors for One-Parameter Models}
\label{sec:refprior}
The goal of reference analysis is to contruct a prior that maximizes the missing
information. This maximization cannot be done directly however, because
$I_{\infty}$ typically diverges. To avoid this problem, one first constructs
the prior $\pi_{k}(\theta)$ that maximizes $I_{k}$, and then takes the limit of
$\pi_{k}(\theta)$ as $k\rightarrow\infty$. Additional care is required when
the parameter space $\Theta$ is unbounded, since in that case the prior that
maximizes $I_{k}$ is often improper, and $I_{k}$ is undefined for improper
priors. The solution is to define reference priors via their restrictions on
arbitrary compact subsets $\Theta_{\ell}$ of $\Theta$. Thus one is led to the
formal definition of a reference prior for $\theta$ as any permissible prior
$\pi_{R}(\theta)$ that satisfies the so-called maximizing missing information
(MMI) property, namely that
\begin{equation}
\lim_{k\rightarrow\infty}\Bigl[ I_{k}\{\pi_{R,\ell}\}-I_{k}\{\pi_{\ell}\}\Bigr]
\;\ge\;0
\label{eq:refdef}
\end{equation}
for any compact set $\Theta_{\ell}$ and candidate prior $\pi(\theta)$, where
$\pi_{R,\ell}$ and $\pi_{\ell}$ are the renormalized restrictions of $\pi_{R}$
and $\pi$ to $\Theta_{\ell}$. A candidate prior is a standard prior that
incorporates any prior knowledge about $\theta$.
A key result is the following constructive definition of the reference prior
$\pi_{R}(\theta)$~\cite{Berger2009},
\begin{eqnarray}
\pi_{R}(\theta) \; & = & \; \lim_{k\rightarrow\infty}
\frac{\pi_{k}(\theta)}{\pi_{k}(\theta_{0})}, \nonumber \\
\mbox{with} \; \pi_k(\theta) \; & = & \; \exp\left\{ \int p(x_{(k)}\,|\,\theta)\,
\ln\left[\frac{p(x_{(k)}\,|\,\theta)\, h(\theta)}
{\int p(x_{(k)}\,|\,\theta)\, h(\theta)\, d\theta}\right]\,
dx_{(k)}\right\},
\label{eq:refformula}
\end{eqnarray}
where $\theta_{0}$ is an arbitrary fixed point in $\Theta$, $h(\theta)$ is any
continuous, strictly positive function, such as $h(\theta)=1$, and
$p(x_{(k)}\,|\,\theta) = \prod_{i=1}^k p(x_i\,|\,\theta)$ is the probability
model for a sample of $k$ independent observations. We emphasize that this
constructive definition only guarantees that the MMI property~\eqref{eq:refdef}
is satisfied. The permissibility part of the reference prior definition must be
separately verified. However, the proponents of reference priors view the MMI
property as considerably more important than permissibility~\cite{Berger2009},
and also believe that it would be highly unusual for a prior satisfying the MMI
property to fail permissibility~\cite{Berger1992b} (counter-examples are known,
but they are rather exotic).
A further useful result that we shall exploit is that, when certain regularity
conditions are met --- essentially those that guarantee asymptotic normality of
the posterior --- the reference prior for models with one continuous parameter
reduces to the well-known Jeffreys prior~\cite{Kass1996},
\begin{equation}
\pi_{R}(\theta)\;=\;\sqrt{\mathbb{E}\left[-\frac{d^{2}}{d\theta^{2}}
\ln p(x\,|\,\theta)\right]},
\label{eq:Jpr}
\end{equation}
where the expectation is taken with respect to the sampling model
$p(x\,|\,\theta)$. In general the analytical derivation of reference priors
can be extremely challenging. However, Eq.~\eqref{eq:refformula} is amenable to
numerical integration~\cite{Berger2009}.
We emphasize that the definition and results described in this section apply
only to the case where the model of interest depends on a single parameter.
A generalization to the multi-parameter case has been formulated and shown to
have the properties listed at the beginning of
Sec.~\ref{sec:RefPriors}~\cite{Bernardo2005}.
We shall not describe it here however, except for when evidence-based priors are
specified for the additional parameters. This is a very common situation in
high energy physics, and will be discussed next.
\section{Nuisance Parameters}
\label{InclPartInfo}
The reference prior algorithm described in Sec.~\ref{sec:refprior} pertains to
models containing no nuisance parameters. In practice, however, every non-trivial
problem must contend with such parameters and the reference prior algorithm
must be generalized accordingly. In this paper we restrict our attention to
nuisance parameters for which partial information is available, which is often
the case in practice.
Depending on the type of partial information that is available, there are two
plausible ways one might choose to incorporate nuisance parameters $\phi$ into
the calculation of the reference priors for a parameter of interest
$\theta$~\cite{Sun1998}:
\begin{description}
\item[Method 1:] Assume that we are given a marginal prior $\pi(\phi)$ for the
nuisance parameters; compute the conditional reference
prior $\pi_{R}(\theta\,|\,\phi)$ for the interest parameter given
a fixed value of $\phi$; the full prior is then
$\pi(\theta, \phi) = \pi_{R}(\theta\,|\,\phi)\, \pi(\phi)$;
\item[Method 2:] Assume that we are given a conditional prior
$\pi(\phi\,|\,\theta)$ for the nuisance parameter given the
interest parameter; marginalize the probability model
$p(x|\theta, \phi)$ with respect to $\phi$ in order to obtain
$p(x|\theta)=\int p(x|\theta,\phi)\,\pi(\phi|\theta)\,d\phi$,
and compute the reference prior $\pi_{R}(\theta)$ for the
marginalized model; the full prior is then
$\pi(\theta,\phi) = \pi(\phi\,|\,\theta)\,\pi_{R}(\theta)$.
\end{description}
In many high energy physics measurements there are often sound reasons for assuming
that the nuisance parameter is independent of the parameter of interest.
Information about a detector energy scale, for example, is typically determined
separately from the measurement of interest, say of a particle mass, and
is therefore considered to be independent \emph{a priori} from one's information about
the particle's mass. When an experimenter is willing to make this assumption,
he or she can declare that $\pi(\phi\,|\,\theta) = \pi(\phi)$ and use Method 2.
When this assumption does not seem fully justified, and it is too difficult to
elicit the $\theta$ dependence of $\pi(\phi\,|\,\theta)$, then it will seem
preferable to use Method 1, which only requires knowledge of the marginal
prior $\pi(\phi)$. When one is unsure of which method to use, one should use
both, and treat the results as part of a test of robustness. An important
practical advantage of Method 1 is that the conditional reference prior is
computed once and for all, for a given model, and can be used with any
evidence-based prior for the nuisance parameters. In contrast, for Method 2
the reference prior must be computed anew every time the priors for the
nuisance parameters change. On the other hand, since Method 2 reduces the problem
to one involving a single parameter, the reference prior algorithm reduces to
Jeffreys' rule~\eqref{eq:Jpr}, which is typically easier to implement.
In the next section we introduce the basic model studied in this paper, and
follow with the application of Methods~1 and~2 to that model.
\subsection{The Single-Count Model}
A very common model for high energy physics measurements is the following.
A number of events $N$ is observed by some apparatus, and it is assumed that
$N$ is Poisson distributed with mean count $\epsilon\,\sigma+\mu$, where
$\sigma$ is the rate of a physics signal process, typically the cross section,
which we detect with an effective integrated luminosity $\epsilon$ --- that is,
the integrated luminosity scaled by the signal efficiency, and $\mu$ is a
background contamination. Thus, $\sigma$ is the parameter of interest, whereas
$\epsilon$ and $\mu$ are nuisance parameters for which we usually have partial
information. For physical reasons none of these three parameters can be
negative. We write the likelihood for this model as
\begin{equation}
p(n|\sigma,\epsilon,\mu)\;=\; \frac{(\epsilon\sigma+\mu)^{n}}{n!}\;
e^{-\epsilon\sigma-\mu}
\quad\textrm{with}\;\; 0\le \sigma <\infty \;\;
\textrm{and}\;\; 0<\epsilon, \mu <\infty.
\label{eq:scm}
\end{equation}
Information about $\epsilon$ and $\mu$ usually comes from a variety of sources,
such as auxiliary measurements, Monte Carlo simulations, theoretical
calculations, and evidence-based beliefs (for example, some sources of
background contributing to $\mu$ may be deemed small enough to ignore, and some
physics effects on $\epsilon$, such as gluon radiation, may be believed to be
well enough reproduced by the simulation to be reliable ``within a factor of
2''). It is therefore natural to represent that information by an evidence-based
prior. Here we will assume that $\epsilon$ and $\mu$ are independent of
$\sigma$ and that their prior factorizes as a product of two gamma densities:
\begin{equation}
\pi(\epsilon,\mu\,|\,\sigma)\;=\;\pi(\epsilon,\mu)\;=\;
\frac{a(a\epsilon)^{x-1/2}\,e^{-a\epsilon}}{\Gamma(x+1/2)}\;
\frac{b(b\mu)^{y-1/2}\,e^{-b\mu}}{\Gamma(y+1/2)},
\label{eq:emuprior}
\end{equation}
where $a$, $b$, $x$, and $y$ are known constants, related to the means
$\bar{\epsilon}$, $\bar{\mu}$ and coefficients of variation $\delta\epsilon$,
$\delta\mu$ by:
\begin{equation}
\bar{\epsilon} = \frac{x+\frac{1}{2}}{a},\quad
\delta\epsilon = \frac{1}{\sqrt{x+\frac{1}{2}}},\quad
\bar{\mu} = \frac{y+\frac{1}{2}}{b},\quad
\delta\mu = \frac{1}{\sqrt{y+\frac{1}{2}}}.
\label{eq:PriorParameters}
\end{equation}
The built-in assumption that $\epsilon$ and $\mu$ are uncorrelated is clearly
an approximation, since they share a dependence on the integrated luminosity,
which is itself uncertain.
There are two ways of interpreting this prior. The first one is appropriate
when information about $\epsilon$ and $\mu$ comes from one or more
non-experimental sources, such as Monte Carlo studies and theoretical
calculations, and takes the form of a central value plus an uncertainty. Since
the $\epsilon$ and $\mu$ components of the prior are each modeled by a
two-parameter density, one can fix the shape of this density in each case by
matching its mean with the central value of the corresponding measurement and
its standard deviation with the uncertainty. It will then be necessary to
check the robustness of the final analysis results to reasonable changes in this
procedure. For example, one may want to replace the gamma distribution by a
log-normal or truncated Gaussian one, and the mean by the mode or median.
The second interpretation of prior~\eqref{eq:emuprior} follows from the analysis
of two independent, auxiliary Poisson measurements, in which the observed number
of events is $x$ for the effective luminosity and $y$ for the background. The
expected numbers of events in these auxiliary measurements are $a\epsilon$ and
$b\mu$, respectively. For a Poisson likelihood with mean $a\epsilon$ the
reference prior coincides with Jeffreys' prior and is proportional to
$1/\sqrt{\epsilon}$. Given a measurement $x$, the posterior will then be a
gamma distribution with shape parameter $x+1/2$ and scale parameter $1/a$. A
similar result holds for the background measurement. In this manner the
prior~\eqref{eq:emuprior} is obtained as a joint reference posterior from two
auxiliary measurements.
The problem we are interested in is finding a prior for $\sigma$, about which
either little is known or one wishes to act as if this is so.
\subsubsection{Application of Method 1 to the Single-Count Model}
\label{Method1SingleCount}
This section serves two purposes: to illustrate the analytical algorithm for
computing reference priors and to apply Method~1 to model~\eqref{eq:scm}.
In Method 1~\cite{Sun1998}, we find first the conditional reference prior
$\pi_{R}(\sigma\,|\,\epsilon,\mu)$ and then multiply by the evidence-based prior
$\pi(\epsilon,\mu)$ to construct the full prior $\pi(\sigma,\epsilon,\mu)$. As
will be illustrated in Sec.~\ref{ValidationStudies}, the single-count model
is regular enough to warrant using Jeffreys' rule in the first step of the
calculation of $\pi_{R}(\sigma\,|\,\epsilon,\mu)$. We therefore apply
Eq.~\eqref{eq:Jpr} to the $\sigma$ dependence of the likelihood~\eqref{eq:scm},
while holding $\epsilon$ and $\mu$ constant; this yields:
\begin{equation}
\pi_{J}(\sigma\,|\,\epsilon,\mu)
\;\propto\; \sqrt{\mathbb{E}\left[-\frac{\partial^{2}}{\partial\sigma^{2}}
\ln p(n|\sigma,\epsilon,\mu)\right]}
\;\propto\; \frac{\epsilon}{\sqrt{\epsilon\,\sigma+\mu}}.
\label{eq:piJ}
\end{equation}
This prior is clearly improper with respect to $\sigma$ and is therefore only
defined up to a proportionality constant. However, this constant could very
well depend on $\epsilon$ and $\mu$, since we kept these parameters fixed in
the calculation. It is important to obtain this dependence correctly, as
examples have shown that otherwise inconsistent Bayes estimators may result.
Reference~\cite{Sun1998} proposes a compact subset normalization procedure.
One starts by choosing a nested sequence $\Theta_{1}\subset\Theta_{2}\subset
\cdots$ of compact subsets of the parameter space $\Theta = \{(\sigma,\epsilon,
\mu)\}$, such that $\cup_{\ell}\Theta_{\ell} = \Theta$ and the integral
$K_{\ell}(\epsilon,\mu)$ of $\pi_{J}(\sigma\,|\,\epsilon,\mu)$ over
$\Omega_{\ell}\equiv\{\sigma: (\sigma,\epsilon,\mu)\in\Theta_{\ell}\}$ is
finite. The conditional reference prior for $\sigma$ on $\Omega_{\ell}$ is then
\begin{equation}
\pi_{R,\ell}(\sigma\,|\,\epsilon,\mu)
\;=\; \frac{\pi_{J}(\sigma\,|\,\epsilon,\mu)}{K_{\ell}(\epsilon,\mu)}.
\label{eq:piRell}
\end{equation}
To obtain the conditional reference prior on the whole parameter space, one
chooses a fixed point $(\sigma_{0},\epsilon_{0},\mu_{0})$ within that space
and takes the limit of the ratio
\begin{equation}
\pi_{R}(\sigma\,|\,\epsilon,\mu)\;\propto\; \lim_{\ell\rightarrow\infty}
\frac{\pi_{R,\ell}(\sigma\,|\,\epsilon,\mu)}
{\pi_{R,\ell}(\sigma_{0}\,|\,\epsilon_{0},\mu_{0})}.
\label{eq:piR}
\end{equation}
By taking the limit in this ratio form, one avoids problems arising from
$K_{\ell}(\epsilon,\mu)$ becoming infinite as $\ell\rightarrow\infty$.
The theory of reference priors currently does not provide guidelines for
choosing the compact sets $\Theta_{\ell}$, other than to require that the
resulting posterior be proper. In most cases this choice makes no difference
and one is free to base the choice of compact sets on considerations of
simplicity and convenience. However, we have found that some care is required
with the single-count model. Indeed, suppose we make the plausible choice
\begin{equation}
\Theta_{\ell} \;=\; \Bigl\{(\sigma,\epsilon,\mu):\; \sigma\in[0,u_{\ell}],\;
\epsilon\in[0,v_{\ell}],\; \mu\in[0,w_{\ell}]\Bigr\},
\label{eq:cset1}
\end{equation}
where $\{u_{\ell}\}$, $\{v_{\ell}\}$, and $\{w_{\ell}\}$ are increasing
sequences of positive constants. If we use these sets in applying
Eqs.~\eqref{eq:piRell} and~\eqref{eq:piR} to the prior~\eqref{eq:piJ},
we obtain:
\begin{equation}
\pi_{R}(\sigma\,|\,\epsilon,\mu)
\;\propto\;\sqrt{\frac{\epsilon}{\epsilon\,\sigma+\mu}}.
\label{eq:piR1}
\end{equation}
Although this prior is still improper with respect to $\sigma$, its dependence
on $\epsilon$ is different from that of the conditional Jeffreys' prior,
Eq.~\eqref{eq:piJ}. This demonstrates the potential importance of the
compact subset normalization. The prior in Eq.~\eqref{eq:piR1} has a serious
problem however. Suppose that the $\epsilon$ marginal of our evidence-based
prior for $\epsilon$ and $\mu$ is $\exp(-\epsilon)/\sqrt{\pi\epsilon}$. It is
then easy to verify that the resulting posterior is improper, since its
$\epsilon$ marginal has the non-integrable form $\exp(-\epsilon)/\epsilon$.
The cause of this problem is the choice of compact sets~\eqref{eq:cset1}.
Fortunately it is not difficult to find a sequence of compact sets that will
provide a proper posterior. Indeed, the $\sigma$ dependence of the
prior~\eqref{eq:piJ} suggests that the compact sets should be based on
the parametrization $(\epsilon\sigma,\epsilon,\mu)$ rather than
$(\sigma,\epsilon,\mu)$~\cite{BergerPC2008}. We therefore set:
\begin{equation}
\Theta_{\ell} \;=\; \Bigl\{(\sigma,\epsilon,\mu):\;
\sigma\in[0,u_{\ell}/\epsilon],\;\epsilon\in[1/v_{\ell},v_{\ell}],\;
\mu\in[0,w_{\ell}]\Bigr\},
\label{eq:NCS2}
\end{equation}
where $u_{\ell}$, $v_{\ell}$, and $w_{\ell}$ are as before. Again using
Eqs.~\eqref{eq:piJ}, \eqref{eq:piRell}, and~\eqref{eq:piR}, we now find:
\begin{equation}
\pi_{R1}(\sigma\,|\,\epsilon,\mu)\;\propto\;
\frac{\epsilon}{\sqrt{\epsilon\sigma+\mu}},
\label{eq:piR1b}
\end{equation}
which is identical to Jeffreys' prior for this problem and yields
well-behaved posteriors. For future use, the subscript $R1$ on the left-hand
side indicates that this reference prior was obtained with Method~1.
We now have all the ingredients needed to calculate the marginal reference
posterior $\pi_{R1}(\sigma\,|\,n)$ for the cross section $\sigma$: the
likelihood~\eqref{eq:scm}, the marginal nuisance prior~\eqref{eq:emuprior},
and the conditional reference prior~\eqref{eq:piR1b}. For calculating
posterior summaries in terms of intervals and upper limits it is convenient
to express the result as a tail probability:
\begin{equation}
\int_{\sigma}^{\infty}\!\!\pi_{R1}(\tau\,|\,n)\;d\tau\;=\;
\int_{\!\frac{\sigma}{a+\sigma}}^{1}
\frac{u^{n+y}\;(1-u)^{x-\frac{1}{2}}}{B\bigl(n+y+1,x+\frac{1}{2}\bigr)}\;
\frac{B_{\frac{b}{b+1}(1+\frac{u-1}{u}\frac{\sigma}{a})}
\bigl(y+\frac{1}{2},n+\frac{1}{2}\bigr)}
{B_{\frac{b}{b+1}}\bigl(y+\frac{1}{2},n+\frac{1}{2}\bigr)}
\;du
\end{equation}
where
\begin{equation}
B_{z}(u,v)\;\equiv\;\int_{0}^{z}\! t^{u-1}\;(1-t)^{v-1}\;dt
\end{equation}
is the incomplete beta function, and $B(u,v)\equiv B_{1}(u,v)\;=\;
\Gamma(u)\Gamma(v)/\Gamma(u+v)$.
\subsubsection{Application of Method 2 to the Single-Count Model}
In contrast with Method~1, Method~2 requires from the start that we specify the
evidence-based prior for the effective integrated luminosity $\epsilon$ and the
background contamination~$\mu$. Furthermore, this specification must be done
conditionally on the signal rate $\sigma$. As mentioned earlier, we will use
expression~\eqref{eq:emuprior} for this prior.
The next step in the application of Method~2 is to marginalize the probability
model~\eqref{eq:scm} with respect to $\epsilon$ and $\mu$:
\begin{align}
p(n\,|\,\sigma)
&\;=\;\iint p(n\,|\,\sigma,\epsilon,\mu)\;
\pi(\epsilon,\mu\,|\,\sigma)\;d\epsilon\,d\mu,\nonumber\\[2mm]
&\;=\;\iint \frac{(\epsilon\sigma+\mu)^{n}}{n!}\;
e^{-\epsilon\sigma-\mu}\;
\frac{a(a\epsilon)^{x-\frac{1}{2}}}{\Gamma(x+\frac{1}{2})}\;
e^{-a\epsilon}\;
\frac{b(b\mu)^{y-\frac{1}{2}}}{\Gamma(y+\frac{1}{2})}\;
e^{-b\mu}\;d\epsilon\,d\mu,\nonumber\\[2mm]
&\;=\;\left[\frac{a}{a+\sigma}\right]^{x+\frac{1}{2}}\;
\left[\frac{b}{b+1}\right]^{y+\frac{1}{2}}\; S_{n}^{0}(\sigma),
\label{eq:MarginalModel}
\end{align}
where
\begin{equation}
S_{n}^{m}(\sigma)\;\equiv\;
\sum_{k=0}^{n} k^{m}\;
\binom{k+x-\frac{1}{2}}{k}\;\binom{n-k+y-\frac{1}{2}}{n-k}\;
\left[\frac{1}{b+1}\right]^{n-k}\;
\left[\frac{\sigma}{a+\sigma}\right]^{k},
\end{equation}
and the binomial coefficients are expressed in terms of gamma functions to
accomodate noninteger values of their arguments. Finally, the reference prior
algorithm must be applied to the marginalized model $p(n\,|\,\sigma)$. As in
the case of Method 1, the conditions for applying Jeffreys' rule are satisfied
here; we therefore obtain:
\begin{equation}
\pi_{R2}(\sigma)\;\propto\;\sqrt{\sum_{n=0}^{\infty}
\frac{\bigl[(x+\frac{1}{2})\,S_{n}^{0}(\sigma)\,-\,
\frac{a}{\sigma}\,S_{n}^{1}(\sigma)\bigr]^{2}}
{(a+\sigma)^{x+5/2}\;S_{n}^{0}(\sigma)}}.
\label{eq:piR2}
\end{equation}
We will use the notation $\pi_{R2}(\sigma)$ to refer to the marginal reference
prior for $\sigma$ obtained with Method~2. Note that the compact subset argument
invoked in the construction of the Method~1 reference prior is not needed here
because all the parameters other than $\sigma$ have already been eliminated by
marginalization.
For Method~2 the marginal reference posterior for $\sigma$ is proportional to
the product of the marginal data probability distribution~\eqref{eq:MarginalModel}
and the marginal reference prior~\eqref{eq:piR2}:
\begin{equation}
\pi_{R2}(\sigma\,|\,n)\;\propto\;p(n\,|\,\sigma)\;\pi_{R2}(\sigma).
\end{equation}
The normalization of $\pi_{R2}(\sigma\,|\,n)$ must be obtained numerically.
\subsection{The Multiple-Count Model}
An important generalization of the single-count model is obtained by considering
$M$ replications of the latter; the likelihood is:
\begin{equation}
p(\vec{n}\,|\,\sigma,\vec{\epsilon},\vec{\mu})\;=\;
\prod_{i=1}^{M}\frac{(\epsilon_{i}\sigma+\mu_{i})^{n_{i}}}{n_{i}!}\,
e^{-\epsilon_{i}\sigma-\mu_{i}}.
\end{equation}
To obtain the Method~1 reference prior for this model, we first calculate
Jeffreys' prior for $\sigma$, while keeping $\vec{\epsilon}$ and $\vec{\mu}$
fixed:
\begin{equation}
\pi_{J}(\sigma\,|\,\vec{\epsilon},\vec{\mu})\;\propto\;
\sqrt{\sum_{i=1}^{M}\frac{\epsilon_{i}^{2}}{\epsilon_{i}\,\sigma+\mu_{i}}}.
\end{equation}
This prior is improper, requiring us to apply the compact subset normalization
described in Sec.~\ref{Method1SingleCount}. Using a straightforward
generalization of the nested compact sets of Eq.~\eqref{eq:NCS2}, we find that
the correct reference prior is identical to Jeffreys' prior.
In order to apply Method~2, we need to specify a proper conditional prior for
the $\mu_{i}$ and $\epsilon_{i}$ given $\sigma$. Neglecting correlations, we
set:
\begin{equation}
\pi(\vec{\epsilon},\vec{\mu}|\sigma)\;=\;
\prod_{i=1}^{M} \frac{a_{i}(a_{i}\epsilon_{i})^{x_{i}-1/2}\,
e^{-a_{i}\epsilon_{i}}}{\Gamma(x_{i}+1/2)}\;
\frac{b_{i}(b_{i}\mu_{i})^{y_{i}-1/2}\,
e^{-b_{i}\mu_{i}}}{\Gamma(y_{i}+1/2)}.
\end{equation}
The marginalized data probability distribution $p(\vec{n}|\sigma)$ is then a
product of expressions of the form~\eqref{eq:MarginalModel}, one for each count
$i$.
Here we no longer attempt to obtain analytical expressions for the Method~1
and~2 reference posteriors. Instead, we use the numerical algorithms described
below.
\subsection{Numerical Algorithms}
\label{NumericalAlgorithm}
In this section we describe numerical algorithms that can be used to compute
Method~1 or~2 reference posteriors for the single- and multiple-count Poisson
likelihoods discussed in the previous sections.
For Method~1 the algorithm starts by generating $(\sigma,\vec{\epsilon},
\vec{\mu})$ triplets from the ``flat-prior posterior'', i.e. the posterior
obtained by setting $\pi(\sigma\,|\,\vec{\epsilon},\vec{\mu})=1$ (line~3 in the
pseudo-code below); the correct reference prior $\pi(\sigma\,|\,\vec{\epsilon},
\vec{\mu})$ is then computed at lines~4--7 and is used at line~9 to weight the
generated $\sigma$ values so as to produce the reference posterior:
{\small
\begin{tabbing}
~~~~~~~~\=~~~~~~~~\=~~~~~~~~\=~~~~~~~~\=\kill
\>{\scriptsize 1}\> Set $\vec{n}_{o}$ to the array of observed event numbers.\\
\>{\scriptsize 2}\> For $i=1,\ldots,I$:\\
\>{\scriptsize 3}\>\>Generate $(\sigma_{i},\vec{\epsilon}_{i},\vec{\mu}_{i}) \sim
p(\vec{n}_{o}\,|\,\sigma,\vec{\epsilon},\vec{\mu})\,
\pi(\vec{\epsilon},\vec{\mu})$. \\
\>{\scriptsize 4}\>\> For $j=1,\ldots,J$:\\
\>{\scriptsize 5}\>\>\>Generate $\vec{n}_{j}\sim p(\vec{n}\,|\,\sigma_{i},
\vec{\epsilon}_{i},\vec{\mu}_{i})$.\\
\>{\scriptsize 6}\>\>\>Calculate $d^{2}[-\ln p(\vec{n}_{j}\,|\,\sigma_{i},
\vec{\epsilon}_{i},\vec{\mu}_{i})]/d\sigma_{i}^{2}$
by numerical differentiation.\\
\>{\scriptsize 7}\>\>Average the $J$ values of
$d^{2}[-\ln p(\vec{n}\,|\,\sigma_{i},\vec{\epsilon}_{i},
\vec{\mu}_{i})]/d\sigma_{i}^{2}$ obtained\\
\>\>\> at line 6, and take the square root. This yields a
numerical\\
\>\>\> approximation to the conditional Jeffreys' prior
$\pi_{J}(\sigma_{i}\,|\,\vec{\epsilon}_{i},\vec{\mu}_{i})$.\\
\>{\scriptsize 8}\> Histogram the $\sigma_{i}$ values generated at line 3,
weighting them by\\
\>\> $\pi_{J}(\sigma_{i}\,|\,\vec{\epsilon}_{i},\vec{\mu}_{i})/
p(\vec{n}_{o}\,|\,\sigma_{i},\vec{\epsilon}_{i},\vec{\mu}_{i})$.
This yields $\pi_{R1}(\sigma)$, the $\sigma$-marginal
prior.\\
\>{\scriptsize 9}\> Histogram the $\sigma_{i}$ values generated at line 3,
weighting them by \\
\>\> $\pi_{J}(\sigma_{i}\,|\,\vec{\epsilon}_{i},\vec{\mu}_{i})$.
This yields $\pi_{R1}(\sigma\,|\,\vec{n}_{o})$, the
$\sigma$-marginal posterior.\\
\end{tabbing}
}
Although not required for the calculation of the reference posterior, an
approximation to the reference prior is provided at line~8. By construction
this approximation is only reliable for $\sigma$ values in the bulk of the
flat-prior posterior. The generation step at line~3 is done via a Markov chain
Monte Carlo procedure~\cite{BAT}. The particular choice of sampling
distribution for the generated $(\sigma,\vec{\epsilon},\vec{\mu})$ triplets is
motivated by the desire to obtain weights with reasonably small variance at
steps 8 and 9. However, the flat-prior posterior $p(\vec{n}_{0}\,|\,\sigma,
\vec{\epsilon},\vec{\mu})\,\pi(\vec{\epsilon},\vec{\mu})$ is not always proper
with respect to $(\sigma,\vec{\epsilon},\vec{\mu})$. When $M=1$ (single-count
model), it is improper if $x\le 1/2$. Propriety can then be restored by
multiplying the flat-prior posterior by $\epsilon$ and correspondingly
adjusting the weights at steps 8 and 9. Another feature of the above algorithm
is that it does not implement the compact subset normalization. In the cases
that we examined, this procedure made no difference, but this may not be true
for more general problems than those our code seeks to solve. Unfortunately
the current lack of guidelines in the choice of compact sets limits our ability
to address this issue in the code.
The algorithm for Method~2 has a simpler structure, since all it does is
apply Jeffreys' rule to a marginalized likelihood $p(\vec{n}_{o}\,|\,\sigma)$
provided by the user. The calculation does not require random sampling of the
parameters and is done at fixed $\sigma$ values. For a given $\sigma$, the
reference prior $\pi_{R2}(\sigma)$ is obtained by Monte Carlo averaging, over
an ensemble of vectors $\vec{n}$ generated from $p(\vec{n}\,|\,\sigma)$, of
an accurate numerical approximation of the second derivative of the negative
log-likelihood~\cite{2derivative}. As already pointed out, Method 2 does not
require a compact subset normalization procedure. The reference posterior is
thus proportional to the product of $p(\vec{n}_{o}\,|\,\sigma)$ and
$\pi_{R2}(\sigma)$, and the normalization with respect to $\sigma$ must be
determined numerically.
\section{Validation Studies}
\label{ValidationStudies}
We have performed a number of studies to validate inferences from the
single-count model, using both the numerical algorithms described in
Sec.~\ref{NumericalAlgorithm} and analytical expressions we obtained for the
marginal Method-1 and~2 posteriors for $\sigma$. To recapitulate, we have two
reference priors for this model:
\begin{align}
\pi_{R1}(\sigma,\epsilon,\mu) &\;=\; \pi_{R1}(\sigma\,|\,\epsilon,\mu)\;
\pi(\epsilon,\mu)\\[2mm]
\pi_{R2}(\sigma,\epsilon,\mu) &\;=\; \pi_{R2}(\sigma)\;
\pi(\epsilon,\mu\,|\,\sigma),
\end{align}
and we have assumed that $\pi(\epsilon,\mu\,|\,\sigma)=\pi(\epsilon,\mu)$ at
Eq.~\eqref{eq:emuprior}. As explained in Sec.~\ref{InclPartInfo}, this
extra assumption affects only the definition of $\pi_{R2}$, which therefore
incorporates more information than $\pi_{R1}$. In the present section we study
and compare the properties of these two reference priors. To begin, we show
some example prior and posterior $\sigma$ marginals in
Fig.~\ref{fig:priorMethod12}.
\begin{figure}
\includegraphics[width=7cm]{prior12}
\hspace*{1cm}
\includegraphics[width=7cm]{posterior12}
\caption{Left: marginal Method-1 and 2 priors, normalized to 1 at $\sigma=1$.
Right: marginal Method-1 and 2 posteriors for 0, 1, and 4 observed events,
together with the posteriors obtained from a flat prior. The $\epsilon$ and
$\mu$ priors have a mean of 1 and a 20\% coefficient of variation
(corresponding to $x=y=24.5$ and $a=b=25$ in
Eq.~\protect\eqref{eq:PriorParameters}). Here and in subsequent plots, the
units of $\sigma$ are arbitrary but consistent with those of $\epsilon$;
e.g., if the latter is expressed in pb$^{-1}$, then $\sigma$ is given in pb so
that, like $\mu$, the product of $\epsilon$ and $\sigma$ is dimensionless.
\label{fig:priorMethod12}}
\end{figure}
As expected, posteriors corresponding to a small observed number of events
favor small cross sections, and posteriors derived from flat priors put less
weight on small cross sections than reference posteriors.
Our derivations of the two reference prior methods made use of Jeffreys' rule
\eqref{eq:Jpr}. As pointed out in Sec.~\ref{sec:refprior}, this approach
assumes that some regularity conditions are satisfied, such that the resulting
posterior is asymptotically normal. We now wish to verify this assumption with
a graphical example. If one adopts the objective Bayesian view that the
parameters $\sigma$, $\epsilon$, and $\mu$ have true values, then the asymptotic
limit can be defined as the result of a large number $N_{R}$ of replications of
the measurement, in the limit where that number goes to infinity. For the case
where each measurement replication $i$ consists of a number of events $n_{i}$
drawn from a probability mass function $f(n\,|\,\sigma,\epsilon,\mu)$, the
reference posterior has the form:
\begin{equation}
\pi_{R}(\sigma,\epsilon,\mu\,|\,n_{1},n_{2},\ldots,n_{N_{R}}) \;\propto\;
\pi_{R}(\sigma,\epsilon,\mu)\,
\prod_{i=1}^{N_{R}}f(n_{i}\,|\,\sigma,\epsilon,\mu),
\label{eq:ReplicatedExp}
\end{equation}
and the reference prior $\pi_{R}(\sigma,\epsilon,\mu)$ is calculated from the
combined likelihood for the $N_{R}$ measurements; it can also be calculated
from a single one of these likelihood functions, since it follows from their
constructive definition~\eqref{eq:refformula} that reference priors are
independent of sample size. For Method 1 the prior is given by the product of
Eqs.~\eqref{eq:emuprior} and~\eqref{eq:piR1b}, and the likelihood component
$f(n\,|\,\sigma,\epsilon,\mu)$ by Eq.~\eqref{eq:scm}. Replicating the
measurement $N_{R}$ times is then equivalent to making a single measurement
with a Poisson likelihood whose mean is $N_{R}$ times the original mean, and
whose observation is the sum of the $N_{R}$ original observations $n_{i}$.
This property of Poisson measurements simplifies the calculations considerably.
For Method~2 the prior is given by Eq.~\eqref{eq:piR2} and the likelihood by
Eq.~\eqref{eq:MarginalModel}. In this case no simplification obtains when
considering multiple replications, and numerical calculations must use
explicitly the full product of likelihood functions.
\begin{figure}
\includegraphics[width=7cm]{m1qq}
\hspace*{1cm}
\includegraphics[width=7cm]{m2qq}
\caption{Q-Q plots of the Method-1 conditional reference posterior (left) and
the Method-2 marginal reference posterior (right) versus the standard normal
distribution. The reference posterior quantiles have been recentered (see
text). $N_{R}$ is the number of measurement replications and $n_{tot}$ is the
observed number of events summed over all replications.
\label{fig:m12qq}}
\end{figure}
Figure~\ref{fig:m12qq} illustrates the calculations with the help of so-called
Q-Q plots, where recentered quantiles from the reference posterior for $\sigma$
are plotted against standard normal quantiles. The posterior quantiles
$Q_{\pi}$ are recentered according to
$\tilde{Q}_{\pi}=(Q_{\pi}-\langle\sigma\rangle)/\Delta\sigma$,
where the posterior mean $\langle\sigma\rangle$ and standard deviation
$\Delta\sigma$ are numerically estimated. For Method~1 we set the true values
of $\sigma$, $\epsilon$, and $\mu$ to $1$. We then randomly generate a
sequence of 100 independent measurements from the probability mass
function~\eqref{eq:scm} and use the subsequences with $N_{R}=1$, $10$, and
$100$ to produce the curves in the left panel. For Method~2 we set the true
value of $\sigma$ to 1 and give the priors for $\epsilon$ and $\mu$ each a
mean of 1 and a coefficient of variation of 20\%. Measurements are then
generated from the probability mass function~\eqref{eq:MarginalModel} in order
to compute the curves in the right panel. Both panels clearly show that the
respective reference posteriors approach a Gaussian shape as the number of
measurement replications increases.
Given the almost negligible difference between Method-1 and 2 posteriors
exhibited in Fig.~\ref{fig:priorMethod12}, and the fact that our analytical
results for Method~1 are computationally more tractable than those for Method~2,
our considerations in the remainder of this section will focus exclusively on
Method~1.
Among the reference prior properties listed in Sec.~\ref{sec:RefPriors}, the
ones of generality, invariance, and coherence are true by construction. The
property of sampling consistency needs more elaboration however, since Bayesian
inferences do not generally coincide with exact frequentist ones, and a proper
evaluation requires first of all the specification of an ensemble of
experiments. A well-known property of Bayesian posterior intervals constructed
from a proper prior is that their coverage is exact when averaged over the
prior~\cite{Pratt1965}. This is an immediate consequence of the law of total
probability. Indeed, given a parameter $\theta$ with proper prior
$\pi(\theta)$, and a measurement $X$, the prior-averaged frequentist coverage
of a $1-\alpha$ Bayesian credibility interval $R(X)$ can be written as:
\begin{equation}
\mathbb{E}_{\pi}[\mathbb{P}(\theta\in R(X)\,|\,\theta)]
\,=\,\mathbb{P}(\theta\in R(X))
\,=\,\mathbb{E}_{m}[\mathbb{P}(\theta\in R(X)\,|\,X)]
\,=\,\mathbb{E}_{m}(1-\alpha)\,=\,1-\alpha,
\end{equation}
where the first expectation is over the prior $\pi(\theta)$ and the second one
over the marginal sampling distribution $m(x)=\int p(x|\theta)\,\pi(\theta)\:d\theta$.
When $\pi(\theta)$ is a reference prior, and especially when it is
improper, there is no natural metric over which the coverage can be averaged.
The only sensible approach in that case is to study the coverage pointwise,
i.e. as a function of the true value of $\theta$. Since the single- and
multiple-count models discussed in this paper combine an improper prior for the
parameter of interest $\sigma$ with proper priors for the nuisance parameters
$\vec{\epsilon}$ and $\vec{\mu}$, we will study interval coverage for a fixed
value of $\sigma$, but averaged over $\pi(\vec{\epsilon},\vec{\mu})$. Our
interest is in how this coverage evolves toward the asymptotic limit. As
before, we take this limit in the sense of an ever-increasing number $N_{R}$
of experiment replications. For a given value of $N_{R}$, the posterior is
formed as in Eq.~\eqref{eq:ReplicatedExp} and its coverage is computed.
\begin{figure}
\includegraphics[width=7cm]{m1_upper_95_cov_l}
\hspace*{1cm}
\includegraphics[width=7cm]{m1_central_68_cov_l}
\caption{Coverage probability of Method-1 posterior credibility upper limits
(left) and central intervals (right), as a function of the number of experiment
replications $N_{R}$. The solid lines indicate the nominal credibility.
\label{fig:m1coverage_xl}}
\end{figure}
Figure~\ref{fig:m1coverage_xl} shows the coverage of 95\% credibility upper
limits and 68\% credibility central intervals as a function of $N_{R}$. As
the latter increases, the coverage converges to the credibility, confirming
the sampling consistency of the method.
Finally, we examine the behavior of reference posterior upper limits on $\sigma$
as a function of the expected background $\bar{\mu}$ (defined in
Eq.~\eqref{eq:PriorParameters}) when the observed number of
events $n$ is small. For comparison, when $n=0$ and there are no uncertainties
on signal efficiency and background, frequentist upper limits decrease
linearly with background. From a Bayesian point of view this result is
surprising. Indeed, when zero events are observed the likelihood function
factorizes exactly into background and signal components, indicating that the
experiment {\em actually} performed can be analyzed as the combination of two
independent experiments, one to measure background and the other signal.
If, in addition, signal and background are {\em a priori} independent, then
posterior inferences about signal will be independent of background. In
particular, upper limits on $\sigma$ will be constant as a function of
$\bar{\mu}$, not linearly decreasing. The reference priors entangle signal and
background however, so that upper limits will not be exactly constant. The
$n=0$ case is illustrated in Fig.~\ref{fig:m1ul_meanbg} for two values of the
relative uncertainties on background and signal efficiency.
\begin{figure}
\includegraphics[width=7cm]{m1ul_meanbg_20}
\hspace*{1cm}
\includegraphics[width=7cm]{m1ul_meanbg_50}
\caption{Variation of the Method 1 reference posterior upper limit with mean
background for several values of the observed number of events $n$. The
relative uncertainty on the background and on the effective luminosity is 20\%
for the left plot and 50\% for the right one.
\label{fig:m1ul_meanbg}}
\end{figure}
For $n>0$ the likelihood function still factorizes approximately since
$(\mu+\epsilon\sigma)^{n}\approx \mu^{n} (1+n\epsilon\sigma/\mu)\approx\mu^{n}$
for $\mu\gg \epsilon\sigma, n$. Thus upper limits will flatten out at large
$\bar{\mu}$, as seen in Fig.~\ref{fig:m1ul_meanbg}. A comparison of the left
and right panels in that figure also shows that upper limits increase with the
uncertainty on background and signal efficiency, as expected.
\section{Measurement of the Single Top Quark Cross Section}
\label{SingleTop}
In this section, we demonstrate the computational feasibility of the methods
described above by applying them to the recent measurement of the single top
cross section by the D0 and CDF collaborations~\cite{D0singletop,CDFsingletop}.
Both collaborations use the same form of likelihood function --- a product of
Poisson distributions over multiple bins of a multivariate discriminant, the
same form of evidence-based priors, namely truncated Gaussians, and flat priors
for the cross section~\cite{MultipleBinFlatPrior}. As a realistic example,
we construct the reference prior for the cross section using one of the data
channels considered by D0.
D0 partitioned their data into 24 channels, defined by lepton flavor (electron
or muon), jet multiplicity (two, three, or four), number of $b$-tagged jets
(one or two), and two data collection periods. The discriminant distribution
is shown in Fig.~3 of Ref.~\cite{D0singletop}. Here we consider the electron,
two-jet, single-tag channel from one of the data taking periods. The
discriminant distribution contains about 500 counts spread over 50 bins,
with a maximum bin count of about 40.
\begin{figure}
\includegraphics[width=7cm]{D0singletop_prior}
\hspace*{1cm}
\includegraphics[width=7cm]{D0singletop_posterior}
\caption{(a) Prior densities computed using Method~1 (histogram) and Method~2
(dashed curve), and (b) corresponding posterior densities, for one of the
channels in the D0 single top measurement and using 2.3 fb$^{-1}$ of data. For
comparison we show the posterior density computed using a flat prior (dotted
curve).
\label{fig:d0results}}
\end{figure}
We model information about the effective integrated luminosity $\epsilon$ and
the background $\mu$ for each bin with the help of the gamma priors of
Eq.~\eqref{eq:emuprior}. These evidence-based priors describe the uncertainty
due to the finite statistics of the Monte Carlo simulations. We do not include
systematic uncertainties in this example. Figure~\ref{fig:d0results}(a) shows
a comparison of the reference prior for the cross section using Methods 1 (the
histogram) and 2 (the dashed curve). The jaggedness of the Method 1 prior
reflects the fluctuations due to the Markov chain Monte Carlo~\cite{BAT}
sampling of the parameters. The increased jaggedness at large $\sigma$ is due
to the fact that the numerical algorithm samples from the flat-prior posterior,
whose density rapidly decreases in this region. It is noteworthy that the
priors computed using the two methods are very similar for this particular
example. This is also reflected in the similarity of the posterior densities,
shown in Fig.~\ref{fig:d0results}(b). In principle fluctuations in the
calculated posterior can be made arbitrarily small by increasing the size of
the Monte Carlo sample. For reference, Fig.~\ref{fig:d0results}(b) also shows
the posterior density using a flat prior for the cross section. An obvious
conclusion can be drawn: when the dataset is large, here of order 500 events,
the precise form of the reference prior is not important. However, for small
datasets --- which is typical of searches for new or rare phenomena, one should
expect the form of the prior to matter. It is then important to use a prior
with provably useful properties, such as the ones enumerated in
Sec.~\ref{sec:RefPriors}.
\section{Discussion}
\label{FinalComments}
Our main purpose in this paper was to propose a set of formal priors with
properties that make them attractive for use in the analysis of high energy
physics data. Aside from the theoretical properties of invariance, coherence
and sampling consistency, the reference prior method has three important
practical advantages: (1) priors can be defined for almost any problem, regardless
of the complexity of the likelihood function and the number of nuisance and
interest parameters, (2) in contrast with flat priors, reference priors have so
far always yielded proper posteriors, and (3) reference priors are
computationally tractable, as shown by the single-top example.
Here we have limited our numerical investigations to the class of likelihood
functions that are derived from Poisson probability mass functions. For this
class the Method-1 reference prior agrees with Jeffreys' rule. For other
classes the compact subset normalization argument may introduce a difference.
A possible generalization of our treatment is to unbinned likelihoods. Since
our Method-1 and 2 numerical algorithms make no assumptions about the likelihood
function, they can be generalized to the unbinned case. However, the Method-1
algorithm does not implement the compact subset normalization and is therefore
only applicable to cases where this procedure makes no difference. Method~2
requires no compact subset normalization but makes an extra assumption about the
conditional nuisance prior.
For problems that involve a single continuous parameter or that can be reduced
to this case by a Method-2-type integration, Ref.~\cite{Bernardo2005}
proposes a numerical algorithm that is based directly on
Eq.~\eqref{eq:refformula} and is therefore very general. However we found
that this algorithm presents some difficulties for the complicated likelihood
functions used in high energy physics. One difficulty is the round-off error
in the product of large numbers of probability densities. Another difficulty
is the assessment of the convergence of the integrals in the formula, and of
the convergence of the finite-sample priors to the reference prior.
Another possible generalization is to problems with more than one parameter of
interest, as for example in the measurement of the individual single top
production cross sections in the $s$ and $t$ channels. For this situation the
reference prior algorithm requires one to sort the parameters of interest by
order of importance~\cite{Bernardo2005}, and the results may depend on this
ordering. A possible interpretation of this dependence is that it is a measure
of the robustness of the result to the choice of prior. This is an area that
requires more study.
We have developed a general-use software package that implements the methods
described in this paper and have released it to the Physics Statistics Code
Repository (phystat.org).
Finally, we note that the main ideas underlying the construction of reference
priors, namely generality, reparametrization invariance, coherence, and
sampling consistency, have motivated the development of methods for
summarizing reference posteriors via point estimates, intervals, and hypothesis
tests. This subfield of objective Bayesianism is known as reference
analysis~\cite{Bernardo2005,Demortier2005}.
\begin{acknowledgments}
We thank the D0 Collaboration for its support of this work and for granting us
permission to use a channel of its single top data. We also thank the members
of the CMS Statistics Committee for many useful discussions. This work was
supported in part by the U.S. Department of Energy under grant
Nos. DE-FG02-91ER40651, DE-FG02-04ER41305, and DE-FG02-95ER40896.
\end{acknowledgments}
|
\section{Introduction}
The spontaneous breaking of chiral symmetry (S$\chi$SB) is one of
the most important phenomena of hadron physics. It defines the properties
of all the light mesons and baryons. Using the general idea of chiral
symmetry, it was proposed in~\cite{Gasser:1983yg} to use a phenomenological
lagrangian, which has a form of the infinite series in the pion momenta
$p^{2}$ and mass $M_{\pi}^{2}$. The low-energy constants of the
series expansion (LEC's) are the free parameters which encode the
low-energy physics in a model-independent way. Up to now they were
extracted phenomenologically from the experimental data, or from the
lattice calculations ((MILC, ETM, JLQCD, RBC/UKQCD, PACS-CS)\cite{Bowman:2004xi,Chu:vi,DeGrand:2001tm,Negele:1998ev}
within so-called Chiral Perturbation Theory (ChPT).
One of the low-energy constants $l_{7}$ is particularly interesting
since it encodes the {}``pure QCD'' part of the $SU(2)$ isospin
symmetry breaking (\emph{i.e.} part which is due to $u-$ and $d-$quark
current mass difference, $m_{u}-m_{d}$).~ For example, the QCD part
of the pion mass difference $m_{\pi^{+}}^{2}-m_{\pi^{0}}^{2}$ has
a form~\cite{Gasser:1983yg}
\begin{equation}
\left(m_{\pi^{+}}^{2}-m_{\pi^{0}}^{2}\right)_{QCD}=\frac{2B^{2}}{F^{2}}l_{7}\left(m_{u}-m_{d}\right)^{2},\label{eq:PiMassSplit_QCD}
\end{equation}
where $B$ and $F$ are the leading order parameters in the chiral
lagrangian, and $m_{u},m_{d}$ are the current quark masses. While
experimentally the isospin breaking effects are known to a very high
precision, separation of these effects on the {}``pure QCD'' and
electromagnetic parts has ambiguities and has been a subject of intensive
debates~\cite{Bijnens:1993ae,Bijnens:1996kk,Rathske,Rusetsky:2009ic,Gasser:1983yg}.
From phenomenology the constant $l_{7}$ is known only with an {}``order
of magnitude'' estimate~\cite{Gasser:1983yg},\begin{equation}
l_{7}\sim5\times10^{-3}.\label{eq:L7-phenom-value}\end{equation}
For this reason it makes sense to estimate this contribution in the
framework of a reliable model.
QCD instanton vacuum model, often refered to as the instanton liquid
model, provides a very natural nonperturbative explanation of the
S$\chi$SB~\cite{'tHooft:1973jz,'tHooft:1976fv,Diakonov:1983hh,Diakonov:1985eg,Diakonov:1995qy,Diakonov:2002fq,Kim:2004hd,Kim:2005jc,LangackerPagels73,Lee:sm,Musakhanov:1998wp,Musakhanov:2002xa}.
It provides a consistent framework for description of the pions and
thus may be used for evaluation of the low energy constants. Due to
instanton-induced nonlinear interaction all the quark and meson loop
integrals are regularized by the natural scale $\mu\sim\rho^{-1}\sim600$~MeV
in Pauli-Villars scheme~\cite{Diakonov:1985eg}, where $\rho$ is
the average size of the instanton. This means that all the scale-dependent
quantities, such as the quark condensate $\left\langle \bar{q}q\right\rangle \equiv\left\langle \bar{u}u\right\rangle +\left\langle \bar{d}d\right\rangle $
and the difference~$\delta\left\langle \bar{q}q\right\rangle \equiv\left\langle \bar{u}u\right\rangle -\left\langle \bar{d}d\right\rangle $,
are given at the scale~$\mu$. Remarkably, the constant $l_{7}$
does not depend on the scale $\mu$. Recently \cite{Goeke:2007bj}
it has been shown that this approach is able to give results consistent
with phenomenological and lattice estimates for the constants $\bar{l}_{3},\bar{l}_{4}$,
providing current quark mass dependencies of the pion mass $m_{\pi}$
and pion decay constant $F_{\pi}$.
In this paper we would like to apply the instanton vacuum model for
the evaluation of the constant $l_{7}$.
We extract the constant $l_{7}$ from the correlator $\langle P^{3}(x)P^{0}(0)\rangle$
using the relation~\cite{Gasser:1983yg} \begin{equation}
\int d^{4}x\, e^{iqx}\langle P^{3}(x)P^{0}(0)\rangle=\frac{G_{\pi}\tilde{G}_{\pi}}{m_{\pi}^{2}-q^{2}}+\mathcal{O}\left(q^{2}\right)=\frac{8B^{3}\left(m_{u}-m_{d}\right)}{q^{2}-m_{\pi}^{2}}l_{7}+\mathcal{O}\left(m,q^{2}\right).\label{eq:P3P0Definition-1}\end{equation}
From the Eqn.~(\ref{eq:P3P0Definition-1}) we may see that evaluations
may be done in the limit $m\equiv\frac{m_{u}+m_{d}}{2}\to0,$ and
make only expansion over \[
\delta m\equiv\left(m_{u}-m_{d}\right).\]
The paper is organized as follows. In Section~\ref{sec:GapEqn} we
discuss the general framework used for evaluation and write out the
next-to-leading order (NLO) gap equation in the presence of the current
mass split $\delta m$, which are needed for evaluation of the dynamical
mass split $\delta M\equiv M_{u}-M_{d}.$ In Section~\ref{sec:Propagators}
we write out explicit expressions for the quark and meson propagators.
In Section~\ref{sec:QuarkCondensate} we evaluate the effects of
the mass split $\delta m$ on the quark condensate, $\delta\langle\bar{q}q\rangle=\langle\bar{q}q\rangle_{u}-\langle\bar{q}q\rangle_{d}$
and extract the constant $h_{3}$. In Section~\ref{sec:P3P0Correlator}
we evaluate the correlator $\langle P^{3}(x)P^{0}(0)\rangle$ and
extract the constant $l_{7}$. In Section~\ref{sec:Conclusion} we
discuss obtained results, their uncertainty limits and draw conclusions.
\section{Instanton vacuum model}
The instanton vacuum model is based on the assumptions that the QCD
vacuum may be considered as a dilute gas of instantons and antiinstantons,
and the number of colors $N_{c}$ is asymptotically large, $N_{c}\to\infty$
(see the reviews \cite{Schafer:1996wv,Diakonov:2002fq}). While in
general the sizes and local density of the instanton gas may be arbitrary,
inter-instanton interaction stabilize these parameters. As it has
been discussed in~\cite{Goeke:2007bj}, the $1/N_{c}$-suppressed
corrections due to the finite size distribution are indeed quite small,
even for $N_{c}=3$. Phenomenological, variational and lattice estimates
lead to average instanton size $\rho\sim0.3\, fm$ and inter-instanton
distance $R\sim1\, fm$~\cite{Diakonov:1983hh}.
The partition function in the field of external scalar and pseudoscalar
currents~$s=\left(s_{0}+\vec{s}\vec{\tau}\right)$ and $p=\left(p_{0}+\vec{p}\vec{\tau}\right)$
has a form~\cite{Goeke:2007bj}
\begin{eqnarray}
Z_{N}[s_{0},\sigma,s_{0},\vec{p},\vec{s},p_{0}] & = & \int d\lambda\exp\left(-\Gamma_{eff}[s_{0},\lambda,\sigma,s_{0},\vec{p},\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},\vec{u}]\right),\label{eq:Zma}\\
\Gamma_{eff} & = & S+\Gamma_{eff}^{mes},\label{eq:Z0}\\
S & = & \frac{N}{V}\,\ln\lambda+2\int d^{4}x\sum\Phi_{i}^{2}(x)-Tr\ln\left(\frac{\hat{p}+is_{0}+\vec{s}\cdot\vec{\tau}+p_{0}\gamma_{5}+i\vec{p}\cdot\vec{\tau}\gamma_{5}+i\, c\, F\Phi\cdot\Gamma F}{\hat{p}+is_{0}+\vec{s}\cdot\vec{\tau}+p_{0}\gamma_{5}+i\vec{p}\cdot\vec{\tau}\gamma_{5}}\right),\label{eq:Z1}\end{eqnarray}
The nonlocal formfactors $F(p)$ in the meson-quark interaction vertices
come from the instanton-induced nonlocal interactions. Together with
the factor $\left(\hat{p}+is_{0}+\vec{s}\cdot\vec{\tau}+p_{0}\gamma_{5}+i\vec{p}\cdot\vec{\tau}\gamma_{5}\right)$
in denominator, which subtracts the divergent high-frequency modes,
they guarantee finite results for all the observables in the instanton
vacuum model. As it was discussed in~\cite{Diakonov:1983hh,Diakonov:1985eg},
the divergent high-ferquency modes are responsible for renormalization
of the parameters of the model. In what follows, we will fix them
at the scale $\mu\sim\rho^{-1}\sim600$~MeV in the Pauli-Villars
scheme~\cite{Diakonov:1983hh,Diakonov:1985eg}.
The meson-loop correction~$\Gamma_{eff}^{mes}$ to the effective
action is given as \begin{align}
\Gamma_{eff}^{mes}[m,\lambda,\sigma] & =\frac{1}{2}Tr\ln\left(4\delta_{ij}+\frac{1}{\sigma^{2}}Tr\left(\frac{c\left(\lambda\right)F^{2}(p)}{\hat{p}+is_{0}+\vec{s}\cdot\vec{\tau}+p_{0}\gamma_{5}+i\vec{p}\cdot\vec{\tau}\gamma_{5}+i\, c\, F\Phi\cdot\Gamma F}\Gamma_{i}\times\right.\right.\label{eq:Z2}\\
& \left.\left.\frac{c\left(\lambda\right)F^{2}(p)}{\hat{p}+is_{0}+\vec{s}\cdot\vec{\tau}+p_{0}\gamma_{5}+i\vec{p}\cdot\vec{\tau}\gamma_{5}+i\, c\, F\Phi\cdot\Gamma F}\Gamma_{j}\right)\right),\nonumber \\
\Phi\cdot\Gamma= & \left(\sigma+i\gamma_{5}\vec{\tau}\vec{\phi}+i\vec{\tau}\vec{\sigma}+\gamma_{5}\eta\right),\end{align}
where $c(\lambda)=\frac{(2\pi\rho)^{2}\sqrt{\lambda}}{2g},$$g^{2}=\frac{(N_{c}^{2}-1)2N_{c}}{2N_{c}-1}$
is a color factor, $\Gamma=\{1,\gamma_{5},i\vec{\tau},i\vec{\tau}\gamma_{5}\}$
is a set of matrices corresponding to quantum numbers of mesons present
in the model, and we will use for the corresponding components of
the field $\Phi$ the notations $\Phi=\{\sigma,\eta,\vec{\sigma},\vec{\phi}\}$.
In contrast to NJL model, the variable $\lambda$ is a dynamical degree
of freedom but not the parameter of the lagrangian. The current masses
of the quarks come into play via constant external currents, viz.
\begin{eqnarray}
s_{0} & = & \frac{m_{u}+m_{d}}{2},\\
s_{3} & = & \frac{m_{u}-m_{d}}{2}.\end{eqnarray}
Notice that with respect to chiral transformations, the mesons may
be separated onto two independent doublets $\left(\sigma,\vec{\phi}\right)$
and $\left(\eta,\vec{\sigma}\right).$ The first doublet~$\left(\sigma,\vec{\phi}\right)$
corresponds to the pion field $U=\left(u_{0},\vec{u}\right)$ in the
notations of~\cite{Gasser:1983yg}, and the second doublet $\left(\eta,\vec{\sigma}\right)$
is an additional degree of freedom which is absent in the chiral lagrangian.
Now we are going to demonstrate explicitly on the example of the constant
$l_{7}$ that this additional degree of freedom $\left(\eta,\vec{\sigma}\right)$
gives an essential contribution to the constant $l_{7}$ . As usual,
the external currents $\left(s_{0},\,\vec{s},\, p_{0},\,\vec{p}\right)$
generate nonzero vacuum averages of the fields $\left\langle \vec{\sigma}\right\rangle =\vec{\sigma}_{v},$
$\left\langle \eta\right\rangle =\eta_{v}$ and $\left\langle \left(\sigma,\vec{\phi}\right)\right\rangle =U=\left(u_{0},\vec{u}\right)$.
Due to the chiral symmetry expansion of the $\Gamma_{eff}$ yields
the general structure \begin{eqnarray}
& & \Gamma_{eff}[\lambda,\sigma,\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},u_{i}]=\Gamma_{eff}[m,\lambda,\sigma,\vec{s}=0,p_{0}=0,\vec{\sigma}_{v}=0,\eta_{v}=0,u_{i}=0]\label{structure}\\
& & +\mathcal{A}\left(\left(\partial u_{0}\right)^{2}+\left(\partial\vec{u}\right)^{2}\right)+\mathcal{B}\left(s_{0}u_{0}+\vec{p}\vec{u}\right)+\mathcal{C}\left(s_{0}p_{0}+\vec{p}\vec{s}\right)^{2}+\mathcal{D}\left(s_{0}\eta_{v}+\vec{p}\vec{\sigma}_{v}\right)^{2}+a\left(p_{0}^{2}+{\vec{s}}^{2}\right)+b\left(p_{0}\eta_{v}+\vec{s}\vec{\sigma}_{v}\right)\nonumber \\
& & +c\left(\eta_{v}^{2}+{\vec{\sigma}_{v}}^{2}\right)+d\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}+e\left(u_{0}p_{0}+\vec{u}\vec{s}\right)\left(u_{0}\eta_{v}+\vec{u}\vec{\sigma}_{v}\right)+f\left(u_{0}\eta_{v}+\vec{u}\vec{\sigma}_{v}\right)^{2}+\mathcal{O}\left(s^{6},p^{6}\right),\nonumber \end{eqnarray}
where we omitted the terms containing derivatives of the fields,
since the external currents are constants, the constants $\mathcal{A}-\mathcal{D},\, a-f$
should be evaluated with account of NLO corrections. The vacuum equations
which follow from (\ref{structure}) are \begin{eqnarray}
& & \frac{\partial\Gamma_{eff}[m,\lambda,\sigma_{v},\vec{s}=0,p_{0}=0,\vec{\sigma}_{v}=0,\eta_{v}=0]}{\partial\lambda}=\frac{\partial\Gamma_{eff}[m,\lambda,\sigma_{v},\vec{s}=0,p_{0}=0,\vec{\sigma}_{v}=0,\eta_{v}=0]}{\partial\sigma_{v}}=0\label{eq:SigmaLambdaGap}\\
& & \frac{\partial\Gamma_{eff}[m,\lambda,\sigma,\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},u_{i}]}{\partial\sigma_{v,i}}=\frac{\partial\Gamma_{eff}[m,\lambda,\sigma,\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},u_{i}]}{\partial\eta_{v}}=\frac{\partial\Gamma_{eff}[m,\lambda,\sigma,\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},u_{i}]}{\partial u_{i}}=0\label{eq:SigmaEtaGap}\end{eqnarray}
The coefficients $\mathcal{A},\mathcal{B}$ are relevant for the
2-point correlators with intermediate pions and $\mathcal{A}\sim F^{2}$
and $\mathcal{B}\propto M_{\pi}^{2}$ in $m_{u}=m_{d}$ limit. The
constants $\mathcal{B},\mathcal{C},\mathcal{D}$ are irrelevant to
our problem since they are constants in front of the term with chiral
doublet $\chi=\left(s_{0},\vec{p}\right)$ which we put to zero in
the current paper.
The Eqns~(\ref{eq:SigmaLambdaGap}) are responsible for the dynamical
mass generation and will be discussed in the next section. The Eqns~(\ref{eq:SigmaEtaGap})
may be explicitly written as \begin{eqnarray}
& & \frac{\partial\Gamma_{eff}[m,\lambda,\sigma,\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},u_{i}]}{\partial\eta_{v}}=bp_{0}+2c\eta_{v}+eu_{0}\left(u_{0}p_{0}+\vec{u}\vec{s}\right)+2fu_{0}\left(u_{0}\eta_{v}+\vec{u}\vec{\sigma}_{v}\right)=0\label{eq:GapEtaExplicit}\\
& & \frac{\partial\Gamma_{eff}[m,\lambda,\sigma,\vec{s},p_{0},\vec{\sigma}_{v},\eta_{v},u_{i}]}{\partial\sigma_{v,i}}=b\vec{s}_{v}+2c\vec{\sigma}_{v}+e\vec{u}\left(u_{0}p_{0}+\vec{u}\vec{s}\right)+2f\vec{u}\left(u_{0}\eta_{v}+\vec{u}\vec{\sigma}_{v}\right)=0\label{eq:GapSigmaExplicit}\end{eqnarray}
Multiplying Eqn~(\ref{eq:GapEtaExplicit}) on $u_{0}$ and Eqn.~(\ref{eq:GapSigmaExplicit})
on $\vec{u}$ and adding results, we may find: \begin{eqnarray}
u_{0}\eta_{v}+\vec{u}\vec{\sigma}_{v} & = & -\frac{b+e}{2\left(c+f\right)}\left(u_{0}p_{0}+\vec{u}\vec{s}\right).\label{eq:GapRes1}\end{eqnarray}
Repeating the same trick with $p_{0}$ and $\vec{s}$, we may get
\begin{eqnarray}
p_{0}\eta_{v}+\vec{s}\vec{\sigma}_{v} & = & -\frac{1}{2c}\left[b\left(p_{0}^{2}+{\vec{s}}^{2}\right)+\left(e-f\frac{b+e}{c+f}\right)\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}\right],\label{eq:GapRes2}\end{eqnarray}
and repeating the same trick with $\eta_{v}$ and $\vec{\sigma}_{v}$,
we may get \begin{eqnarray}
\eta_{v}^{2}+{\vec{\sigma}_{v}}^{2}=-\frac{1}{2c}\left[-\frac{b^{2}}{2c}\left(p_{0}^{2}+{\vec{s}}^{2}\right)+\left(-\frac{b}{2c}\left(e-f\frac{b+e}{c+f}\right)-e\frac{b+e}{2(c+f)}+2f\left(\frac{b+e}{2(c+f)}\right)^{2}\right)\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}\right]\label{eq:GapRes3}\end{eqnarray}
Combining results~(\ref{eq:GapRes1}-\ref{eq:GapRes3}), we may
get for the effective action \begin{eqnarray}
& & \Gamma_{eff}[\lambda,\sigma,\vec{s},p_{0},u_{i}]=...+a\left(p_{0}^{2}+{\vec{s}}^{2}\right)-\frac{b}{2c}\left[b\left(p_{0}^{2}+{\vec{s}}^{2}\right)+\left(e-f\frac{b+e}{c+f}\right)\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}\right]\label{eq:GammaEffChiChi}\\
& & -\left[-\frac{b^{2}}{4c}\left(p_{0}^{2}+{\vec{s}}^{2}\right)+\left(-\frac{b}{4c}\left(e-f\frac{b+e}{c+f}\right)-e\frac{b+e}{4(c+f)}+f\left(\frac{b+e}{2(c+f)}\right)^{2}\right)\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}\right]\nonumber \\
& & +d\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}-e\frac{b+e}{2(c+f)}\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}+f\left(\frac{b+e}{2(c+f)}\left(u_{0}p_{0}+\vec{u}\vec{s}\right)\right)^{2}\nonumber \\
& & =...+\left(a-\frac{b^{2}}{4c}\right)\left(p_{0}^{2}+{\vec{s}}^{2}\right)+\left(d-\frac{b}{4c}\left(e-f\frac{b+e}{c+f}\right)-e\frac{b+e}{4(c+f)}\right)\left(u_{0}p_{0}+\vec{u}\vec{s}\right)^{2}+\mathcal{O}\left(\chi,\chi^{\dagger}\chi\right),\nonumber \end{eqnarray}
where we omitted the terms which are proportional to the chiral doublet
$\chi$. The terms shown in~(\ref{eq:GammaEffChiChi}) are explicitly
chiral invariant and correspond to the terms $\left(\tilde{\chi}^{\dagger}\tilde{\chi}\right)$
and $\left(\tilde{\chi}^{\dagger}U\right)^{2}$ in the chiral lagrangian~\cite{Gasser:1983yg}.
Respectively, for the constant $l_{7}$ we may deduce \begin{eqnarray}
l_{7}=\frac{d-\frac{b}{4c}\left(e-f\frac{b+e}{c+f}\right)-e\frac{b+e}{4(c+f)}}{4B^{2}}\label{eq:L7-abcdef}\end{eqnarray}
Thus we can see that in addition to the term $d$ in numerator there
are three other terms which correspond to contributions of additional
mesons. As we will see from the following sections, these contributions
have different signs and approximately the same order of magnitude
as the term $d$. The formula~(\ref{eq:L7-abcdef}) proves that we
have to consider correlators instead of direct comparison of the terms
in the expansion of the lagrangian.
Below we will not evaluate the constants $\mathcal{A}-\mathcal{D},a-f,$
but instead evaluate the correlators directly.
\section{Gap equation}
\label{sec:GapEqn}
The next-to-leading order (NLO) gap equations which follow from the
effective action~(\ref{eq:Z0}) have a form\begin{eqnarray}
\sigma\frac{\partial S}{\partial\sigma} & = & 4\sigma^{2}-\frac{1}{V}Tr\left(iM(p)\hat{S}(p)\right)-\frac{1}{\sigma^{2}}\int\frac{d^{4}q}{(2\pi)^{4}}\sum V_{3}^{(ij)}(q)\Pi_{ij}(q)=0,\label{eq:GapSigmaNLO}\\
\sigma_{3}\frac{\partial S}{\partial\sigma_{3}} & = & 4\sigma_{3}^{2}-\frac{1}{V}Tr\left(i\delta M(p)\tau_{3}\hat{S}(p)\right)-\frac{1}{\sigma^{2}}\int\frac{d^{4}q}{(2\pi)^{4}}\sum\tilde{V}_{3}^{(ij)}(q)\Pi_{ij}(q)=0\label{eq:GapEquations}\\
\lambda\frac{\partial S}{\partial\lambda} & = & \frac{N}{V}-\frac{1}{2V}Tr\left(\hat{S}(p)\left(iM(p)+i\tau_{3}\delta M(p)\right)\right)+\nonumber \\
& & +\frac{1}{2\sigma^{2}}\int\frac{d^{4}q}{(2\pi)^{4}}\sum_{i}\left(V_{2}^{(ij)}(q)-V_{3}^{(ij)}(q)\right)\Pi_{ij}(q)=0,\label{eq:GapLambdaNLO}\end{eqnarray}
where we used notations\begin{eqnarray}
V_{2}^{(gap)(ij)}(q) & = & \frac{1}{\sigma^{2}}\int\frac{d^{4}p}{(2\pi)^{4}}\, Tr\left(\frac{M(p)}{\hat{p}+i\mu(p)+i\tau_{3}\delta\mu(p)}\Gamma_{i}\frac{M(p+q)}{\hat{p}+\hat{q}+i\mu(p+q)+i\tau_{3}\delta\mu(p+q)}\Gamma_{j}\right),\label{eq:V2gap-definition}\\
V_{3}^{(gap)(ij)}(q) & = & \frac{i}{\sigma^{2}}\int\frac{d^{4}p}{(2\pi)^{4}}\, Tr\left(\left(\frac{M(p)}{\hat{p}+i\mu(p)+i\tau_{3}\delta\mu(p)}\right)^{2}\Gamma_{i}\frac{M(p+q)}{\hat{p}+\hat{q}+i\mu(p+q)+i\tau_{3}\delta\mu(p+q)}\Gamma_{j}\right),\label{eq:V3gap-definition}\\
\tilde{V}_{3}^{(gap)(ij)}(q) & = & \frac{i}{\sigma^{2}}\int\frac{d^{4}p}{(2\pi)^{4}}\, Tr\left(\left(\frac{M(p)\delta M(p)\tau_{3}}{\left(\hat{p}+i\mu(p)+i\tau_{3}\delta\mu(p)\right)^{2}}\right)\Gamma_{i}\frac{M(p+q)}{\hat{p}+\hat{q}+i\mu(p+q)+i\tau_{3}\delta\mu(p+q)}\Gamma_{j}\right),\label{eq:V3tildeGap-definition}\end{eqnarray}
explicit expressions for the vertices~(\ref{eq:V2gap-definition}-\ref{eq:V3tildeGap-definition})
are given in Appendix~\ref{sec:appExplicitExpressions}, and the
propagators used for evaluation are written out in Section~\ref{sec:Propagators}.
In general, equations~(\ref{eq:GapSigmaNLO}-\ref{eq:GapLambdaNLO})
can be solved only numerically.
\subsection{Expansion over $\delta m$ }
\label{sec:Expansion-epsilon} For the special case when $\delta m$
is small, it is possible to solve the equations~(\ref{eq:GapEquations})
making a systematic expansion over the small parameter $\delta m$.
For our purpose it suffices to keep just the first order corrections.
From the first and the third gap equations in~(\ref{eq:GapEquations})
and the structure of the vertices~(\ref{eq:V2gap-definition},\ref{eq:V3gap-definition})
we may conclude that the vacuum expectation values for $\langle\sigma\rangle,\langle\lambda\rangle$
get corrections only in the second order over $\delta m,$ thus in
the first order they remain the same as for $\delta m=0.$ The equation
for the $\langle\sigma_{3}\rangle$ has a form
\begin{align}
\sigma_{3}\frac{\partial S}{\partial\sigma_{3}} & \approx-4\epsilon^{2}\langle\sigma\rangle^{2}-8\epsilon N_{c}\int\frac{d^{4}p}{(2\pi)^{4}}\frac{\left(p^{2}-\mu^{2}(p)\right)M(p)(\delta m+\epsilon M(p))}{\left(p^{2}+\mu^{2}(p)\right)^{2}}-\label{eq:GapSplitApproximate}\\
& \frac{1}{\sigma^{2}}\int\frac{d^{4}q}{(2\pi)^{4}}\sum\tilde{V}_{3}^{(ij)}(q)\Pi_{ij}(q)=0,\end{align}
where $\epsilon=\frac{i\langle\sigma_{3}\rangle}{\langle\sigma\rangle},$
\begin{eqnarray*}
\tilde{V}_{3}^{(ij)}\Pi_{ij}(q) & \approx & \frac{\epsilon}{2\sigma^{2}}\int\frac{d^{4}p}{(2\pi)^{4}}M^{2}(p)M(p+q)\times\left\{ \right.\\
& \left(\Pi_{\sigma\sigma}^{(0)}(q)-\Pi_{\sigma_{3}\sigma_{3}}^{(0)}(q)\right) & Tr\left[S_{+}(p)S_{+}(p)S_{+}(p+q)-S_{-}(p)S_{-}(p)S_{-}(p+q)\right]_{\mathcal{O}(\delta m)}+\\
& 2\Pi_{\sigma\sigma_{3}} & Tr\left[S_{+}(p)S_{+}(p)S_{+}(p+q)+S_{-}(p)S_{-}(p)S_{-}(p+q)\right]_{\delta m=0}-\\
\sum_{i_{\perp}=1,2} & \Pi_{\sigma_{i}}^{(0)}(k) & Tr\left[S_{+}(p)S_{+}(p)S_{-}(p+q)-S_{-}(p)S_{-}(p)S_{+}(p+q)\right]_{\mathcal{O}(\delta m)}-\\
& \left(\Pi_{\eta\eta}^{(0)}(k)-\Pi_{\phi_{3}\phi_{3}}^{(0)}(k)\right) & Tr\left[S_{+}(p)S_{+}(p)\bar{S}_{+}(p+q)-S_{-}(p)S_{-}(p)\bar{S}_{-}(p+q)\right]_{\mathcal{O}(\delta m)}-\\
& 2\Pi_{\eta\phi_{3}}(k) & Tr\left[S_{+}(p)S_{+}(p)\bar{S}_{+}(p+q)+S_{-}(p)S_{-}(p)\bar{S}_{-}(p+q)\right]_{\delta m=0}+\\
\sum_{i_{\perp}=1,2} & \Pi_{\phi_{i}}^{(0)}(k) & Tr\left[S_{+}(p)S_{+}(p)\bar{S}_{-}(p+q)-S_{-}(p)S_{-}(p)\bar{S}_{+}(p+q)\right]_{\mathcal{O}(\delta m)}\\
& & \left.\right\} ,\end{eqnarray*}
the superscript $(0)$ on the propagators and subscripts on $Tr[...]_{\alpha}$
indicates that the proper propagator is to be taken in the limit $\delta m=0$
or just collecting the first $\mathcal{O}(\delta m)$-correction.
One can notice that (\ref{eq:GapSplitApproximate}) has a form
\[
\epsilon\left(\mathcal{X}\epsilon+\mathcal{Y}\delta m\right)=0,\]
where the coefficients~$\mathcal{X},\mathcal{Y}$ should be evaluated
with account of $\mathcal{O}(1/N_{c})$-corrections. A trivial nonzero
solution is $\epsilon=-\delta m\,\mathcal{Y}/\mathcal{X},$ which
corresponds to \begin{equation}
\delta\mu(p)=\delta m\left(1-M\, f^{2}(p)\mathcal{Y}/\mathcal{X}\right).\label{eq:MassSplitBA}\end{equation}
While in general case the explicit expression for the formfactor
has a form\begin{equation}
f(p)=-x\frac{d}{dx}\left(I_{0}(x)K_{0}(x)-I_{1}(x)K_{1}(x)\right)_{x=\frac{p\rho}{2}},\label{eq:FF_BesselExact}\end{equation}
in order to speed up the evaluations here and below we consider two
parameterizations of the formfactor. The first one is a simple {}``dipole''
parameterization~(\cite{Diakonov:1985eg}) of the form \begin{equation}
f(p)=\frac{2}{2+p^{2}\rho^{2}},\label{eq:FF_Dipole}\end{equation}
which coincides with (\ref{eq:FF_BesselExact}) in the region of
small $p\lesssim2/\rho$. The second parameterization has a form \begin{equation}
f(p)=\left.\frac{1}{\sqrt{1+a_{1}x^{2}+a_{2}x^{4}+a_{3}x^{6}}}\right|_{x=\frac{p\rho}{2}},\label{eq:FF_QuasiBessel}\end{equation}
where the free parameters $a_{1}...a_{3}$ are fitted to~(\ref{eq:FF_BesselExact}),
and (\ref{eq:FF_QuasiBessel}) agrees with (\ref{eq:FF_BesselExact})
both for the small and asymptotically large $p.$ We will refer to
(\ref{eq:FF_Dipole}) and (\ref{eq:FF_QuasiBessel}) as dipole and
quasibessel parameterizations respectively. Direct comparison of the
two close parameterizations is important in order to demonstrate that
the results of this paper are very sensitive to the shape of the instanton.
We summarize results obtained for the constants $\mathcal{X},\mathcal{Y}$
with different parameterizations of formfactor in Table~\ref{tab:AB-Parameters}.
As we can see, the $1/N_{c}$-corrections $\mathcal{X}_{NLO}$~and
$\mathcal{Y}_{NLO}$ are small compared to $\mathcal{X}_{LO}$ and
$\mathcal{Y}_{LO}$ respectively, so the $1/N_{c}$-expansion works
very well here. It is important to note that both in the leading order
and next-to-leading order the mass $\delta\mu(p)$ changes sign for
$p\sim0.5$~GeV, so we have a compensation of the small-$p$ and
large-$p$ regions.
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& \textbf{$\mathcal{X}_{LO}\times10^{3}$} & \textbf{$\mathcal{Y}_{LO}\times10^{2}$} & \textbf{$-(\mathcal{Y}/\mathcal{X})_{LO}$} & \textbf{$\mathcal{X}_{NLO}\times10^{3}$} & \textbf{$\mathcal{Y}_{NLO}\times10^{2}$} & \textbf{$-(\mathcal{Y}/\mathcal{X})_{LO+NLO}$}\tabularnewline
\hline
\textbf{Dipole} & -8.50 & --3.53 & -4.16 & 0.80 & 0.61 & -3.80\tabularnewline
\hline
\textbf{QuasiBessel} & -9.04 & -3.33 & -3.68 & 1.02 & 0.69 & -3.30\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:AB-Parameters}Parameters of mass split in different parameterizations
of formfactor. Dipole corresponds to~(\ref{eq:FF_Dipole}). QuasiBessel
corresponds to~(\ref{eq:FF_QuasiBessel}). Dimensions: $[\mathcal{X}_{LO,NLO}]=[GeV^{4}],[\mathcal{Y}_{LO,NLO}]=[GeV^{3}]$.}
\end{table}
\section{Propagators}
\label{sec:Propagators} In this section we would like to discuss
the propagators of the quarks and mesons in presence of the mass split
$\delta m$.
\subsection{Quark propagator}
Here we consider only the leading order quark propagator, the NLO
corrections to the quark propagator will be considered as separate
meson loop corrections to proper correlators. Since the operator $\hat{p}+i\mu(p)+i\delta\mu(p)\tau_{3}$
is diagonal in the flavour space, its inversion is quite straightforward,
with\[
\hat{S}(p)\equiv\frac{1}{\hat{p}+i\mu(p)+i\delta\mu(p)\tau_{3}}=\frac{1-\tau_{3}}{2}\, S_{-}(p)+\frac{1+\tau_{3}}{2}\, S_{+}(p),\]
\[
S_{\pm}(p)=\frac{1}{\hat{p}+i\mu_{\pm}(p)},\]
where $\mu_{\pm}(p)=\mu(p)\pm\delta\mu(p)$.
\subsection{Meson propagator}
For evaluations in this paper we have to evaluate the meson propagator
with account of $1/N_{c}$-corrections. However it is important to
note that NLO evaluations are needed only for $\Pi_{\eta\eta}(0)$,$\underset{q^{2}=-m_{\pi}^{2}}{Res}\Pi_{\phi\phi}(q)$,$\Pi_{\eta\phi}(0)$,
all the other components and expressions for $q\not=0$ may be evaluated
in leading order, which significantly simplifies the task. Due to
$\delta m\not=0$ the propagator is nondiagonal in indices $(ij)$--we
get additional transitions $\sigma\leftrightarrow\vec{\sigma}$ and
$\eta\leftrightarrow\vec{\phi}$. Inversion of the propagator is trivial
and gives:
\begin{equation}
\Pi_{00}=\frac{\left(\Pi^{-1}\right)_{33}}{\left(\Pi^{-1}\right)_{00}\left(\Pi^{-1}\right)_{33}-\left(\Pi^{-1}\right)_{03}\left(\Pi^{-1}\right)_{30}},\qquad\Pi_{33}=\frac{\left(\Pi^{-1}\right)_{00}}{\left(\Pi^{-1}\right)_{00}\left(\Pi^{-1}\right)_{33}-\left(\Pi^{-1}\right)_{03}\left(\Pi^{-1}\right)_{30}},\label{eq:MesonInv_00_33}\end{equation}
\begin{equation}
\Pi_{03}=-\frac{\left(\Pi^{-1}\right)_{30}}{\left(\Pi^{-1}\right)_{00}\left(\Pi^{-1}\right)_{33}-\left(\Pi^{-1}\right)_{03}\left(\Pi^{-1}\right)_{30}},\qquad\Pi_{30}=-\frac{\left(\Pi^{-1}\right)_{03}}{\left(\Pi^{-1}\right)_{00}\left(\Pi^{-1}\right)_{33}-\left(\Pi^{-1}\right)_{03}\left(\Pi^{-1}\right)_{30}},\label{eq:MesonInv_03_30}\end{equation}
\[
\Pi_{ij}=\frac{\delta_{ij}}{\left(\Pi^{-1}\right)_{i}},\,\,\,(i,\, j)\not=3\]
where we used a shorthand notation $(0,3)=(\sigma,\,\sigma_{3})$
for positive parity mesons, and $(0,3)=(\eta,\,\phi_{3})$ for negative
parity mesons.
\subsubsection{Leading order}
In the leading order for the components $\left(\Pi^{-1}\right)_{ij}$
we have \begin{equation}
\left(\Pi^{-1}\right)_{ij}=4\delta_{ij}+\frac{1}{\sigma^{2}}Tr\left(\hat{Q}(p)\Gamma_{i}\hat{Q}(p+q)\Gamma_{j}\right)\label{eq:InversePropagator}\end{equation}
where \[
\hat{Q}(p)=iM(p)\hat{S}(p)\equiv\frac{iM(p)}{\hat{p}+i\mu(p)+i\delta\mu(p)\tau_{3}},\]
and explicit expressions for the components are given in Appendix~\ref{sec:appExplicitExpressions}.
\subsubsection{NLO correction}
As it was discussed earlier, we need a few values for propagators
in the next-to-leading order. Since the NLO evaluations are numerically
slow, from the very beginning we will concentrate on evaluation of
the following quantities:\begin{align*}
\lim_{(q\to0,m\to0,\delta m\to0)}\Pi_{\eta\eta}^{-1}(q),\lim_{(q\to0,m\to0,\delta m\to0)}\frac{\Pi_{\eta\phi}^{-1}(q)}{\delta m},\underset{q^{2}=0}{Res}\Pi_{\phi\phi}(q).\end{align*}
All the terms which do not contribute to one of these limits will
be omitted. For the sake of brevity, below we use notation\[
Q(p)\equiv\frac{Q_{+}(p)+Q_{-}(p)}{2}\approx\frac{iM(p)}{\hat{p}+i\mu(p)}+\mathcal{O}\left(\delta m^{2}\right)\]
For the pion propagator $\Pi_{\phi\phi}(q)$, we may use the chiral
limit and put $m,\delta m$ to zero. The NLO expression for the pion
propagator has a form
\begin{align*}
\Pi_{\phi\phi}^{(ab)-1}(q) & =\left[4\delta^{ab}+\frac{1}{\sigma^{2}}Tr_{p}\left(Q(p)i\gamma_{5}\tau^{a}Q(p+q)i\gamma_{5}\tau^{b}\right)\right]+\\
& +\frac{1}{\sigma^{4}}\int\frac{d^{4}k}{\left(2\pi\right)^{4}}\Pi_{ij}(k)\left(2Tr_{p}\left(Q(p)i\gamma_{5}\tau^{a}Q(p+q)\Gamma_{i}Q(p+q+k)\Gamma_{j}Q(p+q)i\gamma_{5}\tau^{b}\right)\right.+\\
& \left.+Tr_{p}\left(Q(p)i\gamma_{5}\tau^{a}Q(p+q)\Gamma_{i}Q(p+q+k)i\gamma_{5}\tau^{b}Q(p+k)\Gamma_{j}\right)\right)\\
& -\frac{4}{\sigma^{6}}\int\frac{d^{4}k}{\left(2\pi\right)^{4}}\Pi_{i}(k)\Pi_{j}(k+q)Tr_{p}\left(Q(p)i\gamma_{5}\tau^{a}Q(p+q)\Gamma_{i}Q(p+q+k)\Gamma_{j}\right)\times\\
& Tr_{p}\left(Q(p)i\gamma_{5}\tau^{b}Q(p-q)\Gamma_{i}Q(p-q-k)\Gamma_{j}\right).\end{align*}
In complete analogy, for the $\eta$-meson propagator $\Pi_{\eta\eta}(0)$
we have
\begin{align*}
\Pi_{\eta\eta}^{-1}(0) & =\left[4\delta^{ab}+\frac{1}{\sigma^{2}}Tr_{p}\left(Q(p)\gamma_{5}Q(p)\gamma_{5}\right)\right]+\\
& +\frac{1}{\sigma^{4}}\int\frac{d^{4}k}{\left(2\pi\right)^{4}}\Pi_{ij}(k)\left(2Tr_{p}\left(Q(p)\gamma_{5}Q(p)\Gamma_{i}Q(p+k)\Gamma_{j}Q(p)\gamma_{5}\right)+\right.\\
& +\left.Tr_{p}\left(Q(p)\gamma_{5}Q(p)\Gamma_{i}Q(p+k)\gamma_{5}Q(p+k)\Gamma_{j}\right)\right)\\
& -\frac{4}{\sigma^{6}}\int\frac{d^{4}k}{\left(2\pi\right)^{4}}\Pi_{i}(k)\Pi_{j}(k)Tr_{p}\left(Q(p)\gamma_{5}Q(p)\Gamma_{i}Q(p+k)\Gamma_{j}\right)\times\\
& Tr_{p}\left(Q(p)\gamma_{5}Q(p)\Gamma_{i}Q(p-k)\Gamma_{j}\right),\end{align*}
and again we can make evaluations in the chiral limit.
The nondiagonal matrix element $\Pi_{\eta\phi}(0)$ is $\mathcal{O}\left(\delta m\right)$,
so we will extract explicitly $\delta m$ and after that the evaluation
of the constant will be done in the chiral limit. Evaluation is quite
tedious since LO propagators have nondiagonal components. The corresponding
expression has a form\[
\Pi_{\eta\phi}^{-1}(0)=\Pi_{\eta\phi}^{(LO)-1}(0)+\Pi_{\eta\phi}^{(1-meson)-1}(0)+\Pi_{\eta\phi}^{(2-meson)-1}(0),\]
where
\[
\Pi_{\eta\phi}^{(LO)-1}(0)=\left[\frac{1}{2\sigma^{2}}Tr_{p}\left(Q_{+}(p)i\gamma_{5}Q_{+}(p)\gamma_{5}\right)-\frac{1}{2\sigma^{2}}Tr_{p}\left(Q_{-}(p)i\gamma_{5}Q_{-}(p)\gamma_{5}\right)\right],\]
\begin{eqnarray}
\Pi_{\eta\phi}^{(1-mes)-1}(0) & = & \int\frac{d^{4}q}{(2\pi)^{4}}\sum_{ij}\Pi_{ij}(q)V_{ij}^{(1-mes,\eta\phi)}(q)=\int\frac{d^{4}q}{(2\pi)^{4}}\sum_{ij}\Pi_{ij}(q)\frac{i}{2\sigma^{4}}\int\frac{d^{4}p}{(2\pi)^{4}}\times\nonumber \\
& \times & Tr\left(2Q(p)\Gamma_{\eta}Q(p)\Gamma_{\phi}Q(p)\Gamma_{i}Q(p+q)\Gamma_{j}+Q(p)\Gamma_{\eta}Q(p)\Gamma_{i}Q(p+q)\Gamma_{\phi}Q(p+q)\Gamma_{j}\right),\label{eq:PiEta1Loop}\end{eqnarray}
\begin{equation}
\Pi_{\eta\phi}^{(2-mes)-1}(0)=-\frac{4}{\sigma^{6}}\int\frac{d^{4}q}{\left(2\pi\right)^{4}}\Pi_{ij}(q)\Pi_{kl}(q)V_{ik}^{(\eta)}(q)V_{jl}^{(\phi)}(q),\label{eq:PiEta2Loop-0}\end{equation}
\begin{align*}
V_{ik}^{(\eta)}(q) & =\left[Tr_{p}\left(Q(p)\gamma_{5}Q(p)\Gamma_{i}Q(p+q)\Gamma_{k}\right)\right],\\
V_{ik}^{(\phi)}(q) & =\left[Tr_{p}\left(Q(p)i\gamma_{5}\tau^{3}Q(p)\Gamma_{i}Q(p+q)\Gamma_{k}\right)\right],\end{align*}
and explicit expessions for the verices contributing to~(\ref{eq:PiEta1Loop}-\ref{eq:PiEta2Loop-0})
are given in Appendix~\ref{sec:appExplicitExpressions}.
Numerial results of evaluation are presented in Table~\ref{tab:Propagators}.
As we can see, even in the leading order (LO) there is a strong sensitivity
of the propagator $\Pi_{\eta\phi}^{-1}(0)$ to the shape of the instanton
(formfactor $F(p)$). This dependence is discussed in more detail
in Section~\ref{sec:Conclusion}.
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& \textbf{LO} & \textbf{Mass Shift} & \textbf{Mass Split} & \textbf{Meson} & \textbf{All NLO} & \textbf{Total}\tabularnewline
\hline
\textbf{QuasiBessel} & & & & & & \tabularnewline
\hline
$\Pi_{\eta\eta}^{-1}(0)$ & $5.65\times10^{-3}$ & $9.16\times10^{-3}$ & 0 & $-8.88\times10^{-3}$ & $2.86\times10^{-3}$ & $5.93\times10^{-3}$\tabularnewline
\hline
$-i\Pi_{\eta\phi}^{-1}(0)$ & $-4.73\times10^{-3}$ & $0.88\times10^{-3}$ & $0.19\times10^{-3}$ & $-2.30\times10^{-3}$ & $-1.23\times10^{-3}$ & $-1.89\times10^{-3}$\tabularnewline
\hline
$F_{\pi}^{2}$ & $1.24\times10^{-2}$ & $-0.59\times10^{-2}$ & 0 & $0.11\times10^{-2}$ & $-0.49\times10^{-2}$ & $0.76\times10^{-2}$\tabularnewline
\hline
$\langle\bar{q}q\rangle$ & $2.03\times10^{-2}$ & $-0.77\times10^{-2}$ & 0 & $0.31\times10^{-2}$ & $-0.45\times10^{-2}$ & $1.58\times10^{-2}$\tabularnewline
\hline
$B$ & $1.64$ & --- & 0 & --- & --- & $2.09$\tabularnewline
\hline
\textbf{Dipole} & & & & & & \tabularnewline
\hline
$\Pi_{\eta\eta}^{-1}(0)$ & $5.65\times10^{-3}$ & $9.16\times10^{-3}$ & 0 & $-9.72\times10^{-3}$ & $-0.56\times10^{-3}$ & $5.09\times10^{-3}$\tabularnewline
\hline
$-i\Pi_{\eta\phi}^{-1}(0)$ & $-1.72\times10^{-3}$ & $1.47\times10^{-3}$ & $0.17\times10^{-3}$ & $-4.95\times10^{-3}$ & $-3.32\times10^{-3}$ & $-5.20\times10^{-3}$\tabularnewline
\hline
$F_{\pi}^{2}$ & $1.36\times10^{-2}$ & $-0.54\times10^{-2}$ & 0 & $0.31\times10^{-2}$ & $-0.23\times10^{-2}$ & $1.12\times10^{-2}$\tabularnewline
\hline
$\langle\bar{q}q\rangle$ & $2.18\times10^{-2}$ & $-0.72\times10^{-2}$ & 0 & $0.31\times10^{-2}$ & $-0.41\times10^{-2}$ & $1.76\times10^{-2}$\tabularnewline
\hline
$B$ & $1.60$ & --- & 0 & --- & --- & $1.57$\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:Propagators}In this table we give the numbers obtained
for propagators and other relevant constants. $F_{\pi}^{2}$ is used
for evaluation of $Res\left(\Pi_{\phi}\right),$ $\langle\bar{q}q\rangle$
is used for extraction of constant $B$.}
\end{table}
\section{Quark condensate}
\label{sec:QuarkCondensate}
Due to the mass split $\delta m$ there is a flavour difference for
the quark condensate $\delta\langle\bar{q}q\rangle=\langle\bar{u}u\rangle-\langle\bar{d}d\rangle$
. In the leading order this split is\[
\delta\left\langle \bar{q}q\right\rangle _{LO}=\frac{i}{2}Tr\left(\tau_{3}S(p)\right)=4N_{c}\int\frac{d^{4}p}{\left(2\pi\right)^{4}}\left(\frac{\mu_{+}(p)}{p^{2}+\mu_{+}^{2}(p)}-\frac{\mu_{-}(p)}{p^{2}+\mu_{-}^{2}(p)}\right).\]
In the NLO evaluation is also quite straightforward, with\begin{align}
\delta\left\langle \bar{q}q\right\rangle _{meson} & =\int\frac{d^{4}q}{\left(2\pi\right)^{4}}\sum_{ij}\Pi_{ij}(q)V_{ij}^{\left(\delta\bar{q}q\right)}(q),\label{eq:qq_NLO-def}\\
V_{ij}^{\left(\delta\bar{q}q\right)}(q) & =-\int\frac{d^{4}p}{\left(2\pi\right)^{4}}M(p)M(p+q)Tr\left(\frac{i\tau_{3}}{2}S(p)\Gamma_{i}S(p+q)\Gamma_{j}S(p)\right).\nonumber \end{align}
for meson corrections plus corrections from mass shift and mass split
($1/N_{c}$ corrections to $M_{0}$ and $M_{3}$), and explicit expression
for~(\ref{eq:qq_NLO-def}) is given in Appendix~\ref{sec:appExplicitExpressions}.
Results of numerical evaluation are presented in the Table~\ref{tab:qq-split}.
As one can see, due to the large NLO corrections to the mass split
$M_{u}(p)-M_{d}(p)$, the NLO corrections are larger than the LO result.
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& \textbf{LO} & \textbf{Mass Shift} & \textbf{Mass Split} & \textbf{Mesons} & \textbf{All NLO} & \textbf{LO+NLO}\tabularnewline
\hline
\textbf{Dipole} & -0.20 & $6.07\times10^{-2}$ & $1.26\times10^{-2}$ & $1.09\times10^{-3}$ & $7.45\times10^{-2}$ & -0.13\tabularnewline
\hline
\textbf{QuasiBessel} & -0.18 & $6.11\times10^{-2}$ & $1.29\times10^{-2}$ & $2.99\times10^{-3}$ & $7.70\times10^{-2}$ & -0.10\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:qq-split}Different contributions to $\frac{\left\langle \bar{u}u\right\rangle -\left\langle \bar{d}d\right\rangle }{m_{u}-m_{d}}$.
LO: Leading order result Mass Split: Contribution due to NLO correction
to mass split $M_{u}(p)-M_{d}(p)$. Mesons: Contribution of mesons,
All NLO: sum of contributions of mesons and mass shift, LO+NLO--final
result.}
\end{table}
Using formula (11.3) from~\cite{Gasser:1983yg}, it is possible to
get for the constant $h_{3}$ an estimate~%
\footnote{Note that the transition form Minkowsky to Euclid requires to change
the signs of all quark condensates, $\langle\bar{q}q\rangle\to-\langle\bar{q}q\rangle$%
}:\[
h_{3}=\frac{\left.\left(\left\langle \bar{u}u\right\rangle -\left\langle \bar{d}d\right\rangle \right)\right|_{\delta m}^{LO+NLO}}{4B^{2}\delta m}=\frac{0.10\,\delta m}{4B^{2}\delta m}\approx5.48\times10^{-3}.\]
\section{Evaluation of the constant $l_{7}$}
\label{sec:P3P0Correlator}
According to \cite{Gasser:1983yg}, it is possible to evaluate the
constant $l_{7}$ from the correlator $\langle P^{3}(x)P^{0}(0)\rangle$
as\begin{equation}
P_{2}(q)=\int d^{4}x\, e^{iqx}\langle P^{3}(x)P^{0}(0)\rangle=\frac{G_{\pi}\tilde{G}_{\pi}}{m_{\pi}^{2}-q^{2}}+\mathcal{O}\left(q^{2}\right)=\frac{8B^{3}\left(m_{u}-m_{d}\right)}{q^{2}-m_{\pi}^{2}}l_{7}+\mathcal{O}\left(m,q^{2}\right),\label{eq:P3P0Definition}\end{equation}
where the constant $B$ is one of the phenomenological parameters
of the chiral lagrangian (see Table~\ref{tab:Propagators}), the
mass of the pion $m_{\pi}=0$ in the limit $m\to0$ and $m_{u},m_{d}$
are the current quark masses. Since we are interested only in the
residue of the correlator, we should consider only 1-particle reducible
diagrams with pion in the intermediate state.
In the leading order, there are two diagrams shown in the Figure~\ref{fig:P3P0LO}.
Obviously, only the diagram on the right-hand side contributes to
the residue, yielding
\begin{figure}
\includegraphics[scale=0.2]{Figures/LO1}\includegraphics[scale=0.3]{Figures/LO2}
\caption{\label{fig:P3P0LO}Contribution to the $\langle P^{3}P^{0}\rangle$-correlator
in the leading order.}
\end{figure}
\[
P_{2}^{LO}(q)=\sum_{i,j=\eta,\phi}L_{i}^{LO}(q)R_{j}^{LO}(q)\Pi_{ij}(q)=L_{\eta}(q)L_{\phi}(q)\left(\Pi_{\eta\eta}(q)+\Pi_{\phi\phi}(q)\right)+(L_{\eta}^{2}(q)+L_{\phi}^{2}(q))\Pi_{\eta\phi}(q),\]
where
\begin{eqnarray}
L_{\eta}^{LO}(q) & = & -\frac{1}{2}\int\frac{d^{4}p}{(2\pi)^{4}}iMf(p)f(p+q)\left[Tr\left(S_{+}(p)\gamma_{5}S_{+}(p+q)\gamma_{5})\right)+Tr\left(S_{-}(p)\gamma_{5}S_{-}(p+q)\gamma_{5})\right)\right]=\label{eq:P3P0LOLEta}\\
& = & 4iN_{c}\int\frac{d^{4}p}{(2\pi)^{4}}Mf(p)f(p+q)\left[\frac{p^{2}+p\cdot q+\mu_{+}(p)\mu_{+}(p+q)}{\left(p^{2}+\mu_{+}^{2}(p)\right)\left((p+q)^{2}+\mu_{+}^{2}(p+q)\right)}+\frac{p^{2}+p\cdot q+\mu_{-}(p)\mu_{-}(p+q)}{\left(p^{2}+\mu_{-}^{2}(p)\right)\left((p+q)^{2}+\mu_{-}^{2}(p+q)\right)}\right],\nonumber \\
L_{\phi}^{LO}(q) & = & \frac{1}{2}\int\frac{d^{4}p}{(2\pi)^{4}}Mf(p)f(p+q)\left[Tr\left(S_{+}(p)\gamma_{5}S_{+}(p+q)\gamma_{5})\right)-Tr\left(S_{-}(p)\gamma_{5}S_{-}(p+q)\gamma_{5})\right)\right]=\label{eq:P3P0LOLPhi}\\
& =- & 4N_{c}\int\frac{d^{4}p}{(2\pi)^{4}}Mf(p)f(p+q)\left[\frac{p^{2}+p\cdot q+\mu_{+}(p)\mu_{+}(p+q)}{\left(p^{2}+\mu_{+}^{2}(p)\right)\left((p+q)^{2}+\mu_{+}^{2}(p+q)\right)}-\frac{p^{2}+p\cdot q+\mu_{-}(p)\mu_{-}(p+q)}{\left(p^{2}+\mu_{-}^{2}(p)\right)\left((p+q)^{2}+\mu_{-}^{2}(p+q)\right)}\right],\nonumber \end{eqnarray}
and we used identities
\begin{eqnarray}
R_{\eta}^{LO}(q) & = & L_{\phi}^{LO}(q),\label{eq:P3P0LOREta}\\
R_{\phi}^{LO}(q) & = & L_{\eta}^{LO}(q).\label{eq:P3P0LORPhi}\end{eqnarray}
\begin{figure*}
\includegraphics[scale=0.2]{Figures/NLO1}\qquad{}\includegraphics[scale=0.2]{Figures/NLO2}
\includegraphics[scale=0.3]{Figures/NLO3}\qquad{}\includegraphics[scale=0.3]{Figures/NLO4}
\includegraphics[scale=0.3]{Figures/NLO5}\qquad{}\includegraphics[scale=0.3]{Figures/NLO6}
\qquad{}\includegraphics[scale=0.2]{Figures/NLO7}
\caption{\label{fig:P3P0NLO}Contribution to the $\langle P^{3}P^{0}\rangle$-correlator
in the next-to-leading order.}
\end{figure*}
In the next-to-leading order there are seven diagrams shown in the
Figure~\ref{fig:P3P0NLO}. Obviously, only the diagrams 3-6 from
the second and the third row contribute to the residue in pion pole.
The explicit expressions for the corresponding diagrams are given
in Appendix~\ref{sec:appExplicitExpressions}. Using~(\ref{eq:MesonInv_00_33},\ref{eq:MesonInv_03_30}),
one may immediately get
\[
\underset{q^{2}=0}{Res}\Pi_{\eta\phi}(q)\approx-\Pi_{\eta\phi}^{-1}(0)\Pi_{\eta}(0)\underset{q^{2}=0}{Res}\Pi_{\phi\phi}(q),+\mathcal{O}\left(\delta m^{3},m\right)\]
So in evaluation of the residue $\underset{q^{2}=-m_{\pi}^{2}}{Res}\left\langle P_{3}P_{0}\right\rangle \approx\underset{q^{2}=0}{Res}\left\langle P_{3}P_{0}\right\rangle $
one has to keep only the terms
\begin{equation}
\left\langle P_{3}P_{0}\right\rangle =\sum_{i,j=\eta,\phi}L_{i}^{LO}(q)R_{j}^{LO}(q)\Pi_{ij}(q)=L_{\eta}(q)L_{\phi}(q)\Pi_{\phi\phi}(q)+(L_{\eta}^{2}(q)+L_{\phi}^{2}(q))\Pi_{\eta\phi}(q)+non-singulars,\label{eq:ResP3P0Equation}\end{equation}
all the other terms which are not written out explicitly do not contribute
to the residue.
Results of numerical evaluation are presented in Table~\ref{tab:ResP3P0Values}.
As one can see, the model is extremely sensitive to the change of
formfactor. The reasons of such strong dependence will be discussed
in Section~\ref{sec:Conclusion}.
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& \textbf{LO} & \textbf{Mass shift} & \textbf{Mass split} & \textbf{Meson} & \textbf{All NLO} & \textbf{Total}\tabularnewline
\hline
\textbf{QuasiBessel} & & & & & & \tabularnewline
\hline
$-iL_{\eta}$ & $4.07\times10^{-2}$ & $-1.53\times10^{-2}$ & 0 & $-2.33\times10^{-2}$ & $-3.87\times10^{-2}$ & $1.94\times10^{-3}$\tabularnewline
\hline
$L_{\phi}$ & $-0.15\times10^{-3}$ & $-8.03\times10^{-3}$ & $-2.34\times10^{-3}$ & $-6.93\times10^{-3}$ & $-1.73\times10^{-2}$ & $-1.74\times10^{-2}$\tabularnewline
\hline
$l_{7}$ & $0.17\times10^{-4}$ & & & & & $1.198\times10^{-4}$\tabularnewline
\hline
\textbf{Dipole} & & & & & & \tabularnewline
\hline
$-iL_{\eta}$ & $4.35\times10^{-2}$ & $-1.45\times10^{-2}$ & 0 & $-2.38\times10^{-2}$ & $-3.83\times10^{-2}$ & $5.22\times10^{-3}$\tabularnewline
\hline
$L_{\phi}$ & $6.71\times10^{-3}$ & $-1.06\times10^{-2}$ & $-2.10\times10^{-3}$ & $-1.70\times10^{-2}$ & $-2.97\times10^{-2}$ & $-2.30\times10^{-2}$\tabularnewline
\hline
$l_{7}$ & $0.34\times10^{-3}$ & & & & & $1.00\times10^{-3}$\tabularnewline
\hline
\end{tabular}
\caption{\label{tab:ResP3P0Values}Evaluation of the residue $Res_{q^{2}=-m_{\pi}^{2}}\left\langle P_{3}P_{0}\right\rangle $.
See~Eq. (\ref{eq:ResP3P0Equation}) for more details on meaning of
$L_{\eta},L_{\phi}$. The first column is the LO result, columns 2-5
are NLO corrections, column 6 is the total result. In columns 7-8
we give results for $l_{7}$ in LO and NLO (See the Table~\ref{tab:Propagators}
for numbers used in evaluation).}
\end{table}
\section{Conclusion}
\label{sec:Conclusion}
In this paper we evaluated the effects of the current quark mass split
on the dynamical mass, quark condensate and correlator $\left\langle P_{3}P_{0}\right\rangle $.
From these data we extracted the low energy constants $h_{3},l_{7}.$
We found that the dynamical quark mass $\delta M$ is negative, so
as one can see from the left pane of the Figure~\ref{fig:ShapeDep},
the momentum-dependent mass $\delta\mu(p)\equiv\delta m+\delta Mf^{2}(p)$
has different signs for small and large momenta. Due to cancellation
of these contributions, we got very strong sensitivity of all quantities
discussed in this paper to the details of the instanton vacuum model,
such as the shape of instanton (which comes via the formfactor) and
instanton parameters. In the right pane of the Figure~\ref{fig:ShapeDep}
we demonstrate explicitly this fast dependence on the example of the
leading-order integrand of $L_{\phi}(0)$. As it was explained above,
due to different signs of large and small-$p$ contributions, we have
partial (solid line) or almost complete (dashed line) cancellation,
which leads to the strong dependence on parameters of the model. Similar
behaviour is observed for all quantities where the dynamical mass
split $\delta M(p)$ contributes, both in the leading and in the next-to-leading
orders.
It is necessary to note that the instanton vacuum model contains chiral
doublet $(\eta,\vec{\sigma})$-additional degree of freedom which
is absent in the chiral lagrangian, and the cancellation of the different
contributions is due to the dynamics of the field. If we set $-\mathcal{Y}/\mathcal{X}=0$
in (\ref{eq:MassSplitBA}) and thus effectively eliminate the contribution
of the $\sigma_{3}$, we can see that the dynamical mass split $\delta\mu(p)$
is constant for all momenta $p$, and cancellation of different regions
does not happen.
\begin{figure}[h]
\includegraphics[scale=0.4]{Figures/MassDep}\includegraphics[scale=0.4]{Figures/FormDep}
\caption{\label{fig:ShapeDep}Left: dependence of the dynamical mass split
$\delta\mu(p)\equiv\delta m+\delta Mf^{2}(p)$ on the quark momentum
$p$. Right: Instanton shape dependence of the integrand of $L_{\phi}(0)$
in the leading order. $g(p)$ is the integrand of the Eqn~\ref{eq:P3P0LOLPhi}.}
\end{figure}
One of the consequences of the above-mentioned sensitivity of $l_{7}$
to model details is that uncertainty of the instanton vacuum parameters
(average instanton size $\rho$ and inter-instanton distance $R$)
leads to increased uncertainty in the final prediction for $l_{7}$.
As it has been discussed in~\cite{Goeke:2007bj}, different methods
estimate the model parameters are in the range $\rho\sim0.32-0.35$~fm,
$R\sim0.8-1$~fm.While the uncertainty in $\rho,\, R$ is just $\sim10\%$
and is unimportant for most evaluations, for the constant $l_{7}$
it leads to sizeable uncertainty in the final result. Using the Figure~\ref{fig:L7_Rho_R},
we may get for $l_{7}$ an estimate
\begin{equation}
l_{7}\sim(6.6\pm2.4)\times10^{-4}.\label{eq:L7-final-Value}\end{equation}
\begin{figure}
\includegraphics[scale=0.4,angle=-90]{Figures/L7_Rho_r}\caption{\label{fig:L7_Rho_R}{[}Color online{]} Dependence of the constant
$l_{7}$ on the instanton vacuum parameters $\rho$ and $R$}
\end{figure}
The result~(\ref{eq:L7-final-Value}) agrees with a phenomenological
estimate~(\ref{eq:L7-phenom-value}) within uncertainty limits. Using~(\ref{eq:PiMassSplit_QCD}),
we may obtain for the pure QCD contribution to the pion mass difference\[
\left(m_{\pi^{+}}^{2}-m_{\pi^{0}}^{2}\right)_{QCD}\sim1.4\times10^{-5}GeV^{2},\]
i.e. $\sim1\%$ of the experimentally observed difference. This result
does not contradict the well-known fact that the pion mass difference
has electromagnetic origin~\cite{Das:1967it,Langacker:1974nm,Gerstein:1967zz}.
\begin{acknowledgments}
This work was supported in part by Fondecyt (Chile) grant 1090073,
and by the bilateral Funds DFG-436 USB 113/11/0-1 between Germany
and Uzbekistan.
\end{acknowledgments}
|
\section{Introduction}
Over the past few years, a new suite of models for the dark sector has been proposed. They
invoke a vector field which is normally constrained to lie along the time-like direction and
may lead to modifications to the gravitational sector \cite{ZlosnikFerreiraStarkman2007,Maroto2008, DimopoulosEtAl2006,Koivisto:2008xf}. Sometimes called
Einstein-Aether models, they tend to entangle two of the main paradigms currently being
considered: on the one hand modified gravity and on the other dark matter and energy.
Vector field models are attractive because they seem to be able to resolve the problem of
the dark sector (i.e. dark matter {\it and} energy) in a unified way. Most of the emphasis
has gone into constraining vector field models that lead to accelerated expansion
\cite{Maroto2008, DimopoulosEtAl2006, Koivisto:2008xf} although there is a fair amount of work for which the
the vector field leads to a relativistic version of Modified Newtonian Dynamics (MOND)
\cite{Milgrom1983} or can even play the role of dark matter \cite{ZlosnikFerreiraStarkman2007,Zhao2007}.
In fact vector field models seem to incorporate what seems to be a generic feature
of relativistic modified gravity models \cite{Science2009}: that it is impossible to construct
relativistic models that just modify the gravitational sector without introducing new degrees of freedom, which can then behave like either dark matter or dark energy (although for other approaches see, for example, \cite{Milgrom:2009gv,Blanchet:2009zu,Blanchet:2008fj,Soussa:2003sc,Bruneton:2008fk,Sagi:2009kd}).
There has been significant progress in trying to constrain these models. For example, at a fundamental level it
has been shown that a broad class will lead to instabilities and the formation of caustics,
signaling a break down of the fundamental theory \cite{Withers2008}. It has also been shown
that for a general choice of kinetic terms, these theories will be plagued by ghosts or tachyons
\cite{Carroll2009, Eling:2005zq, Lim:2004js}. These pathologies are worrying but do not entirely rule out vector field models-
it has been shown that modifications to the kinetic term, for example, can cure them.
Substantial work has been done on understanding how these fields in these theories behave on macroscopic
scales, either through their interaction with matter to form galaxies and clusters \cite{Chang2008}, or on the largest scales, affecting the growth of structure and its
effect on the CMB \cite{ZlosnikFerreiraStarkman2008}. Indeed for a particular, ``vanilla'' version of the
vector-field model, detailed and definitive constraints have been placed on the
various coupling constants \cite{ZuntzFerreiraZlosnik2008, Jacobson:2008aj}.
In \cite{ZlosnikFerreiraStarkman2008}, it was found that one of the key effects that vectors
would have would be to modify the growth rate of structure. This is not surprising- theories
that modify gravity tend to have this effect. We also found that it lead to a mismatch
between the two gravitational potentials a potentially observable effect \cite{ZhangBean2007}. In
this paper we wish to pursue this analysis and quantify how strong these effects are. Although
we focus on a particular (albeit broad) class of theories, we are interested in extracting
{\it general} lessons from these models. We believe that much of what we learn by looking at these models will shed light on other models of modified gravity (such as, for example, $f(R)$ theories
\cite{Bean2007} and bimetric theories \cite{Bekenstein2004a}).
The structure of this paper is as follows. In Section \ref{theory} we lay out the essential ingredients for a reasonably broad class of vector-like models and its background evolution, and specialize to the form used in the remainder of the paper. In section \ref{linear} we lay out the equations of the perturbed theory, and how we implement them with theoretical constraints. In section \ref{as dark matter} we discuss the problems with modelling dark matter with the theory. In section \ref{as dark energy} we find and constrain the parameters which let the theory act as dark energy. In section \ref{conclusions} we conclude and draw more general lessons about the dark sector.
\section{The Theory}
\label{theory}
\subsection{Theory Definition}
A general action for a vector field $A^{a}$ coupled to gravity can
be written in the form:
\begin{eqnarray}
S=\int d^4x \sqrt{-g}\left[\frac{R}{16\pi G}+{\cal L}(g^{ab},A^{b})\right]
+S_{M} \label{genaction}
\end{eqnarray}
where $g_{ab}$ is the metric, $R$ the Ricci scalar of that metric,
$S_M$ the matter action and $\cal{L}$ is constructed to be generally
covariant and local. By construction $S_M$ only couples to the metric, $g_{ab}$ and
{\it not}
to $A^{a}$.
We will restrict ourselves to consider a Lagrangian that
only depends on covariant derivatives of $A$ and we will consider a unit time-like $A^{a}$. Such a theory can be written
in the form:
\begin{eqnarray}
\label{eq:Lagrangian}
{\cal L}(g^{ab},A^{a})&=&\frac{M^2}{16\pi G}
F(K) +\frac{1}{16\pi G}\lambda(A^a A_a+1)
\\
K&=&M^{-2}K^{ab}_{\phantom{ab}mn}
\nabla_a A^{m}\nabla_b A^{n} \\
\label{tensor}
K^{ab}_{\phantom{ab}mn}&=&c_1g^{ab}g_{mn}
+c_2\delta^a_{\phantom{a}m}\delta^b_{\phantom{b}n}+
c_3 \delta^a_{\phantom{a}n}\delta^b_{\phantom{b}m}
\end{eqnarray}
where $c_i$ are dimensionless constants and $M$
has the dimension of mass. We have removed an additional $c_4$ `acceleration' term in accordance with the transformation described in \cite{ElingJacobson2006}.
As was the case with TeVeS, the constant $G$ may be a different number from
the locally measured value of Newton's gravitational constant. $\lambda$ is a non-dynamical
Lagrange-multiplier
field with dimensions of mass-squared.
The gravitational field equations for this theory, obtained by varying the action with respect to $g^{ab}$ (see \cite{ZlosnikFerreiraStarkman2007}) are
\begin{equation} \label{Einsteineq1}
G_{ab}=\tilde{T}_{ab}+8\pi GT^{\mathrm{matter}}_{ab}
\label{fieldI}
\end{equation}
where the stress-energy tensor for the vector field is given by
\begin{eqnarray}
\tilde{T}_{ab} &=& \frac{1}{2}\nabla_{c}
(F_{K}(J_{(a}^{\phantom{a}c}A_{b)}-
J^{c}_{\phantom{c}(a}A_{b)}-J_{(ab)}A^{c}))
\nonumber \label{lagrangian equation} \\ && -F_{K}Y_{(ab)}
+\frac{1}{2}g_{ab}M^{2}F+\lambda A_{a}A_{b}
\\
F_{K} &\equiv & \frac{dF}{dK}\\
J^{a}_{\phantom{a}c} &=&
(K^{ab}_{\phantom{ab}cd}+
K^{ba}_{\phantom{ba}dc})\nabla_{b}A^{d}
\end{eqnarray}
Brackets around indices denote
symmetrization\footnote[2]{we adopt the convention $X_{(ab)}=\frac{1}{2}(X_{ab}+X_{ba})$, $X_{[ab)}=\frac{1}{2}(X_{ab}-X_{ba})$} and $Y_{ab}$ is the functional derivative
\begin{eqnarray}
Y_{ab} =\nabla_{c}A^{e}\nabla_{d}A^{f}
\frac{\delta(K^{cd}_{\phantom{cd}ef})}{\delta
g^{ab}}
\end{eqnarray}
The equations of motion for the vector field, obtained by varying with respect to $A^{b}$ are
\begin{eqnarray}
\label{eq:veceq1}
\nabla_{a}(F_{K}J^{a}_{\phantom{a}b})
+F_{K} y_{b}&=&2\lambda A_{b}
\label{vectoreom}
\end{eqnarray}
where once again we define the functional derivative
\begin{eqnarray}
y_{b}=\nabla_{c}A^{e}\nabla_{d}A^{f}
\frac{\delta(K^{cd}_{\phantom{cd}ef})}
{\delta A^b}
\end{eqnarray}
Finally, variations of the action with respect to $\lambda$ will fix $A^b A_b=-1$. By inspection, contracting both sides of (\ref{eq:veceq1})
with $A^{b}$ leads to a solution for $\lambda$ in terms of the the vector field
and its covariant derivatives.
These equations allow us to study a general theory of the form presented
in equation \ref{genaction} with a unit time-like vector field.
For our particular, restricted choice
of $K$ we have $Y_{ab}=-c_{1}\left[ (\nabla_{c}A_{a}(\nabla^{c}A_{b})-(\nabla_{a}A_{c})(\nabla_{b}A^{c})\right]$ and $y_b=0$.
\subsection{Background Cosmology}
\label{background}
In this paper we will restrict ourselves to background cosmologies where the spacetime is of the spatially flat Friedmann-Robertson-Walker (henceforth FRW) form:
\begin{equation}
\label{lel}
g_{ab}dx^{a}dx^{b} =-{dt}^{2}+a(t)^{2}\delta_{ij}dx^{i}dx^{j} .
\end{equation}
The energy momentum content of each matter fluid $(i)$ in the background is taken to be of the usual form
\begin{eqnarray}
T^{i}_{ab}=(P^{i}+\rho^{i})u^i_{a}u^i_{b}+P^{i}g_{ab}.
\end{eqnarray}
Where $\rho^{i}$ and $P^{i}$ are the co-moving density and pressure of the $i$th fluid respectively.
We assume that all fluids have identical four velocities $u^{a}=(1,0,0,0)$. Furthermore we define $\rho \equiv \sum_{i}\rho^{(i)}$.
In spacetimes with FRW symmetries, the vector field must align with the direction $\partial_{t}$ and so the vector field is entirely fixed to have components $(1,0,0,0)$ in the co-ordinate system (\ref{lel}). Explicitly, the background value of the scalar $K$ is given by:
\begin{equation}
K_{\mathrm{FRW}}= 3\frac{\alpha H^{2}}{M^{2}}
\label{KFRW equation}
\end{equation}
Where $H\equiv \partial_{t}\ln(a(t))$ and:
\begin{equation}
\alpha \equiv c_{1}+3c_{2}+c_{3}.
\label{alpha definition}
\end{equation}
The Friedmann equation then takes the form \cite{ZlosnikFerreiraStarkman2008}:
\begin{eqnarray}
\label{fried}
\left[1-\alpha K^{1/2}\frac{d}{d K}\left(\frac F{ K^{1/2}}\right)\right]H^2 &=& \frac{8\pi G}{3}\rho \label{00dyn}
\end{eqnarray}
where $\rho$ still includes only the matter components.
We note that the combination of (\ref{00dyn}) and the aether stress energy tensor allows us to write down an effective energy density $\rho_{\mathrm{ae}}$ and pressure $P_{\mathrm{ae}}$ of the aether. From this we may define the fractional energy density $\Omega_{\mathrm{ae}}\equiv 8\pi G\rho_{\mathrm{ae}}/3H_{0}^{2}$, and the aether equation of state parameter $w_{\mathrm{ae}}\equiv P_{\mathrm{ae}}/\rho_{\mathrm{ae}}$:
\begin{eqnarray}
\Omega_{\mathrm{ae}} &=& \frac{M^{2}}{6}\left[\frac{\partial}{\partial H}\left(\frac{F}{H}\right)\right]_{H=H_{0}}
\end{eqnarray}
\begin{equation}
w_{\mathrm{ae}} = -1-\frac{1}{3H^{2}}\frac{\frac{d^{2}}{dtdH}F}{\frac{d}{dH}
\left(\frac{F}{H}\right)}
\end{equation}
The effect of the vector field on the background expansion may be see
as an expansion rate dependent modification to Newton's constant i.e. schematically (\ref{fried}) is an equation of the form $3H^{2}=8\pi G(H^{2})\rho$. It was found in \cite{ZlosnikFerreiraStarkman2008} that various forms of the function $F$ permitted a wide variety of cosmological dynamics:
the presence of the vector field variously leading to accelerated expansion, slowed expansion, rescaling of $G$, and recollapse as summarized in Figure \ref{fig1dyn}
\begin{figure}
\begin{center}
\epsfig{file=fig1.eps,width=8.3cm,height=9.3cm}
\caption{A schematic representation of the types of the late-time background evolution permitted by the functional form $F=\gamma (-K)^{n}$ as a function of $(n,\gamma)$ for $n<1$.}
\label{fig1dyn}
\end{center}
\end{figure}
\subsection{The functional form of $F(K)$ in cosmology}
\label{F}
We must now specify the form of the function $F(K)$ in equation (\ref{lagrangian equation}). There is an obvious set of candidates here - we could attempt to be directly consistent with MOND and use the same branch of $F$ as it uses on small scales. In appendix \ref{mond regime bad} we show that doing so would make it impossible to consistently generate late-time acceleration behavior in the background cosmology. Instead, we will use a simple and reasonable ansatz that works for the regime $|K|\gg1$ that we consider here.
Existing functional forms for F(K) in the MOND regime typically are dominated for a single monomial term for $|K|\gg1$ (see for instance \cite{Famaey:2005fd}) and so it seems reasonable to restrict the function to take this form:
\begin{eqnarray}
F = \gamma (-K)^{n_{\mathrm{ae}}} \ \ &K&<0\nonumber \\
F = \gamma (+K)^{n_{\mathrm{ae}}} \ \ &K& >0
\label{kinetic equation}
\end{eqnarray}
where ${n_{\mathrm{ae}}}\leq 1$. This form has sufficient power to express a wide variety of behavior, and the parameters $\gamma$ and $n$ shall be central to our further analysis.
\section{Linear perturbation Theory}
\subsection{Formalism \& Theory}
\label{linear}
We have seen that the vector field can have a significant effect on the quasistatic, weak field limit and the background cosmological geometry.
Significant evidence for non-baryonic mass persists on the largest cosmological scales \cite{DodelsonLiguori2006} therefore it is vital that a relativistic theory of MOND can account for this. As mentioned, it has been argued \cite{DaiMatsuoStarkman2008b} that even in the quasistatic, weak field limit, a spatial tilt to the vector field may produce significant deviations under some circumstances from the local MOND force law.
Similarly it was shown \cite{ZlosnikFerreiraStarkman2008} that in the context of linear cosmological perturbations the energy density associated with the projection of the vector field onto surfaces of constant conformal time could, to a degree, act as a cosmological `dark matter' .
In this paper we shall comprehensively address the question of whether the Lagrangian (\ref{eq:Lagrangian}) represents a viable model of the dark sector in light of precision cosmology.
In the main body of the paper we will consider scalar perturbations. In Appendix \ref{vt} we derive the equations of motion for the vector field's two divergenceless vector modes as well as the gravitational wave tensor modes (the speed of propagation of which is in general modified by the vector field). Requiring the stability of these modes puts constraints on (\ref{eq:Lagrangian}) (see Section \ref{cons}).
We shall work in the synchronous gauge (see for instance \cite{Ma:1995ey}) and so the metric takes the following form:
\begin{equation}
g_{\mu\nu}dx^{\mu}dx^{\nu}=-a^{2}d\tau^{2}+a^{2}[\gamma_{ij}+h_{ij}]dx^{i}dx^{j}
\end{equation}
where $\tau$ is conformal time, $\gamma_{ij}$ is a spatially flat spacelike 3-metric perturbed by $h_{ij}$ which is built from two scalar potentials $\eta$ and $h$:
\begin{equation}
h_{ij}(\textbf{x},\tau)= \int d^{3}k e^{i\textbf{k}\cdot\textbf{x}}[\hat{\textbf{k}}_{i}\hat{\textbf{k}}_{j}h(\textbf{k},\tau)+(\hat{\textbf{k}}_{i}\hat{\textbf{k}}_{j}-\frac{1}{3}\delta_{ij})6\eta(\textbf{k},\tau)]
\end{equation}
Similarly we will expand the aether field as:
\begin{equation}
A^{\mu}=\frac{1}{a}(1,\partial_{i}V)
\end{equation}
The zeroth component of the aether field is, by virtue of the gauge choice and the constraint,
fixed as equal to $1$ up to second order in perturbations.
A slight complication in the field equations arises because of the presence of the function $F$ which depends nonlinearly on the scalar $K$. We assume that for modes of interest one may consistently regard the perturbation to $K$ as being much less than unity. Thus one can expand $K$ as $K = K^{0}+ K^{\epsilon}$ and $F$ as $F= F^{0}+F_{K}^{0}K^{\epsilon}$ where $K^{\epsilon} \ll 1$ where the superscript $0$ denotes the quantity corresponding to a function's background value. Explicitly we have that:
\begin{eqnarray}
K^{0} &=& 3\frac{\alpha {\cal H}^{2}}{a^{2}M^{2}}\\
{\cal H}K^{\epsilon} &=& -\frac{2}{3}K^{0}(k^{2}V-\frac{h'}{2})
\end{eqnarray}
Where primes denote derivatives with respect to conformal time and we have used the conformal Hubble parameter ${\cal H}\equiv H/a$.
We also make use of the following identity:
\begin{eqnarray}
(K^{0})'{\cal H}=-4K^{0}({\cal H}^{2}-\frac{1}{2}\frac{a''}{a})
\end{eqnarray}
Henceforth we will drop the superscripts on $K^{0}$ and $F^{0}$ i.e. $K$ and $F$ shall be assumed to represent \emph{background} values of the fields.
Towards simplifying the form of the equations, we will will rather use the field $\xi$ instead of $V$, where $\xi$ is defined as:
\begin{equation}
\xi \equiv V-\frac{1}{2k^{2}}(h+6\eta)'
\end{equation}
For further compactness of notation we define the variables:
\begin{eqnarray}
\hat{\alpha} &\equiv &\left(1+2\frac{F_{KK}}{F_{K}}K\right)\alpha \\
\hat{c}_{1} &\equiv & \left(1+2\frac{F_{KK}}{F_{K}}K\right)c_{1}
\end{eqnarray}
The vector field equation of motion (\ref{eq:veceq1}) becomes :
\begin{eqnarray}
0 &=& c_{1}(1+c_{13}F_{K})\frac{(F_{K}\xi')'}{F_{K}}+2{\cal H}c_{1}(1+c_{13}F_{K})\frac{(F_{K}\xi)'}{F_{K}} \nonumber \\
\nonumber &&+ [2c_{1}(1+c_{13}F_{K})(\frac{a''}{a}-{\cal H}^{2})+2(\hat{c}_{1}+\hat{\alpha})({\cal H}^{2}-\frac{1}{2}\frac{a''}{a})\\
\nonumber &&+c_{1}c_{13}(F_{KK}K')'+\frac{1}{3}(\hat{\alpha}+2c_{13})k^{2}]\xi \\
\nonumber && +(c_{1}+\hat{\alpha})\eta'+(\hat{c}_{1}+\hat{\alpha})\frac{1}{k^{2}}({\cal H}^{2}-\frac{1}{2}\frac{a''}{a})(h'+6\eta')\\
&& -\frac{3}{2}\frac{c_{1}}{k^{2}}\frac{(F_{K}\Sigma_{f})'}{F_{K}} \label{eq:veceq}
\end{eqnarray}
The relevant Einstein equations are:
\begin{eqnarray}
\label{eq:ein1}
(1-\frac{1}{2}\hat{\alpha} F_{K})k^{2}\eta' &=& 4\pi G a^{2}
i k^{j}\delta T^{0}_{\phantom{0}j} \nonumber \\
&& +\frac{1}{6}k^{4}(\hat{\alpha}+2c_{13})F_{K}\xi
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:ein2}
(1+\frac{1}{2}c_{1}F_{K})({\cal H}h'-2k^{2}\eta) &=& -8\pi Ga^{2}\delta T^{0}_{\phantom{0}0} \nonumber\\
\nonumber && -\frac{1}{2}F_{K}(c_{1}+\hat{\alpha})6{\cal H}\eta'\\
\nonumber && -2\alpha F_{KK}K{\cal H}k^{2}\xi\\
\nonumber && +\frac{F_{K}c_{1}k^{2}}{a^{2}}\left( a^{2}(1+c_{13}F_{K})\xi \right)'\\
&& -\frac{3}{2}c_{1}F_{K}\Sigma_{f}
\end{eqnarray}
Where we have used the fact that:
\begin{eqnarray}
\label{shear}
(h+6\eta)''+2{\cal H}(h+6\eta)'-2k^{2}\eta &=& -3\Sigma_{f} \nonumber \\
&&+2c_{13}k^{2}[F_{K}(2{\cal H}\xi+\xi')\nonumber \\
&& +F_{KK}K'\xi]
\end{eqnarray}
The functions $\delta T^{0}_{\phantom{0}j}$ and $\delta T^{0}_{\phantom{0}0}$ are the first order perturbations to the corresponding components of the matter fields' stress energy tensors. Summation over field species is assumed. The field $\Sigma_{f}$ is the scalar component of the total fluid shear i.e. $\Sigma_{f}=-8\pi G a^{2}(\hat{k}_{i}\hat{k}^{j}-\frac{1}{3}\delta^{j}_{\phantom{j}i})\Sigma^{i}_{\phantom{i}j}$ and $\Sigma^{i}_{\phantom{i}j}\equiv \delta T^{i}_{\phantom{i}j}-\frac{1}{3}\delta^{i}_{\phantom{i}j}\delta T^{k}_{\phantom{k}k}$. A gauge invariant formulation of the theory's equations may be found in \cite{LiMotaBarrow2007}.
\subsection{Parameter Constraints}
\label{cons}
We can immediately see a number of constraints on the $c_{i}$ and the form of the function $F$. From (\ref{eq:veceq}) it can be shown that for in the limit of timescales shorter than a Hubble time the quantity:
\begin{equation}
C_{S}^{2} = \frac{2}{3}\frac{(\frac{\hat{\alpha}}{2}+c_{13})}{c_{1}(1+c_{13}F_{K})}
\end{equation}
can be interpreted as the squared sound speed of the field $\xi$. The avoidance of exponentially growing subhorizon modes dictates that $C_{S}^{2}$ should be positive definite.
Similarly, one may consider the field equations of the two divergenceless `vector' modes of the vector field and the two transverse traceless `tensor' modes of of the metric. Each respectively has a squared sound speed function (named $C_{V}^{2}$ and $C_{T}^{2}$ respectively) which, as in the scalar case, should be positive definite. These functions are calculated in Appendix \ref{vt} and are as follows:
\begin{equation}
C_T^2=\frac{1}{1+c_{13}F_{K}}
\end{equation}
\begin{equation}
C_V^2=\frac{F_{K}}{2c_1}\frac{2c_1+F_{K}(c_1^2-c_3^2)}{1+c_{13}F_{K}}
\end{equation}
Note that the gravitational wave tensor modes now generically have a time dependent speed of propagation.
Collectively the three positivity constraints imply the following constraints on the $c_{i}$ and function $F$ parameter space:
\begin{eqnarray}
1+F_{K} c_{13} & > & 0\\
(\hat{\alpha}+2c_{13})/2c_{1} & > & 0 \\
F_{K}(1+F_{K}\frac{c_1^2-c_3^2}{2c_1})& > & 0
\end{eqnarray}
Now we turn to the Einstein equations (\ref{eq:ein1}) and (\ref{eq:ein2}). Terms in the perturbed vector field stress energy tensor $\delta \tilde{T}_{ab}$ may contain terms proportional (up to a time-dependent function of the background fields) to $\delta G_{ab}$ and indeed this can be seen in (\ref{eq:ein1}) and (\ref{eq:ein2}) through the appearance of terms proportional to the $c_{i}$ on the left hand sides of the equations.
If the effect $\xi$ is negligible then the effect of the aether stress energy terms proportional to components of the Einstein tensor may always be absorbed into a redefinition of Newton's constant as a time-dependent effective gravitational coupling. Thus, even if $\xi$ has comparatively little effect, there may be a considerable modification to the link between the matter fields and the gravitational field.
The resulting gravitational couplings should be greater than or equal to zero otherwise the gravitational field will interpret normal matter as violating energy conditions, and so risking the appearance of instabilities. This restriction implies the following constraints:
\begin{eqnarray}
(1-\frac{1}{2}\hat{\alpha}F_{K})> 0
\end{eqnarray}
\begin{eqnarray}
(1+\frac{1}{2}c_{1}F_{K}) > 0
\end{eqnarray}
Throughout our analysis, we will only consider regions of the model's parameter space which satisfy these constraints.
\subsection{Computation}
\label{computation}
To study the effects of the vector field in detail we have modified the structure formation Boltzmann code {\sc CMBEASY} \cite{Doran:2003sy}. We add a Newton-Raphson solver for the Hubble parameter, with added aether components. The perturbation evolution is also modified to include the aether components $\xi$ and $\xi'$, and their contribution to the density, pressure and shear perturbations. We also include the altered metric perturbations in the calculation of the CMB source function. We use adiabatic initial conditions\cite{ZlosnikAetherModes_preparation}. Since Boltzmann codes are very highly optimized for $\Lambda$CDM models, care must be taken to ensure that modifications are performed in a consistent manner - for example an unmodified Friedmann equation is often assumed for computational efficiency.
To explore parameter spaces, {\sc CMBEASY} is coupled to a Monte-Carlo Markov Chain (MCMC) engine \cite{Doran:2003ua}. We extended this engine to include our new Aether parameters: $\Omega_{\mathrm{ae}}$, $c_+=(c_1+c_3)$, $c_-=(c_1-c_3)$, $C_S^2$, $n_{\mathrm{ae}}$ and $M$. We include the full ranges of these parameters by allowing the kinetic term to take two branches, for positive and negative $K$ as in (\ref{kinetic equation}).
We constrain models using both CMB data from the WMAP experiment \cite{Komatsu:2008hk} and large-scale structure from the SDSS survey \cite{SDSS:2007wu}, though not always at the same time. As we shall see, some Aether models are extremely poor fits to the combined data sets; to illustrate such problems we want to find models that fit only the large scale structure data. In regimes where fits are extremely poor, MCMC does not work particularly well. We ameliorate such situations by running larger numbers of shorter Markov chains and re-starting from their best-fit positions, and sometimes by abandoning MCMC altogether and simply performing random searches for good parameter combinations.
\section{The vector field as Dark Matter}
\label{as dark matter}
\subsection{Background evolution \& doppler peak positions}
\label{background subsection}
An unusual property of the model considered here is that at the level of cosmological perturbations the field can mimic a perturbed pressureless fluid in the formation of large scale structure whilst behaving entirely differently in the cosmological background \cite{ZlosnikFerreiraStarkman2008} . In this paper we would like to consider not just large scale structure but also other cosmological probes. The anisotropies in the CMB temperature are sensitive to the background dynamics and perturbed dynamics of source of cosmic mass discrepancies in a largely distinguishable manner.
If the aether plays the role of dark matter only in the perturbations, then the background expansion is dark energy-dominated at an earlier time. For a given $H_0$ this reduces the expansion rate of the universe between recombination and now and so decreases the angle subtended on the sky by given distance at last scattering. This moves the CMB doppler peaks to higher $\ell$. Although the aether can give a suitable time dependence to the effective gravitational coupling $G$ in the Friedmann equations so as to yield the same expansion rate as $\Lambda$CDM, the necessary functional forms of $F$ are extremely contrived. For instance these forms essentially contain a new constant scale roughly equal to the Hubble parameter at matter-radiation equality. Even in the event of such a construction, it may not be possible for the squared sound speed of vector field perturbations $C_{S}^{2}$ to remain sufficiently small as to behave like cold dark matter in perturbations \cite{ZlosnikFerreiraStarkman2007}.
The other input to the peak position is the sound horizon at last scattering. The physics of this is sufficiently robust that changes compensatory to the alteration in the distance to last scattering are not feasible without exceptional fine tuning of the perturbational behavior of the aether. Indeed, it has been argued that only additional components that behave like non-relativistic matter in the background might fix this problem \cite{Ferreira:2008ma}.
Could there indeed be such a non-relativistic matter in the background which allows acoustic peak positions which are consistent with the data whilst leaving the aether in its role as the seed of the formation of large scale structure? A appropriate candidate would appear to be massive neutrinos. A suitable mass of such particles so as to account for the right effective contribution to the dust component of the background typically implies that the neutrinos themselves are unable to clump on small enough scales so as to be a good candidate for all of the dark matter.
Such a solution as has been proposed in \cite{SkordisEtAl2006} and \cite{Angus2008b}.
\subsection{Perturbation evolution}
We now argue that even with the inclusion of massive neutrinos, one generically expects the aether to have an unacceptable influence on the large scale CMB anisotropy if it is to also play the dominant role in structure formation.
The first requirement for successful perturbation evolution is that structure can form at all. One necessary condition for this is that the sound speed of the structure seed not be too large, since this would wash out structure. We require that the sound horizon in the model be less than the smallest scales where structure can form linearly: $C_S k_{\mathrm{max}} \tau \lesssim 1$, where $k_\mathrm{max} \sim 0.2 h/M\mathrm{pc}$. For matter power observations at $\tau \sim 3\times 10^{4}$, the present epoch, this yields $C_S \lesssim 10^{-4}$.
There are two underlying physical processes that further constrain the models.
The first is a change to the growth rate of perturbation amplitude. This can cause discrepancies between the amplitudes we expect in the matter power spectrum and the CMB, since the evolution between the two is different. It can also lead to an integrated Sachs-Wolfe (ISW - see below) effect during the matter era since $\Phi$ will accrue a time dependence.
The second is the increased presence of a $\Phi-\Psi$ metric shear. This also leads (directly) to a matter era ISW.
\subsection*{Observable 1: ISW}
\label{ISW subsection}
Under the assumption of adiabaticity we have that
the anisotropy in the CMB, $\Delta T({\hat n})/T$ on large scales in a given spatial direction
${\hat n}$ is given by
\begin{equation}
\label{isw equation}
\frac{\Delta T({\hat n})}{T}\simeq -\frac{1}{3}\Psi(\tau_*,d_*{\hat n}) -
\int_{\tau_*}^{\tau_0}d\tilde\tau({\Psi'}+{\Phi'})[\tilde\tau,(\tau_0-\tilde\tau){\hat n}]
\end{equation}
where $\tau_*$ is the conformal time of last scattering, $\tau_{0}$ is the conformal time today, $d_*$ is the comoving radius of the surface of last scattering, and $\Phi$ and $\Psi$ are the conformal Newtonian gauge gravitational potentials. The integral in (\ref{isw equation}) is the integrated Sachs-Wolfe (ISW) effect. Writing the integrand of (\ref{isw equation}) as $\Psi' + \Phi' = -(\Phi-\Psi)' + 2\Phi'$, we can see this as time derivatives of a shear part and a growth rate part.
In the standard cosmological model, the field $\Phi$ has negligible time dependence during matter domination. It gains a time dependence only when the background starts accelerating, and only then can the resulting growth rate ISW contribution be considerable. In the model considered here, the situation may be rather different. It was found in \cite{ZlosnikFerreiraStarkman2008} that substantial contributions to the ISW may occur even during the matter era. For the aether field to seed structure formation the field $\xi$ must have a suitable growing mode solution in the matter era. Typically the corresponding spatial curvature perturbation $k^{2}\Phi$ will then have a time dependence via the Poisson equation.
The shear part can also gain a time dependence in the aether model, which in the $\Lambda$CDM is very small even during acceleration.
Each of these effects depend on the functional form of $F$, the time-dependence of the $\xi$ growing mode and the choice of the parameters $c_{i}$. It is extremely challenging to find combinations of the parameters which allow for a realistic growth of structure whilst making the ISW acceptably low. This is most easily illustrated by considering the theory TeVeS \cite{Bekenstein2004a} which has many of the same properties as the model considered here. It may be shown that TeVeS can be written as a single metric theory with a timelike vector field of unfixed norm \cite{ZlosnikFerreiraStarkman2006},\cite{Skordis:2009bf}. As in the model considered here, the longitudinal component of the vector field can source the growth of structure \cite{SkordisEtAl2006}, \cite{DodelsonLiguori2006}. We will call this field $V_{T}$. We consider a matter dominated era where $V_{T}$ is responsible for the dominant source in the Poisson and shear equations. These equations respectively then are\cite{Skordis2006}:
\begin{eqnarray}
\label{eq:tevpo}
k^{2}\Phi & \approx& -f_{s}{\cal H}k^{2}V_{T}
-\frac{K_{B}}{2}k^{2}V_{T}' \\
k^{2}(\Psi-\Phi) & \approx& f_{s}k^{2}
(2{\cal H}V_{T}-V_{T}') \\
f_{s}(\tau) &\equiv & \frac{(1-\bar{A}^{4})}{\bar{A}^{4}}
\end{eqnarray}
where $K_{B}$ is a positive constant of the action and $\bar{A}^{4}$ is the norm of the vector squared again (equal to unity in the fixed norm case, but in TeVeS the deviation of $|\bar{A}^{2}|$ from unity is essentially the background variation of the `scalar field' degree of freedom). Therefore:
\begin{eqnarray}
\label{eq:iswi}
\Psi+\Phi = -(f_{s}(\tau)+K_{B})V'_{T}
\end{eqnarray}
In this era the vector field equation is:
\begin{equation}
V_{T}''+b_{1}\frac{V_{T}'}{\tau}+b_{2}\frac{V_{T}}{\tau^{2}}=S[\Phi,\Psi]
\end{equation}
where
\begin{eqnarray}
b_{1} &=& 2(3-\bar{A}^{4})\\
b_{2} &=& 2(2-\bar{A}^{4})+\frac{12\bar{A}^{4}}{K_{B}}(1-\bar{A}^{4})
\end{eqnarray}
and $S[\Phi,\Psi]$ is a source term.
For this situation to arise, there must be a growing mode in $V_{T}$ \cite{DodelsonLiguori2006}. Therefore we require that $b_{2}<0$.
The function $\bar{A}^{4}$ will be rather close to unity \cite{BourliotEtAl2006} so
\begin{equation}
b_{2}\sim 2+12 f_{s}(\tau)/K_{B}
\end{equation}
Therefore we require that $f_{s}(\tau)/K_{B} \lesssim -1/6 $ (recalling that $K_{B}$ is a positive number). By (\ref{eq:iswi}) we see that the contribution to the ISW will be proportional to $f_{s}(\tau)+K_{B}$. Meanwhile we see from (\ref{eq:tevpo}) that the comparative contribution to Poisson's equation is also independently weighted by $f_{s}$ and $K_{B}$. If there is no time dependence in the Poisson equation (\ref{eq:tevpo}) due to the vector field then the vector field will make no contribution to the ISW effect. If there is a time dependence in the Poisson equation, there may yet be no contribution to the ISW if $f_{s}(\tau)\sim -K_{B}$ between last scattering and now (though the overdensities of the baryon field will then not generally grow as $a$, thus contributing to the ISW effect). This would be consistent with the condition for a growing mode in $V_{T}$ but it does not guarantee that the resulting growing mode would be suitable.
Indeed, it was found \cite{SkordisEtAl2006} that in seeding the growth of large scale structure in TeVeS there was a significant danger of incurring unacceptably high temperature anisotropies in the CMB on large scales. Although involving a larger number of terms, the same reasoning carries over to the model considered here - i.e. parameters which allow realistic structure formation will typically lead to an unacceptable ISW effect. This is vividly illustrated in Figure \ref{dark matter spectra} which shows best fit models as compared to SDSS large scale structure data. In every case, the corresponding temperature anisotropy displays a dramatically poor fit to the data at low $\ell$.
\subsection*{Observable 2: Amplitudes}
\label{amplitudes section}
The ratio of the observed amplitudes of the CMB anisotropy and the matter power spectrum is consistent with a growth rate proportional to $a$ (though see \cite{Bean2009}). Any uncompensated change to this growth rate in the aether model over this period would lead to a different ratio. The bias parameter between the galaxy distribution and the underlying density field can be used to rectify this difference, but only if the change is relatively small and unphysically large bias parameters (larger than $\sim10$) are not required.
\subsection{Summary}
We have seen that although the vector field may play a number of the roles that dark matter plays, it seemingly cannot do all at once \footnote[1]{It is interesting to compare this with similar results found in \cite{Li:2008aia} for a framework for generalizations of dark matter. It is not clear whether the model considered in this paper fits within this framework}.
The position of the acoustic peaks in the CMB temperature anisotropy should be taken as a strong indication of additional nonbaryonic nonrelativistic mattter present in the universe during matter domination. Such an effect can be achieved in this model by a rescaling of the value of Newton's constant. However, this rescaling cannot persist into the radiation era \cite{Carroll:2004ai}. Thus, the functional form would have to be approaching the rescaling solution only after one would expect matter (including cold dark matter)-radiation equality to happen.
This implies the presence of a new scale in the theory, roughly corresponding to the Hubble parameter $H_{eq}$ at this time. It seems fair to say in general that a model such as that considered here, is more likely to be a cosmologically viable candidate for dark matter if the scale $M$ in the theory is closer to $H_{eq}$ and not $H_{0}$. It is tempting to speculate
whether a theory where the scale $M$ itself is dynamical may find more success, but that will not be explored in this work.
Furthermore we have seen that even if the background is consistent with observations, the effect on the evolution of perturbations may be unacceptable, notably either through the ISW effect or comparing the respective amplitudes of the CMB anisotropy and matter power spectrum today.
\subsection{Example Problem Spectra}
It has previously been shown that the Einstein Aether can produce acceptable matter power spectra with certain parameter combinations \cite{ZlosnikFerreiraStarkman2008}. Here we show that such combinations do not provide an acceptable fit to CMB measurements. Despite extensive searches we have been unable to find any parameter set within the model that does fit the WMAP data well; this is entirely in line with the problems discussed above.
We use a parameter set which is consistent with BBN limits on $\Omega_B h^2$ and the HST key project measurement of $H_0$. The standard cosmological parameters are: $\Omega_b h^2=0.0193$, $n_s=0.83$, $H_0=89.3\, \mathrm{km/s/Mpc}$, $\Omega_c h^2=0$. The new Aether variables are $c_+=-4.72$, $c_-=-6.11$, $n_{\mathrm{ae}}=0.34$, $\Omega_{\mathrm{ae}}=0.82$, $M_{\mathrm{ae}}=111.3\,\mathrm{km/s/Mpc}$ with $c_2$ set by requiring zero sound speed. This parameter combination is in no sense optimal, but it does provide an illustration of all the problems that arise here.
Figure \ref{dark matter spectra} shows power spectra from our modified Boltzmann code for this parameter set. The matter power is a realistic fit to the SDSS data (this was the criterion for our choice of parameters). The CMB spectra shows various problems. In the low-$\ell$ regime a large ISW effect is clearly present, destroying the fit at large scales, as described in section (\ref{ISW subsection}). The positions of the peaks are poorly fit by the model, as expected and discussed in section (\ref{background subsection}). Finally, in the plot we have rescaled the amplitude of the matter power spectrum by a factor $0.02$, corresponding to a galaxy bias of $0.14$ in order to reconcile the relative amplitudes of the two spectra with the data; such a scaling is unphysically small. This corresponds to the changed growth rate described in section (\ref{amplitudes section}). All these effects cause severe problems when attempting to simultaneously fit the CMB and large scale structure.
Figures \ref{isw1} and \ref{isw2} illustrate the sources of the extreme ISW effects shown in Figure \ref{dark matter spectra}; the time derivatives of the metric quantities plotted create an ISW effect as shown in equation (\ref{isw equation}). The onset of background acceleration in each case is marked by a turnover in the curves at late time. The GEA universe exhibits a dramatically increased $|\Phi-\Psi|$ and time dependence of $|\Phi|$ during the matter era as compared to the $\Lambda$CDM universe. Although the $|\Phi-\Psi|$ has a smaller magnitude its time dependence can be significant for the total ISW effect. Note that values of $\tau$ between the two universes do not correspond to the same physical time or redshift since the universes expand at different rates.
\begin{figure}
\begin{center}
\epsfig{file=isw_k001.eps,width=8.3cm,height=9.3cm}
\caption{Exotic behavior of metric potentials for $k\sim$ $10^{-2}$ $Mpc^{-1}$. The panels show the fields $|\Phi|$ and $|\Phi-\Psi|$ for a $\Lambda$CDM universe (green solid line) and GEA universe (dashed blue line) as a function of $k(\tau-\tau_r)$ where $\tau_r$ is the conformal time of recombination.
}
\label{isw1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsfig{file=isw_k00001.eps,width=8.3cm,height=9.3cm}
\caption{Equivalent of Figure \ref{isw1} for k $\sim$ $10^{-4}$ $Mpc^{-1}$}.
\label{isw2}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsfig{file=dark_matter_spectra.eps,width=8.3cm,height=9.3cm}
\caption{Matter power (top) and CMB (bottom) power spectra for the $\Lambda CDM$ (dashed green) and typical GEA (solid red) models, with WMAP and SDSS constraints.
}
\label{dark matter spectra}
\end{center}
\end{figure}
\section{The vector field as Dark Energy}
\label{as dark energy}
\subsection{Dark Energy Regime}
As shown in figure \ref{fig1dyn}, our vector field can produce late-time acceleration and so play the role of dark energy. Indeed, for the form of the vector field used here, as the index $n_{\mathrm{ae}}\rightarrow0$ the theory becomes the same as a cosmological constant for both the background and the perturbations. Since the model can fit the data well we can use our MCMC engine to find constraints on the parameters of the vector field, telling us exactly how close to the $n_{\mathrm{ae}}=0$ cosmological constant case the theory must be to fit the CMB and LSS data.
If the model can fit the data only extremely close to $n_{\mathrm{ae}}=0$ then it does not provide a compelling alternative to the cosmological constant. If, on the other hand, there is significant flexibility in the model and no fine tuning, or if it can provide a better fit than $\Lambda$CDM, then it is somewhat more interesting.
In this section we will consider the resulting background evolution, CMB temperature anisotropy, and matter power spectrum for a universe containing the vector field, cold dark matter, the conventional matter fields, and no cosmological constant. The acceleration will arise solely from the vector field's modification to the Friedmann equation.
\subsection{Constraints on aether dark energy from data}
Our MCMC generated the constraints on the vector field parameters shown in figures \ref{c1c2c3 constraint} to \ref{alpha constraint}; these curves are the smoothed histograms from our combined Markov chains.
\begin{figure}
\begin{center}
\epsfig{file=c1_c2_c3.eps,width=8.3cm,height=5.12cm}
\caption{Constraints on the three coupling terms of the theory.
}
\label{c1c2c3 constraint}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsfig{file=nae.eps,width=8.3cm,height=5.12cm}
\caption{Constraints on the kinetic term power law index parameter.
}
\label{nae constraint}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsfig{file=cs2.eps,width=8.3cm,height=5.12cm}
\caption{Constraints on the vector field sound speed parameter.
}
\label{cs2 constraint}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsfig{file=alpha.eps,width=8.3cm,height=5.12cm}
\caption{Constraints on the parameter $\alpha=c_1+3c_2+c_3$
}
\label{alpha constraint}
\end{center}
\end{figure}
The most important trend evident in these results is closeness of $n_{\mathrm{ae}}$ to the $\Lambda$ value of zero. We find the best fit value $n_{\mathrm{ae}}=4.2\cdot10^{-3}$, with a 95\% upper limit $n_{\mathrm{ae}}<0.126$.
The best fit value in the MCMC run is very close to the $\Lambda$CDM likelihood of the same data, at the cost of six extra parameters, meaning that it is unlikely to be favored by any model comparison exercise. It does, however, demonstrate the validity of modified gravity-related dark energy candidates.
Having obtained these constraints we can determine their origin. There are two ways in which the vector field must behave like $\Lambda$ to provide a good fit. The first is that the late-time acceleration should be close to that given by $\Lambda$. The second is that any perturbations in the field (which are not present in $\Lambda$) should not affect the observable spectra.
\subsection{Constraint origins - acceleration rates}
The consistency of the acceleration of the universe with the cosmological constant equation of state $w=-1$ is being measured with increasing precision in supernova and baryon acoustic oscillation experiments (which are beyond the scope of this paper). Here, they will be constrained by the late-time ISW effect induced by dark energy, and by the perturbation growth rate.
We can assess how closely vector-induced acceleration mimics $\Lambda$-driven expansion at late times with the equation of state $w_{\mathrm{ae}}$ of the vector field in the background:
\begin{equation}
w_{\mathrm{ae}} = -1-\frac{1}{3H^{2}}\frac{\frac{d^{2}}{dtdH}F}{\frac{d}{dH}
\left(\frac{F}{H}\right)}.
\end{equation}
For the monomial form of $F(K)$ we have that:
\begin{eqnarray}
\label{waez}
w_{\mathrm{ae}} = -1 -\frac{2n}{3(2n-1)}\frac{\dot{H}}{H^{2}}
\end{eqnarray}
Thus the equation of state will generically deviate from $-1$ whenever $n \neq 0$ and so the acceleration for these values will not be degenerate with a cosmological constant. We see immediately from equation (\ref{waez}) that $w_{\mathrm{ae}}(\tau) < -1$ for $ 0 < n < 1/2$ and $w_{\mathrm{ae}}(\tau)> -1 $ for $n> 1/2$. This is clearly visible in Figure \ref{wvsz}.
\begin{figure}
\begin{center}
\epsfig{file=w_varying_nae.eps,width=8.3cm,height=5.12cm}
\caption{The vector field's equation of state as a function of redshift $z$, for various values of the kinetic term index $n_{\mathrm{ae}}.$
}
\label{wvsz}
\end{center}
\end{figure}
\subsection{Constraint origins - perturbation evolution}
Even if the background expansion is rather close to the the $\Lambda$CDM model, the evolution of perturbations need not be.
This is most easily illustrated by considering the Poisson equation on large scales (see \cite{ZlosnikFerreiraStarkman2007} for a derivation). On these large scales there is a time-dependent re-scaling of the the metric perturbation $\Phi$, which we can cast as a modification of the effective gravitational constant $G$:
\begin{eqnarray}
k^{2}\Phi &=& -4\pi G^{(1)}_{\mathrm{eff}} a^{2} \sum_{i}\bar{\rho}_{i}\delta_{i} \\
G^{(1)}_{\mathrm{eff}} &\equiv& \frac{G}{1+\frac{c_{1}}{2}F_{K}}
\end{eqnarray}
where we have assumed that terms proportional to the velocity divergence are ignorable and provisionally considered the effect of the field $\xi$ to be subdominant.
The Friedmann equation may be used to cast the above equation in a more familiar form by eliminating he background $\rho_{i}$ in favour of background expansion rate of the universe and the time-dependent fractional energy density $\Omega_{i}(\tau)\equiv 8\pi G \rho(\tau)/(3H(\tau)^{2})$. This yields:
\begin{eqnarray}
k^{2}\Phi &=& -\frac{3}{2}{\cal H}^{2}\frac{G^{(1)}_{\mathrm{eff}} }{G^{(0)}_{\mathrm{eff}} } \sum_{i}\Omega_{i} (\tau)\delta_{i}
\end{eqnarray}
where
\begin{eqnarray}
G^{(0)}_{\mathrm{eff}} &\equiv& \frac{G}{1-\alpha K^{\frac{1}{2}}\frac{d}{dK}\left(\frac{F}{K^{\frac{1}{2}}}\right)}
\end{eqnarray}
The $n=0$ $\Lambda$CDM Poisson equation may be cast in this form by taking $G^{(1)}_{\mathrm{eff}}= G$ and $G^{(0)}_{\mathrm{eff}}=G/(1-\Lambda/(3H^{2}))$.
For the case where $n$ differs from $0$, the function $G^{(1)}_{\mathrm{eff}}$ will generically possess a time dependence during the background evolution. Therefore the link between the time evolution of the functions $G^{(0)}_{\mathrm{eff}}$, $\delta_{i}$ and $\Phi$ will differ from the case where acceleration is caused by a cosmological constant. We may thus expect the ISW effect to be of a non-standard form. This is vividly illustrated in Figure \ref{nae} where it can be seen that for a given set of $(c_{i},\gamma,M)$, variation of $n$ results in a considerable variation in the large scale CMB temperature anisotropy. Also evident is a variation in the matter power spectrum amplitude with $n$, evident on all scales.
Variation of parameters other than $n$ could also have a significant impact on the success of the models. Given the results of the previous section, it seems unlikely that any influence of the field $\xi$ would tend to improve the models. This indeed seems to be the case. Figure \ref{cs2} depicts various models where the function is varied $C_{S}^{2}$ for fixed values of the other parameters. In particular, $n$ takes the value $0.1$. The function $C_{S}^{2}$ is ultimately a measure of the ability of the field $\xi$ to sustain any homogeneous growing behavior for $(k\tau)>1$; higher values will tend to limit the effect of the vector field to larger and larger scales. The sets of parameters were chosen so such a growing solution indeed existed on superhorizon scales. The figure indicates that a growing $\xi$ field will indeed have deleterious effects on scales where it is not suppressed. There is a significant ISW effect evident for the red curve CMB; the corresponding model must be considered as being at the edge of acceptability. The corresponding
large scale matter power spectrum exhibits exotic oscillations, entirely unrelated to the baryon-acoustic oscillations which occur on other scales. Their presence is in this model is thus reflective of dynamics in the dark energy sector.
\begin{figure}
\begin{center}
\epsfig{file=varying_nae.eps,width=8.3cm,height=9.3cm}
\caption{Matter power (top) and CMB (bottom) power spectra for the various GEA dark energy models, with WMAP and SDSS constraints. The power-law function's exponent $n$ is varied}.
\label{nae}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsfig{file=varying_cs2.eps,width=8.3cm,height=9.3cm}
\caption{Matter power (top) and CMB (bottom) power spectra for the various GEA dark energy models, with WMAP and SDSS constraints. The squared speed of sound of vector field perturbations is varied.
}
\label{cs2}
\end{center}
\end{figure}
\section{Conclusions}
\label{conclusions}
\subsection{Being Dark Matter is hard}
Generalized Einstein-Aether can, with different parameter choices, resemble dark matter in some important ways but never all of them at once. Specifically, the new degrees of freedom introduced by the model may conspire to identically replicate one or more but not all of the following properties of cold dark matter.
\subsubsection{Background dynamics}
To accomplish identical background dynamics to cold dark matter, one must introduce considerable fine tuning into the function $F$ of the theory. Specifically the function must change form either side of (dark) matter-radiation equality. The parameter we tune to make this happen is in the action itself, unlike the usual case where we simply alter the abundance $\Omega_c$. Changes to fit cosmological observations can therefore have a larger impact on the small scale behavior of gravity.
\subsubsection{The speed of sound}
If the speed of sound is too high then structure cannot form on small enough scales. In this model we may reduce the sound speed to be close to zero, at the cost of one of our parameters. When designing new gravity theories this is perhaps the easiest structure formation constraint to investigate, and it should be examined to see if it conflicts with other constraints needed to make the theory useful - for an example, see \cite{Seifert:2007fr}.
\subsubsection{Growth rate of `overdensity'}
Theories of modified gravity designed to replace dark matter must necessarily have growing modes of fluctuation in at least one of the new degrees of freedom they introduce, in order to sufficiently source gravitational collapse and structure formation on scales within their own sound horizon.
There remains some flexibility in the perturbation growth rate, since the bias on the galaxy power spectrum measurements is a free factor. As measurements of weak lensing (which samples gravitation directly) and semi-analytic models (which predict bias) improve this freedom will be reduced.
\subsubsection{Absence of anisotropic stress and contribution the cosmological Poisson equation}
A sufficiently small anisotropic stress associated with the vector field may be implemented by fine tuning the parameter $c_{13}$ to be very small (see equation (\ref{shear})). However, as was discussed in Subsection \ref{ISW subsection}, even this will tend to come at the expense of other desired behavior of the field.
An appreciable time variation of the anisotropic stress over the time from last scattering to today can result in a very poor fit to the
low-$\ell$ CMB $C_\ell$. This problem is likely to be common in theories of modified gravity.
As we have seen, it is a combination of time variation of the anisotropic stress and time variation of the field $\Phi$ via the vector field's effect on the cosmological Poisson equation that contribute to the ISW effect. Though both effects are absent in the cold dark matter case, one may imagine both effects being present in the vector field model but being of equal and opposite sign. As with the case of the isolated anisotropic shear contribution, it seems this is not possible whilst maintaining the other constraints like the existence of a growing mode.
\subsubsection{Effective minimal coupling to the gravitational field}
Even if the vector field growing mode gives an appropriate (dark matter-like) contribution to the Poisson equation, the link between the overdensity and the corresponding $\Phi$ can differ from the CDM case. This difference can come from curvature terms in the vector field stress energy tensor, and its main consequence is a time-dependent rescaling of Newton's constant $G$. Thus $\Phi$ may gain a time dependence during the matter era even there is a completely standard dark matter contribution to the Poisson equation. The converse may also be possible: the time dependence of an incorrect Poisson contribution could be counteracted by a time dependence of the effective $G$.
\subsection{Being Dark Energy is easy}
As a model for dark energy the Generalized Einstein-Aether theory is more successful: we have obtained constraints on its parameters and found that it generates spectra that fit the data across a reasonable range of its parameter space. It is clear that modified gravity approaches to explaining late-time acceleration are viable and can provide motivated explanations for dark energy (though this theory retains the co-incidence problem in the guise of the parameter $M$).
\subsubsection{Closeness to $\Lambda$}
The most interesting constraint on this branch of the theory is on the parameter $n_{\mathrm{ae}}$ and is shown in figure \ref{nae constraint}. In some sense this parameter describes how closely the theory mimics $\Lambda$ (which has $n_{\mathrm{ae}}=0$). The fact that this parameter is rather free, $n_{\mathrm{ae}}<0.126$, (95\% CL), is consistent with the fact that a wide variety of other theories can also explain dark energy: present structure formation data is not very informative about the nature of dark energy, and deeper require expansion probes like baryon acoustic oscillation and supernovae.
\subsubsection{Sound speed}
The other notable constraint on Einstein-Aether dark energy, which may extend to other modified gravity approaches, is the limits on the sound speed, illustrated in figure \ref{cs2 constraint}. As shown in figure \ref{cs2}, an incorrect sound speed can lead to large scale oscillations by modifying the other parameters of the theory and permitting a growing mode excitation at late time.
\subsubsection{Other constraints}
The other constraints on the theory (which are easily fulfilled by choosing the $c$ parameters) come from ensuring that no growing mode can disrupt the power spectra, that the acceleration is close to the $\Lambda$CDM value, and that the value of effective $G$ remains positive at all times.
\subsection{Future issues for modified gravity and structure formation}
Because $\Lambda$CDM is such a good fit to current cosmological data, modified gravity will never be favoured in a model comparison exercise using only current data about linear structure. It is only in combination with physics on galactic and smaller scales that it can be be persuasive. This work highlights a few issues for future model-building in this vein.
The Generalized Einstein Aether model is a member of a class of models in which the scale $M \sim H_0$ associated with Dark Energy is visible to dark matter. Many of the issues raised here will be relevant to any such models which try to use a dark matter scale consistent with small-scale modifications to gravity.
A combination of probes sensitive only to the background (like Type 1A supernovae) and to the behavior of cosmological perturbations is needed to fully constrain these theories. For example, a value $n_{\mathrm{ae}}=0.3$ has $w(z)\sim 1$ at low redshift but is ruled out by our constraints. Similarly there are models with $n_{\mathrm{ae}}\sim 0.75$ that provide reasonable spectra, but they are ruled out by $w(z)$ constraints.
There are, of course, a number of extensions to the theory and features of it that could be change; we could, for example, allow $M$ to vary dynamically, or add more terms to the kinetic component $F$ in equation (\ref{kinetic equation}). There are also myriad possibilities in more changing more general aspects of the primordial conditions or cosmological parameters: what happens if we add add tensors? Can we include an isocurvature mode? Would massive neutrinos help? Or curvature? This leads us to a key caveat that applies to this and all similar work constraining new physics with linear structure: a simple constraint from the data alone is worthless, since any of the numerous other parameters we could change might conspire to counteract whatever problem it solves. We need a physical explanation of a constraint's origin to understand whether it is robust to the cosmologist's tinkering.
{\it Acknowledgments}:
We thank Constantinos Skordis and David Jacobs for useful discussions. GDS was supported by a grant to the CWRU particle/astrophysics theory group from the US DOE. JZ is supported by an STFC rolling grant. TGZ is supported by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research \& Innovation.
|
\section{Introduction}
Understanding the cosmological evolution of active galactic nuclei
(AGNs) is an important issue in modern astrophysics and is closely
related to the evolution of galaxies. One piece of evidence for ``co-evolution" of
AGNs and galaxies is the strong correlation between a central black hole
and its host galaxy (e.g., Magorrian et al. 1998; Marconi \& Hunt 2003).
In an early stage of galaxy evolution, the nuclear region may be hidden
behind rich gas, which leads to active star formation. Indeed
recent models predict that the central black holes in all galaxies
experience a heavily obscured phase (Hopkins et al. 2006,
2008). Therefore, investigating obscured AGNs is of significant
interest to study the co-evolution of black holes and galaxies.
Moreover, previous observations have shown that obscured AGNs are
a major population of the AGN.
Finally, obscured AGN
are predicted to be the main contributors to the unexplained
cosmic X-ray background (CXB) in the hard X-rays
(Comastri et al. 1995; Ueda et al. 2003; Gilli et al. 2007).
Thus, obscured AGN population is a key class
to understand the overall AGN population, AGN evolution, and
relationship between black holes and their hosts.
According to a unified model of the AGN, obscuring matter referred to as a
``torus'' is surrounding a supermassive black hole and gas
photoionized by the AGN is created in the opening part of the torus
(Antonucci 1993). If we observe an AGN from the torus side, absorbed
direct emission from the nucleus and emission scattered by the
photoionized gas are observed in the X-ray spectra. A scattering
fraction is calculated as the fraction of the scattered emission with
respect to the direct emission and reflects the solid angle of the
opening part of the torus and/or an amount of gas responsible for
scattering. {\it Suzaku's} follow-up observations of {\it Swift}
BAT-detected AGNs found a new type of AGN with a very small scattering
fraction ($<$ 0.5\%; Ueda et al. 2007; Eguchi et al. 2009; Winter et
al. 2009). Assuming that the amount of scattering medium does not
differ much from object to object, they would be buried in a
geometrically thick torus with a small opening angle. Noguchi et
al. (2009) found several buried AGN candidates with a very small
scattering fraction from the Second {\it XMM-Newton} Serendipitous
Source Catalogue ($2XMM$) using hardness ratios (HRs) and showed that
such type of AGNs tend to have a low relative
[\ion{O}{3}]$\lambda$5007 luminosity compared to the intrinsic X-ray
luminosity. This implies that the AGN with a small scattering fraction
could constitute the main class of optically elusive obscured AGNs.
In this paper, we construct a new sample of obscured AGNs covering a
broad range of scattering fractions from $2XMM$ Catalogue in the same
way as Noguchi et al. (2009), and investigate multi-wavelength
properties of a buried AGN in comparison with a classical type of AGN
having a large scattering fraction. In Section 2, we describe the
selection method of a new sample of obscured AGNs that covers a broad
range of scattering fractions. Our results of X-ray and optical spectral analysis
are briefly presented in Section 3, and their multiwavelength
properties are discussed in Section 4. Finally, we summarize our
results in Section 5. We adopt ({\it H}$_{\rm 0}$, $\Omega_{\rm m}$,
$\Omega_{\rm \lambda}$) $=$ (70 km s$^{-1}$Mpc$^{-1}$, 0.3, 0.7)
throughout this paper.
\section{Sample Selection}
We used the $2XMM$ Catalogue produced by the {\it XMM-Newton} Survey
Science Centre, which contains $\sim$250,000 detections drawn from
$\sim$3500 {\it XMM-Newton} EPIC observations made between 2000 and
2007 (Watson et al. 2009). The median flux in the full energy band
(0.2$-$12 keV) is $\sim$2.5$\times$10$^{-14}$ erg cm$^{-2}$ s$^{-1}$,
and about 20\% of the sources have total fluxes below
1$\times$10$^{-14}$ erg cm$^{-2}$ s$^{-1}$.
We derived a sample that covers a wide range of scattering fractions
($\sim$0.1\%$-$10\%) from the $2XMM$ Catalogue using the HRs in the same
way as Noguchi et al. (2009) as briefly summarized below. We used HRs
defined as
\begin{displaymath}
{\rm HR}3=\frac{{\rm CR}(2.0-4.5\ {\rm keV})-{\rm CR}(1.0-2.0\ {\rm
keV})}{{\rm CR}(2.0-4.5\ {\rm keV})+{\rm CR}(1.0-2.0\ {\rm keV})}
\end{displaymath}
and
\begin{displaymath}
{\rm HR}4=\frac{{\rm CR}(4.5-12\ {\rm keV})-{\rm CR}(2.0-4.5\ {\rm
keV})}{{\rm CR}(4.5-12\ {\rm keV})+{\rm CR}(2.0-4.5\ {\rm keV})},
\end{displaymath}
where CR(1.0$-$2.0 keV), CR(2.0$-$4.5 keV), and CR(4.5$-$12 keV) are
count rates in the 1.0$-$2.0, 2.0$-$4.5, and 4.5$-$12 keV bands,
respectively. Only 4627 sources, which have count rate for EPIC-pn in
0.2$-$12 keV $>$ 0.05 counts s$^{-1}$, high Galactic latitude ($|b|$ $>$
20$^\circ$), and error of HRs $\leq$ 0.2 at a 90\% confidence level,
were targeted in this selection. The values and errors of HR3 and HR4
given in the $2XMM$ Catalogue were calculated using count rates measured
by the {\tt emldetect} task in the {\it XMM-Newton} Science Analysis
System (SAS). In Figure \ref{figure:HR}, their HR3 and HR4 are
plotted. The five solid lines represent the scattering fractions of
10\%, 5\%, 3\%, 1\%, and 0.5\% from inside to outside. The three dashed lines
correspond to objects with log {\it N}$_{\rm H}$ (cm$^{-2}$) = 23,
23.5, and 24 from lower right to upper left. In calculating these
lines, we assumed an intrinsic spectrum of a power law with a photon
index of 1.9. The details on how to draw these lines are described in
Noguchi et al. (2009). Based on the solid line for 10\% and dashed
lines for log {\it N}$_{\rm H}$ = 23 and 24, we selected objects as
shown by circles in this figure as our sample for which log {\it
N}$_{\rm H}$ and scattering fraction ($f_{\rm scat}$) are in the
range 23$-$24 and less than 10\%, respectively.
\begin{figure}[!b]
\centering
\includegraphics[angle=270,scale=.40]{fig1.eps}
\caption{Distribution of hardness ratio (HR) for the $2XMM$ Catalogue
sources (crosses). Objects satisfying all of the following
conditions are plotted; count rate in 0.2$-$12 keV $>$ 0.05 counts
s$^{-1}$, $|b|$ $>$ 20$^\circ$, and HR error $\leq$ 0.2. Data points with
circles are analyzed and 32 AGNs were selected as our sample from
them. Solid lines show the HRs expected for the scattering fraction
of 10, 5, 3, 1, and 0.5\% from inside to outside. Dashed lines
correspond to log$N$$_{\rm H}$ of 23, 23.5, and 24 cm$^{-2}$ from
lower right to upper left.}
\label{figure:HR}
\end{figure}
We excluded low-luminosity AGNs with the intrinsic 2$-$10 keV
luminosity lower than 10$^{41}$ erg s$^{-1}$ and Compton-thick AGNs
with column densities of log {\it N}$_{\rm H}$ $>$
1.5$\times$10$^{24}$ cm$^{-2}$ from our sample because of the
following reasons. Soft X-rays of low-luminosity AGNs would be
contributed by X-ray emission from hot gas and discrete sources in
their host galaxies, and direct emission of Compton-thick AGNs is
almost completely blocked below 10 keV. Therefore, scattering
fractions of these classes of objects cannot be calculated
correctly. We analyzed X-ray spectra obtained by {\it XMM-Newton} to
identify Compton-thick and low-luminosity AGNs.
If an X-ray continuum spectrum above 4 keV is fitted by a pure
reflection component
and the equivalent width of Fe-K$\alpha$ with respect to the
reflection component is greater than 1 keV (Matt et al. 1991), we
classified such objects as a Compton-thick AGN. The absorption
corrected 2$-$10 keV luminosities were calculated from the best-fit
models obtained in the same ways as discussed in Section 3.1, and
objects with the luminosity lower than 10$^{41}$ erg s$^{-1}$ were
regarded as a low-luminosity AGN. Objects listed as a Seyfert 1, 1.5
or a star in NASA/IPAC Extragalactic Database (NED) were
excluded. Objects with a scattering fraction more than 10\%, which was
calculated from a formula described in Section 4.1, were also removed
from our sample. Furthermore, NGC 1052 was excluded from the sample
because this object shows complicated spectrum requiring a double
partial covering model, and a component absorbed by $N_{\rm H}$ $\sim$
2$\times$10$^{22}$ cm$^{-2}$ is relatively strong, which makes it
difficult to measure the scattering fraction. We finally selected 32
objects as members of our sample, of which 10 objects are newly
selected AGNs and the others are included in Noguchi et al. (2009)
sample.
A summary of the newly selected AGNs and their properties is found in
Table \ref{table:sample}. Most of the AGNs in our sample are at $z$ $<$
0.1 because we assumed $z$ = 0 in the simulation of AGN spectra used
to calculate the expected hardness ratios and our sample was selected
from bright sources (count rate in 0.2$-$12 keV $>$ 0.05 counts
s$^{-1}$). The Galactic column densities are calculated from 21 cm
measurements (Kalberla et al. 2005) using the {\tt nh} tool at the
NASA's High Energy Astrophysics Science Archive Research Center. Since
{\it XMM-Newton} results of some of the objects in our sample have
been published, references for them are also listed in Table
\ref{table:sample}.
{\tabletypesize{\scriptsize}
\begin{deluxetable*}{llcccccccc}
\setlength{\tabcolsep}{-3pt}
\tablewidth{0pt}
\tablecaption{Newly Selected AGNs}
\tablehead{
\colhead{$2XMM$ Name} &
\colhead{Other Name} &
\colhead{Class} &
\colhead{Reference$^a$} &
\colhead{Redshift} &
\colhead{$N_{\rm H}$$^b$} &
\colhead{Start Date} &
\colhead{Exposure$^c$} &
\colhead{Count Rate$^d$} &
\colhead{Reference$^e$} \\
\colhead{} &
\colhead{} &
\colhead{} &
\colhead{} &
\colhead{} &
\colhead{($10^{20}$ cm$^{-2}$)} &
\colhead{} &
\colhead{(s)} &
\colhead{(counts s$^{-1}$)}&
\colhead{}
}
\startdata
$2XMM$ J030030.5$-$112456 & MCG $-$02$-$08$-$039 & Sy2 & 1 & 0.030 & 5.13 & 2006 Jan 23 & 5335 & 0.14 & \nodata\\
$2XMM$ J031000.0+170559 & 3C 79 & Sy2 & 1 & 0.256 & 8.72 & 2004 Feb 14 & 9302 & 0.035 & 3\\
$2XMM$ J033336.3$-$360825 & NGC 1365 & Sy1.8 & 1 & 0.005 & 1.34 & 2003 Jan 17 & 15142 & 0.66 & 4\\
$2XMM$ J085537.7+781324 & NGC 2655 & Sy2 & 2 & 0.005 & 2.18 & 2005 Sep 04 & 3117 & 0.20 & 5\\
$2XMM$ J104943.4+583750 & 2MASX J10494334+5837501 & \nodata & \nodata & 0.115 & 0.667& 2005 Oct 10 & 21908 & 0.044 & \nodata \\
$2XMM$ J113240.2+525701 & Mrk 176 & Sy2 & 1 & 0.027 & 1.09 & 2004 May 02 & 9639 & 0.053 & 6\\
$2XMM$ J134442.1+555312 & Mrk 273 & Sy2 & 1 & 0.038 & 0.890& 2002 May 07 & 18018 & 0.067 & 7\\
$2XMM$ J135602.7+182218 & Mrk 463 & Sy2 & 1 & 0.051 & 2.03 & 2001 Dec 22 & 21081 & 0.12 &8, 9\\
$2XMM$ J201657.3$-$704459 & IC 4970 & \nodata & \nodata & 0.016 & 4.09 & 2002 Mar 31 & 19129 & 0.015 &\nodata\\
$2XMM$ J224937.0$-$191627 & MCG $-$03$-$58$-$007 & Sy2 & 1 & 0.0315 & 2.06 & 2005 May 09 & 2024 & 0.13 & \nodata
\enddata
\label{table:sample}
\tablenotetext{a}{References for optical classification.}
\tablenotetext{b}{Galactic column density by 21 cm measurement (Kalberla et al. 2005).}
\tablenotetext{c}{Cleaned exposure of EPIC-pn.}
\tablenotetext{d}{Count rate in the 0.4$-$10 keV band.}
\tablenotetext{e}{References for published {\it XMM-Newton} results.}
\tablerefs{(1)Veron-Cetty \& Veron 2006 (2) Ho et al. 1997; (3) Hardcastle et al. 2006; (4) Risaliti et al. 2005; (5) Akylas \& Georgantopoulos 2009; (6) Guainazzi et al. 2005; (7) Balestra et al. 2005; (8) Awaki et al. 2006; (9) Bianchi et al. 2008.}
\end{deluxetable*}
}
\section{Spectral Analysis}
\subsection{X-ray Spectra}
We analyzed X-ray spectra of newly selected sources to calculate the
scattering fraction. In the current analysis, the spectra were grouped
to a minimum of 15 counts per bin and the $\chi^2$ minimization
technique was used. The quoted errors are given at the 90\% confidence
level for one interesting parameter (i.e., $\Delta\chi^2$ = 2.71). We
performed spectral fits for our sample in the 0.4$-$10 keV range with
various models by using XSPEC version 11.2.
First, all spectra were fitted with a baseline model consisting of
absorbed and unabsorbed power laws, along with a Gaussian
corresponding to an Fe-K$\alpha$ line. {\tt zpowerlw} and {\tt zgauss}
models in XSPEC were used to account for the power laws and the
Gaussian line, respectively. All the components are modified by the
Galactic absorption. Our baseline model is shown as {\tt
phabs*(zpowerlw + zphabs*(zpowerlw + zgauss))} in XSPEC, where {\tt
zphabs} and {\tt phabs} are models corresponding to photoelectric
absorption by cold matter at the redshift of the source and in our
Galaxy, respectively. The photon indices ($\Gamma$) of the two power
laws were linked.
None of the spectra of the 10 objects were satisfactorily fitted with
the baseline model. The spectrum of IC 4970 was explained by adding
another absorption for the less absorbed power law to the baseline
model. For the other objects, we used a complex model, in which an
optically thin thermal plasma model ({\tt mekal} model in XSPEC; Mewe
et al. 1985, Kaastra. 1992, Liedahl et al. 1995) and/or a Compton
reflection model ({\tt pexrav} model in XSPEC) were added to the
baseline model.
For the {\tt mekal} model, the abundance was fixed at 0.5 solar, where
the solar abundance table by Anders and Grevesse (1989) was
assumed. For the {\tt pexrav} model, the inclination angle of the
reflector and the high-energy cutoff of the incident power law were
fixed at 60$^{\circ}$ (0$^\circ$ corresponds to face-on) and at 300
keV, respectively. The solar abundance table by Anders \& Grevesse
(1989) was assumed. {\tt rel\_refl} parameter was set to $-$1 to
produce a reflection component only. The normalization and the photon
index of {\tt pexrav} were assumed to be the same as that of the
heavily absorbed power law. Since the reflection component of NGC
1365 required further absorption, we applied a {\tt zphabs} model to
the {\tt pexrav} model. Furthermore, if needed, we added another {\tt
mekal} component and/or Gaussian lines as many as required to model
the soft part of the spectrum.
Table \ref{table:model} shows the best-fit models. The spectra with
the models are shown in Figure \ref{figure:sp}.
The spectral
parameters for the models are summarized in Tables \ref{table:para1}$-$\ref{table:para3}. $\Gamma$ were distributed
between $\sim$1.5 and 2.0, which is the range of values for Seyfert 2s
(Smith \& Done 1996), although we fixed the value at 1.9 for some
objects since uncertainties of $\Gamma$ became large ($>$ 20\% at the
90\% level of confidence for one interesting parameter) or $\Gamma$
was far from a typical range (about 1.4$-$2.4), if $\Gamma$ was left
free. The obtained $N_{\rm H}$ were in the range of 10$^{23-24}$
cm$^{-2}$. We calculated absorption-corrected luminosities for the
power-law components in 2$-$10 keV and of soft components in 0.5$-$2
keV using the best-fit model. The 0.5$-$2 keV luminosities were
calculated using all the components except for the heavily absorbed
power law corresponding to direct emission. The calculated
luminosities are summarized in Table \ref{table:data}. The
luminosities for our sample are in the range of Seyferts
($\sim$10$^{41-44}$ erg s$^{-1}$).
\begin{figure*}
\centering
\includegraphics[width=7cm,height=8cm,clip,angle=270]{fig2a.eps}
\includegraphics[width=7cm,height=7.5cm,clip,angle=270]{fig2b.eps}
\includegraphics[width=7cm,height=8cm,clip,angle=270]{fig2c.eps}
\includegraphics[width=7cm,height=7.5cm,clip,angle=270]{fig2d.eps}
\includegraphics[width=7.5cm,height=8cm,clip,angle=270]{fig2e.eps}
\includegraphics[width=7.5cm,height=7.5cm,clip,angle=270]{fig2f.eps}
\caption{X-ray spectra ({\it upper panels}) and residuals in units of
$\sigma$ ({\it lower panels}). Model components are shown with
dashed, dotted, dot-dashed, and triple-dot-dashed lines.}
\label{figure:sp}
\end{figure*}
\begin{figure*}
\figurenum{2}
\centering
\includegraphics[width=7cm,height=8cm,clip,angle=270]{fig2g.eps}
\includegraphics[width=7cm,height=7.5cm,clip,angle=270]{fig2h.eps}
\includegraphics[width=7.5cm,height=8cm,clip,angle=270]{fig2i.eps}
\includegraphics[width=7.5cm,height=7.5cm,clip,angle=270]{fig2j.eps}
\caption{Continued}
\end{figure*}
\begin{deluxetable}{ll}
\tabletypesize{\scriptsize}
\tablecaption{The Best-Fit Models of Newly Selected AGNs}
\tablewidth{0pt}
\tablehead{\colhead{Name}& \colhead{Model$^a$}
}
\startdata
MCG $-$02$-$08$-$039 & BM + MEKAL + two lines \\
3C 079 & BM + MEKAL \\
NGC 1365 & BM + MEKAL + nine lines + abs-Ref \\
NGC 2655 & BM + MEKAL \\
2MASX J10494334+5837501 & BM + two MEKAL + Ref \\
Mrk 176 & BM + MEKAL + abs-Ref \\
Mrk 273 & BM + MEKAL \\
Mrk 463 & BM + MEKAL \\
IC 4970 & BM$^b$\\
MCG $-$03$-$58$-$007 & BM + two MEKAL
\enddata
\tablenotetext{a}{BM: baseline model; PL: power law, MEKAL: thin thermal plasma model ({\tt mekal}), Ref: cold refection model ({\tt pexrav}), abs-Ref: absorbed ref. All components are absorbed by the Galactic absorption.}
\tablenotetext{b}{Additional absorption was applied to the less absorbed power law.}
\label{table:model}
\end{deluxetable}
{\tabletypesize{\scriptsize}
\begin{deluxetable*}{lccccccccc}
\tablecolumns{10}
\tablecaption{Spectral Parameters for Absorption, Power Law, and Gaussian in the Best-fit Models}
\setlength{\tabcolsep}{-3pt}
\tablewidth{0pt}
\tablehead{
\colhead{}\\
\colhead{Name} & \colhead{$N_{\rm H}$} &
\colhead{$\Gamma$} & \colhead{$A_{\rm int}$$^a$} &
\colhead{$E_{\rm line}$} & \colhead{$\sigma$} &
\colhead{EW} &\colhead{$A_{\rm ga}$$^b$} &
\colhead{$A_{\rm scat}$$^c$} & \colhead{$\chi^2_{\rm \nu}$(dof)} \\
\colhead{} &\colhead{($10^{22}$ cm$^{-2}$)}&
\colhead{} & &
\colhead{(keV)} & \colhead{(eV)} &
\colhead{(eV)} & &
& \\
}
\startdata
MCG $-$02$-$08$-$039 & 33$^{+13}_{-10}$ & 1.9(f) & 0.92$^{+0.44}_{-0.28}$ & 6.435$^{+0.061}_{-0.066}$ & 10(f) & 320$\pm200$ & 0.88$\pm0.54$ & 2.30$^{+0.47}_{-0.49}$ & 1.12(33)\\
3C 79 & 48$^{+17}_{-13}$ & 1.9(f) & 0.69$^{+0.30}_{-0.20}$ & 6.4(f) & 10(f) & 0($<$110) & 0($<$0.28) & 1.18$\pm0.34$ & 0.93(17)\\
NGC 1365 & 65.3$^{+2.2}_{-2.3}$ &2.056$\pm0.018$ & 12.4$^{+0.38}_{-0.28}$ & 6.390$^{+0.051}_{-0.055}$&0($<$124)&83$\pm40$&1.7$\pm1.1$ &10.09$^{+0.69}_{-0 .72}$&1.08(186)\\
NGC 2655 & 31$^{+10}_{-7}$ & 1.9(f) & 1.11$^{+0.40}_{-0.30}$ & 6.4(f) & 10(f) & 65($<$280) & 0.21($<$0.90) & 3.19$^{+0.85}_{-0.86}$ & 0.79(36)\\
2MASX J10494334+5837501 & 28.9$^{+7.8}_{-5.8}$ & 1.9(f) & 0.367$^{+0.077}_{-0.062}$ & 6.403$^{+0.063}_{-0.072}$ & 60($<$195) & 300$^{+190}_{-150}$ & 0.36$^{+0.23}_{-0.18}$ & 0.40$^{+0.14}_{-0.13}$ & 1.14(65) \\
Mrk 176 & 62.4$^{+15.2}_{-7.5}$ & 1.9(f) & 1.081$^{+0.17}_{-0.19}$ & 6.4(f) & 10(f) & 200$\pm150$ & 0.66$\pm0.49$ & 0.83$^{+0.29}_{-0.23}$ & 1.11(27) \\
Mrk 273 & 82$^{+20}_{-16}$ & 1.59$^{+0.13}_{-0.16}$ & 0.52$^{+0.38}_{-0.23}$ & 6.4(f) & 10(f) & 240$^{+190}_{-160}$ & 0.68$^{+0.53}_{-0.44}$ & 2.05$^{+0.36}_{-0.28}$ & 1.21(72) \\
Mrk 463 & 46.4$^{+6.5}_{-5.8}$ & 2.02$^{+0.13}_{-0.09}$ & 1.12$^{+0.46}_{-0.31}$ & 6.414$^{+0.040}_{-0.037}$ & 10(f) & 250$\pm100$ & 0.70$\pm0.28$ & 3.21$^{+0.35}_{-0.36}$ & 1.02(136) \\
IC 4970 & 32$^{+15}_{-11}$ & 1.9(f) & 0.34$^{+0.21}_{-0.13}$ & 6.410$^{+0.050}_{-0.062}$ & 0($<$100) & 470$\pm300$ & 0.47$\pm0.30$ & 1.13$^{+0.24}_{-0.25}$ & 0.65(20)\\
MCG $-$03$-$58$-$007 & 23.5$^{+9.6}_{-7.8}$ & 1.9(f) & 0.78$^{+0.30}_{-0.28}$ & 6.4(f) & 10(f) & 0($<$480) & 0($<$1.14) & 1.12$^{+0.82}_{-0.97}$ & 0.68(10)
\enddata
\tablecomments{Photon index of the power law with only Galactic absorption was assumed to be the same value as power law absorbed by cold matter at the redshift of the source. (f) indicates fixed parameter.}
\tablenotetext{a}{Normalization of the absorbed power law in units of $10^{-3}$ photons cm$^{-2}$ s$^{-1}$ at 1 keV.}
\tablenotetext{b}{Normalization of the Gaussian line in units of $10^{-5}$ photons cm$^{-2}$ s$^{-1}$ in the line.}
\tablenotetext{c}{Normalization of the less absorbed power law in units of $10^{-5}$ photons cm$^{-2}$ s$^{-1}$ at 1 keV.}
\label{table:para1}
\end{deluxetable*}
}
{\tabletypesize{\scriptsize}
\begin{deluxetable*}{lccccccc}
\setlength{\tabcolsep}{-3pt}
\tablecaption{Spectral Parameters for MEKAL and the Additional Absorptions in the Best-fit Models}
\tablewidth{0pt}
\tablehead{
\colhead{}\\
&&&&&&\multicolumn{1}{c}{Additional} \\
&\multicolumn{2}{c}{MEKAL$^a$} &&\multicolumn{2}{c}{MEKAL$^a$}&\multicolumn{1}{c}{Absorption} \\ \cline{2-3}\cline{5-6}\\
\colhead{Name} &\colhead{\it kT} & \colhead{$A_{\rm m}$$^b$} &&\colhead{\it kT} & \colhead{$A_{\rm m}$$^b$} &\colhead{$N_{\rm H}$} \\
&\colhead{(keV)} & &&\colhead{(keV)} & &\colhead{($10^{22}$ cm$^{-2}$)} \\
}
\startdata
MCG $-$02$-$08$-$039 & 0.197$^{+0.021}_{-0.016}$ & 14.3$\pm3.3$ &&\nodata &\nodata &\nodata \\
3C 79 & 0.35$^{+0.38}_{-0.16}$ & 1.14$^{+0.98}_{-0.76}$ &&\nodata &\nodata &\nodata \\
NGC 1365 &0.635$\pm0.018$&22.26$^{+0.92}_{-0.94}$&&\nodata&\nodata&0.63$^{+0.54}_{-0.45}$$^c$\\
NGC 2655 & 0.535$^{+0.096}_{-0.077}$ & 9.7$\pm1.4$ &&\nodata &\nodata &\nodata\\
2MASX J10494334+5837501 & 0.54$^{+0.13}_{-0.14}$ & 0.612$^{+0.17}_{-0.16}$ && 0.081($<$0.085) & 28.8$^{+6.4}_{-6.2}$ &\nodata\\
Mrk 176 & 0.61$^{+0.12}_{-0.14}$ & 1.66$^{+0.45}_{-0.35}$ && 0.081($<$0.095) & 18$^{+10}_{-16}$ &\nodata\\
Mrk 273 & 0.695$^{+0.083}_{-0.057}$ & 3.00$^{+0.54}_{-0.59}$ &&\nodata &\nodata &\nodata \\
Mrk 463 & 0.664$\pm0.043$ & 4.15$^{+0.58}_{-0.56}$ &&\nodata &\nodata&\nodata \\
IC 4970 &\nodata&\nodata&&\nodata&\nodata&0.13$^{+0.11}_{-0.08}$$^d$\\
MCG $-$03$-$58$-$007 & 0.36$^{+0.26}_{-0.10}$ & 5.2$^{+2.9}_{-2.1}$ && 0.081($<$0.11) & 125$^{+56}_{-100}$ &\nodata
\enddata
\tablenotetext{a}{Metal abundances were fixed at 0.5 solar.}
\tablenotetext{b}{Normalization of {\tt mekal} in units of 10$^{-19}$/(4$\pi$({\it D}$_{\rm A}\times$(1+{\it z}))$^2$) $\int$ {\it n}$_{\rm e}$ {\it n}$_{\rm H}$ {\it dV}, where {\it D}$_{\rm A}$ is the angular size distance to the source (cm), {\it n}$_{\rm e}$ is the electron density (cm$^{-3}$), and {\it n}$_{\rm H}$ is the hydrogen density (cm$^{-3}$).}
\tablenotetext{c}{Absorption for a {\tt pexrav} model.}
\tablenotetext{d}{Absorption for a less absorbed power law.}
\label{table:para2}
\end{deluxetable*}
}
\newcounter{1}
\setcounter{1}{1}
\newcounter{4}
\setcounter{4}{4}
\newcounter{5}
\setcounter{5}{5}
\newcounter{6}
\setcounter{6}{6}
\newcounter{7}
\setcounter{7}{7}
\newcounter{8}
\setcounter{8}{8}
\newcounter{17}
\setcounter{17}{17}
\newcounter{9}
\setcounter{9}{9}
\newcounter{10}
\setcounter{10}{10}
\newcounter{11}
\setcounter{11}{11}
\newcounter{13}
\setcounter{13}{13}
\newcounter{16}
\setcounter{16}{16}
\newcounter{23}
\setcounter{23}{23}
\newcounter{25}
\setcounter{25}{25}
\newcounter{26}
\setcounter{26}{26}
{\tabletypesize{\scriptsize}
\begin{deluxetable}{lcccc}
\tablewidth{\columnwidth}
\tablecaption{Spectral Parameters for Additional Gaussians in the Best-fit Models}
\tablewidth{0pt}
\tablehead{
\colhead{}\\
\colhead{Name} &\colhead{$E_{\rm line}$} &\colhead{$\sigma$} &\colhead{$A_{\rm ga}$$^a$} &\colhead{Identification} \\
&\colhead{(keV)} & (eV) & & \\
}
\startdata
MCG $-$02$-$08$-$039 & 0.467$^{+0.023}_{-0.043}$ & 10(f) &3.5$^{+3.9}_{-1.7}$ & N \Roman{6} K$\alpha$, N \Roman{7} K$\alpha$ \\
& 0.959$\pm0.024$ & 10(f) &0.70$\pm0.33$ & Ne \Roman{10} K$\alpha$ $?$, Ne \Roman{10} K$\alpha$ $?$ \\
\\
NGC 1365 & 0.544$\pm0.014$ & 10(f) &2.42$\pm0.72$ & O \Roman{7} K$\alpha$\\
& 1.202$^{+0.019}_{-0.020}$ & 10(f) &0.67$^{+0.23}_{-0.22}$ & Ne \Roman{5} Ly$\alpha$ \\
& 1.354$\pm0.19$ & 10(f) &0.61$^{+0.14}_{-0.26}$ & Mg \Roman{4} K$\alpha$\\
& 1.820$^{+0.045}_{-0.028}$ & 10(f) &0.37$\pm0.16$ & Si \Roman{8}\\
& 2.65$^{+0.05}_{-0.12}$ & 10(f) &0.29$\pm0.17$ & S \Roman{16} Ly$\alpha$\\
& 6.687$^{+0.037}_{-0.039}$ & 10(f) &$-$2.66$^{+0.73}_{-0.74}$ & Fe \Roman{25} K$\alpha$\\
& 7.002$^{+0.040}_{-0.050}$ & 10(f) &$-$1.48$\pm0.69$ & Fe \Roman{26} K$\alpha$\\
& 7.98$^{+0.09}_{-0.16}$ & 10(f) &$-$1.6$\pm1.0$ & Fe \Roman{25} K$\beta$\\
& 8.287$^{+0.086}_{-0.069}$ & 10(f) &$-$1.63$^{+0.94}_{-0.95}$ & Fe \Roman{26} K$\beta$
\enddata
\tablecomments{(f) indicates fixed parameter.}
\tablenotetext{a}{Normalization of the Gaussian line in units of $10^{-5}$ photons cm$^{-2}$ s$^{-1}$.}
\label{table:para3}
\end{deluxetable}
}
{\tabletypesize{\scriptsize}
\begin{deluxetable*}{lccccccccccc}
\setlength{\tabcolsep}{-3pt}
\tablewidth{0pt}
\tablecaption{Multi-wavelength Properties of Our Sample}
\tablehead{
\colhead{Name} &
\colhead{Hard$^a$} &
\colhead{Soft$^b$} &
\colhead{$\log L_{\rm FIR}$} &
\colhead{$f_{\rm 60}$/$f_{\rm 25}$$^c$} &
\colhead{$\log L_{\rm [O III]}^{\rm int}$$^d$} &
\colhead{Ref.$^e$} &
\colhead{$M_{\rm BH}$$^f$} &
\colhead{Ref.$^g$} &
\colhead{$L_{\rm bol}/L_{\rm Edd}$$^h$} &
\colhead{$f_{\rm scat}$$^i$}&
\colhead{$f_{\rm scat}^{\rm corr}$$^j$}
}
\startdata
\multicolumn{12}{l}{{\bf AGN in Noguchi et al. (2009)}}\\
Mrk 348 & 43.49 & 40.82 & 43.49 & 1.5 & 41.95 & 1 & 6.7 & 12 & 0.2 & 0.4 & 0.4 \\
3C 33 & 44.07 & 41.55 & \nodata & \nodata & 42.52 & 2 & \nodata & \nodata & \nodata & 0.3 & \nodata \\
2MASX J02281350$-$0315023 & 43.46 & 41.01 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & 0.5 & \nodata \\
NGC 1142 & 43.51 & 41.25 & 44.77 & 8.4 & 41.87 & 3 & 8.2 & 12 & $-$1.3 & 0.8 & 0.3 \\
3C 98 & 43.01 & 41.00 & \nodata & \nodata & 41.91 & 4 & \nodata & \nodata & \nodata & 2.0 & \nodata \\
B2 0857+39 & 44.12 & 41.77 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & 0.6 & \nodata \\
IC 2461 & 41.83 & 38.93 & 43.00 & \nodata & 40.85 & 5 & \nodata & \nodata & \nodata & 0.2 & \nodata \\
2MASX J10335255+0044033 & 43.75 & 41.77 & \nodata & \nodata & 43.62 & 6 & \nodata & \nodata & \nodata & 1.4 & \nodata \\
MCG +08$-$21$-$065 & 42.65 & 39.95 & 43.51 & \nodata & 39.78 & 7 & \nodata & \nodata & \nodata & 0.3 & 0.1 \\
NGC 4074 & 42.88 & 40.88 & \nodata & \nodata & 42.05 & 8 & 7.9 & 13 & $-$1.6 & 1.4 & \nodata \\
NGC 4138 & 41.21 & 38.88 & \nodata & \nodata & 38.75 & 9 & 7.6 & 14 & $-$3.0 & 1.2 & \nodata \\
NGC 4388 & 42.89 & 40.63 & 43.94 & 2.9 & 41.77 & 1 & 7.0 & 13 & $-$0.7 & 1.4 & 0.8 \\
NGC 4507 & 43.09 & 41.19 & 43.81 & 3.1 & 41.69 & 1 & 7.5 & 12 & $-$1.0 & 3.1 & 2.9 \\
ESO 506$-$G027 & 43.69 & 40.58 & 43.58 & 1.9 & \nodata$^k$ &\nodata & \nodata & \nodata & \nodata & 0.2 & 0.1 \\
2MASX J12544196$-$3019224 & 43.07 & 41.00 & \nodata & \nodata & 41.60 & 5 & \nodata & \nodata & \nodata & 1.4 & \nodata \\
NGC 4939 & 42.33 & 40.31 & 43.59 & 5.4 & 41.43 & 1 & 7.6 & 15 & $-$1.8 & 1.3 & 0.8 \\
ESO 383$-$G18 & 42.60 & 40.27 & \nodata & 1.4 & 40.36 & 5 & \nodata & \nodata & \nodata & 1.3 & \nodata \\
ESO 103$-$G035 & 43.33 & 40.35 & 43.54 & 1.0 & 41.65 & 1 & 7.0 & 15 & $-$0.3 & 0.1 & 0.1 \\
IC 4995 & 41.96 & 40.65 & 43.40 & 2.5 & 41.96 & 8 & 7.1 & 13 & $-$1.8 & 8.1 & 7.2 \\
NGC 7070A & 41.75 & 39.33 & 42.40 & \nodata & \nodata$^l$ & \nodata & 6.8 & 14 & $-$1.7 & 0.9 & 0.7 \\
NGC 7172 & 43.06 & 40.18 & 43.77 & 7.1 & 39.83 & 1 & 7.5 & 12 & $-$1.1 & 0.3 & 0.1 \\
NGC 7319 & 42.96 & 40.92 & \nodata & \nodata & 41.44 & 1 & 7.2 & 13 & $-$0.9 & 1.3 & \nodata \\
\hline
\multicolumn{12}{l}{{\bf Newly selected AGN}} \\
MCG $-$02$-$08$-$039 & 42.71 & 41.48 & \nodata & 1.1 & 41.62 & 10 & 7.7 & 12 & $-$1.6 & 8.2 & \nodata \\
3C 079 & 44.45 & 42.69 & \nodata & 2.8 & \nodata & \nodata & \nodata & \nodata & \nodata & 2.4 & \nodata \\
NGC 1365 & 42.29 & 40.58 & 44.53 & 6.6 & 40.97 & 1 & \nodata & \nodata & \nodata & 2.1 & \nodata \\
NGC 2655 & 41.20 & 39.99 & 42.78 & 6.0 & 39.91 & 9 & 7.6 & 14 & $-$3.1 & 8.3 & 7.2 \\
2MASX J10494334+5837501 & 43.49 & 41.83 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & 3.0 & \nodata \\
Mrk 176 & 42.73 & 40.87 & 43.96 & 3.0 & 42.51 & 11 & 8.0 & 13 & $-$1.9 & 1.9 & 1.4 \\
Mrk 273 & 42.90 & 41.42 & 45.53 & 9.5 & 42.45 & 1 & 8.1 & 12 & $-$1.8 & 7.3 & \nodata \\
Mrk 463 & 43.18 & 41.84 & 44.76 & 1.4 & 42.87 & 1 & \nodata & \nodata & \nodata & 5.3 & 4.4 \\
IC 4970 & 41.71 & 39.96 & \nodata & \nodata & \nodata$^l$ & \nodata & \nodata & \nodata & \nodata & 2.4 & \nodata \\
MCG $-$03$-$58$-$007 & 42.70 & 41.39 & 44.44 & 3.0 & 41.78 & 10 & \nodata & \nodata & \nodata & 6.6 & 5.1
\enddata
\label{table:data}
\tablenotetext{a} {{\rm L}ogarithm of absorption-corrected 2$-$10 keV luminosity (erg s$^{-1}$). }
\tablenotetext{b} {{\rm L}ogarithm of absorption-corrected 0.5$-$2 keV luminosity of soft X-ray components (erg s$^{-1}$). }
\tablenotetext{c} {60 $\mu$m to 25 $\mu$m flux ratio. }
\tablenotetext{d} {{\rm L}ogarithm of [O III] luminosity (erg s$^{-1}$) corrected for reddening using Balmer decrement. }
\tablenotetext{e} {References for [O III] luminosity. }
\tablenotetext{f} {{\rm L}ogarithm of black hole mass calculated from stellar velocity dispersion. }
\tablenotetext{g} {References for stellar velocity dispersion. }
\tablenotetext{h} {{\rm L}ogarithm of Eddington ratio. }
\tablenotetext{i} {Scattering fraction (\%).}
\tablenotetext{j} {Scattering fraction corrected for the contribution of starburst (\%).}
\tablenotetext{k} {{\rm Lower} limit ($\log L_{\rm [O III]}^{\rm int}$$>43.23$ (erg s$^{-1}$)) is given in Landi et al. (2007).}
\tablenotetext{l} {[O III] lines are not detected.}
\tablerefs{(1) Bassani et al. 1999; (2) Yee \& Oke 1978; (3) Shu et
al. 2007; (4) Costero \& Osterbrock 1977; (5) This work; (6) Dong et
al. 2005; (7) Line flux measurement based on the Sloan Digital Sky
Survey data at MPA/JHU (http://www.mpa-garching.mpg.de/SDSS/);
Kauffmann et al. 2003; (8) Polletta et al. 1996; (9) Ho et al. 1997;
(10) de Grijp et al. 1992; (11) Mulchaey et
al. 1994; (12) Garcia-Rissmann et al. 2005; (13) Nelson \& Whittle
1995; (14) McElroy 1995; (15) Cid Fernandes et al. 2004}
\end{deluxetable*}
}
\subsection{Optical Spectra}
Optical spectroscopic observations of four targets in our sample,
2MASX~J12544196--3019224, IC~4970, NGC~7070A, and ESO~383--G18, were
performed during several nights between 2008 August 9 and 11 by using
the South African Astronomical Observatory (SAAO) 1.9 m telescope with the
Cassegrain spectrograph. The grating six, which has a spectral range
of about 3500$-$5300 \AA ~at a resolution around 4 \AA, was used with
$\sim$ 2 arcsec slit placed on the center of each galaxy, for a total
integration time ranging from 750 to 3600 s. The data reduction and
analysis was made in a standard manner with the IRAF package to derive
the flux calibrated spectra. To obtain the sensitivity curve, we
fitted the observed spectral energy distribution (SED) of standard stars
with low-order polynomials. We detected [\ion{O}{3}]$\lambda$5007 as
well as H$\alpha$ and H$\beta$ narrow lines from
2MASX~J12544196--3019224 and ESO~383--G18. Their narrow lines and line
intensity ratios indicate that 2MASX~J12544196--3019224 and
ESO~383--G18 are a Seyfert 2 and \ion{H}{2} nucleus, respectively.
For IC~4970 and NGC~7070A, we derived the 90\% upper limits on the
[\ion{O}{3}] intensity, based on an estimated rms noise in the
continuum flux around the corresponding wavelength.
The fluxes were corrected for the slit loss, estimated from
the image extent of the target or a nearby point-like source projected
onto the spatial direction. We also analyzed
calibrated optical spectrum of IC 2461 taken from the Sloan Digital
Sky Survey (SDSS) Data Release 7. The
narrow lines and line intensity ratios indicate a Seyfert 2. The line
fluxes and the spectra of the five objects are shown in Table
\ref{table:optical} and Figure \ref{figure:sp_opt}, respectively.
\begin{deluxetable}{lcccc}
\tabletypesize{\scriptsize}
\tablecaption{[O III]$\lambda$5007, H$\alpha$, and H$\beta$ Fluxes}
\tablewidth{0pt}
\tablehead{\colhead{Name}& \colhead{[O III]} & \colhead{H$\alpha$} & \colhead{H$\beta$} & \colhead{Class}
}
\startdata
IC 2461 & 4.2 & 3.9 & 0.50 & Sy2 \\
2MASX J12544196$-$3019224 & 4.1 & 1.9 & 0.58 & Sy2 \\
ESO 383$-$G18 & 14 & 6.5 & 2.8 & H II \\
NGC 7070A & $<$1 & \nodata &\nodata & \nodata \\
IC 4970 & $<$2 & \nodata &\nodata & \nodata
\enddata
\tablecomments{Fluxes are in units of 10$^{-14}$ erg cm$^{-2}$ s$^{-1}$. Upper limits are at a 90\% confidence level.}
\label{table:optical}
\end{deluxetable}
\newpage
\section{Discussion}
\subsection{Scattering Fraction}
We first calculated a scattering fraction ({\it f}$_{\rm scat}$) for our sample
consisting 32 AGNs with the equation
\begin{displaymath}
{\it f}_{\rm scat}=\frac{L_{0.5-2}^{\rm soft}}{L_{0.5-2}^{\rm int}} ,
\end{displaymath}
where $L_{0.5-2}^{\rm int}$ and $L_{\rm 0.5-2}^{\rm soft}$ are
absorption corrected fluxes in the 0.5$-$2 keV band for the absorbed
power law and all the components except for the heavily absorbed power
law, respectively. The calculated values are shown in Table
\ref{table:data}. Various components other than scattered emission,
however, may contribute to the soft X-rays in obscured AGNs such as
thermal emission originated from hot plasma collisionally heated by
starburst activity. In fact, high-resolution images available with
{\it Chandra} show that the soft X-ray emission is dominated by the
starburst component in some cases (NGC 4945, Schurch et al. 2002),
although the {\it XMM-Newton} Reflection Grating Spectrometer (RGS)
high-resolution spectra show evidence that the soft X-ray emission
appears to be dominated by AGN emission for most Seyfert 2 galaxies
with high-quality data (Guainazzi \& Bianchi 2007). Therefore, the
value calculated from the above equation is regarded as an upper limit
on the scattering fraction.
We examined the starburst contribution to the soft X-rays for our
sample using far-infrared luminosities ({\it L}$_{\rm FIR}$), and
calculated scattering fractions by subtracting the estimated
contribution. {\it L}$_{\rm FIR}$ is often used to estimate starburst
activity although we cannot rule out the contribution of AGNs and host
galaxies to {\it L}$_{\rm FIR}$. We calculated {\it L}$_{\rm FIR}$ for
our sample using the formula defined in Helou et al. (1985), based on
flux densities at 60 $\mu$m and 100 $\mu$m. We collected infrared
fluxes (60 $\mu$m and 100 $\mu$m) measured with {\it Infrared
Astronomical Satellite} ({\it IRAS}) for 18 of 32 objects from NED.
{\it IRAS} fluxes for the rest of the 14 objects are not available.
Observations with upper limits were not used. {\it L}$_{\rm FIR}$ for
our sample is shown in Table \ref{table:data}. Ranalli et al. (2003)
show that there is a strong correlation between {\it L}$_{\rm FIR}$
and $L_{\rm 0.5-2}^{\rm soft}$ for starburst and normal galaxies as
expressed by
\begin{displaymath}
{\rm log}{\it L}_{\rm 0.5-2}^{\rm soft} = {\rm log}{\it L}_{\rm
FIR}^{\rm int} - 3.70.
\end{displaymath}
We estimated a 0.5$-$2 keV luminosity created by starburst ({\it
L}$_{0.5-2}^{\rm SB}$) using this correlation and calculated the
scattering fraction as
\begin{displaymath}
{\it f}_{\rm scat}^{\rm corr}=\frac{L_{0.5-2}^{\rm soft} - {\it
L}_{0.5-2}^{\rm SB}}{L_{0.5-2}^{\rm int}}.
\end{displaymath}
The calculated values are shown in Table \ref{table:data}. In Figure
\ref{figure:fs}, $f_{\rm scat}^{\rm corr}$ values are compared with
$f_{\rm scat}$. IC 2461, NGC 1365, and Mrk 273 were not plotted in
this figure since the values of $L_{0.5-2}^{\rm soft}$ $-$
$L_{0.5-2}^{\rm SB}$ are negative. This means that starburst
contribution to their soft X-rays is substantial. In fact, recent
{\it XMM-Newton} RGS spectra of NGC 1365 show soft X-ray emission
dominated by collisionally ionized plasma (Guainazzi et al. 2009).
Figure \ref{figure:fs} shows a relatively tight correlation between
$f_{\rm scat}^{ \rm corr}$ and $f_{\rm scat}$, although the values of
$f_{\rm scat}^{\rm corr}$ are lower than $f_{\rm scat}$ and the
scatter for objects with a small $f_{\rm scat}$ is larger than for a
large $f_{\rm scat}$. The tight correlation, particularly at large
$f_{\rm scat}$, is consistent with the idea that the soft X-rays
dominated by emission from photoionized gas are likely to be a common
characteristic of Seyfert 2 galaxies. Therefore, in the following
discussions, we use $f_{\rm scat}$ since the values can be calculated
for all objects in our sample.
From our spectral analysis combined with our previous results, we
found that $f_{\rm scat}$ for our sample are in the range of
$\sim$0.1\%$-$10\%. In particular, those of eight objects are very small
($<$ 0.5\%) as reported in Noguchi et al. (2009), whereas a typical
value for Seyfert 2s previously studied is about 3\% (Bianchi \&
Guainazzi 2007; Turner et al. 1997). The scattering fraction is
proportional to both the solid angle subtended by the scattering
electrons and a scattering optical depth. Thus, such a very small
$f_{\rm scat}$ implies that they would be a new type of AGN buried in
a very geometrically thick torus with a small opening angle (Ueda et
al. 2007; Winter et al. 2008, 2009; Eguchi et al. 2009; Noguchi et
al. 2009) on the assumption that the optical depth of scatterer is
nearly constant among the objects. Our sample that covers a broad
range of ${\it f}_{\rm scat}$ allows us to investigate properties of
buried AGNs with a small ${\it f}_{\rm scat}$ by comparing with those
of a classical type of Seyfert 2s with a large ${\it f}_{\rm scat}$.
\begin{figure*}
\centering
\includegraphics[width=7cm,height=8cm,clip,angle=270]{fig3a.eps}
\includegraphics[width=7cm,height=7.5cm,clip,angle=270]{fig3b.eps}
\includegraphics[width=7.5cm,height=8cm,clip,angle=270]{fig3c.eps}
\includegraphics[width=7.5cm,height=7.5cm,clip,angle=270]{fig3d.eps}
\includegraphics[width=7.5cm,height=8cm,clip,angle=270]{fig3e.eps}
\caption{Optical spectra of IC 2461, 2MASX J12544196$-$3019224,
ESO 383$-$G18, NGC 7070A, and IC 4970 in the 4500-7500 \AA \ wavelength range.}
\label{figure:sp_opt}
\end{figure*}
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=.40]{fig4.eps}
\caption{Comparison of scattering fractions ($f_{\rm scat}$) with corrected scattering
fractions ($f_{\rm scat}^{\rm corr}$) calculated by subtracting the starburst contribution
estimated from FIR luminosity. The solid line represents $f_{\rm
scat}$$=$$f_{\rm scat}^{\rm corr}$. }
\label{figure:fs}
\end{figure}
\subsection{Geometry of Obscuring Material}
Fabian et al. (1998) proposed a model in which low-luminosity AGNs
were obscured by starburst in the inner 100 pc of a central massive
black hole and suggested that supernovae from nuclear starburst input
the energy to the circumnuclear gas to create a torus-like structure.
This suggestion was investigated in more detail using
three-dimensional hydrodynamic simulations (Wada \& Norman 2002,
2007). Seyfert galaxies with nuclear starburst within a few hundred pc
from the center were indeed found observationally (Imanishi
2002, 2003; Rodriguez-Ardila \& Viegas 2003). Therefore, nuclear
starburst is considered as a key factor to keep the shape of a
torus-like absorber either theoretically or observationally. If this
is the case, the scale height of the torus would expand with an
increase of the star-forming activity, and starburst activity in
objects with a small $f_{\rm scat}$ should be stronger than those in
objects with a large $f_{\rm scat}$. To confirm this expectation, we
investigated a relation between $f_{\rm scat}$ and {\it L}$_{\rm FIR}$
as shown in Figure \ref{figure:FIR}. We found no significant
correlation between them, in contrast to our expectation that there is an
anti-correlation.
This means that the scale height of the torus seems
not to be primarily determined by starburst activity and there are
other contributing factors to support the shape of a torus-like
absorber.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=.40]{fig5.eps}
\caption{Far infrared luminosity ($L_{\rm FIR}$) plotted against scattering fraction.
}
\label{figure:FIR}
\end{figure}
The absence of a correlation between $f_{\rm scat}$ and $L_{\rm FIR}$
might in part be due to a selection bias against infrared luminous
objects such as ultraluminous infrared galaxies (ULIRGs), which have
most of their luminosity emerging in the infrared with $L$(8$-$1000
$\mu$m) $>$ 3$\times$10$^{45}$ erg s$^{-1}$ and contain a
substantially large amount of gas and dust in their nuclei compared
with classical Seyferts (Sanders \& Mirabel 1996). Imanishi et
al. (2007) found buried AGNs signatures in a significant fraction of
ULIRGs using {\it Spitzer} Infrared Spectrograph even though they are
optically classified as non-Seyferts, and suggested that AGNs in ULIRGs
are almost fully buried in surrounding gas and dust, for which $f_{\rm
scat}$$\sim$ 0 is expected. In our sample, however, only Mrk 273 and
Mrk 463 are classified as a ULIRG. The paucity of infrared luminous
objects in our sample is possibly due to our selection criteria since
many of ULIRGs are heavily obscured by Compton-thick matter $>$
10$^{24}$ cm$^{-2}$ (Risaliti et al. 2000; Teng et al. 2009). If we
had plotted them in Figure \ref{figure:FIR}, they would be located in
the upper left portion since they have {\it f}$_{\rm scat}$$\sim$ 0
and are bright in the infrared, and the anti-correlation may appear in
Figure \ref{figure:FIR} as expected.
\subsection{Infrared Color}
Warm infrared colors of AGNs are usually explained as emission from
hot dust in the vicinity of a nucleus. For a buried AGN, the cooler
component from the outer region of the torus dominates, since a
geometrically thick torus almost covers the center region unless the
viewing angle is near face-on. Hence, we
predict that buried AGNs with a small scattering fraction show cooler
infrared colors. 60 $\mu$m to 25 $\mu$m flux ratios ($f_{\rm 60}$/$f_{\rm
25}$) for our sample are shown in Table \ref{table:data}. Figure
\ref{figure:color} compares $f_{\rm scat}$ with $f_{\rm 60}$/$f_{\rm
25}$ and shows that there is no correlation between them. This
result is inconsistent with above prediction. However, this may be
explained by the effect of the difference of inclination of the torus to
our line of sight because $f_{\rm 60}$/$f_{\rm 25}$ depends also on
the viewing angle (Heisler et al. 1997). If we view a Seyfert galaxy
nearly face-on, the value of $f_{\rm 60}$/$f_{\rm 25}$ becomes smaller
because infrared emission from the hot dust near the nucleus can be
seen. On the other hand, an edge-on view makes $f_{\rm 60}$/$f_{\rm
25}$ larger since the hot dust is obscured and cooler component from
the outer part dominates. Expected infrared SEDs are calculated by
several authors for smooth (Pier \& Krolik 1992, 1993; Granato
\& Danese 1994) and clumpy (Nenkova el al. 2008) distribution of
torus. According to these calculations, objects with small $f_{\rm
scat}$ may be viewed from a face-on angle, in spite of the fact that
their absorption column densities along the line of sight is large
(10$^{23-24}$ cm$^{-2}$). A similar situation is found in {\it Swift}
BAT-selected AGNs (Ueda et al. 2007). If we observe them from an
edge-on side, their direct emission would be completely blocked by
Compton-thick matter ($\gg$ 10$^{24}$ cm$^{-2}$). Thus, if our results
are explained in terms of face-on geometry, this implies the existence of a
yet to be discovered very Compton-thick AGN.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=.40]{fig6.eps}
\caption{Infrared color $f_{\rm 60}$/$f_{\rm 25}$ plotted against
scattering fraction.
}
\label{figure:color}
\end{figure}
\subsection{Scattering Fraction and [\ion{O}{3}] }
If a scattering fraction reflects the opening angle of the torus,
objects with a small $f_{\rm scat}$ should have also a small narrow-line
region (NLR), which is considered to exist in the opening region
of the torus. We predicted that their [\ion{O}{3}] $\lambda$5007
emission luminosities produced in the NLR are weaker than those for
objects with a large $f_{\rm scat}$ at a given X-ray luminosity (Ueda
et al. 2007; Noguchi et al. 2009). To test this prediction, Noguchi et
al. (2009) compared ratios of reddening corrected [\ion{O}{3}] to
intrinsic 2$-$10 keV luminosities ({\it L}$_{\rm [O\ III]}^{\rm
int}$/{\it L}$_{\rm 2-10}^{\rm int}$) for a sample of the buried AGN
consisting of 22 objects with a small $f_{\rm scat}$ with those for a
large sample of Seyfert2s compiled by Bassani et al. (1999), and found
that there is a clear difference in their distributions, with the buried
AGNs having smaller ratios of {\it L}$_{\rm [O\ III]}^{\rm int}$/{\it
L}$_{\rm 2-10}^{\rm int}$. This result is in good agreement with
the above prediction. In this subsection, to confirm the result we
examine whether there is a similar relationship between {\it L}$_{\rm
[O\ III]}^{\rm int}$/{\it L}$_{\rm 2-10}^{\rm int}$ and $f_{\rm
scat}$ using our sample, which covers a broad range of $f_{\rm
scat}$.
We used [\ion{O}{3}] luminosities collected from the literature as
shown in Table \ref{table:data}. These values are corrected for the
extinction by using the relation
\begin{displaymath}
{\it L}_{\rm [O\ III]}^{\rm int}={\it L}_{\rm [O\ III]}^{\rm obs}\biggl[\frac{{\rm
H}\alpha/{\rm H}\beta}{({\rm H}\alpha/{\rm
H}\beta)_0}\biggr]^{2.94},
\end{displaymath}
assuming an intrinsic Balmer decrement (H$\alpha$/H$\beta$)$_0$ = 3.0,
where $L_{[{\rm O III}]}^{\rm obs}$ and {\rm H}$\alpha/${\rm H}$\beta$
are an observed [\ion{O}{3}] luminosity and a ratio between observed
{\rm H}$\alpha$ and {\rm H}$\beta$ line fluxes, respectively (Bassani
et al. 1999). The calculated values are shown in Table \ref{table:data}.
In Figure \ref{figure:opt}, we plot {\it L}$_{\rm [O\ III]}^{\rm
int}$/{\it L}$_{\rm 2-10}^{\rm int}$ against $f_{\rm scat}$. As
expected, we find there is a positive correlation between them. The
best-fit liner line (dashed line in Figure \ref{figure:opt}) is given
as
\begin{displaymath}
{\rm log} \frac{{\it L}_{\rm [O\ III]}^{\rm int}}{{\it L}_{\rm
2-10}^{\rm int}} = (0.89\pm0.43) {\rm log}\ {\it f}_{\rm scat} -
(1.42\pm0.23).
\end{displaymath}
We computed a Spearman's rank correlation coefficient ($\rho$) and
Kendall's rank correlation coefficient ($\tau$) to determine a level
of significance, and found $\rho$ $=$ 0.62 ({\it p} $=$ 0.0011) and
$\tau$ $=$ 0.45 ({\it p} $=$ 0.0014), where {\it p} is null hypothesis
probability. This correlation clearly supports the result obtained by
Noguchi et al. (2009) showing {\it L}$_{\rm [O\ III]}^{\rm int}$/{\it
L}$_{\rm 2-10}^{\rm int}$ for buried AGNs tend to be smaller than
other Seyfert 2s. Therefore, as suggested in Noguchi et al. (2009),
estimation of intrinsic luminosities of buried AGNs based on
[\ion{O}{3}] luminosities would result in large uncertainties, and
surveys using optical emission lines could be subject to biases
against buried AGNs.
\begin{figure}[!t]
\centering
\includegraphics[angle=270,scale=.40]{fig7.eps}
\caption{Ratio of intrinsic 2$-$10 keV to
reddening corrected [\ion{O}{3}] line luminosities plotted against
scattering fraction. The dashed line is a linear fit to the data.
Spearman's and Kendall's rank correlation coefficients are $\rho$ $=$ 0.62 ({\it p} $=$ 0.0011) and
$\tau$ $=$ 0.45 ({\it p} $=$ 0.0014), respectively.
}
\label{figure:opt}
\end{figure}
\subsection{Black Hole Mass and Eddington Ratio}
The black hole mass ($M_{\rm BH}$) is one of the most important pieces
of information to represent properties of an AGN, and we compare
$f_{\rm scat}$ with $M_{\rm BH}$. Black hole masses for our sample
were estimated from the relation found by Tremaine et al. (2002)
\begin{displaymath}
M_{\rm BH}=10^{8.13}\times\biggl[\frac{\sigma_{\ast}}{200\ {\rm
km\ s^{-1}}}\biggr]^{4.02}M_{\sun},
\end{displaymath}
where $\sigma_{\ast}$ is a stellar velocity dispersion. The values of
$\sigma_{\ast}$ for 16 among 32 objects are collected from the
literature as shown in Table \ref{table:data}. The black hole masses
for our sample are in the range 6.7 $\leq$ log($M_{\rm BH}/M_{\sun}$)
$\leq$ 8.2 (Table \ref{table:data}). In Figure \ref{figure:MBH}, we
plot $M_{\rm BH}$ against $f_{\rm scat}$ and found only a hint of very
weak positive correlation between them ($\rho$ $=$ 0.36 ($p$ $=$ 0.17)
and $\tau$ $=$ 0.25 ($p$ $=$ 0.19)). If this trend is real, it is
consistent with an X-ray absorption model for a compton-thin AGN
described in Lamastra et al. (2006), in which a higher black hole mass
leads to a larger opening angle.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=.40]{fig8.eps}
\caption{Black hole mass plotted against scattering fraction. The
dashed line is a linear fit to the data. Spearman's and Kendall's rank correlation
coefficients are $\rho$ $=$ 0.36 ({\it p} $=$ 0.17) and $\tau$ $=$ 0.25 ({\it p} $=$ 0.19),
respectively.
}
\label{figure:MBH}
\end{figure}
The Eddington ratio ($L_{\rm bol}/L_{\rm Edd}$) is also an important
parameter in studies of accretion and evolution of AGNs. Eddington
luminosities ($L_{\rm Edd}$) are calculated from the black hole mass
as $L_{\rm Edd}$=1.26$\times$10$^{38}$($M_{\rm
BH}/M_{\sun}$). Bolometric luminosities ($L_{\rm bol}$) are
calculated as $L_{\rm bol}=$30$\times$$L_{\rm 2-10}^{\rm int}$, where
a typical bolometric correction factor for luminous AGNs ($\sim$ 30) is
applied (Elvis et al. 1994; Risaliti \& Elvis 2004; Vasudevan \&
Fabian 2007) and $L_{\rm 2-10}$ is an intrinsic luminosity in the 2$-$
10 keV band. We note that [\ion{O}{3}] luminosities may not be a good
estimator of intrinsic luminosities as discussed in Section 4.2,
although they are often used to derive bolometric luminosities
(Heckman et al. 2004). The values of $L_{\rm bol}/L_{\rm Edd}$ are in
the range 10$^{-4}$ $<$ $L_{\rm bol}/L_{\rm Edd}$ $<$ 2. In Figure
\ref{figure:Edd}, the $L_{\rm bol}/L_{\rm Edd}$ ratios are plotted
against $f_{\rm scat}$, and we find an anti-correlation between them
($\rho$ $=$ $-$0.54 ($p$ $=$ 0.032) and $\tau$ $=$ $-$0.40 ($p$ $=$
0.033)). This correlation could indicate that buried AGNs are rapidly
growing compared with AGNs surrounded by the torus with a large opening
part. {\it f}$_{\rm scat}$ could be an indicator of the growing phase
of black holes and a useful parameter to understand the evolution of
the active nucleus.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=.40]{fig9.eps}
\caption{Eddington ratio plotted against scattering fraction. The
dashed line is a linear fit to the data. Spearman's and Kendall's rank correlation
coefficients are $\rho$ $=$ $-$0.54 ({\it p} $=$ 0.032) and $\tau$ $=$ $-$0.40 ({\it p} $=$ 0.033),
respectively.
}
\label{figure:Edd}
\end{figure}
\subsection{Host Galaxy}
Studying properties of AGN host galaxies is of particular interest to
understand a relationship between an AGN and its host. We made $gri$ composite
images of 12 galaxies in our sample, for which imaging data were taken at
Kitt Peak (M. Koss et al. in preparation) or the SDSS data are available. An arcsinh
stretch was used as described in Lupton et al. (2004) with color scaled by
flux, and resulting images are shown in Figure \ref{figure:image}.
The morphology of the galaxies spans a wide range including early-type,
spiral, and strongly interacting galaxies. The inclination of the spiral
hosts also ranges from edge-on to near face-on.
Winter et al. (2009) studied host galaxies of {\it Swift}/BAT-selected
AGNs and argued that AGNs with small absorption tend to
be near face-on, while highly absorbed AGNs are found regardless of the
inclination. The absorption columns for our sample are large ($N_{\rm H} =
10^{23-24}$ cm$^{-2}$), and the observed wide
variety of the inclination angles of the hosts is in agreement with the
results obtained by Winter et al. (2009). The three objects with the lowest values
of $f_{\rm scat}$ (IC 2461, ESO 506$-$G027, and MCG+08$-21-065$) are
edge-on spirals, while Mrk 348, which has similarly small $f_{\rm scat}$,
resides in nearly face-on spiral. Thus, the relation between
the inclination angle of the host and $f_{\rm scat}$ is not clear.
Four among 12 objects show clear signature of galaxy interaction
(Mrk 176, NGC 1142, Mrk 463, and Mrk 273). 2MASS~J10335255+0044033
is also possibly interacting. Three objects (Mrk 176, Mrk 463, and Mrk 273)
show intermediate to large values of the scattering fractions, while
$f_{\rm scat}$ for NGC 1142 is 0.8\%. The hosts of objects with
small $f_{\rm scat}$ ($<0.5$\%) are spirals without signature of
interaction (IC 2461, Mrk 348, ESO 506$-$G027, and MCG+08$-21-065$;
note that signatures of a double nucleus Mrk 348 were found in higher
resolution $Hubble Space Telescope$ ($HST$) images by Gorjian (1995)).
Through interactions of galaxies, gas inside galaxies could be transported to
the central region. Such gas is a candidate for the source of obscuring
matter around AGNs. Our findings, however, suggest that geometrically thick
obscuring matter is not directly related to galaxy interaction.
Galaxy interaction also leads to active star formation in the galaxies,
and we expect intense soft X-rays from such starburst activity. Therefore,
we should note that $f_{\rm scat}$ could be overestimated in such systems.
Among the objects showing interaction, Mrk 463 and Mrk 273 are ULIRGs,
and significant starburst activity is likely to be taking place.
The comparison between $f_{\rm scat}$ and $f^{\rm corr}_{\rm scat}$ in
Section 4.1 shows that the soft X-ray emission in Mrk 273 is indeed
dominated by starburst. The differences between $f_{\rm scat}$ and
$f^{\rm corr}_{\rm scat}$ for the other objects (Mrk 176, NGC 1142, and Mrk 463)
are not very large, and the contribution of starburst to the soft X-rays does
not alter the above discussion.
\begin{figure*}[!htb]
\centering
\includegraphics[scale=0.85]{fig10a.eps}
\includegraphics[scale=0.85]{fig10b.eps}
\includegraphics[scale=0.85]{fig10c.eps}\\
\includegraphics[scale=0.85]{fig10d.eps}
\includegraphics[scale=0.85]{fig10e.eps}
\includegraphics[scale=0.85]{fig10f.eps}\\
\includegraphics[scale=0.85]{fig10g.eps}
\includegraphics[scale=0.85]{fig10h.eps}
\includegraphics[scale=0.85]{fig10i.eps}
\includegraphics[scale=0.85]{fig10j.eps}\\
\includegraphics[scale=0.85]{fig10k.eps}
\includegraphics[scale=0.85]{fig10l.eps}
\caption{$gri$ composite images of galaxies in the sample taken at
Kitt Peak from M. Koss et al. (in preparation) and the SDSS. An arcsinh
stretch was used as described in Lupton et al. 2004 with color
scaled by flux. Galaxies are from left to right$-$first row: IC 2461,
Mrk 176, NGC 1142; second row: Mrk 348, Mrk 463, Mrk 273; third row:
NGC 4138, ESO 506$-$G027, NGC 4074, MCG +08$-$21$-$065; fourth row: NGC
4388, 2MASX J10335255+0044033.}
\label{figure:image}
\end{figure*}
\section{Conclusions}
We derived a new sample of obscured AGNs from the $2XMM$
Catalogue paying attention to the strength of the
scattered emission. Our sample covers a wide range of scattering
fractions, which is a fraction of scattered emission with respect to
direct emission and reflects the opening angle of the obscuring torus, and
allows us to investigate relations between geometrical structure
around a nucleus (the opening angle of the torus) and multiwavelengh
properties of AGNs. To calculate the scattering fractions
quantitatively, we analyzed X-ray spectra obtained with {\it XMM-Newton} and
found that our sample covers the range $f_{\rm scat}$$\sim$0.1\%$-$10\%.
Optical spectra of five objects were also analyzed.
We investigated multiwavelength properties for our sample, and found
the followings.
\begin{itemize}
\item There is no significant correlation between a scattering
fraction and a far-infrared luminosity, though the absence of a
correlation may be partly due to selection biases. This result
implies that the energy generated by nuclear starburst activity is
not a major source of energy supply to maintain the torus-like
shape of obscuring matter around the nucleus. There is no
significant correlation between an infrared color $f_{\rm
60}$/$f_{\rm 25}$ and a scattering fraction.
\item We found that the ratios of extinction-corrected
[\ion{O}{3}]$\lambda$5007 to intrinsic 2$-$10 keV luminosities for
objects with a small scattering fraction tend to be smaller than
those with a large scattering fraction. This result is in
accordance with our prediction that objects with a small opening
angle (or a small scattering fraction) also have a small amount of
narrow line region gas. Surveys using optical emission lines could
be biased against buried AGNs.
\item We compared black hole masses and Eddington ratios with
scattering fractions. The comparison with black hole masses showed
that there is only a hint of very weak correlation. The Eddington
ratio of buried AGNs tends to be larger for objects with a small
scattering fraction. The scattering fraction could be a useful
parameter to study growth and evolution of supermassive black
holes.
\item We examined optical images of 12 galaxies in our sample.
No clear relationships between $f_{\rm scat}$ and the inclination
angle or signatures of interaction of the hosts.
\end{itemize}
\acknowledgments We are grateful to Tohru Nagao and Yoshiaki Taniguchi
for useful discussions. This paper is based on observations obtained
with {\it XMM-Newton}, an ESA science mission with instruments and
contributions directly funded by ESA Member States and the USA
(NASA). The Kitt Peak National Observatory images were obtained using
MD-TAC time for program 0417. Kitt Peak National Observatory,
National Optical Astronomy Observatory, is operated by the Association
of Universities for Research in Astronomy (AURA) under cooperative
agreement with the National Science Foundation. This research made use
of the NASA/IPAC Extragalactic Database, which is operated by the Jet
Propulsion Laboratory, Calfornia Institute of Technology, under
contract with the National Aeronautics and Space Administration. This
work is supported by Grants-in-Aid for Scientific Research 20740109
(Y.T.), 20540230 (Y.U.), and 21244017 (H.A.) from the Ministry of
Education, Culture, Sports, Science, and Technology of Japan.
Facilities: \facility{{\it XMM-Newton}, {\it Swift}, {\it Sloan}, SAAO:1.9m, KPNO:2.1m}
|
\section{Introduction}
Recently, investigations of topological phases and phase transitions
has attracted great attention in condensed matter
physics\cite{Qi2010a}. The quantum Hall state\cite{Klitzing1980a} is
the first example of a topological state of quantum matter, with a
fully gap ground state in the bulk, and gapless excitations at the
edge. The chiral edge state is a holographic mirror of bulk
topology\cite{Thouless1982a}. In the recently discovered time
reversal invariant topological
insulators\cite{Bernevig2006a,Kane2005a}, helical edge states are
confined at the edge by the bulk energy gap, and states with
opposite spins counter-propagate. In the case of the quantum spin
Hall state realized in HgTe quantum wells, the topologically trivial
and non-trivial states are separated by a topological quantum phase
transition, tunable by the thickness of the quantum well.
Quantum spin chain is another example where topological quantum
phase transition is found. The low energy dynamics of 1D large-spin
Heisenberg antiferromagnet can be described as O(3) nonlinear sigma
model\cite{Haldane1988a}. Half-integer-spin chains are generally
gapless, while integer spin chains are gapped; they are described by
the O(3) nonlinear sigma model with and without the topological
term. This distinction bears strong similarity to the topological
insulators, which is distinct from the conventional insulators by
the presence of a topological term\cite{Qi2008a}.
In this paper, we investigate a spin-2 chain model with two
topologically distinct, transactionally invariant ground states. One
model has edge state $S=1$, while the other model has edge spin
$S=3/2$. The Berry's phase associated with the edge spins differ by
$\pi$. The square of the time reversal operator $T$ gives $T^2=1$
for the first case, whereas it gives $T^2=-1$ in the second case.
Due to this topological difference, the two ground states must be
separated by one or several topological quantum phase transitions
where the spin gap closes.
This paper is constructed as follows. In the next section, we will
review two exact solvable quantum spin model in one dimension. Some
materials can be systematically found in a brilliant work by Tu
et.al\cite{Tu2008}. Afterwards, our new model Hamiltonian is
presented according to the topological argument. Corresponding
numerical results are shown in the third section, where phase
diagram is discussed as well. Conclusions are drawn in the end.
\section{Model Hamiltonian}
Our starting point is an integer spin model introduced by Affleck,
Kennedy, Lieb, and Tasaki, namely the AKLT
model\cite{Affleck1987,Affleck1988}. It is proved that the AKLT
model has a unique infinite volume ground state, with an exponential
decay spin-spin correlation\cite{Affleck1988}. In agreement with the
Haldane conjecture\cite{Haldane1983}, the excitation gap of AKLT
model is nonzero. The ground state of AKLT model can be written down
exactly in terms of Schwinger bosons:
\begin{equation}\label{eq:AKLT_ground_state}
|\Psi ^{AKLT}\rangle=\prod_{\langle ij\rangle}(a_{i}^{\dagger
}b_{j}^{\dagger }-b_{i}^{\dagger }a_{j}^{\dagger })^{S}|0\rangle,
\end{equation}%
where $S$ stands for site spin, and $a^\dagger_i$, $b^\dagger_i$ are
creation operators of Schwinger bosons at the $i$th site. This
ground state can be rephrased in a pictorial form. The spin-2 AKLT
ground state can be schematically shown in Fig\ref{Fig:Spin2AKLT},
where the circles stand for sites on the chain. Each spin-2 can be
decomposed into totally symmetric combinations of four spin-1/2
states, and each state is represented by a solid dot in the figure.
Two pairs of neighboring dots form singlet states, shown in red
bonds. These bonds are usually called \textit{valence bonds}, and in
this sense, AKLT ground state is usually referred as \textit{Valence
Bond State}. It is proved rigorously that this ground state is
unique under this periodic boundary condition. Due to the symmetric
intra-site coupling and anti-symmetric inter-site coupling, the
parent Hamiltonian of this ground state is given by
\begin{equation} \label{AKLTHam}
H^{AKLT}=\sum_{\langle
ij\rangle}K_3P_3(\mathbf{S}_i,\mathbf{S}_j)+K_4P_4(\mathbf{S}_i,\mathbf{S}_j),\quad
K_3,K_4>0
\end{equation}
where $P_3$ and $P_4$ are the projection operators onto the spin-3
and spin-4 subspaces respectively. The positive coefficients ensure
the state in Eq.(\ref{eq:AKLT_ground_state}) to be the corresponding
groundstate.
In 1998, Scalapino, Zhang, and Hanke (SZH) introduced a SO(5)
symmetric spin model\cite{Scalapino1988} with an exact valence bond
ground state. The original motivation for the model is to illustrate
the SO(5) theory of high Tc superconductivity, which unifies the
antiferromagnetic (AF) and the d-wave superconducting (SC)
phases\cite{Zhang1997,Demler2004}. However, it was soon found later
that the SO(5) symmetry can also be interpreted as an enhanced spin
rotational symmetry\cite{Wu2003a}. The SZH model contains five
quantum states at each site, forming the vector representation of
the SO(5) group. However, these five states can also be interpreted
as the quantum states of the spin $S=2$ of the SO(3) spin chain. SZH
presented an exact ground state wave function expressed as a matrix
product state of the Dirac $\Gamma$ matrices, and showed that the
edge states of the SZH model are 4 fold degenerate at each edge.
Interpreted as an $S=2$ chain language, the edge spin contains
$S=3/2$ spin quantum numbers. Following this work, more valence bond
states with higher symmetry groups have been
constructed\cite{Kluemper1991,Affleck1991,Schuricht2008,Tu2008,Arovas2009}.
\begin{figure}[tbp]
\includegraphics[width=5cm]{Fig1_Spin2AKLT.eps} \caption{The
sketch of spin-2 AKLT wavefunction. The circles represent sites, and
each spin-2 is decomposed into four spin-1/2 solid dots. Red lines
connecting the dots stand for singlet states.}\label{Fig:Spin2AKLT}
\end{figure}
The basic idea of the SZH model is the following. The tensor product
of two SO(5) spinor can be decomposed into SO(5) singlet,
antisymmetric tensor, and symmetric traceless tensor, namely,
\begin{equation}
5\times5=1+10+14.
\end{equation}
In analogy with the spin-1 AKLT model, the largest subspace is
projected out, leading to the desired SZH model:
\begin{equation}\label{SZHSO5}
H=J\sum_{\langle xy\rangle}P_{14}(xy),\quad J>0.
\end{equation}
Due to the Clifford algebra of the five $\Gamma$-matrics:
$\Gamma^a\Gamma^b=2\delta^{ab}+2i\Gamma^{ab}$, no symmetric
traceless components are involved in the product of two
$\Gamma$-matrices. As a consequence,
\begin{equation} \label{SZHwave}
|\Psi^{SZH}\rangle=\sum_{m_1,\ldots
m_N}Tr(\Gamma^{m_1}\Gamma^{m_2}\ldots\Gamma^{m_N})|m_1m_2\ldots
m_N\rangle,
\end{equation}
is the ground state of the above SZH model, where $m$ is a vector
label of the SO(5) group, which can also be interpreted as the
$m_i=-2,-1,0,1,2$ quantum numbers of $S=2$ spin chain.
Due to the relationships between the SO(5) and the SO(3) groups,
there exists a natural deformation of the SZH model to an SU(2)
spin-2 SZH model. The required map from SO(5) group onto SU(2) group
with spin-2 is given by:
\begin{eqnarray}
10(SO(5))&=&3(SO(3))\oplus7(SO(3)) \\
14(SO(5))&=&5(SO(3))\oplus9(SO(3))
\end{eqnarray}
And therefore the SZH Hamiltonian deforms to;
\begin{equation} \label{SZHHam}
H^{SZH}=\sum_{\langle
ij\rangle}J_2P_2(\mathbf{S}_i,\mathbf{S}_j)+J_4P_4(\mathbf{S}_i,\mathbf{S}_j),\quad
J_2,J_4>0.
\end{equation}
The ground state is unchanged up to an SO(5) rotation.
To see if the SZH state is gapped or not, we can evaluate the ground
state spin correlation function. The correlation of matrix product
state can be easily derived by the transfer matrix technique, given
by,
\begin{equation}
\langle S^\mu_1S^\mu_r\rangle=(Tr
G^L)^{-1}Tr[Z(S^\mu)G^{r-2}Z(S^\mu)G^{L-r}],
\end{equation}
where $\mu=x,y,z$\cite{Tu2008}. Define $g=\sum_m\Gamma^m|m\rangle$,
then $G=g^\dagger\otimes g =\sum_{m}\Gamma^m\otimes \Gamma^m$, and
$Z(S^\mu)=g^\dagger\otimes S^\mu g$. As it's an isotropic magnet,
the correlation functions are the same in any directions. After some
detail calculation, we derive $\langle S^x_1S^x_r\rangle=\langle
S^y_1S^y_r\rangle=\langle S^z_1S^z_r\rangle=-20\times5^{-r}$ for
integer $r>1$. Therefore, the correlation length $\xi$ of the SZH
model equals to $1/\ln5$. This finite correlation length indicates
that the low-lying excitation in the SZH model is gapped, consistent
with the Haldane conjecture.
It is interesting to note that the correlation function of the SZH
model is negative-definite, with a correlation length
$\xi=1/ln5\sim0.61$. Consequently, the lattice constant is almost
twice of the correlation length, and the spins at neighboring sites
correlate extremely weakly. On the other hand, according to Arovas
\textit{et.al}'s work\cite{Arovas1988B}, the correlation length for
spin-2 AKLT is $1/\ln2$ which is roughly the lattice constant.
Therefore, although AKLT model is also a strongly disordered
antiferromagnet, the neighboring spins are closely correlated,
leading to the conventional staggering correlation function, say,
$\mathrm{S}_1\cdot\mathrm{S}_r\propto(-1)^r2^{-r}$.
Now we have two sets of models of 1-dimensional spin-2 chain with
exactly known Hamiltonians and ground state wavefunctions. The
differences between the AKLT and SZH model are not only the analytic
forms as they appear, but also the topological distinctions. The
same as topological insulator, the bulk topology is relevant to the
edge state of an open chain. For the spin-2 AKLT model, two solid
dots at each edge remain free. Symmetrical combination of these two
spin-1/2 dots results an edge spin with $S=1$. This boson-like edge
state is consistent with large-N theory of SU(N) quantum
antiferromagnets\cite{Ng1994}. However, the SZH model serves a
complement of the large-N analysis. For an open chain, the SZH
ground state is given by
\begin{equation} \label{SZHopen}
|\Psi;i,j\rangle=\sum_{m_1,\ldots
m_N}(\Gamma^{m_1}\Gamma^{m_2}\ldots\Gamma^{m_N})_{ij}|m_1m_2\ldots
m_N\rangle.
\end{equation}
It explicitly shows that at each edge, there are four degrees of
freedom since the matrix product state is four dimensional.
Therefore, the edge state of SZH model is spin-3/2, \textit{ie.},
fermion-like. That is completely different from the AKLT model. As
the edge state is protected by topology, and is robust under
perturbation, the AKLT and SZH models belong to different
topological classes. It can be easily understood from the Berry
phase's language. Berry phase $\Phi_{BP}$ is the additional phase
when the spin winds around, which relates to the expectation value
of $T^2$ by $\exp(-i\Phi_{BP})=\langle T^2\rangle$, where $T$ is the
time reversal operator. It's well-known that $T^2=-1$ for half
integer spins such as $S=3/2$, while $T^2=1$ for integer spins such
as $S=1$. As a consequence, the Berry phases of the two models under
investigation differ by an angle of $\pi$. It is this difference
that makes topological distinction of AKLT and SZH models. From
another points of view, the edge spin determines the ground state
degeneracy (GSD)\cite{Wen1990a}, and naturally serves as a topology
index in analogy to fractional quantum Hall effect. These two models
thus have different topology index due to different edge spins.
Given the topological distinction of the two ground states, we
construct a model Hamiltonian interpolating between the AKLT and SZH
models:
\begin{equation}\label{Hamiltonian}
H(\alpha) =(1-\alpha )H^{AKLT}+\alpha H^{SZH}.
\end{equation}
Without loss of generality, we set $J_2=K_3=1$, $J_4=K_4=\beta$ in
the following. As the edge state is robust unless the gap closes,
there must exist one or several topological quantum phase
transitions (TQPT) where the gap closes and reopens in the evolution
of $\alpha$ from $0$ to $1$. This TQPT can be addressed by studying
the behavior of energy spectrum and correlation function at each
$\alpha$.
\begin{figure}[tbp]
\centerline{
\includegraphics[height=4.0in,width=4.0in]{Fig2_Energy_Gradient.eps}
}
\caption{(color online) Ground state energy per site $E_0/N$ (a) and
the absolute value of its corresponding second derivative
$|d^2E/d\alpha^2|$ (b), as a function of $\alpha$, got by DMRG with
600 states and $\beta=1.0$ at different system sizes.
\label{Fig:Energy_Gradient}}
\end{figure}
\section{Numerical Results}
The {\it density matrix renormalization group} (DMRG) method is
employed\cite{White1992} in our study. For this purpose, it's
helpful to rewrite the projection operators explicitly in terms of
spin operators. Applying the identity
\begin{equation}
{\bf S}_i\cdot{\bf S}_j=\sum_{J=0}^{2S}[\frac{1}{2}J(J+1)-S(S+1)]P_J(ij),
\end{equation}
one can easily get
\begin{eqnarray}
P_2(ij)&=&\frac{1}{126}[-120({\bf S}_i\cdot{{\bf S}_j})-14({\bf S}_i\cdot{{\bf S}_j})^2+7({\bf S}_i\cdot{{\bf S}_j})^3\nonumber\\
&&+({\bf S}_i\cdot{{\bf S}_j})^4]\\
P_3(ij)&=&\frac{1}{360}[162({\bf S}_i\cdot{{\bf S}_j})-7({\bf S}_i\cdot{{\bf S}_j})^2-10({\bf S}_i\cdot{{\bf S}_j})^3\nonumber\\
&&-({\bf S}_i\cdot{{\bf S}_j})^4]+1\\
P_4(ij)&=&\frac{1}{2520}[90({\bf S}_i\cdot{{\bf S}_j})+63({\bf S}_i\cdot{{\bf S}_j})^2+14({\bf S}_i\cdot{{\bf S}_j})^3\nonumber\\
&&+({\bf S}_i\cdot{{\bf S}_j})^4]
\end{eqnarray}
For present study, we keep $m=600-1000$ states in the DMRG block
with more than $16$ sweeps to get a converged result, and the
truncation error is less than $10^{-7}$ in near the critical point,
and much less than $10^{-10}$ away from the critical point. We make
use of the open boundary condition (OBC) and the total number of
sites is $N=600$. To check the finite-size effect of the system, we
have studied the ground state energy per site $E_0/N$ and the
absolute value of the corresponding second derivative
$|d^2E/d\alpha^2|$ with respect to $\alpha$, with system size
$N=10-1000$ and $\beta=1$, as shown in
Fig.\ref{Fig:Energy_Gradient}. Such the ground state energy $E_0/N$
starts to converge, and the sharp peak of $|d^2E_0/d\alpha^2|$
appears at $N\geq 60$. Besides the ground state energy, we have also
calculated the correlation functions, and find that the finite-size
effect can be neglected at $N\geq 600$. Therefore, in the following
calculation, $N=600$ is adopted, for which the finite-size effect
can be completely neglected.
\begin{figure}[tbp]
\centerline{
\includegraphics[height=4.2in,width=3.6in]{Fig3_SCor_N600_A10.eps}
}
\caption{(color online) The real space spin-spin correlation
functions $\langle\textbf{S}_1\cdot \textbf{S}_r\rangle$ at
different $\alpha$, obtained by the DMRG with $N=600$ and
$\beta=1.0$. \label{Fig:SCor_N600}}
\end{figure}
Consistent with the discussion above, the existence of TQPT is
clearly supported by the numerical calculation, as shown in Fig.
\ref{Fig:Energy_Gradient}. The second derivative of the ground
energy has a sharp peak around $\alpha_c=0.80$, which indicates the
second order quantum phase transition addressed above. This is
further confirmed by the real-space spin-spin correlation function,
as shown in Fig.\ref{Fig:SCor_N600}. At $\alpha=0$ and $1$, the
correlation function is exactly the same with the analytical result.
When $\alpha<0.80$, the correlation behaves similarly with the AKLT
model, showing an oscillating behavior with respect the lattice
separation. However, when $\alpha>0.80$, the correlation behaves
similarly with the SZH model, showing an negative-definite behavior.
At the critical point, the correlation function undergoes an
qualitative change from AKLT to SZH.
To get the ground state phase diagram, we also calculate the peak
position $\alpha_c$ of $|d^2E/d\alpha^2|$ as a function of $\beta$,
as shown in Fig.\ref{Fig:dE_Dimer}. Above the red line, the system
is topologically connected to SZH model, and therefore belongs to
the same topological class as SZH phase. One feature of this phase
is translational symmetric and has spin-$3/2$ excitations on the
edge. Similarly, the regions where $\alpha$ is small belong to the
same topological class as AKLT phase. Translational symmetry is also
respected, but with spin-$1$ excitations on the edge.
\begin{figure}[tbp]
\centerline{
\includegraphics[height=3in,width=3.6in]{dimer.eps}
}
\caption{(color online) Dimer order parameters of the model
Hamiltonian with different choices of $\alpha$ and $\beta$.
(\textbf{a})The existence of dimerization is established by choosing
different system sizes $N$. (\textbf{b})Dimer order parameters for
different $\beta$. This parameter disappears for the critical
$\beta=0.27$.}\label{dimer}
\end{figure}
\begin{figure}[tbp]
\centerline{
\includegraphics[height=2.4in,width=3.6in]{Fig4_dE_Dimer.eps}
}
\caption{(color online) Ground state phase diagram of the model
Hamiltonian got by DMRG with $N=600$. The location of second order
phase transition for different $\beta$ is indicated by the red line.
The yellow region above red line belongs to SZH topological class,
while the light blue region belongs to AKLT class. Possible
dimerization is shown in the green region sandwiched between these
two phases. \label{Fig:dE_Dimer}}
\end{figure}
However, one thing we should keep in mind that even the model
Hamiltonian is translational invariant, spontaneous symmetry
breaking is also possible. One well known example is the
biliear-biquadratic model for spin-1 chain, whose Hamiltonian is
also written in terms of local projections, and translational
invariant. However, the emergence of the dimerized phase is
addressed in previous
works\cite{Klumper1990}\cite{Batchelor1990}\cite{Kennedy1992a}. In
order to quantitatively describe this issue, the following dimer
order parameter
\begin{equation}
D_\alpha=|\langle \mathbf{S}_i\mathbf{S}_{i+1}\rangle - \langle
\mathbf{S}_{i+1}\mathbf{S}_{i+2}\rangle|
\end{equation}
as a function of $\alpha$ is introduced, where $i$ labels the center
site in the spin chain so that one can minimize the possible
finite-size effect induced by open boundary condition. In the
calculation, the total site number $N$ is set to be an even number
to avoid potential ambiguity in the definition, in which case $i$ is
simply $N/2$. The indications of dimerization is plotted in
Fig.\ref{dimer}. It is shown explicitly that dimerization appears
with proper choices of parameters $\alpha$ and $\beta$. One can rule
out possible finite size effect as the dimer order parameter doesn't
scale with the system size $N$, shown in Fig.\ref{dimer}(a). When
$\beta$ decreases, the maximum amplitude of dimer order parameter
decreases as well, and the dimerization expands less and less
regions of $\alpha$ monotonously. At the critical point
$\beta=0.27$, dimer order parameter vanishes for any $\alpha$, and
the corresponding dimerization phase shrinks as well. The system
undergoes a quantum phase transition without spontaneous breaking of
translational symmetry. During this phase transition, the energy and
spin-spin correlation function have the same form as shown in Fig.
(\ref{Fig:Energy_Gradient}) and Fig. (\ref{Fig:SCor_N600}). However
it's worth mentioning that for Neel state of spin-2 chain, the
corresponding dimer order parameter would be 8. Therefore our dimer
order parameter is pretty small compared to strict antiferromagnets,
and the dimerization phase cannot be confirmed definitely. One
possibility is the gap between excitation state and ground state for
the present model is too small to be distinguished numerically. As a
result, certain dimerized excitation state enters into our results
and lead to such finite but tiny dimer order parameter. A promising
solution is to apply the periodic boundary condition (PBC) here, so
that one can rule out the possibility of dimerization acquired from
open boundary condition. Therefore, we have also done some
calculation by DMRG for system with PBC. For the system size
$N=100-200$ sites, we keep up to $m=3000$ states with truncation
error smaller than $10^{-8}$. Finally, we find that both OBC and PBC
systems give us consistent results.
Observed the possible presence of dimerization, one can readily work
out the phase diagram, see Fig.\ref{Fig:dE_Dimer}. The red curve
stands for critical points of $|d^2E/d\alpha^2|$, while the two
dashed dark curves are upper/lower bounds of dimerized phase of each
$\beta$ value. It shows the onset value for second order derivative
coincides with the upper bound of dimerized phase, so that they
describe the same phase transition between dimerized phase and SZH
phase. While on the AKLT side, the phase transition would be of
higher order that is invisible in Fig.\ref{Fig:Energy_Gradient}.
Despite the possible presence of dimerized phase, the system
undergoes a quantum phase transition without spontaneous breaking of
translational symmetry at $\alpha_c$ when $\beta=0.27$. As the
symmetry is unchanged in this case, this phase transition is
originated from the topology only.
In conclusion, the topological distinction of the AKLT model and SZH
model for the $S=2$ spin chain is presented in this work. A model
Hamiltonian as an interpolation between these two models is
introduced. The quantum phase transition of the model is protected
by topology, and established by DMRG calculation. The results
indicate the presence of a dimered phase between two topological
phases of AKLT and SZH. This dimered phase shrinks at critical value
of $\beta_c=0.27$. One would like to realize this topological phase
transition in real materials. On the AKLT side, we have a well
defined example already. $S=1$ edge spin is observed in $S=2$ chain
of CsCr$_{1-x}$Mg$_x$Cl$_3$\cite{Yamazaki1996a}. However, examples
on SZH side are still missing. Probably one can employ cold atom
techniques to realize TQPT proposed in this paper.
We thank Zheng-Cheng Gu, Hosho Katsura, Naoto Nagaosa, and Yong-Shi
Wu for insightful discussions. This work is supported by Ministry of
Education of China under the Grant No. B06011, the NSF of China, the
National Program for Basic Research of MOST-China, and SCZ is
supported by the NSF under grant numbers DMR-0904264. HCJ
acknowledges funding from Microsoft Station Q. In finishing this
work, we became aware of a parallel work\cite{Zheng2010a} which has
reached similar conclusion based on different approaches. They get a
brilliant result about the universality class of phase transition,
but the possible dimer order phase is not mentioned.
|
\section{Introduction}\label{introsct}
In this paper we investigate the distribution of zeros of higher derivatives of the Riemann zeta function.
In order to put our main results in perspective,
we first give a summary of some of the main results and conjectures in this area.
Let $s = \sigma + it$. For all $k \in \mathbb{N}$ the $k$-th derivative of the Riemann zeta function $\zeta^{(k)}(s)$ is
\begin{equation}\label{eq1}
\zeta^{(k)}(s) = (-1)^k \sum_{n=2}^\infty \frac{\log^k n}{n^s}, \; \; \; \mathrm{for} \; \; \; \sigma > 1,
\end{equation}
and can be extended to a meromorphic function on $\mathbb{C}$, with a single pole (of order $k$) at the point $s=1$. However, unlike
$\zeta(s)$ itself, the functions $\zeta^{(k)}(s)$ have neither Euler products nor functional equations.
Their non-trivial zeros do not lie on a line, but appear to be distributed (seemingly at random) to the right of the critical line $\sigma = \frac12$.
In 1934 Speiser \cite{s:5} was the first to show that the Riemann Hypothesis (denoted RH everywhere below) implies that $\zeta'(s)$ has no zeros
to the left of the critical line $\sigma = \frac12$.
Unfortunately, for higher derivatives this particular property does not stay true. But, in 1974, Levinson and Montgomery \cite{lm:1} showed (again, assuming the RH) that the number of zeros of $\zeta^{(k)}(s)$ in the left half-plane is always finite. More recently
Y\i ld\i r\i m \cite{y:2} proved that both $\zeta''(s)$ and $\zeta'''(s)$ have exactly one pair of non-trivial zeros with $\sigma<0$,
namely $\zeta''(s)$ has a zero at approximately $-0.35508433021 \pm 3.590839324398 i$. He also showed that the RH implies that neither $\zeta''(s)$ nor $\zeta'''(s)$ have any zeros
$\rho$ with $0 < \Re(\rho) <\frac12$.
\begin{figure}[ht]
\caption{Zeros of $\zeta'(s)$ in $\mathbb{C}$, with the zero-free region.}\label{figd3}
\includegraphics[width=12cm]{zeta-zerofree-k1.png}
\end{figure}
In regions to the right of the critical line, i.e. for $\sigma\geq\frac12$, the total number
of zeros of $\zeta^{(k)}(s)$ does not differ by much from the number of zeros of $\zeta(s)$.
In fact, if we let $N(T)$ and $N_k(T)$ denote
the number of such zeros $\rho$ with $0 \leq \Im(\rho) \leq T$ of $\zeta(s)$ and $\zeta^{(k)}(s)$, respectively, then according to Berndt \cite{b:2}
\begin{equation}\label{eqberndt}
N_k(T) = N(T) - \frac{T}{2 \pi} \log 2 + O(\log T),
\end{equation}
where, by the classical Riemann-von Mangoldt formula (see Landau \cite{l:1}),
\[
N(T) = \frac{T}{2 \pi} \log \frac{T}{2 \pi} - \frac{T}{2 \pi} + O(\log T).
\]
Most non-trivial zeros of $\zeta^{(k)}(s)$ are located close to the line $s = \frac{1}{2} + it$.
Soundararajan \cite{s:4} showed that, for $k=1$, a positive portion of the zeros $\rho$ of $\zeta'(s)$ satisfies $\Re(\rho) < \frac{1}{2} + c/\log T$.
Nevertheless many of the zeros of $\zeta^{(k)}(s)$ lie further to the right, even though their real parts can be bounded from above.
For $k\geq 3$ such upper bounds were given by Spira \cite{s:5a} in 1965. They were improved by Verma and Kaur \cite{vk:1} (see \ourref{Table}{tablebound}).
\begin{table}[ht]\label{tablebound}
\caption{Lower real bounds for zero-free regions in the right half-plane.}
\begin{tabular}{l|c|c|c|c}
& $\;\;\zeta\;\;$ & $\zeta'$ &$\zeta''$ & $\zeta^{(k)}$ for $k\ge 3$\\
\hline
Hadamard \cite{h:1}, de la Vall{\' e}e-Poussin \cite{v:1} & $1$ &&&\\
Titchmarsh \cite{t:1}& & $3$ &&\\
Spira \cite{s:5a}& & & & $\frac{7}{4}k+2$\\
Verma \& Kaur \cite{vk:1}& & & & $1.13588 k+2$\\
Skorokhodov \cite{s:3} & & $2.93938$ & $4.02853$ &\\
\end{tabular}
\end{table}
In this paper we explicate some new, unexpected properties
of the location of zeros of$\zeta^{(k)}(s)$ with $\frac{1}{2} \leq \Re(s) <1\point13588 k+2$.
In particular we prove the existence of zero-free regions for $\zeta^{(k)}$ and
show that the zeros exhibit a fascinating vertical periodicity between these zero-free regions, which we call critical strips in analogy to the critical strip of zeta.
This enables us to give exact formulas for their number, while also proving that all zeros of $\zeta^{(k)}(s)$ inside them are {\em simple}.
\ourref{Figure}{figd100} illustrates these phenomena for $\zeta^{(38)}(s)$.
\begin{figure}[ht]
\caption{Zeros of $\zeta^{(38)}(s)$ in $\mathbb{C}$, with zero-free regions (characterized by the dominance of $Q_M^{38}(s)$ for $M=2$ and $3$)} \label{figd100}
\includegraphics[width=13cm]{zeta-zerofree-k38.png}
\end{figure}
\section{Statement of Main Results}
In order to state our results precisely, we need to introduce some notation and definitions.
Let $Q^{k}_n(s):=(\log n)^k/n^s$ denote the $n$-th term of the Dirichlet series (\ref{eq1}) for $\zeta^{(k)}(s)$.
All the previously known zero-free regions for $\zeta^{(k)}(s)$ have been obtained by finding solutions to
\[
\left|\zeta^{(k)}(s)\right|=\left|\sum_{n=2}^\infty Q^{k}_n(s)\right|\ge Q^{k}_2(\sigma) - \sum_{n=3}^\infty Q^{k}_n(\sigma)>0,
\]
or some variation thereof (see \cite{s:3,t:1,vk:1}); that is, by finding the regions of the complex plane where the first non-zero term $Q^{k}_2(s)$ dominates all the other terms of the expansion \eqref{eq1} of $\zeta^{(k)}(s)$
(in the sense that $Q^{k}_2(s)$ is greater {\em in modulus} than the rest of the terms combined). Evidently, these conditions imply $\zeta^{(k)}(s)\not=0$. However, $Q^{k}_2(s)$ is not always the dominant term; any other term can take this role.
This is clear from the fact that $|Q^{k}_n(s)|=Q^{k}_n(\sigma)$ viewed as a function of $n$ has its global maximum at $n=e^{k/\sigma}$.
Using this argument one can show the existence of regions where $Q^{k}_n(s)$, $n\ge2$ is the dominant term of \eqref{eq1},
which then provides us with a new zero-free region of $\zeta^{(k)}(s)$, for each $n$, for every sufficiently large $k$.
By $Q^{k}_M(s)$ we denote the term of \eqref{eq1} which has the largest modulus. As we prove in \ourref{Lemma}{mon}, one important property is that if we fix some such M, then the moduli of the terms of \eqref{eq1} on the left side of $Q^{k}_M(s)$ grow, and on the right side of $Q^{k}_M(s)$ decrease, always in monotone fashion. Since no term $Q^{k}_M(s)$ can attain dominance on a line where its absolute value is equal to that of another term (and by the aforesaid property this can only happen when $Q^{k}_M(\sigma) = Q^{k}_{M+1}(\sigma)$ or $Q^{k}_M(\sigma) = Q^{k}_{M-1}(\sigma)$), it is reasonable to expect that the zeros of $\zeta^{(k)}(s)$ will be located close to these lines. For this purpose, we define
\begin{equation}\label{eq4}
q_M:=\frac{\log \left( \frac{\log M}{\log M+1} \right)}{\log \left( \frac{M}{M+1} \right)},
\end{equation}
so that $Q^{k}_M(\sigma)=Q^{k}_{M+1}(\sigma)$ precisely at the line $\sigma=k q_M$. In particular, we have:
\[
q_2 \approx 1.13588, \quad
q_3 \approx 0.808484, \quad
q_4 \approx 0.668855.
\]
Note that $q_2$ is the constant of \cite{vk:1} that appears in \ourref{Table}{tablebound}.
In the above notation, our first main result can be stated as follows.
\begin{theorem}\label{thmone}
(a) {\em (The case of $M = 3$)} Let $k\in\N$. We have $\zeta^k(s)\ne 0$ for
\[
q_3 k + 4\log 3 \le\sigma\le q_2 k - 2.
\]
(b) {\em (The case of $M > 3$)} If $M\in\N$ and $M\ge 3$, then $\zeta^k(s)\ne 0$ for
\[
q_M k+(M+1)\log 3\le\sigma\le q_{M-1} k -M\log 3.
\]
\end{theorem}
\begin{figure}[ht]
\caption{Zero-free regions of $\zeta^{(k)}(\sigma+it)$,
for $M=2,\dots,9$.}\label{fzf1}
\includegraphics[width=14cm]{zeta-zerofree-wedges.png}
\end{figure}
For $k\ge 3$ and $M\ge 2$ we define the {\em critical strip} $S_M^{k}$ of $\zeta^{(k)}(s)$ as the region between
the lines $\sigma = q_{M} k - (M+1)\log 3$ and $\sigma = q_M k + (M+1)\log 3$,
as long as $q_{M+1} k + (M+2)\log 3<q_{M} k - (M+1)\log 3$.
The {\em critical line} of $S_M^k$ is given by $\sigma = q_M k$.
A way to visualize the critical strips $S_M^k$ is to consider their location in the $\sigma k$-plane (see \ourref{Figure}{fzf1}).
In this graphical representation, the wedge-shaped regions correspond to the zero-free regions, i.e.\ the regions of dominance of the terms $\mtau{M}{s}{k}$ (for $M = 2$ this
is Verma and Kaur \cite{vk:1}, for $M \geq 3$ it is new), while the critical strips $S_M^k$ are the narrow regions centered around the lines $\sigma = q_M k$ that separate the wedges.
The tips of the wedges are at
\[
k_M = \frac{(2M+1)\log 3}{q_{M-1}-q_{M}},
\]
which means that the first critical strips $S_1^k$ can be observed for all $k\ge 14$, the second $S_2^k$ for all $k \geq 41$, and the third $S_3^k$ for all $k \geq 87$.
It is an interesting corollary to \ourref{Theorem}{thmone} that, for
the number $c(k)$ of critical strips of $\zeta^{(k)}(s)$ inside the region $1/2 \leq \sigma < q_2 k + 2$, we have
\begin{equation}\label{eq6}
\frac{\sqrt{k}}{3\log k} < c(k) < \frac{2\sqrt{k}}{\log k}.
\end{equation}
If we also consider the imaginary parts of $Q_M^{k}(q_M k+it)=Q_{M+1}^{k}(q_M k+it)$, then we obtain the solutions:
\begin{equation}\label{eqconjt}
t= (2j+1) \frac{\pi}{\log(M+1)-\log(M)}
\end{equation}
for $j\in\Z$, showing that the precise location of the zeros $\rho$ inside $S_M^k$ should be close to
\[
k\cdot q_{M}+\frac{(2j+1)\pi}{\log \left( \frac{M+1}{M} \right) }\cdot i
\]
for some $j\in\N$. This suggests existence of an amazing vertical periodicity (in the limit) of the zeros of $\zeta^{(k)}(s)$ at the critical lines, with the {\em periods} $\frac{\pi}{\log(M+1)-\log(M)}$.
Although it is virtually impossible to give exact location of every transcendental zero in a given critical strip (and describe the way it approaches the limiting values with growing $k$),
we are at least able to separate the zeros by horizontal line segments, whose imaginary parts lie between the values for $t$ in (\ref{eqconjt}).
That is, we first establish that $\zeta^{(k)}(s)\ne 0$, for
\[
s=\sigma+\frac{2\pi j}{\log(M+1)-\log M},
\]
where $q_M k-(M+1)\log 3\le \sigma \le q_M k+(M+1)\log 3$, and then (with the help of Rouch\'e's theorem) we show that between every two consecutive lines that horizontally partition the critical strip $S_M^k$ this way there is exactly one zero of $\zeta^{(k)}(s)$. In other words:
\begin{theorem}\label{thmboxzero}
(a) Let $j\in\N$. Then each rectangular region $R \subset S_M^k$, consisting of all $s = \sigma + it$ with
\[
q_M k-(M+1)\log 3<\sigma < q_M k+(M+1)\log 3
\]
and
\[
\frac{2\pi j}{\log(M+1)-\log(M)}<t<\frac{2\pi(j+1)}{\log(M+1)-\log(M)}
\]
contains exactly one zero of $\zeta^{(k)}(s)$.
(b) For all $M \geq 2$ and $k \in \N$, all the zeros of $\zeta^{(k)}(s)$ inside $S_M^k$ are simple.
\end{theorem}
Clearly, \ourref{Theorem}{thmboxzero}~(a) can be converted into an exact formula for the number of zeros of $\zeta^{(k)}(s)$ (for carefully chosen values of $T$) inside any
given critical strip.
\begin{corollary}\label{corcount}
Let $N_M^k(T)$ denote the number of zeros $\rho$ of $\zeta^{(k)}(s)$ (with $\Im (\rho) \leq T$) inside the critical strip $S_M^k$. Then, for all $j \geq 1$,
\[
N_M^k \left(\frac{2\pi j}{\log(M+1)-\log(M)} \right)=j.
\]
\end{corollary}
\begin{remark}
As an immediate consequence of this result we have: For all $k \geq 3$, and all $T > 0$,
\[
N_M^k(T) = \frac{\log(M+1)-\log(M)}{\pi} T + O(1).
\]
This implies that, for any given $k \geq 3$, the total number of zeros contained within all the critical strips is $O(T)$,
so always $o(N_k(T))$.
\end{remark}
Finally, noticing -- as we have in our last remark -- that the important formula (\ref{eqconjt}), that describes the vertical quasi-periodicity of zeros of $\zeta^{(k)}(s)$, only contains $M$, and is independent of $k$, we realize that, with growing $k$, the critical strips $\{ S_M^k \}_{k=2}^{\infty}$ can undergo a shift in one direction only: to the right, and with the length of the shift very close to $q_M$ for each increment of $k$.
In other words, from \ourref{Theorem}{thmboxzero} we can see that all zeros of $\zeta^{(k)}(s)$ contained in a given critical strip $S_M^k$ will keep shifting (almost) linearly, and with a (almost) fixed shift, the period, $q_M$, to the right, as $k$ grows to infinity.
A simple consequence of this observation is the following:
\begin{conjecture}
For all $k\in\N$ there is a one-to-one correspondence between the non-trivial zeros of
$\zeta^{(k)}(s)$ and $\zeta^{(k+1)}(s)$, where the zeros of $\zeta^{(k+1)}(s)$ always stay to the right of the
corresponding zeros of $\zeta^{(k)}(s)$.
\end{conjecture}
\begin{remark}
Spira \cite{s:5a} had already noticed that the zeros of $\zeta'(s)$ and $\zeta''(s)$ seem to come in pairs,
where the zero of $\zeta''(s)$ was always located to the right of the zero of $\zeta'(s)$. With the help of extensive
computations Skorokhodov \cite{s:3} observed this behavior for higher derivatives.
We conjecture not only the one-to-one correspondence, but also (for every fixed $M$ -- the existence of a quasi-lattice of zeros
of $\zeta^{(k)}(s)$, created as $k=1, 2, 3, \cdots$
\end{remark}
\begin{remark}
The zero-free regions obtained in \ourref{Theorem}{thmone} easily generalize to a large class of Dirichlet series.
Since, in our proofs of the zero-free regions for $\zeta^{(k)}(s)$, we only consider the absolute values of its coefficients,
it follows that if $L(s)=\sum_{n=1}^\infty \frac{a_n}{n^s}$, and $|a_M| \ge |a_n|$ for some $M\ge 3$ and
all $n\ge 2$, then $L^{k}(s)\ne 0$ for
$ q_M k + cM \le\sigma\le q_{M-1} k - c(M-1)$, for some computable constant $c > 0$. Our other results are slightly harder
to extend and state in full generality. We relegate those investigations to a future project.
\end{remark}
\section{Two Auxiliary Lemmas}
We consider the $\sigma k-$plane interpretation of \ourref{Theorem}{thmone}.
In general, the wedges in \ourref{Figure}{fzf1} are the sets containing all points $(k, \sigma)$ that satisfy
\[
q_M k + b_1 < \sigma < q_{M-1} k + b_2.
\]
for some $M\in\N$ and constants $b_1$ and $b_2$. This implies that if $q_{M}k+b_1\le\sigma\le q_{M-1}k+b_2$, then also
\begin{equation}\label{eq9}
k\ge \frac{b_1-b_2}{q_{M-1}-q_{M}},
\end{equation}
with equality holding precisely when $k = k_M$, a point where the tip of the wedge is located.
This fact will be of importance in the proof of (\ref{eq6}) (see \ourref{Corollary}{corcountstrips}).
The growth properties of $q_n$ play an important role in understanding the critical strips $S_m^k$:
\begin{lemma}\label{lemqlog}
For all $n \geq 3$ we have
\[
\frac{1}{\log n} \; \leq \; q_{n-1} \; \leq \; \frac{1}{\log (n - 1)}.
\]
\begin{proof}
In order to prove the lower bound, we write:
\begin{align*}
\alpha_{n-1} &: = \frac{ \log (n-1)}{\log n} = 1 + \frac{\log (n -1) - \log n}{\log n} = 1 + \frac{\log(\frac{n-1}{n})}{\log n}, \\
\beta_{n-1} &:= \log (\alpha_{n-1}) = \log \left(1 + \frac{\log(\frac{n-1}{n})}{\log n} \right) < \frac{\log(\frac{n -1}{n})}{\log n}, \\
\end{align*}
the last inequality holds because $\log (1 + x) < x$, for all $x > -1$. The desired lower bound follows from $ q_{n-1} = \beta_{n-1}/\log((n-1)/n)$.
In order to prove the upper bound, we write:
\begin{align*}
q_{n-1} &:= \frac{ \log \left(\frac{\log (n-1)}{\log n} \right) }{\log \left(\frac{n-1}{n} \right)} = \frac{ \log \left(1 - \frac{- \log \left( \frac{n-1}{n} \right)}{\log n} \right)}{\log \left(\frac{n-1}{n} \right)} = \frac{ \log \left(1 - \frac{- \log \left( \frac{n-1}{n} \right)}{\log n} \right)}{\log \left(\frac{n-1}{n} \right)} \\
&= \frac{1}{\log n} +
\frac{1}{2}
\frac{\log \left(\frac{n-1}{n} \right)}
{(\log n)^2}
+ \frac{1}{3}
\frac{\log \left(\frac{n-1}{n} \right)^2}
{(\log n)^3}+ \cdots < \frac{1}{\log n} + \frac{2}{3(\log n)^2} < \frac{1}{\log (n-1)}, \\
\end{align*}
for all $n \geq 3$, giving us the desired bound.
\end{proof}
\end{lemma}
\begin{corollary}\label{corcountstrips}
For all $k$, the number $c(k)$ of critical strips of $\zeta^{(k)}(s)$ is bounded by
\[
\frac{\sqrt{k}}{3 \log k} < c(k) < \frac{2 \sqrt{k}}{\log k}.\eqno{(\ref{eq6})}
\]
\begin{proof}
We count the number of wedges given by $q_m k + (m+1)\log 3 \leq \sigma \leq q_{m-1} k - m\log 3$, which is equal to the number $c(k)$ of critical strips $S_m^k$.
The tips of the wedges are at
\[
k_m = \frac{(2m+1)\log 3}{q_{m-1}-q_{m}} \geq
\frac{(2m+1)\log 3}{\frac{1}{\log (m-2)} - \frac{1}{\log m}} =
\frac{(2m+1)\log 3}{\frac{\frac{2}{m} + \frac{4}{2m^2} + \frac{8}{3m^3} + \cdots }{\log (m-2) \log m}} > \frac{1}{2} \log 3\cdot m^2 (\log m)^2.
\]
Since every $m$ that satisfies the above inequality corresponds to a unique critical strip
$S_m^k$, it follows that inverting this relationship will give us the wanted upper bound
(\ref{eq6}) on $c(k)$. For the intrinsically more interesting lower bounds, we similarly have:
\[k_m = \frac{(2m+1)\log 3}{q_{m-1} - q_{m}} \leq \frac{(2m+1)\log 3}{\frac{1}{\log (m-1)} - \frac{1}{\log m} - \frac{2}{3m(\log m)^2}} <
6\log 3\cdot m^2 (\log m)^2,
\]
from which, by the same inversion, we obtain the desired result.
\end{proof}
\end{corollary}
For a fixed $M \geq 3$ and $k \in \mathbb{N}$, with the help of the above lemma we can now zoom in on the lines $\sigma = q_M k$ (the expected critical lines), and prove that in their viscinity one has a monotonically (in the modulus) growth of terms of the Dirichlet series (\ref{eq1}):
\begin{lemma}\label{mon}
Let $M$ be fixed, let $2 \leq n < M-1$, and $Q_n^k(\sigma) = \frac{\log^k n}{n^{\sigma}}$ be the $n$-th term of (\ref{eq1}). Then at the line
$\sigma = q_M k$, either $Q_M^k(\sigma)$ or $Q_{M-1}^k(\sigma)$ is the term with the largest modulus.
\begin{proof}
First: $0 = Q_1^k(\sigma) < Q_2^k(\sigma) < Q_3^k(\sigma)$, since for all $k > 1$, by \ourref{Lemma}{lemqlog}, we clearly have
\[
\frac{\log 2}{\log 3} < \left( \frac{2}{3} \right)^{q_M}
\; \; \; \Longleftrightarrow \; \; \; Q_2^k(q_M k) = \left( \frac{\log 2}{2^{q_M}} \right)^k < \left( \frac{\log 3}{3^{q_M}} \right)^k = Q_3^k(q_M k).
\]
Moreover, for $\sigma = q_M k$, the function $z(x) = \frac{\log^k x}{x^{\sigma}}$ has a single maximum, precisely at:
\begin{align*}
0 = z^{\prime}(x) = \left( \left( \frac{\log x}{x^{q_M}} \right)^k \right)^{\prime} &=
k \left( \frac{\log x}{x^{q_M}} \right)^{\prime} \left( \frac{\log x}{x^{q_M}} \right)^{k-1} \\
&= k \left( \frac{x^{q_M - 1} - q_M (\log x) x^{q_M - 1}}{x^{2q_M}} \right) \left( \frac{\log x}{x^{q_M}} \right)^{k-1},
\end{align*}
i.e.\ at an $x$ for which we have
\[
x^{q_M - 1} - q_M (\log x) x^{q_M - 1} = 0 \; \; \; \Longleftrightarrow \; \; \; q_M = \frac{1}{\log x}
\; \; \; \Longleftrightarrow \; \; \; x = \exp(q_M^{-1}),
\]
which, by \ourref{Lemma}{lemqlog}, implies $M-1 < x$, and that proves the result.
\end{proof}
\end{lemma}
\section{Proof of \ourref{Theorem}{thmone}}
Now we are ready to prove our first main result. We will show that $\zeta^{(k)}(s)$ has no zeros if $(k,\sigma)$ in the $\sigma k$-plane lies in one of the wedges
given by an inequality of the form
\[
q_M k+b_1\le\sigma\le q_{M-1}k+b_2
\]
for suitably chosen $b_1,b_2\in\mathbb{R}} \newcommand{\Z}{\mathbb{Z}$.
We choose $b_1,b_2$ such that these wedges are the regions where $Q^{k}_M(s)=\frac{\log^k M}{M^s}$ is
the dominant term (in the modulus) of $\zeta^{(k)}(s)$.
Everywhere hereafter we write $H^{k}_M(s)$ for the ``head'' and $T^{k}_M(s)$ for the ``tail'' of the series $\zeta^{(k)}(s)$ split by $Q_M^k(s)$:
\[
H^{k}_M(s) :=\sum_{n=2}^{M-1}Q_n^k(s)=\sum_{n=2}^{M-1}\frac{\log^k n}{n^s}\quad\mbox{ and }\quad
T^{k}_M(s) :=\sum_{n=M+1}^{\infty}Q_n^k(s)=\sum_{n=M+1}^{\infty}\frac{\log^k n}{n^s}.
\]
Our goal will be to show that
\[
|\zeta^{(k)}(s)|\ge Q^{k}_M(\sigma)-H^{k}_M(\sigma)-T^{k}_M(\sigma)=
Q^{k}_M(\sigma)\left(1-\frac{H^{k}_M}{Q^k_M}(\sigma)-\frac{T^{k}_M}{Q^k_M}(\sigma)\right)>0
\]
for our choice of $b_1$ and $b_2$, keeping in mind that
\[
\frac{Q_{M+1}}{Q_M}(q_Mk+b_1)=\left(\frac{M}{M+1}\right)^{b_1}
\mbox{ and }
\frac{Q_{M-1}}{Q_M}(q_{M-1}k+b_2)=\left(\frac{M}{M-1}\right)^{b_2},
\]
as one can easily verify.
\subsection*{The Tails}
We first find an upper bound for the tails $T^{k}_M(\sigma)$.
\begin{lemma}\label{intbd}
Fix some integer $M\ge2$, and assume $k - 1 < (\sigma - 1)\logM.$ Then
\begin{equation}\label{intbdi}
T_M^{k}(\sigma) = \sum_{n=M+1}^\infty\frac{\log^kn}{n^\sigma}\le \int_{M}^\infty \mtau{x}{\sigma}{k} dx < Q_M^k(\sigma)R_M^k(\sigma).
\end{equation}
where
\[
R_M^{k} (\sigma) = \frac{M}{\sigma-1} \left( 1 + \frac{k}{(\sigma-1)\logM - k + 1} \right) .
\]
\end{lemma}
\begin{proof}
For $k \in \Z$, the integral in \eqref{intbdi} can be written in a closed form.
Applying recursively the general formula (for all $b, -a \neq -1$):
$ \int \frac{(\log x)^a}{x^b} \; dx = - \frac{(\log x)^a}{(b-1)x^{b-1}} + \frac{a}{b-1} \int \frac{(\log x)^{a-1}}{x^b} \; dx, $
we obtain
\begin{align*}
\int_{M}^\infty \mtau{x}{\sigma}{k} dx &=
\mtau{M}{\sigma}{k} \frac{M}{\sigma-1}
\sum_{r=0}^k \frac{k!}{(k-r)!} \frac{\log^{-r} M}{(\sigma-1)^r}\\
&\le Q_M^k(\sigma) \frac{M}{\sigma-1}
\Bigl(
1 + \sum_{r=1}^k k (k-1)^{r-1}
\Bigl( \frac{1}{(\sigma-1) \logM} \Bigr)^r
\Bigr)\\
&< Q_M^k(\sigma) \frac{M}{\sigma-1}
\Bigl(
1 + \frac{k}{(\sigma-1)\logM} \sum_{r=0}^\infty
\Bigl(
\frac{k-1}{(\sigma-1)\logM}
\Bigr)^r
\Bigr)\\
&= Q_M^k(\sigma) \frac{M}{\sigma-1}
\Bigl(
1 + \frac{k}{(\sigma-1)\logM - k + 1}
\Bigr),
\end{align*}
where the convergence of the geometric series is implied by $k - 1 < (\sigma - 1)\logM$.
\end{proof}
It is clear why estimating $R_M^{k}(\sigma)$ will be vital for the proofs of our theorems. We note:
\begin{lemma}\label{3}
If $a_1k+b_1\le\sigma$, and $k \geq k_M$, then
\begin{equation}\label{9}
R_M^{k}(\sigma) \le
R_M^{k}(a_1 k + b_1) \le
R_M^{k_M}(a_1 k_M + b_1),
\end{equation}
as long as the following two conditions are satisfied:
\
a_1 > \frac{1}{\logM} \; \; \; \mathrm{and} \; \; \; (a_1\logM - 1) k_M + 1 + (b_1 - 1)\logM > 0,
\]
and in the case of $b_1 < 1 - 1/\logM$ also:
\
k_M \ge \frac{1}{a_1\logM}\left(
-(b_1-1)\logM - 1 + \sqrt{\frac{|(b_1-1)\logM + 1|}{a_1\logM - 1}}
\right).
\]
\begin{proof}
The left-hand inequality of (\ref{9}) is evident from the fact that $T_M^k(\sigma)$ is decreasing when viewed as a function of $\sigma$ alone.
The right-hand inequality of (\ref{9}) is equivalent to saying that $R^{k}_M(\sigma)$ is decreasing as a function of $k$.
To see why this is the case, just notice that if we rewrite $T_M^k(q_M k + b_1)$ in the form
\[ y(k) = \frac{1}{(c+1)k+d-1} \frac{(c+1)k+d}{ck+d}, \]
where $c:=a_1\logM-1>0$ and $d:=1 + (b_1-1)\logM$, then clearly
\[
y'(k) = - \frac{c (1 + c)^2 k^2 + 2 c d k (1 + c) + d(1 + c d)}{((c+1)k + d - 1)^2 (ck+d)^2},
\]
from which it is easy to see that $y'(k)$ can change sign only if $d<0$ (otherwise it remains nonpositive).
However, the condition $d<0$ translates to $b_1<1-1/\logM$, in which case one requires $k_M \ge z_0$, where
\[
z_0 := -\frac{d}{1+c} + \frac{1}{1+c} \sqrt{\frac{|d|}{c}}
\]
is the right zero of the numerator of the above expression for $y'(k)$.
\end{proof}
\end{lemma}
The way the estimate for $T_M^{k}(\sigma)$ from \ourref{Lemma}{intbd} will be used in the proof of \ourref{Theorem}{thmone} is via the separation:
\[
T_M^{k}(\sigma)=Q_M^{k}(\sigma)+T_{M+1}^k(\sigma)\le Q_M^k(\sigma)(1+R_{M+1}^k(\sigma))
\le Q_M^k(q_M k+b_1)(1+R_{M+1}^k(q_M +b_1)),
\]
where $R_{M+1}^k(q_M +b_1)$ will converge because $q_M>\frac{1}{\log (M+1)}$, by \ourref{Lemma}{lemqlog}.
Also, a corollary of \ourref{Lemma}{intbd} and \ourref{Lemma}{3} is a proof of the result of \cite{vk:1}. We include it here because it exemplifies several of the important ideas and illustrates some of the key workings of our general method, being the special case of $M=2$ (representing the dominance of the term $Q_2^k(\sigma)$).
\begin{theorem}[\cite{vk:1}]\label{thmvk}
For all $\sigma \geq q_2 k+2$ we have $\zeta^k(s)\ne 0$.
\begin{proof}
First write
\[
|\zeta^{(k)}(s)|\ge\frac{2^\sigma}{\log^k 2}-T_2^{k}(\sigma) \ge
Q_2^k(\sigma)
\left(1-
\frac{Q_3^k}{Q_2^k}(\sigma)-\frac{Q_4^k}{Q_2^k}(\sigma)\left(1+R_4^{k}(\sigma)\right)\right).
\]
By \ourref{Lemma}{3} we have $R_4^{k}(\sigma)\le R_4^{k}(q_2 k+2)<0.68$, for $k\ge 3$.
Furthermore,
\[
\frac{Q_4^k}{Q_2^k}(\sigma)
=\frac{2^\sigma}{\log^k 2}\frac{\log^k 4}{4^\sigma}
=\frac{2^k \log^k 2}{\log^k 2}\frac{1}{2^\sigma}
=2^{k-\sigma}\le 2^{k-q_2k+2}
\le 2^{3(1-q_2)-2}
\le 0.19.
\]
The quotient
$\frac{Q_3^k}{Q_2^k}(\sigma)$ is decreasing in $\sigma$, and hence
$\frac{Q_3^k}{Q_2^k}(\sigma)\le \frac{Q_3^k}{Q_2^k}(q_2 k+2)=\frac{4}{9}$. So we obtain
\[
1-\frac{Q_3^k}{Q_2^k}(\sigma)-\frac{Q_4^k}{Q_2^k}(\sigma)\left(1+R_4^{k}(\sigma)\right)
\ge 1-\frac{4}{9}-0.19(1+0.68)>0,
\]
which establishes the result.
\end{proof}
\end{theorem}
Since \ourref{Theorem}{thmone}~(a) deals with the next case of $M=3$ (corresponding to the dominance of the term $Q_3^k(\sigma)$), and only a little bit of extra effort is needed to cover it, we give a proof of it right now. The result states: $\zeta^k(s)\ne 0$, for
\[
q_3 k + 4\log 3 \le\sigma\le q_2 k - 2.
\]
\begin{proof}[{\bf Proof of \ourref{Theorem}{thmone}~(a)}]
Separating the dominant term $Q_3^k(\sigma)$, we get
\[
|\zeta^{(k)}(s)|\ge
Q_3^k(\sigma)-Q_2^k(\sigma)-T_3^{k}(\sigma)\ge
Q_3^k(\sigma)\left(1-\frac{Q_2^k}{Q_3^k}(\sigma)-\frac{Q_4^k}{Q_3^k}(\sigma)\left((1+R_4^{k}(\sigma)\right)\right).
\]
Therefore we only need to show that
\[
1-\frac{Q_2^k}{Q_3^k}(\sigma)-\frac{Q_4^k}{Q_3^k}(\sigma)\left((1+R_4^{k}(\sigma)\right) >0.
\]
From \ourref{Lemma}{3}, $R_4^{k}(\sigma)\le R_4^{k}(q_3 k+4)\le R_4^{k_3}(q_3 k_3+4)<0.72$, for $q_3 k+4\log 3\le \sigma$ and
$k\ge k_3=\frac{4\log 3+2}{q_2-q_3}$.
Also,
$\frac{Q_4^k}{Q_3^k}(\sigma)\le\frac{Q_4^k}{Q_3^k}(q_3 k + 4\log 3)<0.29$
and
$\frac{Q_2^k}{Q_3^k}(\sigma)\le\frac{Q_2^k}{Q_3^k}(q_2 k -2)<0.29$.
Hence
\[
1-\frac{Q_2^k}{Q_3^k}(\sigma)-\frac{Q_4^k}{Q_3^k}(\sigma)\left((1+R_4^{k}(\sigma)\right)
>1-0.45-0\point29(1+0.72)>0,
\]
as desired.
\end{proof}
\ourref{Theorem}{thmone}~(b) deals with the dominance of the general term $Q_M^k(\sigma)$, and consequently requires knowledge of the behavior of the sum of all the terms preceding it.
\subsection*{The Heads}
We rewrite the heads of the series \eqref{eq1} in the following form:
\begin{align*}
H^{k}_M(\sigma)
&=
Q^{k}_M(\sigma)\left({\frac{Q^{k}_{M-1}}{Q^{k}_{M}}(\sigma)+\frac{Q^{k}_{M-2}}{Q^{k}_{M}}(\sigma)+\dots+\frac{Q^{k}_{2}}{Q^{k}_{M}}(\sigma)}\right)\\
&=
Q^{k}_M(\sigma)\left({\frac{Q^{k}_{M-1}}{Q^{k}_{M}}(\sigma)
\left(1+\frac{Q^{k}_{M-2}}{Q^{k}_{M-1}}(\sigma)\left(1+\dots\left(1+\frac{Q^{k}_{2}}{Q^{k}_{3}}(\sigma)\right)\dots\right)\right)}\right),
\end{align*}
and we will find upper bounds for all the above quotients $\frac{Q^{k}_{n-1}}{Q^{k}_{n}}(\sigma)$ of consecutive terms.
Here, notice that $\frac{Q^{k}_{n-1}}{Q^{k}_{n}}(\sigma)$ is always increasing with $\sigma$.
For $2\le n\le M$, in the wedges (see \ourref{Figure}{fzf1}) delimited by $q_M k+b_2\le\sigma\le q_{M-1}k+b_2$, this yields
\[
\frac{Q^{k}_{n-1}}{Q^{k}_{n}}(\sigma)
\le
\frac{Q^{k}_{n-1}}{Q^{k}_{n}}(q_{M-1} k+b_2)
\le
\frac{Q^{k}_{n-1}}{Q^{k}_{n}}(q_{n-1} k+b_2)
=
\left( \frac{n}{n-1} \right)^{b_2},
\]
the second inequality holds because we have $q_{M-1}<q_n$ for all $n<M-1$, while the equality holds due to the fact that $\sigma=q_{n-1} k$ is the solution of
$
Q^{k}_n(\sigma)=
Q^{k}_{n-1}(\sigma)
$.
Therefore
\[
\frac{Q^{k}_{n-1}}{Q^{k}_n}(q_{n-2} k+b_2) = \left(\frac{n}{n-1}\right)^{b_2} = \left(\frac{n-1}{n}\right)^{-b_2}.
\]
Thus, in order for $H^{k}_M (\sigma)$ to stay bounded, we must choose $b_2<0$. It is not difficult to see that a choice of $b_2$ as a linear function of $M$ (i.e.\ $-b_2$ of the form $cM+d$, with $d\ge c$) is likely to work, since:
\begin{lemma}\label{lemcMd}
Let $c,d\in\mathbb{R}} \newcommand{\Z}{\mathbb{Z}$ and $0<c\le d$.
Then $y(M)=\left(\frac{M}{M+1}\right)^{c M+d}$ is monotonously increasing with the asymptote $1/e^c$.
\begin{proof}
As $\lim_{M\to\infty} \left(1+\frac{1}{M}\right)^{cM}=e^c$ and $\lim_{M\to\infty}\left(\frac{M}{M+1}\right)^d=1$, for all fixed $d > 0$,
we evidently have $\lim_{M\to\infty}\left(\frac{M}{M+1}\right)^{cM+d}=1/e^c$; and the function $y(M)$ is monotonically increasing because, for all $d>c>0$, we have
\begin{align*}
y'(M)&=
{{\left( \frac{M}{M + 1} \right)}^{{c M} + d} \left(-c \log \left( \frac{M+1}{M} \right) + \frac{c M + d}{M} -\frac{cM+d}{M + 1}\right)}\\
&>
{\left( \frac{M}{M + 1} \right)}^{{c M} + d}\left(-\frac{c}{M}+\frac{cM+d}{M}-\frac{cM+d}{M+1}\right) =
{\left( \frac{M}{M + 1} \right)}^{c M}\frac{-c+d}{M(M+1)}\ge 0,
\end{align*}
proving the result.
\end{proof}
\end{lemma}
Due to a technical nature of our arguments, the proof of \ourref{Theorem}{thmone}~(b) will be divided into two cases.
The first case: $M \geq 11$, will be handled in full generality, while the second case: $4 \leq M \leq 10$, will have to be
treated separately for each value of $M$.
\begin{lemma}\label{lem11-}
If $11\le M$, then $\zeta^{(k)}(\sigma)\ne 0$, for
\[
q_M k+(M+1)\log 3 \le \sigma \le q_{M-1} k-M \log 3.
\]
\begin{proof}
In a way similar to the approach we took in part (a) of the theorem, here we write
\[
\left|\zeta^{(k)}(s)\right|\ge Q^{k}_M(\sigma) - H^{k}_M(\sigma)- T^{k}_M(\sigma)\ge
Q^{k}_M(\sigma)
\left(1-\frac{H^{k}_M}{Q^{k}_M}(\sigma)-
\frac{Q^{k}_{M+1}}{Q^{k}_M}(\sigma)\left(1+R^{k}_M(\sigma)\right)\right).
\]
By \ourref{Lemma}{3}, for $q_M k+(M+1)\log 3\le \sigma$,
$k\ge k_M=\frac{(2M+1)\log 3}{q_{M-1}-q_M}$, and $M\ge 11$, we have
\[
R_M^{k}(\sigma)\le R_M^{k}(q_M k+(M+1)\log 3)\le R_M^{k_M}(q_M k_M +(M+1)\log 3)< \frac{1}{2}.
\]
On the other hand, from \ourref{Lemma}{lemcMd}, we obtain the nice upper bound $\frac{Q^{k}_{M+1}}{Q^{k}_M}(\sigma)<\frac{1}{3}$ for $q_M k+(M+1)\log 3\le \sigma$ (this is precisely where the choice of the constant $\log 3$ is necessary); while, from \ourref{Lemma}{lemcMd}, we also get:
\[
\frac{H^{k}_M}{Q^{k}_M}(\sigma)\ge \sum_{n=1}^{\infty}\frac{1}{3^n}=\frac{1}{1-\frac{1}{3}}-1=\frac{1}{2}.
\]
Thus, for $11\le M$ and $q_M k+(M+1)\log 3\le \sigma \le q_{M-1} k+M\log 3$, we have
\[
1- \frac{H^{k}_M}{Q^{k}_M}(\sigma)-\frac{Q^{k}_{M+1}}{Q^{k}_M}(\sigma)
\left(1+R^{k}_M(\sigma)\right)
>
1-\frac{1}{2}-\frac{1}{3}\left(1+\frac{1}{2}\right)=0,
\]
which proves the lemma.
\end{proof}
\end{lemma}
\begin{lemma}\label{lem4-10}
If $4\le M\le 10$, and $q_M k+(M+1)\log 3\le \sigma \le q_{M-1} k+M\log 3$, then $\zeta^{(k)}(\sigma)\ne 0$.
\begin{proof}
Unfortunately, for $4\le M\le 10$, we have to consider each $M$ separately. We proceed as above, but since here the estimate
$\frac{Q^{k}_{M+1}}{Q^{k}_M}(\sigma)<\frac{1}{3}$
is not quite sharp enough to give us sufficiently good results, we will need to list upper bounds for $R_{M+1}^{k}$ and $T_M^k$ for individual $M$. We have
(keeping, as above, $\sigma_M:=q_M k_M+(M+1)\log 3$):
\begin{center}
\begin{tabular}{l||c|c|c|c|c|c|c}
$M$ & 4 & 5 & 6 & 7 & 8 & 9 & 10\\\hline
$R_{M+1}^{k}(\sigma_M)$ & $<0.6$ & $<0.57$ & $<0.55$ & $<0.53$ & $<0.52$ & $<0.51$ & $<0.51$\\[2pt]
$\frac{Q^{k}_{M+1}}{Q^{k}_M}(\sigma_M)$ & $<0.3$ &$<0.31$ &$<0.31$ &$<0.31$ & $<0.32$ & $<0.32$ & $<0.32$\\[2pt]
$\frac{T_M^k}{Q^{k}_M}(\sigma_M)$ & $<0.47$ &$<0.47$ &$<0.48$ &$<0.48$ & $<0.48$ & $<0.48$ & $<0.48$
\end{tabular}
\end{center}
As it is now easy to verify, for each $M$ that satisfies $4\le M\le 10$, and for all $\sigma$ is the range $q_M k+(M+1)\log 3\le \sigma \le q_{M-1} k+M\log 3$, we again conclude:
\[
1-\frac{H^{k}_M}{Q^{k}_M}(\sigma) - \frac{T_M^k}{Q^{k}_M}(\sigma) \ge
1-\frac{1}{2}- \frac{T_M^k}{Q^{k}_M}(\sigma_M)
>0,
\]
which proves the lemma.
\end{proof}
\end{lemma}
Combining \ourref{Lemma}{lem11-} and \ourref{Lemma}{lem4-10} gives us a proof of \ourref{Theorem}{thmone}~(b).
\begin{remark}
The zero-free regions we have given are not the largest possible.
For example, if one considered the lines $\sigma=\frac{1}{2}(q_M+q_{M+1})k+\frac{1}{2}\log 3$
through the centers of the wedges and searched for the lowest $k$ for which there were no zeros on
those lines, then one would obtain the following values for $k_M$ (which are lower than the values we have for the tips
of the wedge-shaped regions):
\begin{center}
\begin{tabular}{l||c|c|c|c|c|c|c|c|c}
$M$ & 3 & 4&5&6&7&8&9&10\\\hline
$k_M$ on line &14&41&87&154&247&368&519&703\\
$k_M$ at the tip &20&71&151&269&429&638&898&1214
\end{tabular}
\end{center}
\end{remark}
\section{Proof of \ourref{Theorem}{thmboxzero}}
In this last section we describe a counting technique that will allow us to obtain very precise (for many heights even exact) estimates for the number of zeros of $\zeta^{(k)}(s)$ in all critical strips $S_M^k$. It differs from the classical quantitative methods (notably Berndt's, in \cite{b:2}, for $N_k(T)$) in its use of
Rouch{\' e}'s theorem. Because of the property of the quasi-periodicity of the zeros of $\zeta^{(k)}(s)$ inside $S_M^k$ we are able to count the zeros by individual separation. In order for our approach to work, we first find horizontal, periodically-spaced zero-free line segments within the critical strips (in \ourref{Lemma}{lemline}). Then we show that there is always exactly
one zero of $\zeta^{(k)}(s)$ in the rectangles $R_j$ (for $j \in \mathbb{N}$) that are delimited by the vertical edges of two of the neighbouring zero-free regions and two of those horizontal zero-free lines (see \ourref{Figure}{figbox}).
As already mentioned above, in the critical strips $S_M^k$, which are located between the two zero-free regions, where the expansion of $\zeta^{(k)}(s)$ is
dominated by the terms $Q_M^k(s)$ and $Q_{M+1}^k(s)$ respectively, one can obtain values of the imaginary parts $t$ of expected zeros by solving the equation
$Q_M^k(\sigma+it)= Q_{M+1}^k(\sigma+it)$, an act of balancing the real and imaginary parts of two largest terms, and then choosing the horizontal lines of separation exactly halfway between them,
thus managing to avoid even the most irregular of zeros inside $S_M^k$. That is exactly what we do below.
A nice consequence of this method is that all the zeros of $\zeta^{(k)}(s)$ inside $S_M^k$ are simple.
\begin{lemma}\label{lemline}
Let $M \geq 2$ and $k \in \mathbb{N}$. If $s \in S_M^k$, then $\zeta^{(k)}(s)\ne 0$ for
\[
s=\sigma+i\cdot\frac{2\pi j}{\log(M+1)-\log M}.
\]
\end{lemma}
\begin{figure}[ht]
\caption{\label{figbox} The curve $\gamma$ is the boundary of the rectangle $R_j$.
The point $\color{red}\bullet$ represents a zero of $Q_M^k(s)+Q_{M+1}^k(s)$ on the critical line $\sigma=q_M k$.\vspace{1em}}
\includegraphics[width=16cm]{zeta-curve-rouche.png}
\end{figure}
\begin{proof}
For the sake of simplicity we consider the slightly wider rectangles $R_j$, with the vertical boundaries: $q_M k-(M+1)\log 3\le \sigma \le q_M k+(M+1)\log 3$, composition of which contains $S_M^k$.
In the center of the critical strip $S_M^k$, that is on the critical line $\sigma=q_M k$ we have $|Q_M^k(s)|=|Q_{M+1}^k(s)|$.
We consider the line segments in $S_M^k$ with
\[
t=\frac{2\pi j}{\log(M+1)-\log M}, \mbox{ where }j\in\Z,
\]
so that $Q_M^k(q_M k+it)=Q_{M+1}^k(q_M k+it)$.
We set $s=\sigma+it$, with $t$ as above, and consider the real and imaginary parts of the expression
\[
\zeta^{(k)}(s)=\sum_{n=2}^\infty \left(\cos(t\log n)-i\cdot\sin(t\log n)\right)Q_n^k(\sigma).
\]
With $|\Im(Q_{n}^k(s)|\le Q_n^k(\sigma)$ and $|\Re(Q_{n}^k(s)|\le Q_n^k(\sigma)$ we obtain
\begin{align*}
|\Re(\zeta^{(k)}(s))|&\ge |\cos(t\log M)Q_M^k(\sigma)+\cos(t\log(M+1))Q_{M+1}^k(\sigma)|-H_M^k(\sigma)-T_{M+1}^k(\sigma),\\
|\Im(\zeta^{(k)}(s))|&\ge |\sin(t\log M)Q_M^k(\sigma)+\sin(t\log(M+1))Q_{M+1}^k(\sigma)|-H_M^k(\sigma)-T_{M+1}^k(\sigma).
\end{align*}
Now, if $t = 0$, the situation is trivial, while if $t\ne 0$, then we either have $|\sin(\tau)|\ge\sin(\pi/2)=1/\sqrt{2}$ or $|\cos(\tau)|\ge\cos(\pi/2)=1/\sqrt{2}$.
Therefore, because obviously $|\zeta^{(k)}(s)|\ge |\Re(\zeta^{(k)}(s))|$ and $|\zeta^{(k)}(s)|\ge |\Im(\zeta^{(k)}(s))|$ and by our choice of $t$, we obtain:
\begin{align*}
|\zeta^{(k)}(s)|
&\ge \frac{1}{\sqrt{2}}\left(Q_M^k(\sigma)+Q_{M+1}^k(\sigma)\right)-H_M^k(\sigma)-T_{M+1}^k(\sigma)\\
&= Q_M^k(\sigma)\left(\frac{1}{\sqrt{2}}+\frac{1}{\sqrt{2}}\frac{Q_{M+1}^k}{Q_M^k}(\sigma)
-\frac{H_M^k}{Q_M^k}(\sigma)-\frac{Q_{M+2}^k}{Q_M^k}(\sigma)-\frac{T_{M+2}^k}{Q_M^k}(\sigma)\right)\\
&= Q_M^k(\sigma)\left(\frac{1}{\sqrt{2}}-\frac{H_M^k}{Q_M^k}(\sigma)
+\frac{Q_{M+1}^k}{Q_M^k}(\sigma)
\left(\frac{1}{\sqrt{2}}-\frac{Q_{M+2}^k}{Q_{M+1}^k}(\sigma)-\frac{T_{M+2}^k}{Q_{M+1}^k}(\sigma)\right)\right).
\end{align*}
Recall that $\frac{Q_{M+2}^k}{Q_{M+1}^k}(\sigma)$ and $\frac{T_{M+2}^k}{Q_{M+1}^k}(\sigma)$ are decreasing in $\sigma$.
From the proofs of \ourref{Lemma}{lem11-} and \ourref{Lemma}{lem4-10} we know that
\begin{align*}
\frac{1}{\sqrt{2}}&-\frac{Q_{M+2}^k}{Q_{M+1}^k}(\sigma)-\frac{T_{M+2}^k}{Q_{M+1}^k}(\sigma)
\ge \frac{1}{\sqrt{2}} -\frac{Q_{M+2}^k}{Q_{M+1}^k}(\sigma)\left(1+R_{M+2}(\sigma)\right)\\
&\ge \frac{1}{\sqrt{2}} -\frac{Q_{M+2}^k}{Q_{M+1}^k}(q_{M+1}k+(M+2)\log3)
\left(1+R_{M+2}(q_{M+1}k+(M+2)\log3)\right)>0.
\end{align*}
Similarly, since $\frac{H_M^k}{Q_M^k}(\sigma)$ is decreasing in $\sigma$, we obtain
\[
\frac{1}{\sqrt{2}}-\frac{H_M^k}{Q_M^k}(\sigma)\ge\frac{1}{\sqrt{2}}-\frac{H_M^k}{Q_M^k}(q_{M-1}k-M\log3)\ge\frac{1}{\sqrt{2}}-\frac{1}{2}>0,
\]
which concludes the proof of the lemma.
\end{proof}
\begin{proof}[Proof of \ourref{Theorem}{thmboxzero}]
Let $j\in\N$. Now that the non-vanishing of $\zeta^{(k)}(s)$ along the horizontal division lines has been established (and those vertical lines, by definition, cannot contain any zeros), we can integrate along the entire boundary $\gamma$ of each of the rectangles $R_j$ described above (see \ourref{Figure}{figbox}). Moreover, \ourref{Theorem}{thmone}
and \ourref{Lemma}{lemline} assert that we have
\begin{align*}
\left|\left(Q_M^k(s)+Q_{M+1}^k(s)\right)-\zeta^{(k)}(s)\right|
&=\left|H_M^k(s)+T_{M+1}^k(s)\right|\\
&\le H_M^k(\sigma)+T_{M+1}^k(\sigma)\le\left|Q_M^k(s)+Q_{M+1}^k(s)\right|,
\end{align*}
along the vertical and horizontal parts of the boundary curve $\gamma$, respectively. Thus, along the entire boundary of $R_j$, the function $\zeta^{(k)}(s)$ is closely approximated by the
function $Q_M^k(s)+Q_{M+1}^k(s)$; and therefore, by Rouch\'e's Theorem, it has to have the same number
of zeros inside $R_j$, for every $j \in \mathbb{N}$. However, it is now very easy to check that the approximation function $Q_M^k(s)+Q_{M+1}^k(s)$ has always
exactly one zero in $R_j$, namely
\[
s=q_M k+ i\cdot\frac{(2j+1)\pi}{\log(M+1)-\log M}.
\]
This proves both the simplicity of all zeros of $\zeta^{(k)}(s)$ inside $S_M^k$, and the sharp formula for $N_M^k(T)$, as given in \ourref{Corollary}{corcount}.
\end{proof}
\begin{figure}
\includegraphics[width=8cm]{zeta-zerofree-regions-k100.png}
\includegraphics[width=8cm]{zeta-zerofree-regions-k200.png}
\includegraphics[width=8cm]{zeta-zerofree-regions-k400.png}
\includegraphics[width=8cm]{zeta-zerofree-regions-k800.png}
\caption{Zero-free regions and zero-free line segments for $\zeta^{(100)}$, $\zeta^{(200)}$, $\zeta^{(400)}$, and $\zeta^{(800)}$.}
\end{figure}
\section{Acknowledgments}
This research was supported in part by a New Faculty Grant from UNC Greensboro.
Most of the work on the paper was conducted while Thomas Binder was a visiting researcher at UNC Greensboro, in Fall 2008.
The authors would like to thank Prof.\ Garry J.\ Tee from the University of Auckland for several helpful comments and many useful remarks.
All computations and plots were done with the computer algebra system Sage \cite{sage}.
|
\section{Introduction}\label{introduction}
As discussed in \cite{assem}, the study of cluster algebras naturally leads to the special case $k=2$ of the notion of $\SL{k}$-tiling introduced in this paper. Our $\SL{k}$-tilings are simply $\mathbb{Z}\times \mathbb{Z}$ arrays of numbers (or elements of a commutative ring) having all adjacent $k\times k$ minors equal to one. Not only are they a natural extension of notions already considered, but one can recast in their guise such notions as \defn{$T$-systems} of Theoretical Physics (see \cite{di_Francesco}), or \defn{frieze patterns} of Coxeter (see \cite{coxeter}).
An instance of a positive integer $\SL{2}$-tiling is given in Figure~\ref{fig1}.
\begin{figure}[ht]
$$\begin{array}{ccccccccccccccc} \setlength{\unitlength}{3mm}
\ddots & \vdots & \vdots& \vdots& \vdots& \vdots& &\vdots & \vdots& \vdots& \vdots& \vdots& \vdots& \vdots &\revddots \\[3pt]
\cdots &887&567&247&174&101&\dr{28}&11&5&4&3&2&1&1&\cdots\\
\cdots &158&101&44&31&18& \r{5}&2&1&1&1&1&1&2&\cdots \\
\cdots &61&39&17&12&7& \r{2}&1&1&2&3&4&5&11&\cdots \\
\cdots &25&16&7&5&3& \r{1}&1&2&5&8&11&14&31&\cdots \\
&\dr{14}&\r{9}&\r{4}&\r{3}&\r{2}&\r{1}&\r{2}&\r{5}&\dr{13}&\dr{21}&\dr{29}&\dr{37}&\dr{82}& \\
\cdots &3&2&1&1&1&\r{1}&3&8&21&34&47&60&133&\cdots \\
\cdots &1&1&1&2&3&\r{4}&13&35&92&149&206&263&583&\cdots \\
\cdots &1&2&3&7&11&\dr{15}&49&132&347&562&777&992&2199&\cdots \\
\revddots & \vdots & \vdots& \vdots& \vdots& \vdots& &\vdots & \vdots& \vdots& \vdots& \vdots& \vdots& \vdots & \ddots
\end {array} $$
\begin{picture}(0,0)(0,0)\setlength{\unitlength}{3.5mm}
\put(-21,7.49){\textcolor{red}{\linethickness{1pt}\line(1,0){3}}}
\put(18,6.85){\textcolor{red}{\Huge$\longrightarrow$}}
\put(20,6.5){\textcolor{red}{$+$}}
\put(-2.6,14.5){\textcolor{red}{\linethickness{1pt}\line(0,1){2}}}
\put(-3.2,0.4){\textcolor{red}{\Huge$\downarrow$}}
\put(-2,0){\textcolor{red}{$+$}}
\end{picture}
\caption{A $\SL{2}$-tiling with values in ${\mathbb N}^*$.}\label{fig1}
\end{figure}
Clearly any $\SL{k}$-tiling $\mathcal{A}$ is of rank at least $k$ (when considered as a bi-infinite matrix). As we will see, the $\SL{k}$-tilings that are of minimal rank are of particular interest, not only by themselves, but as well as for the cases when they correspond to frieze patterns or $T$-systems. We call \defn{tame} such minimal rank $\SL{k}$-tilings, and we give several general results regarding them. Among these interesting results, we show that to any tame $\SL{k}$-tiling there corresponds another interesting tame $\SL{k}$-tiling, that we call its \defn{dual}. The entries of these dual tilings are obtained by computing adjacent $(k-1)\times (k-1)$-minors. It is striking that this duality is actually an involution. We also re-derive, in a new an elegant manner, all the results of Conway-Coxeter concerning frieze-patterns. Indeed, our approach allows new tools to bear on this subject, especially because we can now make use of linear algebra and particular presentations for $\SL{2}(\mathbb{Z})$.
Our approach also opens the door for the study of generalized frieze patterns, including those that have already been considered in \cite{cordes}. This corresponds to the study of $\SL{k}$-tilings that afford two periods (in two linearly independent directions). We call \defn{toric} such $\SL{k}$-tilings, since they are evidently characterized by their value on a torus. Once again the tame situation is of particular interest. More on this will be the subject of a planed sequel for this paper.
\section{Definitions}\label{sec_defn}
We consider arrays ${\mathcal A}=(a_{ij})_{i,j\in\mathbb{Z}}$, with values $a_{ij}$ lying
in a field $\mathbb{K}$. For two equal cardinality finite subsets $I$ and $J$ of $\mathbb{Z}$, we denote by ${\mathcal A}_{IJ}$ the submatrix of ${\mathcal A}$ obtained by selecting the rows indexed by the elements of $I$ and columns indexed by the elements of $J$. The corresponding minor is denoted\footnote{Without explicit reference to the underlying bi-infinite matrix $\mathcal{A}$.} by ${M}_{IJ}$, that is: ${M}_{IJ}:=\det {\mathcal A}_{IJ}$. Since we often need to write down {\sl adjacent} $k\times k$ minors, we introduce the short hand notations:
\begin{equation} {{\mathcal A}_{ij}^{(k)}:={\mathcal A}_{\{i,\ldots, i+k-1\},\{j,\ldots, j+k-1\}}}, \qquad {\rm and}\qquad {{M}_{ij}^{(k)}:=\det {\mathcal A}_{ij}^{(k)}}.\end{equation}
We say that ${\mathcal A}$ is a \defn{$\SL{k}$-tiling} of the plane if all its adjacent $k\times k$ minors of ${\mathcal A}$ are equal to one. This is to say that it satisfies the \defn{$\SL{k}$-property}:
\begin{equation}\label{num_condition}
{ {M}_{ij}^{(k)}=1,\qquad {\rm for\ all\ } i\ {\rm and\ } j\ {\rm in}\ \mathbb{Z}},
\end{equation}
We sometimes consider \defn{partial $\SL{k}$-tilings}, only defined on some subset $S$ (called \defn{shape}) of $\mathbb{Z}\times \mathbb{Z}$, with condition~(\ref{num_condition}) applying only if all the entries considered belong to the underlying subset. As usual, a \defn{rectangle} in $\mathbb{Z}\times \mathbb{Z}$ is a (possibly infinite) shape $S$ such that $(u,v+s)$ and $(u+r,v)$ lie in $S$, whenever $(u,v)$ and $(u+r,v+s)$ both lie in $S$.
A partial $\SL{k}$-tiling is said to be a \defn{$\SL{k}$-array} if its shape is a rectangle. In particular, ${\mathbb N}\times {\mathbb N}$ and $\mathbb{Z}\times \mathbb{Z}$ shaped $\SL{k}$-tilings are $\SL{k}$-arrays. Clearly linear combinations of rows (or columns) make sense for $\SL{k}$-arrays, so that we may consider the notion of rank of a such $\SL{k}$-tilings. In particular any $\SL{k}$-array is at least of rank $k$, since any $k$ consecutive rows have to be linearly independent in view of the $\SL{k}$-property. We say that a (partial) $\SL{k}$-tiling is \defn{tame} if it has rank $k$. Otherwise we call it \defn{wild}.
A word of warning is in order concerning our convention for the underlying coordinate system. Indeed, as in the example of Figure~\ref{fig1}, we use the usual matrix convention for coordinates, so that the $x$-axis points downwards, and the $y$-axis points to the right.
\subsection*{A family of examples}
The \defn{positive integer frieze patterns} of Coxeter (see \cite{coxeter,conway,propp}) give rise to an interesting family of nonzero {\it partial} $\SL{2}$-tilings. Up to a $45^\circ$ degree tilting, the original description of Coxeter may be formulated as follows. One considers partial $\SL{2}$-tilings\footnote{The $\SL{2}$ condition applies only when it makes sense.} such as the one illustrated in Figure~\ref{fig1_5}, assuming that all $a_{ij}$ are positive integers.
\begin{figure}[ht]
$$\begin{array}{l}
\begin{array}{ccccccccccccccccc}
&\, \ddots &\, \textcolor{red}{\ddots}&\, \textcolor{red}{\ddots} &\, \textcolor{red}{\cdots} &\, \textcolor{red}{\ddots} &\, \ddots&
\end{array}\\[5pt]
\begin{array}{ccccccccccccccccc}
&\hbox{\hskip18pt}& & 1& \textcolor{red}{a_{10}} & \textcolor{red}{a_{20}} &\textcolor{red}{\cdots} & \textcolor{red}{a_{n-2,0}} & 1 & \\
& && & 1& \textcolor{red}{a_{11}} & \textcolor{red}{a_{21}} &\textcolor{red}{\cdots} & \textcolor{red}{a_{n-2,1}} & 1 & \\
&&&& &\, \ddots &\, \textcolor{red}{\ddots} &\, \textcolor{red}{\ddots} &\, \textcolor{red}{\cdots} &\, \textcolor{red}{\ddots}&\, \ddots &\\
&&&&& & 1& \textcolor{red}{a_{1j}} & \textcolor{red}{a_{2j}} &\textcolor{red}{\cdots} & \textcolor{red}{a_{n-2,j}} & 1 & \\
\end{array}\\[-2pt]
\begin{array}{ccccccccccccccccc}
&\hbox{\hskip55pt}&&&&&& & \ddots & \textcolor{red}{\ddots}&\ \textcolor{red}{\ddots} & \textcolor{red}{\cdots} &\ \textcolor{red}{\ddots} & \ddots&
\end{array}\\
\end{array}$$
\caption{Conway-Coxeter frieze patterns.}\label{fig1_5}
\end{figure}
Note that the number of ``diagonals`` is $n$.
As shown in \cite{coxeter}, one of the striking property of frieze patterns is that they are necessarily periodic along the direction $y=x$. This is to say that there exists some $p$ in $\mathbb{Z}$ such that $a_{i+p,j+p}=a_{ij}$ for all $i$ and $j$, with $p=n+1$ ($n$ being the number of diagonals as above).
We may turn frieze patterns into full $\SL{2}$-tilings, by the simple device of extending them (skew) periodically both along the $x$ and $y$ directions, i.e.: setting
$${a_{i+p,j}=-a_{ij}},\quad {\rm and}\quad {a_{i,j+p}=-a_{ij}}.$$
One needs only check that this is consistent with the $\SL{2}$-condition at the ``boundary''.
Such a tiling is illustrated in Figure~\ref{fig2} in the case of a generic\footnote{All positive frieze patterns of width $2$ may be obtained from it by specialization.} frieze pattern having $4$ diagonals. In this $\SL{2}$-tiling, $a$ and $b$ may assume any value as long as we have
$${c=\frac{1+b}{a}},\qquad {d=\frac{1+a+b}{a\,b}},\qquad {e=\frac{1+a}{b}}.$$
Observe the further symmetry corresponding to a transposition followed by a diagonal translation.
\begin{figure}[ht]
$$\begin {array}{rrrrrrrrrrrrrrrr}
\textcolor{red}{a}&\textcolor{red}{b}&1&0&-1&-a&-b&-1&0&1&\textcolor{red}{a}&\textcolor{red}{b}\\
1&\textcolor{red}{c}&\textcolor{red}{d}&1&0&-1&-c&-d&-1&0&1&\textcolor{red}{c}\\
0&1&\textcolor{red}{e}&a&1&0&-1&-e&-a&-1&0&1\\
-1&0&1&b&c&1&0&-1&-b&-c&-1&0\\
-e&-1&0&1&d&e&1&0&-1&-d&-e&-1\\
-a&-b&-1&0&1&\textcolor{red}{a}&\textcolor{red}{b}&1&0&-1&-a&-b\\
-1&-c&-d&-1&0&1&\textcolor{red}{c}&\textcolor{red}{d}&1&0&-1&-c\\
0&-1&-e&-a&-1&0&1&\textcolor{red}{e}&a&1&0&-1\\
1&0&-1&-b&-c&-1&0&1&b&c&1&0\\
e&1&0&-1&-d&-e&-1&0&1&d&e&1\\
\textcolor{red}{a}&\textcolor{red}{b}&1&0&-1&-a&-b&-1&0&1&\textcolor{red}{a}&\textcolor{red}{b}\\
1&\textcolor{red}{c}&d&1&0&-1&-c&-d&-1&0&1&\textcolor{red}{c}\end {array}$$
\caption{Skew-periodic extension of a Conway-Coxeter frieze pattern.}\label{fig2}
\end{figure}
The number of frieze patterns having $n$ diagonals, and for which all entries are positive integers, has been shown in \cite{conway} to be another incarnation of the ubiquitous Catalan numbers
$$C_{n-1}=\frac{1}{n}\binom{2\,n-2}{n-1}.$$
A nice exposition of classical results regarding frieze patterns, as well as many new results tying their study to the type-$A$ cluster algebras of Fomin and Zelevinsky, is given by Propp in \cite{propp}.
It may readily be checked that the example of Figure~\ref{fig2} is a rank $2$ bi-infinite matrix. We will show in Section~\ref{sec_applications} how we may construct all frieze patterns using our theory of $\SL{k}$-tilings (for $k=2$), by extending them to complete $\SL{k}$-tilings. Moreover, using our theory, we give new proofs of all the results obtained by Coxeter and Conway.
Although our exploration of this point of view will mainly be for the case $k=2$, many of our results actually hold (with the necessary adaptations) in the general context of a suitable notion of $\SL{k}$-frieze patterns (see Section~\ref{sec_closing}).
\section{Tame $\SL{k}$-tilings}\label{sec_general}
Not all $\SL{k}$-tiling are tame, as seen in Example~(\ref{not_zero_free}) for $k=2$.
\begin{equation}\label{not_zero_free}
\begin {array}{cccccccccccr}
\ddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\revddots\\
\cdots&1&{x_{{11}}}&-1&{x_{{12}}}&1&{x_{{13}}}&-1&{x_{{14}}}&1&{x_{{15}}}&\cdots\\
\cdots&0&1&0&-1&0&1&0&-1&0&1&\cdots\\
\cdots&-1&{x_{{21}}}&1&{x_{{22}}}&-1&{x_{{23}}}&1&{x_{{24}}}&-1&{x_{{25}}}&\cdots\\
\cdots&0&-1&0&1&0&-1&0&1&0&-1&\cdots\\
\cdots&1&{x_{{31}}}&-1&{x_{{32}}}&1&{x_{{33}}}&-1&{x_{{34}}}&1&{x_{{35}}}&\cdots\\
\cdots&0&1&0&-1&0&1&0&-1&0&1&\cdots\\
\cdots&-1&{x_{{41}}}&1&{x_{{42}}}&-1&{x_{{43}}}&1&{x_{{44}}}&-1&{x_{{45}}}&\cdots\\
\cdots&0&-1&0&1&0&-1&0&1&0&-1&\cdots\\
\cdots&1&{x_{{51}}}&-1&{x_{{52}}}&1&{x_{{53}}}&-1&{x_{{54}}}&1&{x_{{55}}}&\cdots\\
\cdots&0&1&0&-1&0&1&0&-1&0&1&\cdots\\
\revddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\ddots\end {array}
\end{equation}
Here the $x_{ij}$ may be chosen at will (or as independent variables). In particular, this example shows that there are $\SL{2}$-tilings of any rank $\geq 2$.
In part, the interest of considering tame tilings comes from the fact that they are easily characterized by their value on relatively small subsets of $\mathbb{Z}\times\mathbb{Z}$. But we will also make evident that tame tilings have very nice properties.
We first illustrate tameness with the following special case.
\subsection*{$0$-free tilings} We say that a $\SL{k}$-tiling is \defn{$0$-free} if all its $(k-1)\times (k-1)$ adjacent subminors are nonzero. Note that in the case $k=2$: a $\SL{2}$-tiling is 0-free if its values are nonzero; in particular if they are positive integers, as in Figure~\ref{fig1} or the $\SL{2}$-tilings constructed in \cite{assem}.
The proof of the following proposition uses Dodgson\footnote{a.k.a. Lewis Carrol} ``Condensation Law of Determinants'' \cite{alice}, that can be stated in the format:
\begin{equation}\label{carrol}
{{M}_{ij}^{(r+1)} {M}_{i+1,j+1}^{(r-1)} = \det
\left(\begin{array}{lll} {M}_{ij}^{(r)} & {M}_{i,j+1}^{(r)}\\[6pt]
{M}_{i+1,j}^{(r)} & {M}_{i+1,j+1}^{(r)}
\end{array} \right)}
\end{equation}
for all $r$.
In fact this is a direct consequence of a result of Desnanot and Jacobi (see \cite[Th. 3.12, page 111]{bressoud}).
For instance, with $r=2$, we get the identity
$$\begin{array}{l}
\det
\left(\begin{array} {lll}a_{ij} & a_{i,j+1} & a_{i,j+2}\\
a_{i+1,j} & a_{i+1,j+1} & a_{i+1,j+2}\\
a_{i+2,j} & a_{i+2,j+1} & a_{i+2,j+2}\\
\end{array}\right) \det
\begin{pmatrix} a_{i+1,j+1}
\end{pmatrix} =\\[25pt]
\hskip150pt \det \begin{pmatrix} \left| \begin{array}{llll} a_{ij} & a_{i,j+1}\\
a_{i+1,j}&a_{i+1,j+1}
\end{array}\right| &
\left| \begin{array}{llll} a_{i,j+1} & a_{i,j+2}\\
a_{i+1,j+1} & a_{i+1,j+2}
\end{array}\right| \\[14pt]
\left| \begin{array}{llll} a_{i+1,j} & a_{i+1,j+1}\\
a_{i+2,j}& a_{i+2,j+1}
\end{array}\right| &
\left| \begin{array}{llll} a_{i+1,j+1}& a_{i+1,j+2}\\
a_{i+2,j+1}& a_{i+2,j+2}
\end{array}\right|
\end{pmatrix}
\end{array}$$
\begin{proposition}\label{rankk}
{Any $0$-free $\SL{k}$-array is tame}.
\end{proposition}
\begin{proof}[\bf Proof.]
Consider any adjacent $(k-1)\times (k-1)$ subarray of a $\SL{k}$-array, we observe that the determinant of the corresponding submatrix does not vanish, since this is precisely the $0$-free condition. On the other hand, for $r=k$, the right-hand side of (\ref{carrol}) is zero in all instances, since the four $k\times k$ minors considered are all equal to $1$. We thus conclude that any adjacent $(k+1)\times (k+1)$ subarray of a $0$-free $\SL{k}$-array must necessarily have vanishing determinant. The proof then follows from Lemma~\ref{det_rank} (see Section~\ref{preuves}).
\end{proof}
We may construct all tame $\SL{k}$-tilings as follows. Given a $\SL{k}$-tiling ${\mathcal A}$, let us denote by $R_i$ and $C_j$ its rows and columns. Then each $C_j$ is a linear combination of the $k$ preceding columns, that is, of $C_{j-1},\ldots, C_{j-k}$. The linear combination may be written as
\begin{equation}\label{colonnes}
{ (-1)^kC_{0}-(-1)^{k}a_1C_{1}+\cdots -a_{k-1}C_{k-1}+C_k=0}.
\end{equation}
Indeed this follows from the $\SL{k}$-property and from the next lemma, which is an exercise in linear algebra (expansion of the determinant with respect to the rows), left to the reader.
\begin{lemma}
{Let $A$ be a rank $k$ matrix with $k+1$ columns $C_0,\ldots,C_k$ and rows indexed by $\mathbb{Z}$. Then}
\begin{equation}\begin{array}{ll}
{(-1)^k M_{01}^{(k)}C_{0}-(-1)^{k}\det A_{\{0,\ldots,k-1\},\{0,2,\ldots,k-1\}} C_{1}+\cdots}\\[6pt]
\qquad \qquad\qquad\qquad\qquad
{-\det A_{\{0,\ldots,k-1\},\{0,\ldots,k-2,k\}} C_{k-1}+ M_{00}^{(k)}C_k=0}.
\end{array}
\end{equation}
\end{lemma}
Hence, to each $j$ in $\mathbb{Z}$ we associate a row vector $\lambda_j=(a_1,...,a_{k-1})$ of dimension $m=k-1$ over the field $\mathbb{K}$, and we simply denote by $\lambda$ the resulting element of $ (\mathbb{K}^{1\times m})^{\mathbb{Z}}$. Similarly, we associate to each $i$ a column vector $\gamma_i$ of dimension $m$ over $\mathbb{K}$, which expresses the linear dependence of $R_i$ on the $m$ preceding rows, and denote by $\gamma$ the resulting element of $(\gamma_i)\in (\mathbb{K}^{m\times 1})^{\mathbb{Z}}$. These row and columns vectors are called the \defn{linearization coefficients} of the $\SL{k}$-tiling ${\mathcal A}$.
We call \defn{linearization data} the triple
\begin{equation}\label{data_lin}
{({\mathcal A}_{00}^{(k)}, \lambda, \gamma)} \quad {\rm in}\quad { \SL{k}(\mathbb{K})\times (\mathbb{K}^{1\times m})^{\mathbb{Z}} \times (\mathbb{K}^{m\times 1})^{\mathbb{Z}}},\quad \rm {where}\ {m=k-1}.
\end{equation}
\begin{proposition}\label{coefflinear}
{The mapping}
$${{\mathcal A} \mapsto ({\mathcal A}_{00}^{(k)}, \lambda, \gamma)},$$
{which associates to a tame $\SL{k}$-tiling its linearization data, is a bijection between the set of tame $\SL{k}$-tilings and the set $\SL{k}(\mathbb{K})\times (\mathbb{K}^{1\times m})^{\mathbb{Z}} \times (\mathbb{K}^{m\times 1})^{\mathbb{Z}}$, with $m=k-1$}.
\end{proposition}
\begin{proof}[\bf Proof.]
The fact that this mapping is well-defined and injective follows from the remarks preceding the proof. For surjectivity, let the data in $\SL{k}(\mathbb{K})\times (\mathbb{K}^{1\times m})^{\mathbb{Z}} \times (\mathbb{K}^{m\times 1})^{\mathbb{Z}}$ be given. Then clearly there exists a $\mathbb{Z} \times \mathbb{Z}$ array ${\mathcal A}$ of rank $k$ which maps onto this data.
We have only to verify that the $\SL{k}$-property holds. This is a consequence of the following easy fact (and its dual): let $M$ be a $k\times (k+1)$-matrix with columns $C_0,\ldots,C_k$ such that (\ref{colonnes}) holds.
Then the matrix of its first $k$ columns is of determinant 1 if and only if the matrix of its $k$ last columns is of determinant 1.
Indeed,
\begin{eqnarray*}
\det(C_1,\ldots,C_{k})&=&\det(C_1,\ldots ,C_{k-1},-(-1)^kC_{0}+(-1)^{k}a_1C_{1}-\ldots +a_{k-1}C_{k-1})\\
&=&\det(C_1,\ldots ,C_{k-1},-(-1)^kC_{0})\\
&=&\det(C_{0},\ldots,C_{k-1}).
\end{eqnarray*}
\end{proof}
\begin{proposition}\label{rank1}
{Let ${\mathcal A}$ be an $\SL{k}$-tiling. Then ${\mathcal A}$ is tame if and only if the infinite matrix $\mathcal{M}:=({M}_{IJ})_{I,J}$,
with $I$ and $J$ varying in $k$-subsets of $\mathbb{Z}$, is of rank $1$. In particular, if $I_0, J_0$ are intervals in $\mathbb{Z}$ and $I, J$ are any $k$-subsets, then we have}
\begin{equation}\label{mineursmultiplicatifs}
{{M}_{IJ}={M}_{IJ_0} \, {M}_{I_0J}}.
\end{equation}
\end{proposition}
\begin{proof}[\bf Proof.]
If ${\mathcal A}$ is tame, then the fact that $\mathcal{M}$ has rank $1$ is a consequence of the study of ${\mathbb N}\times{\mathbb N}$ tilings in Section~\ref{NbyN}. The converse follows from Lemma~\ref{rang1}.
Now take $I$, $J$, $I_0$, and $J_0$ as in the statement. Then
\begin{equation}\label{etape}
{
\begin{pmatrix}
{M}_{I_0J_0} & {M}_{I_0J} \\
{M}_{IJ_0} & {M}_{IJ}
\end{pmatrix}
}
\end{equation}
is a submatrix of $\mathcal{M}$. But ${M}_{I_0J_0} =1$ (since ${\mathcal A}$ is a $\SL{k}$-tiling), and the determinant of (\ref{etape}) mush vanish, i.e.:
$${M}_{IJ}-{M}_{I_0J}{M}_{IJ_0} =0,$$
since $\mathcal{M}$ is of rank $1$.
Thus we get the desired equality.
\end{proof}
The easy direct proof of the next result is left to the reader. It will be useful in the sequel.
\begin{lemma}\label{suppression_colonne}
{Let $C_1$, $C_2$, and $C_3$ be three consecutive columns {\rm (}resp. rows{\rm )}, of a tame $\SL{2}$-tiling, that are such that $C_2=C_1+C_3$. Then a new tame $\SL{2}$-tiling may be constructed by suppressing the column {\rm (}resp. row{\rm )} $C_2$}.
\end{lemma}
\subsection*{Group actions on tilings}
There is a natural translation action of $\mathbb{Z}^2$ on $\SL{2}$-tilings.
Formally, the action of the vector $(p,q)$ replaces the tiling ${\mathcal A}=(a_{ij})$ by
$${(p,q)\cdot {\mathcal A}=(a_{i+p,j+q})}.$$
We denote by ${\mathcal A}_x$ the translate of ${\mathcal A}$ by $(1,0)$, and by ${\mathcal A}_y$ the translate by $(0,1)$.
We may describe these last translates intrinsically in terms of the data given by the bijection of Proposition~\ref{coefflinear}.
More specifically, let $(\mathcal{S}, \lambda, \gamma)$ be the linearization data corresponding to ${\mathcal A}$ via this bijection. Then the linearization data corresponding to ${\mathcal A}_x$ is $(\mathcal{S}_x, \lambda, \gamma')$, with
$$
\mathcal{S}_x = \begin{pmatrix} 0&1&0 &\ldots&0 \\ 0&0&1&\ldots&0 \\ \vdots&\ddots&\ddots &\ddots&\vdots\\ 0&\cdots&0&0&1\\(-1)^{k-1} & (-1)^k(\gamma_k)_1&\ldots&&(\gamma_k)_{k-1}
\end{pmatrix} \mathcal{S},
$$
and $\gamma_i'=\gamma_{i+1}$
Transposition of matrices (which amounts to exchanging rows and columns) also preserves $\SL{k}$-tilings, as is easily observed. In terms of linearization data, it amounts to transposing the initial matrix and exchanging $\lambda$ and $\gamma$. We can thus easily describe ${\mathcal A}_y$ using these remarks.
Observe that if $k\equiv 0$ or $1\mod 4$, then a vertical or an horizontal symmetry also preserves $\SL{k}$-tilings, since in this case, such a symmetry preserves the determinant of $k\times k$ matrices (for $k\equiv 2$ or $3 \mod 4$, the determinant is clearly multiplied by $-1$).
\section{Dual tilings}
To any array ${\mathcal A}$, we associate the \defn{$m$-derived array}:
\begin{equation}\label{rule}
{\partial_m {\mathcal A}:=\left({M}_{ij}^{(m)}\right)_{i,j}},
\end{equation}
consisting of the adjacent $m\times m$ minors of ${\mathcal A}$. For $\SL{k}$-arrays, we are specially interested in the case $m=k-1$, in which case the resulting array is called the \defn{dual} array of ${\mathcal A}$. We also write ${\mathcal A}^*$ for $\partial_{k-1} {\mathcal A}$.
Clearly $\partial_ 1$ is the identity operator, and it is natural to set $\partial_0 {\mathcal A}$ equal to the tiling whose value is $1$ in all positions.
As an illustration of the above definition, the dual of the $\SL{3}$-tiling
\begin{equation}\label{pavage_fibo3}
{\begin {array}{ccccccccccccccc}
\ddots&\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\revddots\\
\cdots & 8997&1782&353&70&14&3&1&1&\cdots \\
\cdots & 1782&353&70&14&3&1&1&2&\cdots\\
\cdots & 353&70&14&3&1&1&2&5&\cdots\\
\cdots & 70&14&3&1&1&2&5&14&\cdots\\
\cdots & 14&3&1&1&2&5&14&42&\cdots\\
\cdots & 3&1&1&2&5&14&42&131&\cdots\\
\cdots & 1&1&2&5&14&42&131&417&\cdots\\
\revddots&\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\ddots
\end {array}}
\end{equation}
is the tiling
\begin{equation}\label{une_derivee}
{ \begin {array}{ccccccccccccccccc}
\ddots&\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\revddots\\
\cdots & 417&131&42&14&5&2&1&\cdots\\
\cdots &131&42&14&5&2&1&1&\cdots\\
\cdots & 42&14&5&2&1&1&3&\cdots\\
\cdots &14&5&2&1&1&3&14&\cdots\\
\cdots &5&2&1&1&3&14&70&\cdots\\
\cdots &2&1&1&3&14&70&353&\cdots\\
\revddots&\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\ddots
\end {array} }
\end{equation}
We have the following property of derivation of arrays, that will be proved in Section~\ref{preuves}.
\begin{proposition}\label{derive_equation}
{The dual of a tame $\SL{k}$-tiling is a tame $\SL{k}$-tiling}.
{Moreover, for any natural integers $r,s$ such that $r+s=k$, we have}
\begin{equation}\label{condensation_devive}
{\partial_r {\mathcal A}^*=(r-1,r-1)\cdot (\partial_{s} {\mathcal A})}
\end{equation}
{In particular, $({\mathcal A}^*)^*$ and ${\mathcal A}$ coincide up to translation}.
\end{proposition}
Observe also that, with $r=k-1$, identity (\ref{carrol}) gives $({\mathcal A}^*)^*=\textcolor{red}{a_{11}}\det({\mathcal A})$ for any $3\times 3$ matrix
$${{\mathcal A}=\begin{pmatrix}
a_{00}& a_{01} & a_{02}\\
a_{10}& \textcolor{red}{a_{11}} & a_{12}\\
a_{20}& a_{21} & a_{22}
\end{pmatrix}}.$$
Hence, for any $\SL{3}$-tiling ${\mathcal A}$, we have
\begin{equation}\label{rec_trois}
{ ({\mathcal A}^*)^* = \left(a_{i+1,j+1}\, {M}_{ij}^{(3)}\right)_{i,j\in \mathbb{Z}}}.
\end{equation}
It follows that $({\mathcal A}^*)^*={\mathcal A}$ (up to the necessary translation) for any $\SL{3}$-tiling, wether they be tame or wild. However, for $k\geq 4$, it may be checked that the tameness condition is necessary for (\ref{condensation_devive}) to hold.
\section{Tilings associated to paths}\label{words_tilings}
We consider \defn{paths} $\pi$
as lists of \defn{points}, i.e.: elements of $\mathbb{Z}\times \mathbb{Z}$,
$$\pi=(i_0,j_0),(i_1,j_1),\ldots, (i_{N},j_{N}),$$
starting at ${\rm s}(\pi):=(i_0,j_0)$ and ending at ${\rm e}(\pi):=(i_{N},j_{N})$, and such that
$$(i_{s+1},j_{s+1})=\begin{cases} (i_s,j_s-(1,0) & \text{or}, \\[6pt]
(i_s,j_s)+(0,1),
\end{cases}$$
for points along the path. To understand why signs appear here, it may be good to recall our convention for the orientation of the $x$ and $y$ axis (see Section~\ref{sec_defn}).
If we fix the start and end points $(i_0,j_0)$ and $(i_{N},j_{N})$, it is well known that these paths number $\binom{m+n}{m}$ (with $(-m,n)=(i_{N},j_{N})-(i_0,j_0)$), and that they are in bijection with words
\begin{equation}\label{le_mot}
w=w_1w_2\cdots w_{n+m}
\end{equation}
on the alphabet ${\mathcal A}=\{x,y\}$, having $m$ occurrences of the letter $x$, and $n$ occurrences of the letter $y$. Recall that the corresponding classical bijection, between paths and words, is realized by choosing $(i_{s+1},j_{s+1})=(i_s,j_s)-(1,0)$ if $w_s=x$, and $(i_{s+1},j_{s+1})=(i_s,j_s)+(0,1)$ if $w_s=y$. We denote by $\pi_w$ the resulting path.
We also consider words, and associated paths, that are infinite in both directions,
\begin{equation}\label{le_mot}
w=\cdots w_{-3}w_{-2}w_{-1} w_0w_1w_2w_3 \cdots
\end{equation}
and say that they are \defn{bi-infinite} words (or paths). Such a word (and the associated path) is said to be \defn {admissible} if there are infinitely many $x$' s and $y$'s in both directions.
Let us now associate to a given word $w$ (finite or bi-infinite) a tiling ${\mathcal A}={\mathcal A}_{w;k}$, whose entries are obtained by the (weighted) enumeration of paths starting and ending at some points of $\pi_w$. We restrict these paths to stay within some ``distance'' $k$ of $\pi_w$. This is made more precisely below after the introduction of more notation and terminology (some of which will only be used later).
\begin{figure}[ht]\setlength{\unitlength}{5mm}
\begin{picture}(0,15)(-1.5,-8)
\put(-17,0){$\textcolor{red}{x}$}
\put(-12,5.5){$\textcolor{red}{y}$}
\put(-16,5){\textcolor{red}{\vector(0,-1){10}}}
\put(-16,5){\textcolor{red}{\vector(1,0){10}}}
\put(6,-6.5){\linethickness{4pt}\textcolor{yellow}{\line(0,1){10.5}}}
\put(5.5,-7){\linethickness{4pt}\textcolor{yellow}{\line(-1,0){13.5}}}
\put(-7,-7){\coude(-1,0)41\coude(3,1)33\coude(6,4)22\coude(8,6)11\coude(9,7)44}
\X(-10,-8){0}{2}
\Y(-8,-8){0}{1}
\Y(-8,-7){1}{1}
\Y(-7,-7){1}{1}
\Y(-7,-6){2}{1}
\X(-7,-5){2}{4}
\Y(-3,-5){2}{2}
\X(-3,-3){2}{3}
\Y(0,-3){2}{3}
\X(0,0){2}{4}
\Y(4,0){2}{1}
\X(4,1){2}{3}
\Y(6,2){2}{3}
\X(6,4){2}{3}
\X(9,4){1}{1}
\X(10,4){0}{1}
\put(5.8,-7){$p=(i,j)$}
\put(-16,-7.2){$(i,\beta_\pi(p))=\gamma_w(p)\rightarrow$}
\put(5,5.5){$\chi_w(p)=(\alpha_\pi(p),j)$}
\put(5.8,4.5){$\downarrow$}
\put(12.5,4.5){\textcolor{red}{$\pi_w$}}
\put(10.8,4.2){\textcolor{red}{\vector(3,1){1}}}
\put(-13,-9){\textcolor{red}{$\pi_w$}}
\put(-11.8,-8.4){\textcolor{red}{\vector(3,1){1}}}
\end{picture}
\caption{Path from $\gamma_w(p)$ to $\chi_w(p)$ in the $3$-fringe of $\pi_w$.}\label{fig_dessous}
\end{figure}
Given a path $\pi_w$ as above, for each point $p=(i,j)$ in $\mathbb{Z}\times \mathbb{Z}$ we denote by $\gamma_w(p)=(i,\beta_\pi(p))$ (resp. $\chi_w(p)=(\alpha_\pi(p),j)$) the leftmost (resp. topmost) point
that lies on the path $\pi_w$, which has the same first (resp. second) coordinate as $p$. We say that these are respectively the \defn{horizontal projection} and \defn{vertical projection} of $p$ on $\pi_w$.
A point $p=(i,j)$ is said to lie \defn{below} the path $\pi$ if we have the inequalities $\beta_\pi(p)\leq j$, or equivalently, $\alpha_\pi(p)\leq i$. Otherwise, we say that $p$ lies \defn{above} the path.
We now consider the word $w(p)$:
associated to the portion of the path $\pi_w$ going from the horizontal projection $\gamma_w(p)$ of $p$ on $\pi_w$, to its vertical projection $\chi_w(p)$. This word is used to define the notion of \defn{projection word} of a point $p$, denoted by $w_p$, as follows. We simply set $w_p:=w(p)$ whenever $p$ lies below the path. Otherwise, when $p$ lies above the path, we set $w_p:=\overline{w(p)}$. Here $\overline{w}$ is the operation corresponding to reading the letters of a word $w$ in reverse order, replacing each $x$ by $\overline{x}$ and each $y$ by $\overline{y}$. For example, with $w=yyxyxyyyx$, we get $\overline{w}=\overline{x}\,\overline{y}\,\oy\overline{x}\,\overline{y}\,\overline{x}\,\ox\overline{y}\,\oy$.
Let $p$ be a point lying below the path $\pi_w$, and suppose that $w_p$ factors as $x^i\,u\,y^j$ (with $i$ and $j$ maximal). Then we say that $u_p:=u$ is the \defn{short projection word} of $p$ on $w$.
Illustrating with the tiling of Figure~\ref{Fig_mot_slk}, one may check that for the points $p$ corresponding to the entries with value equal to $6$ (lying below the path), one has $w_p=xxyyxxyy$ and $u_p=yyxx$ (for all instances of $6$); whereas for the points $p$ corresponding to the entries $30$ (lying above the path), one has $w_p= \bar {y}\bar{y}\bar{x}\bar{x}$ (likewise for all instances of $30$).
For a point $p$ lying below a path $\pi$, the \defn{distance} between $p$ and $\pi$ is the unique integer $k\in{\mathbb N}$ such that $(i-k+1,j-k+1)$ lies on $\pi$. Observe that there is but one point of $\pi$ lying on any given diagonal $x=y+c$. Our definition makes it so that points lying on the path are considered to be at distance $1$ of it (this will make our life easier later).
We further consider the notion of \defn{$k$-fringe}, $\Phi_k(\pi)$ of a path $\pi$, i.e.: the points lying below the path that are within distance $k$ of it. Thus, we have
\begin{equation}\label{def_fringe}
\Phi_k(\pi):=\{(i+m,j+m)\ |\ (i,j)\in\pi,\quad{\rm and}\quad 0\leq m<k\}
\end{equation}
Some of these notions are illustrated in Figure~\ref{fig_dessous}.
Given two points $p$ and $q$ on the path $\pi_w$ of a word $w$, we consider the sets of paths
\begin{equation}\label{defn_ense_chemins}
{{\mathcal{P}}_{w;k}(p,q):=\{\theta\ {\rm a\ path}\ |\ {\rm s}(\theta)=p,\ {\rm e}(\theta)=q,\ {\rm and} \ \theta\subset \Phi_k(\pi_w)\}},
\end{equation}
that start at $p$, end at $q$, while staying inside the $k$-fringe of $\pi_w$. We also denote by ${\mathcal{P}}_{w;k}$ the set of all such paths, disregarding start and end points.
The tiling ${\mathcal A}_{w;k}=\left(a_{ij}\right)_{i,j}$ is then defined, for points $p=(i,j)$ lying below the path $\pi_w$, by setting
\begin{equation}\label{def_tiling}
{a_{ij}:=\# {{\mathcal{P}}_{w;k}(\chi_w(p),\gamma_w(p))}}.
\end{equation}
For instance, for the word $w=\ldots yyxxyxyyyx\ldots $ and $k=2$, the resulting (partial) tiling is as follows:
$$\{\mathcal A}_{w;2}=\begin {array}{ccccccc}
&&&&&&\r{1}\\
&&&\r{1}&\r{1}&\r{1}&\r{1}\\
&&\r{1}&\r{1}&2&3&4\\
&&\r{1}&2&5&8&11\\
\r{1}\ &\r{1}\ &\r{1}&3&8&13&18\end {array}$$
Observe that, for fixed $d$, any path $\theta$ in the $k$-fringe of $\pi$ goes through at most one of the the $k$ points of the set $\Phi_k(\pi)\cap \triangle_d$, where $\triangle_d$ denotes the \defn {diagonal}
\begin{equation}
\triangle_d:=\{\ (i,j)\in \mathbb{Z}\times \mathbb{Z} \ |\ i-j=d\ \}.
\end{equation}
It is useful to have the following terminology: given a bi-infinite path and a tiling ${\mathcal A}$, let us call \defn {principal minors of order} $m$ (relative to $\pi_w$), the minors of ${\mathcal A}$ that are of the form $M_{ij}^{(m)}$, with $(i,j)$ lying on the path $\pi_w$. In other word, a principal minor is an adjacent minor of ${\mathcal A}$ whose upper left corner lies on the path (in particular, it is contained in the $k-1$ fringe of the path when $m\leq k$).
For $h\in\mathbb{Z}$, we also say that a minor $M_{ij}^{(m)}$ is \defn{located on the $h$-th diagonal} if we have $h=j-i$.
To tie all this to our study of $\SL{k}$-tilings, we now give entirely combinatorial arguments for the following statements.
\begin{proposition}\label{prop_word_tiling}
{The partial tiling ${\mathcal A}_{w;k}$ is a $0$-free $\SL{k}$-tiling with principal minors of order $<k$ all equal to $1$.}
{Moreover, for any $k$-subsets $I$ and $J$ such that $I\times J$ is contained in the shape of ${\mathcal A}$ {\rm (}all the points lying below $\pi_w${\rm )},
we have}
\begin{equation}\label{mineurs_multiplicatifs_chemins}
{{M}_{IJ}={M}_{I,\{j,\ldots,j+k-1\}} \, {M}_{\{i,\ldots,i+k-1\},J}}.
\end{equation}
{Finally, if the path $\pi_w$ is admissible, then ${\mathcal A}_{w;k}$ extends uniquely to a complete tame $\SL{k}$-tiling.}
\end{proposition}
There is some redundance here, since the 0-free property implies that ${\mathcal A}$ is tame (Proposition~\ref{rankk}), and thus (\ref{mineurs_multiplicatifs_chemins}) holds by Proposition~\ref{rank1}. Notwithstanding, we want to make evident that nice combinatorial methods may be used to understand all this.
\begin{proof}[\bf Proof.]
It follows from a theorem of Gessel-Viennot (see \cite{viennot}) that
we may interpret combinatorially any minor of ${\mathcal A}_{w;k}$ as follows.
Recall that a family of paths is said to be non-crossing if no pair of paths in the family has a common point.
Given equal cardinality subsets $I$ of rows and $J$ of columns, such that $I\times J$ lies below $\pi_w$, we denote by ${\rm NoX}_{IJ}$ the set of all non-crossing families of paths in ${\mathcal{P}}_{w;k}$ linking the horizontal projection of $I$ on $\pi_w$ to the vertical projection of $J$ on $\pi_w$. More precisely, let
$$I=\{i_1<i_2<\ldots <i_m\}, \qquad{\rm and}\qquad J=\{j_1<j_2<\ldots <j_m\},$$
and denote by $p_1,\ldots, p_m$ and $q_1,\ldots, q_m$ the respective horizontal and vertical projections on $\pi_w$.
Then the elements of ${\rm NoX}_{IJ}$ are ``configurations'' $\{\pi_1,\pi_2,\ldots,\pi_m\}$ of paths $\pi_s $ in ${\mathcal{P}}_{w;k}$, with
\begin{equation}
{ {\rm s}(\pi_s)=p_s}, \quad {\rm and}\quad {{\rm e}(\pi_s)=q_s},
\end{equation}
no two of which cross. In our context, the aforementioned theorem of \cite{viennot} states that we have
\begin{equation}\label{mineurs_chemins}
{{M}_{IJ} = \# {\rm NoX}_{IJ}}.
\end{equation}
Recall that this is shown by constructing a sign changing involution on the set of crossing configurations, thus showing that they can be eliminated from a global signed counting that clearly corresponds to the evaluation of the determinant considered.
Observe that non-crossing path configurations $\{\pi_1,\pi_2,\ldots,\pi_m\}$ intersect any given diagonal $\Delta_d$ in at most $m$ points. In fact, this intersection number is exactly equal to $m$ for the diagonals that pass through points of $\pi_w$ lying between $p_1$ and $q_1$. This forces
all the sets $ {\rm NoX}_{IJ}$ to be empty whenever $\#I=\#J>k$. Hence the corresponding minors all vanish, so that the tiling is of rank $k$.
To continue with our combinatorial argument, let us write $ {\rm NoX}_{ij}^{(m)}$
when
$$I=\{i,\ldots,i+m-1\}, \quad {\rm and}\quad J=\{j,\ldots,j+m-1\},$$
so that we have
\begin{equation}\label{mineurs_chemins_ij}
{{M}_{ij}^{(m)} = \# {\rm NoX}_{ij}^{(m)}}.
\end{equation}
In the $k$-fringe of $\pi_w$, there is room for one and exactly one configuration of $k$ non-crossing paths having adjacent starting points and adjacent end points, so that
we must necessarily have ${M}_{ij}^{(k)} =1$ for all point $(i,j)$ in the tiling, hence the $\SL{k}$-condition is verified.
The tiling is $0$-free, and in fact we have ${M}_{ij}^{(m)}\geq 1$ for all $m$ between $1$ and $k$, since there exists corresponding non-crossing path configurations for all these $m$.
Finally, the multiplicative property (\ref{mineurs_multiplicatifs_chemins}) of $k\times k$-minors can easily be explained as follows, in terms of configurations of $k$ non-crossing paths.
With $I$ and $J$ satisfying the hypothesis of the proposition, the required identity follows from a simple bijection
\begin{equation}\label{chemins_multiplicatifs}
{{\rm NoX}_{IJ} \rightarrow {\rm NoX}_{I,\{j,\ldots,j+k-1\}} \times {\rm NoX}_{\{i,\ldots,i+k-1\},J}},
\end{equation}
obtained by breaking up the paths considered into three portions as follows. For a path $\pi_s$ (starting at $p_s$ and ending at $q_s$) in $\Phi_k(\pi_w)$, denote respectively by $p$ and $q$ the points of $\pi_s$ that lie on the diagonals respectively containing $p_1$ and $q_1$. These exist since $p_s\leq p_1\leq q_1\leq q_s$. We decompose $\pi_s$ as the concatenation
$$\pi_s= \pi_s^{(1)} \pi_s^{(2)} \pi_s^{(3)},$$
with
\begin{enumerate}
\item[$\bullet$] $\pi_s^{(1)}$ being the portion of $\pi_s$ going from $p_s$ to $p$,
\item[$\bullet$] $\pi_s^{(2)}$ being the portion of $\pi_s$ going from $p$ to $q$, and
\item[$\bullet$] $\pi_s^{(3)}$ being the portion of $\pi_s$ going from $q$ to $q_s$.
\end{enumerate}
In particular, all the paths $\pi_s^{(2)}$ start on the same diagonal (the one that contains $p_1$) and end on the same diagonal (the one that contains $q_1$). Since these $k$ paths are non-crossing and all lie in the $k$-fringe, there is but one possibility for the resulting configuration $\{\pi_1^{(2)},\ldots, \pi_k^{(2)}\}$. We easily identify the configurations $\{\pi_1^{(1)},\ldots, \pi_k^{(1)}\}$ with elements of ${\rm NoX}_{I,\{j,\ldots,j+k-1\}}$ (by application of the same decomposition as above to these last elements, observing that in this cases third components are trivial). Likewise we identify the configurations $\{\pi_1^{(3)},\ldots, \pi_k^{(3)}\}$ with elements of ${\rm NoX}_{\{i,\ldots,i+k-1\},J}$. This establishes the bijection.
The $\SL{k}$-tiling ${\mathcal A}_{w;k}$ may be uniquely completed into a tame $\SL{k}$-array by Lemma~\ref{lem_unique} below.
\end{proof}
For example, with $k=4$ and the word $w=\cdots xxyyxxyyxxyy\cdots$, we first get the partial array (lying below the path) of Figure~\ref{Fig_mot_slk} by path enumeration, and then complete it to get a $\SL{4}$-tiling of $\mathbb{Z}\times \mathbb{Z}$.
\begin{figure}[ht]
\begin{center}
\begin{picture}(-50,130)(0,0)
\put(-160,70){\textcolor{red}{$\begin {array}{ccccccccc}
\ddots&\vdots&\vdots&\vdots&\vdots&\ \vdots\ &\ \, \vdots\ \,&\quad \vdots\quad &\revddots\\
\cdots& 1437&457&30&10\\
\cdots& 457&146&10&\ \ 4\ \ \\
\cdots& 30&10\\
\cdots& 10&\ 4\ \\
\cdots\\
\cdots\\
\cdots
\end {array}$}}
\put(-150,55){{$\begin {array}{cccccccccc}
&&&&&&&& \\
&& & & & \r{1}&\ \r{1}\ &\ \r{1}\ &\cdots\\
& & & & & \r{1}& 2&3&\cdots\\
& & &\r{1}&\r{1}&\r{1}&3&\ 6&\cdots\\
& & &\r{1}&2&3& 10&22&\cdots\\
&\ \r{1}\ &\r{1}&\r{1}&3&6&\ 22\ &53&\cdots\\
&\r{1}& 2&3&10&22&84&211&\cdots\\
& \r{1}&\quad 3\quad &6&\ 22\ & 53& 211 &553&\cdots\\
\revddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots& \ddots
\end {array}$}}
\end{picture}
\end{center}
\caption{The $\SL{4}$-tiling associated to ${\ldots xxyyxxyyxxyy \ldots}$}\label{Fig_mot_slk}
\end{figure}
\begin{lemma}\label{lem_unique}
{Consider a partial $\SL{k}$-tiling which is defined on every point below a given path, and such that all of its adjacent $(k+1)\times (k+1)$-subminor (lying entirely in its shape) vanishes. If the path is admissible, then the partial $\SL{k}$-tiling extends uniquely to a complete tame $\SL{k}$-tiling.}
\end{lemma}
\begin{proof}[\bf Proof.]
Let $i$, $j$ and $k$ be integers such that $\{i,i+1,i+2,\ldots\} \times \{j-k,\ldots,j-1,j\}$ is contained in the shape of ${\mathcal A}$. denote by $C_{j-k},\ldots,C_{j-1},C_j$ the columns of the corresponding submatrix of ${\mathcal A}$. Then, it follows from the vanishing $(k+1)\times (k+1)$-subminor condition, that we have a relation of the form
$${ (-1)^kC_{j-k}-(-1)^{k}a_1C_{j-k+1}+\cdots -a_{k-1}C_{j-1}+C_j=0}.$$
for some $a_1,\ldots, a_{k-1}$ in $\mathbb{K}$. Note that the coefficients $a_h$ are independent of the $i$ chosen.
Considering the analogous argument for rows, and assuming that the origin of the plane is in the shape of ${\mathcal A}$, we obtain a linearization data (see Section~\ref{sec_general}). Using this linearization data, we may apply Proposition~\ref{coefflinear} to get a complete tame tiling of the plane. Call it ${\mathcal A}'$. It follows from the construction that ${\mathcal A}$ and ${\mathcal A}'$ coincide on the shape of ${\mathcal A}$, which proves the lemma.
\end{proof}
\subsection*{Weighted word tilings} We now extend the previous construction to the situation where paths are given Laurent monomial weights.
At a point $p=(i,j)$, along a path $\theta$, we say that we have a \defn{right-turn} (resp. \defn{left-turn}) if both $(i+1,j)$ and $(i,j+1)$) (resp. $(i,j-1)$ and $(i-1,j)$) belong to the path $\theta$. This is illustrated in Figure~\ref{Fig_turn}.
\begin{figure}[ht]\setlength{\unitlength}{8mm}
$$ \begin{matrix}
\begin{picture}(4,2)(-1,-.5)
\anticoude(0,0)11
\D(0,0){1}\X(0,1){0}{1}\D(1,1){1}
\put(1.2,1.2){$\revddots$}
\put(-0.9,-0.6){$\revddots$}
\put(1.4,.9){\small${(i,j+1)}$}
\put(-2,0){\small${(i+1,j)}$}
\put(-2.2,1.3){$\textcolor{red}{(i,j)}=p$}
\end{picture}&\qquad\qquad
\begin{picture}(4,2)(-1,-.5)
\coude(0,0)11
\D(0,0){1}\X(1,0){0}{1}\D(1,1){1}
\put(1.2,1.2){$\revddots$}
\put(-0.9,-0.6){$\revddots$}
\put(1.3,.9){\small${(i-1,j)}$}
\put(-2.2,0){\small ${(i,j-1)}$}
\put(1.3,-0.3){$p=\textcolor{red}{(i,j)}$}
\end{picture}\\
\nu(p)^{-1}& \nu(p)
\end{matrix}$$
\caption{Right and left turns at $p$, and corresponding weight.}\label{Fig_turn}
\end{figure}
For a given (bi-infinite) word $w$, we start by giving a weight $\nu(p)$ to each point $p=(i,j)$ in the $k$-fringe of $\pi_w$, setting
$$\nu(p):= \frac{t_{j-i,r}}{t_{j-i,r-1}},$$
where $r$ is the distance between $p$ and $\pi_w$, We assume here that the $t_{m,r}$ are independent commuting variables, setting $t_{m,r}=1$ whenever $r\leq 0$ or $r\geq k$.
With all this at hand, define the \defn{weight} $\omega(\theta)$ (with respect to the word $w$) of a nonempty path $\theta$ to be the product
\begin{equation}\label{defn_poids_chemin}
{\omega(\theta):=\alpha\, \beta \prod_{p\ {\rm left-turn\ of\ }\theta} \nu(p)\ \cdot \prod_{p\ {\rm right-turn\ of\ }\theta} \nu(p)^{-1}},
\end{equation}
where we set $\alpha:= \nu(p_s)$, if $\theta$ starts at $p_s$ by a vertical step. Otherwise we set $\alpha=1$. Likewise, we set $\beta:=\nu(p_e)$, if $\theta$ ends at $p_e$ after an horizontal step. Otherwise, we set $\beta=1$. Finally, when $\theta $ is the empty path, both starting and ending at $p$, we simply set $\omega(\theta):=\nu(p)$.
We then consider the partial tiling $\mathcal{B}_{w;k}:=(b_{ij})_{i,j}$, for point $(i,j)$ lying below the path, obtained by setting
\begin{equation}\label{pavage_pondere}
{b_{ij}:=\sum_{\theta} \omega(\theta)},
\end{equation}
for $\theta$ varying in the set ${\mathcal{P}}_{w;k}(\chi_w(i,j),\gamma_w(i,j))$, of paths starting at $\chi_w(i,j)$ and ending at $\gamma_w(i,j)$.
\begin{proposition} \label{pavage_variables_t}
{There is a unique tame $\SL{k}$-tiling of $\mathbb{Z}\times \mathbb{Z}$ extending $\mathcal{B}_{w;k}$, with entries Laurent polynomials in the variables $t_{hr}$. More precisely, the values are in the subsemiring generated by these variables and their inverses. Moreover, each principal minor of order $r$, $r=1,\ldots,k-1$, located on the $h$-th diagonal, is equal to $t_{hr}$.}
\end{proposition}
\begin{proof}[\bf Proof.]
Again, we simply apply the Gessel-Viennot technique, verifying that the involution (as in their original proof), required to show that crossing path configurations may be eliminated, is weight preserving. There are several cases, left to the reader.
It follows, as in the proof of Proposition~\ref{prop_word_tiling}, that $b_{ij}$ is a tame $\SL{k}$-tiling. It is clear that (\ref{pavage_pondere}) is in the described semiring.
Moreover, by the non-crossing path description, each principal minor of order $r<k$ is equal to $t_{hr}$, if the minor is located on the $h$-th diagonal.
\end{proof}
This proposition may be used to construct $\SL{k}$-tilings having arbitrary values (variables) as entries in the $(k-1)$-fringe of the path $\pi_w$. It turns out that the entries of the resulting tiling actually lie in the polynomial ring generated by these variables as well as the inverses of all principal minors (relative to $\pi_w$) of order at most $k-1$. This is a Laurent-like phenomenon (see \cite{fomin}) of a new kind. Moreover, we may in fact replace ``ring'' by ``semiring'', so that we actually get a positivity result, just as is the case in the theory of cluster algebras.
\begin{corollary}
{An admissible path $\pi_w$ being given, associate to each point in its $(k-1)$-fringe, a distinct commutative variable. Then this assigment extends uniquely into a complete tame $\SL{k}$-tiling of the plane whose values are in the semiring generated by the principal minors of order $<k$ and their inverses}.
\end{corollary}
\begin{proof}[\bf Proof.]
Consider the tiling of Proposition~\ref{pavage_variables_t}. Let $a_{ij}$ denote its value at the point $(i,j)$. Then, by the same proposition, each $a_{ij}$ is in the semiring generated by the variables $t_{hr}$ and their inverses, where $h\in \mathbb{Z}$ and $r=1,\ldots,k-1$.
Let $s_{hr}=t_{hr}/t_{h,r-1}$ for $h\in \mathbb{Z}$ and $r=1,\ldots,k-1$. Recall that the $t_{hr}$ are distinct commuting variables, and that $t_{h,0}=1$. The field of fractions $\mathbb{K}$ in the variables $t_{hr}$ is also generated by the $s_{hr}$, and the mapping $t_{hr}\mapsto s_{hr}$ is an automorphism of this field.
For $h\in \mathbb{Z}$ and $r$ going from $1$ to $k-1$, denote by $\alpha_{hr}$ the entry $a_{ij}$, if the point $(i,j)$ lies at distance $r$ below the path $\pi_w$, on the $h$-th diagonal.
By the path description, we see that, $\alpha_{hr}$ is the sum of $s_{hr}$ and of a fraction in the $t_{h'r'}$ with $(h',r')<(h,r)$ for the natural order on $\mathbb{Z}^2$. The latter fraction, when expressed in the $s_{h'r'}$, involves only variables $s_{h'r'}$ with the same condition. Hence, the function $t_{hr}\mapsto \alpha_{hr}$ defines an automorphism of $\mathbb{K}$.
Now, let $x_{hr}$ be a family of distinct commuting variables, for $h\in \mathbb{Z}$ and $r=1,\ldots,k-1$. Let $\mathbb{L}$ be its field of fractions. The fields $\mathbb{K}$ and $\mathbb{L}$ are isomorphic (e.g. by the mapping $t_{ij}\mapsto x_{ij}$). By what we have just seen, the mapping $\alpha_{hr}\mapsto x_{hr}$, $h\in \mathbb{Z}$ and $r=1,\ldots,k-1$ defines an isomorphism from $\mathbb{K}$ onto $\mathbb{L}$. If we map each $a_{ij}$ under this isomorphism, we obtain a tame $\SL{k}$-tiling $X=(b_{ij})$ such that $b_{hr}=x_{hr}$ for $h\in \mathbb{Z}$ and $r=1,\ldots,k-1$. This also implies that we may find elements $\tau_{hr}$ in the field $\mathbb{L}$ such that $b_{ij}$ is in the semiring generated by the $\tau_{hr}$ and their inverses, $h\in \mathbb{Z}$ and $r=1,\ldots,k-1$.
Furthermore, by Proposition~\ref{pavage_variables_t}, the principal $r\times r$-minor of $X$ ($r<k$), located on the $h$-th diagonal, is equal to $\tau_{hr}$.
Unicity follows from the following lemma, which of independent interest.
\end{proof}
\begin{lemma}
{An admissible path $\pi_w$ being given, associate to each point in its $(k-1)$-fringe, an element of some field. Suppose that the $(k-1)$-principal minors relative to $\pi_w$ are nonzero. Then this partial tiling extends uniquely to a tame $\SL{k}$-tiling of the plane.}
\end{lemma}
\begin{proof}[\bf Proof.]
Indeed, under the nonzero $(k-1)$-principal minor hypothesis, the $\SL{k}$ property imposes that we have a unique extension of the partial tiling to its $k$-fringe. This furnishes enough $k\times (k+1)$ and $(k+1)\times k$ submatrices so that we may compute the linearization data for any tame $\SL{k}$-tiling that would extend the $k$-fringe shaped partial tiling (see the remark following the proof of Proposition~\ref{coefflinear}). This proves unicity, in view of the same proposition.
Existence, which will not be used here, is left to the reader.
\end{proof}
In \cite{assem}, one may find many $\SL{2}$-tilings associated to paths, both over the integers, and with arbitrary variables on the path.
The case $\SL{3}$ has an extra interesting feature. Indeed, a consequence of (\ref{rec_trois}) is that we can very elegantly characterize any $\SL{3}$-tiling in tandem with its dual tiling. Indeed, under the assumption that ${\mathcal A}$ is $\SL{3}$ and writing ${\mathcal A}^*=\left(a^{*}_{ij}\right)_{i,j}$, the tiling identity considered is equivalent to the family of equalities
\begin{equation}\label{outils_SL3}
\begin{array}{rcl}
{a_{ij}}&=&{\displaystyle \frac{1}{a_{i-1,j-1}}(\textcolor{red}{a^{*}_{i-1,j-1}} + a_{i-1,j}a_{i,j-1})},\\[12pt]
\textcolor{red}{a^{*}_{ij}}&=&{\displaystyle\frac{1}{\textcolor{red}{a^{*}_{i-1,j-1}}}(a_{i-1,j-1} + \textcolor{red}{a^{*}_{i-1,j}a^{*}_{i,j-1}})}.
\end{array}
\end{equation}
This makes it evident (in another fashion) that the tiling constructed from a path is positive (and non-zero) for points lying below the path, since entries of ${\mathcal A}$ and ${\mathcal A}^*$ may be calculated recursively in parallel using the positive expression on the right-hand-side of (\ref{outils_SL3}). In the case of integer tilings, this is illustrated in Figure~\ref{pavage_et_dual}. Large entries correspond to the ${a_{ij}}$'s, and smaller ones correspond to the $\textcolor{red}{a^*_{ij}}$'s. The entry $\textcolor{red}{a^*_{ij}}$ sits immediately to the south-east of ${a_{ij}}$. Clearly the recursion process may be continued where it is left off. It corresponds to the statement that each number is obtained as the determinant of the $4$ numbers that immediately surround it.
\begin{figure}[ht]\setlength{\unitlength}{6mm}
{\begin{picture}(8,8)(0,0)
\put(2,8){ \put(0,0){$1$}\put(1,0){$1$}\put(2,0){$1$}\put(3,0){$1$}\put(4,0){$1$}\put(5,0){$1$}}
\put(1,6){ \put(0,0){$1$}\put(1,0){$1$}\put(2,0){$2$}\put(3,0){$3$}\put(4,0){$4$}\put(5,0){$\cdots$}}
\put(1,4){ \put(0,0){$1$}\put(1,0){$2$}\put(2,0){$5$}\put(3,0){$9$}\put(4,0){$\cdots$}}
\put(0,2){ \put(0,0){$1$}\put(1,0){$1$}\put(2,0){$3$}\put(3,0){$9$}\put(4,0){$\cdots$}}
\put(0,0){$1$}\put(1,0){$2$}\put(2,0){$7$}\put(3,0){$\cdots$}
\end{picture}}
\textcolor{red}{\begin{picture}(8,8)(7.8,-1)\tiny
\put(2,6){\put(0,0){$1$} \put(1,0){$1$} \put(2,0){$1$}\put(3,0){$1$}\put(4,0){$1$}}
\put(1,4){\put(0,0){$1$} \put(1,0){$1$} \put(2,0){$3$} \put(3,0){$6$}\put(4,0){$\cdots$}}
\put(1,2){\put(0,0){$1$} \put(1,0){$3$} \put(2,0){$14$} \put(3,0){$\cdots$}}
\put(0,0){$1$} \put(1,0){$1$} \put(2,0){$6$}\put(3,0){$\cdots$}
\end{picture}}
\caption{Joint calculation of a $\SL{3}$-tiling and its dual.}\label{pavage_et_dual}
\end{figure}
\section{Matrix description}
Consider the morphism $\mu$, from the free group $F_{x,y}$ (on the letters $x$ and $y$) to the group $\SL{k}$, which is obtained by setting
\begin{equation}
{\mu(x):=\mathrm{Id} + N}, \qquad{\rm and}\qquad
{\mu(y):=\mathrm{Id}+N^{\rm tr}},
\end{equation}
where we denote by $N$ the matrix nilpotent $k\times k$ matrix
$$N:=\begin{pmatrix}
0& 1& 0 &&\ldots &0\\
& 0 & 1 &0 & & \\
& & \ddots &\ddots&\ddots &\vdots \\
& & &0&1 &0\\
0& &\ldots & &0 &1
\end{pmatrix}$$
(Recall that $N$ is nilpotent of order $k$, so that $N^k=0$, and $N^i\not=0$ when $i<k$.)
We denote $\overline{x}$ the inverse of $x$ in the free group $F_{x,y}$, and likewise for $y$.
We then define the function ${\mathcal T}_w:\mathbb{Z}\times \mathbb{Z}\rightarrow {\mathbb N}$ as
\begin{equation}\label{pavage_mot_matrice}
{{\mathcal T}_w(p):=
{\bf e}_k\, \mu(w_p)\, {\bf e}_k^{\rm tr}}
\end{equation}
where ${\bf e}_k$ denotes the unit $k$-vector $(0,\ldots,0,1)$. Recall that the projection word $w_p$ has been defined in Section~\ref{words_tilings}.
\begin{proposition}\label{theo_neuf}
{For any admissible bi-infinite word $w$, the function ${\mathcal T}_w$ is a tame $\SL{k}$-tiling, whose principal minors of order $<k$ are all equal to $1$. It coincides with the tiling of Proposition~\ref{prop_word_tiling}.}
\end{proposition}
For the proof of Proposition~\ref{theo_neuf}, see Section~\ref{preuves}. Observe that this result easily implies the following:
\begin{corollary}
{With the same hypothesis as in Proposition~\ref{theo_neuf}, we have}
\begin{itemize}
\item[(i)] {If $p$ lies below $\pi_w$ then}
\begin{equation}
{ {\mathcal T}_w(p)= {\bf e}_k\, \mu(u_p)\, {\bf e}_k^{\rm tr}}.
\end{equation}
{In other words we can replace the projection word $w_p$ by the short projection word $u_p$ in our calculations}.
\item[(ii)] {If $p$ lies above $\pi_w$,with $w_p=\overline{x_1}\cdots \overline{x_n}$, $x_1,\ldots,x_n \in \{x,y\}$, then}
\begin{equation}
{ {\mathcal T}_w(p)={\bf e}_k\, \mu'({x_1}\cdots x_n)\, {\bf e}_k^{\rm tr}},
\end{equation}
{ where $\mu'$ is the morphism such that}
$${\mu'({x}):=(Id-N)^{-1}}, \qquad{\rm and}\qquad
{ \mu'({y}):=\mu'({x})^{\rm tr}}.$$
\end{itemize}
{In particular, we conclude that ${\mathcal T}_w(p)$ is positive for all $p$.}
\end{corollary}
\begin{proof}[\bf Proof.] Assume that $w=w_p=x^i\,u\,y^j$ with $u=u_p$. To show (i), we first observe that
\begin{eqnarray*}
\mu(x^i)&=&(\mathrm{Id}+N)^i\\
&=& \mathrm{Id}+\sum_{j=0}^i \binom{i}{j} N^i
\end{eqnarray*}
is upper unitriangular, hence ${\bf e}_k \mu(x^i) = {\bf e}_k$.
Likewise, $\mu(y^i)$ is lower unitriangular, so that $\mu(y^j) {\bf e}_k^{\rm tr}= {\bf e}_k^{\rm tr}$. Thus we directly calculate that
\begin{eqnarray*}
{\mathcal T}_w(p) &=& {\bf e}_k\, \mu(x^i u y^j)\, {\bf e}_k^{\rm tr}\\
&=& {\bf e}_k\, \mu(x^i)\mu(u)\mu( y^j)\, {\bf e}_k^{\rm tr}\\
&=& {\bf e}_k\, \mu(u)\, {\bf e}_k^{\rm tr}
\end{eqnarray*}
as announced.
For (ii), we make use of the matrix isomorphism
$$\alpha(A):=D_k\, A\, D_k^{-1},$$
with $D_k$ standing for the diagonal matrix with entry equal to $(-1)^{i+1}$ on the diagonal. Clearly, $\alpha(a_{ij})=((-1)^{i+j}a_{ij})$. Thus $\alpha(\mu'(x))=\mu(\overline{x})$ and $\alpha(\mu'(y))=\mu(\overline{y})$ as is easily verified. Hence, for any $x_1,\ldots,x_n \in \{x,y\}$, we have $\alpha(\mu(x_1\cdots x_n))=\mu(\overline{x_1}\cdots \overline{x_n})$. We conclude since $\alpha(A_{kk})={\mathcal A}_{kk}$. For the final assertion, note that $\mu'$ has nonnegative coefficients.
\end{proof}
\section{Proofs}\label{preuves}
To prove some of our previous assertions we first need a few
linear algebra lemmas.
\begin{lemma}\label{det_rank}
{If a matrix has all its adjacent $(k+1)\times (k+1)$ minors vanishing, whereas no adjacent $k\times k$ vanishes, then it is of rank $k$}.
\end{lemma}
\begin{proof}[\bf Proof.]
It is enough to show that for any choice of $k+1$ successive columns $C_0,\ldots, C_k$ of this matrix, $C_0$ (resp. $C_k$) is a linear combination of $C_1, \ldots, C_k$ (resp. of $C_0, \ldots, C_{k-1}$). By symmetry, the property with $C_0$ will suffice. Let $v_i$ denote the rows of the matrix $(C_0,\ldots,C_k)$. Note that $v_i$ is of length $k+1$.
To show our assertion, let us construct a non-vanishing linear form $\varphi$ that annihilates all $v_i$. The existence of such a linear form implies the existence of scalars $a_0,\ldots, a_{k}$ such that $\sum_{j=0,\ldots,k} a_jC_j=0$. Moreover $a_0$ has to be nonzero, since otherwise it would contradict the assumption on the non-vanishing $k$-minors.
Such a linear form exists for $k+1$ successive rows of $M$, since $\det(M)=0$ by assumption. Consider $k+2$ successive rows of $M$, and two non-vanishing linear forms $\varphi$ and $\psi$ such that the first $k+1$ rows are in ${\mathrm Ker}(\varphi)$ and the $k+1$ last are in ${\mathrm Ker}(\psi)$. Then we argue as follows to show that $\varphi$ and $\psi$ must necessarily be proportional. If we restrict the two linear forms to the $k$ intermediate rows, $v_1,\ldots, v_k$ say, we see that $\varphi$ and $\psi$, considered as column vectors of length $k+1$, are both annihilated by the $(k\times (k+1)$-matrix
\begin{equation}
\begin{pmatrix}
v_1\\v_2\\\vdots\\v_k
\end{pmatrix},
\end{equation}
whose rows are the vectors $v_i$.
By assumption, this matrix is of rank $k$, hence it has a kernel of dimension $1$. It follows that its columns vectors are proportional, and thus so are $\varphi$ and $\psi$.
\end{proof}
\begin{lemma}\label{rang1}
{Let $A$ be a square matrix of order $k+1$ such the matrix of its $k\times k$-minors $(\det(A_{IJ})_{I,J}$, with $I$ and $J$ running through all $k$-subsets of $\{1,\ldots,k+1\}$, is of rank $1$. Then $\det(A)=0$}.
\end{lemma}
\begin{proof}[\bf Proof.]
If the central $(k-1)\times (k-1)$-minor $\det A_{\{2,\ldots,k\},\{2,\ldots,k\}}$ of $A$ is nonzero, then (\ref{carrol}), with $r=k$, implies that $\det(A)=0$. If some $(k-1)\times (k-1)$-minor of $M$ is nonzero, we may bring it into central position by row and column permutations;
these operations amount to row and column permutations of the matrix of $k\times k$-minors of $A$; hence, by the previous argument, $\det(A)=0$. Finally, if all the $(k-1)\times (k-1)$-minors of $A$ vanish, then so does $\det(A)$.
\end{proof}
\begin{lemma}\label{lemme_1}
{Let $V$ be a vector space, and consider a finite ordered set of indices $K$ for which we have selected vectors
$v_k,v'_k$ in $V$, as well as $u_k, u'_k$ in the dual space $V^*$. Assuming that for all $k$ in $K$ we have the relations\footnote{Here, as in the sequel of this section, the stars {\rm (}$\rouge{\star}${\rm )} stand for some coefficients that we do not actually need to specify.}
\begin{eqnarray*}
v'_k&=& v_k +\sum_{\ell<k} \rouge{\star}\, v_\ell,\\
u'_k&=& u_k +\sum_{\ell<k} \rouge{\star}\, u_\ell.
\end{eqnarray*}
Then we have the equality}
\begin{equation}
{ \det\left(u_k(v_\ell)\right)_{k,\ell\in K} =\det \left(u'_k(v'_\ell)\right)_{k,\ell\in K}} .
\end{equation}
\end{lemma}
\begin{proof}[\bf Proof.]
We simply pass from one matrix to the other by multiplication on the left and on the right by uni-triangular matrices.
\end{proof}
Consider now
intervals of cardinality $k-1$ of the set $\{2,\ldots, 2k-1\}$, of the form
$$I_q:=\{q+1,q+2,\ldots, q+k-1\},\quad q=1,\ldots, k.$$
For convenience sake, we write $K_q:=[k]\setminus \{q\}$ (with $[k]$ standing as usual for $\{1,\ldots,k\}$). Let $e_1,\ldots, e_{2\,k-1}$ be elements of some vector space $V$. For $J=\{j_1,\ldots,j_s\}$ such that
$$1\leq j_1\leq \ldots\leq j_s\leq 2k-1,$$
we denote by $e_J$ the wedge product
$$e_J:=e_{j_1}\wedge e_{j_2}\wedge\ldots\wedge e_{j_s}.$$
Then the following holds.
\begin{lemma}\label{lemme_2}
{ For vectors $e_1,\ldots, e_{2\,k-1}$ in $V$ which are such that}
$${e_j=(-1)^{k-1}e_{j-k}+\rouge{\star}\, e_{j-k+1} +\ldots + \rouge{\star}\, e_{j-1},\qquad {\rm for}\quad j=k+1,\ldots, 2k-1},$$
{the following identity holds for all $q=1,\ldots, k$}:
\begin{equation}\label{formule_lemme_2}
{ e_{I_q}= (-1)^{q+1} e_{K_q}+ \rouge{\star}\, e_{K_{q-1}}+\ldots+ \rouge{\star}\, e_{K_1}}.
\end{equation}
\end{lemma}
\begin{proof}[\bf Proof.]
Writing $E:=e_{q+1}\wedge e_{q+2}\wedge\ldots\wedge e_k$ and $\varepsilon:=(-1)^{k-1}$, we calculate that
\begin{eqnarray*}
e_{I_q}&=& E\wedge e_{k+1}\wedge \ldots\wedge e_{k+q-1}\\
&=& E\wedge e_{k+1}\wedge \ldots\wedge e_{k+q-2}\wedge(\varepsilon e_{q-1}+\rouge{\star}\,e_q+\ldots + \rouge{\star}\, e_{k+q-2} ) \\
&=& E\wedge e_{k+1}\wedge \ldots\wedge e_{k+q-2}\wedge(\varepsilon e_{q-1}+\rouge{\star}\,e_q ) \\
&&\qquad\hbox{\rm (since $e_{q+1},\ldots e_{k+q-2}$ appear as factors in the product $E\wedge e_{k+1}\wedge \ldots \wedge e_{k+q-2}$.)}\\
&=& E\wedge e_{k+1}\wedge \ldots\wedge e_{k+q-3}\wedge(\varepsilon e_{q-2}+\rouge{\star}\, e_{q-1} +\ldots + \rouge{\star}\, e_{k+q-3})\\
&&\hskip1.7in \wedge(\varepsilon e_{q-1}+\rouge{\star}\,e_q ) \\
&=& E\wedge e_{k+1}\wedge \ldots\wedge e_{k+q-3}\wedge(\varepsilon e_{q-2}+\rouge{\star}\, e_{q-1} +\rouge{\star}\, e_{q})\\
&&\hskip1.7in \wedge(\varepsilon e_{q-1}+\rouge{\star}\,e_q ) \\
&=&\ldots\\
&=&E\wedge (\varepsilon e_1+\rouge{\star}\, e_2+\ldots+\rouge{\star}\,e_q) \wedge\ldots \\
&&\qquad \wedge(\varepsilon e_{q-2}+\rouge{\star}\, e_{q-1} +\rouge{\star}\, e_{q})\\
&&\qquad \wedge(\varepsilon e_{q-1}+\rouge{\star}\,e_q ) \\
&=&(-1)^{(q-1)(k-q)} (\varepsilon e_1+\rouge{\star}\, e_2+\ldots+\rouge{\star}\,e_q) \wedge\ldots \wedge(\varepsilon e_{q-2}+\rouge{\star}\, e_{q-1} +\rouge{\star}\, e_{q})\\
&&\qquad \wedge(\varepsilon e_{q-1}+\rouge{\star}\,e_q )\wedge E.
\end{eqnarray*}
The product that precedes $E$ is evidently in the $(q-1)^{\rm th}$-exterior power of the span of $e_1,\ldots,e_q$. It is thus a linear combination of the $e_{[q]\setminus \{i\}}$, for $i=1,\ldots,q$. It follows (as we are multiplying these $e_{[q]\setminus \{i\}}$ by $E$ on the right) that we have expressed $e_{I_q}$ as a linear combination of the $E_{K_i}$, for $i=1,\ldots, q$. Moreover, $e_{K_q}$ appears only once in the resulting expression. Its coefficient is thus $(-1)^{(q-1)(k-q)}(-1)^{(q-1)(k-1)}$. We conclude that (\ref{formule_lemme_2}) holds, since $(q-1)(k-q)+(q-1)(k-1)\equiv (q-1)(-q-1)\equiv (q+1)^2\equiv(q+1)$ modulo $2$.
\end{proof}
In the next result, the first row and first column of matrices are indexed by $1$.
\begin{proposition}\label{Condense}
{Consider a $(2k-1)\times(2k-1)$ matrix $A$ of rank $k$ having all of its adjacent $k\times k$ minors equal to $1$, and write $B=A_{11}^{(k)}$, $C=\partial_{k-1}A$, and $D=C_{22}^{(k)}$. Then, for all $h\leq k$, we have
$$\det D_{11}^{(h)}=\det B_{h+1,h+1}^{(k-h)}.$$}
\end{proposition}
Observe that the square matrices $A,B,C,D$ are respectively of order $2k-1$, $k$, $k+1$ and $k$ (as illustrated in Figure~\ref{fig_ABCD}). Recall also that $\partial_{k-1}A$ is the matrix of adjacent $k-1$-minors of $A$.
\begin{figure}[ht]
\setlength{\unitlength}{2mm}
\begin{center} \label{dessin}
\begin{picture}(15,19)(-3,-1.5)
\put(3.5,10.5){$\textcolor{red}{B}$}
\put(3.8,17){$\textcolor{red}{k}$}
\put(0.2,15.5){\textcolor{red}{$\overbrace{\hskip1.6cm}$}}
\put(-2.5,6.5){${A}$}
\put(0.25,6.8){\textcolor{red}{\carre(0,0)(8)}}
{\carre(0,0)(15)}
\put(4.5,-3){${2\,k-1}$}
\put(0,-0.3){{$\underbrace{\hskip3cm}$}}
\end{picture}\qquad \qquad
\begin{picture}(15,16)(-3,-1)
{\carre(0,6)(9.5)}
\put(1.2,6.2){\textcolor{red}{\carre(0,0)(8)}}
\put(4.5,9.5){$\textcolor{red}{D}$}
\put(4.8,2.8){$\textcolor{red}{k}$}
\put(1.3,5.7){\textcolor{red}{$\underbrace{\hskip1.6cm}$}}
\put(2.3,17.5){${k+1}$}
\put(0,16){{$\overbrace{\hskip 1.9cm}$}}
\put(-2.5,10){${C}$}
\end{picture}
\caption{The square matrices $A$, $B$, $C$, and $D$ of Proposition~\ref{Condense}.} \label{fig_ABCD}
\end{center}
\end{figure}
\begin{proof}[\bf Proof.]
Let $e_1,\ldots,e_{2k-1}$ be the column vectors of $A$, and consider the vector space $V$ that they span. Dually, let $\varphi_1,\ldots,\varphi_{2k-1}$ be the restriction to $V$ of the
$2k-1$ projections of column vectors on the underlying field of scalars.
We clearly have $\varphi_i(e_j)=a_{ij}$. Using the usual duality\footnote{$\langle \psi_1\wedge\ldots\wedge \psi_k, v_1\wedge\ldots \wedge v_k\rangle
=\det\left(\psi_i(v_j)\right)_{1\leq i,j\leq k}$.} between $(V^*)^{\wedge k}$ and $V^{\wedge k}$, we see that $d_{ij}=c_{i+1,j+1}=\langle\varphi_{I_i},e_{I_j}\rangle$, with the notations introduced before Lemma~\ref{lemme_2}. Thus, the determinant of $D_{11}^{(h)}$ is equal to $\det\left(\varphi_{I_i}(e_{I_j})\right)_{1\leq i,j\leq h}$.
In view of the hypotheses on $A$, we have $$e_j= \varepsilon\, e_{j-k}+\rouge{\star}\, e_{j-k+1} +\ldots + \rouge{\star}\, e_{j-1},\qquad ({\rm where\ as\ before}\ \varepsilon:=(-1)^{k-1})$$
for all $j=k+1,\ldots, 2k-1$ (as in the hypothesis of Lemma~\ref{lemme_2}). Dually we have
$$\varphi_j= \varepsilon\, \varphi_{j-k}+\rouge{\star}\, \varphi_{j-k+1} +\ldots + \rouge{\star}\, \varphi_{j-1},$$
for all $j=k+1,\ldots, 2k-1$.
Applying Lemma~\ref{lemme_2}, we get for $1\leq i,j\leq h$ that
\begin{eqnarray*}
\varphi_{I_i}&=&(-1)^{i+1} \varphi_{K_i}+\rouge{\star}\, \varphi_{K_{i-1}}+\ldots+\rouge{\star}\,\varphi_{K_1},\quad{\rm and}\\
e_{I_j}&=&(-1)^{j+1} e_{K_j}+\rouge{\star}\, e_{K_{j-1}}+\ldots+\rouge{\star}\,e_{K_1}.
\end{eqnarray*}
Using Lemma~\ref{lemme_1}, we conclude that the above determinant is equal to
$$\det\left((-1)^{i+j}\varphi_{K_i}(e_{K_j})\right)_{1\leq i,j\leq h},$$
which is exactly the $h\times h$-minor of the adjoint matrix of $B$, corresponding to rows and columns going from $1$ to $h$.
To finish the argument, we apply a result Jacobi
stating that (in the case of matrices of determinant $1$) a minor is equal to the complementary minor of the adjoint matrix.
\end{proof}
\begin{proof}[\bf Proof of Proposition~\ref{derive_equation}.]
Proposition~\ref{Condense} implies (\ref{condensation_devive}). This equation, for $r=k$ and $s=0$ implies that the dual is a $\SL{k}$-tiling. For $r=k-1$ and $s=1$, it implies that the tiling coincides with its bidual, up to the necessary translation.
In order to show that the dual is tame, we proceed as follows. Observe that for any matrix (finite or infinite) $(a_{ij})$ of rank at most $k$, there exist a vector space $E$ of dimension at most $k$, vectors $u_j\in E$, and linear forms $\varphi_i$ on $E$, all such that $a_{ij}=\varphi_i(u_j)$ (take the space spanned by the columns and the linear function obtained by projections of the columns). Conversely, such a data gives a matrix $(a_{ij})$ of rank at most $k$.
Now we form the matrix $(\langle \varphi_I,u_J\rangle)_{IJ}$, over some family of $k-1$-subsets of the row and column indices. Then $u_j$ is in the $(k-1)$-th exterior power of $E$, which is of dimension at most $k$. Hence this new matrix is of rank at most $k$. This implies that the dual is at rank at most $k$, since it is obtained from the original tiling by such a construction.
\end{proof}
Our proof of Proposition~\ref{theo_neuf} relies on the following two lemmas.
\begin{lemma}\label{thm9lemma1} {Let $p$ and $q$ be two points that are adjacent horizontally, i.e.: $p=(a,b)$ and $q=(a,b+1)$.
Then $w_q=w_p\,x\,y^i$, where $i+1$ is the number of points lying on the path $\pi_w$ in the same vertical as $q$.}
\end{lemma}
\begin{proof}[\bf Proof.]
Denote by $q_0,\ldots,q_i$ these $i+1$ points (starting from the top), and by $r_1,\ldots,r_j$ all the points of $\pi_w$ lying to the left of $q_i$ (labelled from left to right). This corresponds to the portion of the path $\pi_w$ illustrated in Figure~\ref{portion}.
\begin{figure}[h]
$$ \setlength{\unitlength}{6mm}\setlength{\carrelength}{5mm} \begin{picture}(4,5)(0,-1.5)
\Case{1,3}{$\!q_0$} \Case{1.9,3}{}
\put(.9,1.5){$\vdots$}
\Case{-3,0}{$\!r_1$} \put(-1.8,0){$\cdots$} \Case{0.1,0}{$\!r_j$} \Case{1,0}{$\!q_i$}
\Case{-3,-1}{}
\end{picture}$$
\caption{Points of the path $\pi_w$ that lie on the same row and column as $q$.}\label{portion}
\end{figure}
Clearly we have $w_{q_k}=y^k$, for $k<i$, and $w_{q_i}=x^j y^i$. Each of the following cases is clear (it helps to consider Figure~\ref{portion}), using the definition of $w_p$ in Section~\ref{words_tilings}.
\begin{enumerate}
\item[1)] If $q$ lies strictly above $q_0$, we have
$$w_p=w_q\,\overline{y}^i\, \overline{x},$$
implying that $w_q=w_p\, x\, y^i$ as required.
\item[2)] When $q=q_j$, for $0\leq j\leq i-1$, then
$$w_p= \overline{y}\,^{i-j} \overline{x}=y^j\overline{y}^i\overline{x}$$
so that again we have $w_q=y^j = w_p\,x\,y^i$.
\item[3)] Finally when $q$ lies below $q_i$, we evidently have $w_q=w_p\,x\,y^i$, thus the assertion is verified for all possible cases.
\end{enumerate}
\end{proof}
\begin{lemma}\label{thm9lemma2}
{For $m_1,\ldots,m_k$ in ${\mathbb N}$, let
$$u_i:=x^{m_{i-1}} y\ldots x^{m_1}y, $$
when $0\leq i\leq k$. {\rm (}In particular $u_0=1$.{\rm )}
Then we have
$$\begin{pmatrix} {\bf e}_k \,\mu(u_0)\\
{\bf e}_k \,\mu(u_1)\\
\vdots\\
{\bf e}_k\,\mu(u_{k-1})
\end{pmatrix} =
\begin{pmatrix}
0 & 0 &\ldots &0& 1\\
0 & 0 &\ldots &1& \rouge{\star}\\
\vdots & \vdots &\ddots & \vdots& \vdots\\
1 & \rouge{\star} &\ldots &\rouge{\star}& \rouge{\star}\\
\end{pmatrix}$$
}\end{lemma}
\begin{proof}[\bf Proof.]
We recursively show that ${\bf e}_k\,\mu(u_i)=(0,\ldots,0,1,\rouge{\star},\ldots,\rouge{\star})$ with $1$ sitting in position $(k-i)$.
If $i=0$, we have ${\bf e}_k \,\mu(u_0)={\bf e}_k=(0,\ldots,0,1)$, so that $1$ indeed sits in position $k$.
By induction we may assume that the first nonzero value of the vector
$${\bf v}:={\bf e}_k\, \mu(x^{m_{i-1}} y\ldots x^{m_2}y)$$
is a $1$ sitting in position $(k-i)$. Then the the first nonzero value of the vector
\begin{eqnarray*}
{\bf e}_k\ \mu(x^{m_{i-1}} y\ldots x^{m_1})&=& {\bf v}\, \mu(x^{m_1})\\
&=& (0,\ldots,0,1,\rouge{\star},\ldots,\rouge{\star}),
\end{eqnarray*}
also sits in position $(k-i)$, since $\mu(x^{m_1})$ is upper unitriangular. We can thus easily conclude
since ${\bf e}_k\,\mu(u_i)$ is obtained by multiplying (on the right) this last vector by $\mu(y)=\mathrm{Id}+N^{\rm tr}$,
hence its first nonzero value lies in position $(k-i-1)$.
\end{proof}
\begin{proof}[\bf Proof of Proposition~\ref{theo_neuf}.]
Let us first check that ${\mathcal T}_w$ is indeed a $\SL{k}$-tiling. Consider any set of points $p_{ij}$, $0\leq i,j\leq k-1$, forming an adjacent $k\times k$ sub-array of $\mathbb{Z}\times \mathbb{Z}$, and let us write $w_{ij}$ for the projection word $w_{p_{ij}}$ associated to these points $p_{ij}$. From Lemma~\ref{thm9lemma1} and its symmetric statement, there exists integers $m_0,\ldots ,m_{k-1}$ and $n_0,\ldots,n_{k-1}$ such that
$$w_{ij}= u_i\,w_{00}\,v_j,$$
with
$$u_i= x^{m_{i-1}} y\ldots x^{m_1}y,\qquad {\rm and}\qquad v_j=x\,y^{n_1} \ldots x\,y^{m_{j-1}}.$$
We have the matrix identity
\begin{eqnarray*}
\Big({\bf e}_k\ \mu(u_i\,w_{00}\,v_j)\ {\bf e}_k^{\rm tr}\Big)_{0\leq i,j\leq k-1}
&=& \Big({\bf e}_k\ \mu(u_i)\,\mu(w_{00})\, \mu(v_j)\ {\bf e}_k^{\rm tr}\Big)_{0\leq i,j\leq k-1}\\
&=& \begin{pmatrix} {\bf e}_k \ \mu(u_0)\\
\vdots\\
{\bf e}_k\ \mu(u_{k-1})
\end{pmatrix}\,
\mu(w_{00})\,
\big( {\bf e}_k \ \mu(v_0),\
\ldots\
,\ {\bf e}_k\ \mu(v_{k-1}\big).
\end{eqnarray*}
The fact that $\det(w_{00})=1$, together with Lemma~\ref{thm9lemma2}, implies that this matrix has determinant $1$ as announced.
To show that $\mathcal{T}_w$ is tame we argue as follows. First, observe that any given row of $\mathcal{T}_w$ is of the form
$$({\bf e}_k\ \mu(m)\ {\bf e}_k^{\rm tr})_{m\in \mathcal{M}},$$
where $\mathcal{M}$ is the (ordered) set of projection words of the points on this row.
Choose $k$ other rows, lying below the given row. These $k$ rows are (successively) of the form
$$({\bf e}_k\ \mu(m_i\cdots m_1\,m)\ {\bf e}_k^{\rm tr})_{m\in \mathcal{M}},$$
for $i$ running from $1$ to $k$, and suitable words $m_1$, $\ldots$, $m_k$. Now, the $k+1$ row vectors ${\bf e}_k\, \mu(m_i\cdots m_1)$, $i=0,\ldots, k$, are perforce linearly dependent, since
they are all of length $k$. Multiplying, this linear combination by $\mu(m)\,{\bf e}_k^{\rm tr}$ on the right, we find that the $k+1$ chosen rows of $\mathcal{T}_w$ are linearly dependent, and hence $\mathcal{T}_w$ is of rank $\leq k$.
The proof that $\mathcal{T}_w$ is the same $\SL{k}$-tiling as the one described in Proposition~\ref{prop_word_tiling}, using Lemma~\ref{lem_unique}, is left to the reader.
\end{proof}
\section {Applications}\label{sec_applications}
\subsection{$\SL{2}$-Frieze patterns revisited}\label{revisit}
The aim of this section is to show that the frieze patterns of Coxeter may be realized in terms of $\SL{2}$-tilings. This gives a new slant on their study, with emphasis on their link with representations of $\SL{2}(\mathbb{Z})$.
\begin{proposition} \label{extension}
{Let $a_i$, $i\in\mathbb{Z}$ be non-zero elements in the field $K$. There exists a unique tame $\SL{2}$-tiling that extends the partial tiling of {\rm (\ref{frieze_esquisse})}}.
\begin{equation}\label{frieze_esquisse}{
\begin{array}{rrrrrrrrrrrrrrrrrrrrrr}
&&&\ddots\\
&&&\ddots& 1\\
&&&& a_{-1} & 1 \\
&&&& &a_{0} & 1 \\
&&&& & &a_{1} & 1 \\
&&&& & &&\ddots & \ddots \\
&&&& & && & a_i& 1 \\
&&&&&&&&&\ddots&\ddots
\end{array}
}\end{equation}
\end{proposition}
To better study such tilings, let us consider the notion of \defn{signed continuant polynomials} $q_n(x_1,\ldots,x_n)$ defined by the recurrence
\begin{equation}\label{continuants}
{
q_n(x_1,\ldots,x_n):=x_n\, q_{n-1}(x_1,\ldots,x_{n-1})- q_{n-2}(x_1,\ldots,x_{n-2}),
}\end{equation}
whenever $n>0$, setting $q_{-1}:=0$ and $q_0:=1$. We omit indices when possible, writing simply $q(x_1,\ldots,x_n)$ for $q_n(x_1,\ldots,x_n)$.
Let us now consider the particular $\SL{2}$ matrices
$${Y(t):=\begin{pmatrix} 0 & -1\\ 1 & t \end{pmatrix}},$$
for which one may easily show by induction that
\begin{equation}\label{matricecontinuants}
{Y(x_1) Y(x_2) \cdots Y(x_n)= \begin{pmatrix} -q(x_2,\ldots,x_{n-1}), & -q(x_2,\ldots,x_{n})\\
q( x_1,\ldots,x_{n-1}), & q(x_1,\ldots,x_{n})\end{pmatrix}}.
\end{equation}
\begin{proof} [\bf Proof of Proposition~\ref{extension}.]
To prove unicity, we exploit the fact that the $\SL{2}$-tiling contains subarrays of the form
\begin{equation}\label{sous_matrice}
{
\begin{matrix}
1&0&*\\
a_i&1&0
\end{matrix}}
\end{equation}
Indeed, this follows directly from the $\SL{2}$-property.
Let $C_1$, $C_2$, and $C_3$ be the three corresponding columns, from left to right.
Then, since the tiling is of rank 2, we have $C_1-\alpha C_2+C_3=0$, which forces $\alpha=a_i$. Thus the coefficients of linearization are completely determined. Moreover, we are given at least one adjacent $2\times 2$-subarray, namely lower left $2\times 2$-submatrix of (\ref{sous_matrice}). Thus, unicity of the tiling follows by Proposition~\ref{coefflinear}.
For the existence of the tiling, we check that we may define its entries as follows:
$$ {\begin{matrix}
0 & -1 & -a_{\textcolor{red}{i}} &&\ldots & & \beta\\
1 & 0 & -1 &\ddots &&&\vdots\\
a_{\textcolor{red}{i}} & 1 & 0 &\ddots&\ddots&&\vdots \\[3pt]
\vdots & \ddots&\ddots & \ddots &\ddots&\ddots&\vdots\\
\vdots&&\ddots&\ddots&0&-1& -a_{\textcolor{red}{j}}\\
\vdots&&&\ddots&1 & 0 & -1\\
\alpha & \cdots&\cdots &\cdots&a_{\textcolor{red}{j}} & 1 & 0 \end{matrix}}$$
with $\beta=-\alpha$, and
\begin{equation}\label{valeurfrieze}
{ \alpha =q(a_i,\ldots,a_j)}.
\end{equation}
We then need only verify that the resulting tiling has the right the linearization coefficients (as in the first part of the proof). To this aim, let us denote by $\alpha'$ and $\alpha''$ the two entries of the tiling that sit immediately to the right of $\alpha$, so that we have
$$
\begin{matrix}
\alpha & \alpha' & \alpha''
\end{matrix}
$$
and therefore $\alpha'=q(a_{i+1},\ldots,a_j)$, and $\alpha''=q(a_{i+2},\ldots,a_j)$. But the recurrence~(\ref{continuants}) implies that
$${\alpha-a_i\,\alpha'}+\alpha''=0.$$
Hence, since the proof for $\beta$ is analoguous, the tiling defined above has the desired linearization coefficients.
\end{proof}
Consider now any frieze pattern, as below, with the $a_i$ positive integers, having $n$ diagonals (see Figure~\ref{fig1_5}).
\begin{equation}\label{frieze}
\begin{array}{ccccccccccccccc}
&&&&&&&\ddots\\
&&&&&&&\ddots&1 \\
&&&&&&&&a_{-2}&1\\
&&&&&&&\revddots&&a_{-1}&1\\
&&&&\ddots&&\revddots&&&&a_0&1\\
&&&&&1&&&&&&a_1&1\\
&&&&&&1&&&&&&a_2&1\\
&&&&&&&1&&&&\revddots&&\ddots&\ddots\\
&&&&&&&&1&&\revddots\\
&&&&&&&&&1\\
&&&&&&&&&&\ddots\\
\end{array}
\end{equation}
Observe that, since the coefficients of the frieze pattern are positive, they are completely characterized by the $a_i$, in view of the $\SL{2}$-property. Hence this frieze pattern extends uniquely to the same complete $\SL{2}$-tiling ${\mathcal A}$ as the one given by Proposition~\ref{extension}. Note that this extension has a few values more, immediately deduced from the positivity of the entries of the frieze pattern, and the $\SL{2}$-property, without resulting to tameness. These are the $0$'s and $-1$'s given below. We may therefore extract from the tiling ${\mathcal A}$ the following subarray, where $n$ is the number of diagonals of the frieze pattern and $i\in\mathbb{Z}$:
$$
\begin{array}{ccccccccccccccccccc}
a_{i+1}\\
*&a_{i+2}\\
\vdots&\ddots&\ddots\\
\vdots&&&\ddots\\
1&\hdots&\hdots&*&a_{i+n-1}\\
0&1&\hdots&\hdots&*&a_{i+n}\\
-1&0&\hdots&\hdots&\hdots&*&a_{i+n+1}
\end{array}
$$
By Formula~(\ref{valeurfrieze}) for the entries of the tiling ${\mathcal A}$, we obtain
$$\begin{array}{lll}
q(a_{i+1},\ldots,a_{i+n+1})=-1, & q(a_{i+1},\ldots,a_{i+n})=0,\\
q(a_{i+2},\ldots,a_{i+n})=1,& q(a_{i+2},\ldots,a_{i+n+1})=0
\end{array},$$
and, using (\ref{matricecontinuants}), we conclude that
\begin{equation}\label{=-1}
{Y(a_{i+1}) Y(a_{i+2}) \cdots Y(a_{i+n+1})= \begin{pmatrix} -1 & 0\\
0 & -1\end{pmatrix}}.
\end{equation}
Using this we may prove
\begin{corollary}\label{antiperiod}
{Let ${\mathcal A}=(a_{ij})$ be the unique tame $\SL{2}$-tiling ${\mathcal A}=(a_{ij})$ extending a given frieze pattern $\mathcal{F}$ with $n$ diagonals}.
\begin{enumerate}
\item[(i)] {${\mathcal A}$ has diagonal period $n+1$, that is to say $a_{i+n+1,j+n+1}=a_{ij}$}.
\item[(ii)] {\rm (Coxeter \cite{coxeter}, Conway-Coxeter\cite{conway})} {In particular, the frieze pattern has diagonal period $n+1$}.
\item[(iii)] { Moreover, ${\mathcal A}$ has horizontal and vertical skew-period $n+1$, this is to say that}
$${a_{i+n+1,j}=-a_{ij}=a_{i,j+n+1}}.$$
\item[(iv)] {\rm (Coxeter \cite{coxeter}, Conway-Coxeter\cite{conway})} { Finally, ${\mathcal A}$ and $\mathcal{F}$ are invariant under a glided symmetry, which is the symmetry with respect to the middle diagonal of $\mathcal{F}$ followed by the diagonal translation of length $\frac{n+1}{2}$}.
\end{enumerate}
\end{corollary}
\begin{proof}[\bf Proof.]
Referring to (37), let $C_i$ denote the column containing the coefficient $a_i$. Then we have $C_{i}-a_iC_{i+1}+C_{i+2}=0$, as is shown at the beginning of the proof of Proposition~\ref{extension}. Thus we have the following recurrence between the $\mathbb{Z}\times 2$ matrices $(C_i,C_{i+1})$:
$$
(C_i,C_{i+1})=(C_{i+1},C_{i+2})\, Y(a_i).
$$
Thus, (\ref{=-1}) implies that ${\mathcal A}$ has horizontal skew-period $n+1$. Vertical periodicity follows by symmetry, and the diagonal periodicity follows at once.
In order to prove (iv), note that (\ref{=-1}) implies
$$
Y(a_{i+2}) \cdots Y(a_{i+n+1})=-Y(a_{i+1}^{-1})= \begin{pmatrix} -a_{i+1} & -1\\
1 & 0\end{pmatrix}.
$$
Thus $a_{i+1}=q(a_{i+3},\ldots,a_{i+n})$ by (\ref{matricecontinuants}). This shows, by taking $i=n,0,1,\ldots$ and recalling that we have the diagonal period $n+1$ (hence $a_i=a_{i+n+1}$) that: $a_{n+1}=q(a_{n+3},\ldots, a_{2n})=q(a_2,\ldots, a_{n-1})$, $a_1=q(a_3,\ldots,a_n)$, \ldots.
Hence, using (\ref{valeurfrieze})
we see
that $R$ has the following form, extending (\ref{frieze_esquisse}):
$$
{
\begin{array}{cccccccccccccccccccccc}
&&&\ddots\\
&&&\ddots& 1\\
&&&& a_{1} & 1 \\
&&&& &a_{2} & 1 \\
&&&& & &a_{3} & 1 \\
&&&\ddots&\ddots & &&\ddots & \ddots \\
&&&& 1& a_{n+1}&& & a_{n-1}& 1 \\
&&&&&1&a_1&&&a_n&1\\
&&&&&&1&a_2&&&a_{n+1}&1\\
&&&&&&&\ddots&\ddots&&&\ddots&\ddots
\end{array}
}
$$
We conclude by using a symmetric version of Proposition~\ref{extension}.
\end{proof}
Following Conway-Coxeter (in \cite{conway}) we call \defn{quiddity} a sequence $a_1,\ldots,a_{n+1}$, where $a_i$ gives the number of triangle incident to the vertex $i$ in a triangulation of a convex $(n+1)$-gone, whose vertex are successively labeled $1$ to $n+1$ turning around the $n$-gone. They show \cite[p. 180]{conway} that any quiddity may be obtained from the particular quiddity $1\,1\,1$ by successive applications of the local rewriting rule
$$ {\cdots\,\textcolor{red}{a\,b}\,\cdots \quad\rightarrow\quad \cdots\,\textcolor{red}{a+1\,1\,b+1}\,\cdots}.$$
We prove below their result that quiddities and frieze patterns are in one-to-one correpondence. For this, we make a detour through presentations of the group $\SL{2}(\mathbb{Z})$.
\begin{proposition}\label{rewriting}
{
Consider the rewriting rule in the free monoid $\mathbb{P}^*$ generated by $\mathbb P$}
\begin{equation}
{ (a+1)\,\,1\,\,(b+1) \quad \rightarrow \quad a\,b},
\end{equation}
{where $a,b \in \mathbb P$. Then
$$Y(w):=Y(n_1)\ldots Y(n_k)=\varepsilon\,\begin{pmatrix} 1 & 0 \\ 0 & 1\end{pmatrix},\qquad (\varepsilon =\pm 1),$$
if and only if $w\rightarrow^* 1^k$, with $k\equiv 0\ (\mod 6)$ when $\varepsilon =1$, and $k\equiv 3\ (\mod 6)$ when $\varepsilon =-1$. In this case, if $w$ is not a power of 1, then it contains a factor $(a+1)\,\,1\,\,(b+1)$.
}
\end{proposition}
One direction of the proposition easily follows from the identities
\begin{equation}\label{relations1}
{Y(1)^3=-\begin{pmatrix} 1 & 0 \\ 0 & 1\end{pmatrix}},\qquad {\rm and}\qquad {Y(a+1)Y(1)Y(b+1)=Y(a)Y(b)},
\end{equation}
both of which can be easily checked by direct computation.
Moreover, for further use, it is easily checked that
\begin{equation}\label{relation2}
{Y(1)Y(2)Y(1)Y(2)=-1}
\end{equation}
and also, recursively, that
\begin{equation}\label{Yn}
{Y(n)=(-1)^n (Y(2)\,Y(1)^2)^{n-2}\,Y(2)}.
\end{equation}
We now give a proof of Proposition~\ref{rewriting} after recalling some facts regarding presentations of $\SL{2}(\mathbb{Z})$. To simplify our discussion, let us informally\footnote{This can easily be made formally correct by adding a generator, with straightforward relations, to our presentations.} write ``$-1$'' for a central element of $\SL{2}$ whose square is the identity (denoted by $1$).
\begin{lemma}
{Denoting $Y(1)$ by $y_1$, and $Y(2)$ by $y_2$},
\begin{enumerate}
\item[(i)] {$\SL{2}(\mathbb{Z})$ affords the presentation }
\begin{equation}\label{first_pres}
{ \langle\ y_1, y_2\ |\ y_1^3=-1,\quad (y_1y_2)^2=-1\ \rangle.}
\end{equation}
\item[(ii)] {$\SL{2}(\mathbb{Z})$ affords the confluent presentation}
\begin{equation}\label{conf_pres}
{ \langle\ y_1, y_2\ |\ y_1^3\rightarrow -1,\quad y_2y_1y_2\rightarrow y_1^2\ \rangle.}
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}[\bf Proof.]
To show (i), let $$
a=\begin{pmatrix}1&1\\0&1 \end{pmatrix}, b=\begin{pmatrix}1&0\\1&1 \end{pmatrix}.
$$
It is well-known that $\SL{2}(\mathbb{Z})$ has the presentation
\begin{equation}\label{presentationSL2}
{ \langle\ a,b\ |\ \bar a b\bar a=b\bar a b,(\bar a b\bar a)^4=1\rangle}.
\end{equation}
Direct calculations show that
\begin{equation}\label{ab_en_fonction}
a=\bar y_1y_2,\quad b=y_1 \bar y_2.
\end{equation}
Thus $y_1,y_2$ generate $\SL{2}(\mathbb{Z})$. The relations in (i) hold by (\ref{relations1}) and (\ref{relation2}). Therefore, it is enough to show that these relations imply the relations in (\ref{presentationSL2}), once $a,b$ have been replaced using (\ref{ab_en_fonction}).
By direct substitution, we get
$$\bar a b\bar a=\bar y_2 y_1y_1\bar y_2\bar y_2y_1,\quad{\rm and}\quad b\bar a b=y_1\bar y_2\bar y_2y_1y_1\bar y_2.$$
Now, since $y_1^3=-1$, we have $-\bar y_1=y_1^2$, and hence $y_1^2\bar y_2^3(-\bar y_1)=(-\bar y_1)\bar y_2^3y_1^2$. Thus $y_1^2\bar y_2^2(-\bar y_2\bar y_1)=(-\bar y_1\bar y_2)\bar y_2^2y_1^2$. But, since $y_1y_2y_1y_2=-1$, we also have $y_2y_1y_2y_1=-1$, and therefore $-\bar y_2\bar y_1=y_1y_2$ and $-\bar y_1\bar y_2=y_2y_1$. Hence
$$ y_1^2\bar y_2^2y_1y_2=y_2y_1\bar y_2^2y_1^2.$$
Multiplying this both on the left and on the right by $\bar y_2$, we obtain
$$\bar y_2y_1^2\bar y_2^2y_1=y_1\bar y_2^2y_1^2\bar y_2,$$
so that $\bar a b\bar a=b\bar a b$.
On the other hand, we have $\bar a b=\bar y_2y_1^2\bar y_2$, and we have seen that $y_1y_2=-\bar y_2\bar y_1$. Thus
$$y_1=-\bar y_2\bar y_1\bar y_2=-\bar y_2(-y_1^2)\bar y_2=\bar y_2y_1^2 \bar y_2,$$
since $\bar y_1=-y_1^2$. Thus $(\bar y_2y_1^2\bar y_2)^6=1$. It follows, using $\bar a b\bar a=b\bar a b$, that
$$ (\bar a b\bar a)^4=\bar a b\bar a b\bar a b \bar a b\bar a b\bar a b=(\bar a b)^6=(\bar y_2y_1^2\bar y_2)^6=1.$$
(ii) We conclude from the first part that
\begin{equation}
{ \langle\ y_1, y_2\ |\ y_1^3= -1,\quad y_2y_1y_2=y_1^2\ \rangle.}
\end{equation}
is a presentation of $\SL{2}(\mathbb{Z})$. Orienting the equalities, we obtain a rewriting system, whose confluence we must prove. This follows since
the only non-trivial critical pair that needs to be examined is
$$\begin{matrix} & & y_2y_1y_2y_1y_2\\
& \swarrow & & \searrow\\
(y_1^2)\,y_1y_2 && && y_2y_1\, (y_1^2)\\
&\searrow && \swarrow\\
&& -y_2
\end{matrix}$$
This ends our proof.
\end{proof}
\begin{proof}[\bf Proof of Proposition~\ref{rewriting}.]
We need only show that if $w\not=1^n$ is such that $Y(w)=\pm Id$, then $w$ must contain a factor of the form $(a+1)\,\,1\,\,(b+1)$. To see this, formally replace each letter $n$ in ${\mathbb P}$ by $y_n$ in words $w$ in ${\mathbb P}^*$. Then, assuming that $w$ is different from $y_1^n$, we may consider the canonical expansion
$$w= y_1^{n_0}\,y_{m_1}\,y_1^{n_1}\,y_{m_2}\,\cdots y_{m_k}\,y_1^{n_k},$$
where each $m_i\geq 2$ and $k\geq 1$. Using (\ref{Yn}), we replace in this expansion each $y_m$ by $(-1)^m\, (y_2y_1^2)^{m-2} y_2$, we obtain a word in $\{y_1,y_2\}^*$ containing at least one instance of $y_2$. Since the system~(\ref{conf_pres}) is confluent, this word must contain $y_2y_1y_2$, hence one of the $n_i$ must be equal to $1$, thus proving our assertion.
\end{proof}
\begin{corollary}[Conway-Coxeter \cite{conway}]
{For each frieze pattern of the form {\rm (}\ref{frieze}{\rm )}, the bi-infinite sequence of positive integers $\cdots a_{-2}a_{-1}a_0a_1a_2a_3\cdots$ is equal to $\cdots wwwwwww \cdots$ for some quiddity $w$}.
\end{corollary}
\begin{proof}[\bf Proof.]
Denote by $\mathcal{F}$ this frieze pattern, let $n$ be the number of diagonals of $\mathcal{F}$ and denote by ${\mathcal A}$ the $\SL{2}$-tiling obtained through Proposition~\ref{extension}. Then, by the discussion before Corollary~\ref{antiperiod}, the coefficients $a_i$ satisfy
$${Y(a_{1}) Y(a_{2}) \cdots Y(a_{n+1})= \begin{pmatrix} -1 & 0\\
0 & -1\end{pmatrix}}
.$$
Thus, by Corollary~\ref{antiperiod}, $\cdots a_{-2}a_{-1}a_0a_1a_2a_3\cdots=\cdots wwwwwww \cdots$, with $w=a_1\cdots a_{n+1}$. Moreover, by Proposition~\ref{rewriting}, we have
$${a_{i-1}>1,\quad a_i=1},\quad {\rm and}\quad {a_{i+1}>1},$$
for some $i$, $1<i<n+1$. Thus we find in ${\mathcal A}$ the subarray
$$
\begin{array}{ccccccccc}
a_{i-2}&1&0\\
&a_{i-1}&1&0\\
&(a_{i-1}-1)&1&1&0\\
&&(a_{i+1}-1)&a_{i+1}&1\\
&&&&a_{i+2}
\end{array}
$$
Let $C_j$ and $R_j$ denote the columnn and row containing $a_j$. Then, as in the discussion at the beginning of the proof of Proposition~\ref{rewriting}, we have $C_i-C_{i+1}+C_{i+2}=0$ and $R_i-R_{i-1}+R_{i-2}=0$. Thus, by Lemma~\ref{suppression_colonne}, we may suppress both the column $C_{i+1}$ and the row $R_{i-1}$, to get the tame $\SL{2}$-tiling
$$
\begin{array}{cccccccc}
a_{i-2}&1\\
&(a_{i-1}-1)&1&\\
&&(a_{i+1}-1)&1\\
&&&a_{i+2}
\end{array}
$$
We may clearly do this periodically for each column $C_{i+1+p(n+1)}$ and each row $R_{i-1+p(n+1)}$, for $p\in \mathbb{Z}$. Since both ${\mathcal A}$ and $\mathcal{F}$ have the diagonal period $n+1$, by Corollary~\ref{antiperiod}, we obtain a tame $\SL{2}$-tiling ${\mathcal A}'$ and a frieze pattern $\mathcal{F}'$ with $n-1$ diagonals, such that ${\mathcal A}'$ is the unique extension of $\mathcal{F}'$ according to Proposition~\ref{extension}. This proves the corollary, once it is noted that the initial case corresponds to the frieze patterns reduced to $n=2$ diagonals containing only $1$'s (here considered as having a diagonal period equal to $3$).
\end{proof}
\subsection{${\mathbb N}\times {\mathbb N}$ $\SL{k}$-tilings}\label{NbyN}
When we restrict ourselves to ${\mathbb N}\times {\mathbb N}$ arrays, we may apply tools from matrix algebra and generating series.
Assume that, for an invertible $k\times k$ matrix, we have a $(k+h)\times (k+h)$ matrix of rank $k$ that decomposes into blocks in the following manner
\begin{equation}\begin{pmatrix}\label{matrice_bloc}
\mathcal{S} & \Lambda\\
\Gamma & \mathcal{X}
\end{pmatrix},
\end{equation}
with $h$ possibly infinite.
Then we must necessarily have ${\mathcal{X}=\Gamma\, \mathcal{S}^{-1}\,\Lambda}$.
Indeed, it is clear in the following simple matrix identity
$$\begin{pmatrix}
\mathcal{S} & \Lambda\\
\Gamma & \mathcal{X}
\end{pmatrix}\,
\begin{pmatrix}
\mathrm{Id}_k & -\mathcal{S}^{-1}\,\Lambda\\
0 & Id_h
\end{pmatrix}=
\begin{pmatrix}
\mathcal{S} & 0\\
\Gamma & \mathcal{X}-\Gamma\, \mathcal{S}^{-1}\,\Lambda
\end{pmatrix}
$$
that the right-hand side is also a matrix of rank $k$, since we are multiplying our original rank $k$ matrix by an invertible one. However, we already know that $\mathcal{S}$ is of rank $k$. This forces $\mathcal{X}-\Gamma\, \mathcal{S}^{-1}\,\Lambda$ to vanish, and we have the required identity.
Let us assume that ${\mathcal A}$ is a tame \defn{quarter-plane} (of shape ${\mathbb N}\times {\mathbb N}$) $\SL{k}$-array. Choose $\mathcal{S}:={\mathcal A}_{00}^{(k)}$, and let $\Gamma$ (resp. $\Lambda$) stand for the subarray consisting of the first $k$ columns (resp. rows) of ${\mathcal A}$. Then, we deduce from the above identity that we have
\begin{equation}\label{decomposition}
{ {\mathcal A}=\Gamma\, \mathcal{S}^{-1}\,\Lambda},
\end{equation}
whenever ${\mathcal A}$ and $\mathcal{S}$ are both of rank $k$.
It follows that for any subset $I$ (resp. $J$) of rows (resp. columns, with $\#I=\#J$), we have
\begin{equation}
{ {\mathcal A}_{IJ}=\Gamma_I\, \mathcal{S}^{-1}\, \Lambda_J}.
\end{equation}
The simplest possible case of this identity allows the calculation of entries of ${\mathcal A}$ in the form
\begin{equation}\label{entrees_de_A}
{a_{ij}= \Gamma_i\, \mathcal{S}^{-1}\, \Lambda_j}.
\end{equation}
Now, if we choose both $I$ and $J$ to be of cardinality $k$, and take the determinant of both sides, we deduce from the fact that $\det(\mathcal S)=1$, the
identity
\begin{equation}
{{M}_{IJ}={M}_{I,\{1,\ldots,k\}} \, {M}_{\{1,\ldots,k\},J}}.
\end{equation}
A straightforward encoding of the ${\mathbb N}\times {\mathbb N}$-array ${\mathcal A}$ is through its bivariate \defn{generating function}:
\begin{equation}\label{gen_fonct}
{{\mathcal A}(x,y):=\sum_{(i,j)} a_{ij}\, x^i y^j},
\end{equation}
with the sum running over all pairs $(i,j)$ belonging to the shape of ${\mathcal A}$. An equivalent description may be given in terms of matrices, considering $X=\left(x^i\right)_{0\leq i}$ as an infinite one-line matrix, and likewise
$Y=\left(y^j\right)_{0\leq j}$ as an infinite one-column matrix. We then have
${\mathcal A}(x,y)=X\,{\mathcal A}\,Y$.
Now, when ${\mathcal A}$ is tame, it follows from (\ref{decomposition}) that
we have
\begin{eqnarray}
{\mathcal A}(x,y)&=&X\,{\mathcal A}\,Y \nonumber\\
&=& X\, \Gamma\, \mathcal{S}^{-1}\, \Lambda\, Y\nonumber\\
&=& {\begin{pmatrix} C_1(x) & \ldots & C_k(x) \end{pmatrix}
\, \mathcal{S}^{-1}\, \begin{pmatrix} L_1(y) \\ \vdots \\ L_k(y)\end{pmatrix}},\label{gen_fonct_matrice}
\end{eqnarray}
where the $C_i(x)$ are respectively the generating functions of the first $k$ columns of ${\mathcal A}$. Likewise
the $L_j(x)$ are the respective generating functions of the first $k$ rows of ${\mathcal A}$. Often, as below, we have $\mathcal{S}={\rm Id}$.
To illustrate the situation considered above, one may show that the $\SL{k}$-property holds for the matrix of binomial coefficients
$$\Gamma:=\left(\binom{j}{i}\right)_{\monatop{0\leq i,}{0\leq j <k}}.$$
From this, we get a $\SL{k}$-array ${\mathcal A}:=\Gamma\Lambda$, with $\Lambda$ equal to the transpose of $\Gamma$. Observe that
the generating function of the $j^{\rm th}$-column (resp. $i^{\rm th}$) of $\Gamma$ (resp. $\Lambda$) is evidently
$$\frac{1}{(1-x)^j} = \sum_{i\geq 0} \binom{i+j}{i} x^i,\qquad {\rm for}\ j=0,1,\ldots, k-1$$
(resp. $(1-y)^i$). After calculation, using (\ref{gen_fonct_matrice}), we get that the generating function of ${\mathcal A}$ is
\begin{equation}\label{gen_fonct_k}
{{\mathcal A}(x,y)=\sum_{\ell=1}^{k} \frac{x^{\ell-1} y^{\ell-1}}{(1-x)^\ell (1-y)^\ell}}.
\end{equation}
The following result follows from the constructions in Section~\ref{words_tilings}.
\begin{proposition}\label{prop_rectangle}
{The tiling given by {\rm (}\ref{gen_fonct_k}{\rm )} has all minors of the from ${M}_{i0}^{(m)}$ and ${M}_{0j}^{(m)}$ equal to $1$, whenever $m< k$, and it is a $\SL{k}$-tiling.}
\end{proposition}
Using (\ref{entrees_de_A}), or directly from (\ref{gen_fonct_k}), one may calculate that the individual entries of ${\mathcal A}$ are given by the formula
\begin{equation}\label{coeff_rect}
{a_{ij}=\sum_{\ell=0}^{k-1} \binom{i}{\ell}\binom{j}{\ell}}.
\end{equation}
It follows also from Section~\ref{words_tilings}, that for $(i,j)$ in the $k$-fringe, $a_{ij}=\binom{i+j}{i}$ .
For example, with $k=3$, we get the array of Figure~\ref{fig_quarter}.
\begin{figure}[ht]
$$
\begin{picture}(0,0)(0,0)\setlength{\unitlength}{5mm}
\put(1.7,3.5){\put(0,0){\vector(1,0){17}}
\put(0,0){\vector(0,-1){6.5}}}
\end{picture}
{\mathcal A}= \begin {array}{cccccccccc}
1&1&1&1&1&1&1&1&1&\cdots\\
1&2&3&4&5&6&7&8&9&\cdots\\
1&3&6&10&15&21&28&36&45&\cdots\\
1&4&10&19&31&46&64&85&109&\cdots\\
1&5&15&31&53&81&115&155&201&\cdots\\
\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\ddots
\end {array}
$$
\caption{A $\SL{3}$-tiling of ${\mathbb N}\times {\mathbb N}$.} \label{fig_quarter}
\end{figure}
It may readily be shown that the dual of ${\mathcal A}$ affords the generating function
\begin{equation}\label{gen_fonct_derk}
{{\mathcal A}^*(x,y)=\frac{1}{(1-x)(1-y)}+\sum_{\ell=2}^k \frac{x\,y}{(1-x)^\ell (1-y)^\ell}}.
\end{equation}
\subsection{Zigzag path}
It is shown in \cite{assem} that the $\SL{2}$-tiling associated to the bi-infinite word $\cdots xyxyxyxyx \cdots$ has entries equal to the Fibonacci numbers of even rank (if we set $F_{n+2}=F_{n+1}+F_n$, $F_0=F_1=1$), see the left part of Figure~\ref{fig_zigzag}. If we let $k$ go to infinity, then by Section~\ref{words_tilings}, the entries of the resulting tiling are the Catalan numbers. In particular, it is noteworthy that the value of the $(k\times k)$-principal minors given by Proposition~\ref{prop_word_tiling} corresponds in this situation to the classical result stating that for any natural integer $k$, the Hankel matrix $(C_{h+i+j})_{i,j=0,\ldots, k}$ (with either $h=0$, or $h=1$) has determinant equal to $1$ (here, as usual, we have $C_n=\frac{1}{n+1}\binom{2n}{n}$), see the right part of Figure~\ref{fig_zigzag}.
\begin{figure}[ht]
$$
\begin{array}{c}
\begin{array}{ccccccccccccccccccccccccccccccccccccccccccccccc}
&&&&&&&&\revddots&&&&&&&&&&&&\revddots\\
&&&&&&1&1&&&&&&&&&&&1&1\\
&&&&&1&1&2&&&&&&&&&&1&1&2\\
&&&&1&1&2&5&&&&&&&&&1&1&2&5\\
&&&1&1&2&5&13&&&&&&&&1&1&2&5&14\\
&&1&1&2&5&13&34&&&&&&&1&1&2&5&14&42\\
&1&1&2&5&13&34&89&&&&&&1&1&2&5&14&42&132\\
\revddots&&&&&&&&&&&\revddots
\end{array}\\
k=2 \hskip6cm k\rightarrow \infty
\end{array}
$$\vskip-10pt
\caption{Zigzag path tiling}\label{fig_zigzag}
\end{figure}
\section{Closing remarks}\label{sec_closing}
\subsection*{A converse}
Experiments suggest that a ``converse'' of Proposition~\ref{derive_equation} holds, namely that for any tame tiling, if {\rm (}\ref{condensation_devive}{\rm )} holds for some pair $(r,s)$ for which $r+s=k$, then the tiling is necessarily a $\SL{k}$-tiling. Special cases, for small values of $k$, are easy to prove using generic value tilings and Gr\"obner basis computations.
\subsection*{Generalized frieze patterns}
A notion of generalized frieze patterns, for $k>2$, has been considered in \cite{cordes}. These are best understood in terms of certain tame ``toric'' $\SL{k}$-tilings $\mathcal{A}$.
More precisely, we say that a tiling has a \defn{skew-period} $(p,q)$ in $\mathbb{Z}\times \mathbb{Z}$ ($\not=(0,0)$), if and only if
\begin{equation}\label{defn_skew_periodic}
\mathcal{A}(i+p,j+q) =(-1)^k\mathcal{A}(i,j),\qquad {\rm for\ all}\ (i,j)\in\mathbb{Z}\times \mathbb{Z},
\end{equation}
and we then say that the tiling is \defn{skew-periodic}. A \defn{toric} tiling is one that has two linearly independent skew-periods. A \defn{$\SL{k}$-frieze patterns} $\mathcal{A}$ is a tame $\SL{k}$-tiling such that
\begin{equation}\label{defn_cond_frise}
\mathcal{A}(i,j)= \begin{cases}
1, & \text{if}\ i=j,\\
0, & \text{if}\ i-j<k,\\
(-1)^{k-1}, & \text{if}\ i=j+k,
\end{cases}
\end{equation}
which is periodic (not skew) with a period of the form $(p,-p)$, for $p>k$.
In other words, on top of being periodic, the tiling is prescribed to have a diagonal of $1$'s, another diagonal filled with $(-1)^{k-1}$ below, with these two diagonals separated by $(k-1)$ diagonals of $0$'s.
Condition (\ref{defn_cond_frise}) and periodicity (together with tameness) ensure that the whole tiling is determined by its values along a band $\{(i,j)\ j\leq i< j+p,\ j\in\mathbb{Z}\}$, with $p$ as above. Moreover, one may show that any such tiling is toric, with skew-periods $(p,0)$ and $(0,p)$. This implies that it exhibits a frieze-like behaviour, since the tiling must necessarily have period $(p,p)$. The generalized frieze patterns of \cite{cordes} appear as special cases of this notion.
It is interesting to observe (and easy to prove) that the $\SL{k}$-frieze patterns, with $p=k+2$, that are of the form illustrated in Figure~\ref{fig_slk_frises}, include those for which the sequence $\ldots,a_{-1},a_0,a_1,\ldots a_i\,\ldots$ is a quiddity. Indeed, applying the $\SL{k}$-condition to the submatrix with main diagonal corresponding to $a_i$'s, one easily check that we have
$$q_k(a_i,a_{i+1},\ldots,a_{i+k-1})=1,$$
with $q_k$ denoting the signed continuant polynomials considered in Section~\ref{revisit}.
\begin{figure}[ht]
\begin{equation}{
\begin{array}{ccccccccccccccccccc}
\ddots &\ddots&\ddots&\ddots&\ddots\\
&0&0&1& a_{-1} & 1 \\
&&0&0&1&a_{0} & 1 \\
&&&0& 0&1&a_{1} & 1 \\
&&&&\ddots& \ddots&\ddots&\ddots & \ddots \\
&&&& &0&0& 1& a_i& 1 \\
&&&&&&\ddots&\ddots&\ddots&\ddots&\ddots
\end{array}
}\end{equation}
\caption{Positive integer $\SL{k}$-frize patterns of ``width'' $1$.}\label{fig_slk_frises}
\end{figure}
Other interesting toric $\SL{k}$-tilings seem to abound. For example, with $k=3$ and $p=4$, we have the following positive integer valued toric $\SL{k}$-tilings:
$$\mathcal{A}=\begin {array}{cccccccccccccc}
\ddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\revddots\\ \noalign{\medskip}
\cdots&1&1&1&1&1&1&1&1&1&1&1&\cdots\\ \noalign{\medskip}
\cdots&1&2&3&2&1&2&3&2&1&2&3&\cdots\\ \noalign{\medskip}
\cdots&1&3&6&4&1&3&6&4&1&3&6&\cdots\\ \noalign{\medskip}
\cdots&1&2&4&3&1&2&4&3&1&2&4&\cdots\\ \noalign{\medskip}
\cdots&1&1&1&1&1&1&1&1&1&1&1&\cdots\\ \noalign{\medskip}
\cdots&1&2&3&2&1&2&3&2&1&2&3&\cdots\\ \noalign{\medskip}
\cdots&1&3&6&4&1&3&6&4&1&3&6&\cdots\\ \noalign{\medskip}
\revddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\ddots\end {array}
\begin{picture}(0,0)(0,0)
\textcolor{red}{\put(-195,-4){\framebox(60,72){}}}
\end{picture}
$$
It may be checked that this is a tame tiling,
however the entries of the corresponding dual tiling are not all positive, since:
$$\partial\mathcal{A}=\begin {array}{rrrrrrrrrrrrrrrrrrrrrrrr}
\ddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\revddots\\ \noalign{\medskip}
\cdots&1&1&-1&-1&1&1&-1&-1&1&1&\cdots\\ \noalign{\medskip}
\cdots&1&3&0&-2&1&3&0&-2&1&3&\cdots\\ \noalign{\medskip}
\cdots&-1&0&2&1&-1&0&2&1&-1&0&\cdots\\ \noalign{\medskip}
\cdots&-1&-2&1&2&-1&-2&1&2&-1&-2&\cdots\\ \noalign{\medskip}
\cdots&1&1&-1&-1&1&1&-1&-1&1&1&\cdots\\ \noalign{\medskip}
\cdots&1&3&0&-2&1&3&0&-2&1&3&\cdots\\ \noalign{\medskip}
\revddots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\vdots&\ddots
\end {array}$$
General properties of tame toric tilings, as well as results concerning $\SL{k}$-frieze patterns similar to those of Subsection~\ref{revisit}, will be the subject of a planed sequel to this paper.
\subsection*{T-systems}
On a closing note, it is interesting to observe that there is a close tie between tame $\SL{k}$-tilings and the notion of $T$-systems, which appear as solutions of the discrete Hirota equation (see \cite{di_Francesco}) of mathematical physics. Indeed, up to a simple relabelling, one can characterize the entries of $T$-systems in terms of derivatives of suitably chosen tame $\SL{k}$-tilings. Recall that a $T$-system $T:\{0,\ldots, r+1\}\times\mathbb{Z}\times \mathbb{Z}\rightarrow {\mathbb N}$ must satisfy the equation
\begin{equation}\label{T_def}
T_{\{\alpha,j,k+1\}}T_{\{\alpha,j,k-1\}}=T_{\{\alpha,j+1,k\}}T_{\{\alpha,j-1,k\}}+T_{\{\alpha+1,j,k+1\}}T_{\{\alpha-1,j,k+1\}},
\end{equation}
with boundary conditions
\begin{equation}\label{T_cond}
T_{\{0,j,k\}}=T_{\{r+1,j,k\}}=1,
\end{equation}
for all $j$, and $k$ in $\mathbb{Z}$. It is shown in \cite{di_Francesco} that
\begin{equation}\label{T_formule}
T_{\{\alpha,j,k\}}= \det \begin{pmatrix}
T_{\{1,j-a+b,k+a+b-\alpha-1\}}
\end{pmatrix} _{1\leq a,b\leq \alpha}.
\end{equation}
From this, one can readily see that
$$T_{\alpha,j,k}=(\partial^\alpha \mathcal{A})_{s,t}$$
for $s$, $t$ simple linear functions of $j$ and $k$, and $\mathcal{A}$ a $\SL{r+1}$-tiling directly obtained from $(T_{1,j,k})_{j,k}$.
|
\section{Bulk duality}
\label{secbulk}
\setcounter{equation}{0}
This section presents a brief review of the bulk non-Abelian duality \cite{SYdual}
and introduces all relevant notation (which is also summarized in \cite{SYrev}).
The bulk theory is ${\mathcal N}=2\;$ SQCD with the U($N$) gauge group and
$N_f$ flavors of fundamental quark hypermultiplets ($N<N_f<2N$).
\subsection{Bulk theory at large $\xi$}
The field content is as follows. The ${\mathcal N}=2\;$ vector multiplet
consists of the U(1)
gauge field $A_{\mu}$ and the SU$(N)$ gauge field $A^a_{\mu}$,
where $a=1,..., N^2-1$, and their Weyl fermion superpartners
plus
complex scalar fields $a$, and $a^a$ and their Weyl superpartners.
The $N_f$ quark multiplets of the U$(N)$ theory consist
of the complex scalar fields
$q^{kA}$ and $\tilde{q}_{Ak}$ (squarks) and
their fermion superpartners, all in the fundamental representation of
the SU$(N)$ gauge group.
Here $k=1,..., N$ is the color index
while $A$ is the flavor index, $A=1,..., N_f$. We will treat $q^{kA}$
as a rectangular matrix with $N$ rows and $N_f$ columns.
This theory is endowed with
the FI term $\xi$ which singles out the vacuum in which $r=N$ squarks condense.
Consider, say, the (1,2,...,$N$) vacuum in which the first $N$ flavors develop
vacuum expectation values
(VEVs),
\begin{eqnarray}
\langle q^{kA}\rangle &=&\sqrt{
\xi}\,
\left(
\begin{array}{cccccc}
1 & \ldots & 0 & 0 & \ldots & 0\\
\ldots & \ldots & \ldots & \ldots & \ldots & \ldots\\
0 & \ldots & 1 & 0 & \ldots & 0\\
\end{array}
\right),
\qquad \langle \bar{\tilde{q}}^{kA}\rangle =0,
\nonumber\\[4mm]
k&=&1,..., N\,,\qquad A=1,...,N_f\, .
\label{qvev}
\end{eqnarray}
In this vacuum the
adjoint fields also develop
VEVs, namely,
\begin{equation}
\left\langle \left(\frac12\, a + T^a\, a^a\right)\right\rangle = - \frac1{\sqrt{2}}
\left(
\begin{array}{ccc}
m_1 & \ldots & 0 \\
\ldots & \ldots & \ldots\\
0 & \ldots & m_N\\
\end{array}
\right),
\label{avev}
\end{equation}
where $m_A$ are quark mass parameters.
For generic values of $m_A$'s, the VEVs (\ref{avev}) break the SU$(N)$ subgroup of the gauge
group down to U(1)$^{N-1}$. However, in the special limit
\begin{equation}
m_1=m_2=...=m_{N_f},
\label{equalmasses}
\end{equation}
the SU$(N)\times$U(1) gauge group remains unbroken by the adjoint field.
In this limit the theory acquires a global flavor SU$(N_f)$ symmetry.
While the adjoint VEVs do not break the SU$(N)\times$U(1) gauge group in the limit
(\ref{equalmasses}),
the quark condensate (\ref{qvev}) results in the spontaneous
breaking of both gauge and flavor symmetries.
A diagonal global SU$(N)$ combining the gauge SU$(N)$ and an
SU$(N)$ subgroup (which rotates first $N$ quarks) of the flavor SU$(N_f)$
group survives, however. Below we will refer to this diagonal
global symmetry as to $ {\rm SU}(N)_{C+F}$.
More exactly, the pattern of breaking of the
color and flavor symmetry
is as follows:
\begin{equation}
{\rm U}(N)_{\rm gauge}\times {\rm SU}(N_f)_{\rm flavor}\to
{\rm SU}(N)_{C+F}\times {\rm SU}(\tilde{N})_F\times {\rm U}(1)\,,
\label{c+f}
\end{equation}
where $\tilde{N}=N_f-N$.
The phenomenon of color-flavor locking takes place in the vacuum.
The global SU$(N)_{C+F}$ group is responsible for
formation of the non-Abelian strings (see below).
For unequal quark masses in (\ref{avev}) the global symmetry (\ref{c+f}) is broken down to
U(1)$^{N_f-1}$.
Since the global (flavor) SU$(N_f)$ group is broken by the quark VEVs anyway we can consider
the following mass splitting:
\begin{equation}
m_P=m_{P'}, \qquad m_K=m_{K'}, \qquad m_P-m_K=\Delta m
\label{masssplit}
\end{equation}
where $P,P'=1, ..., N$ and $K,K'=N+1, ..., N_f$.\footnote{A generic mass difference $m_A-m_B$
(for all $A,B = 1,2, ..., N_f$)
will be referred to as $\Delta m_{AB}$ below,
while $\Delta m$ is reserved for $m_P-m_K$, ($P=1,2, ..., N$, $K=N+1, ..., N_f$).
In Ref.~\cite{SYcrossp}
the mass differences inside the first group
(or inside the second group) were called $\Delta M_{\rm inside}$.
The mass differences $m_P-m_K$ were referred to as
$\Delta M_{\rm outside}$.}
This mass splitting respects the global
group (\ref{c+f}) in the $(1,2,...,N)$ vacuum. This vacuum then becomes isolated.
No Higgs branch develops. We will often use this limit below.
Now let us discuss the mass spectrum in our theory. Since
both U(1) and SU($N$) gauge groups are broken by squark condensation, all
gauge bosons become massive. In fact, at nonvanishing $\xi$, both the quarks and adjoint scalars
combine with the gauge bosons to form long ${\mathcal N}=2\;$ supermultiplets \cite{VY}, for a review see
\cite{SYrev}.
Note that all states come in representations of the unbroken global
group (\ref{c+f}), namely, the singlet and adjoint representations
of SU$(N)_{C+F}$
\begin{equation}
(1,\, 1), \quad (N^2-1,\, 1),
\label{onep}
\end{equation}
and in the bifundamental representations
\begin{equation}
\quad (\bar{N},\, \tilde N), \quad
(N,\, \bar{\tilde N})\,,
\label{twop}
\end{equation}
where in (\ref{onep}) and (\ref{twop}) we mark representation with respect to two
non-Abelian factors in (\ref{c+f}). The singlet and adjoint fields are the gauge bosons,
and the first $N$ flavors of the squarks $q^{kP}$ ($P=1,...,N$), together with their fermion superpartners.
The bifundamental fields are the quarks $q^{kK}$ with $K=N+1,...,N_f$.
These quarks transform in the two-index representations of the global
group (\ref{c+f}) due to the color-flavor locking.
At large $\xi$ this theory is at weak coupling. Namely, the condition
\begin{equation}
\xi\gg \Lambda\, ,
\label{weakcoupling}
\end{equation}
ensures weak coupling in the SU$(N)$ sector because
the SU$(N)$ gauge coupling does not run below the scale of the quark VEVs
which is determined by $\sqrt\xi$. Here $\Lambda$ is the dynamical scale of the SU($N$) gauge theory. More explicitly,
\begin{equation}
\frac{8\pi^2}{g^2_2 (\xi)} =
(N-\tilde N )\ln{\frac{g_2\sqrt{\xi}}{\Lambda}}\gg 1 \,, \rule{8mm}{0mm}
\label{4coupling}
\end{equation}
where $g_2^2$ is the coupling constant of the SU$(N)$ sector.
\subsection{Duality}
\label{secbulkduality}
As was shown in \cite{SYdual}, at $\sqrt\xi \sim \Lambda$ the theory goes through a crossover transition
to the strong coupling regime. At small $\xi$ ($\sqrt\xi \ll \Lambda$) this regime can be described
in terms of weakly coupled dual ${\mathcal N}=2\;$ SQCD, with the gauge group
\begin{equation}
{\rm U}(\tilde N)\times {\rm U}(1)^{N-\tilde N}\,,
\label{dualgaugegroup}
\end{equation}
and $N_f$ flavors of
light {\em dyons}.
This non-Abelian ${\mathcal N}=2\;$ duality is similar to Seiberg's duality in ${\mathcal N}=1\;$ supersymmetric QCD
\cite{Sdual,IS}. Later a dual non-Abelian gauge group SU$(\tilde N)$ was identified on the Coulomb branch
at the root of a baryonic Higgs branch in the ${\mathcal N}=2\;$ supersymmetric SU($N$) gauge theory with massless quarks
\cite{APS}.
Light dyons are in the fundamental representation of the gauge group
U$(\tilde N)$ and are charged under Abelian factors in (\ref{dualgaugegroup}). In addition, there are
light dyons $D^l$ ($l=\tilde N+1, ..., N$) neutral under
the U$(\tilde N)$ group, but charged under the
U(1) factors. A small but nonvanishing $\xi$ triggers condensation of all these
dyons,
\begin{eqnarray}
\langle D^{lA}\rangle \! \! &=&\!\!\sqrt{
\xi}\,
\left(
\begin{array}{cccccc}
0 & \ldots & 0 & 1 & \ldots & 0\\
\ldots & \ldots & \ldots & \ldots & \ldots & \ldots\\
0 & \ldots & 0 & 0 & \ldots & 1\\
\end{array}
\right),
\quad \langle \bar{\tilde{D}}^{lA}\rangle =0,\quad l=1,...,\tilde N,
\nonumber\\[4mm]
\langle D^{l}\rangle &=& \sqrt{\xi}, \qquad \langle\bar{\tilde{D}}^{l}\rangle =0\,,
\qquad l=\tilde N +1, ..., N\,.
\label{Dvev}
\end{eqnarray}
Now, consider either equal quark masses or the mass choice (\ref{masssplit}).
Both, the gauge and flavor SU($N_f$) groups, are
broken in the vacuum. However, the color-flavor locked form of (\ref{Dvev}) guarantees that the diagonal
global SU($\tilde N)_{C+F}$ survives. More exactly, the unbroken global group of the dual
theory is
\begin{equation}
{\rm SU}(N)_F\times {\rm SU}(\tilde N)_{C+F}\times {\rm U}(1)\,.
\label{c+fd}
\end{equation}
Here SU$(\tilde N)_{C+F}$ is a global unbroken color-flavor rotation, which involves the
last $\tilde N$ flavors, while SU$(N)_F$ factor stands for the flavor rotation of the
first $N$ dyons.
Thus, a color-flavor locking takes place in the dual theory too. Much in the same way as
in the original theory, the presence of the global SU$(\tilde N)_{C+F}$ group
is the reason behind formation of the non-Abelian strings.
For generic quark masses the global symmetry (\ref{c+f}) is broken down to
U(1)$^{N_f-1}$.
In the equal mass limit or for the mass choice (\ref{masssplit})
the global unbroken symmetry (\ref{c+fd}) of the dual theory at small
$\xi$ coincides with the global group (\ref{c+f}) present in the
$r=N$ vacuum of the original theory at large
$\xi$. Note
however, that this global symmetry is realized in two distinct ways in two dual theories.
As was already mentioned, the quarks and U($N$) gauge bosons of the original theory at large $\xi$
come in the $(1,1)$, $(N^2-1,1)$, $(\bar{N},\tilde N)$, and $(N,\bar{\tilde N})$
representations of the global group (\ref{c+f}), while the dyons and U($\tilde N$) gauge
bosons form $(1,1)$, $(1,\tilde N^2-1)$, $(N,\bar{\tilde N})$, and
$(\bar{N},\tilde N)$ representations of (\ref{c+fd}). We see that the
adjoint representations of the $(C+F)$
subgroup are different in two theories. A similar phenomenon was detected in \cite{SYcross}
for the Abelian dual theory (i.e. $\tilde N=0$).
This means that
the quarks and gauge bosons
which form the adjoint $(N^2-1)$ representation
of SU($N$) at large $\xi$ and the dyons and gauge bosons which form the adjoint $(\tilde N^2-1)$ representation of SU($\tilde N$) at small $\xi$ are, in fact, {\em distinct} states.
The $(N^2-1)$ adjoints of SU($N$) become heavy
and decouple as we pass from large to small $\xi$
along the line (\ref{masssplit}). Moreover, some
composite $(\tilde N^2-1)$ adjoints of SU($\tilde N$), which are
heavy and invisible in the low-energy description at large $\xi$ become light
at small $\xi$ and form the $D^{lK}$ dyons
($K=N+1,...,N_f$) and gauge bosons of U$(\tilde N)$. The phenomenon of level crossing
takes place (Fig.~\ref{figevol}). Although this crossover is smooth in the full theory,
from the standpoint of the low-energy description the passage from large to small $\xi$ means a dramatic change: the low-energy theories in these domains are
completely
different; in particular, the degrees of freedom in these theories are different.
This logic leads us to the following conclusion. In addition to light dyons and gauge bosons
included in the low-energy theory at small $\xi$ we have
heavy fields which form the adjoint representation
$(N^2-1,1)$ of the global symmetry (\ref{c+fd}). These are screened quarks
and gauge bosons from the large $\xi$ domain.
Let us denote them as $M_P^{P'}$ ($P,P'=1,...,N$).
As was already explained in Sec.~\ref{intro},
at small $\xi$ they decay into the monopole-antimonopole
pairs on the curves of marginal stability (CMS). \footnote{Strictly speaking,
such pairs
can be formed by monopole-antidyons and
dyon-antidyons as well,
the dyons carrying root-like electric charges. In this paper we will call all these states
``monopoles". This is to avoid confusion with dyons which appear in Eq.~(\ref{Dvev}). The
latter dyons carry weight-like electric charges and, roughly speaking, behave as
quarks, see \cite{SYdual} for further details.}
This is in accordance with results obtained
for ${\mathcal N}=2\;$ SU(2) gauge theories \cite{SW1,SW2,BF} on the Coulomb branch at zero $\xi$
(we confirm this result for the theory at hand in Sec.\ref{secmoral}).
The general rule is that the only states which exist at strong coupling inside CMS are those which can become massless on the Coulomb branch
\cite{SW1,SW2,BF}. For our theory these are light dyons shown in Eq.~(\ref{Dvev}),
gauge bosons of the dual U$(\tilde N)$ theory and monopoles.
At small nonvanishing $\xi$ the
monopoles and antimonopoles produced in the decay process of adjoints $(N^2-1,1)$
cannot escape from
each other and fly off to separate because they are confined. Therefore, the quarks or gauge bosons in the
strong coupling domain of small $\xi$ evolve into stringy mesons $M_P^{P'}$ ($P,P'=1, ..., N$)
-- the monopole-antimonopole
pairs connected by two strings \cite{SYdual} as shown in Fig.~\ref{figmeson}.
By the same token, at large $\xi$, in addition to the light quarks and gauge bosons,
we have heavy fields $M_K^{K'}$ ($K,K'=N+1, ..., N_f$), which form the adjoint $(\tilde N^2-1)$ representation of SU($\tilde N$).
This is schematically depicted in Fig.~\ref{figevol}.
The $M_K^{K'}$ states are (screened) light
dyons and gauge bosons of the dual theory. At large $\xi$ they decay into
monopole-antimonopole
pairs and form stringy mesons \cite{SYdual} shown in Fig.~\ref{figmeson}.
\begin{figure}
\epsfxsize=7cm
\centerline{\epsfbox{evol.eps}}
\caption{\small Evolution of the SU$(N)$ and SU$(\tilde N)$ gauge bosons and light quarks (dyons) vs. $\xi$.
On both sides of the level crossing at $\xi=\Lambda^2$ the global groups are SU$(N)\times$SU$(\tilde N)$,
however, above $\Lambda^2$ it is SU$(N)_{C+F}\times$SU$(\tilde N)_F$ while
below $\Lambda^2$ it is SU$(N)_F\times$SU$(\tilde N)_{C+F}$.}
\label{figevol}
\end{figure}
In \cite{SYdual} we also conjectured that the fields $M_P^{P'}$ and $M_K^{K'}$ are Seiberg's meson fields
\cite{Sdual,IS},
which occur in the dual theory upon breaking of ${\mathcal N}=2\;$ supersymmetry by the mass-term
superpotential $\mu[A^2 +(A^a)^2]$ for the adjoint fields in the limit
$\mu\to\infty$. In this limit our theory becomes ${\mathcal N}=1\;$ SQCD.
We see that the picture of the non-Abelian confinement obtained in \cite{SYdual} is based on the presence of
extra stringy meson states -- the
monopole-antimonopole pairs -- bound by confining strings both in the weak and strong coupling
domains of the bulk theory. These meson states fill representations $(N^2-1,1)$
and $(1,\tilde N^2-1)$ of the global unbroken group at small and large $\xi$,
respectively. Below we confirm the presence of these stringy mesons by
studying the global quantum numbers of confined monopoles in both domains. To this end
we explore kinks in the ${\mathcal N}=(2,2)\;$ supersymmetric weighted CP model on the string world sheet.
We remind that the confined monopoles of the bulk theory are presented by the junctions of
two elementary non-Abelian strings \cite{T,SYmon,HT2}. These elementary
strings corresponds to different
vacua of the effective sigma model on the world sheet.
See also
the review paper \cite{SYrev} for details.
\section{World sheet theory}
\label{secWCP}
\setcounter{equation}{0}
In this section we briefly describe the world-sheet low-energy sigma models on the non-Abelian strings
at large and small $\xi$. Non-Abelian strings in ${\mathcal N}=2\;$ SQCD with $N_f=N$
where first found and studied in \cite{HT1,ABEKY,SYmon,HT2}. Then we discuss
how the bulk duality translates into the world-sheet duality~\cite{SYdual}.
\subsection {World sheet theory at large \boldmath{$\xi$}}
\label{strings}
To warm up, we start from $N_f=N$.
The Abelian $Z_N$-string solutions break the SU$(N)_{C+F}$ global group. As a result,
the non-Abelian strings have orientational zero modes associated with rotations of their color
flux inside the non-Abelian SU($N$ group.
The global group is broken on the $Z_N$ string solution down to
${\rm SU}(N-1)\times {\rm U}(1)$.
Hence,
the moduli space of the non-Abelian string is described by the coset space
\begin{equation}
\frac{{\rm SU}(N)}{{\rm SU}(N-1)\times {\rm U}(1)}\sim CP(N-1)\,,
\label{modulispace}
\end{equation}
in addition to $C$ spanned by the translational modes. The translational moduli totally decouple.
They are sterile free fields, and we can forget about them.
Therefore, the
low-energy effective theory on the non-Abelian string
is the two-dimensional ${\mathcal N}=2\;$ CP$(N-1)$ model \cite{HT1,ABEKY,SYmon,HT2}.
Now we add ``extra" quark flavors with degenerate masses, increasing $N_f$ from $N$ up to a certain value
$N_f >N$.
The strings emerging in such theory are semilocal.
In particular, the string solutions on the Higgs branches (typical
for multiflavor theories) usually are not fixed-radius strings, but, rather,
possess radial moduli, aka size moduli, see \cite{AchVas} for a comprehensive review of
the Abelian semilocal strings. The transverse size of such a string is not fixed.
Non-Abelian semilocal strings in ${\mathcal N}=2\;$ SQCD with $N_f>N$ were studied in
\cite{HT1,HT2,SYsem,Jsem}.
The orientational
zero modes of the semilocal non-Abelian string are parametrized by a complex vector $n^P$ ($P=1, ..., N$),
while its $\tilde N=(N_f-N)$ size moduli are parametrized by a complex vector
$\rho^K$ ($K=N+1, ..., N_f$). The effective two-dimensional theory
which describes the internal dynamics of the non-Abelian semilocal string is
the ${\mathcal N}=(2,2)\;$ weighted CP model on a ``toric" manifold, which includes both types of fields. The bosonic
part of the action
in the gauged formulation (which assumes taking the limit $e^2\to\infty$)
has the form\,\footnote{Equation (\ref{wcp}) and similar expressions below are given in Euclidean notation.}
\begin{eqnarray}
&&S = \int d^2 x \left\{
\left|\nabla_{\alpha} n^{P}\right|^2
+\left|\tilde{\nabla}_{\alpha} \rho^K\right|^2
+\frac1{4e^2}F^2_{\alpha\beta} + \frac1{e^2}\,
\left|\partial_{\alpha}\sigma\right|^2
\right.
\nonumber\\[3mm]
&+&\left.
2\left|\sigma+\frac{m_P}{\sqrt{2}}\right|^2 \left|n^{P}\right|^2
+ 2\left|\sigma+\frac{m_{K}}{\sqrt{2}}\right|^2\left|\rho^K\right|^2
+ \frac{e^2}{2} \left(|n^{P}|^2-|\rho^K|^2 -2\beta\right)^2
\right\},
\nonumber\\[4mm]
&&
P=1,...,N\,,\qquad K=N+1,...,N_f\,.
\label{wcp}
\end{eqnarray}
The fields $n^{P}$ and $\rho^K$ have
charges +1 and $-1$ with respect to the auxiliary U(1) gauge field;
hence, the corresponding covariant derivatives in (\ref{wcp}) are defined as
\begin{equation}
\nabla_{\alpha}=\partial_{\alpha}-iA_{\alpha}\,,\qquad
\tilde{\nabla}_{\alpha}=\partial_{\alpha}+iA_{\alpha}\,,
\label{covarder}
\end{equation}
respectively.
If only charge $+1$ fields were present, in the limit
$e^2\to\infty$ we would get a conventional twisted-mass deformed
CP $(N-1)$ model.
The presence of charge $-1$ fields $\rho^K$ converts the CP$(N-1)$
target space into that of the a weighted
CP$(N_f-1)$ model.
As in the CP$(N-1)$ model, small mass differences
$\left| m_A-m_B\right|$ lift orientational and size zero modes generating a shallow potential on the modular space.
The $D$-term condition
\begin{equation}
|n^P|^2 - |\rho^K|^2=2\beta
\label{unitvec}
\end{equation}
is implemented in the limit $e^2\to\infty$. Moreover, in this limit
the gauge field $A_{\alpha}$ and its ${\mathcal N}=2\;$ bosonic superpartner $\sigma$ become
auxiliary and can be eliminated.
The two-dimensional coupling constant $\beta$ is related to the four-dimensional
one as
\begin{equation}
\beta= \frac{2\pi}{g_2^2}\,.
\label{betag}
\end{equation}
This relation is obtained at the classical level
\cite{ABEKY,SYmon}.
In quantum theory
both couplings run. In particular, the model (\ref{wcp}) is asymptotically free
\cite{W93} and develops its own scale $\Lambda_{\sigma}$.
The ultraviolet cut-off in the sigma model on the string world sheet
is determined by $g_2\sqrt{\xi}$.
Equation~(\ref{betag}) relating the two- and four-dimensional couplings
is valid at this scale.
At this scale the four-dimensional coupling is given by (\ref{4coupling})
while the two-dimensional one
\begin{equation}
4\pi\beta(\xi) =
\left(N-\tilde N \right)\ln\,{\frac{g_2\sqrt{\xi}}{\Lambda_\sigma}}\gg 1 \,.
\label{2coupling}
\end{equation}
Then Eq.~(\ref{betag}) implies
\begin{equation}
\Lambda_{\sigma} = \Lambda\, ,
\label{lambdasig}
\end{equation}
and from now on we will omit the subscript $\sigma$.
In the bulk, the running of the coupling constant is frozen at
$g_2\sqrt{\xi}$, because of the VEVs of
the squark fields.
The logarithmic evolution of the coupling constant in the string world-sheet theory
continues uninterrupted below this point, with the same running law.
As a result, the dynamical scales of the bulk and world-sheet
theories turn out to be the same, much in the same way as in the $N_f=N$ theory \cite{SYmon}.
We remind that the scale $g_2\sqrt{\xi}$ determines the ultraviolet cut-off
in the sigma model on the
string world sheet. Therefore we can consider the theory (\ref{wcp}) as an effective
theory on the string only at energies below $g_2\sqrt{\xi}$. This is fulfilled if
\begin{equation}
\sqrt{\xi}\gg {\rm max}\left(|m_A-m_B|,\Lambda\right).
\label{cpcond}
\end{equation}
Summarizing,
if the quark mass differences are small, the
$(1 , ... ,\, N)$ vacuum of the original U$(N)$ gauge theory supports non-Abelian
semilocal strings. Their internal dynamics is
described by the effective two-dimensional low-energy ${\mathcal N}=(2,2)\;$ sigma model (\ref{wcp}).
The model has $N$ orientational moduli $n^P$ with the U(1) charge $+1$ and masses $m_P=\{m_1,...,m_N\}$,
plus
$\tilde N$ size moduli $\rho^K$, with the U(1) charge $-1$ and masses $(-m_K)=-\{m_{N+1},...,m_{N_f}\}$.
A final remark is in order here. The
strict semilocalality achieved at $\Delta m_{AB}=0$
destroys confinement of monopoles
\cite{EY,SYsem}. The reason is that the string transverse size (determined by $\rho^K$'s )
can grow indefinitely.
When it becomes of the order of the distance $L$ between sources of the magnetic flux (the monopoles),
the linearly rising confining potential between these sources is replaced by a Coulomb-like
potential. To have confinement of monopoles we should
lift the size zero modes keeping $\Delta m$ nonvanishing.
That's exactly what we will do, eventually sticking to (\ref{masssplit}),
preserving both confinement and the global symmetry.
\subsection{Dual world-sheet theory}
\label{dwsth}
The dual bulk U$(\tilde N)$ theory at small $\xi$ also supports
non-Abelian semilocal strings. The $(1,...,N)$ vacuum of the original theory
(\ref{qvev}) transforms into the vacuum (\ref{Dvev}) of the dual theory. Therefore, the internal
string dynamics on the string world sheet
is described by a similar ${\mathcal N}=(2,2)\;$ sigma model. Now it has $\tilde N$ orientational moduli with the
U(1) charge $+1$ and masses $m_K=\{m_{N+1},...,m_{N_f}\}$.
To make contact with (\ref{wcp}) we call them $\tilde{\rho}^K$.
In addition, it has $N$ size moduli with the U(1) charge $-1$ and masses $(-m_P)=-\{m_1,...,m_N \}$. We
refer to these size moduli as
$\tilde{n}^P$.
The bosonic part of the action of the world-sheet model
in the gauged formulation (which assumes taking the
limit $\tilde{e}^2\to\infty$) has the form
\begin{eqnarray}
S_{{\rm dual}} \! \! \! &=& \! \! \! \int d^2 x \left\{
|\nabla_{\alpha} \tilde{\rho}^{K}|^2 +|\tilde{\nabla}_{\alpha} \tilde{n}^P|^2
+\frac1{4e^2}F^2_{\alpha\beta} + \frac1{e^2}\,
|\partial_{\alpha}\sigma|^2
\right.
\nonumber\\[3mm]
&+&\! \! \!
\left.
2\left|\sigma+\frac{m_P}{\sqrt{2}}\right|^2 \left|\tilde{n}^{P}\right|^2
+ 2\left|\sigma+\frac{m_{K}}{\sqrt{2}}\right|^2\left|\tilde{\rho}^K\right|^2
+ \frac{e^2}{2} \left(|\tilde{\rho}^K|^2-|\tilde{n}^{P}|^2 -2\tilde{\beta}\right)^2
\right\},
\nonumber\\[3mm]
&&
P=1,...,N\,,\qquad K=N+1,...,N_f\,,
\label{dcp}
\end{eqnarray}
where the covariant derivatives are defined in (\ref{covarder}).
We see that the roles of the orientational
and size moduli interchange in Eq.~(\ref{dcp}) compared with (\ref{wcp}).
As in the model (\ref{wcp}), small mass differences
$\Delta m_{AB}$ lift orientational and size zero modes of the non-Abelian semilocal string generating a
shallow potential on the moduli space.
Much in the same way as in the model (\ref{wcp}), the dual coupling constant $\tilde{\beta}$ is
determined by the bulk dual coupling $\tilde{g}_2^2$,
\begin{equation}
4\pi\tilde{\beta}(\xi) = \frac{8\pi^2}{\tilde{g}_2^2}(\xi)=
(N-\tilde N )\ln{\frac{\Lambda}{\tilde{g_2}\sqrt{\xi}}}\,,
\label{d2coupling}
\end{equation}
see Eqs.~(\ref{betag}) and (\ref{2coupling}).
The dual theory makes sense at $\tilde{g}_2 \sqrt\xi\ll \Lambda$ where $\tilde\beta$ is positive and
$$
4\pi\tilde{\beta}(\xi)\gg 1
$$
(weak coupling).
The bulk and world-sheet
dual theories have identical $\beta$
functions, with the first coefficient $(\tilde N-N)<0$. They are both infrared (IR) free. As
in the model (\ref{wcp}), the coincidence of the $\beta$ functions in the bulk
and world-sheet theories implies that the scale of the dual model (\ref{dcp}) is equal to that of the
bulk theory,
$$\tilde{\Lambda}_{\sigma}=\Lambda\,,$$
cf. Eq.~(\ref{lambdasig}).
Comparing (\ref{d2coupling}) with (\ref{2coupling}) we see that
\begin{equation}
\tilde{\beta}=-\beta\,.
\label{tbb}
\end{equation}
At $\xi\gg\Lambda^2$
the original theory is at weak coupling, and $\beta$ is positive. Analytically continuing to the domain
$\xi\ll\Lambda^2$, we formally make $\beta$ negative, which signals, of course, that the low-energy description
in terms of the original model is inappropriate. At the same time, $\tilde\beta$ satisfying
(\ref{tbb}) becomes positive, and the dual model assumes the role of the legitimate low-energy description
(at weak coupling). A direct inspection of
the dual theory action (\ref{dcp}) shows that
the dual theory can be interpreted as a continuation of the sigma model (\ref{wcp})
to negative values of the coupling constant $\beta$.
Both world-sheet theories (\ref{wcp}) and (\ref{dcp}) give the {\em effective low-energy}
descriptions of string dynamics valid at the energy scale below $g_2\sqrt{\xi}$.
Let us note that
the world-sheet duality between
two-dimensional sigma models (\ref{wcp}) and (\ref{dcp}) was previously noted in Ref.~\cite{Jsem}.
In this paper two bulk theories, with the U$(N)$ and U$(\tilde N)$ gauge groups, were considered
(these theories were referred to as a dual pair in \cite{Jsem}). Two-dimensional
sigma models (\ref{wcp}) and (\ref{dcp})
were presented as effective low-energy descriptions of the non-Abelian strings for
these two bulk theories.
\section{Semiclassical description of the world-sheet theories}
\label{secclasslim}
\setcounter{equation}{0}
At $N<N_f<2N$ the original model (\ref{wcp}) is asymptotically free, see (\ref{2coupling}). Its coupling $\beta$
continues running below $g_2\sqrt\xi$ until it stops at the scale of the mass differences
$\Delta m_{AB}$.
If all mass differences are large,
$|\Delta m_{AB}| \gg \Lambda$, the model is at weak
coupling. From (\ref{wcp}) we see that the model has $N$ vacua (i.e. $N$
strings from the standpoint of the bulk theory),
\begin{equation}
\sqrt{2}\sigma=-m_{P_0},\qquad |n^{P_0}|^2=2\beta\,,\qquad n^{P\neq P_0}=\rho^K=0\,,
\label{classvac}
\end{equation}
where $P_0=1, ..., N$.
In each vacuum there are $2(N_f-1)$ elementary excitations, counting
real degrees of
freedom. The action (\ref{wcp}) contains $N$ complex fields
$n^P$ and $\tilde N$ complex fields $\rho^K$.
The phase of $n^{P_0}$ is eaten by the Higgs mechanism.
The condition $|n^{P_0}|^2 = 2\beta$ eliminates one extra field.
The physical masses of the elementary excitations
\begin{equation}
M_A = |m_A-m_{P_0}|\,,\qquad A\neq P_0\,.
\label{elmass}
\end{equation}
In addition to the elementary excitations, there are kinks (domain ``walls" which are particles in two
dimensions) interpolating between different vacua.
Their masses scale as
\begin{equation}
M^{\rm kink} \sim \beta\, M_A \,.
\label{kinkmasscl}
\end{equation}
The kinks are much heavier than elementary
excitations at weak coupling.\footnote{ Note that they have nothing to do
with Witten's $n$ solitons \cite{W79} identified as the $n^P$ fields at
strong coupling. In the next section we present a general formula for the kink spectrum
outside CMS (at weak coupling).}
Now we pass to the dual world-sheet
theory (\ref{dcp}). It is {\em not} asymptotically free (rather, IR-free) and, therefore,
is at weak coupling at small mass differences, $|m_{AB}|\ll \Lambda$.
From (\ref{dcp}) we see that this model has $\tilde N$ vacua
\begin{equation}
\sqrt{2}\sigma=-m_{K_0},\qquad |\rho^{K_0}|^2=2\tilde{\beta}\,,
\qquad n^{P}=\rho^{K\neq K_0}=0,
\label{classvacd}
\end{equation}
where $K_0=N+1,...,N_f$.
In each vacuum there are $2(N_f-1)$ elementary excitations
with the physical masses
\begin{equation}
M_A = |m_A-m_{K_0}|\,,\qquad A\neq K_0\,.
\label{elmassd}
\end{equation}
The dual model has kinks too; their masses scale as (\ref{kinkmasscl}).
It is important to understand
that the dual theory (\ref{dcp}) is not asymptotically free
at energies much larger than the mass differences. At energies smaller than some
mass differences certain fields decouple, and the theory may or may not become
asymptotically free. Keeping in mind the desired limit (\ref{masssplit}) we will consider
the following mass choice in the dual theory
\begin{equation}
m_P\sim m_{P'}, \qquad m_K\sim m_{K'}, \qquad m_P-m_K\sim\Delta m
\label{massdef}
\end{equation}
where $P,P'=1,...,N$ and $K,K'=N+1,...,N_f$. Moreover, we will often consider
the mass hierarchy
\begin{equation}
|\Delta m_{PP'}|\sim |\Delta m_{KK'}|\ll |\Delta m|\ll \Lambda,
\label{masshier}
\end{equation}
where $\Delta m_{PP'}=m_P-m_{P'}$ and $\Delta m_{KK'}=m_K-m_{K'}$.
\vspace{1mm}
Clearly, the dual model is not asymptotically free only if
the mass differences
$\Delta m_{KK'}$ are not too small.
If we take
\begin{equation}
|\Delta m_{KK'}| \ll |\Delta m|\ll\Lambda
\label{dop1}
\end{equation}
the model becomes asymptotically free below $|\Delta m|$.
In fact,
the model then reduces to the CP$(\tilde N-1)$ model with an
effective scale
\begin{equation}
\tilde{\Lambda}_{\rm LE}^{\tilde N}\equiv \frac{(\Delta m)^N}{\Lambda^{N-\tilde N}}\,,\qquad
\tilde{\Lambda}_{\rm LE}\ll|\Delta m|\,.
\label{duallambdale}
\end{equation}
In particular, if $|\Delta m_{KK'}|\lower.7ex\hbox{$\;\stackrel{\textstyle<}{\sim}\;$} \tilde{\Lambda}_{LE}$, descending down to $\tilde{\Lambda}_{LE}$ we
enter (more exactly, the dual CP$(\tilde N-1)$ enters) the strong coupling regime.
Thus, there are {\em two} strong coupling regimes in the dual model. One is at large mass differences
$|m_{AB}|\gg \Lambda$ where the original model (\ref{wcp}) is
at weak coupling and provides an adequate description, while the other is at very small mass differences $\Delta m_{KK'}\lower.7ex\hbox{$\;\stackrel{\textstyle<}{\sim}\;$} \tilde{\Lambda}_{LE}$ where the dual model effectively reduces to the strongly coupled
CP$(\tilde N-1)$ model.
\section{Exact superpotential}
\label{secsup}
\setcounter{equation}{0}
The CP$(N-1)\;$ models are known to be described by an exact superpotential \cite{AdDVecSal,ChVa,W93,Dorey} of
the Veneziano-Yankielowicz type \cite{VYan}.
This superpotential was generalized to the case of the weighted CP models in
\cite{HaHo,DoHoTo}. In this section we will briefly outline this method.
Integrating out the fields $n^P$ and $\rho^K$ we can describe
the original model (\ref{wcp}) by the following
exact twisted superpotential:
\begin{eqnarray}
{\cal W}_{\rm eff}
& =&
\frac1{4\pi}\sum_{P=1}^N\,
\left(\sqrt{2}\,\Sigma+{m}_P\right)
\,\ln{\frac{\sqrt{2}\,\Sigma+{m}_P}{\Lambda}}
\nonumber\\[3mm]
&-&
\frac1{4\pi}\sum_{K=N+1}^{N_F}\,
\left(\sqrt{2}\,\Sigma+{m}_K\right)
\,\ln{\frac{\sqrt{2}\,\Sigma+{m}_K}{\Lambda}}
\nonumber\\[3mm]
&-& \frac{N-\tilde N}{4\pi} \,\sqrt{2}\,\Sigma\, ,
\label{2Dsup}
\end{eqnarray}
where $\Sigma$ is a twisted superfield \cite{W93} (with $\sigma$ being its lowest scalar
component).
Minimizing this superpotential with
respect to $\sigma$ we get the vacuum field formula,
\begin{equation}
\prod_{P=1}^N(\sqrt{2}\,\sigma+{m}_P)
=\Lambda^{(N-\tilde N)}\,\prod_{K=N+1}^{N_f}(\sqrt{2}\,\sigma+{m}_K)\,.
\label{sigmaeq}
\end{equation}
Note that the roots of this equation coincide with the double roots of the Seiberg--Witten curve of
the bulk theory \cite{Dorey,DoHoTo}. This is, of course, a manifestation of coincidence
of the Seiberg--Witten solution of the bulk theory with the exact solution of (\ref{wcp})
given by the superpotential (\ref{2Dsup}). As was mentioned in Sec.~\ref{intro},
this coincidence was observed in \cite{Dorey,DoHoTo} and explained later
in \cite{SYmon,HT2}.
Now, let us consider the effective superpotential of the dual world-sheet theory (\ref{dcp}).
It has the form
\begin{eqnarray}
\widetilde{\mathcal W} _{\rm eff}
&= &
\frac1{4\pi}\sum_{K=N+1}^{N_f}\,
\left(\sqrt{2}\,\Sigma+{m}_K\right)
\,\ln{\frac{\sqrt{2}\,\Sigma+{m}_K}{\Lambda}}
\nonumber\\[3mm]
& -&
\frac1{4\pi}\sum_{P=1}^{N}\,
\left(\sqrt{2}\,\Sigma+{m}_P\right)
\,\ln{\frac{\sqrt{2}\,\Sigma+{m}_P}{\Lambda}}
\nonumber\\[3mm]
&-& \frac{\tilde N-N}{4\pi} \,\sqrt{2}\,\Sigma \, .
\label{2Ddsup}
\end{eqnarray}
We see that it coincides with the superpotential (\ref{2Dsup}) up to a sign. Clearly, both, the
root equations and the BPS spectra, are the same for the two sigma
models, as was expected \cite{SYdual}.
Although classically the dual pair of the weighted CP models at hand are given by different
actions (\ref{wcp}) and (\ref{dcp}), in the quantum regime they reduce to the one and the same theory.
This is, of course, expected. Classically the couplings of both theories are determined
by the ultraviolet (UV) cut-off scale $\sqrt{\xi}$, see (\ref{2coupling}) and (\ref{d2coupling}).
However, in quantum theory these couplings run and, in fact, are determined by the mass differences.
Therefore, if $|\Delta m_{AB}|\gsim \Lambda$, the coupling $\beta$ is positive and we use
the original theory (\ref{wcp}).
On the other hand, If $|\Delta m_{AB}|\lower.7ex\hbox{$\;\stackrel{\textstyle<}{\sim}\;$} \Lambda$, the coupling
$\beta$ becomes negative, we use the dual theory (\ref{dcp}) which has positive
$\tilde{\beta}$, see Sec.~\ref{secclasslim}.
The bulk FI parameter $\xi$ no longer plays a role. Only the values of the
mass differences matter.
\vspace{1mm}
It is instructive to summarize the situation.
The theory has three distinct regimes, namely,
\vspace{1mm}
(i) The weak coupling domain in the original description at large mass differences,
\begin{equation}
|\Delta m_{AB}|\gg \Lambda;
\label{largemasses}
\end{equation}
(ii) The mixed regime in the dual description at intermediate masses,
\begin{equation}
\tilde{\Lambda}_{\rm LE} \ll |\Delta m_{AB}|\ll \Lambda\,,
\label{intermasses}
\end{equation}
where all mass differences above are assumed to be of the same order.
Certain vacua (namely, $\tilde N$ vacua) are at weak coupling
and can be seen classically, see (\ref{classvacd}), while $N-\tilde N$ other vacua are at strong coupling.
If, instead, we
assume the mass hierarchy (\ref{masshier}) then in order to keep $\tilde N$ vacua at weak coupling
we have to impose the condition
\begin{equation}
|\Delta m_{KK'}|\gg \tilde{\Lambda}_{LE}\,,
\label{dop2}
\end{equation}
see (\ref{duallambdale}).
This is the reason why we call this region ``intermediate mass" domain.
\vspace{1mm}
(iii) The strong coupling regime in the dual description at hierarchically small masses,
\begin{equation}
|\Delta m_{PP'}|\sim |\Delta m_{KK'}|\lower.7ex\hbox{$\;\stackrel{\textstyle<}{\sim}\;$} \tilde{\Lambda}_{LE}|\ll |\Delta m|\ll \Lambda\,.
\label{smallmasses}
\end{equation}
\vspace{3mm}
The masses of the BPS kinks interpolating between the
vacua $\sigma_{I}$ and $\sigma_{J}$ are given by the appropriate
differences of the superpotential (\ref{2Dsup}) calculated at distinct roots \cite{HaHo,Dorey,DoHoTo},
\begin{eqnarray}
M^{\rm BPS}_{I J}
&=&
2\left|{\cal W}_{\rm eff}(\sigma_{J})-{\cal W}_{\rm eff}(\sigma_{I})\right|
\nonumber\\
&=&
\left|\frac{N-\tilde N}{2\pi} \,\sqrt{2}(\sigma_{J}-\sigma_{I})
- \frac1{2\pi}\sum_{P=1}^{N}\,
{m}_P\,\ln{\frac{\sqrt{2}\,\sigma_{J}+{m}_P}{\sqrt{2}\,\sigma_{I}+{m}_P}}
\right.
\nonumber\\
&+&
\left.
\frac1{2\pi}\sum_{K=N+1}^{N_f}\,
{m}_K\,\ln{\frac{\sqrt{2}\,\sigma_{J}+{m}_K}{\sqrt{2}\,\sigma_{I}+{m}_K}}
\right|
\, .
\label{BPSmass}
\end{eqnarray}
The masses obtained from (\ref{BPSmass}) were shown \cite{SYmon,HT2}
to coincide with those of monopoles and dyons in the bulk theory. The latter are
given by the period integrals of the Seiberg--Witten curve \cite{Dorey,DoHoTo}.
Now we will consider the vacuum structure and the kink spectrum in more detail in
two quasiclassical regions -- at large mass differences (the original theory) and
at intermediate mass differences (the dual theory).
\subsection{Large \boldmath{$|\Delta m_{AB}|$}}
\label{seclargemasses}
Consider the vacuum structure of the theory (\ref{sigmaeq})
in the weak coupling regime $|\Delta m_{AB}|\gg \Lambda$. In this domain the model has $N$ vacua
which in the leading
(classical) approximation are determined by Eq.~(\ref{classvac}). Equation (\ref{sigmaeq}) reproduces
this vacuum structure.
Namely, VEVs of $\sigma$ in each of the $N$ vacua (say, at $P=P_0$) are given by the corresponding mass $m_{P_0}$,
plus a small correction,
\begin{equation}
\sqrt{2}\sigma_{P_0} \approx -m_{P_0}
+\Lambda^{N-\tilde N}\,\frac{\prod_{K=N+1}^{N_f}({m}_K-m_{P_0})}{\prod_{P\neq P_0}({m}_P-m_{P_0})}
\,.
\label{sigmaclass}
\end{equation}
The spectrum of kinks is given by Eq.~(\ref{BPSmass}). To be more specific, let us
consider the kinks interpolating between
the neighboring vacua\,\footnote{If the mass parameters $m_P$
are randomly scattered in the complex plane, how one should define
the ``neighboring vacua"? In the regime under consideration, for all $P$ the vacuum values
$\sigma_P$ are close to $-m_P/\sqrt{2}$. Assume $\sigma_{P_0}$ is chosen.
Then the neighboring vacuum $\sigma_{P_0+1}$ is defined in such a way that the difference
$|m_{P_0} - m_{P_0+1}|$ is the smallest in the set $|\Delta m_{P_0P}|$.}
$P_0$ and $\mbox{$(P_0+1)$}$. Then we have
\begin{equation}
m^{\rm kink}=|m_D^{P_0+1}-m_D^{P_0}|\approx
\left| \left(m_{P_0}-m_{P_0+1}\right)\,\frac{N-\tilde N}{2\pi}\,\ln{\frac{\bar\Delta m_{AB}}{\Lambda}}\right|,
\label{mD}
\end{equation}
where
$\bar\Delta m_{AB}$
is a certain average value of the mass differences (it depends holomorphically on all
mass differences in the problem).
Here we use (\ref{sigmaclass}).
If all mass
differences are of the same order
so is $\bar\Delta m_{AB}$, although $\bar\Delta m_{AB}$ does not coincide with any of
the individual mass differences. For a generic choice of the mass differences
$\bar\Delta m_{AB}$ has a nonvanishing phase.
We see that in the logarithmic approximation the kink mass
is proportional to the mass difference $(m_{P_0}-m_{P_0+1})$ times the coupling constant
$\beta$, as one would expect at weak coupling, cf. Eq.~(\ref{kinkmasscl}).
This is not the end of the story, however. The logarithmic functions
in (\ref{BPSmass}) are multivalued, and we have to carefully choose their branches.
Each logarithm in (\ref{2Dsup}) can be written in the integral form as
\begin{equation}
\frac{m_{A}}{2\pi}\,\ln{\frac{\sqrt{2}\,\sigma_{P_0+1}+{m}_A}{\sqrt{2}\,\sigma_{P_0}+{m}_A}}
=\frac{m_{A}}{2\pi}\,\int_{\sigma_{P_0}}^{\sigma_{P_0+1}}\,
\frac{\sqrt{2}\,d\sigma}{\sqrt{2}\,\sigma +m_A}
\,,\label{integral}
\end{equation}
with the integration contour to be appropriately chosen. Distinct choices
differ by pole contributions
\begin{equation}
{\rm integer}\times im_A
\label{dop3}
\end{equation}
for different $A$.
These different mass predictions
for the BPS states correspond to dyonic kinks. In addition to the topological charge,
the kinks can carry Noether charges with respect to
the the global group (\ref{c+f}) broken down to U(1)$^{N_f}$ by the mass differences.
This produces a whole family of dyonic kinks.\footnote{ They represent confined
monopoles and dyons with the root-like electric charges in the bulk theory.}
We stress that all these kinks with the imaginary part (\ref{dop3}) in the mass formula
interpolate between the same pair of vacua: $P_0$ and $(P_0+1)$. Our aim is
to count their number and calculate their masses.
Generically there are way too many choices of the integration contours in (\ref{BPSmass}).
Not all of them are realized. Moreover, the kinks present in the quasiclassical domain
decay on CMS or form new bound states, cf. e.g. \cite{svz,OlmezS}. Therefore the quasiclassical
spectrum outside CMS and quantum spectrum inside CMS are different.
We have to use an additional input on the structure of kink solutions to find out
the correct form of the BPS spectrum. In this section
we will summarize information on the classical spectrum while in the remainder of the
present paper we will use
the mirror representation \cite{HoVa} of the model at hand to obtain the quantum spectrum.
The general formula for the BPS spectrum can be written as follows \cite{Dorey}:
\begin{equation}
M^{\rm BPS}=\left| \sum_I m_{D}^I T_I + i\sum_A m_A S_A\right|,
\label{TSmass}
\end{equation}
where the first term is a nonperturbative contribution and $T_I$ is the topological charge
$N$-vector, while the second term
represents the dyonic (the Noether charge) ambiguity discussed above, with $S_A$ describing a global
U(1) charge of the given BPS
state with respect to the U(1)$^{N_f}$ group.
The topological charge is given by
\begin{equation}
T_P=\delta_{PP_0+1}-\delta_{PP_0}
\label{topcharge}
\end{equation}
for kinks interpolating between the vacua $P_0$ and $(P_0+1)$, while $m_D$'s are
approximately given by the logarithmic terms in (\ref{mD}),
\begin{equation}
m_D^{P_0}\approx m_{P_0}\,\frac{N-\tilde N}{2\pi}\,\ln{\frac{\bar\Delta m_{AB}}{\Lambda}}.
\end{equation}
Equation~(\ref{mD}) corresponds to
the kink with $S_A=0$.
At weak coupling the BPS kinks can be studied using the classical solutions of the
first-order equations. Each given kink solution breaks the global U(1)$^{N_f}$ group. Therefore, the
kinks acquire zero modes associated with rotations in this internal group.
Quantization of the corresponding dynamics gives rise to dyonic kinks which carry global charges
$S_A$. This program was carried out for the CP$(N-1)\;$ model in \cite{Dorey} and
for the weighted CP model (\ref{wcp}) in \cite{DoHoTo}. The result is
\begin{equation}
S_P=sT_P\,,\qquad S_K=0
\label{classdyoncharge}
\end{equation}
for $P=1,...,N$ and $K=N+1,...N_f$, where $s$ is integer. Thus, at large $|\Delta m_{AB}|$ we have an
infinite tower of the
dyonic kinks with masses
\begin{eqnarray}
M^{\rm kink}
&\approx&
\left| \left(m_{P_0}-m_{P_0+1}\right)\right|
\nonumber\\[3mm]
&\times&
\left| \,\frac{N-\tilde N}{2\pi}\,\ln{\frac{\bar\Delta m_{AB}}{\Lambda}}
- is\right|.
\label{kinkspectrlm}
\end{eqnarray}
The expression in the second line under the sign of the absolute value
has both, real and imaginary parts. The real part is obtained from the logarithm by replacing
$\bar\Delta m_{AB}$ under the logarithm by $|\bar\Delta m_{AB}|$. The imaginary part
includes the phase of $\bar\Delta m_{AB}$, which, in principle,
could be obtained for any given set of the mass differences,
but in practice this is hard to do for generic mass choices. In addition, the imaginary part
includes $is$, where $s$ is an integer (positive, negative or zero). When we change $s$,
we scan all possible values of the U(1) charge (i.e. go through the entire set of dyons).
In addition to the above monopoles/dyons,
in this domain of $\Delta m_{AB}$ there are elementary excitations, see Eq. (\ref{elmass}).
These excitations are BPS-saturated too and can be described by the general formula (\ref{TSmass})
with $T=0$ and $S_A=\delta_{AB}-\delta_{A P_0}$ for any $B=1, ..., N_f$ in
the $P_0$-vacuum \footnote{ The actual kink spectrum at weak coupling is more complicated
than the one in (\ref{kinkspectrlm}). the kink states from the tower (\ref{kinkspectrlm}) can form
bound states with different elementary states in certain special domains of the mass
parameters \cite{DoHoTo,LeeYi}.}.
The above spectrum changes upon passing through CMS. In particular, we will see
that elementary excitations do not exist inside CMS. All excitations that survive inside CMS are
the kinks
with nonvanishing topological charges. This is a two-dimensional counterpart of the
bulk picture: the quarks and gauge bosons decay inside the strong coupling domain giving rise to the
monopole-antimonopole pairs, see Sec.~\ref{secbulk}.
\subsection{Intermediate masses}
\label{secintmass}
Now let is consider the domain of intermediate mass differences, see Eq.~(\ref{intermasses}). Then
Eq.~(\ref{sigmaeq}) has $(N-\tilde N)$ solutions with
\begin{equation}
\sqrt{2}\sigma=\Lambda\,\exp\left({\frac{2\pi i}{N-\tilde N}\,l}\right),\qquad l=0,...,(N-\tilde N-1).
\label{Lvac}
\end{equation}
We will refer to these vacua as the $\Lambda$-vacua. They are at strong coupling, and will be studied later.
In this section we consider other $\tilde N$ vacua, which are
at weak coupling and seen classically in the dual
theory, see Eq.~(\ref{classvacd}).
For these vacua Eq.~(\ref{sigmaeq}) gives
\begin{equation}
\sqrt{2}\sigma_{K_0} \approx -m_{K_0}
+\frac{1}{\Lambda^{N-\tilde N}}\,\frac{\prod_{P=1}^{N}({m}_P-m_{K_0})}{\prod_{K\neq K_0}({m}_K-m_{K_0})}
\,,
\label{sigmaclassd}
\end{equation}
where $K_0=N+1,...,N_f$. These $\tilde N$ vacua will be referred to as the {\em zero-vacua} since
in these vacua, with
small masses, the $\sigma$ vacuum expectation values are much smaller than in the
$\Lambda$-vacua.
Substituting this in Eq.~(\ref{BPSmass}) we get the spectrum of the kinks interpolating
between the neighboring vacua
$K_0$ and $K_0+1$,
\begin{equation}
M^{\rm kink}\approx
\left| \left(m_{K_0}-m_{K_0+1}\right)\,\frac{N-\tilde N}{2\pi}\,\ln\frac{\Lambda}{\bar\Delta m_{AB}}
+ is\left(m_{K_0}-m_{K_0+1}\right)\right|,
\label{kinkspectrim}
\end{equation}
where we take into account the U(1) charges parallelizing the derivation of
Eq.~(\ref{kinkspectrlm}) and applying the quantization
procedure of \cite{Dorey,DoHoTo} to the dual theory (\ref{dcp}).
As previously,
all mass differences are assumed to be of the same order.
If, instead, we consider a stricter mass hierarchy (\ref{masshier}) (still requiring that
we are at weak coupling $|\Delta m_{KK'}|\gg \tilde{\Lambda}_{LE}$) then Eq.~(\ref{kinkspectrim})
must be modified. With this stricter hierarchy the product in the numerator of the
second term in (\ref{sigmaclassd}) reduces to $(\Delta m)^N$ to form $\tilde{\Lambda}_{LE}$,
and the kink spectrum takes the form
\begin{equation}
M^{\rm kink}\approx
\left| \left(m_{K_0}-m_{K_0+1}\right)\,\frac{\tilde N}{2\pi}\,
\ln\frac{ \bar\Delta m_{KK'}}{\tilde{\Lambda}_{LE}}
+ is\left(m_{K_0}-m_{K_0+1}\right)\right|,
\label{kinkspectrim2}
\end{equation}
where $\tilde{\Lambda}_{LE}$ is given by (\ref{duallambdale}). This is just the kink spectrum
in the CP$(\tilde N-1)$ model at weak coupling.
In addition to the $T\neq 0$ kinks,
there are elementary excitations with masses given in Eq.~(\ref{elmassd}).
They correspond to $T=0$ and $S_A=\delta_{AB}-\delta_{A K_0}$ for any $B=1,...,N_f$ in the $K_0$-vacuum
in (\ref{TSmass}).
Confronting Eqs.~(\ref{kinkspectrlm}) and (\ref{kinkspectrim})
we see that the kinks
have different Noether charges in the domains of large and intermediate mass differences.
At large masses they have charges with respect to the first $N$ factors of the
global U(1)$^{N_f}$ group, while the kinks at the intermediate masses are charged with respect to
the last $\tilde N$ factors (this would correspond to SU$(N)$ and SU$(\tilde N)$ factors
of the global group (\ref{c+f}) in the limit (\ref{masssplit})). Therefore, they are,
in fact, absolutely distinct states.
The BPS states decay/form new bound states upon passing from one domain to
another. The restructuring happens on CMS which are surfaces located at $|\Delta m_{AB}|\sim \Lambda$
in the mass parameter space.\footnote{Of course, this restructuring
is a reflection of the same phenomenon in
the bulk
theory, see Sec.~\ref{secbulk}.} As we will see shortly,
in the weighted CP model at hand we
have another set of CMS at much smaller mass differences $|\Delta m_{KK'}|\sim \tilde{\Lambda}_{LE}$.
This additional CMS separates the domain of intermediate masses from
that at strong coupling, see (\ref{smallmasses}).
\section{Mirror description}
\label{secmirror}
\setcounter{equation}{0}
Now we turn to the study of the quantum BPS spectrum inside CMS. We will determine
the BPS spectrum in the $\Lambda$-vacua (\ref{Lvac}) at intermediate and small masses,
as well as the spectrum in the zero-vacua in the strong coupling domain at
hierarchically small masses (\ref{smallmasses}).
\subsection{Mirror superpotential}
To this end we will rely on the mirror
formulation \cite{HoVa} of the weighted CP model (\ref{wcp}). In this formulation one
describes the CP model as a Coulomb gas of instantons (see \cite{FFS} where it was first
done in the nonsupersymmetric CP(1) model). In supersymmetric setting this description leads
to an affine Toda theory with an exact superpotential. The exact mirror superpotentials
were found for the ${\mathcal N}=(2,2)\;$ CP$(N-1)\;$ model and its various generalizations with toric
target spaces in \cite{HoVa}. For the model (\ref{wcp}) the mirror superpotential has the form
\begin{equation}
W_{\rm mirror}=-\frac{\Lambda}{4\pi}\left\{\sum_{P=1}^{N} X_P -\sum_{K=N+1}^{N_f} Y_K
-\sum_{P=1}^{N} \frac{m_P}{\Lambda}\,\ln{X_P}+\sum_{K=N+1}^{N_f}\frac{m_K}{\Lambda}\,\ln{Y_K}
\right\}
\label{mirror}
\end{equation}
supplemented by the constraint
\begin{equation}
\prod_{P=1}^{N} X_P=\prod_{K=N+1}^{N_f} Y_K.
\label{constraint}
\end{equation}
This representation can be checked by a straightforward calculation. Indeed, add the term
\begin{equation}
\frac{\Lambda}{4\pi} \,\sqrt{2}\Sigma\left(\sum_{P=1}^{N} \ln{X_P}-\sum_{K=N+1}^{N_f} \ln{Y_K}
\right)
\end{equation}
to the superpotential (\ref{mirror}), which takes into account the constraint
(\ref{constraint}). Here $\Sigma$ plays a role of the Lagrange multiplier.
Then integrate over $X_P$ and $Y_K$ ignoring their kinetic terms. In this way one arrives at
\begin{equation}
X_P=\frac1{\Lambda}\left(\sqrt{2}\,\sigma+{m}_P\right),
\qquad Y_K=\frac1{\Lambda}\left(\sqrt{2}\,\sigma+{m}_K\right).
\label{XandY}
\end{equation}
Substituting (\ref{XandY}) back into (\ref{mirror}) one gets the superpotential
(\ref{2Dsup}).
Clearly for the dual model (\ref{dcp}) the mirror superpotential coincides
with that in (\ref{mirror}) up to an (irrelevant) sign.
Below we will
find the critical points of the superpotential (\ref{mirror}) and discuss
the vacuum structure of the model in the mirror representation. Since our goal is
to study the domain of intermediate or hierarchically small masses, see (\ref{intermasses})
and (\ref{smallmasses}), respectively, we will assume that
\begin{equation}
|\Delta m_{AB}|\ll \Lambda
\end{equation}
and keep only terms linear in $|\Delta m_{AB}|/ \Lambda$. As a warm-up exercise
we will start with the
CP$(N-1)\;$ model.
\subsection{\boldmath{CP$(N-1)\;$} model}
For CP$(N-1)\;$ model $\tilde N=0$ and the superpotential (\ref{mirror}) reduces to
\begin{equation}
W_{\rm mirror}^{{\rm CP}(N-1)}=-\frac{\Lambda}{4\pi}\left\{\sum_{P=1}^{N} X_P
-\sum_{P=1}^{N} \frac{m_P}{\Lambda}\,\ln{X_P}\,,
\right\}
\label{mirrorcp}
\end{equation}
while the constraint (\ref{constraint}) reads
\begin{equation}
\prod_{P=1}^{N} X_P=1\,.
\label{constraintcp}
\end{equation}
Expressing, say $X_1$ in terms of $X_P$ with $P=2,3, ..., N$ by virtue of this constraint
and substituting the result in (\ref{mirrorcp}), we get the
vacuum equations,
\begin{equation}
X_P=X_1+\frac{ m_{P}-m_1}{\Lambda} = X_1+\frac{\Delta m_{P1}}{\Lambda}\,.\qquad P=2, ..., N\,.
\label{vaceqcp}
\end{equation}
Substituting this in (\ref{constraintcp}) we obtain $X_1$. This procedure
leads us to the following VEV's of the $X_P$ fields:
\begin{equation}
X_P\approx \exp{\left(\frac{2\pi i}{N}\,l\right)} + \frac{1}{\Lambda}\left(
m_P-m\right)\,,\qquad \forall P\,,
\label{xvevcp}
\end{equation}
where we ignore quadratic in mass
differences terms,
\begin{equation}
m\equiv\frac1N\sum_{P=1}^{N} m_{P}\,,
\label{m}
\end{equation}
and
each of the $N$ vacua of the model is labeled by a value of $l$, namely
$l=0$ in the first vacuum, $l=1$ in the second, and so on till we arrive at $l=(N-1)$.
\subsection{\boldmath{$\Lambda$}-vacua}
\label{lambdava}
Now we turn to the strong coupling vacua in the weighted CP model at intermediate
or small masses
(see Eq.~(\ref{smallmasses})), using
the mirror representation (\ref{mirror}).
These vacua were defined as $\Lambda$-vacua in Sec.~\ref{secintmass}, Eq.~(\ref{Lvac}).
Again expressing, say, $X_1$
in terms of other fields by virtue of the constraint (\ref{constraint}) we get the vacuum equations
\begin{equation}
X_P=X_1+\frac{\Delta m_{P1}}{\Lambda},
\qquad Y_K=X_1+\frac{\Delta m_{K1}}{\Lambda}\, .
\label{vaceqL}
\end{equation}
Substituting this in the constraint (\ref{constraint}) and resolving for $X_1$ we get
the following VEVs:
\begin{eqnarray}
X_P
&\approx &
\exp{\left(\frac{2\pi i}{N-\tilde N}\,l\right)} + \frac{1}{\Lambda}\left(
m_P-\hat{m}\right),\quad P=1,...,N,
\nonumber\\[3mm]
Y_K
&\approx &
\exp{\left(\frac{2\pi i}{N-\tilde N}\,l\right)} + \frac{1}{\Lambda}\left(
m_K-\hat{m}\right),\quad K=N+1,...,N_f\,,
\nonumber\\[3mm]
l&=&0, ..., (N-\tilde N-1)
\label{vevL}
\end{eqnarray}
for $(N-\tilde N)$ vacua of the theory. Here
\begin{equation}
\hat{m}\equiv\frac1{N-\tilde N}\left(\sum_{P=1}^{N} m_{P}-\sum_{K=N+1}^{N_f}m_{K}\right)\,.
\label{hatm}
\end{equation}
As in (\ref{xvevcp}), in Eq.~(\ref{vevL}) we neglect
the quadratic in mass difference terms, cubic, and so on.
\subsection{Zero-vacua}
\label{zerovacu}
Now consider other $\tilde N$ vacua of the model (\ref{dcp}) (which were termed zero-vacua
in Sec.~\ref{secintmass}) using the mirror description (\ref{mirror}). At intermediate masses,
VEVs of $\sigma$ are given, in the classical approximation, by the mass parameters $m_K$ in the dual
theory, see (\ref{classvacd}). The corrections are given by (\ref{sigmaclassd}).
Since the mirror representation is particularly useful at strong coupling, in this section we will
focus on a very small hierarchical region of the mass parameters (\ref{smallmasses}).
It is convenient to express one of $Y_K$'s, say, $Y_{N_f}$ in terms of other fields
using (\ref{constraint}). Then vacuum equations the take the form
\begin{equation}
X_P=Y_{N_f}+\frac{\Delta m_{PN_f}}{\Lambda}\,,
\qquad Y_K=Y_{N_f}+\frac{\Delta m_{KN_f}}{\Lambda}\,.
\label{vaceqz}
\end{equation}
From the first equation we see that, given the hierarchical masses (\ref{smallmasses}),
all $X_P$ fields are equal to each other to the leading order,
\begin{equation}
X_P^{(0)}\approx \frac{\Delta m}{\Lambda}\,, \qquad P=1, ..., N\,.
\label{Xvevzero}
\end{equation}
With these $X_P$'s the constraint (\ref{constraint}) takes the form
\begin{equation}
\prod_{K=N+1}^{N_f} Y_K=\left(\frac{\tilde{\Lambda}_{LE}}{\Lambda}\right)^{\tilde N},
\label{constraintY}
\end{equation}
where $\tilde{\Lambda}_{LE}$ is the scale of the effective low energy CP$(\tilde N-1)$ model
(\ref{duallambdale}).
Substituting the fields $Y_K$ from
the second equation in (\ref{vaceqz}) we get
\begin{equation}
Y_K\approx \frac{\tilde{\Lambda}_{LE}}{\Lambda}\,\left\{
\exp{\left(\frac{2\pi i}{\tilde N}\,l\right)} + \frac{1}{\tilde{\Lambda}_{LE}}\left(
m_K-\tilde{m}\right)\right\},
\label{Yvev}
\end{equation}
where $l=0,...,\tilde N-1$,
\begin{equation}
\tilde{m}\equiv\frac1{\tilde N}\sum_{K=N+1}^{N_f}m_{K}\,.
\label{tildem}
\end{equation}
As usual, we neglect quadratic, cubic, etc. mass-difference terms.
Finally, we are ready to solve the vacuum equations.
Substituting (\ref{Yvev}) in the first expression in (\ref{vaceqz})
we get $O(\tilde{\Lambda}_{LE}/\Lambda )$ corrections to (\ref{Xvevzero}),
\begin{equation}
X_P\approx \frac{1}{\Lambda}(m_P-\tilde m)+\frac{\tilde{\Lambda}_{LE}}{\Lambda}\,
\exp{\left(\frac{2\pi i}{\tilde N}\,l\right)}, \qquad l=0,...,\tilde N-1\,.
\label{Xvev}
\end{equation}
We see that there are exactly $\tilde N$ vacua with very small values of $Y_K$'s. The VEV structure
of $Y_K$'s reduces to that of the CP$(\tilde N-1)$ model with the scale
parameter $\tilde{\Lambda}_{LE}$,
see (\ref{xvevcp}).
\section{Kinks inside CMS}
\label{seckink}
\setcounter{equation}{0}
In this section we use the mirror representation to find
the kink spectrum inside CMS in the
weighted CP model on the string world sheet. First, we briefly
review the kink solutions and their spectrum \cite{HoVa}
in the $\mbox{CP$(N-1)\;$}$ model and only then turn to the weighted CP model
(\ref{wcp}).
\subsection{Kinks in the \boldmath{CP$(N-1)\;$} model}
\label{seccpnkinks}
As was shown in \cite{HoVa}, in the strong coupling regime (inside CMS) the number
of kinks interpolating between the vacua $P$ and $P+k$ of ${\mathcal N}=(2,2)\;$ supersym\-metric CP$(N-1)\;$ model is
\begin{equation}
\nu(N,k)=C_N^k\equiv\frac{N!}{k!(N-k)!}\, .
\end{equation}
In particular, the number of kinks interpolating between the neighboring vacua ($k=1$)
is $N$, and they
form a fundamental representation of the SU$(N)$ group. They carry the minimal charge with respect to
the gauge U(1) and, therefore, were identified \cite{W79}
as $n^P$ fields in terms of the original description, see (\ref{wcp}) for $\tilde N=0$.
Consider a kink interpolating between the neighboring $l=0$ and $l=1$ vacua, see (\ref{xvevcp}).
The kink solution has the following structure \cite{HoVa}. All $X_P$'s start in the vacuum
with $l=0$ and end in the vacuum with $l=1$.
Moreover, all $X_P$ with $P\neq P_0$
(were $P_0$ is fixed for a given kink solution) are equal to each other and have
phases which wind by the angle $2\pi/N$ in the anti-clockwise direction. Then the constraint
(\ref{constraintcp}) ensures that $X_{P_0}$ winds in the opposite (clockwise) direction
by the angle $[-2\pi (N-1)/N]$, see Fig.~\ref{figkinkcp}. (If one considers
nonelementary kinks
interpolating between non-neighboring vacua, say $l=0$ and $l=2$, then two fields $X_{P_0}$ and
$X_{P_0^\prime}$ would have
opposite windings with respect to all others, and so on \cite{HoVa}.)
\begin{figure}
\epsfxsize=8cm
\centerline{\epsfbox{kinkcp.eps}}
\caption{\small Windings of the fields $X_P$ for the kink interpolating between
the $l=0$ and $l=1$ vacua.}
\label{figkinkcp}
\end{figure}
The kink mass is given by
\begin{eqnarray}
M^{\rm BPS}
&=&
2\left|{\cal W}_{\rm mirror}^{{\rm CP}(N-1)}(l=1)-{\cal W}_{\rm mirror}^{{\rm CP}(N-1)}(l=0)\right|
\nonumber\\[3mm]
&\approx&
\left|\frac{N}{2\pi}\,\Lambda\left(e^{\frac{2\pi}{N}\,i}-1\right) +i\left(m_{P_0}-
m\right)\right|,
\label{cpspectr}
\end{eqnarray}
where we use (\ref{mirrorcp}) and neglect quadratic in mass differences terms.
The parameter $m$ is defined in Eq.~(\ref{m}).
If we look at the absolute values of the fields $X_P$ rather than at their phases,
we will see that, generically, their absolute values differ from unity. This is discussed
in Appendix A, cf. also \cite{RitzSV}.
The explicit profile functions of the kink solutions are irrelevant
for determination of the kink spectrum,
since the latter is given by
central charges --
the differences of the superpotential on the initial and
final vacua. The phases of $X_P$ are important, however, because logarithms in (\ref{mirrorcp})
are multivalued functions. In particular, the term $im_{P_0}$ appeared in (\ref{cpspectr})
because $X_{P_0}$ has the relative winding angle $(-2\pi)$ with respect to other $X_P$'s.
We see that we have exactly $N$ dyonic kinks associated for the given choice of $P_0$
and its neighbor.
(In addition, $P_0$ can be chosen arbitrarily from the set $P_0=1,...,N$).
The above dyonic kinks
have different charges with respect to
the global U(1)$^{N}$ and are split at generic masses, but become degenerate in the equal mass limit.
Clearly, they form a fundamental representation of the global SU$(N)$ in this limit.
We stress again that all $N$ kinks here interpolate between two fixed vacua, $l=0$ and $l=1$.
The BPS spectrum inside CMS, see (\ref{cpspectr}), is very different from that
outside CMS, see (\ref{kinkspectrlm}) with $\tilde N=0$. The weak coupling spectrum has an
infinite tower of dyonic kinks, all associated with the same mass difference
$(m_{P_0}-m_{P_0+1})$. Also, the weak coupling spectrum has elementary states
with $T_P=0$. The quantum spectrum has only a finite number of states ($N$), with masses
which depend on all mass differences present in the theory. Moreover, all these states
are topological (they are kinks), no elementary states are present. The majority of the BPS states
present at weak coupling (in particular, all elementary excitations) decay on CMS and are
absent at strong coupling.
In conclusion we note that Eq.~(\ref{BPSmass}) is exact and,
in principle, can be used to calculate the BPS spectrum in any domain of the parameter space
of the theory. We can put $\tilde N=0$ in this formula (descending down to the CP$(N-1)\;$ model) and
apply (\ref{BPSmass}) to the kinks which interpolate between two vacua $l=0$ and $l=1$ with
VEVs of $\sigma$ given by
\begin{equation}
\sqrt{2}\sigma\approx \Lambda\,\exp\left({\frac{2\pi i}{N}\,l}\right),\qquad l=0,...,N
\label{cpvac}
\end{equation}
at small masses, cf. (\ref{Lvac}) with $\tilde N=0$.
Say, the main contribution in (\ref{cpspectr}) proportional to $\Lambda$
comes from the first nonlogarithmic term in the second line in
(\ref{BPSmass}). Moreover, now the result in (\ref{cpspectr}) shows how ambiguities
related to the choice of the logarithm branches in (\ref{BPSmass}) should be
resolved at strong coupling. Namely, we get exactly the same BPS spectrum
as in (\ref{cpspectr}) from (\ref{BPSmass}) (with $\tilde N=0$) if we choose the
integration contours in (\ref{integral}) as shown in Fig.~\ref{figcontourcp}.
For the $P_0$-th dyonic kink , the integration contour should pick up exactly one pole contribution
located at $\sqrt{2}\sigma=-m_{P_0}$ in the clockwise direction.
This shows, in fact, how the kink solutions look in terms of the field $\sigma$.
\begin{figure}
\epsfxsize=10cm
\centerline{\epsfbox{contourcp.eps}}
\caption{\small Integration contour in the $\sigma$ plane. Dots denote two vacua
$l=0$ and $l=1$, while filled squires denote poles located at $\sqrt{2}\sigma=-m_{P}$
for $P=1, ..., N$.}
\label{figcontourcp}
\end{figure}
The above recipe was obtained in the CP$(N-1)\;$ model in \cite{HaHo} using a
brane construction,
see also \cite{DoHoTo}. Here we confirm it directly in field theory
using the mirror description of the model. See also \cite{RitzSV}.
\subsection{Kinks in the \boldmath{$\Lambda$}-vacua}
\label{kilva}
In this section we work out a similar procedure to obtain the BPS spectrum in
the weighted CP model
(\ref{dcp}). We will focus on the $\Lambda$-vacua
at intermediate or small masses, see (\ref{intermasses}) and (\ref{smallmasses}).
Consider kinks interpolating between the
neighboring $l=0$ and $l=1$ vacua (\ref{vevL}).
Much in the same way as in the CP$(N-1)\;$ model (Sec.~\ref{seccpnkinks}), all $X_P$'s and $Y_K$'s are equal
to each other and
wind by the angle $2\pi/(N-\tilde N)$, except one field whose winding angle is determined by the
constraint (\ref{constraint}). There are two different types of kinks depending on
the choice of the latter variable: $X_{P_0}$ or $Y_{K_0}$, ($P_0=1, ..., N$ and
$K_0=(N+1), ..., N_f$). We refer to these two types
of solutions as to the $P$- and $K$-kinks, respectively.
\subsubsection{\boldmath{$P$}-kinks}
\label{subsub1}
Solutions for $P$-kinks are very similar to those for kinks in the $\mbox{CP$(N-1)\;$}$ model.
In a given $P$-kink the variable $X_{P_0}$ winds in the clockwise direction by
the angle $[-2\pi (N-\tilde N-1)/(N-\tilde N)]$, see Fig.~\ref{figkinkcp}.
All other fields, i.e. $X_{P\neq P_0}$ and $Y_K$, are equal to each other and wind
counterclockwise by
the angle $2\pi/(N-\tilde N)$.
The superpotential (\ref{mirror}) implies that the mass
of this kink is
\begin{equation}
M^{\rm BPS}_{P_{0}} \approx
\left|\frac{N-\tilde N}{2\pi}\,\Lambda\left(e^{\frac{2\pi}{N-\tilde N}\,i}-1\right) +i\left(m_{P_0}-
\hat{m}\right)\right|,
\label{LspectrP}
\end{equation}
where $\hat{m}$ is given in Eq.~(\ref{hatm}).
We see that in the transition $l=0 \longrightarrow l=1$
we have exactly $N$ kinks associated with the arbitrary choice of of the contour (with the loop around the pole
$P_0=1,...,N$).
They are split for generic masses, but become degenerate in the limit (\ref{masssplit})
we are interested in.
They form the fundamental representation of the global SU$(N)$ in this limit.
\subsubsection{ \boldmath{$K$}-kinks}
\label{subsub2}
In the $K$-kink solutions all fields $X_{P}$ and $Y_{K\neq K_0}$ are equal to each other and have
the winding
angle $2\pi/(N-\tilde N)$, while the field $Y_{K_0}$ winds in the anticlockwise direction by
the angle $[2\pi (N-\tilde N+1)/(N-\tilde N)]$, see Fig.~\ref{figkinkL}, as dictated\,\footnote{ In fact,
for these
solutions to exist $(N-\tilde N)$ is required to be large enough.
If $(N-\tilde N)$ is not large enough, the functions of $|X_P|$ and $|Y_K|$
develop singularities; the singular solutions must be discarded, cf. Appendix A. The reason
behind this phenomenon is that for
$(N-\tilde N)$ not large enough the vacuum which is the closest neighbor to $l=0$ is, in fact,
one of the zero-vacua (Sec.~\ref{zerovacu}) rather than the $l=1$ $\Lambda$-vacuum.}
by the constraint
(\ref{constraint}). Thus, the $Y_{K_0}$ field has the relative winding $+2\pi$ with respect
to all other fields.
\begin{figure}
\epsfxsize=8cm
\centerline{\epsfbox{kinkL.eps}}
\caption{\small Windings of f
the fields $X_P$ and $Y_K$ for the $K$-kink interpolating between the
$l=0$ and $l=1$ $\Lambda$-vacua.}
\label{figkinkL}
\end{figure}
\noindent
Substituting this
information in the superpotential (\ref{mirror}) we obtain the $K$-kink mass,
\begin{equation}
M^{\rm BPS}_{K_0} \approx
\left|\frac{N-\tilde N}{2\pi}\,\Lambda\left(e^{\frac{2\pi}{N-\tilde N}\,i}-1\right) +i\left(m_{K_0}-
\hat{m}\right)\right|.
\label{LspectrK}
\end{equation}
Clearly, we have $\tilde N$ kinks of this type due to the possibility of choosing the contour
encompassing any of the $\tilde N$ poles,
$K_0=(N+1), ..., N_f$.
They are split at generic masses, but become degenerate in the limit (\ref{masssplit}).
They form the fundamental representation of the global SU$(\tilde N)$ group in this limit.
Altogether we have $N_f$ kinks interpolating between each pair of neighboring $\Lambda$-vacua.
They form the fundamental representation of the global group (\ref{c+fd}) in the
limit (\ref{masssplit}). More exactly, they form the $(N,1)+(1,\tilde N)$ representation of this
group.
Much in the same way as in the CP$(N-1)\;$ model, we can verify that Eq.~({\ref{BPSmass}) reproduces
this spectrum with the appropriate choice of the integration contours. Namely,
for the $P_0$-kink the contour in the $\sigma$ plane encircles the pole at
$\sqrt{2}\sigma=-m_{P_0}$ in the clockwise direction, see Fig.~\ref{figcontourcp}.
For the $K_0$-kink
the contour encircles the pole at
$\sqrt{2}\sigma=-m_{K_0}$ in the anticlockwise direction.
\subsection{Kinks in the zero-vacua}
\label{seckinkZ}
Now we consider kinks in the
zero-vacua in the domain of
small hierarchical masses (\ref{smallmasses}).
These vacua of the weighted CP model are most interesting since they correspond
to $\tilde N$ non-Abelian strings of the dual bulk theory. Clearly, the number of kinks
does not depend on which pair of neighboring vacua we pick up. Thus, we expect to have
altogether $N_f$ kinks, as was the case in the $\Lambda$-vacua. We check this explicitly below.
Much in the same way as in the $\Lambda$-vacua, the kinks interpolating between
the neighboring
$l=0$ and $l=1$ zero-vacua (see (\ref{Yvev})) fall into two categories: the $P$-
and $K$-kinks, respectively, depending on the choice of the particular $X_{P_0}$ or $Y_{K_0}$ field
with an opposite winding with respect to other fields.
\subsubsection{\boldmath{$K$}-kinks}
\label{subsub3}
Let us start from the $K$-kinks. The kink solution looks very similar to that in the
CP$(\tilde N-1)$ model. All $Y_{K\neq K_0}$ fields are approximately equal to each other and
have the winding angles $2\pi/\tilde N$, while the $Y_{K_0}$ field has the winding angle $-2\pi(\tilde N-1)/\tilde N$,
see (\ref{Yvev}). Correction terms in $X_P$, proportional to $\tilde{\Lambda}_{LE}/\Lambda$
also all have the same windings by the angle $2\pi/\tilde N$, see (\ref{Xvev}).
This gives the following expression for the mass of the $K_0$-kink:
\begin{equation}
M^{\rm BPS}_{K_0} \approx
\left|\frac{N-\tilde N}{2\pi}\,\tilde{\Lambda}_{LE}\left(e^{\frac{2\pi}{\tilde N}\,i}-1\right) -i\left(m_{K_0}-
\tilde{m}\right)\right|.
\label{ZspectrK}
\end{equation}
The factor $(N-\tilde N)$ in the first term appears from two first terms
in (\ref{mirror}) due to winding of both $X_P$ and
$Y_K$ fields. The $im_{K_0}$ term is due to the relative winding $-2\pi$ of
the $Y_{K_0}$ field.
We have $\tilde N$ kinks of this type associated with the arbitrary choice of
$K_0=(N+1), ..., N_f$.
They are split with generic masses, but become degenerate in the limit (\ref{masssplit}). In this limit they
form the fundamental representation of the global SU$(\tilde N)$ group.
\subsubsection{\boldmath{$P$}-kinks}
\label{subsub4}
In this case all $Y_K$ fields are approximately equal to each other and have
the winding angles $2\pi/\tilde N$. The fields $X_{P\neq P_0}$ are all equal to $\Delta m/\Lambda$,
to the leading order, and
do {\em not} wind; however, they have windings $2\pi/\tilde N$ in the correction terms, see (\ref{Xvev}).
The field $X_{P_0}$ does wind. Its absolute value is $\Delta m/\Lambda$ . The winding angle is
$2\pi$, as enforced by the constraint (\ref{constraint}), see Fig.~\ref{figkinkz}. A more detailed
description of the kink solutions is presented in Appendix B. The windings above imply
the following mass of the $P_0$-kink:
\begin{eqnarray}
M^{\rm BPS}_{P_0}
&\approx&
\left|\frac{N-\tilde N}{2\pi}\,\tilde{\Lambda}_{LE}\left(e^{\frac{2\pi}{\tilde N}\,i}-1\right) -i\left(m_{P_0}-\tilde{m}\right)\right|
\nonumber\\[4mm]
&=&
\left|-i\Delta m+\frac{N-\tilde N}{2\pi}\,\tilde{\Lambda}_{LE}\left(e^{\frac{2\pi}{\tilde N}\,i}-1\right) -i\left(m_{P_0}-m\right)\right|,
\label{ZspectrP}
\end{eqnarray}
where $m$ and $\tilde{m}$ are given by (\ref{m}) and (\ref{tildem}), while
$\Delta m=m-\tilde m$, see (\ref{massdef}).
In the last expression the first term is the leading contribution, the second one is
a correction, while the third term accounts for still smaller splittings.
\begin{figure}
\epsfxsize=8cm
\centerline{\epsfbox{kinkZ.eps}}
\caption{\small Windings of fields $X_{P_0}$ and $Y_K$ for the $P_0$-kink interpolating between $l=0$ and $l=1$ zero-vacua.}
\label{figkinkz}
\end{figure}
There are $N$ kinks of this type associated with the arbitrary choice of
$P_0=1,...,N$.
They are split at generic masses, but become degenerate in the limit (\ref{masssplit}).
They form the fundamental representation of the global SU$(N)$ group in this limit.
Much in the same way as in the $\Lambda$-vacua, the total number
of the kinks interpolating between each pair of the neighboring zero-vacua
is $N_f$.
They form the $(N,1)+(1,\tilde N)$ representation of the global group (\ref{c+fd}) in the
limit (\ref{masssplit}). Note, that the $P$-kinks in the zero-vacua are heavier than the $K$-kinks.
Their masses are given by $\Delta m$ (to the leading order) while the $K$-kink masses
are of the order of $\tilde{\Lambda}_{LE}$. Still, given
small hierarchical masses (\ref{smallmasses}),
all kinks in the zero-vacua are lighter
than
those in the $\Lambda$-vacua which have masses of the order of $\Lambda$, see Eqs.~(\ref{LspectrP})
and (\ref{LspectrK}).
In parallel with the $\Lambda$-vacua, we can verify that Eq.~({\ref{BPSmass}) reproduces
the above BPS spectrum with the appropriate choice of the integration contours. Namely,
for the $P_0$-kink the contour in the $\sigma$ plane encircles the pole at
$\sqrt{2}\sigma=-m_{P_0}$ in the anticlockwise direction, cf. Fig.~\ref{figcontourcp}.
For the $K_0$-kink
the contour encircles the pole at
$\sqrt{2}\sigma=-m_{K_0}$ in the clockwise direction.
\section{Lessons for the bulk theory}
\label{secmoral}
\setcounter{equation}{0}
This section carries a special weight and is, in a sense, central for the present investigation,
since here
we translate our results for the kink spectrum in the weighted two-dimensional CP model
(\ref{wcp}) in the language of strings and confined monopoles of the bulk four-dimensional theory.
We start from the most interesting strong-coupling domain
\begin{equation}
\xi\ll \Lambda
\label{strongcoupling}
\end{equation}
which can be described in terms of weakly-coupled dual bulk theory \cite{SYdual},
see Sec.~\ref{secbulkduality}. At this point we take the limit (\ref{masssplit}) to ensure
the presence of the unbroken global group (\ref{c+fd}).
As was mentioned previously, the elementary non-Abelian strings of the bulk theory correspond
to various vacua of the world-sheet two-dimensional theory, see, e.g. the review paper~\cite{SYrev}
for a detailed discussion. The weighted CP model (\ref{dcp}) has two types of
vacua, namely: $(N-\tilde N)$ $\Lambda$-vacua and $\tilde N$ zero-vacua. The former are not-so-interesting
from the standpoint of the bulk theory. Indeed, they yield just Abelian $Z_{N-\tilde N}$ strings associated with
the winding of the $(N-\tilde N)$ singlet dyons $D^l$ charged with respect to U(1)
factors of the gauge group of
the dual theory (\ref{dualgaugegroup}) \cite{SYdual}, see (\ref{Dvev}).
Moreover,
the weighted CP model (\ref{dcp}) is, in fact, inapplicable in the description of these strings.
This model is an effective low-energy theory which can be used below the scale $\sqrt{\xi}$.
However, the energy scale in the $\Lambda$-vacua is of order of $\Lambda$, i.e. much
larger than $\sqrt{\xi}$ in the domain (\ref{strongcoupling}).
Below we focus on $\tilde N$ zero-vacua which correspond to $\tilde N$ elementary {\em non}-Abelian
strings associated with the winding of the light dyons $D^{lA}$ of the dual bulk theory.
The latter are
charged with respect to both Abelian and non-Abelian factors \cite{SYdual} of the dual
gauge group (\ref{dualgaugegroup}). The energy scale in these vacua
of the world-sheet theory is of the order of
$$
{\rm max}(\Delta m_{KK'},\tilde{\Lambda}_{LE})\,,
$$
which we assume to be much less than $\sqrt{\xi}$. Thus, in these vacua the weighted CP model (\ref{dcp})
can be applied to describe the internal dynamics of the non-Abelian strings of the bulk theory.
The confined monopoles of the bulk theory are
seen as kinks in the world-sheet theory.
The results presented in Sec.~\ref{seckink} demonstrate that in
the weighted CP model
there are $N_f$ elementary kinks interpolating between
the neighboring zero-vacua. More exactly, we found $N$ $P$-kinks with masses (\ref{ZspectrP}) and
$\tilde N$ $K$-kinks with masses (\ref{ZspectrK}). In the limit (\ref{masssplit}) they
form the $(N,1)+(1,\tilde N)$ representations of the global group (\ref{c+fd}).
This means that the total number of stringy mesons $M_A^B$ formed by
the monopole-antimonopole
pairs connected by two different elementary non-Abelian strings (Fig.~\ref{figmeson}) is $N_f^2$.
The mesons $M_P^{P'}$ form the
singlet and $(N^2-1,1)$ adjoint representations of the global group (\ref{c+fd}),
the mesons $M_P^{K}$ and $M_K^{P}$ form bifundamental representations
$(N,\bar{\tilde N})$ and $(\bar{N},\tilde N)$, while the mesons
$M_K^{K'}$ form the singlet and $(1,\tilde N^2-1)$ adjoint representations.
(Here as usual, $P=1,...,N$ and $K=(N+1),...,N_f$.)
All these mesons have masses of the order of $\sqrt{\xi}$, determined by the
string tension
\begin{equation}
T=2\pi\xi\,.
\label{ten}
\end{equation}
They are heavier than the
elementary states, namely, dyons and dual gauge bosons which form
the (1,1), $(N,\bar{\tilde N})$, $(\bar{N},\tilde N)$,
and $(1,\tilde N^2-1)$ representations and have masses $\sim \tilde{g}_2\sqrt{\xi}$.
Therefore, the (1,1), $(N,\bar{\tilde N})$, $(\bar{N},\tilde N)$, and $(1,\tilde N^2-1)$ stringy mesons
decay into elementary states, and we are left with $M_P^{P'}$ stringy mesons
in the representation
$(N^2-1,1)$.
This is exactly what was predicted in \cite{SYdual} from the bulk perspective, see
Sec.~\ref{secbulkduality}. Thus, our world sheet picture nicely matches the bulk analysis.
Note also that the $M_P^{P'}$ stringy mesons with strings corresponding to the
$\Lambda$-vacua of the weighted CP model (the ``$\Lambda$-strings") are heavy and decay into
$M_P^{P'}$ stringy mesons with strings corresponding to
the zero-vacua (the ``zero-strings"). To see that this is indeed the case, observe that
the confined monopoles (i.e. kinks
of the weighted CP model) on the $\Lambda$-string have masses of the order of $\Lambda$,
see (\ref{LspectrP}). Therefore, the $\Lambda$-stringy mesons also have
masses of the order of $\Lambda$. The $M_P^{P'}$ mesons with the zero-strings
are much lighter in the domain (\ref{strongcoupling}). Their masses are of the order of
${\rm max}(\Delta m,\sqrt{\xi})$.
Now, let us discuss yet another match of the world-sheet and bulk pictures.
In Sec.~\ref{seclargemasses} we saw that there are elementary excitations
($T=0$ and $S_A=\delta_{AP_1}-\delta_{AP_2}$) on the
string at weak coupling (this is attainable with large
$|\Delta m_{AB}|$). These excitations would form the adjoint representation
$(N^2-1,1)$ of the global group if the limit (\ref{masssplit}) could
be taken. However, in the strong coupling domain of hierarchically
small masses (\ref{smallmasses}) the kink spectrum is very different, see
Sec.~\ref{seckink}. In particular, no elementary excitations are left: these
states decay
on CMS into a $P_1$-kink plus a $P_2$-antikink, see Eqs.~(\ref{LspectrP}) and (\ref{ZspectrP}).
Since the BPS
spectra of the ${\mathcal N}=(2,2)\;$ two-dimensional theory on the string and
the ${\mathcal N}=2\;$ four-dimensional
bulk theory on the Coulomb branch (at zero $\xi$) coincide \cite{Dorey,DoHoTo,SYmon,HT2},
the decay process above is in one-to-one correspondence with the decay of the bulk states
identified in \cite{SYdual}.
Namely,
the quarks $q^{kP_1}$ (with $k=P_2$ due to
the color-flavor locking) and the gauge bosons present in the bulk theory at weak coupling
decay into monopole-antimonopole pairs. If $\xi$ is small but nonvanishing, the monopoles
and antimonopoles cannot move apart: they are bound together by pairs of
confining strings and form \cite{SYdual}
the $M_{P_2}^{P_1}$ mesons shown in Fig.~\ref{figmeson}. Thus, our results from two dimensions confirm the
decay of the quarks and gauge bosons in the strong-coupling domain of the bulk theory.
This decay
process is a crucial element of our mechanism of non-Abelian confinement.
To explain this in more detail we present a ``phase diagram" of the bulk
theory, see Fig.~\ref{figphdiag}. The vertical and horizontal solid
lines in this figure schematically represent the bulk theory
CMS. The horizontal axis gives the masses
$\Delta m_{PP'}\sim \Delta m_{KK'}$ which we force to be of the same order,
while $\Delta m$ is fixed, $\Delta m \ll \Lambda$. The vertical axis gives
the FI parameter $\xi$. The vertical dashed lines depict CMS of the world-sheet theory.
On these lines the spectrum of the stringy mesons of the bulk theory rearranges itself.
On the solid lines the ``perturbative"
spectrum of the bulk theory rearranges itself. Different domains inside CMS
(where the spectra change continuously) are denoted by capital letters A, B, ..., F.
\begin{figure}
\epsfxsize=10cm
\centerline{\epsfbox{phdiag.eps}}
\caption{\small ``Phase diagram" of the bulk theory. Various domains
are separated by CMS on which physical spectrum is rearranged.}
\label{figphdiag}
\end{figure}
The domains $A$, $B$ and $C$ are at weak coupling in the original U$(N)$ gauge
theory. The elementary (``perturbative") states in these domains are the $q^{kA}$
quarks and gauge bosons
of the U($N$) gauge group. The domains $D$ and $E$ are at weak coupling in the dual theory,
see Sec.~\ref{secbulkduality}. The elementary (``perturbative") states in these domains
are the light dyons
$D^{lA}$ and $D^l$ in Eq.~(\ref{Dvev}) plus the gauge bosons of the dual gauge group (\ref{dualgaugegroup}).
Masses of all these states smoothly evolve with the change of parameters inside these
domains. The domain $F$ is at strong coupling in the dual theory.
With small mass differences in the domain
(\ref{smallmasses}), $N$ flavors of dyons in Eq.~(\ref{Dvev}) decouple, the dual theory becomes
asymptotically free, and at $\xi\ll \tilde{\Lambda}_{LE}$ passes into the strong coupling regime.
The dual gauge group (\ref{dualgaugegroup}) gets broken down to U(1)$^N$ by the
Seiberg--Witten mechanism.
Keeping in mind the limit (\ref{masssplit}), our task is to pass from the domain $A$ into the domain
$D$ and prove that the quarks and gauge bosons of the former domain decay into
the monopole-antimonopole
pairs connected by confining strings in the domain $D$. We do it in three steps.
First,
we pass from the domain $A$ to $C$ at large $\xi$
and then move towards small $\xi$ (preserving large mass differences) inside $C$.
The original bulk theory is in the weak-coupling regime in these domains and the quarks and gauge boson spectra
evolve smoothly.
Next,
we pass from the domain $C$ to the domain $E$ at small (or vanishing) $\xi$. Here we
use correspondence between the
BPS spectra of the bulk and world-sheet theories. The
$q^{kP_1}$ quarks (with $k=P_2$ due to the
color-flavor locking) and gauge bosons of the domain $C$ correspond to elementary states
with $T=0$ and $S_A=\delta_{AP_1}-\delta_{AP_2}$, see (\ref{elmass}). As was already
explained, these states decay on CMS of the two-dimensional theory into a
$P_1$-kink and a $P_2$-antikink, interpolating between the $\Lambda$-vacua, see (\ref{LspectrP}).
Since the spectra
of the massive BPS states in the bulk (at $\xi=0$) on the one hand, and in the world-sheet theory
on the other hand, are identical, both the
quarks and gauge bosons from domain $C$ do not exist in the domain $E$. They decay
into a $P_1$-monopole and a $P_2$-antimonopole.
At small but nonvansishing $\xi$ the latter
states are confined by $\Lambda$-strings which gives rise to $M_{P_2}^{P_1}$ stringy mesons.
Furthermore, as we pass from the domain $E$ to our final destination -- the domain $D$ -- these $M_{P_2}^{P_1}$
mesons ``glued" by $\Lambda$-strings, decay on the world-sheet theory CMS located
at $\Delta m_{PP'}\sim \Delta m_{KK'}\sim \tilde{\Lambda}_{LE}$.
They decay into lighter $M_{P_2}^{P_1}$ mesons glued by zero-strings. The latter stringy mesons were
absent in the domain
$E$ at intermediate masses (see (\ref{kinkspectrim})), but emerge in the domain $D$
at small masses, see (\ref{ZspectrP}). This completes our proof.
In conclusion it is worth noting that
the stringy mesons $M_P^{P'}$ in the adjoint representation $(N^2-1,1)$
of the global group
are metastable, strictly speaking. An extra monopole-antimonopole pair can be
created on the string, making the $(N^2-1,1)$ meson to decay into a pair of
stringy mesons in the bifundamental representation, $(N,\bar{\tilde N})$ and $(\bar{N},\tilde N)$.
During the subsequent stage these stringy bifundamentals decay into elementary
bifundamental dyons. One can suppress the rate of this decay, however. If we
keep $\Delta m\ll \sqrt{\xi}$ and take a limit of large $N$, while $\tilde N$ is fixed
this decay rate is of the order of $\tilde N/N\ll 1$. To see that this indeed the case, note that the above
decay process goes through creation of a monopole-antimonopole pair from
the fundamental representation of SU($\tilde N$) on the string which selects
$\tilde N$ channels out of $N_f$.
Similar considerations can be applied to the weak-coupling domain of large
$\xi$,
$$\sqrt{\xi}\gg \Lambda\,,
$$ of the bulk theory. We still have kinks
in the $(N,1)+(1,\tilde N)$ representations of the global group (\ref{c+f})
in the limit (\ref{masssplit}). Thus, all types of mesons $M^A_B$ are formed
with masses $\sim\sqrt{\xi}$.
However, in this regime the
(1,1), $(N,\bar{\tilde N})$, $(\bar{N},\tilde N)$, and $(N^2-1,1)$ stringy mesons
can decay into quarks and gauge bosons
with the same quantum numbers and masses $\sim g_2^2\sqrt{\xi}$. Therefore,
we are left with elementary states and stringy adjoint mesons $(1,\tilde N^2-1)$.
The latter are metastable and decay in pairs
of bifundamental states.
\section{Conclusions}
\label{concl}
Both ${\mathcal N}=2\;$ four-dimensional theories belonging to the dual pair discussed above
support non-Abelian strings. The world-sheet theory on the strings is given by the weighted CP
model which appears in two varieties depending on which side of duality we are. These two
weighted CP
models also form a dual pair. We explore the kink spectra
(which represent confined monopoles of the bulk theory) and their evolution
in passing through the crossover domain. Of most interest is
small-$\xi$ dynamics. In fact, at small $\xi$ we find {\em two} weak coupling subdomains and
a strong-coupling one depending on the values of the differences of the mass parameters
$\Delta m_{AB}$.
We have shown that in the limit (\ref{masssplit}) where the global group
(\ref{c+f}) is unbroken confined monopoles form the fundamental representation
of the global group. Therefore, stringy mesons
(shown in Fig.~\ref{figmeson}) formed by pairs of monopoles and antimonopoles
belong to the adjoint or singlet representations of the global group. This nicely matches
global quantum numbers of mesons in the ``real world."
We proved the statement proposed in \cite{SYdual} that quarks and gauge
bosons present in the original theory at large $\xi$ decay on CMS into
monopole-antimonopole pairs confined by non-Abelian strings as we enter the small-$\xi$ domain.
This result is a crucial element of our mechanism of non-Abelian confinement.
In summary, in this paper we used the world-sheet theory to confirm the picture of non-Abelian
confinement obtained in \cite{SYdual} from the bulk perspective. Non-Abelian confinement
is {\em not} associated with formation of chromoelectric strings connecting quarks, as
a naive extrapolation of the Abelian confinement picture suggests.
Rather, it is due to the decay on CMS of the Higgs-screened quarks and
gauge bosons into monopole-antimonopole pairs confined by non-Abelian strings
in the strong coupling domain of small $\xi$. We stress again that the non-Abelian strings confine
monopoles
both in the original and dual theories.\footnote{ Similar results were recently obtained
in \cite{MarsY} for ${\mathcal N}=2\;$ supersymmetric QCD with $N_f>2N$.}
Analysis of the mass spectra presented in the bulk of the paper raises a number of intriguing questions.
One of them refers to typical sizes of the objects considered. At the moment, in the absence of a detailed analysis,
one can address this issue only at the qualitative level. At small $\xi$ and $\Delta m
\ll \sqrt{\xi}\ll \Lambda$
we expect that
the smallest size $\sim \Lambda^{-1}$ is that of the elementary (``perturbative") states,
namely dyons from Eq.~(\ref{Dvev}) and dual gauge bosons.
The sizes of the stringy mesons in the representation $N^2-1$
are expected to be of the order $(\tilde g \sqrt{\xi})^{-1}$, i.e of order of the thickness of the non-Abelian strings.
Finally, the largest sizes $(\Delta m)^{-1}$ and $\sim \tilde\Lambda_{LE}^{-1}$
belongs to the $P$- and $K$-kinks (confined monopoles on a string), respectively, provided that the mass difference
$\Delta m_{KK^\prime} \to 0$.
\section*{Acknowledgments}
We are grateful to A. Gorsky for valuable discussions.
This work is supported in part by DOE grant DE-FG02-94ER408.
The work of A.Y. was supported
by FTPI, University of Minnesota,
by RFBR Grant No. 09-02-00457a
and by Russian State Grant for
Scientific Schools RSGSS-11242003.2.
\vspace{1cm}
\section*{Appendix A: \\
Kink solutions in \boldmath{CP$(N-1)\;$} model}
\renewcommand{\thesection.\arabic{equation}}{A.\arabic{equation}}
\setcounter{equation}{0}
\renewcommand{\thesubsection}{A.\arabic{subsection}}
\setcounter{subsection}{0}
In this appendix we discuss in more detail the kink solutions in
the $\mbox{CP$(N-1)\;$}$ model (see Sec.~\ref{seccpnkinks}).
For simplicity we set $\Delta m_{PP'}=0$. Small mass differences will
slightly deform the kink profile functions but will not change the very
fact of its existence. The kink masses are given by differences
of the mirror superpotential evaluated at the initial and final vacua, see (\ref{cpspectr}).
We look for the kink solution interpolating between $l=0$ and $l=1$ vacua
using the {\em ansatz}
\begin{equation}
X_{P\neq P_0}=r\,e^{i\theta},\qquad X_{P_0}=\frac{e^{-i(N-1)\theta}}{r^{(N-1)}}\,,
\label{cpansatz}
\end{equation}
where $r(z)$ and $\theta(z)$ are kink profile functions, subject to boundary
conditions
\begin{eqnarray}
\theta(z=-\infty)
&=&
0,\qquad r(z=-\infty)=1,
\nonumber\\[3mm]
\theta(z=\infty)
&=&
\frac{2\pi}{N},\qquad r(z=\infty)=1,
\label{cpbc}
\end{eqnarray}
see (\ref{xvevcp}). The last expression in (\ref{cpansatz}) is dictated by
(\ref{constraintcp}).
The explicit form of the kink profile functions depends on the form of the kinetic term
which is not known. Therefore, both profile functions $r(z)$ and $\theta(z)$
cannot be determined. Still we can obtain the kink solution up to a single unknown
profile function $\theta(z)$ expressing $r$ as a function of $\theta$.
To this end we exploit the fact that the kink trajectory in the complex plane of superpotential
goes along the straight line connecting the points $W_{\rm mirror}(l=0)$ with
$W_{\rm mirror}(l=1)$ \cite{ChVa}. The difference
\begin{eqnarray}
&&
W_{\rm mirror}(l=1)-W_{\rm mirror}(l=0)=
-\frac{N\Lambda}{4\pi}\left(e^{\frac{2\pi}{N}\,i}
-1\right)
\nonumber\\[3mm]
&&
=-\frac{N\Lambda}{4\pi}\,e^{\frac{\pi}{N}\,i}\,2i\sin{\left(\frac{\pi}{N}\right)},
\label{cpdiff}
\end{eqnarray}
where we used (\ref{mirrorcp}) and (\ref{xvevcp}).
\vspace{2mm}
On the other hand,
\begin{eqnarray}
&&
W_{\rm mirror}(\theta)-W_{\rm mirror}(\theta=0)=
-\frac{\Lambda}{4\pi}\left[ (N-1)re^{i\theta}+\frac1{r^{N-1}}e^{-i(N-1)\theta}-N\right]
\nonumber\\[3mm]
&&
=-\frac{r\Lambda}{4\pi}\,e^{\frac{\pi}{N}\,i}\left\{(N-1)e^{i(\theta-\frac{\pi}{N})}
+\frac1{r^{N}}e^{-i(N-1)\theta-i\frac{\pi}{N}}-\frac{N}{r}e^{-\frac{\pi}{N}}\right\}.
\label{cpdiffp}
\end{eqnarray}
Comparing Eqs.~(\ref{cpdiff}) and (\ref{cpdiffp}) we see that for the kink
trajectory to go along the straight line, the real part of the expression in the curly brackets
in (\ref{cpdiffp}) must vanish. As a result,
\begin{equation}
(N-1)\cos{\left(\theta-\frac{\pi}{N}\right)}-\frac{N}{r}\cos{\frac{\pi}{N}}
+\frac1{r^{N}}\cos{\left[(N-1)\theta+\frac{\pi}{N}\right]}=0.
\label{cpkinkeq}
\end{equation}
This equation determines the function $r(\theta)$. It has
a nonvansihing nonsingular solution at $0\leq \theta\leq 2\pi/N$ which satisfies
the boundary conditions $r(2\pi/N)=r(0)=1$. Say, for $N=2$, $r=1$. For large $N$
the profile function $r$ is approximately given by
\begin{equation}
r(\theta)\approx 1+\frac1{N}\left\{1-\cos{\left[(N-1)\theta+\frac{\pi}{N}\right]}
\right\}+\cdots .
\label{largeNrcp}
\end{equation}
\section*{Appendix B: \\
Kink solutions in the zero-vacua}
\renewcommand{\thesection.\arabic{equation}}{B.\arabic{equation}}
\setcounter{equation}{0}
Here we consider the
kink solutions interpolating between the zero-vacua with $l=0$ and $l=1$,
see (\ref{Yvev}), at hierarchically small masses (\ref{smallmasses}). The
$K$-kink solutions in the zero-vacua are quite similar to kinks in the CP$(N-1)\;$ model.
Therefore, we focus on $P$-kinks.
\begin{figure}
\epsfxsize=8cm
\centerline{\epsfbox{kinkeq.eps}}
\caption{\small The right- and left-hand sides of Eq.~(\ref{kinkeq}). The closed circle denotes the regular
solution. }
\label{figkinkeq}
\end{figure}
Analyzing the $P$-kinks we
assume the limit (\ref{masssplit}) (for simplicity). The {\em ansatz} for the kink
solution takes the form
\begin{eqnarray}
Y_{K}
&=&
r\,e^{i\theta}\,\,\frac{\tilde{\Lambda}_{LE}}{\Lambda}\,,
\nonumber\\[3mm]
X_{P\neq P_0}
&\approx&
\frac{\Delta m}{\Lambda},\qquad \quad \,\,\,
X_{P_0}\approx\frac{\Delta m}{\Lambda}\,r^{\tilde N}e^{i\tilde N \theta}\,,
\label{ansatz}
\end{eqnarray}
where we restrict ourselves to the leading order contributions ($\sim \Delta m/\Lambda$)
ignoring next-to-leading $O(\tilde{\Lambda}_{LE}/\Lambda )$ terms. The profile functions
$r(z)$ and $\theta(z)$ are subject to the boundary conditions
\begin{eqnarray}
\theta(z=-\infty)
&=&
0,\qquad r(z=-\infty)=1\,,
\nonumber\\[3mm]
\theta(z=\infty)
&=&
\frac{2\pi}{\tilde N},\qquad r(z=\infty)=1\,,
\label{bc}
\end{eqnarray}
see (\ref{Yvev}) and (\ref{Xvevzero}). The last expression in (\ref{ansatz}) is dictated by
(\ref{constraint}).
The superpotential
(\ref{mirror}) then gives
\begin{eqnarray}
&&
W_{\rm mirror}(\theta)-W_{\rm mirror}(\theta=0)
\nonumber\\[3mm]
&&
\approx-\frac{\Delta m}{4\pi}\left\{r^{\tilde N}e^{i\tilde N \theta}
-\tilde N\,(i\theta +\ln{r})-1\right\}\,,
\label{supoftheta}
\end{eqnarray}
while the difference of the superpotential in the initial and final vacua is
\begin{equation}
W_{\rm mirror}(l=1)-W_{\rm mirror}(l=0)\approx
\frac{\Delta m}{4\pi}\times 2\pi i\,.
\label{supofthetap}
\end{equation}
For the kink trajectory to go along the straight line in the complex plane of
superpotential, the real part of the expression in the curly brackets
in (\ref{supoftheta}) must vanish. This requirement implies
\begin{equation}
r^{\tilde N}\cos{(\tilde N \theta)}-\tilde N\ln{r}=1\,.
\label{kinkeq}
\end{equation}
The latter equation always has a finite nonvanishing solution
in the interval $0\leq \theta\leq 2\pi/N$, subject to the
boundary conditions $r(2\pi/N)=r(0)=1$. To
check that this is indeed the case
we rewrite it as
\begin{equation}
\cos{(\tilde N \theta)}=\frac{1+\tilde N\ln{r}}{r^{\tilde N}}\,.
\label{dop9}
\end{equation}
The right- and left-hand sides of (\ref{dop9}) are schematically plotted
in Fig.~\ref{figkinkeq}. For any $-1<\cos{(\tilde N \theta)}<1$ we have only one
nonsingular solution $r(\theta)$. In particular, at $\tilde N\gg 1$
\begin{equation}
r(\theta)\approx 1-\frac1{\tilde N}[1-\cos{(\tilde N \theta)}]+\cdots\,.
\label{dop9p}
\end{equation}
\newpage
\small
|
\section{Introduction}
\noindent
A $K$-quasiconformal mapping is an orientation preserving homeomorphism $f:\Omega\to\Omega'$ between domains $\Omega,\Omega'\subset\mathbb R^n$ that belongs to the Sobolev space $W^{1,n}_{loc}(\Omega;\Omega')$ and satisfies the distortion inequality
\begin{equation}
\max_{|\xi|=1}|\partial_\xi f(x)|\leq K\,\min_{|\xi|=1}|\partial_\xi f(x)|
\end{equation}
at almost every point $x\in\Omega$. If $K=1$, then $f$ is indeed a conformal mapping. If one does not require $f$ to be a homeomorphism, then we simply say that $f$ is $K$-quasiregular. For more background on these mappings, see the monograph \cite{astalaiwaniecmartin}.\\
\\
In the planar setting ($n=2$), Astala's Theorem \cite{astalaareadistortion} solved the long standing Gehring-Reich conjecture on the area distortion of quasiconformal mappings, namely
\begin{equation}\label{area}
|\phi(E)|\leq C_K\,|E|^{1/K},
\end{equation}
where $\phi:\Omega\to\Omega'$ is a conveniently normalized $K$-quasiconformal mapping between planar domains, and $E$ is a measurable subset of $\Omega$. Other related questions, like the optimal integrability and the sharp Hausdorff dimension distortion, were solved as a consequence of \eqref{area}. In particular, one has
\begin{equation}\label{dimension}
\frac{1}{\dim\phi(E)}-\frac12\leq\frac{1}{K}\left(\frac{1}{\dim E}-\frac12\right).
\end{equation}
Furthermore, in \cite{astalaareadistortion} it was shown that equality can be attained for some sets $E$ and mappings $\phi$.\\
\\
In the last years renewed interest has arisen in connection with these questions, and deep advances have been made, improving the above inequality in several directions. The sharp
quasiconformal distortion of Hausdorff contents $\mathcal M^ t$, $0<t<2$, was obtained by Lacey, Sawyer and Uriarte-Tuero \cite{Lacey-Sawyer-Uriarte}. They showed that if $E$ is contained in some ball $B$, $0<t<2$, and $t'=\frac{2Kt}{2+(K-1)t}$, then
\begin{equation}\label{contents}
\frac{\mathcal M^{t'}(\phi(E))}{\operatorname{diam}(\phi(B))^{t'}} \leq C(K) \,\left(\frac{\mathcal M^t(E)}{\operatorname{diam}(B)^t}\right)^{\frac{t'}{Kt}},
\end{equation}
which in particular proves the following implication about the corresponding Hausdorff measures
$\mathcal H^t$:
\begin{equation}\label{zerotozero}
\mathcal H^t(E)=0\hspace{1cm}\Rightarrow\hspace{1cm}\mathcal H^{t'}(\phi(E))=0.
\end{equation}
This extends \eqref{dimension}, and answers in the affirmative a conjecture by Astala \cite{astalaareadistortion}. Previously, in \cite{ACMOU}, the particular case $t'=1$ had been solved. In any case, notice that from \rf{contents} it is not clear if $\mathcal H^{t'}(\phi(E))<\infty$ whenever
$\mathcal H^t(E)<\infty$.\\
\\
The optimal quasiconformal distortion of analytic capacity has also been a topic of deep research (see for instance \cite{astalaareadistortion} or \cite{ACMOU}). In a recent joint work of Tolsa and Uriarte-Tuero \cite{Tolsa-Uriarte}, it is shown that, for $K>1$,
\begin{equation}\label{gammavsriesz}
\gamma(\phi(E))\leq C_K\,\,\left(\dot {\mathcal C}_{\frac{2K}{2K+1},\frac{2K+1}{K+1}}(E)\right)^\frac{K+1}{2K},
\end{equation}
where $\dot{\mathcal C}_{\alpha,p}$ is the classical Riesz capacity of nonlinear potential theory (see \rf{defcap}), and $\gamma$ denotes the analytic capacity. This estimate has remarkable consequences in the determination of removable sets for bounded $K$-quasiregular mappings. For the holomorphic case, see \cite{tolsasemiadditivityanalyticcapacity} and \cite{davidunrectifiable1setszeroanalyticcapacity}. To get \eqref{gammavsriesz}, the authors first show a sharp bound for the distortion of a Hausdorff content $\mathcal M^h$ (see \cite[Lemma 2.11]{Tolsa-Uriarte}), where $h$ is a gauge function which
is not invariant under translations. As a matter of fact, it turns out that
Riesz capacities can be recovered as a supremum of Hausdorff contents $\mathcal M^h$ with $h$ running within some precise class. This allowed the authors to prove a second estimate, now concerning quasiconformal distortion of Riesz capacities. More precisely, for each $q>1$, they showed that
\begin{equation}\label{1dimensriesz}
\dot {\mathcal C}_{\frac1q,q}(\phi(E))\leq C_{K,q}\,\,\dot {\mathcal C}_{\alpha,p}(E)^\frac{K+1}{2K}
\end{equation}
where $p=1+\frac{2K}{K+1}(q-1)$ and $2-\alpha p=\frac{2}{K+1}$. Note that the left hand side is a $1$-dimensional quantity for every $q>1$.
\\
\\
In the present paper, we extend \eqref{1dimensriesz} to all other indices $\alpha,p$ and obtain a general version for the quasiconformal distortion of all Riesz capacities.
\begin{theorem} \label{teocap2}
Let $1<q<\infty$ and $0<\beta q<2$. Let $t'=2-\beta q$, and $t$ be such that
$$
\frac1t-\frac12 = K\left(\frac{1}{t'}-\frac12\right).
$$
Let $E\subset\mathbb C$ be compact, and let $\phi:\mathbb C\to\mathbb C$ be a $K$-quasiconformal map. If $E$ is contained in a ball $B$, then
\begin{equation}\label{eq11}
\frac{\dot{{\mathcal C}}_{\beta,q}(\phi(E))}{\operatorname{diam}(\phi(B))^{t'}}\leq C(\beta,q,K)\,\left(\frac{\dot{{\mathcal C}}_{\alpha,p}(E)}{\operatorname{diam}(B)^t}\right)^\frac{t'}{Kt}
\end{equation}
where
$$
p=1+\frac{Kt}{t'}\,(q-1)\hspace{1cm}\text{and}\hspace{1cm} 2-\alpha p=t.
$$
The constant in \rf{eq11} depends only on $\beta$, $q$, $K$.
\end{theorem}
\noindent
This result follows by combining some of the ideas from \cite{Tolsa-Uriarte} with others from
\cite{Lacey-Sawyer-Uriarte}. Following this approach, in Lemma \ref{lemkgran} below one obtains distortion estimates in terms of the $h$-contents $\mathcal M^h$, with $h$ of the form $h(B(x,r))=r^t\,
{\varepsilon}(B(x,r)),$ with $0<t<2$ and ${\varepsilon}(\cdot)$ satisfying some appropriate conditions.
This result extends \cite[Lemma 2.11]{Tolsa-Uriarte} (which only dealt with the case $t'=1$).
Theorem \ref{teocap2} is a direct consequence of the distortion estimates in terms of $h$-contents.
\\
\\
A second main result that we establish, using the $h$-contents $\mathcal M^h$, is the following distortion theorem involving Hausdorff measures.
\begin{theorem} \label{teopri}
Let $0<t<2$ and denote $t' =\frac{2Kt}{2+(K-1)t}$. Let $\phi:\mathbb C\to\mathbb C$ be $K$-quasiconformal. For any ball $B$ and any compact set $E\subset B$, we have
\begin{equation}\label{eqhaus}
\frac{\mathcal H^{t'}(\phi(E))}{\operatorname{diam}(\phi(B))^{t'}} \leq C(K) \,\left(\frac{\mathcal H^t(E)}{\operatorname{diam}(B)^t}\right)^{\frac{t'}{Kt}}.
\end{equation}
In particular, if $\mathcal H^t(E)$ is finite, then also $\mathcal H^{t'}(\phi(E))$ is finite.
\end{theorem}
\noindent
Notice that \rf{eqhaus} is the estimate that one gets replacing Hausdorff contents by Hausdorff measures in \rf{contents}.
This result may seem somewhat surprising, because in the arguments used in
\cite{Lacey-Sawyer-Uriarte} (and in \cite{ACMOU} in the case $t'=1$) to prove \rf{contents} it
is essential that one works with Hausdorff contents, and not with Hausdorff
measures: for instance, many estimates
in \cite{Lacey-Sawyer-Uriarte} and \cite{ACMOU} involve some packing conditions which hold for Hausdorff
contents but not for Hausdorff measures.\\
\\
Let us remark that Theorem \ref{teopri} was proved in \cite{Tolsa-preprintqc}. However, since this result follows
using the distortion estimates for $h$-contents in Lemma \ref{lemkgran}, X. Tolsa has preferred to include this in the present paper (see Section \ref{tolsapreprint}).\\
\\
An immediate corollary of Theorem \ref{teopri} is the following.
\begin{coro}\label{teosigma}
Let $E\subset\mathbb C$ be compact and $\phi:\mathbb C\to\mathbb C$ a $K$-quasiconformal map.
If $\mathcal H^t(E)$ is $\sigma$-finite for some $0<t<2$, then $\mathcal H^{t'}(\phi(E))$ is $\sigma$-finite for $t'=\frac{2Kt}{2+Kt-t}$.
\end{coro}
As usual, the letters $c,C$ denote constants (often, absolute constants) that may change
at different occurrences, while constants with a subscript, such as $C_1$,
retain their values. The notation $A\lesssim B$ means that there is a positive constant $C$ such that $A\leq CB$, and $A\simeq B$ means that $A\lesssim B\lesssim A$.
\bigskip
\section{Measures, gauge functions and Hausdorff contents} \label{sec1}
\subsection{Strategy for the proof of Theorem \ref{teocap2}}\label{strategy}
\noindent
To motivate the introduction of the $h$ contents below, we will describe
the main ideas in the proof of Theorem \ref{teocap2}. Recall that the homogeneous Riesz capacity $\dot{\mathcal C}_{\alpha,p}$ is defined as
\begin{equation}\label{defcap}
\dot{\mathcal C}_{\alpha,p}(E)=\inf\{\|g\|_{L^p(\mathbb C)}; g\in L^p(\mathbb C),\,I_\alpha\ast g\geq\chi_E \},
\end{equation}
where $I_\alpha(z)=\frac{1}{|z|^{2-\alpha}}$ is the usual planar Riesz kernel of order $\alpha$. Further, by Wolff's Theorem (see for instance \cite{adamshedberg}), it turns out that
$$
\dot{\mathcal C}_{\alpha,p}(E)\simeq\sup\{\mu(E); \operatorname{supp}(\mu)\subset E, \dot{W}^{\mu}_{\alpha,p}(z)\leq 1\,\,\forall z\in\mathbb C\}
$$
where
$$\dot{W}^\mu_{\alpha,p}(z)=\int_0^\infty\left(\frac{\mu(B(z,r))}{r^{2-\alpha p}}\right)^{p'-1}\frac{dt}{t}$$
is the homogeneous Wolff potential of $\mu$. In this paper, we prove that $\dot{\mathcal C}_{\alpha,p}(E)$ coincides with the following supremum of generalized Hausdorff contents, modulo multiplicative constants:
$$
\dot{{\mathcal C}}_{\alpha,p}(E)
\simeq\sup\left\{\!\mathcal M^h(E): h(x,r)=r^{2-\alpha\,p}\,{\varepsilon}(x,r),\;{\varepsilon}\in{\mathcal G}_2,\int_0^\infty\!\left(\frac{h(x,r)}{r^{2-\alpha\,p}}\right)^{p'-1}\frac{dr}{r}\leq 1\!
\right\}
$$
(see Subsection \ref{Gp} for the definition of ${\mathcal G}_2$). Then the problem is reduced to see how quasiconformal mappings distort the generalized Hausdorff contents $\mathcal M^h$ whenever $h$ is an admissible gauge function for $\dot{{\mathcal C}}_{\alpha,p}$.
In particular, to each finite Borel measure $\mu$ with bounded Wolff potential, we can associate a gauge function $h=h_\mu$ admissible for the Riesz capacity. Actually, in
the supremum above one may restrict to such gauge functions $h_\mu$. This fact is very useful because the Hausdorff measures $\mathcal H^{h_\mu}$ and contents $\mathcal M^{h_\mu}$ can be seen as \emph{regularized} versions of $\mu$, whence easier to work with. Such gauge functions were already introduced in \cite{Tolsa-Uriarte}, but for the reader's convenience we remember their definition and main properties in the next subsections.
\noindent
\subsection{The gauge functions $h_{\mu,a,t}$}
Let $0<t<2$ and $a>0$ be fixed parameters. We consider the function
\begin{equation}\label{eqpsia}
\psi_{a,t}(x) = \frac1{|x|^{t+a}+1},\qquad x\in\mathbb C.
\end{equation}
Given a compactly supported finite Borel measure $\mu$, let us define for every ball $B=B(x,r)$ the gauge function
\begin{equation}\label{defhx}
h(x,r)=h_{\mu,a,t}(x,r) = h_{\mu,a,t}(B) =\int \psi_{a,t}\left(\frac{|x-y|}{r}\right)\,d\mu(y).
\end{equation}
Now, for our fixed $a,t$, set
\begin{equation}\label{defvex}
{\varepsilon}(x,r)={\varepsilon}_{\mu,a,t}(x,r)= \frac{h_{\mu,a,t}(x,r)}{r^ t}=\frac1{r^t} \int \psi_{a,t}\Bigl(\frac{y-x}r\Bigr)d\mu(y),
\end{equation}
so that $h_{\mu,a,t}(x,r)=r^t\,{\varepsilon}_{\mu,a,t}(x,r)$. One should view $t$ as a dimensional parameter,
while the role of $a$ is to provide enough decay at $\infty$ of $\psi_{a,t}$.
Notice that, by construction,
$\mu(B)\leq 2\,h_{\mu,a,t}(B)$, and that $h_{\mu,a,t}(B)$ can be seen as a
smooth version of $\mu(B)$. Similarly, ${\varepsilon}_{\mu,a,t}(B)$ is a kind of smooth substitute of the
$t$-dimensional density $\theta_\mu^t(B)=\mu(B)/r(B)^t$. One of the advantages of
${\varepsilon}_{\mu,a,t}(x,r)$ over $\theta_\mu^t(B(x,r))$ is that, for $C=2^a$,
$${\varepsilon}_{\mu,a,t}(x,2r)\leq C{\varepsilon}_{\mu,a,t}(x,r)$$
for any $x$ and $r>0$, which fails in general for $\theta_\mu^t(x,r)$. Analogously, we have
$$h_{\mu,a,t}(x,2r)\leq C\,h_{\mu,a,t}(x,r),$$
for $C=2^{at}$,
while $\mu(B(x,r))$ and $\mu(B(x,2r))$ may be very different.
\bigskip
\subsection{The measures $\mathcal H^h$ and the contents $\mathcal M^h$}\label{Hh} Let $\mathcal B$ denote the family of all closed balls contained in $\mathbb C$, and let ${\varepsilon}:\mathcal B\to[0,\infty)$ be any function defined on $\mathcal B$. We set ${\varepsilon}(x,r)={\varepsilon}(B(x,r))$, and we define $h(x,r)={\varepsilon}(x,r)\,r^t$. We assume that
$$\lim_{r\to 0}h(x,r)=0\hspace{1cm}\text{ for all }x\in\mathbb C.$$
We introduce the measure $\mathcal H^h$ following Carath\'eodory's construction (see \cite{mattila}, p.54): given $0<\delta\leq\infty$ and a set $F\subset\mathbb C$,
$$ \mathcal H^h_\delta(F) = \inf\sum_i h(B_i),$$
where the infimum is taken over all coverings $F\subset \bigcup_i B_i$ with balls $B_i$ with radii smaller that $\delta$. Finally, we
define
$$ \mathcal H^h(F) = \lim_{\delta\to0} \mathcal H^h_\delta(F).$$
The above limit exists, because $\mathcal H^h_\delta(F)$ is a non-increasing function of $\delta$. For $\delta=\infty$, we obtain the $h$-content, and we simply write $\mathcal M^h(E)=\mathcal H_\infty^h(E)$. Recall also that $ \mathcal H^h$ is a Borel regular measure (see \cite{mattila}), although it is not a ``true'' Hausdorff measure. It is clear that $\mathcal M^h(F)\leq\mathcal H^h(F)$. On the other hand, the implication
$$\mathcal M^h(F)=0\hspace{1cm}\Rightarrow\hspace{1cm}\mathcal H^h(F)=0$$
also holds if the function $r\mapsto h(x,r)$ is non-decreasing for all $x\in F$. This is the case, for instance, if $h=h_{\mu,a,t}$ for certain measure $\mu$.
\begin{lemma}\label{Mmu}
For any Borel set $A$, we have
$$\mu(A)\leq 2\,\mathcal M^{h_{\mu,a,t}}(A).$$
\end{lemma}
\begin{proof}
Given any $\eta>0$, consider a covering $A\subset \bigcup_i B_i$ by balls so that
$$\sum_i h_{\mu,a,t}(B_i) \leq \mathcal M^{h_{\mu,a,t}}(A) +\eta.$$
Since $\mu(B_i)\leq 2 h_{\mu,a,t}(B_i)$, we have
$$\mu(A)\leq \sum_i \mu(B_i)\leq 2\sum_i h_{\mu,a,t}(B_i) \leq 2 \mathcal M^{h_{\mu,a,t}}(A) + C\eta.$$
\end{proof}
\subsection{The families $\mathcal G_1$ and $\mathcal G_2$}
\label{Gp}. We say that the function ${\varepsilon}:\mathcal B\to[0,\infty)$ belongs to $\mathcal G_1$ if there exists a constant $C_2$ such that
\begin{equation}\label{eqeq1}
C_2^{-1}\,{\varepsilon}(x,r)\leq {\varepsilon}(y,s)\leq C_2\,{\varepsilon}(x,r)
\end{equation}
whenever $|x-y|\leq 2r$ and $r/2\leq s\leq 2r$. Note that \rf{eqeq1} also holds with a different constant $C_2$ if one assumes $|x-y|\leq Cr$ and $C^{-1}r\leq s\leq Cr$, by applying \rf{eqeq1} finitely many
times. It is easy to check that every ${\varepsilon}_{\mu,a,t}$ belongs to $\mathcal G_1$, due to the properties of the function $\psi_{a,t}$.\\
\\
\noindent
It was noticed in \cite{Tolsa-Uriarte} that if ${\varepsilon}\in{\mathcal G}_1$ then Frostman's Lemma holds for $\mathcal M^h$, where $h(x,r)=r^t\,{\varepsilon}(x,r)$:
\begin{lemma}\label{frostman}
If ${\varepsilon}\in\mathcal G_1$ and $h(x,r)=r^t{\varepsilon}(x,r)$, then Frostman's Lemma holds for $ \mathcal M^h$. That is,
given a compact set $F\subset \mathbb C$, then $ \mathcal M^h(F)>0$ if and only if
there exists a Borel measure $\nu$ supported on $F$ such that $\nu(B)\leq h(B)$ for any ball $B$. Furthermore, we can find $\nu$ so that $\nu(F)\geq c^{-1} \mathcal M^h(F)$.
\end{lemma}
\noindent
The proof is almost the same as the one of the usual Frostman's Lemma (for instance, see \cite{mattila}, p.112), taking into account the regularity properties of the gauge functions $h\in\mathcal G_1$.\\
\\
Now we introduce the class $\mathcal G_2$. For each fixed $0<t<2$, the class $\mathcal G_2=\mathcal G_2(t)$ consists on functions ${\varepsilon}={\varepsilon}(x,r)$ such that
\begin{equation}\label{eqsum4}
\sum_{k\geq0} 2^{-k(2-t)}\,{\varepsilon}(x,2^kr)\leq C_3 \,{\varepsilon}(x,r), \text{ for all }x\in\mathbb C,
\end{equation}
for certain constant $C_3>0$. Notice that \eqref{eqsum4} is equivalent to saying that
$$\int_r^\infty\frac{\varepsilon(x,s)}{s^{2-t}}\,\frac{ds}{s}\leq C_3\,\frac{\varepsilon(x,r)}{r^{2-t}},$$
or in terms of $h(x,r)=r^t\,{\varepsilon}(x,r)$,
$$\int_r^\infty\frac{h(x,s)}{s^{2}}\,\frac{ds}{s}\leq C_3\,\frac{h(x,r)}{r^{2}}.$$
Observe that this estimate does not hold for the area, that is, for $h(x,r)=r^2$, neither for gauges $h$ too close to $r^2$, like $h(r)=r^2\log\frac1r$.
\begin{lemma}\label{emug2}
Let $a>0$, $\alpha,\beta>0$ and $m=\min(\alpha,\beta)$. If $\alpha\neq \beta$, then for all $x>0$ we have
\begin{equation}\label{lemtec1}
\sum_{k\geq 0} \frac{2^{-\beta k}}{\bigl(2^{-k}x\bigr)^{\alpha} + 1}\leq \frac C{x^{m} + 1},
\end{equation}
with $C$ depending only on $\alpha, \beta$. As a consequence, if $\mu$ is a finite Borel measure, then ${\varepsilon}_{\mu,a,t}\in \mathcal G_2(t)$ whenever $0<a<2-t$.
\end{lemma}
\begin{proof}
The estimate \eqref{lemtec1} is just a numerical inequality which can be proved by splitting the
sum according to whether $2^{-k}x\geq 1$ or $2^{-k}x< 1$, and then approximating the denominator
inside the sum by $2^{-k}x$ in the first case and by $1$ in the second, for instance.
We skip the details.
To deal with the last
statement, we just have to combine \eqref{lemtec1} with the definitions above to get
$$\aligned
\sum_{k=0}^\infty \frac{{\varepsilon}_{\mu,a,t}(x,2^kr)}{2^{(2-t)k}}
&=r^{-t}\sum_{k=0}^\infty \frac{h_{\mu,a,t}(x,2^kr)}{2^{2k}}= r^{-t}\,\int \sum_{k=0}^\infty 2^{-2k}\,\psi_{a,t}\left(2^{-k}\,\frac{|z-x|}{r}\right)\,d\mu(z)\\
&\leq C\,r^{-t}\,\int \psi_{a,t}\left(\frac{|z-x|}{r}\right)d\mu(z)=C\,r^{-t}\,h_{\mu,a,t}(x,r)=C\,{\varepsilon}_{\mu,a,t}(x,r),
\endaligned$$
where \eqref{lemtec1} was used with $\alpha=t+a$, $\beta=2$ (hence $m=t+a$). In particular, ${\varepsilon}_{\mu,a,t}\in\mathcal G_2(t)$.
\end{proof}
\begin{remark}
If $a=2-t$, then we cannot ensure that ${\varepsilon}_{\mu,a,t}\in \mathcal G_2(t)$. Indeed, if
we set $\alpha=\beta$ in the left hand side \eqref{lemtec1}, then one gets a worse estimate.
One easily checks that in this case,
$$\sum_{k\geq 0} \frac{2^{-\alpha k}}{\bigl(2^{-k}x\bigr)^{\alpha} + 1}
\simeq\frac{1}{\alpha}\,\frac{\log(1+x^\alpha)}{x^\alpha+1},$$
with absolute constants. Hence a logarithmic term appears, which implies that
$$
\aligned
\sum_{k=0}^\infty \frac{{\varepsilon}_{\mu,2-t,t}(x,2^kr)}{2^{(2-t)k}}&= r^{-t}\,\int \sum_{k=0}^\infty \frac{2^{-2k}}{\left(2^{-k}\,\frac{|z-x|}{r}\right)^2+1}\,d\mu(z)\\
&\simeq r^{-t} \int\frac{1}{\left(\frac{|z-x|}{r}\right)^2+1}\,\log\left(1+\left(\frac{|z-x|}{r}\right)^2\right)\,d\mu(z)
\endaligned
$$
and so we cannot infer that ${\varepsilon}_{\mu,2-t,t}\in\mathcal G_2(t)$.
\end{remark}
\subsection{Behavior of $\mathcal G_1$ and $\mathcal G_2$ under quasiconformal mappings}
Need will arise of evaluating gauge functions $h=h(B(x,r))=r^t\,{\varepsilon}(x,r)$ on sets that are not necessarily balls. To do this, given an arbitrary bounded set $A\subset\mathbb C$, let $B$ a ball with minimal diameter that contains $A$. Then we set
$${\varepsilon}(A)={\varepsilon}(B).$$
If may happen that $B$ is not unique, but this does not cause any harm. In this case, for definiteness, we can define ${\varepsilon}(A)$ as the infimum of the values ${\varepsilon}(B)$ over all balls $B$ with minimal diameter containing $A$. Analogously, if $h(x,r)=r^t\,{\varepsilon}(x,r)$, we define $h(A)$ as the infimum the $h(B)$'s. \\
Our next objective consists in showing that if $\phi$ is a $K$-quasiconformal planar homeomorphism and $0\leq d\leq1$, then the function defined by
$${\varepsilon}(B) = {\varepsilon}_{\mu,a,t}(\phi(B))^d$$
for any ball $B\subset\mathbb C$, also belongs to $\mathcal G_1\cap\mathcal G_2$. In fact, because of the geometric properties of quasiconformal mappings and the smoothness of $\psi_{a,t}$, it is easily seen that ${\varepsilon}$ satisfies \rf{eqeq1}. To show that \rf{eqsum4} also holds requires some more effort.
\begin{lemma}\label{lemtec5}
Let ${\varepsilon}_{\mu,a,t}$ be as above, and let $\phi:\mathbb C\to\mathbb C$ be a $K$-quasiconformal mapping. For every $d>0$ and $K\geq 1$ there exist two positive constants $C=C(K,d)$ and $C_1=C_1(k,d)$ such that
\begin{equation}\label{eqfo2}
\sum_{j\geq0}\frac{{\varepsilon}_{\mu,a,t}(\phi(B(x,2^jr)))^d}{2^{bj}}\leq C(K,d)\,{\varepsilon}_{\mu ,a,t}(\phi(B(x,r)))^d
\end{equation}
whenever $0<a<C_1b$. In particular, if $b=2-t$ and $a$ is chosen small enough, the function ${\varepsilon}$ defined by ${\varepsilon}(B) = {\varepsilon}_{\mu,a,t}(\phi(B))^d$ for
any ball $B$, belongs to $\mathcal G_1\cap\mathcal G_2(t)$.
\end{lemma}
\begin{proof}
Set $d_j={\rm \operatorname{diam}}(\phi(B(x,2^jr)))$. We have
\begin{align*}
S & =\sum_{j\geq0}\frac{{\varepsilon}_{\mu,a,t}(\phi(B(x,2^jr)))^d}{2^{bj}} \simeq \sum_{j\geq0}\frac{{\varepsilon}_{\mu ,a,t}(B(\phi
(x),d_j))^d}{2^{bj}}\\
& \lesssim \sum_{k\geq0}
\sum_{j:d_0 2^{k}\leq d_j< d_02^{k+1}}\frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))^d}{2^{bj}}.
\end{align*}
For each $j\geq0$ we have
$$\frac{d_j}{d_0} = \prod_{i=1}^j \frac{d_i}{d_{i-1}} = \prod_{i=1}^j \frac{{\rm \operatorname{diam}}(\phi
(B(x,2^ir)))}{{\rm \operatorname{diam}}(\phi(B(x,2^{i-1}r)))}
\leq C(K)^j = 2^{C_2j},$$
with $C_2$ depending on $K$. Thus, for $j,k$ such that $d_0 2^{k}\leq d_j< d_02^{k+1}$,
$$2^j\geq \Bigl(\frac{d_j}{d_0}\Bigr)^{1/C_2}\simeq 2^{k/C_2}.$$
Since the number of $j$s such that $d_0 2^{k}\leq d_j< d_02^{k+1}$ depends only on $K$,
we obtain
$$S\lesssim \sum_{k\geq0}
\sum_{j:d_0 2^{k}\leq d_j< d_02^{k+1}} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))^d}{2^{C_1bk}} \leq C
\sum_{k\geq0} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))^d}{2^{C_1bk}},$$
with $C_1=1/C_2$.
First we consider the case $d=1$:
from Lemma \ref{emug2}, if $0<a<C_1b$, we infer that
\begin{align}\label{eqd41}
\sum_{k\geq0} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))}{2^{C_1bk}} & =
\sum_{k\geq0} \frac{1}{2^{(t+C_1b)k}d_0^{\,t}}\int \frac1{\biggl(\dfrac{|\phi(x)-y|}{2^kd_0}\biggr)^{t+a} + 1}\,
d\mu(y) \nonumber\\
& \lesssim \frac1{d_0^{\,t}}\int
\frac 1{\biggl(\dfrac{|\phi(x)-y|}{d_0}\biggr)^{t+a} + 1}\,d\mu(y) \nonumber\\
& =
{\varepsilon}_{\mu ,a,t}(B(\phi(x),d_0))\lesssim {\varepsilon}_{\mu ,a,t}(\phi(B(x,r))),
\end{align}
and so we are done in this case.
If $d>1$, we set
$$S\lesssim \left( \sum_{k\geq0} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))}{2^{\frac{C_1}d bk}} \right)^d
\lesssim {\varepsilon}_{\mu ,a,t}(\phi(B(x,r)))^d,$$
by \rf{eqd41}, replacing $C_1$ there by $C_1/d$ (and thus assuming now that $0<a<\frac{C_1}d b$).
When $0<d<1$ we use H\"older inequality, with $p=1/d$:
\begin{align*}
\sum_{k\geq0} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))^d}{2^{C_1bk}} & \leq
\biggl(\sum_{k\geq0} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))}{2^{C_1bk}}\biggr)^d
\biggl(\sum_{k\geq0} \frac1{2^{C_1bk}}\biggr)^{1/p'}\\
& \leq C \biggl(\sum_{k\geq0} \frac{{\varepsilon}_{\mu ,a,t}(B(\phi(x),2^kd_0))}{2^{C_1bk}}\biggr)^d.
\end{align*}
If we plug in this inequality the estimate obtained in \rf{eqd41}, then \rf{eqfo2} follows.
The fact that ${\varepsilon}$ defined by ${\varepsilon}(B) = {\varepsilon}_{\mu,a,t}(\phi(B))^d$ for
any ball $B$ belongs to $\mathcal G_2(t)$ is a consequence of the definition of $\mathcal G_2(t)$ in
\rf{eqsum4} and the estimate \rf{eqfo2}, choosing $b=2-t$ and $0<a<C_1b$.
\end{proof}
\bigskip
\section{Weighted bounds for the Beurling transform}
\subsection{The weight $\omega$ and the Beurling transform}
In this section, $0<t<2$ is fixed. We will prove weighted estimates for the Beurling transform. To describe the class of weights we refer to, let $\mathcal{P}=\{ P_i\}_{i=1}^N$ be a family of dyadic squares such that $3P_i\cap 3P_j=\varnothing$ if $i\neq j$, and satisfying the $h$-packing condition
\begin{equation}\label{pack}
\sum_{P\in\mathcal{P}:P\subset Q} h(P)\leq C_{pack}\, h(Q)\quad \mbox{ for every
dyadic square $Q$}.
\end{equation}
Here, $h(x,r)=r^t\,{\varepsilon}(x,r)$ is any gauge function with ${\varepsilon}\in{\mathcal G}_2$. Then the weights we are interested in are precisely the following:
\begin{equation}\label{defomega}
\omega = \sum_{P\in\mathcal{P}} \frac{h(P)}{\ell(P)^2}\,\chi_P.
\end{equation}
These weights already appeared in \cite{Lacey-Sawyer-Uriarte} in the particular case
${\varepsilon}(x,r)\equiv1$. It is easy to see that if $\overline P = \cup_{i=1}^N P_i$, then $\omega$ belongs to $A_{1,\overline P}$, the local $A_1$ Muckenhoupt class. That is, for every square $Q\subset \mathbb C$,
\begin{equation}\label{eqmax2}
\frac{\omega(Q)}{\ell(Q)^2}\leq C\,\omega(x),\hspace{1cm}\text{ for almost every }x\in\overline P\cap Q.
\end{equation}
Indeed, let $Q$ be a square containing $x\in
P_i$ such that $\ell(Q)\geq\ell(P_i)$. We have $\omega(Q)\leq C\,h(Q)$ because of the packing condition on the squares from $\mathcal P$, and then
using that ${\varepsilon}\in\mathcal G_2$ we infer that $\frac{h(x,r)}{r^2}\leq C\,\frac{h(x,s)}{s^2}$ if
$s\leq r$, and thus
$$\frac{\omega(Q)}{\ell(Q)^2}\leq C\,\frac{h(Q)}{\ell(Q)^2} \leq C\,\frac{h(P_i)}{\ell(P_i)^2} = C\,\omega(x).$$
If $\ell(Q)<\ell(P_i)$, then it is also clear that \rf{eqmax2} holds.
From \rf{eqmax2} we obtain that if $M$ is the classical Hardy-Littlewood maximal operator, then
\begin{equation}\label{eqmax3}
M\omega(x)\leq C\,\omega(x)\hspace{1cm}\text{ for almost every }x\in\overline P.
\end{equation}
Let $M_\omega$ be the centered Hardy-Littlewood maximal function with respect to the $\omega$ measure. That is,
$$M_\omega f(x) = \sup_{r>0} \frac1{\omega(Q(x,r))}\,\int_{Q(x,r)}f(y)\,\omega(y)\,dm(y),$$
where $Q(x,r)$ stands for the square centered at $x$ with side length $2r$, and $m$ denotes the planar Lebesgue measure.
It is well known that $M_\omega$ is of weak type $(1,1)$ and strong type
$(p,p)$, for $1<p\leq\infty$, with respect to the measure $\omega$.
From the following lemma it turns out that the same is also true for $M$:
\begin{lemma}\label{lemmaxomega}
Let $\omega$ be as above.
There exists some constant $C$ such that
$$M f(x)\leq C\, M_\omega f(x)\qquad\mbox{for $f\in L^1_{loc}(\overline P)$ and $x\in\overline P$.}$$
As a consequence, $M$ is of weak type $(1,1)$ and strong type
$(p,p)$, $1<p\leq\infty$, with respect to $\omega$.
\end{lemma}
\begin{proof}
Let $f\in L^1_{loc}(\overline P)$
and $Q$ a square containing $x\in P_i$. Consider the minimal square $Q'$ centered at $x$
containing $Q$. Since $\ell(Q')\simeq \ell(Q)$, using \rf{eqmax2} we get
\begin{align*}
\frac1{m(Q)}\int_Q|f|\,dm&\lesssim \frac1{m(Q')}\int_{Q'}|f|\,dm \\
& \lesssim \frac{\inf_{y\in Q'\cap \overline P}\omega(y)}{\omega(Q')}\,\int_{Q'}|f|\,dm\leq
\frac1{\omega(Q')}\int_{Q'}|f|\,\omega\,dm\leq M_\omega f(x).
\end{align*}
Thus $M f(x)\leq C\, Mf_\omega(x)$ and the lemma follows.
\end{proof}
\bigskip
Recall that the Beurling transform of a function $f:\mathbb C\to\mathbb C$ is given by
$$\mathcal{S} f(z) = \frac{-1}{\pi}\,{\rm p.v.} \int_\mathbb C \frac{f(\xi)}{(z-\xi)^2}\,dm(\xi).$$ The $\epsilon$-truncated Beurling transform is
$$\mathcal{S}_\epsilon f(z) = \frac{-1}{\pi} \int_{|z-\xi|>\epsilon} \frac{f(\xi)}{(z-\xi)^2}\,dm(\xi),$$
and the maximal Beurling transform, $\mathcal{S}_* f(z)=\sup_{\epsilon>0}|\mathcal{S}_\epsilon f(z)|.$
\medskip
\begin{proposition}\label{propo9}
Let $\mathcal{P}$ be a family of dyadic squares as above, and set
$\overline P = \bigcup_{i=1}^N P_i$. If ${\varepsilon}\in\mathcal G_2$ and $\omega$ is the
weight defined by \rf{defomega}, then the Beurling transform, is bounded in $L^p(\omega)$, for $1<p<\infty$,
and of weak type $(1,1)$ with respect to $\omega$. That is,
\begin{equation}\label{strong}
\|\mathcal{S}(f\chi_{\overline P})\|_{L^p(\omega)}\leq C\,\|f\|_{L^p(\omega)}\quad \mbox{for all $f\in L^p(\omega)$},
\end{equation}
and
\begin{equation}\label{weak**}
\|\mathcal{S}(f\chi_{\overline P})\|_{L^{1,\infty}(\omega)}\leq C\|f\|_{L^1(\omega)}\quad \mbox{for all $f\in L^1(\omega)$},
\end{equation}
for some $C>0$ depending on $p$ and $C_{pack}$.
\end{proposition}
It is possible to prove the estimate \rf{strong} (which, for $p=2$, is the one needed in connection with quasiconformal distortion) by an appropriate modification of the arguments of
\cite{Lacey-Sawyer-Uriarte}.
However we have preferred to follow a new approach:
first
we will show the following weak type inequality, which is stronger than \rf{weak**}:
\begin{equation}\label{weak}
\|\mathcal{S}_*(f\chi_{\overline P})\|_{L^{1,\infty}(\omega)}\leq C\|f\|_{L^1(\omega)}
\quad \mbox{for all $f\in L^1(\omega)$}.
\end{equation}
Then, by means of a \emph{good lambda} inequality, we will deduce that the maximal Beurling transform is bounded in $L^p(\omega)$, for $1<p<\infty$, that is
\begin{equation}\label{eqstrong*}
\|\mathcal{S}_*(f\chi_{\overline P})\|_{L^p(\omega)}\leq C\,\|f\|_{L^p(\omega)}.
\end{equation}
Clearly, \eqref{strong} follows from this estimate.
We prove \eqref{weak} in the next Lemma.
\begin{lem}\label{weakbstar}
We have
$$\omega\left(\left\{z\in\mathbb C:|\mathcal{S}_* f(z)|>\lambda\right\}\right)\leq\frac{C}{\lambda}\,\|f\|_{L^1(\omega)}$$
for every $f\in L^1(\omega)$ and $\lambda>0$.
\end{lem}
\begin{proof}
We have
$$\aligned
&\omega\left\{z\in\mathbb C:|\mathcal{S}_* f(z)|\geq \lambda\right\}=\sum_i\omega\left\{z\in P_i: |\mathcal{S}_* f(z)|\geq \lambda\right\}\\
&\leq
\sum_i \omega\left\{z\in P_i: |\mathcal{S}_*(f\chi_{2P_i})(z)|\geq\frac{\lambda}{2}\right\}+
\sum_i \omega\left\{z\in P_i: |\mathcal{S}_*(f\chi_{\mathbb C\setminus 2P_i})(z)|\geq\frac{\lambda}{2}\right\}\\
&=A+B.
\endaligned$$
For $A$ we just use the boundedness of $\mathcal{S}_*:L^1\to L^{1,\infty}$ with respect to Lebesgue measure,
$$
\aligned
A=\sum_i \omega\Bigl\{z\in P_i: |\mathcal{S}_*(f\chi_{2P_i})(z)| & \geq\frac{\lambda}{2}\Bigr\}
\\ &
= \sum_i \frac{h(P_i)}{\ell(P_i)^2}\,\Bigl|\Bigl\{z\in P_i: |\mathcal{S}_*(f\chi_{2P_i})(z)|\geq\frac{\lambda}{2}
\Bigr\}\Bigr|
\\
&\leq 2\|\mathcal{S}_*\|_{L^1\to L^{1,\infty}}\sum_i \frac{h(P_i)}{\ell(P_i)^2}\, \frac1\lambda
\int |f\chi_{2P_i}|\,dm\\
&= 2\|\mathcal{S}_*\|_{L^1\to L^{1,\infty}}\, \frac{\|f\|_{L^1(\omega)}}{\lambda}
\endaligned
$$
because the squares $3P_i$ are disjoint and $\omega$ coincides with the Lebesgue measure times
$\frac{h(P_i)}{\ell(P_i)^2}$ on every $P_i$. For the remaining term, denoting the center of $P_i$
by $z_i$, one has
$$\aligned
B&=\sum_i \omega\left\{z\in P_i: |\mathcal{S}_*(f\chi_{\mathbb C\setminus 2P_i})(z)|\geq\frac{\lambda}{2}\right\}\\
&\leq\frac2\lambda\sum_i \int_{P_i} |\mathcal{S}_*(f\chi_{\mathbb C\setminus 2P_i})(z)|\,d\omega(z)\\
&\leq\frac2\lambda\sum_i\int_{P_i} \sum_{j\neq i}\int_{P_j}\frac{|f(\xi)|}{|\xi-z|^2}\,dm(\xi)\,d\omega(z)\\
&\leq C\,\frac2\lambda\sum_{j}\left(\sum_{i\neq j}\frac{\omega(P_i)}{|z_j-z_i|^2}\right)\int_{P_j}|f(\xi)|\,dm(\xi).
\endaligned$$
Using the $h$-packing condition \eqref{pack} and the ${\mathcal G}_2$ condition for $h$, we get
$$
\aligned
\sum_{i\neq j}\frac{\omega(P_i)}{|z_j-z_i|^2}&=\sum_{k=2}^\infty\,\sum_{i: z_i\in 2^kP_j\setminus 2^{k-1}P_j}\frac{h(P_i)}{|z_i-z_j|^2}\leq C\,\sum_{k=2}^\infty\,\sum_{i: P_i\subset 2^{k+1}P_j}\frac{h(P_i)}{(2^k\ell(P_j))^2}\\
&= C\,\sum_{k=2}^\infty \frac{1}{(2^k\ell(P_j))^{2}}\sum_{i: P_i\subset 2^{k+1}P_j}h(P_i)\leq C\,\sum_{k=2}^\infty \frac{h(2^kP_j)}{(2^k\ell(P_j))^{2}}\leq C\,\frac{h(P_j)}{\ell(P_j)^{2}},\\
\endaligned
$$
where we used the fact that if $z_i\in 2^kP_j$, then $P_i\subset 2^{k+1}P_j$ (because
$3P_i\cap3P_j=\varnothing$). Thus,
$$B\leq \frac{C}{\lambda}\,\|f\|_{L^1(\omega)}.$$
The Lemma follows since both $A$, $B$ are bounded by constant multiples of $\frac1\lambda\,\|f\|_{L^1(\omega)}$.
\end{proof}
\bigskip
\begin{proof}[\bf Proof of the boundedness of $\mathcal S_*$ in $L^p(\omega)$, for $1<p<\infty$]
Our goal here is to obtain the following \emph{good lambda} inequality,
\begin{equation}\label{gl}
\omega\bigg(\left\{z: \mathcal{S}_* f(z)>10\lambda, M_\omega f(z)\leq\gamma\lambda\right\}\bigg)\leq C\,\gamma\,\omega\bigg(\left\{z:\mathcal{S}_* f(z)>\lambda\right\}\bigg),
\end{equation}
for every $\lambda>0$, and some $\gamma$ small enough. Recall that $M_\omega$ denotes the centered Hardy-Littlewood maximal function with respect to the $\omega$ measure.
By standard arguments, the preceding estimate implies that
$$\|{\mathcal S}_*f\|_{L^p(\omega)}\leq C_p
\|M_\omega f\|_{L^p(\omega)},$$
for $0< p <\infty$. Since $M_\omega$ is bounded in $L^p(\omega)$
for $p>1$, this implies that
$\mathcal S_*$ is bounded in $L^p(\omega)$, $1<p<\infty$.
\noindent To get \eqref{gl}, let us denote $\Omega_\lambda=\{\mathcal{S}_* f>\lambda\}$, and let
$$\Omega_\lambda=\bigcup_{j=1}^\infty Q_j$$
be a Whitney decomposition of $\Omega_\lambda$. That is, $\sum_j\chi_{10Q_j}\leq C$, and for every $j$ we have that $100 Q_j\subset \Omega_\lambda$ but $1000 Q_j\nsubseteq\Omega_\lambda$. Let $Q_j$ be a fixed Whitney cube, and assume that there exists $z_j\in Q_j$ such that $M_\omega f(z_j)\leq \gamma\lambda$ (otherwise there is nothing to prove). Let $t_j\in\mathbb C\setminus\Omega_\lambda$ the closest point to $Q_j$ in $\mathbb C\setminus\Omega_\lambda$. Let $B=B(t_j,c_0\ell(Q_j))$ be a ball centered at $t_j$ and such that $10Q_j\subset B$. We can decompose
$$f=f\,\chi_B+f\,\chi_{\mathbb C\setminus B}.$$
For every $z\in Q_j$, the truncated singular integral
$\mathcal{S}_\epsilon (f\chi_B)(z)$ can be written as the sum of two terms,
$$\aligned
\mathcal{S}_\epsilon (f\chi_B)(z)&=\int_{|t-z|\geq \epsilon}\frac{f(t)\,\chi_{B\cap 3B_j}(t)}{(t-z)^2}\,dm(t)+\int_{|t-z|\geq \epsilon}\frac{f(t)\,\chi_{B\setminus 3B_j}(t)}{(t-z)^2}\,dm(t)=I+II,
\endaligned$$
where $B_j=B(z_j,\ell(Q_j))$. For $II$, we use that $|t-z|\simeq \ell(Q_j)$
and Lemma \ref{lemmaxomega} to get
\begin{equation}\label{eq1}
II\leq \frac{C}{\ell(Q_j)^2}\int_{B\setminus 3B_j}|f(t)|\,dm(t)\leq C\,M f(z_j)
\leq C\,M_\omega f(z_j)\leq C\,\gamma\lambda,
\end{equation}
and this is uniform in $\epsilon$. Therefore, since $\mathcal{S}_* f(z)\leq \mathcal{S}_*(f\chi_B)(z)+\mathcal{S}_*(f\chi_{\mathbb C\setminus B})(z)$, we have
\begin{equation}\label{A+B}
\aligned
\omega \bigg(\big\{z\in Q_j: &\mathcal{S}_* f(z)>10\lambda, M_\omega f(z)\leq\gamma\lambda\big\}\bigg)\\
\leq
&\;\omega\bigg(\left\{z\in Q_j: \mathcal{S}_* (f\chi_B)(z)>2\lambda, M_\omega f(z)\leq\gamma\lambda\right\}\bigg)\\
&+\omega\bigg(\left\{z\in Q_j: \mathcal{S}_* (f\chi_{\mathbb C\setminus B})(z)>8\lambda,
M_\omega f(z)\leq\gamma\lambda\right\}\bigg)=A+B.
\endaligned
\end{equation}
Choosing $\gamma$ so that $C\gamma<1$ in \eqref{eq1}, we get that
$$
A\leq \omega\bigg(\left\{z\in Q_j: \mathcal{S}_* (f\chi_{B\cap 3B_j})(z)>\lambda, M_\omega f(z)\leq\gamma\lambda\right\}\bigg).
$$
On the other hand, by Lemma \ref{weakbstar},
$$
\omega\left\{\mathcal{S}_* (f\chi_{B\cap 3B_j})(z)>\lambda\right\}\leq \frac{C}{\lambda}\,\|f\chi_{B\cap 3B_j}\|_{L^1(\omega)}\leq\frac{C\,M_\omega f(z_j)\,\omega(B\cap 3B_j)}{\lambda}.
$$
Therefore
\begin{equation}\label{eqA}
A\leq C\,\gamma\,\omega(B\cap 3B_j)\leq C\,\gamma\,\omega (7Q_j).
\end{equation}
To estimate $B$, notice that, for $z\in Q_j$,
$$
\aligned
\left|\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(z)-\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(t_j)\right|
&\leq\int_{\mathbb C\setminus B}|f(t)|\left|\frac{1}{(t-z)^2}-\frac{1}{(t-t_j)^2}\right|\,dm(t)\\
&\leq C\,\ell(Q_j)\int_{\mathbb C\setminus B}\frac{|f(t)|\,|t-\frac{z+t_j}{2}|}{|t-z|^2\,|t-t_j|^2}dm(t)\\
&\leq C\,\ell(Q_j)\int_{\mathbb C\setminus B}\frac{|f(t)|}{|t-z_j|^3}dm(t)\\
&\leq C\,\ell(Q_j)\int_{\mathbb C\setminus 3B_j}\frac{|f(t)|}{|t-z_j|^3}\,dm(t)\\
&\leq C\,Mf(z_j)\leq C\,M_\omega f(z_j).
\endaligned
$$
The next to the last step above follows as usually after decomposing into dyadic annuli. Since $B\supset B(z,3\ell(Q_j))$, to compute $\mathcal{S}_*(f\chi_{\mathbb C\setminus B})(z)=\sup_{\epsilon>0}|\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(z)|$ it suffices to take $\epsilon>3\ell(Q_j)$. Therefore
$$
\aligned
|\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(t_j)-\mathcal{S}_\epsilon(f)(t_j)|&=\left|\int_{|t-t_j|>\epsilon}\frac{f(t)\,\chi_B(t)}{(t-t_j)^2}dm(t)\right|\\
&\leq\frac{C}{\ell(Q_j)^2}\int_{|t-t_j|>\epsilon}|f(t)|\,\chi_B(t)dm(t)\\
&=\frac{C}{\ell(Q_j)^2}\int_{\epsilon<|t-t_j|<c_0\ell(Q_j)}|f(t)|\,dm(t)\\
&\leq\frac{C}{\ell(Q_j)^2}\int_{|t-z_j|\leq 2c_0\ell(Q_j)}|f(t)|dm(t)\leq C\,Mf(z_j).
\endaligned
$$
Summarizing, we get
$$
\aligned
|\mathcal{S}_\epsilon&(f\chi_{\mathbb C\setminus B})(z)|\\
&\leq |\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(z)-\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(t_j)|+|\mathcal{S}_\epsilon(f\chi_{\mathbb C\setminus B})(t_j)-\mathcal{S}_\epsilon f(t_j)|+|\mathcal{S}_\epsilon f(t_j)|\\
&\leq C M_\omega f(z_j)+|\mathcal{S}_\epsilon f(t_j)|.
\endaligned
$$
Therefore, if $z$ belongs to $\left\{z: \mathcal{S}_* (f\chi_{\mathbb C\setminus B})(z)>8\lambda, M_\omega f(z)\leq\gamma\lambda\right\}$, we get that
$$
\aligned
8\lambda<|\mathcal{S}_*(f\chi_{\mathbb C\setminus B})(z)|\leq C M_\omega f(z_j)+|\mathcal{S}_* f(t_j)|\leq C\gamma\lambda+\lambda.
\endaligned
$$
because $t_j\notin\Omega_\lambda$. In particular, back to \eqref{A+B}, for small enough $\gamma$ we must have $B=0$. Now, with the help of \eqref{eqA} we get
$$
\omega\bigg(\left\{z\in Q_j: \mathcal{S}_* f(z)>10\lambda, M_\omega f(z)\leq\gamma\lambda\right\}\bigg)\leq C\,\gamma\,\omega(7Q_j).
$$
Since the squares $7Q_j$ have bounded overlap, summing in $j$ we obtain
$$\aligned
\omega\bigg(\left\{z: \mathcal{S}_* f(z)>10\lambda, M_\omega f(z)\leq\gamma\lambda\right\}\bigg)
\leq C\,\gamma\,\omega\left(\left\{z:\mathcal{S}_* f(z)>\lambda\right\}\right),
\endaligned$$
which is \eqref{gl}.
\end{proof}
\bigskip
\section{Quasiconformal distortion of $h$-contents}
\subsection{Conformal outside}
Let $\phi$ be a $K$-quasiconformal mapping on $\mathbb C$, and let ${\varepsilon}_0\in\mathcal G_1$. Consider the associated gauge function $h_0(x,r) = r^t\,{\varepsilon}_0(x,r)$, for a fixed $0<t<2$.\\
\\
For every $x\in\mathbb C$ and $r>0$, denote
$${\varepsilon}(x,r) = {\varepsilon}_0(\phi(B(x,r)),\qquad
h(x,r) = r^t\,{\varepsilon}(x,r),$$ and suppose that ${\varepsilon}\in\mathcal G_2$.
This fact is crucial in this subsection.\\
\\
Let now ${\mathcal P}=\{P_i\}$ be a finite family of dyadic squares, with disjoint triples, and satisfying the packing condition
\begin{equation}\label{packingconstant}
\sum_{P\in\mathcal{P}:P\subset Q} h(P)\leq C_{pack}\, h(Q)\quad \mbox{ for every dyadic square $Q$}.
\end{equation}
If we introduce the weight
$$\omega=\sum_{P\in{\mathcal P}}\frac{h(P)}{|P|}\,\chi_P\simeq\sum_{P\subset\mathcal{P}}\frac{{\varepsilon}(P)}{\ell(P)^{2-t}}\,\chi_P$$
then it follows from Proposition \ref{propo9} that the Beurling transform is bounded in $L^2(\omega)$,
$$\|{\mathcal S}(f\chi_{\overline P})\|_{L^2(\omega)}\leq \|\mathcal S\|_{L^2(\omega)}\,\|f\|_{L^2(\omega)}$$
and the norm $\|{\mathcal S}\|_{L^2(\omega)}$ depends only on the packing constant $C_{pack}$. Therefore there exists a number $\delta>0$, depending only on $C_{pack}$, such that
\begin{equation}\label{Kpetit}
1\leq K<1+\delta\hspace{1cm}\Rightarrow \hspace{1cm}\frac{K-1}{K+1}\,\|{\mathcal S}\|_{L^2(\omega)}<1.
\end{equation}
In establishing quasiconformal distortion estimates for the $h$-contents we need to
normalize the mappings. A very convenient normalization is obtained by requiring that $\phi$ is
conformal outside the unit disk and has the development
$$\phi(z) = z + {\mathcal O}(1/z),\qquad |z| > 1.$$
We call such $\phi$ as {\it principal quasiconformal mapping}.
\begin{lemma} \label{lemkpetit}
Let $\phi,{\varepsilon}_0,{\varepsilon},\mathcal{P},\omega$ be as above and suppose moreover that the quasiconformal mapping $\phi$ is principal and conformal
outside $\overline P=\bigcup_{P\in\mathcal{P}} P$. Then we have
$$\sum_{P\in\mathcal{P}}h_0(\phi(P))\leq C\,\sum_{P\in\mathcal{P}} h(P),$$
for some constant $C=C(K)>0$.
\end{lemma}
\begin{proof}
First of all, by the definition of ${\varepsilon}_0$ on arbitary sets, we see that
$$
h_0(\phi(P_i))=\inf_{\phi(P_i)\subset B} h_0(B)=\inf_{\phi(P_i)\subset B(x,r)} r^t\,{\varepsilon}_0(x,r)\simeq \operatorname{diam}(\phi(P_i))^t\,{\varepsilon}(P_i),
$$
with constants that only depend on $K$. Thus, by H\"older's inequality we get
\begin{align*}
\sum_i h_0(\phi(P_i))
&\leq C_K\,\sum_i \operatorname{diam}(\phi(P_i))^t{\varepsilon}(P_i)\\
&\leq C_K\Bigl(\sum_i \operatorname{diam}(\phi(P_i))^2\ell(P_i)^{t-2}{\varepsilon}(P_i)\Bigl)^{t/2}\Bigl(\sum_i\ell(P_i)^t{\varepsilon}(P_i)\Bigr)^{\frac{2-t}2}\\
& = C_K\biggl(\sum_i \operatorname{diam}(\phi(P_i))^2\,\frac{\omega(P_i)}{\ell(P_i)^2}\biggl)^{t/2}\Bigl(\sum_i \omega(P_i)\Bigr)^{\frac{2-t}2}
=: C_K \,A^{t/2}\, \omega(\overline P)^{\frac{2-t}2}.
\end{align*}
To estimate $A$, we start by getting from quasisymmetry that
\begin{align*}
A & \leq C_K \sum_i \int_{P_i} J(z,\phi)\,\omega(z)\,dm(z)
\end{align*}
Now, as $\phi$ is a principal quasiconformal mapping, $\phi(z)-z={\mathcal C}g(z)$, where ${\mathcal C}g$ is the Cauchy transform of the Neumann series
$$g=\sum_{n=0}^\infty(\nu\mathcal S)^n(\nu),$$
and $\nu(z)=\frac{\overline\partial\phi(z)}{\partial\phi(z)}$ whenever $\partial\phi(z)\neq 0$ (otherwise we simply set $\nu(z)=0$).
Since $K<1+\delta$, \eqref{Kpetit} says that the above series converges absolutely in $L^2(\omega)$
(using the key fact that $\operatorname{supp}(\nu)\subset \overline P$ by the conformality of $\phi$ off $\overline{P}$), and moreover one easily gets
$$\|g\|_{L^2(\omega)}\leq
\frac{\|\nu\|_\infty}{1-\|\nu\|_\infty\,\|{\mathcal S}\|_{L^2(\omega)}}\,\omega({\overline{P}})^{1/2}
.$$
Then arguing as in \cite{Lacey-Sawyer-Uriarte} (see also \cite{astalaareadistortion})
$$\aligned
\sum_i\int_{P_i} J(z,\phi)\,\omega(z)\,&dm(z)=\int_{\overline{P}}
\Bigl(|\partial \phi(z)|^2-|\bar\partial\phi(z)|^2\Bigr)\,\omega(z)\,dm(z)
\\
&= \int_{\overline{P}}\left( 1+|{\mathcal S}g(z)|^2+2\text{\rm Re\,}({\mathcal S}g(z))-|g(z)|^2\right)\,\omega(z)\,dm(z)\\
&\leq \left(\omega({\overline{P}}) + \|{\mathcal S}\|_{L^2(\omega)}^2\,\|g\|_{L^2(\omega)}^2+2 \omega({\overline{P}})^\frac12\,\|{\mathcal S}\|_{L^2(\omega)}\,\|g\|_{L^2(\omega)}\right)\\
&\leq \omega({\overline{P}})\left( 1 + \,\frac{\|{\mathcal S}\|_{L^2(\omega)}^2\,\|\nu\|_\infty^2}{\left(1-\|\nu\|_\infty\,\|{\mathcal S}\|_{L^2(\omega)}\right)^2}+\frac{2\,\|{\mathcal S}\|_{L^2(\omega)}\,\|\nu\|_\infty}{1-\|\nu\|_\infty\,\|{\mathcal S}\|_{L^2(\omega)}}\right)\\
&\leq C(K)\,\omega({\overline{P}}),
\endaligned
$$
by \rf{Kpetit}. Thus,
$$\sum_{P\in\mathcal{P}}h_0(\phi(P))\leq C(K)\, \omega(\overline P)=C(K)\,\sum_{P\in\mathcal{P}}h(P),$$
and so the lemma follows.
\end{proof}
\subsection{Conformal inside}
We will prove now an h-version of \cite[Theorem 2.2]{ACMOU}. Here
the point is to use quasiconformal mappings that are conformal inside a finite disjoint union
of quasidisks, allowing improved integrability for the gradient \cite{astalanesi}. Let us
emphasize that no $\mathcal G_2$ assumption will be needed here.
\begin{theo}\label{nesi}
Let $\phi:\mathbb C\to\mathbb C$ be $K$-quasiconformal, principal, and conformal outside $\mathbb D$. Assume that $Q_i\subset\mathbb D$ are pairwise disjoint $K$-quasidisks, and that $\phi$ is conformal in $\Omega=\cup_iQ_i$. For a fixed $\varepsilon_0\in{\mathcal G}_1$ and $0<t<2$, let $t'=\frac{2Kt}{2+(K-1)t}$ and $h_0(r)=r^{t'}{\varepsilon}_0(x,r)$. Set
$${\varepsilon}(B)={\varepsilon}_0(\phi(B))^\frac{Kt}{t'}\hspace{1cm}\text{and}\hspace{1cm} h(r)=r^{t}\,{\varepsilon}(x,r).$$
Then
$$
\sum_i h_0(\phi(Q_i))\leq C(K,t)\,\left(\sum_i h(Q_i)\right)^\frac{t'}{Kt}.
$$
\end{theo}
\begin{proof}
From the quasisymmetry of $\phi$, and the doubling properties of ${\varepsilon}_0$ and ${\varepsilon}$,
$$
h_0(\phi(Q_i))
=\inf_{\phi(Q_i)\subset B} h_0(B)
=\inf_{\phi(Q_i)\subset B(x,r)}r^{t'} {\varepsilon}_0(x,r)
\simeq\operatorname{diam}(\phi(Q_i))^{t'}\,{\varepsilon}(Q_i)^\frac{t'}{Kt}$$
with constants that depend only on $K$. By quasisymmetry again and H\"older's inequality, we get
$$\aligned
\operatorname{diam}(\phi(Q_i))
&\leq C_K\,\left(\int_{Q_i}J(z,\phi)\,dm(z)\right)^\frac12\\
&\leq C_K\left(\int_{Q_i}J(z,\phi)^\frac{K}{K-1}\,dm(z)\right)^\frac{K-1}{2K}\,\operatorname{diam}(Q_i)^\frac1K
\endaligned$$
since quasidisks have diameter comparable to the square root of its area. Therefore, by H\"older's inequality and the improved borderline integrability of quasiconformal mappings (\cite{astalanesi}),
$$\aligned
\sum_i h_0(\phi(Q_i))
&\leq C(K)\,\sum_i \operatorname{diam}(\phi(Q_i))^{t'}\,{\varepsilon}(Q_i)^\frac{t'}{Kt}\\
&\leq C(K)\sum_i\left(\int_{Q_i}J(z,\phi)^\frac{K}{K-1}\,dm(z)\right)^{\frac{K-1}{2K}\,t'}\,\operatorname{diam}(Q_i)^\frac{t'}{K}\,{\varepsilon}(Q_i)^\frac{t'}{Kt}\\
&\leq C(K)\left(\sum_i\int_{Q_i}J(z,\phi)^\frac{K}{K-1}\,dm(z)\right)^{\frac{K-1}{2K}\,t'}\,
\left(\sum_i\operatorname{diam}(Q_i)^t\,{\varepsilon}(Q_i)\right)^\frac{t'}{Kt}\\
&\leq C(K)\, \left(\sum_i h(Q_i)\right)^\frac{t'}{Kt}\\\endaligned
$$
as claimed.
\end{proof}
\subsection{The main lemma on distortion of $h$-contents. }We are now ready to prove the main estimate on the distortion of $h$-contents by quasiconformal mappings with small distortion.
\begin{lemma}\label{mainlem}
Let $E\subset B(0,1/2)$ be compact and $\phi:\mathbb C\to\mathbb C$ a principal $K$-quasiconformal mapping, conformal on $\mathbb C\setminus \bar \mathbb D$. Let ${\varepsilon}\in\mathcal G_1$, and assume that $({\varepsilon}\circ\phi)^\frac{Kt}{t'}\in \mathcal G_2(t)$. Set
$$h(x,r)=r^{t'}{\varepsilon}(x,r)$$
and
\begin{equation*}
\wt {\varepsilon}(x,r)= {\varepsilon}(\phi(B(x,r)))^\frac{Kt}{t'},\qquad \wt h(x,r)= r^{t}\wt{\varepsilon}(x,r).
\end{equation*}
If $\delta$ is as in \eqref{Kpetit} and $K<1+\delta$, then
$$ \mathcal M^{h}(\phi(E)) \leq C(K) \, \mathcal M^{\wt h}(E)^\frac{t'}{Kt}.$$
\end{lemma}
\begin{proof}
Let us fix $\eta>0$. As in \cite{Lacey-Sawyer-Uriarte} we find a finite family ${\mathcal P}=\{P_1,\dots,P_N\}$ of dyadic cubes, with disjoint triples, such that $E\subset\cup_i 12P_i$ and
$$\sum_i \wt h(P_i)\leq C( \mathcal M^{\wt h}(E)+\eta).$$
Further, we may also assume that the packing condition \eqref{packingconstant} is satisfied, for instance, with constant $C_{pack}=1$, so that $\delta>0$ in \eqref{Kpetit} is fixed. We now decompose $\phi=\phi_2\circ\phi_1$, where both $\phi_1$, $\phi_2$ are principal $K$-quasiconformal
mappings. Moreover, we require $\phi_1$ to be conformal in $\mathbb C\setminus\cup_iP_i$, and $\phi_2$ to be conformal on $\cup_i\phi_1(P_i)$ . \\
\\
Now, we note that $\phi(E)$ can be covered by the quasidisks $\phi(12P_i)$. We can then estimate the $h$-content of $\phi(E)$ with the help of the quasisymmetry,
$$\aligned
\mathcal M^{h}(\phi(E))&\leq \sum_i h(\phi(12P_i))\leq C(K)\,\sum_i h(\phi(P_i))\\
&= C(K)\,\sum_ih(\phi_2\circ\phi_1(P_i)).
\endaligned$$
Since $\phi_1$ is a global $K$-quasiconformal mapping, each $\phi_1(P_i)$ is a $K$-quasidisk. Hence, by Lemma \ref{nesi}, if we define the new gauge function $ h_0(x,r)=r^{t}\,{\varepsilon}_0(x,r)$, with
${\varepsilon}_0(B)={\varepsilon}(\phi_2(B))^\frac{Kt}{t'}$, then we have
$$
\sum_ih(\phi_2\circ\phi_1(P_i))\leq C(K)\,\left(\sum_i h_0(\phi_1(P_i))\right)^\frac{t'}{Kt}.
$$
To estimate the sum on the right hand side above, we use Lemma \ref{lemkpetit}. Indeed, the composition ${\varepsilon}_0\circ\phi_1=({\varepsilon}\circ\phi)^\frac{Kt}{t'}$ certainly belongs to $\mathcal G_1$, and by assumption it also belongs to $\mathcal G_2(t)$. Hence Lemma \ref{lemkpetit} gives us another gauge $h_1(x,r)=r^t\,{\varepsilon}_1(x,r)$, with
${\varepsilon}_1(D)={\varepsilon}_0(\phi_1(D))$,
such that
$$
\sum_i h_0(\phi_1(P_i))\leq C(K)\,\sum_i h_1(P_i).
$$
But then
$$
{\varepsilon}_1(D)={\varepsilon}_0(\phi_1(D))\simeq{\varepsilon}(\phi_2\circ\phi_1(D))^\frac{Kt}{t'}={\varepsilon}(\phi(D))^\frac{Kt}{t'}=\wt{\varepsilon}(D),
$$
so that $h_1(P_i)\simeq \wt h(P_i)$. Summarizing,
$$\aligned
\mathcal M^{h}(\phi(E)) &\leq C(K)\,\left(\sum_i h_0(\phi_1(P_i))\right)^\frac{t'}{Kt}
\leq C(K)\,\left(\sum_i h_1(P_i)\right)^\frac{t'}{Kt}
\\
&\leq C(K)\,\left(\sum_i \wt h(P_i)\right)^\frac{t'}{Kt}\leq C(K)\,\left( \mathcal M^{\wt h}(E)+\eta\right)^\frac{t'}{Kt}.
\endaligned$$
Now letting $\eta\to 0$, the claim follows.
\end{proof}
\noindent
Our next goal is to remove the smallness assumption $K<1+\delta$ in the previous Lemma.
This is done by means of a standard factorization argument.
\begin{lemma}\label{lemkgran}
Let $0<t<2$. Let ${\varepsilon}\in\mathcal G_1$ and set $h(x,r)=r^{t'}{\varepsilon}(x,r)$. Suppose that for any principal $K$-quasiconformal mapping $\psi:\mathbb C\to\mathbb C$ the function $({\varepsilon}\circ\psi)^d$ belongs to $\mathcal G_2$ for any $\frac{t'}{Kt}\leq d\leq 1$. Let $E\subset B(0,1/2)$ be compact and $\phi:\mathbb C\to\mathbb C$ a principal $K$-quasiconformal mapping, conformal on $\mathbb C\setminus \bar \mathbb D$. Set
\begin{equation*}
\wt{\varepsilon}(x,r)= {\varepsilon}(\phi(B(x,r)))^{\frac{Kt}{t'}},\qquad \wt h(x,r)= r^{t}\wt{\varepsilon}(x,r).
\end{equation*}
Then we have
$$ \mathcal M^{h}(\phi(E)) \leq C(K) \, \mathcal M^{\wt h}(E)^\frac{t'}{Kt}.$$
\end{lemma}
\begin{proof}
We factorize $\phi$ so that $\phi = \phi_n\circ\cdots \phi_1$, where each $\phi_i$ is a $K^{1/n}$-quasiconformal mapping, conformal on $\mathbb C\setminus \phi_{i-1}\circ\dots\circ\phi_1(\mathbb D)$. We can further do this so that
$$K^{1/n}<1+\delta$$
for $\delta$ as in \eqref{Kpetit}. Of course, such $n$ will depend on $K$ and also on the packing constant $C_{pack}$ in \eqref{packingconstant}. So we have
$$E = E_0 \stackrel{\phi_1}{\longrightarrow} E_1
\stackrel{\phi_2}{\longrightarrow} \ldots \stackrel{\phi_{n-1}}{\longrightarrow} E_{n-1} \stackrel{\phi_n}{\longrightarrow} E_n
=\phi(E).$$
We now denote $t_0=t$, and for $0\leq j\leq n-1$, we take $t_{j}$ so that
$$\frac{1}{t_{j+1}}-\frac12=\frac{1}{K^\frac1n}\left(\frac{1}{t_j}-\frac12\right).$$
In particular, $t_n=t'$. For any ball $B$, we also set ${\varepsilon}_n(B)={\varepsilon}(B)$, $h_n(B)=h(B)$, and for $j=n-1,n-2,\dots,1,0$ let
$${\varepsilon}_j(B) = {\varepsilon}_{j+1}(\phi_{j+1}(B))^{\frac{K^\frac1n t_j}{t_{j+1}}}
= {\varepsilon}_j(\phi_j(B))^{\frac{t_j}{K^{1/n}\,t_{j-1} }}$$
and $$h_j(x,r)= r^{t_j}\,{\varepsilon}_j(x,r).$$
Note that therefore $${\varepsilon}_0(B)={\varepsilon}_1(\phi_1(B))^\frac{K^{1/n}t_0}{t_1}={\varepsilon}_2(\phi_2\circ\phi_1(B))^\frac{K^{2/n}t_0}{t_2}={\varepsilon}_n(\phi(B))^\frac{Kt}{t'}=\wt{\varepsilon}(B).$$
By recursively using Lemma \ref{mainlem}, we have
\begin{align*}
\mathcal M^{h}(\phi(E)) & = \mathcal M^{h_n}(E_n)\leq C(K)\, \mathcal M^{h_{n-1}}(E_{n-1})^{\frac{t_n}{K^{1/n} t_{n-1}}} \\
&\leq C(K)\,C(K)^\frac{t_n}{K^{1/n}t_{n-1}}\, \mathcal M^{h_{n-2}}(E_{n-2})^{\frac{t_n}{K^{2/n} t_{n-2}}}\\
&\leq C(K)^{1+\frac{t_n}{K^{1/n}t_{n-1}}+\dots+\frac{t_n}{K^{(n-1)/n}t_{1}}}\, \mathcal M^{h_0}(E_0)^{\frac{t_n}{K\,t_0 }}\\
&=C(K)^{1+(n-1)\left(1-\frac{t'}{2}\right)+\frac{t'}{2}\frac{K-K^{1/n}}{K^{1/n}-1}}\, \mathcal M^{h_0}(E)^{\frac{t'}{K\,t }}.
\end{align*}
Now, since $C_{pack}$ is fixed, we see that the constant above depends only on $K$. Therefore we can rewrite this in the following way
$$
\mathcal M^{h}(\phi(E))\leq C(K)\, \mathcal M^{\wt h}(E)^\frac{t'}{Kt}
$$
as claimed.
\end{proof}
\bigskip
\section{Quasiconformal distortion of Riesz capacities.}
The following lemma describes the relationship between Riesz capacities and $h$-contents.
\begin{lemma}\label{capMh}
Let $1<p<\infty$ and $0<\alpha p<2$, and let $E\subset\mathbb C$ be compact. Then
\begin{equation}\label{supr}
\dot{{\mathcal C}}_{\alpha,p}(E)\simeq \sup\left\{ \mathcal M^h(E)\right\}
\end{equation}
where the supremum on the right hand side runs over all gauge functions
$$h(x,r)=r^{2-\alpha p}\,{\varepsilon}(x,r)$$
with ${\varepsilon}\in{\mathcal G}_1$ and such that
\begin{equation}\label{normaliz}
\int_0^\infty\frac{{\varepsilon}(x,r)^{p'-1}}{r}\,dr\leq 1,\qquad x\in\mathbb C.
\end{equation}
\end{lemma}
\begin{proof}
We will use the characterization of $\dot{\mathcal C}_{\alpha,p}$ in terms of Wolff potentials (see Subsection \ref{strategy}). Let us consider a measure $\mu$ supported on $E$ and such that $\dot{W}_{\alpha,p}^\mu(x)\leq 1$ for all $x\in\mathbb C$, and let $t=2-\alpha p$, so that $0<t<2$. For small enough $a>0$, construct $h_{\mu,a,t}$ as we did in \eqref{defhx}. Recall that $h_{\mu,a,t}$ belongs to the class ${\mathcal G}_1$. By Lemma \ref{Mmu}, $ \mathcal M^{h_{\mu,a,t}}(E)\geq C\,\mu(E)$. Decomposing the integrals into annuli, for all $x\in\mathbb C$ we get
\begin{align*}
\int_0^\infty {\varepsilon}_{\mu,a,t}(x,r)^{p'-1}\frac{dr}r &= \int_0^\infty\frac1{r^{t(p'-1)}} \biggl(\int \psi_{a,t}\Bigl(\frac{y-x}r\Bigr)d\mu(y)\biggr)^{p'-1}\frac{dr}r\\
& \leq C \sum_{j\in \mathbb Z} 2^{-t(p'-1)j} \Bigl(\sum_{k>j}\mu(B(x,2^k)) 2^{(t+a)(j-k)}\Bigr)^{p'-1}\\
& \leq C \sum_{j\in \mathbb Z} 2^{-t(p'-1)j} \sum_{k>j}\mu(B(x,2^k))^{p'-1} 2^{(p'-1)(t+\frac a2)(j-k)},
\end{align*}
where we applied H\"older's inequality for $p'-1>1$, and the fact that $(c+d)^{p'-1}\leq c^{p'-1}+ d^{p'-1}$ otherwise. Thus,
\begin{align}\label{ed**}
\int _0^\infty{\varepsilon}_{\mu,a,t}(x,r)^{p'-1}\frac{dr}r &
\lesssim
\sum_{k\in \mathbb Z} \mu(B(x,2^k))^{p'-1}\,
2^{-(p'-1)(t+\frac a2)k}
\sum_{j<k} 2^{t(p'-1)\frac a2 j} \nonumber \\
& \simeq \sum_{k\in \mathbb Z} \mu(B(x,2^k))^{p'-1}\,
2^{-(p'-1)tk}
\lesssim \dot W^\mu_{\alpha,p}(x)\lesssim 1.
\end{align}
Above we allow constants in the estimates to depend on $\alpha,p,t,a$, but not on $\mu$. Therefore, if $\mu$ is admissible for $\dot{\mathcal C}_{\alpha,p}(E)$ then $h_{\mu,a,t}$ is admissible for the supremum in \eqref{supr} and
$$\sup_h\left\{ \mathcal M^h(E)\right\}\geq C\, \mathcal M^{h_{\mu,a,t}}(E)\geq C\,\mu(E).$$
Conversely, let us fix any gauge function $h(x,r)=r^{2-\alpha p}\,{\varepsilon}(x,r)$ in ${\mathcal G}_1$ satisfying \eqref{normaliz} and such that $ \mathcal M^h(E)>0$. By Lemma \ref{frostman}, there exists a measure $\mu$ supported on $E$ such that
$$\mu(B)\leq C\,h(B),$$
and furthermore, we can choose $\mu$ so that $\mu(E)\geq C\, \mathcal M^h(E)$. But then
$$
\dot W_{\alpha,p}^\mu(x)=\int_0^\infty\left(\frac{\mu(B(x,r))}{r^{2-\alpha p}}\right)^{p'-1}\frac{dr}{r}\leq\int_0^\infty\frac{{\varepsilon}(x,r)^{p'-1}}{r}\,dr\leq 1.
$$
Therefore $\mu$ is admissible for $\dot{{\mathcal C}}_{\alpha,p}(E)$, and $\dot{{\mathcal C}}_{\alpha,p}(E)\geq C\,\mu(E)\geq C\, \mathcal M^h(E).$
\end{proof}
\noindent
Let us remark that we can further restrict the class of admissible functions $h$ in the above supremum. In fact, it follows from the proof above that
\begin{equation}\label{caphmu}
\dot{{\mathcal C}}_{\alpha,p}(E)\simeq \sup\left\{ \mathcal M^{h_{\mu,a,t}}(E): \operatorname{supp}(\mu)\subset E,
\int_0^\infty\left(\frac{h_{\mu,a,t}(x,r)}{r^t}\right)^{p'-1}\,\frac{dr}{r}\leq 1\right\},
\end{equation}
where $t=2-\alpha p$, as above. On the other hand, we emphasize the fact that Lemma \ref{capMh} does not hold if we restrict the supremum to gauge functions $h$ invariant under translations (see \cite[Remark 5.6.4]{adamshedberg}).
Finally, if condition \eqref{normaliz} is replaced by
$$\lim_{r\to 0}\frac{h(x,r)}{r^{2-\alpha p}}= \lim_{r\to 0}{\varepsilon}(x,r)=0\hspace{1cm}\text{ uniformly in }x$$
then we obtain the lower $t$-dimensional Hausdorff content, which vanishes exactly on sets having $\sigma$-finite $t$-dimensional Hausdorff measure $ \mathcal H^t$ (see \cite{sionsjerve} for more details). \\
\\
\noindent
Before proving Theorem \ref{teocap2}, we need the following auxiliary result.
\begin{lemma} \label{lemcg}
Let $\phi:\mathbb C\rightarrow\mathbb C$ be a $K$-quasiconformal mapping, and
${\varepsilon}_0\in\mathcal G_1$.
Define ${\varepsilon}_1(B) = {\varepsilon}_0(\phi(B))$ for any ball $B\subset\mathbb C$.
For any $s>0$ we have
$$\int_0^\infty {\varepsilon}_1(x,r)^s\,\frac{dr}r\leq C(K,s)\int_0^\infty {\varepsilon}_0(\phi(x),r)^s\,\frac{dr}r.
$$
\end{lemma}
\begin{proof}
We have
$$
\int_0^\infty
{\varepsilon}_0(\phi(B(x,r)))^s\, \frac{dr}r
\leq C(s)\sum_{j\in\mathbb Z} {\varepsilon}_0(\phi(B(x, 2^j)))^s.$$
Denote now $r_j=\operatorname{diam}(\phi(B(x, 2^j))$. We obtain
\begin{align*}
\sum_{j\in\mathbb Z} {\varepsilon}_0(\phi(B(x, 2^j))^s & = \sum_{k\in\mathbb Z} \sum_{j:2^k\leq r_j< 2^{k+1}}\!\!\!\!{\varepsilon}_0(\phi(B(x, 2^j))^s \\ &\lesssim \sum_{k\in\mathbb Z} \sum_{j:2^k\leq r_j< 2^{k+1}}\!\!\!\!{\varepsilon}_0(B(\phi(x), r_j))^s \\
& \lesssim C(K) \sum_{k\in\mathbb Z} {\varepsilon}_0(B(\phi(x), 2^k))^s\leq C(K,s)\int_0^\infty
{\varepsilon}_0(\phi(x),r)^s\, \frac{dr}r,
\end{align*}
where we took into account that $\#\{j:2^k\leq r_j< 2^{k+1}\}\leq C(K)$. This follows easily from the fact that the moduli of the annuli $B(x,2^{j+1})\setminus B(x,2^j)$ is
$K$-quasi-invariant by $\phi$.
\end{proof}
\medskip
\begin{proof}[\bf Proof of Theorem \ref{teocap2}] By standard methods, we may assume that $\phi$ is a principal quasiconformal mapping, conformal on $\mathbb C\setminus \bar \mathbb D$, and that $E\subset B(0,1/2)=\frac12 B$ (and so $\operatorname{diam}(\phi(B))\simeq 1$). We can further assume that $\dot{\mathcal C}_{\beta,q}(\phi(E))>0$, since otherwise the statement is obvious. Let $a>0$ be small enough, and $t'=2-\beta q$. By \eqref{caphmu}, we can find a finite Borel measure $\mu$ supported on $\phi(E)$ such that
$$\dot{\mathcal C}_{\beta,q}(\phi(E))\simeq\mathcal M^{h_{\mu,a,t'}}(\phi(E)),$$
and
$$\int_0^\infty\left(\frac{h_{\mu,a,t'}(x,r)}{r^{t'}}\right)^{q'-1}\frac{dr}{r}\leq 1$$
for all $x\in\mathbb C$. Writing $h_{\mu,a,t'}(x,r)=r^{t'}\,{\varepsilon}_{\mu,a,t'}(x,r)$, we proved in Lemma \ref{lemkgran} that if
$$\wt h(x,r)=r^t\,\wt{\varepsilon}(x,r)\hspace{1cm}\text{and}\hspace{1cm}\wt{\varepsilon}(B)={\varepsilon}_{\mu,a,t'}(\phi(B))^\frac{Kt}{t'}$$
then we have the inequality
\begin{equation}\label{ineq}
\mathcal M^{h_{\mu,a,t'}}(\phi(E)) \leq C(K) \mathcal M^{\wt h}(E)^\frac{t'}{Kt},
\end{equation}
with a constant $C(K)>0$ depending only on $K$. Furthermore, using our choice
$$
\frac{1}{p'-1}=p-1=\frac{Kt}{t'}\,(q-1)=\frac{Kt}{t'}\,\frac{1}{q'-1},
$$
together with Lemma \ref{lemcg}, we get that
\begin{align*}
\int_0^\infty\left(\frac{\wt{h}(x,r)}{r^{2-\alpha p}}\right)^{p'-1}\,\frac{dr}{r}
&=\int_0^\infty \wt{\varepsilon}(x,r)^{p'-1}\,\frac{dr}r=\int_0^\infty{\varepsilon}_{\mu,a,t'}(\phi(B(x,r)))^{\frac{Kt(p'-1)}{t'}} \,\frac{dr}r \\
& \leq C\int_0^\infty{\varepsilon}_{\mu,a,t'}(\phi(x),r)^{\frac{Kt(p'-1)}{t'}} \,\frac{dr}r\\
&= C\, \int_0^\infty{\varepsilon}_{\mu,a,t'}(\phi(x),r)^{q'-1} \,\frac{dr}r\leq C
\end{align*}
for all $x\in\mathbb C$. Therefore, by Lemma \ref{capMh}, $\wt h$ is admissible for $\dot{\mathcal C}_{\alpha,p}(E)$, and taking supremum in \eqref{ineq} we get that
$$\dot{\mathcal C}_{\beta,q}(\phi(E))\leq C\,\dot{\mathcal C}_{\alpha,p}(E)^\frac{t'}{Kt},$$
as desired.
\end{proof}
\bigskip
\section{Proof of Corollary \ref{teosigma} using Lemma \ref{lemkgran}}
Although Corollary \ref{teosigma} is an immediate consequence of Theorem \ref{teopri},
which will be proved in next section, we would like to show that it also follows rather easily from
Lemma \ref{lemkgran}.
\\
\\
Suppose that $\mathcal H^{t'}(\phi(E))$ is non $\sigma$-finite. Then it supports some non zero measure $\mu$ such that
$\theta_\mu^{t'}(x)=0$ at $\mu$-a.e. $x\in\phi(E)$. We can assume $\theta_\mu^{t'}(x)=0$ for all $x\in\mathbb C$,
replacing $\mu$ by its restriction to some nice subset if necessary . It is easy to check that
this implies that ${\varepsilon}_{\mu,a,t'}(x,r)\to 0$ as $r\to 0$, for all $x\in\mathbb C$ (one only has to write ${\varepsilon}_{\mu,a,t'}(x,r)$
as a convex combination of $\mu(B(x,s))/s^{t'}$, $s\in(0,\infty)$).
As in Theorem \ref{teocap2}, using Lemma \ref{lemkgran} and Frostman Lemma, we deduce that
there exists another non zero measure $\nu$ supported $E$ such that $\nu(x,r) \leq r^{t}\wt {\varepsilon}(x,r)$
for all $x\in\mathbb C$ and $r>0$, with
$$\wt{\varepsilon}(x,r)= {\varepsilon}_{\mu,a,t'}(\phi(B(x,r)))^{\frac{Kt}{t'}}.$$
Since $\phi$ is continuous, we have
$$\lim_{r\to0}\wt{\varepsilon}(x,r)= \lim_{r\to0}{\varepsilon}_{\mu,a,t'}(\phi(B(x,r)))^{\frac{Kt}{t'}}
=0 \quad\mbox{ for all $x\in\mathbb C$,}$$
which implies that $\theta_\nu^{t}(x)=0$ for all $x\in\mathbb C$. The fact that $E$ supports a non zero measure with zero $t$-density $\nu$-a.e. implies that $\mathcal H^{t}(E)$ is non $\sigma$-finite.
{\hspace*{\fill} $\square$ \vspace{2pt}}
\bigskip
\section{Quasiconformal distortion of Hausdorff measures.}\label{tolsapreprint}
\noindent
First we need the following technical lemma.
\begin{lemma}\label{lemdens}
Let $0<s\leq 2$.
Let $\mu$ be a finite Borel measure, and let $x\in\mathbb C$ and $\theta_1>0$ be such that
\begin{equation}\label{eq21}
\frac{\mu(B(x,r))}{r^s} \leq \theta_1 \quad\mbox{ if $0< r\leq \delta$.}
\end{equation}
Then there exists $\delta'>0$, depending only on $\delta,a,s, \theta_1$ and $\mu(\mathbb C)$, such that
$${\varepsilon}_{\mu,a,s}(x,r)\leq \theta_2 \quad\mbox{ if $0< r\leq \delta'$,}
$$
with $\theta_2=C\,\theta_1$, with $C$ depending only on $a,s$.
\end{lemma}
\begin{proof}
By the definition of ${\varepsilon}_{\mu,a,s}$ and $\psi_{a,s}$,
\begin{align*}
{\varepsilon}_{\mu,a,s}(x,r) &= \frac1{r^s} \biggl[ \int_{|x-y|\leq r} + \sum_{j\geq1} \int_{2^{j-1}r<|x-y|\leq 2^jr} \biggr]\psi_{a,s}
\Bigl(\frac{y-x}r\Bigr)d\mu(y)\\
& \leq C \sum_{j\geq0} \frac{\mu(B(x,2^jr))}{2^{j(s+a)}\,r^s}.
\end{align*}
If $2^jr\leq\delta$, we use the estimate \eqref{eq21}. Otherwise, we take into account that
$$\frac{\mu(B(x,2^jr))}{2^{js}\,r^s} \leq \frac{\mu(\mathbb C)}{\delta^s}.$$
So if $N$ denotes the biggest integer such that $2^Nr\leq\delta$, then
\begin{align*}
{\varepsilon}_{\mu,a,s}(x,r) & \leq C \theta_1 \sum_{0\leq j\leq N} 2^{-ja}+ C
\frac{\mu(\mathbb C)}{\delta^s} \sum_{j\geq N+1} 2^{-ja} \\
& \leq C \biggl(\theta_1
+ 2^{-Na} \,\frac{\mu(\mathbb C)}{\delta^s}\biggr)
\leq C \biggl(\theta_1
+ \frac{r^a\mu(\mathbb C)}{\delta^{s+a}}\biggr).
\end{align*}
If we take $\delta'$ small enough so that
$$\frac{(\delta')^a\mu(\mathbb C)}{\delta^{s+a}} \leq \theta_1,$$
the lemma follows.
\end{proof}
\medskip
\begin{lemma}
Let $0<t<2$ and set $t'=\frac{2Kt}{2+(K-1)t}$. Let $\phi:\mathbb C\to\mathbb C$ be a principal $K$-quasiconformal mapping, conformal outside the unit disk, and let $E\subset B(0,1/2)$. Then
\begin{equation}\label{eqhaus2}
\mathcal H^{t'}(\phi(E))\leq C(K)\,\mathcal H^t(E)^\frac{t'}{Kt}.
\end{equation}
\end{lemma}
\begin{proof}
To prove \rf{eqhaus2}, we may assume that $\mathcal H^{t'}(\phi(E))>0$. Because of the estimates on the upper density of Hausdorff measures (see \cite[p.89]{mattila}), there exists $\delta>0$ and $F\subset E$ compact such that
$\mathcal H^{t'}(\phi(F))\geq \mathcal H^{t'}(\phi(E))/2$, and
\begin{equation}\label{eqdens}
\frac{\mathcal H^{t'}(\phi(E)\cap B(x,r))}{r^{t'}} \leq 5\quad \mbox{for all $x\in\phi(F)$ and $0<r\leq \delta.$}
\end{equation}
Let us denote $\mu=\mathcal H^{t'}_{|\phi(F)}$, and consider the associated gauge function
$$h_{\mu,a,t'}(x,r)=r^{t'}\,{\varepsilon}_{\mu,a,t'}(x,r)=\int\psi_{a,t'}\left(\frac{|x-y|}{r}\right)d\mu(y)$$
where $\psi_{a,t'}(r)=\frac1{1+r^{a+t'}}$, for $a>0$ small enough. Recall that, by Lemma \ref{Mmu}, $\mu\leq 2\mathcal M^{h_{\mu,a,t'}}$. Further, we can apply Lemma \ref{lemkgran} to $h_{\mu,a,t'}$ and ${\varepsilon}_{\mu,a,t'}$ (since they fulfill the required assumptions if $a$ is chosen small enough, by Lemma \ref{lemtec5}), and then we get
\begin{equation}\label{eq2}
\mathcal M^{h_{\mu,a,t'}}(\phi(F)) \leq C(K,t) \,\mathcal M^{\wt h}(F)^{\frac{t'}{Kt}},
\end{equation}
where $\wt h(x,r)= r^{t}\wt{\varepsilon}(x,r)$ and
$$\wt{\varepsilon}(x,r)= {\varepsilon}_{{\mu,a,t'}}(\phi(B(x,r)))^\frac{Kt}{t'}.$$
In particular, $\mathcal M^{\wt h}(F)>0$, and by Frostman's Lemma, there exists a measure $\nu$, supported on $F$, such that
\begin{equation}\label{eq61}
\nu(B(x,r))\leq \wt h(B(x,r)) = r^{t}{\varepsilon}_{\mu,a,t'}(\phi(B(x,r)))^\frac{Kt}{t'}.
\end{equation}
Furthermore, we can choose $\nu$ so that $\nu(F)\geq C\,\mathcal M^{\wt h}(F)$.
It now suffices to show that ${\varepsilon}_{\mu,a,t'}(\phi(B(x,r)))$ is uniformly bounded for $r$ small enough, as then $\nu(F)\leq C\mathcal H^t(F)$.
\\
\\
From \rf{eqdens} and Lemma \ref{lemdens} we infer that ${\varepsilon}_{\mu,a,t'}(y,s)\leq C_1$ for all $y\in \phi(F)$ and $0<s<\delta'$, with $\delta'=\delta'(\delta, a,t',\mu(F))$, and $C_1=C_1(a,t')$. As a consequence, if $\delta''>0$ is taken small enough, then
\begin{equation}\label{growtheps}
{\varepsilon}_{\mu,a,t'}(\phi(B(x,r))\leq C_1,\hspace{1cm}\text{for all $x\in F$ and $0<r<\delta''$.}
\end{equation}
To see this, first by quasisymmetry
$$B(\phi(x),r_1)\subset \phi(B(x,r))\leq B(\phi(x),r_2)\hspace{1cm}\text{ for some }r_2\leq C(K) r_1.$$
Moreover, from the local H\"older continuity of $K$-quasiconformal mappings,
$$\frac{2r_1}{\operatorname{diam}\phi(\mathbb D)}\leq C(K)\,\left(\frac{r}{\operatorname{diam}\mathbb D}\right)^\frac1K,$$
and since $\phi$ is conformal on $\mathbb C\setminus\mathbb D$, $|\phi(\mathbb D)|\leq C(K)\,|\mathbb D|$. Hence $r_1\leq C(K)r^{1/K}$, and therefore ${\varepsilon}_{\mu,a,t'}(\phi(B(x,r))\leq{\varepsilon}_{\mu,a,t'}(B(\phi(x),r_2))\leq C_1$ whenever $r_2<\delta'$, which immediately follows if $r<\delta''=C(K)\,(\delta')^K$. This proves \eqref{growtheps}.\\
\\
From \eqref{growtheps} and \rf{eq61}, we immediately get that
$$\nu(B(x,r))\leq C_3r^{t}\quad\mbox{all $x\in F$ and $0<r<\delta''$,}$$
with $C_3=C_3(K,t,t')$. It is easy to check that this implies that
$\mathcal H^{t}(F)\gtrsim \nu(F)$.
Indeed, if
$F\subset \bigcup_i A_i$, with $\operatorname{diam}(A_i)\leq d\leq \delta''$ and $A_i\cap
F\neq\varnothing$, we take a ball
$B_i$ centered on $F\cap A_i$ with radius $r(B_i)=\operatorname{diam}(A_i)\leq \delta''$ for each $i$, and then
$$\sum_i\operatorname{diam}(A_i)^{t}= \sum_i r(B_i)^{t}\gtrsim \sum_i \nu(B_i) \geq \nu(F),$$
and so $\mathcal H^{t}_{d}(\phi(F))\gtrsim \nu(F)$ for all $0<d<\delta''$. Letting $d\to0$ our claim follows.
\\
\\
Summarizing,
$$\aligned
\mathcal H^t(E)\geq \mathcal H^t(F)&\geq C\,\nu(F)\geq C\,\mathcal M^{\wt h}(F)\geq C\,\mathcal M^{h_{\mu,a,t'}}(\phi(F))^\frac{Kt}{t'}\\
&\geq C\,\mu(\phi(F))^\frac{Kt}{t'}\geq C\,\mathcal H^{t'}(\phi(F))^\frac{Kt}{t'}\geq C\,\mathcal H^{t'}(\phi(E))^\frac{Kt}{t'}.
\endaligned$$
\end{proof}
\medskip
\begin{proof}[\bf Proof of Theorem \ref{teopri}]
The theorem follows from the preceding lemma by standard arguments in quasiconformal theory. However, for completeness we give the details. We factorize $\phi=\phi_2\circ \phi_1$, where $\phi_1$, $\phi_2$ are both $K$-quasiconformal maps, with $\phi_1$ principal and conformal on $\mathbb C\setminus 2B$, and
$\phi_2$ is conformal on $\phi_1(2B)$. Let $g(z)=dz+ b$ be the linear function that maps the unit disk to $2B$ (so $d=\operatorname{diam}(B)$). The function $h=g^{-1}\circ \phi_1\circ g$ verifies the assumptions of the main lemma, so that
$$\mathcal H^{t'}(g^{-1}\circ\phi_1(E)) \leq C(K) \,\mathcal H^t(g^{-1}(E))^{\frac{t'}{Kt}}.$$
On the other hand,
$$\mathcal H^{t'}(g^{-1}\circ\phi_1(E)) = \frac{\mathcal H^{t'}(\phi_1(E))}{\operatorname{diam}(B)^{t'}},\qquad
\mathcal H^t(g^{-1}(E))=\frac{\mathcal H^t(E)}{\operatorname{diam}(B)^t}.$$
Using also quasisymmetry and Koebe's distortion theorem, we get that $\operatorname{diam}(\phi_1(B))\simeq\operatorname{diam}(\phi_1(2B))\simeq\operatorname{diam}(2B)$ with constants depending only on $K$. Hence
\begin{equation}\label{eqc4}
\frac{\mathcal H^{t'}(\phi(E))}{\operatorname{diam}(\phi_1(B))^{t'}}\leq C(K)\,\left(\frac{\mathcal H^t(E)}{\operatorname{diam}(B)^t}\right)^{\frac{t'}{Kt}}.
\end{equation}
Now, since $\phi_2$ is conformal on $\phi_1(2B)$, by Koebe's distortion theorem and quasisymmetry, for each ball $B_0$ contained in $B$ we have
$$\frac{\operatorname{diam}(\phi_2(\phi_1(B_0)))}{\operatorname{diam}(\phi_2(\phi_1(2B)))}\simeq \frac{\operatorname{diam}(\phi_1(B_0))}{\operatorname{diam}(\phi_1(2B))}.$$
From this estimate and quasisymmetry again, it is straightforward to check that
$$\frac{\mathcal H^{t'}(\phi_1(E))}{\operatorname{diam}(\phi_1(B))^{t'}}\simeq \frac{\mathcal H^{t'}(\phi(E))}{\operatorname{diam}(\phi(B))^{t'}},$$
with constants depending on $K$, which together with \rf{eqc4} yields \rf{eqhaus}.
\end{proof}
\bigskip
\section{Examples showing sharpness of results}\label{SectionExamples}
In \cite[Thm 2.2]{uriartesharpqcstretching}, an example was constructed of a $K$-quasiconformal map $\phi: \mathbb C \to \mathbb C$ and a compact set $E \subset \mathbb D$ such that $diam(E) \simeq diam (\phi E) \simeq 1$ and such that, for $0<t<2$ and $t' = \frac{2Kt}{2+(K-1)t}$, $\mathcal H^t(E) \simeq \mathcal H^{t'}(\phi E) \simeq 1$. In the same paper, \cite[Cor 3.5]{uriartesharpqcstretching}, an example is constructed with the same hypotheses, except that both $\mathcal H^t(E)$ and $\mathcal H^{t'}(\phi E)$ are sigma-finite (but infinite.) These prove the sharpness of Corollary \ref{teosigma} and Theorem \ref{teopri}.
We will show next that Theorem \ref{teocap2} is sharp. This was already shown for the case $\beta = \frac{2}{3}$, $q = \frac{3}{2}$ in \cite[Thm 8.8]{Tolsa-Uriarte}. We will follow the scheme in \cite{Tolsa-Uriarte} very closely, repeating some of the arguments from \cite{Tolsa-Uriarte} for the convenience of the reader.
\subsection{Basic construction for the subsequent examples}\label{BasicConstructionCantorSets}
Following the scheme of \cite{Tolsa-Uriarte}, we argue as in \cite{uriartesharpqcstretching}. We assume the reader is familiar with the latter paper and we will use the notation from it without further reference. The formulae look slightly nicer if we assume in the construction that $\varepsilon_n = 0$ for all $n$, i.e. that we take infinitely many disks in each step, completely filling the area of the unit disk $\mathbb D$ (see equations (2.1), (2.2) and (2.3) in \cite{uriartesharpqcstretching}.) It is not strictly needed to set in that construction $\varepsilon_n = 0$ for all $n$, and we will later indicate the corresponding formulae if $\varepsilon_n > 0$ for all $n$ (which is the case in \cite{uriartesharpqcstretching}.) The construction in \cite{uriartesharpqcstretching} works as well if we set $\varepsilon_n = 0$ for all $n$, the only point that the reader might wonder about is whether the resulting map is $K$-quasiconformal.
However, this can be easily seen by a compactness argument (approximating the desired map by maps with finitely many circles in each step which are $K$-quasiconformal and have more and more disks in each step of the construction).
So we get (see equations (2.5) and (2.6) in \cite{uriartesharpqcstretching}) a Cantor type set $E$ and a $K$-quasiconformal map $\phi$ so that a building block in the $N$-th step of the construction of the source set $E$ is a disk with radius
\begin{equation}\label{RadiusSourceNthStep}
s_{j_1,...,j_N}=\left( (\sigma_{1,j_1})^K \, R_{1,j_1} \right) \dots \left( (\sigma_{N,j_N})^K R_{N,j_N} \right) ,
\end{equation}
and a building block in the $N$-th step of the construction in the target set $\phi(E)$ is a disk with radius given by
\begin{equation}\label{RadiusTargetNthStep}
t_{j_1,...,j_N}=\left( \sigma_{1,j_1} \, R_{1,j_1} \right) \dots \left( \sigma_{N,j_N} \, R_{N,j_N} \right) .
\end{equation}
Now we consider a measure $\mu$ supported on $\phi(E)$ (which will be the ``large" set of dimension $t'$) and its image measure $\nu = \phi^{-1}_\ast \mu$ supported on $E$ (which will be the ``small" set of dimension $t$) given by splitting the mass according to area. More explicitly,
\begin{equation}\label{DefinitionOfMuStep0}
\mu (\mathbb D) = 1,
\end{equation}
for any disk $B_{1,j_1} = \psi^{i_1}_{1,j_1} \left( \, \overline{\mathbb D} \, \right)$ of the first step of the construction with radius $t_{j_1} = \left( \sigma_{1,j_1} \, R_{1,j_1} \right)$,
\begin{equation}\label{DefinitionOfMuStep1}
\mu (B_{1,j_1}) = \left( R_{1,j_1} \right)^2,
\end{equation}
and in general, for any disk $B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} = \psi^{i_1}_{1,j_1} \circ \dots \circ \psi^{i_N}_{N,j_N} \left( \, \overline{\mathbb D} \, \right)$ of the $N^{th}$ step of the construction with radius
$t_{j_1,...,j_N}=\left( \sigma_{1,j_1} \, R_{1,j_1} \right) \dots \left( \sigma_{N,j_N} \, R_{N,j_N} \right) $,
\begin{equation}\label{DefinitionOfMuStepN}
\mu (B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N}) = \left( R_{1,j_1} \dots R_{N,j_N} \right)^2\ .
\end{equation}
Since we took $\varepsilon_N = 0 $ for all $N$, the total mass of $\mu$ is $1$ in every step. (If one prefers to take $\varepsilon_N > 0 $ for all $N$, the definition should be changed to $\mu (B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N}) = \left( R_{1,j_1} \dots R_{N,j_N} \right)^2\ \ \prod_{n=N+1}^{\infty} \left( 1- \varepsilon_n \right) $, and the total mass of $\mu$ is renormalized by the factor $\prod_{n=1}^{\infty} \left( 1- \varepsilon_n \right) >0
$, but otherwise the rest of the construction we are about to describe works well.)
Since $\nu$ is the image measure, for any disk $D_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} = \varphi^{i_1}_{1,j_1} \circ \dots \circ \varphi^{i_N}_{N,j_N} \left( \, \overline{\mathbb D} \, \right) = \phi^{-1}( B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} = \varphi^{i_1}_{1,j_1} \circ \dots \circ \varphi^{i_N}_{N,j_N} \left( \, \overline{\mathbb D} \, \right) )$
we get
\begin{equation}\label{DefinitionOfNuStepN}
\nu (D_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N}) = \left( R_{1,j_1} \dots R_{N,j_N} \right)^2\ .
\end{equation}
The following lemma simplifies the computation of the Wolff potentials for the Cantor type sets just described. It was proved in \cite{Tolsa-Uriarte}, but we recall it here (as well as its proof) for the convenience of the reader.
\begin{lemma}\label{ComputingWolffPotentialInOurCantorSets}
For the Cantor type sets just described (in subsection \ref{BasicConstructionCantorSets}), for any $\alpha, p >0$ with $\alpha p < 2$, and for $x \in \phi (E)$, the Wolff potentials satisfy
$$\dot W^\mu_{\alpha,p}(x) \simeq \sum_n \biggl(\frac{\mu(B(x,2^{n}))}{2^{n(2-\alpha p)}}\biggr)^{p'-1}\, \simeq
\sum_{ N : x \in B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} } \biggl(\frac{\mu( B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} ) }{ \left( t_{j_1,...,j_N} \right)^{(2-\alpha p)}}\biggr)^{p'-1} \ ,
$$
and analogously for $\nu$, $D_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N}$ and $s_{j_1,...,j_N}$.
\end{lemma}
\begin{proof}
We first introduce some convenient notation.
For any multiindexes $I=(i_1,...,i_N)$ and $J=(j_1,...,j_N)$, where $1\leq i_k ,
j_k \leq \infty $ (since we are taking infinitely many disks in each step of the construction),
we will denote by
\begin{equation}\label{DefinitionProtectingDisk}
P^{N}_{I;J} = \frac{1}{\sigma_{N,j_N}}\, \psi^{i_1}_{1,j_1} \circ \dots \circ \psi^{i_N}_{N,j_N}(\mathbb D)
\end{equation}
a {\it{protecting}} disk of generation $N$. Then, $P^N_{I;J}$ has radius
$$r(P^{N}_{I;J}) = \frac{1}{ \sigma_{N,j_N} } t_{j_1,...,j_N}=\left( \sigma_{1,j_1} \, \dots \sigma_{N-1,j_{N-1}} \right) \left( R_{1,j_1} \dots R_{N,j_N} \right).$$
Analogously, we will write
\begin{equation}\label{DefinitionGeneratingDisk}
G^{N}_{I;J} = \psi^{i_1}_{1,j_1} \circ \dots \circ \psi^{i_N}_{N,j_N}(\mathbb D)
\end{equation}
to denote a {\it{generating}} disk of generation $N$, which has radius $$r(G^{N}_{I;J}) = t_{j_1,...,j_N}=\left( \sigma_{1,j_1} \, \dots \sigma_{N,j_N} \right) \left( R_{1,j_1} \dots R_{N,j_N} \right).$$
Notice that, since all values of $\sigma_{n,j_n}$ and $R_{n,j_n}$ are $\leq \frac{1}{100}$, then $\mu (G^{N}_{I;J}) = \mu (2 G^{N}_{I;J})$, so we can pretend without loss of generality that the radii $t_{j_1,...,j_N}$ are of the
form $2^k$, $k\in\mathbb Z$.
Now, if $ r(G^{N}_{I;J}) \lesssim t \lesssim r(P^{N}_{I;J}) $, and $x \in \phi (E)$ so that $B(x, t) \subseteq P^{N}_{I;J}$, then $\mu ( B(x,t)) = \mu ( G^{N}_{I;J} )$, so that
$$
\sum_{n : G^{N}_{I;J} \subseteq B(x,2^{n}) \subseteq P^{N}_{I;J}} \biggl(\frac{\mu(B(x,2^{n}))}{2^{n(2-\alpha p)}}\biggr)^{p'-1}
$$
is a geometric series with sum comparable, with constants depending on $\alpha$ and $p$, to its largest term, namely
to $ \biggl(\frac{\mu( G^{N}_{I;J} ) }{ \left( t_{j_1,...,j_N} \right)^{(2-\alpha p)}}\biggr)^{p'-1} $.
And if $ r(P^{N}_{I;J}) \lesssim t \lesssim r(G^{N-1}_{I';J'}) $, where $G^{N-1}_{I';J'}$ is the unique generating disk of generation $N-1$ containing $P^{N}_{I;J}$, and $x \in \phi (E)$, so that $P^{N}_{I;J} \subseteq B(x, t) \subseteq G^{N-1}_{I';J'}$, then
\begin{equation}\label{UpperBoundForMuOfBallIfBallInBetweenProtectingNAndGeneratingNMinus1}
\mu ( B(x,t)) \lesssim \frac{ t^2}{ \left( \sigma_{1,j_1} \, \dots \sigma_{N-1,j_{N-1}} \right) \left( R_{1,j_1} \dots R_{N-1,j_{N-1}} \right)} \left( R_{1,j_1} \dots R_{N-1,j_{N-1}} \right)^2 \ ,
\end{equation}
i.e., the mass that $\mu$ assigns to $B(x,t)$ is proportional to its area once $G^{N-1}_{I';J'}$ is renormalized to $\mathbb D$, but multiplied by the mass that $\mu$ assigns to $G^{N-1}_{I';J'}$, namely $\left( R_{1,j_1} \dots R_{N-1,j_{N-1}} \right)^2$. Hence
$$
\sum_{n : P^{N}_{I;J} \subseteq B(x,2^{n}) \subseteq G^{N-1}_{I';J'}} \biggl(\frac{\mu(B(x,2^{n}))}{2^{n(2-\alpha p)}}\biggr)^{p'-1}
$$
is dominated by a geometric series (if $n$ appears in the above sum and $2^{n} = \frac{ r(G^{N-1}_{I';J'}) }{2^k}$ with $k >0$, then $$ \biggl(\frac{\mu(B(x,2^{n}))}{2^{n(2-\alpha p)}}\biggr)^{p'-1} \lesssim \biggl(\frac{\mu( G^{N-1}_{I';J'})}{r(G^{N-1}_{I';J'})^{(2-\alpha p)}} \ \frac{ 2^{k(2-\alpha p)} }{2^{2k} } \biggr)^{p'-1},$$ and hence the above sum is $ \lesssim \biggl(\frac{\mu( G^{N-1}_{I';J'} )}{r(G^{N-1}_{I';J'})^{(2-\alpha p)}}\biggr)^{p'-1}$, with constants depending only on $\alpha$ and $p$.)
\end{proof}
\subsection{Example}\label{CantorSetExample
In order to see that Theorem \ref{teocap2} is sharp, it is useful to recall Theorem 5.5.1 (b) in \cite{adamshedberg} adapted to our situation (and combined with Proposition 5.1.4):
\begin{theorem}\label{TheoremRelatingCapacities}
Let $E \subset \mathbb C$. Then there is a constant $A$ such that
$$
\dot C_{\beta, q} (E) \leq A \dot C_{\alpha, p} (E) \ ,
$$
for $\beta q = \alpha p = 2-\frac{2}{K+1} = \frac{2K}{K+1}$, $p<q$.
Moreover, there exist sets $E$ such that $\dot C_{\beta, q} (E) =0$ but $\dot C_{\alpha, p} (E) >0$.
\end{theorem}
Hence it is conceivable that Theorem \ref{teocap2} might be strengthened to a statement of the form
$$
\frac{\dot{C}_{\beta,q}(\phi(E))}{\operatorname{diam}(\phi(B))^{t'}}\leq C(\beta,q,K)\,\left(\frac{\dot{C}_{\widetilde{\alpha}, \widetilde{p}}(E)}{\operatorname{diam}(B)^t}\right)^\frac{t'}{Kt} \;,
$$
for some $\widetilde{\alpha}, \widetilde{p}$ such that $\widetilde{\alpha} \widetilde{p} = \alpha p = 2-t$ and $\widetilde{p} > p$. The following theorem shows that the answer to this question is negative.
\begin{theorem}\label{teosharp}
For any $\widetilde{\alpha}, \widetilde{p} >0$ such that $\widetilde{\alpha} \widetilde{p} = \alpha p = 2-t$ and $ \widetilde{p} > p$, there exists a compact $E \subset \mathbb C$ and a $K$-quasiconformal map $\phi$ such that $\dot C_{ \beta, q}(\phi E) >0$ (and hence $\dot C_{ \alpha , p }(E) >0$, due to Theorem \ref{teocap2}), but $ \dot C_{ \widetilde{\alpha}, \widetilde{p}}(E) = 0$.
\end{theorem}
\begin{proof}
For $E$ and $\phi$ as in Subsection \ref{BasicConstructionCantorSets},
notice that by Lemma \ref{ComputingWolffPotentialInOurCantorSets}, for $x \in \phi (E)$
\begin{eqnarray*}
\dot W^{\mu}_{ \beta, q } (x) & \simeq & \sum_{ N : x \in B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} } \biggl(\frac{\mu( B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} ) }{ \left( t_{j_1,...,j_N} \right)^{2-\beta q} } \biggr)^{q'-1} \\
& = & \sum_{ N : x \in B_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} } \biggl(\frac{ \left( R_{1,j_1} \dots R_{N,j_N} \right)^2 }{ \left( \sigma_{1,j_1} \dots \sigma_{N,j_N} R_{1,j_1} \dots R_{N,j_N} \right)^{2-\beta q} }\biggr)^{q'-1}.
\end{eqnarray*}
Since on the one hand $E$ is very ``close" to satisfying $0 < \mathcal H^{t}(E) < \infty$
and $0 < \mathcal H^{t'}(\phi E) < \infty$ (see (3.9), (3.10) and (4.5) in \cite{uriartesharpqcstretching}) and, on the other hand, an important element in
the proof of the semiadditivity of analytic capacity is that the potential is ``approximately
constant" on each scale (see \cite{tolsasemiadditivityanalyticcapacity}), the above equation
suggests the choice
\begin{equation}\label{ChoiceOfSigma}
\sigma_{N,j_N} = \left( R_{N,j_N} \right)^{\frac{2-t}{tK}} \, d_N\quad\mbox{ for all $N$,}
\end{equation}
where $d_N \in [1,2]$ is a parameter to be determined, independent of $j_N$.
If we take
\begin{equation}\label{ChoiceOfDjForExample2}
d_j = \left( \frac{j+1}{j} \right)^\delta,
\end{equation}
for an appropriate $\delta >0$ to be chosen later, then for $x \in \phi E$, we have
\begin{equation}\label{FirstConditionForDelta}
\dot W^{\mu}_{ \beta, q } (x) \simeq \sum_n \left\{ \prod_{j=1}^{n} \frac{1}{\left( d_j \right)^{t' (q' -1)}} \right\} = \sum_{n=2}^{\infty} \frac{1}{n^{t' (q' -1) \delta}} < \infty,
\end{equation}
once $\delta >0$ is appropriately chosen, so that $ \dot C_{ \beta, q } (\phi E) >0$ (and hence $\dot C_{ \alpha , p }(E) >0$, due to Theorem \ref{teocap2}.)
By Lemma \ref{ComputingWolffPotentialInOurCantorSets} and
\rf{RadiusSourceNthStep}, for $x \in E$,
\begin{eqnarray*}
\dot W^{\nu}_{ \widetilde{\alpha}, \widetilde{p} } (x) & \simeq & \sum_{ N : x \in D_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} } \biggl(\frac{\nu( D_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} ) }{ \left( s_{j_1,...,j_N} \right)^{ 2- \widetilde{\alpha} \widetilde{p} }}\biggr)^{\widetilde{p}' - 1} \\
& = & \sum_{ N : x \in D_{N ; j_1, \dots , j_N}^{i_1, \dots , i_N} } \biggl(\frac{ \left( R_{1,j_1} \dots R_{N,j_N} \right)^2 }{ \left[ \left( \sigma_{1,j_1} \dots \sigma_{N,j_N} \right)^{K} \left( R_{1,j_1} \dots R_{N,j_N} \right) \right]^{ 2- \widetilde{\alpha} \widetilde{p} }
}\biggr)^{\left( \widetilde{p}'-1 \right) } \ ,
\end{eqnarray*}
so that, substituting $\sigma_{N,j_N} = \left( R_{N,j_N} \right)^{\frac{2-t}{tK}} \, d_N$ and $d_j = \left( \frac{j+1}{j} \right)^\delta$ we get, for $x \in E$,
\begin{equation}\label{SecondConditionForDelta}
\dot W^{\nu}_{ \widetilde{\alpha}, \widetilde{p} } (x) \simeq \sum_n \left\{ \prod_{j=1}^{n} \frac{1}{\left( d_j \right)^{tK}} \right\}^{\left( \widetilde{p}'-1 \right) } =
\sum_{n=2}^{\infty} \frac{1}{n^{t K (\widetilde{p}' -1) \delta}} = \infty,
\end{equation}
once $\delta >0$ is appropriately chosen, so that $ \dot C_{ \widetilde{\alpha}, \widetilde{p} } ( E) =0$.
In order that both \eqref{FirstConditionForDelta} and \eqref{SecondConditionForDelta} be satisfied, it is enough to choose $\delta > 0$ so that $ t K (\widetilde{p}' -1) \delta \leq 1 < t K (p' - 1) \delta $ (since $t K (p' - 1) = t' (q' -1)$, which by the way implies that, in our example, $ \dot W^{\mu}_{ \beta, q } ( \phi x) \simeq \dot W^{\nu}_{\alpha, p } (x) $ for $x \in E$.) Since $\widetilde{p} > p $, this can be achieved by setting
$$
\delta = \frac{ 1}{ t K (\widetilde{p}' -1) } \;.
$$
\end{proof}
Let us remark that the above example also gives that $\dot C_{\gamma,r}(E)=0$ if
$\gamma\,r<\widetilde{\alpha}\,\widetilde{p} = \alpha\,p = 2-t$. This due to the fact that there is some constant $A$
independent of $E$ such that
$$\dot C_{\gamma,r}(E)^{1/(2-\gamma r)}\leq A \,\dot C_{\alpha,p}(E)^{1/(2-\alpha p)}.$$
See Theorem 5.5.1 of \cite{adamshedberg}.
\bigskip
\bigskip
\noindent{\bf Acknowledgements:}
Part of this paper was done while all authors were attending the research semester
``Harmonic Analysis, Geometric Measure Theory and Quasiconformal Mappings'' in
the CRM (Barcelona) in 2009.
\noindent
K.A.\ is supported by Academy of Finland, grants SA-134757
and CoE in Analysis and dynamics research, grant SA-18634, and EU-network CODY, grant
400630.
A.C.\ and J.V.\
are partially supported by grants 2009-SGR-420 (Catalonia) and MTM2007-60062 (Spain). A.C. is also partially supported by EU-network CODY.
X.T.\ is partially supported by grants 2009-SGR-420 (Catalonia) and MTM2007-62817 (Spain). I.U.\
was a postdoctoral fellow at the University of Missouri, Columbia, USA, and at Centre de Recerca Matem\`{a}tica, Barcelona, Spain, for some periods of time during the elaboration of this paper. He was partially supported by NSF grant DMS-0901524.
\bigskip
\bibliographystyle{alpha}
|
\section{Introduction}
The existence of long-lived massive particles is always welcome in
theories beyond the Standard Model (SM) if their lifetime is longer
than the age of the universe (or completely stable) as they provide a
candidate for the cold Dark Matter (DM) component of the universe to
account for the abundance, $\Omega_{\rm CDM} h^2 \simeq 0.11$,
inferred by cosmological observations. However, long-lived particles
decaying after Big Bang Nucleosynthesis (BBN) place the theory in a
very difficult situation when confronted with cosmological and
astrophysical observations\footnote{Decay products of particles with
lifetimes as long as $10^{27}$ sec., see e.g. \cite{Picciotto:2004rp,Hooper:2004qf,Yuksel:2007dr}, can still
produce observable signatures at present. Particles with longer
lifetimes are only constrained from the measured relic
density.}. The very successful predictions of standard BBN are
spoiled by the energetic products of the decay that can dissociate the
produced light elements for lifetimes from $10^2$ to $10^{10}$
seconds. If the lifetime is between $10^{10}$ and $10^{13}$ seconds,
very stringent constraints come from the shape of the Cosmic Microwave
Background (CMB) spectrum. For longer lifetimes, the emitted photons
can reach us today and be observed as part of the Diffuse
Extragalactic Background RAdiation (DEBRA). These observations
restrict strongly the released energy in the decays of the long-lived
particle at different times and hence its abundance, mass and
lifetime.
The paradigm of such long-lived particle conflicting with cosmological
and astrophysical observations is the gravitino in supersymmetric
theories. The gravitino is the spin $3/2$ supersymmetric partner of
the graviton and the couplings of the gravitino to ordinary matter are
gravitational couplings suppressed by the Planck mass, $M_{\rm Pl}=(8
\pi G_N)^{-1/2}\simeq 2.4\times 10^{18}\gev$. Such small couplings
make the typical gravitino lifetime (with electroweak scale masses and
neglecting the masses of the decay products) of the order of $10^8$
seconds. Therefore, they decay after BBN and the gravitino abundance
at BBN is severely constrained \cite{Moroi:1995fs}. This has very
important consequences in the phenomenology of this model. In fact,
these BBN constraints forbid reheating temperatures larger than $10^6$
or $10^7$ GeV, which precludes thermal leptogenesis from generating a
large enough baryon asymmetry.
It is interesting to check whether it is possible to evade these
stringent bounds. However, given that the gravitino couplings are
completely fixed by supergravity, it is very difficult to change the
gravitino production or its decay. The only ``free'' parameter available
(from an effective theory point of view) is the gravitino mass
itself and similarly the masses of the decay products. The gravitino
gets a mass after supersymmetry breaking and the different
soft-breaking terms receive a contribution proportional to the
gravitino mass. If supergravity is the mechanism of mediation of the
supersymmetry breaking from the hidden sector to the visible sector,
we can expect the SUSY masses and the gravitino to be of the same
order\footnote{In principle, in gauge mediation or anomaly mediation models, constraints from gravitino decays can be evaded as gravitino mass can be, respectively, much smaller or much larger than typical SUSY masses}.
However, most phenomenological analyses of BBN constraints assume that
the released energy in the thermal plasma is of the same order as the
gravitino mass itself. Although this is correct in most of the cases
where the mass difference between the gravitino and the SUSY particles
(typically the lightest MSSM supersymmetric particle (MLSP)) is sizeable, we
can ask what happens if the gravitino and the MSSM LSP masses are much
closer, $\Delta M = m_{3/2} - m_{\rm MLSP} \ll m_{3/2}$. This has two different
consequences, first it is evident that the released energy in the
thermal plasma is much smaller and this can help to relax the previous
bounds. On the other hand, this small mass difference increases the
gravitino lifetime and other constraints as the CMB spectrum or
diffuse gamma rays come into play.
In this work, we will analyze this scenario that we call ``degenerate
gravitino'' scenario, including both the case where the gravitino is
the NLSP decaying to the MSSM LSP and the case where the gravitino
itself is the LSP and all SUSY particles decay into the gravitino. In
the following, we define $\delta$ as the degree of degeneracy between
NLSP and LSP
\begin{equation}
\delta\equiv \frac{\mnlsp-\mlsp}{\mlsp} =
\frac{\Delta M}{\mlsp}
\label{delta}
\end{equation}
In the degenerate gravitino scenario, the total dark matter abundance
has two sources: one is the usual LSP component from thermal
production and decoupling and a second one from the NLSP non-thermal
decays to LSP and SM particles. The sum of both components should
reproduce the observed CDM relic density. The phenomenology of the
degenerate gravitino scenario depends on both the identity of the
lightest MSSM supersymmetric particle and on whether the gravitino is
the LSP or the NLSP. The BBN, CMB and DEBRA bounds apply to the NLSP
abundance while the LSP abundance is only constrained by the total
dark matter abundance. In most cases, the lightest MSSM supersymmetric particle can
be either the lightest neutralino or the lightest stau. Given that
dark matter can not be a charged particle, we are left with three
possibilities: gravitino LSP with neutralino or stau NLSP and gravitino
NLSP with neutralino LSP. We will see that, for lifetimes smaller than the age of the universe, the strongest constraints
on the reheating temperature arise in the case of gravitino NLSP,
where all the previous constraints apply to the gravitino abundance
and thus on the reheating temperature. On the other hand, when the
gravitino is the LSP, it is possible to reach reheating temperatures
compatible with thermal leptogenesis provided that the NLSP abundance
at decoupling is sufficiently suppressed. Clearly, if the NLSP lifetime is
much longer than the age of the universe, corresponding to small enough $\delta$,
the maximal reheating temperatures consistent with the observed DM abundance
($\treh \simeq 4.1 \times 10^9$ GeV) can be reached.
In this work, we
will reanalyze the different constraints on the energy injected by
NLSP decays and the reheating temperature in terms of
this parameter $\delta$ in the three above mentioned cases.
In the next section, we will present the constraints on the
energy release for different lifetimes of the NLSP in a model
independent way. In Section \ref{sec3} we apply these
constraints to the gravitino case where the energy release and the
lifetime are related. In section \ref{sec4} we analyse the case of
the CMSSM looking for the new constraints on the reheating temperature
in the degenerate gravitino scenario. Finally in section
\ref{sec:conclusions} we present our conclusions.
\section{Model-independent bounds}
\label{sec1}
If the NLSP lifetime is smaller than the age of the Universe, the
energetic decay products can significantly affect cosmology. Even for
lifetimes slightly larger than the age of the universe, observation of
diffuse gamma rays can constraint this scenario.
In this section, we consider the model-independent constraints on the
degenerate NLSP-LSP scenario. As we will see, depending on the NLSP
lifetime, these constraints originate from BBN observations, CMB
spectral distortion or searches of diffuse gamma rays. There are additional
constraints if the NLSP is charged. We focus on the case that the NLSP and LSP
masses are nearly degenerate i.e. $\delta\ll 1$ in \eq{delta}.
\subsection{Relic abundance constraint}
The first constraint that should be considered is that the total relic density of cold dark matter must match the observed value. We will consider the situation where only the LSP and the NLSP are relevant; that is, heavier MSSM particles have already decayed to LSP/NLSP before BBN takes place. In general, the final abundance of the LSP will have two components. The first one is the thermal abundance, $\Omega^{\rm TP}_{\rm LSP}$, which comes from thermal processes occurring in the plasma like scatterings and freeze-out. The second one is the non-thermal component, $\Omega^{\rm NTP}_{\rm LSP}$, that includes the contribution from LSP particles produced in NLSP decay. Therefore, the total cold dark matter relic density, will be \footnote{Recall that we are assuming that the NLSP lifetime is smaller than the age of the Universe. If it were not the case, Eq.~(\ref{abundance}) is still valid as we are considering $m_{\rm NLSP}\simeq m_{\rm LSP}$.}
\begin{equation}
\Omega_{\rm CDM}\,h^2 = \Omega^{\rm TP}_{\rm LSP}\,h^2+
\frac{1}{1+\delta} \, \Omega^{\rm TP}_{\rm NLSP}\,h^2\,,
\label{abundance}
\end{equation}
where the last term stands for $\Omega^{\rm NTP}_{\rm LSP}$ and we have used \eq{delta}. Assuming no late entropy release, Eq.~(\ref{abundance}) represents the present amount of dark matter, which should be equated with the observed value \cite{Komatsu:2008hk}
\be
\Omega_{\rm WMAP}\,h^2= 0.1131\pm0.0034
\label{wmap}
\ee
It is useful to define the new parameter
\be
\omega
\equiv\frac{\ynlsp}{Y_{\rm CDM}}\;,
\label{omega}
\ee
which quantifies the amount of present cold dark matter coming from the NLSP decay. In this equation $\ynlsp$ refers to the NLSP yield just before its decay \footnote{Here the yield of a species $i$ is defined as the ratio of the number density $n_i$ to entropy, $Y_i\equiv n_i/s$. Recall also that the yield and the relic density of a massive particle species $i$ are related through $Y_i \simeq 4.1\times 10^{-12} \(100\gev/ m_i\)\(\Omega_i\, h^2/ 0.11\)
$.}. Combining \eq{abundance} and \eq{omega} we get that
\be
\omega=1-\frac{\Omega^{\rm TP}_{\rm LSP}\,h^2}{\Omega_{\rm WMAP}\, h^2}\, ,
\ee
which implies that $\omega\le 1$ independently of $\delta$.
\subsection{BBN constraints}
At temperatures of order $T\sim 1$ MeV, the light nuclei are synthesized in the primordial plasma. These temperatures corresponds to times between 1 sec and $10^3$ sec. The obtained abundances in standard BBN calculations are in striking agreement with observation \footnote{There are possible discrepancies in Lithium abundances, which might be
explained using Gravitino dark matter ~\cite{Jedamzik:2005dh,Bailly:2008yy}.
}.
However, the injection of energetic particles in the primordial plasma at BBN or later can disrupt the standard
BBN processes~\cite{Cyburt:2002uv,Jedamzik:2004er,kkm04,Jedamzik:2006xz}, leading to a disagreement between theory and observation.
Thus, any additional particle decaying at BBN or later is subject to the strong constraints from light nuclei abundances.
Usually, the energy of injected particles is assumed to be of the same order of the LSP/NLSP mass and stringent bounds were derived for
various scenarios from BBN~\cite{Ellis:1984eq,fst04-sugra,Ellis:2003dn,Roszkowski:2004jd,Kohri:2005wn,Cerdeno:2005eu,
Steffen:2006hw,Pradler:2006qh,Kawasaki:2008qe,Bailly:2009pe}.
These constraints are obtained by solving the full set of Boltzmann equations for BBN with a late-decaying particle. This requires the detailed study of the spectrum of decay products with all the relevant
nuclear cross sections~\cite{Jedamzik:2004er,kkm04,Jedamzik:2006xz}.
The constraints apply to the released electromagnetic or hadronic energy, parametrised by $\xi_{i}$ defined as \cite{fst04-sugra}
\dis{
\xi_{i}\equiv E_{i}\,B_{i}\,\ynlsp \label{xiem}\;,
}
where $E_i$ is the released energy per decay with $i=$ em for electromagnetic decays and $i=$ had for the hadronic ones and $B_i$ stands for the respective branching ratios. The constraints from hadronic processes are important when the lifetime is relatively
short $\tau_{\rm NLSP}\lesssim 10^7$ sec \cite{kkm04,Kohri:2005wn}. The typical lifetimes considered in this work are
larger that $10^7$ sec and therefore, in the following, we will consider only the BBN constraints on $\xi_{\rm em}$.
In the 2-body electromagnetic decay, the corresponding released energy is
\dis{
E_{\rm em}= \frac{\mnlsp^2-\mlsp^2}{2\mnlsp}\, ,
}
where we assumed that the visible particle mass is negligible. In the degenerate mass limit, $\delta \ll 1$, this reduces to $E_{\rm em}\simeq \mlsp \,\delta$.
Notice that, as emphasized before, $E_{\rm em}$ is much smaller than the usually\footnote{Small mass differences has been considered in \cite{Bailly:2009pe,Cembranos:2007fj} and for completely different motivations than ours.} assumed value $\mnlsp/2$.
Given the above discussion, for our purposes, we can apply the constraints on $\xi_{\rm em}$ from Ref.~\cite{Jedamzik:2006xz} directly to our scenario. Then, the constraint on $\xi_{\rm em}$ reads
\dis{
\xi_{\rm em}
\simeq
4.1 \times 10^{-10} \gev \bfrac{\abundwmap}{0.11}\, \omega\, B_{\rm em}\,\delta
< \xi_{\textrm {upper limit}}\,, \label{xiBBN}
}
where the right hand side in the above equation can be read off from the upper limit derived in \cite{Jedamzik:2006xz} for $B_{\rm had}=0$.
For general values, the upper limit on the product $\omega\, B_{\rm em}\,\delta$ from BBN is shown in Figure~\ref{fig:BBNCMB} with red (dashed) line.
From this figure, we can see that for lifetimes between $10^7$ and $10^{10}$ sec, this constraint requires $\omega B_{\rm em}\delta$ to be between $10^{-3}$ and $10^{-4}$. Notice that BBN constraints apply equally to the case of charged NLSP decaying to gravitino and electromagnetic showers.
\begin{figure}[!t]
\begin{center}
\vspace{-1cm}
\includegraphics[width=0.6\textwidth]{plot.eps}
\vspace{-0.7cm}
\caption{ \small Constraints on the
combined parameters $\omega B_{\rm em}\delta$ versus lifetime
of NLSP using the result of~\cite{Jedamzik:2006xz} with $B_h=0$.
The red (dashed) line come from em BBN constraints and blue (solid) line
from CMB distortion as explained in the text.
Regions above the lines are excluded.}
\label{fig:BBNCMB}
\end{center}
\end{figure}
\subsection{CMB spectral distortion}
In addition to BBN constrains, for long lifetimes $\tau_{\rm NLSP} \gsim 10^7$ sec., there are strong bounds from the shape of the CMB black-body spectrum. As pointed out in~\cite{Ellis:1984eq}, the late injection of
electromagnetic energy may distort the frequency dependence of the CMB
spectrum from its observed blackbody shape. At late times of
interest in our scenario, energetic photons from
NLSP decays lose energy through processes such as $\gamma
e^-\rightarrow \gamma e^-$, but photon number remains conserved since
other processes, like double Compton scatterings and thermal
bremsstrahlung, become inefficient. As a result, the spectrum follows
the Bose--Einstein distribution function
\begin{equation}
f_\gamma(E)=\frac{1}{e^{E/(kT)+\mu}-1},
\label{eq:cmbbedist}
\end{equation}
where $\mu$ here denotes the chemical potential.
Then, the chemical potential of the distorted CMB spectrum has to satisfy the constraint, $|\mu|< 9\times 10^{-5}$~\cite{mubound}, which, for decay lifetimes $\taunlsp \lsim 8.8\times10^{9} \sec$, translates into an upper bound on
the released energy from NLSP decay, $\xi_{\rm em}$ defined in \eq{xiem}
\cite{Hu:1993gc,Roszkowski:2004jd},
\dis{
\xi_{\rm em} <& 1.59\times 10^{-8} \,
e^{(\tau_{dC}/\taunlsp)^{5/4}}\left({1 \sec \over \taunlsp} \right)^{1/2}\gev
\label{xiCMB},
}
where $\tau_{dC}\simeq 6.085\times 10^6$ sec.
For longer lifetimes ($\taunlsp \gsim 8.8\times 10^{9} \sec$),
the spectral distortions in
the CMB spectrum can be described in terms of the integral of the fractional
contributions to the energy $\epsilon$ of the CMB per comoving volume during
decay through the Compton $y$ parameter, $4y=\delta\epsilon/\epsilon$, given by
\begin{eqnarray}
\frac{\delta\epsilon}{\epsilon}=7.04\times\frac{1}{T(t_{\rm eff})}\xi_{\rm em},
\label{eq:cmbypar}
\end{eqnarray}
where $T(t)$ is the CMB temperature and $t_{\rm
eff}=[\Gamma(1-\beta)]^{1/\beta}\taunlsp$, for a time--temperature relation
$T\propto t^{-\beta}$, with $\Gamma$ the usual Gamma function.
In the radiation dominated era in the early Universe, for $T
< 0.1 \mev$,
\begin{eqnarray}
T=1.15\times 10^{-3}\left(\frac{t}{1\sec}\right)^{-1/2} \gev,
\end{eqnarray}
which gives
$\beta=1/2$. Thus $t_{eff}= [\Gamma(1/2)]^2 \taunlsp=\pi\taunlsp$.
The observational limit $|y|<1.2\times 10^{-5}$~\cite{Hagiwara:2002fs}
gives the constraint
\dis{
\xi_{\rm em}\lesssim 4.42\times 10^{-9}
\gev\sqrt{\frac{1\sec}{\taunlsp}}.
\label{eq:yboundonxiem}
}
Using Eqs.~(\ref{eq:cmbypar},\ref{eq:yboundonxiem}), this CMB constraint
is plotted in Figure~\ref{fig:BBNCMB} with blue (solid) line.
We can see that for lifetimes $\taunlsp \gtrsim 10^{10}$ sec. the CMB
constraint becomes more stringent than the BBN ones and sets the constraint on $\xi_{\rm em}$ until recombination time.
Similarly to the BBN, CMB constraints apply also to charged NLSP decaying to gravitino and electromagnetic showers.
\subsection{Diffuse Gamma-ray observation}
After recombination, at the cosmic time around $10^{13}$ sec,
the number density of free electrons drops quickly and the photons are almost
free from the interactions.
Therefore the photons from the decaying particles can reach us now and
contribute to the Cosmic Gamma-ray Background (CGB).
The observed CGB highly constrains any extra contribution
including the photons from late decaying particles
\footnote{Under appropriate conditions, the late decays of WIMPs to gravitinos
and MeV photons
may explain the MeV CGB anomalies~\cite{Cembranos:2007fj}.}.
The present photon flux from two-body decay can be written as
\dis{
\frac{d\Phi}{dE_\gamma}=\frac{c}{4\pi}\int^{t_0}_{t_i}\frac{dt}{\taunlsp}\frac{\rho_c\,\Omega_{\rm WMAP}\,\omega\,
B_{\rm em}}{\mnlsp}\,e^{-t/\taunlsp}\delta(E_\gamma
-a E_{\rm em}),
}
where $E_{\rm em}\simeq \mlsp \delta$ is the energy of the photon at
production, $\taunlsp$ and $\mnlsp$ are the lifetime and mass of NLSP,
$\rho_c=3H_0^2/8\pi G_N=8.0992h^2\times 10^{-47}\gev^4$ and $a=a(t)$ is
the time-dependent scale factor with $a(t_0)=1$ at present time $t_0$.
The delta function can be integrated using the formula
$\delta(f(t))=\delta(t-t_a)/|f'(t=t_a)|$ where $t_a$ is the solution which
satisfies $E_\gamma=a(t_a) E_{\rm em}$. After integration, the flux reads
\dis{
\frac{d\Phi}{dE_\gamma}=\frac{c}{4\pi}\frac{\rho_c\,\Omega_{\rm WMAP}\,\omega \,B_{\rm em}}{\mnlsp\,\taunlsp}\,\frac{e^{-t_a/\taunlsp}}{E_\gamma
H(E_\gamma/E_{\rm em})}\Theta(E_{\rm em}-E_\gamma),
}
where the $\Theta$ function simply cuts energies larger than the initial energy
and each observed photon with energy $E_\gamma$ is produced at a time $t_a$ which satisfies
$a(t_a)=E_\gamma/E_{\rm em}$. Assuming the dark energy is a cosmological constant,
this function $t_a(a)$ is given by~\cite{Cembranos:2007fj}
\dis{
t_a\equiv t (a=E_\gamma/E_{\rm em})=\frac{2\log [(\sqrt{\Omega_\Lambda a^3} +\sqrt{ \Omega_M +\Omega_\Lambda a^3})/\sqrt{\Omega_M} ]}{3H_0 \sqrt{\Omega_\Lambda}},
}
where $H(a)=H_0\sqrt{\Omega_M a^{-3} + \Omega_\Lambda}$.
Taking into account that $c/H_0=1.3\times 10^{28} \cm$,
$\rho_c=5.6\times 10^{-6} \gev/\cm^3$, $H_0=70 \km\,\sec^{-1}\, \mpc^{-1}$,
$\Omega_\Lambda=0.7$, $\Omega_{M}=\Omega_{\rm CDM}+\Omega_B=0.3$ and
$\Omega_{\rm CDM}=\Omega_{\rm WMAP}=0.25$, we find that the flux is
\dis{
\frac{d\Phi}{dE_\gamma}=1.37\times 10^{21}~ \gev \cm^{-2} \frac{\omega B_{\rm em}}{\mnlsp \taunlsp}
\frac{e^{-t_a/\taunlsp}}{E_\gamma \sqrt{0.7+0.3(E_{\rm em}/E_\gamma)^3 } },
\label{CGBspec}
}
with
\dis{
t_a=3.51\times 10^{17} \sec \log[\sqrt{2.33a^3}+\sqrt{1+2.33 a^3}].
}
This differential flux must be compared to the observation of
diffuse gamma ray flux. At each photon energy, we must require that this
flux is smaller than the flux, $E_\gamma^2 \frac{d \Phi}{d E_{\gamma}}$,
observed by SPI, COMPTEL and EGRET \cite{Yuksel:2007dr}.
Notice that, unlike the BBN and CMB constraints where we can find a bound on $\xi_{\rm em}$ for a given value of $\taunlsp$, now we need to specify both $\taunlsp$ and $E_\gamma$ to obtain a bound on $\omega B_{\rm em} \delta$.
On the other hand, if the lifetime of NLSP is longer than the age of Universe,
the line spectrum from the galactic center can be observed without
cosmological redshift and this provides a further constraint on the emitted radiation from NLSP decays. For this we apply the bounds from
Ref.~\cite{Yuksel:2007dr}.
Taking into account of the exponential decay of NLSP we have
\dis{
\frac{\rho_{sc}\omega B_{\rm em}}{4\pi\mnlsp \taunlsp}e^{-t_0/\taunlsp} \zeta_{lim} < \mathcal{F}(E_\gamma=E_{\rm em}),
}
where $\rho_{sc}=0.3 \gev \cm^{-3}$ is the dark matter density at the solar
distance from the Galactic center, $R_{sc}=8.5\kpc$, $\zeta_{lim}$ is a
dimensionless integral of the line-of-sight intensity in the galactic center
which ranges between $0.5 -1.5$ for various dark matter halo
profiles~\cite{Yuksel:2007dr}. The function $\mathcal{F}$ is given in Figure 2
of Ref.~\cite{Yuksel:2007dr}. Once again, as in
the case of the CGB constraint in \eq{CGBspec}, this constraint depends on the photon energy and therefore we need to specify both $\taunlsp$ and $E_\gamma$ to obtain bound on $\omega B_{\rm em} \delta$. Thus, there is no simple analog of Figure~\ref{fig:BBNCMB} for a
constraint on $\xi_{\rm em}$ from Diffuse gamma rays observations. Moreover, the constraints on $\xi_{\rm em}$ from the galactic center gamma rays
are of the same order of magnitude (although slightly stronger) as the ones from the diffuse extragalactic emission. Therefore, for
simplicity, we will only consider the diffuse gamma rays constraints, which apply to a broader range of energy.
\subsection{Catalyzed BBN}
Heavy long-lived negatively charged particles, $X^-$, present during BBN can
bind with light nuclei modifying standard BBN reactions. These
catalyzed reactions, called CBBN, result in a change of light element
abundances, and in particular lead to the overproduction of
$^{6}\textrm{Li}$ through CBBN reactions ~\cite{CBBN}.
For lifetimes longer than $5\times 10^3 \sec$, the observed light-nuclei abundances result on a constraint on the abundance of the charged relic, $Y_{X^-} < 2\times10^{-16}$~\cite{Hamaguchi:2007mp,Pradler:2007is}.
However, taking more conservative $^{6}\textrm{Li}/ ^{7}\textrm{Li}$ constraints, it is possible to relax slightly the previous bound to $Y_{X^-} < 10^{-14}$--$10^{-15}$ ~\cite{Jedamzik:2007qk,Jedamzik:2009uy}.
Using \eq{omega}, we can translate the constraint on the yield of charged NLSP
from catalyzed BBN to a bound on $\omega$
\dis{
\omega
\lesssim
2.44\times10^{-3}\bfrac{\mlsp}{100\gev}\bfrac{Y_{\textrm{CBBN}}}{10^{-14}}
\label{CBBNomega}
}
where we used $\abundwmap=0.11$ and $Y_{\textrm{CBBN}}$ is the maximum value
allowed from catalyzed BBN for lifetimes larger than $10^5\sec$.
As can be see from this equation, the catalyzed BBN bound is very stringent
and indeed, it is very difficult to obtain such small yield at freeze-out in the MSSM \cite{Berger:2008ti}. However, it is still possible to find small
allowed regions in the MSSM, or even in the CMSSM, where the $\stau$ is the NLSP with large $\tan \beta$ \cite{Ratz:2008qh,Pradler:2008qc} with relaxed $^{6}\textrm{Li}/ ^{7}\textrm{Li}$ bounds~\cite{Bailly:2009pe}.
\section{Bounds on the degenerate gravitino scenario}
\label{sec3}
In this section, we will apply the model-independent bounds derived in the last section to the case of degenerate gravitino scenario. In our scenario, the gravitino can be either the LSP or the NLSP. In both cases it is thermally produced at reheating. After inflation, the Universe is reheated at a temperature
$T_{\rm RH}$ and gravitinos are produced through thermal scatterings in the plasma \footnote{We assume that at $\treh$, the Universe is composed of a thermal bath of MSSM degrees of freedom and that gravitinos production by inflaton decay is negligible \cite{Nilles:2001ry}.}. Their resulting relic density is linear in the reheating temperature and it is given by \cite{Bolz:2000fu}\footnote{Taking into account the result of Ref.~\cite{Rychkov:2007uq} the gravitino relic density would roughly increase a factor 2.}
\begin{equation}
\Omega^{\rm TP}_{3/2}\,h^2\simeq 0.27 \(\frac{T_{\rm RH}}{10^{10}\gev}\)\(\frac{100\gev}{m_{3/2}}\)\(\frac{M_3}{\tev}\)^2\,,
\label{omega1}
\end{equation}
where $M_3$ is the gluinos mass. Assuming no late entropy release, \eq{omega} sets an absolute bound on the reheating temperature, i.e. $\Omega^{\rm TP}_{3/2}\lesssim \Omega_{\rm WMAP}$, implies
\begin{equation}
T_{\rm RH}\lesssim 4.1\times 10^9 \gev\, \(\frac{m_{3/2}}{100\gev}\)
\(\frac{\tev}{M_3}\)^2\,.
\label{max}
\end{equation}
This has to be compared with the minimum reheating temperature for successful thermal leptogenesis \cite{Davidson:2002qv,Giudice:2003jh, Antusch:2006gy} $T_{\rm RH} \gtrsim 2\times 10^9 \gev$. From this equation it is clear that the reheating temperature can not reach values much above $\sim 10^{10}$ GeV. \footnote{A possibility to increase $T_{\rm RH}$ is to consider a squeezed gaugino spectrum reducing the gluino mass $M_3$ (see e.g. \cite{Olechowski:2009bd}).} In addition to this constraint, we have to implement the constraints on the released energy considered in the last section.
The relevant particles in the analysis of NLSP decays in the degenerate scenario are the gravitino and the lightest MSSM particle, which can be either the neutralino or the stau\footnote{Sneutrinos LSP are marginally allowed \cite{Arina:2007tm} due to their too large direct detection cross-sections. We do not consider them in our analysis.}.
We have two different situations depending on the particle nearly degenerate with the gravitino, namely gravitino-neutralino and gravitino-stau degeneracy. In each case the gravitino can be either the LSP or the NLSP.
\subsection{Gravitino-neutralino degeneracy}
As usual, in the MSSM, the lightest neutralino eigenstate $\chi^0_1$ is
parametrised as
$\chi^0_i = N_{i 1} (-i\widetilde{B})+ N_{i 2} (-i \widetilde{W}_3) + N_{i 3}\widetilde{H}^0_U + N_{i 4}\widetilde{H}^0_D$, where the unitary matrix $N$ defines the composition of neutralinos in terms of the Bino, Wino and Higgsinos. Since we are considering mass splittings that are smaller than the $Z$ mass to suppress the hadronic branching ratio, the dominant (2-body) decay channel will be $\chi^0_1\to \gamma\,\widetilde{G}$ or $\widetilde{G}\to\chi^0_1 \gamma$. As we will see below, the typical lifetime of NLSP is of order $10^{13}\sec \times \left( 1 \gev/ \Delta M\right)^3$. Therefore, depending on $\Delta M$ we will have to consider different constraints.
\subsubsection{Gravitino LSP}
In this case, the lifetime of neutralinos is given by
\begin{equation}
\tau_{\chi}\simeq \frac{1.78\,\times 10^{13}\sec}{|N_{11}\cos\theta_W+N_{12}\sin\theta_W |^2} \left(\trac{ 1 \gev}{ \Delta M}\right)^3\,.
\end{equation}
Notice that, in the limit of $\Delta M \ll m_{3/2}$, the lifetime depends only on the mass splitting $\delta\timesm_{3/2}=\Delta M$, but not on the overall mass scale $m_{3/2}$ \cite{Moroi:1995fs,fst04-sugra}.
If the gravitino is the LSP, the neutralino will decay into gravitino and photon and there are strong constraints on the released energy $\xi_{\rm em} = \omega\, B_{\rm em} \,\delta$. Different constraints will apply depending on the lifetime which, in turn is fixed by $\Delta M$.
\begin{itemize}
\item For $10 \gev \lesssim
\Delta M \lesssim 90 \gev$, the relevant constraint is BBN and from Fig.~\ref{fig:BBNCMB} we have roughly $\omega \delta \lesssim 10^{-3}$.
\item For smaller splittings, from $1 \gev \lesssim \Delta M \lesssim 10 \gev$, the constraints from CMB spectrum are much stronger than the BBN ones and $\omega \delta$ is between $10^{-4}$ and $10^{-6}$
\item For $ 30 \mev \lesssim \Delta M \lesssim 1 \gev$ we have to take into account the diffuse gamma ray observations. Although the constrains on $\omega \delta$ depend on $m_{3/2}$, for $m_{3/2}\simeq 100$ GeV, typical values range from $\omega \delta = 10^{-6}$ to $\omega \delta = 10^{-9}$.
\item Mass differences from $ 2 \mev \lesssim \Delta M \lesssim 30 \mev$ correspond to NLSP with lifetimes $ 2 \times 10^{21} \sec \gtrsim \tau_{\rm NLSP} \gtrsim 5 \times 10^{17} \sec$. These NLSP are already decaying at present, therefore diffuse gamma ray observations constrain their abundance. In this case $\omega \delta$ ranges from $10^{-9}$ to $10^{-5}$.
\item Finally, for smaller $\Delta M$, the neutralino is still present in the
universe and it is completely stable for practical purposes. The only constraint comes from the WMAP measurement of the dark matter abundance.
\end{itemize}
These constraints are summarized in Figure~\ref{fig:combinedneutr}. We can see that for a given $\Delta M$ the constraints on $\omega$ are very strong for $\Delta M > 30 \mev$, and, as we will see, it is difficult to reach such small neutralino abundances at decoupling in the MSSM. For smaller $\Delta M<2 \mev$, corresponding to neutralino lifetimes longer than the age of the universe, $\omega \simeq 1$ is evidently allowed. For $2 \mev<\Delta M <30\mev$, $\omega$ ranges from 1 to $10^{-6}$.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.6\textwidth]{NLSP_neut_V.eps}
\end{center}
\vspace{-1cm}
\caption{\small Constraints on the
combined parameters $\omega B_{\rm em}$ versus $\Delta M$
for given mass of neutralino NLSP (for $N_{11} \simeq 1$) with Gravitino LSP.
Solid line is for $\mnlsp=100\gev$ and dashed line for $\mnlsp=200\gev$.
The blue lines for $\Delta M > 2 \gev$ come from BBN and CMB constraints
and the green lines for $\Delta M< 2 \gev$ come from the diffuse gamma ray observations.
Regions above the lines are excluded. }
\label{fig:combinedneutr}
\end{figure}
The only constraint on the gravitino abundance comes from \eq{abundance} and the dark matter abundance measured by WMAP. Given $\omega$, we can use Eq.~(\ref{delta}) and Eq.~(\ref{omega}) to calculate the required reheating
temperature
\begin{eqnarray}
\treh&=& 4.1\times10^{9}\gev\left(\frac{m_{3/2}}{100\gev} \right)\left(\frac{1\tev}{M_3}\right)^2 \left(1-\omega\right).
\label{treh_GLSP}
\end{eqnarray}
From Eq.~(\ref{treh_GLSP}), we see that, provided $\omega\ll 1$, one gets the maximal allowed reheating temperature $\treh = 4.1 \times 10^9$ GeV.
\subsubsection{Gravitino NLSP}
If the gravitino is the NLSP, the dominant decay channel is $\widetilde{G}\to \gamma\,\chi^0_1$ and its lifetime is given by
\begin{equation}
\tau_{3/2}\simeq \frac{3.56\,\times 10^{13}\sec}{|N_{11}\cos\theta_W+N_{12}\sin\theta_W |^2} \left(\trac{ 1 \gev}{ \Delta M}\right)^3\,,
\end{equation}
which is only a factor 2 larger than the gravitino LSP case. However, unlike the gravitino LSP case, the bounds on the released energy from this decay constrain strongly the initial thermal abundance of gravitinos. Notice that, here, $\omega$ represents the fraction of neutralinos coming from gravitino decay and therefore the initial abundance of gravitinos.
The constraints on $\omega \delta$ for different ranges of $\Delta M$ seen in the previous section apply equally in this case. Now, given $\Delta M$ and $m_{3/2}$, which fix $\delta$, we have a direct constraint on the gravitino thermal abundance, and hence on $\omega$. As we know, the gravitino thermal abundance is directly proportional to $\treh$ which in this case can be written
\dis{
\treh\simeq 4.1\times10^{9}\gev\left(\frac{m_{3/2}}{100\gev} \right)\left(\frac{1\tev}{M_3}\right)^2 \omega \left(\frac{1}{1+\delta}\right) .
\label{treh_GNLSP}
}
From this equation, we see that, opposite to the case of gravitino LSP, in order to maximize the reheating temperature, one needs $\omega$ as large as possible. The constraints on $\treh$ as a function of $\delta$ are shown in Figure~\ref{fig:treh}. Given the strong constraints on $\omega$ for the different $\delta$s,
$\treh$ is considerably smaller than the maximal value allowed by WMAP. The only exception to this are the cases with very small $\Delta M \lesssim 2 \times 10^{-3}$ GeV, corresponding to the NLSP still present as a dark matter component, where $\omega\lesssim 1$ and the LSP abundance is smaller that the observed $\Omega_{\rm WMAP}\,h^2$.
\begin{figure}[!t]
\begin{center}
\begin{tabular}{c c}
\includegraphics[width=0.6\textwidth]{GNLSP_Treh_V.eps}
&
\end{tabular}
\end{center}
\vspace{-1cm}
\caption{\small Constraints on the maximal reheating temperature after inflation
for the Gravitino NLSP with neutralino LSP (with $N_{11} \simeq 1$)
for $\mnlsp= 100\gev$ (solid)
and $200\gev$ (dashed). Regions above the lines are excluded. Notice that, as we can see from \eq{treh_GNLSP}, these maximal reheating temperatures are only possible if the corresponding bounds on $\omega$ are satisfied}
\label{fig:treh}
\end{figure}
\subsection{ Gravitino-stau degeneracy}
In the MSSM, the lightest stau state $\stau_1$ can also be the LSP. However this situation is usually discarded as staus cannot play the role of cold dark matter. The remaining possibility is then that $\stau_1$ is the NLSP with the gravitino as the LSP. The dominant decay channel in this case is $\stau_1 \to \tau\,\widetilde{G}$ if the mass-difference is larger than the tau mass, i.e. $\Delta M \geq 1.77 \gev$~~\footnote{Notice that for $\Delta M \leq 1.77 \gev$ the two body flavour-conserving channel is closed and the stau can decay only through lepton flavour violating channels, $\stau_1 \to \mu\,\widetilde{G}$ or $\stau_1 \to e\,\widetilde{G}$, where the lifetime would be inversely proportional to the lepton-flavour violating coupling \cite{Masiero:2004js,Kaneko:2008re}: $\tau_{\stau_1}\simeq 2\times 10^{14}\sec |\delta^{\rm LFV}_{\tau i}|^{-2}$ for $m_{3/2} =100 \gev$ and $\Delta M = 2 \gev$. This means that, if these flavour violating couplings, the so-called Mass Insertions, are sizable, $\delta^{\rm LFV}_{\tau i}\geq 0.03$, the stau could decay before the present time. However, we do not consider this possibility in this paper. }. In this case, the lifetime of the stau NLSP is then given by
\begin{equation}
\tau_{\stau_1}\simeq 6.69\times 10^{15}\sec \left(\trac{ m_{3/2}}{300 \gev}\right) \left(\trac{2 \gev }{ \Delta M}\right)^4 \left(\trac{0.21}{ 1 - m_\tau^2/(\Delta M)^2}\right)^{3/2} \,.
\end{equation}
Notice that now, the stau lifetime, in contrast with the degenerate
gravitino-neutralino case, does depend on the overall mass scale $m_{3/2}$.
Moreover for a mass difference similar to the tau mass, $\Delta M\simeq 1.78 \gev$, the stau lifetime is equal to the age of the universe.
In principle, the stau has to satisfy similar constraints as in
the degenerate gravitino-neutralino case with the exception of diffuse
gamma ray constraints. Notice that the stau does not decay directly to
photons and therefore does not contribute to the gamma ray
background. Despite this fact, all the decays of the stau produce
electromagnetic cascades that affect both BBN and CMB observables.
Nevertheless, the main difference with the neutralino case is that
stau, being charged, can form bound states with light elements and
affect BBN predictions \cite{Pradler:2007is, Jedamzik:2007qk}. Then,
from \eq{CBBNomega} we obtain a strong constraint on $\omega$ which
for $Y_{\rm CBBN}= 10^{-15}$ is $\omega < 7.32 \times 10^{-4}
m_\stau/(300 {\rm GeV})$.
Finally, if the stau lifetime is longer than the age of the universe,
the stau yield is very strongly bounded by
the presence of anomalously heavy Hydrogen in deep sea water \cite{Yamagata:1993jq}. In terms of $\omega$ the bound reads, $\omega \leq 2.2 \times 10^{-27}\left(\mstau/100 \gev\right)$, for stau masses between 5 GeV and 1.7 TeV.
Thus, in practice, this possibility can be completely discarded. In our analysis,
we require that staus have already decayed at present and therefore we
eliminate the staus with a lifetime longer than the age of the universe which
corresponds to $\Delta M \lesssim 2 \gev$.
\begin{figure}[!t]
\begin{center}
\includegraphics[width=0.6\textwidth]{NLSP_stau.eps}
\end{center}
\vspace{-1cm}
\caption{\small Constraints on $\omega$ versus $\Delta M$ with $m_{3/2} = 300 \gev$
for stau NLSP with Gravitino LSP.
The solid blue line represents the BBN and CMB constraints.
The vertical line at $\Delta M \simeq 2 \gev$ corresponds to a stau lifetime equal to the age of the universe and therefore the region on the right is excluded from heavy-water searches. The dashed pink lines correspond to the CBBN constraint for
$Y_{\rm CBBN}= 10^{-14}$ and $Y_{\rm CBBN}= 10^{-15}$.}
\label{fig:combinedstau}
\end{figure}
These constraints are presented in Figure~\ref{fig:combinedstau} where we see
that for $Y_{\rm CBBN}= 10^{-15}$ the CBBN is the most stringent of all the
constraints and requires $\omega \lesssim 7\times 10^{-4}$.
This corresponds to stau relic density well below the
observed relic density, \eq{wmap}, and the thermal relic density at freeze-out
for purely RH staus \cite{Berger:2008ti}. In order to get such small abundance,
the staus must have a substantial LR mixing. This requires
both large $\mu$ and large $\tan\beta$ and moderate $m_{\stau_1}$
\cite{Ratz:2008qh, Pradler:2008qc}, which as we will see in section \ref{sec4}, is difficult, but still possible in the CMSSM.
Furthermore, the vertical line at $\Delta M \simeq 2 \gev$ corresponds to our requirement that all staus have already decayed at the present age.
The bound on $\treh$ here is analogous to the case of
gravitino-neutralino degeneracy with gravitino LSP. Again, the
reheating temperature is given in \eq{treh_GLSP} and given that, in this case,
the allowed points in parameter space require $\omega \lesssim
10^{-2}$, one gets the maximal reheating temperature $\treh = 4.1
\times 10^9$ GeV.
\section{The degenerate gravitino scenario in the CMSSM }
\label{sec4}
Perhaps the most appealing mechanism to transmit SUSY breaking
from a hidden sector, where SUSY breaking occurs, to the visible
sector is to use gravitational interactions, that are suppressed by the Planck scale \footnote{Of course, in gravity mediated scenarios, there can be an associated cosmological moduli problem.
In this work, as this issue is outside the focus of the paper, we do not address it and we will just assume that it is
solved by some unspecified mechanism. } \cite{Nilles:1983ge}. This scenario is commonly called
gravity mediation, and the CMSSM is one of its simplest and most popular
realizations for phenomenological studies \cite{kkrw94}.
It is defined in terms of only five free parameters:
common scalar ($\mzero$), gaugino ($\mhalf$) and
tri--linear ($\azero$) mass parameters
(all specified at the GUT scale) plus the ratio of Higgs vacuum
expectation values $\tanb$ and $\text{sign}(\mu)$, where $\mu$ is
the Higgs/higgsino mass parameter whose square is computed from
the conditions of radiative electroweak symmetry breaking. In addition to these parameters, in this analysis we have the gravitino mass, $m_{3/2}$, that we keep as a free parameter.
The gravitino LSP (or NLSP) scenario in the CMSSM has been investigated
thoroughly in the literature for the past years, both from the point of view
of cosmological implications for dark matter \cite{Roszkowski:2004jd} and
for implications in collider searches \cite{Choi:2007rh}.
One of the remarkable results of this scenario from cosmology is that the
limits imposed on the reheating temperature of the Universe after
an inflationary epoch have got down to a few $10^7$ GeVs \cite{Bailly:2009pe}.
This constraint basically rules out the thermal leptogenesis mechanism as the
mean to produce the observed baryon asymmetry of the universe.
\begin{table}[t]
\centering
\begin{tabular}{|l | l l l | l|}
\hline
Observable & Mean value & \multicolumn{2}{c|}{Uncertainties} & ref. \\
& $\mu$ & ${\sigma}$ (exper.) & $\tau$ (theor.) & \\\hline
$\delta a_\mu \times 10^{10}\; (e^+ e^-)$ & 29.5 & 8.8 & 2.0 &
\cite{Miller:2007kk} \\
$\BR(\overline{B}\rightarrow X_s\gamma) \times 10^{4}$ & 3.52 & 0.33 & 0.3 & \cite{hfag} \\
$\abundchi$ & 0.1099 & 0.0062 & $0.1\,\abundchi$& \cite{wmap5yr}
\\\hline\hline
& Limit (95\%~\text{CL}) & \multicolumn{2}{r|}{$\tau$ (theor.)} & ref.
\\ \hline
$\mhl$ & $>114.4\gev$ & \multicolumn{2}{r|}{$3 \gev$}
& \cite{Barate:2003sz} \\
Sparticle masses & \multicolumn{3}{c|}{As implemented in
Micromegas} & \cite{micromegas} \\ \hline
\end{tabular}
\caption{ \small Summary of the observables used in the analysis to
constrain the CMSSM parameter space. Upper part:
Observables for which a positive measurement has been made.
$\delta a_\mu= a_\mu^{\rm exp}-a_{\mu}^{\text{SM}}$ denotes the discrepancy between
the experimental value and the SM prediction of the anomalous magnetic
moment of the muon $(g-2)_{\mu}$.
Lower part: Observables for which only limits currently
exist.
\label{tab:obs}}
\end{table}
In this section, we apply the results of the previous sections for the
degenerate gravitino scenario
to investigate the highest reheating temperature that can be reached in the
CMSSM fulfilling the WMAP constraint on the CDM abundance and the relevant
collider bounds; namely, direct SUSY searches, $(g-2)_\mu$ using
($e^+e^- \rightarrow $hadrons) data\footnote{Notice that we obtain the hadronic contribution to $(g-2)_\mu$ using only the data from ($e^+e^- \rightarrow $hadrons), inclusion of the $\tau$ data can decrease the discrepancy between the SM prediction and the experimental result.} and the $BR(B\rightarrow X_s \gamma)$. All
the constraints imposed are summarized in Table~\ref{tab:obs}.
We use the fortran package SUSPECT \cite{Djouadi:2002ze} to solve the RGEs
and to calculate the spectrum of physical sparticles and Higgs bosons,
following the procedure outlined in \cite{Djouadi:2001yk}.
To evaluate the CDM abundance in each point of the CMSSM parameter
space, we employ the \texttt{MicrOMEGAs} code \cite{micromegas}.
The branching ratio for the $B \rightarrow X_s \gamma$ decay
has been computed with the numerical code \texttt{SusyBSG}
\cite{Degrassi:2007kj} using the full NLO QCD contributions, including
the two-loop calculation of the gluino contributions presented
in \cite{Degrassi:2006eh} and the results of \cite{D'Ambrosio:2002ex}
for the remaining non-QCD $\tanb$-enhanced contributions.
Finally we compute $\delta_{\rm had}^{\rm SM}a_\mu$ at full
one-loop level adding the logarithmic piece of the quantum
electrodynamics two-loop calculation plus two-loop contributions
from both stop-Higgs and chargino-stop/sbottom \cite{Heinemeyer:2004yq}.
Then, the effective two-loop effect due
to a shift in the muon Yukawa coupling proportional to $\tan^2 \beta$
has been added as well \cite{Marchetti:2008hw}. Recall that the
communication among the different codes is done via the SLHA
accord \cite{Skands:2003cj}.
In particular, regarding the direct constraints on new particle searches,
given that the theoretical error in computing the lightest
Higgs mass $\mhl$ by SUSPECT is about $3 \, \gev$ \cite{Allanach:2004rh}, we require the calculated value of $\mhl$ to exceed $111 \, \gev$. In the case of
observables for which a positive measurement has been made, we require our predictions to be within the $2 \sigma$ range, for which we have added the theoretical and
experimental errors, found in Table~\ref{tab:obs}, in quadrature.
In our numerical analysis we take $m_t = 173.1 \, \gev$ \cite{:2009ec}.
For the case of the stau NLSP, in addition, we completely exclude points that do not satisfy the conservative
bound on $Y_\stau < 10^{-14}$ \cite{Jedamzik:2007qk} from the
catalyzed nuclear reactions \cite{CBBN} (see Fig.~\ref{fig:combinedstau}).
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.5\linewidth]{a0_tb10_mu+_w.eps}
\caption[text]{ \small Values of w in the $m_{1/2}$–-$m_0$ plane of the
CMSSM parameter space for $m_t=173.1 \, \gev$, $A_0=0$, $\tanb=10$ and
$\mu>0$. The red (medium grey) region is forbidden by the Higgs bound from LEP and in the very light grey region no correct electroweak symmetry breaking is obtained. The light brown (light grey) band below the NEWB area corresponds to the region forbidden by the LEP chargino bound. The regions below the dashed green lines satisfy the $(g-2)_\mu$ constraint at the 2
$\sigma$ ($a_\mu^{\rm SUSY} > 2.43 \times 10^{-10}$) or 3 $\sigma$ level ($a_\mu^{\rm SUSY} > 11.45 \times 10^{-10}$).}
\label{fig:cmssm-tb10w}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.5\linewidth]{a0_tb10_mu+_tr_lsp.eps}
\caption[text]{ \small Maximal reheating temperatures in the $m_{1/2}$–-$m_0$ plane
corresponding to the values of w in Fig.~\ref{fig:cmssm-tb10w}.}
\label{fig:cmssm-tb10tr}
\end{center}
\end{figure}
We are now ready to present some representative numerical results. We are going to calculate the abundance of the MLSP (MSSM LSP), ${\rm w} = Y_{\rm MLSP}/Y_{\rm CDM}$, at freeze-out. Notice that ${\rm w} = \omega$ if the MLSP is the NLSP, i.e. with gravitino LSP, but ${\rm w}= (1-\omega)$ when the gravitino is the NLSP. As a result, the reheating temperature will
be proportional to $(1-{\rm w})$ in both cases.
In the following we will focus on three representative $\tanb$ values: one moderate, $\tanb=10$, and two large values, $\tanb=50$ and $\tanb=55$, and we
take $A_0 =0$. Indeed, we have performed
several other scans varying both $A_0$ and $\tanb$ and we found that there
are no significant differences to these cases.
\subsection{Low-medium $\tan \beta = 10$}
In Fig.~\ref{fig:cmssm-tb10w} we show the region of ${\rm w} \leq 1$ (corresponding to $\Omega_{\chi}h^2 \leq 0.115$) in the
($m_{1/2}$, $m_0$) plane for $A_0 = 0$, $\tanb = 10$, whereas we have chosen
$\mu > 0$ motivated by $(g-2)_\mu$ data.
In the CMSSM, the only mechanism which provides a dark matter abundance consistent with the WMAP value, if we require agreement with $(g-2)_\mu$ at 3 $\sigma$, is stau-neutralino
coannihilations \cite{Nihei:2002sc}.
This region is located in a narrow band above the stau-neutralino
degeneracy line ($\mstau= m_\chi$).
In addition to this relic abundance constraint, the only effective constraint
in this region comes from direct Higgs searches
at LEP which excludes low gaugino masses $m_{1/2} \lesssim 300 \gev$. In this figure, the allowed values for ${\rm w}$ are shown in different colours.
As we can see, in this case, we can only obtain ${\rm w}$ between 1 and $\mathcal{O}(10^{-1})$. In the region of $\mstau\leq m_\chi$, although the annihilation
mechanisms are more efficient than in the neutralino case due to fact that
the stau is a charged particle, the lowest ${\rm w}$ we can reach is
${\rm w}_{\min} \sim 7 \times 10^{-2}$, whereas from Eq.~(\ref{CBBNomega})
we see that the most conservative bound $Y_\stau < 10^{-14}$ from catalyzed
nuclear reactions would require ${\rm w} \lesssim 2.7 \times 10^{-3}$.
Hence, this region of $\mstau\leq m_\chi$ is completely excluded for all values of $\Delta M$. If we eliminate the requirement of a non-zero SUSY
contribution to $(g-2)_\mu$, we obtain two small regions. First, there is a vertical strip at high $m_0$ corresponding to a pole in the annihilation $\chi \chi \to b \bar{b}$ via the
lightest Higgs. In this region we can obtain values of ${\rm w}$ as low as ${\rm w} \simeq 0.01$. Then, there is a long strip, below the non-electroweak
symmetry breaking (NEWB) region, at very large $m_0$, where
the neutralino is a mixed Bino-Higgsino state and the annihilation to $W^+ W^-$ is efficient enough to get a right relic abundance. However, in this region we obtain always ${\rm w} \gtrsim 0.1$.
Using these allowed values for w and the corresponding $\omega$ values, we obtain a bound on the mass difference $\Delta M$ in our degenerate gravitino scenario from Figure \ref{fig:combinedneutr}. According to Figure \ref{fig:combinedneutr}, for $\omega \geq 0.03$ the relevant constraint is diffuse gamma ray observations, and this requires $\Delta M \leq 10^{-2}$ GeV. This means that the lifetime of the NLSP (neutralino or gravitino) is longer than the age of the universe.
Notice that, in the case of gravitino LSP, the fact that we cannot obtain w~$=\omega<0.1$ implies that this scenario would be completely ruled out by BBN, CMB and diffuse gamma ray observations, unless $\Delta M < 10^{-2}$ GeV and the neutralinos are beginning to decay at present. However, if the gravitino is the NLSP, we have two possibilities. The first possibility would be that the observed dark matter abundance is provided completely by the neutralino w~$\simeq 1$, with only a small fraction, at the level of $10^{-3}$, due to gravitinos. This corresponds to $\treh$ of the order of $10^7$ GeV, which is the usual situation in previous studies. The second possibility would be that there is a sizeable fraction of the dark matter due to gravitinos but, again, this would require $\Delta M < 10^{-2}$ GeV corresponding to a very long gravitino lifetime which permits to evade the cosmological bounds.
The maximum values of $\treh$ in these scenarios are shown in Fig.~\ref{fig:cmssm-tb10tr}. As it can be seen from this figure, values of $\treh > 10^9 \gev$ consistent with thermal leptogenesis are accessible in the model, although this is only possible if $\Delta M \leq 10^{-2}$ GeV.
\subsection{Large $\tan \beta = 50$}
\label{tgb50}
In Fig.~\ref{fig:cmssm-tb50w}, we analyze the case of $\tanb = 50$.
For large values of $\tanb$, in the neutralino MSSM LSP region ($m_\chi\leq \mstau$), we have different mechanisms to get a correct relic density consistent with WMAP. These mechanisms are i) stau-neutralino coannihilations, ii) ``A-pole'' region,
where the s-channel exchange of the CP-odd Higgs boson, $A$, can become nearly
resonant \cite{Ellis:2001msa, Roszkowski:2001sb}
and iii) the ``focus point'' or ``hyperbolical branch'' where a significant
higgsino component, enhances
its annihilation cross sections into final states containing gauge
and/or Higgs bosons \cite{Chan:1997bi, Feng:1999mn, Feng:1999zg}.
In fact, the focus point region occurs at $m_0$ much larger than $m_{1/2}$ and therefore multi-TeV scalar masses which implies that
$\delta a_\mu^{MSSM}$ is well below the lower $2\sigma$ limit due to SUSY
decoupling \cite{Moroi:1995yh}. However, if we accept $\delta a_\mu^{MSSM}$ at the $3\sigma$ level we can also reach the ``focus point'' region for $m_0 \simeq 2 \times 10^3 \gev$.
Both the neutralino-stau coannihilation band and the ``A-pole'' region merge for $\tanb =50$ as we can see in this figure. This region is cut from the small $m_{1/2}$ values by the lower $2\sigma$ limit of the $\BR(\overline{B}\rightarrow X_s\gamma)$ constraint \footnote{The reason for this is the destructive interference of the
chargino/squark loops which grows with $\tanb$ \cite{Degrassi:2000qf}.
This effect is enhanced for chargino masses $ \leq 100 \, \gev$
whereas is never dramatic for masses around $1$ TeV or larger.}.
This constraint limits the minimum possible value of w obtainable through the ``A-pole''. Therefore we can only reach ${\rm w}_{\min} \sim 0.1$ both in the coannihilation and ``A-pole'' regions. In the ``focus point'' region the situation is again similar and we can reach only values ${\rm w}_{\min} \sim 0.1$--$0.01$.
In the case of $\mstau\leq m_\chi$, we can observe an allowed narrow band
close to the tachyonic region in which the staus annihilate to
$b \bar b$ via a pole in the s-channel exchanging a lightest Higgs.
This process allows for values of ${\rm w}_{\min} \sim 10^{-5}$ which are
consistent with the conservative CBBN constraint,
$Y_\stau < 10^{-14}$--$10^{-15}$. The broader band at large $m_{1/2}$
corresponds to the annihilation into a pair of light Higgses exchanging
a light Higgs via a s-channel and here we
obtain ${\rm w}_{\min} \sim 10^{-3}$ and therefore $Y_\stau < 10^{-14}$.
On the other hand, we have to keep in mind that if we had imposed the more severe constraint $Y_\stau < 10^{-16}$, both regions would be ruled out.
Likewise, with these values for w, we obtain, from Figure \ref{fig:combinedneutr}, the allowed values for the mass difference $\Delta M$.
In the neutralino-stau coannihilation or A-pole regions only $\Delta M \leq 10^{-2}$ GeV are allowed, similarly to the $\tanb = 10$ case. However, in the region of $\mstau\leq m_\chi$ with values of w~$\gtrsim 10^{-5}$, we can see from Fig.~\ref{fig:combinedstau} that values of the mass difference, $90 \gev > \Delta M \gtrsim 1$ GeV, would be still allowed for these points \footnote{We are considering only $\Delta M < 90 \gev$ where the hadronic BBN constraints are not efficient.}.
The maximum values of $\treh$ are shown in Fig.~\ref{fig:cmssm-tb50tr}. As we have seen before, both for the gravitino LSP and gravitino NLSP cases $\treh \propto(1- {\rm w})$. The Maximal reheating temperatures for $m_{\chi}<m_\stau$ are always $\treh > 10^9 \gev$ if the condition of $\Delta M \lesssim 10^{-2}$ GeV is satisfied, as in the case of $\tan \beta=10$. However, the region below the
stau--neutralino degeneracy line are only allowed in the case of gravitino LSP when the maximal reheating temperatures are again $\treh > 10^9 \gev$, although
in this case we do not require a tight degeneracy between gravitino and stau.
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.5\linewidth]{a0_tb50_mu+_w.eps}
\caption[text]{ \small Values of w in the $m_{1/2}$–-$m_0$ plane of the
CMSSM parameter space for $m_t=173.1 \, \gev$, $A_0=0$, $\tanb=50$ and
$\mu>0$. In this case, the red region represents the bound from BR$(b \to s \gamma)$ at $2\sigma$ and there is a new region in dark brown (dark grey) marked ``TACHYONS'' corresponding to the presence of tachyonic masses. The other colored regions have the same meaning as in the case of $\tan \beta =10$. Regions below the dashed green lines satisfy the $(g-2)_\mu$ constraint at the 2
$\sigma$ or 3 $\sigma$ level.}
\label{fig:cmssm-tb50w}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.5\linewidth]{a0_tb50_mu+_tr_lsp.eps}
\caption[text]{\small Maximal reheating temperatures in the $m_{1/2}$–-$m_0$ plane
corresponding to the values of w in Fig.~\ref{fig:cmssm-tb50w}. Notice that the region below the stau--neutralino degeneracy line is not allowed in the case of gravitino NLSP.}
\label{fig:cmssm-tb50tr}
\end{center}
\end{figure}
\subsection{Large $\tan \beta = 55$}
In Fig.~\ref{fig:cmssm-tb55w}, the case of $\tanb = 55$ is explored.
Similarly to the case of $\tan \beta=50$, in the neutralino MSSM LSP region, the main annihilation mechanisms are
the neutralino-stau coannihilation and the ``A-pole'' resonance, however the ``focus point'' region is absent in this case. Again the $\BR(\overline{B}\rightarrow X_s\gamma)$ constraint limits the minimum possible value of w obtainable through the ``A-pole''. Therefore we can only reach ${\rm w}_{\min} \sim 0.1$, both in the coannihilation and ``A-pole'' regions.
In the case of $\mstau\leq m_\chi$, as for $\tan \beta=50$, staus annihilate through an s-channel
exchange of the light Higgs to a pair of light Higgses, although the annihilation to $b \bar b$ is suppressed in this case (it does not satisfy the required constraint $Y_\stau < 10^{-14}$). The allowed process, $\stau \stau \to h h $,
allows for values of
${\rm w}_{\min} \sim 6 \times 10^{-4}$ which are consistent with the conservative CBBN constraint for $Y_\stau < 10^{-15}$.
Likewise, with these values of w, from Figure \ref{fig:combinedneutr} we obtain the allowed values for the mass difference $\Delta M$.
As before, in the neutralino-stau coannihilation or A-pole regions only $\Delta M \leq 10^{-2}$ GeV are allowed and in the region of $\mstau\leq m_\chi$ mass differences $90 \gev > \Delta M >2$ GeV would be still allowed.
The corresponding values of $\treh$ are shown in Fig.~\ref{fig:cmssm-tb55tr}.
The discussion of section \ref{tgb50} applies also in this case.
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.5\linewidth]{a0_tb55_mu+_w.eps}
\caption[text]{\small Values of w in the $m_{1/2}$–-$m_0$ plane of the
CMSSM parameter space for $m_t=173.1 \, \gev$, $A_0=0$, $\tanb=55$ and
$\mu>0$. The meaning of the different regions is the same as in Fig.~\ref{fig:cmssm-tb50w}. The regions below the dashed green lines satisfy the $(g-2)_\mu$ constraint at the 2
$\sigma$ or 3 $\sigma$ level.}
\label{fig:cmssm-tb55w}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[angle=0,width=0.5\linewidth]{a0_tb55_mu+_tr_lsp.eps}
\caption[text]{\small Maximal reheating temperatures in the $m_{1/2}$–-$m_0$ plane
corresponding to the values of w in Fig.~\ref{fig:cmssm-tb55w}. As before the region below the stau--neutralino degeneracy line is not allowed in the case of gravitino NLSP.}
\label{fig:cmssm-tb55tr}
\end{center}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
Even though they are attractive candidates for cold dark matter,
gravitinos are usually a problem in standard cosmology. This is due to
the fact that they typically decay at or just after the BBN putting
in danger the successful light-elements abundances. Requiring that
gravitinos do not conflict with big-bang nucleosynthesis implies that
the reheating temperature should be low. In general, there is no
experimental constraint on how low $T_{\rm RH}$ should be, so in
principle a reheating temperature as low as MeV, so to permit BBN, is
perfectly allowed. However, successful thermal leptogenesis requires a
high reheating temperature $T_{\rm RH}\gtrsim 2 \times 10^9$ GeV.
In this paper, we propose a solution that alleviates this tension by
making the gravitino degenerate with the lightest MSSM particle. This
has the direct consequence that the injected energy is suppressed,
making the decay products less dangerous for BBN. Due to the small
mass splitting, the gravitino and the MSSM lightest particle are
typically long-lived. We analysed this scenario (the ``degenerate
gravitino'' scenario) by confronting it to cosmological and
astrophysical constraints. Since the NLSP decays at or after BBN, we
considered in addition to BBN, constraints from CMB spectral
distortions and diffuse gamma rays observations. First we performed a
model-independent analysis by considering a generic NLSP-LSP
degenerate scenario where the NLSP decays through NLSP$\to$LSP+ $X$
($X = \gamma, \tau$). Since the final cold dark matter relic density
is the sum of both thermal and non-thermal contributions, we required
that the total cold dark matter is consistent with cosmological
observation. Then, using the results of this analysis, we studied the
degenerate gravitino scenario in the context of the CMSSM where three
types of spectra arise, they are: gravitino NLSP with neutralino LSP
and gravitino LSP with neutralino or stau NLSP. Each of these cases
has been analysed in this framework defined by the usual high energy
parameters ($m_0$, $m_{1/2}$, $A_0$, $\tan\beta$, sign$(\mu)$) and the
gravitino mass $m_{3/2}$, and confronted with low-energy
observables. We find that high reheating temperatures consistent with
thermal leptogenesis can be found in all three scenarios if a sizable
part of cold dark matter comes from gravitinos produced at
reheating. In this case, depending on $\tan\beta$, we are led to
regions in the parameter space where the relic density of stau and
neutralinos are somewhat suppressed. In general to satisfy all the
constraints, the mass splitting between the NLSP and LSP should be
$\Delta M\simeq 10^{-2}$ GeV for the degenerate neutralino-gravitino
scenario, which implies very long-lived NLSPs which are beginning to
decay at present. On the other hand, in the gravitino-stau scenario,
splittings in the range 10 GeV $\lesssim\Delta M \lesssim 90$ GeV are
still consistent with reheating temperatures of the order of $10^9$
GeV if we consider the conservative CBBN constraint $Y_{\rm CBBN} \leq
10^{-15}$.
Let us comment on the required degeneracy in the ``degenerate gravitino'' scenario. Although a degeneracy of the order of $\Delta M\simeq 10^{-2}$ GeV certainly implies a certain amount of fine-tuning, this tuning is only two orders of magnitude stronger than the usual tuning required in the coannihilation or funnel regions to obtain the right relic density in the MSSM. On the other hand, notice also that the fine tuning in our scenario is much softer that the tuning required in other scenarios like inelastic dark matter \cite{TuckerSmith:2001hy}.
Finally, it is also important to consider the phenomenological consequences of
this scenario in colliders. In the case of neutralino LSP or NLSP, the only
indirect signal of this scenario will be that the relic density of neutralinos inferred from the measurements of supersymmetric masses and couplings at LHC, will not match the observed cold dark matter abundance and will be smaller.
However, the measurements at direct detection experiments will agree with the
cross sections obtained from colliders.
On the other hand, for stau NLSP, the collider signatures would be spectacular,
as the staus would be completely stable on collider scales and slow charged
tracks will appear in the detector \cite{champ}. Notice that similar signatures can arise in other scenarios like
for instance in degenarate neutralino-stau scenario \cite{Kaneko:2008re} or in gauge-mediation scenarios. However, since the typical stau lifetimes
in these scenarios are smaller than 1 sec, some of the staus will decay inside the detector. In contrast,
stau lifetimes range from $10^9-10^{15}$ seconds in our scenario, making it very difficult to observe. Nevertheless, following the analysis
of \cite{Asai:2009ka}, stau lifetimes could be measured at LHC for mass splittings $30 \gev \lesssim \Delta M\lesssim 90 \gev$, corresponding
to lifetimes $10^{10} \sec \gtrsim \tau_\stau \gtrsim 10^8 \sec$. In this case, direct detection experiments will give a null results as all the dark matter at present times is made of gravitinos. Therefore, the "degenerate gravitino" scenario will be probed at colliders and direct detection experiments if SUSY is discovered at LHC. \\[.0cm]
\noindent{\bf Note added:}~
While completing this work we noticed the preprint \cite{Peter:2010au}
where the astrophysical consequences in dark matter halo properties of
a neutral long lived decaying-dark matter particle were studied. The required
parameters in our ``degenerate gravitino'' scenario in the case of
neutralino LSP or NLSP are such that the constraints of \cite{Peter:2010au} are
satisfied.
\section*{Acknowledgments}
We thank T.~Moroi for useful discussions.
O.~V. and L.~B. acknowledge financial support from spanish MEC and FEDER (EC) under grant FPA2008-02878 and Generalitat Valenciana under the grant PROMETEO/2008/004.
O.~V was supported in part by European program MRTN-CT-2006-035482
``Flavianet''. L.~B. thanks the Abdus Salam ICTP, the Service de Physique Th\'eorique of ULB Brussels and the CERN-TH division for hospitality
during the completion of this work.
K.Y. Choi was partly supported by the Korea Research
Foundation Grant funded by the Korean Government (KRF-2008-341-C00008)
and by the second stage of Brain Korea 21 Project in 2006.
The work of R. Ruiz de Austri has been supported in part by MEC (Spain)
under grant FPA2007-60323, by Generalitat Valenciana under grant
PROMETEO/2008/069 and by the Spanish Consolider-Ingenio 2010 Programme
CPAN (CSD2007-00042).
The use of the ciclope cluster of the IFT-UAM/CSIC is also acknowledged.
|
\section{Introduction}
\label{INTRO}
The asymptotic conformal invariance of QCD has fostered a number of
theoretical attempts to push the original AdS/CFT duality \cite{MA98}
beyond its conjectured domain of validity. The AdS/CFT correspondence
postulates more generally a relation, through a well-defined set of
prescriptions \cite{WI98, GKP98}, between weakly coupled string
theories living in the bulk of an anti-de Sitter (AdS) space and
strongly coupled conformally invariant field theories defined on its
boundary. The efforts within the so-called AdS/QCD approach rely on
the assumption that the AdS/CFT dictionary can still describe the
strong coupling regime of a confining gauge theory like QCD, despite
the breaking of conformal invariance, by computing correlation
functions within a semi-classical field theory formulated in a
five-dimensional AdS spacetime.
The various AdS/QCD models usually fall into two main categories. In
the so-called ``top-down'', or ``gauge/string'', approach one sticks
as much as possible to the original formulation of the AdS/CFT
correspondence, and starts from a superstring theory living on
$AdS_5\times M_5$ (where $M_5$ is some five-dimensional compact
manifold) and tries to derive an effective theory which describes the
low energy phenomena in QCD. The second, ``bottom-up'', or
``gauge/gravity'', approach puts aside the string-theoretic motivation
of the correspondence and starts from a phenomenological Lagrangian in
an appropriate five dimensional metric background which
incorporates as much as possible the known properties of QCD. Contrary
to some claims, the former approach is no more rigourous, in its
present stage, than the latter.
The simplest way to break conformal invariance is by introducing a
hard cut-off in AdS space which can be interpreted as an infrared mass
scale of the gauge theory. The first ``hard-wall'' model \cite{PS01}
was able to generate a power-law scaling of glueball elastic
scattering amplitudes at fixed angles. It is also possible to break
conformal invariance softly through a background dilaton field which
can be chosen \cite{KKSS} so as to reproduce the linear Regge behavior
of the meson trajectories. Inside the bottom-up approach, it has
proved possible to reproduce qualitatively within both hard-wall and
soft-wall models the spectra of low-lying hadron states as well as
various decay constants and coupling constants
\cite{BOS02,TER05,HAM06,COL08}.
A frequent criticism claims that AdS/QCD models are nothing but some
kind of ``bag models'' since quite a few other phenomenological models
of QCD are able to reproduce the static parameters of hadronic states
at the same level of accuracy. In fact such criticisms are pretty
superficial because they ignore that AdS/QCD can incorporate in a
coherent framework several of these models (vector meson dominance,
1/N expansion, sum rules, $\cdots$, see \cite{EH09} for an up-to-date
review of the pros and cons of the AdS/QCD approach).
Moreover the AdS/QCD models can also be used to study in a completely
relativistic Lorentz invariant manner dynamical non-perturbative
aspects like hadronic form factors or structure functions. In
particular one can expect that conformal symmetry is most relevant for
describing the electromagnetic interactions of hadrons, since the
photon is a massless particle \footnote{e.g. there is a dynamical
$O(4,2)$ symmetry which is sufficient to determine completely the
spectrum and eigenfunctions of the relativistic hydrogen atom.}. A
very important process that reveals the internal structure of hadrons
is the deep inelastic scattering (DIS) of a highly energetic lepton
off a hadron. This process was first investigated in a hard-wall
model in \cite{PS03}. Since then, there
have been a number of related studies within the different flavors of
AdS/QCD models \cite{HAT07,BAY08-1,BAY08-2,COR08,ALB08,HAT08,BAY08-3,
GAO09,HAT09,LEV09,COR10,KOV09}. However the semi-classical density
of states of the hard-wall or soft-wall models with canonical
dimensions does not agree \cite{PRSW} with the power-law behavior of
the structure functions observed at high energy. It is presently far
from clear how a correct partonic description of hadrons can emerge
from the stringy corrections expected in the kinematical regime of
DIS.
On the other hand the AdS/QCD models provide a low-energy description
of the electromagnetic form factors which does agree with the
dimensional counting rules of hadrons up to a few GeV
\cite{GR07-1,GR07-2,BRO07,KL07,AC08,WARD08}. Getting additional
information about the electromagnetic structure of hadrons requires
the study of four-point functions. There is one distinguished
electromagnetic process which supplies all experimentally accessible
information, namely Compton scattering. There has been a lot of
theoretical work about Compton scattering. Strong interactions can
significantly modify the amplitude but there is a low-energy theorem
\cite{LOW,GG54} which guarantees that the Born contribution dominates
near threshold. The two leading orders of an expansion of the real
Compton scattering amplitude off a nucleon in terms of the frequency
of the photon are entirely given in terms of the charge, mass and
anomalous magnetic moment of the nucleon. For spinless targets like
the pion, there is no linear term in the energy of photons. This
theorem is based only on Lorentz invariance, gauge invariance and
crossing symmetry. Quadratic corrections to the Born terms (electric
and magnetic polarizabilities as well as generalized polarizabilities
with virtual photons) are not specified by symmetry arguments alone
and characterize non point-like elementary particles. Measuring these
quantities allows to test the different models of strong interactions
(see \cite{PP95} for a review of the theoretical predictions on
charged and neutral pion polarizabilities).
The purpose of this work is to study the Compton amplitude off a
spinless target in the AdS/QCD formalism. We shall focus on the
kinematical region where the photons are soft. The semi-classical
approximation in the AdS/QCD duality should apply best in this region
and there exist experimental measurements of the pion polarizabilities
to compare with \cite{MAMI,COMPASS}. But we shall also study the
kinematical region with one deeply virtual photon (DVCS) since it is
straightforward to extract the corresponding structure functions in
the Bjorken limit.
The plan of the paper is as follows. We shall begin by a brief summary
of the standard lore about the general structure of the virtual
Compton amplitude off an unpolarized target. Next we shall describe
the generic soft-wall model we work with. In section \ref{4PT} we
shall explain how to calculate four-point functions in AdS/QCD and in
section \ref{SCALAR} we shall compute the Compton amplitude generated
by the minimal coupling of a bulk scalar field with a bulk U(1) gauge
field. We also compare our results with the recent hard-wall
calculation of \cite{GX09}. In the subsequent section we shall make
explicit the structure of this Compton amplitude in the deep inelastic
region, and in section \ref{DVCS} we shall clarify its Lorentz and
gauge-invariant structure in the DVCS kinematical region. Then we
shall extract the corresponding structure functions in the Bjorken
limit. Section \ref{POL} is devoted to the calculation of the
polarizabilities of a spinless target. We shall comment on the
implications of our results in the conclusion.
\section{Virtual Compton amplitude off an unpolarized target}
\label{COMPTON}
The amplitude of the virtual Compton scattering, $\gamma^{\star}(q_1)
+ A(p_1) \rightarrow \gamma^{\star}(q_2) + A(p_2)$, where $A$ is a
spinless, or spin-averaged, hadron is defined through the off-forward
matrix element of the time-ordered product of two electromagnetic
currents,
\begin{gather}
T_{\mu\nu} = i\int d^4x\,e^{iq\cdot x}\,\langle p_2
\vert T\{J_{\mu}(x/2)J_{\nu}(-x/2)\}\vert p_1\rangle\,,\quad
q=\frac{1}{2}(q_1+q_2)\,.
\end{gather}
In general, a two-to-two scattering amplitude depends on six
independent kinematical invariants, namely the external virtualities
$q_1^2, q_2^2, p_1^2, p_2^2$ and the usual Mandelstam variables
\begin{gather}
s = (p_1+q_1)^2\,,\ t=(p_1-p_2)^2\,,\ u= (p_2-q_1)^2\,,
\end{gather}
obeying the constraint
\begin{gather}
s + t + u = q_1^2 + q_2^2 + p_1^2 + p_2^2\,.
\end{gather}
It is convenient, for calculating the Compton form factors, to choose
$q_1$, $q_2$ and $p=p_1+p_2$ as the three independent momenta of the
process. At most thirteen independent tensors can contribute to the
Compton amplitude,
\begin{gather}
\begin{split}
g^{\mu\nu},\,p^{\mu}p^{\nu},\,q_1^{\mu}q_1^{\nu},\,
q_2^{\mu}q_2^{\nu},\,q_1^{\mu}q_2^{\nu},\,q_2^{\mu}q_1^{\nu},\,
p^{\mu}q_1^{\nu},\,q_2^{\mu}p^{\nu},\,p^{\mu}q_2^{\nu},\,q_1^{\mu}p^{\nu}, \\
\epsilon_{\mu\nu\rho\sigma}p^{\rho}q_1^{\sigma},\,
\epsilon_{\mu\nu\rho\sigma}p^{\rho}q_2^{\sigma},\,
\epsilon_{\mu\nu\rho\sigma}q_1^{\rho}q_2^{\sigma}\,.\qquad\qquad\quad
\end{split}
\end{gather}
The antisymmetric tensors are parity-violating. Electromagnetic gauge
invariance implies that
\begin{align}
q_1^{\mu}T_{\mu\nu} = T_{\mu\nu}\,q_2^{\nu} = 0\,.
\end{align}
From the vanishing of the six coefficients of the linearly independent
vectors, one can deduce five linearly independent conditions. Hence
the most general spin-averaged, gauge-invariant, and parity-conserving
Compton amplitude has five independent form factors:
\begin{align}
\label{Vmunu}
\begin{split}
T^{\mu\nu} &= V_1\left(g^{\mu\nu} - \frac{q_1^{\mu}q_1^{\nu}}{q_1^2}
- \frac{q_2^{\mu}q_2^{\nu}}{q_2^2}
+ q_1^{\mu}q_2^{\nu}\frac{(q_1.q_2)}{q_1^2q_2^2} \right) \\
&+ V_2\left( p^{\mu} - q_1^{\mu}\frac{(p.q_1)}{q_1^2}\right)
\left(p^{\nu} - q_2^{\nu}\frac{(p.q_2)}{q_2^2}\right) \\
&+ V_3\left(q_2^{\mu} - q_1^{\mu}\frac{(q_1.q_2)}{q_1^2}\right)
\left(q_1^{\nu} - q_2^{\nu}\frac{(q_1.q_2)}{q_2^2}\right) \\
&+ V_4\left( p^{\mu} - q_1^{\mu}\frac{(p.q_1)}{q_1^2}\right)
\left(q_1^{\nu} - q_2^{\nu}\frac{(q_1.q_2)}{q_2^2}\right) \\
&+ V_5\left(q_2^{\mu} - q_1^{\mu}\frac{(q_1.q_2)}{q_1^2}\right)
\left(p^{\nu} - q_2^{\nu}\frac{(p.q_2)}{q_2^2}\right)\,.
\end{split}
\end{align}
The form factors $V_1$, $V_2$, $V_3$, $V_4$ and $V_5$ can be readily
identified as the coefficients of the tensors $g^{\mu\nu}$,
$p^{\mu}p^{\nu}$, $q_1^{\nu}q_2^{\mu}$, $p^{\mu}q_1^{\nu}$ and
$p^{\nu}q_2^{\mu}$ respectively. They are in general functions of the
six independent scalar invariants. The off-shell Compton form factors
are not directly measurable. The Compton form factors of on-shell
virtual Compton amplitudes, defined by the conditions,
\begin{align}
p_1^2 = p_2^2 = -M_H^2\,,
\end{align}
depend only on four independent scalar invariants. The gauge-invariant
tensors in \eqref{Vmunu} will be denoted respectively as
$\V_i^{\mu\nu}(p,q_1,q_2)\,,\ i=1,\cdots 5$.
We shall be interested more particularly in two kinematical regimes,
according to whether one or two photons are real. In
electrophotoproduction the outgoing photon is real, $q_2^2=0$, and
thus transversely polarized. One can contract the Compton amplitude
with the polarization $\epsilon_2$, set $\epsilon_2\cdot q_2=0$ and
still impose a gauge condition on the outgoing photon, e.g.
$\epsilon_2\cdot p = 0$, by choosing the Coulomb gauge
$\epsilon_2^0=0$ and the frame $\v{p}=\b{0}$. Then the contracted
amplitude becomes,
\begin{align}
\label{Vmu}
\begin{split}
A^{\mu}_{\text{VCS}} &= T^{\mu\nu}\,\epsilon_{2\nu}^{\star} =
V_1\left(\epsilon_2^{*\mu} -
\frac{q_1^{\mu}}{q_1^2}(\v{\epsilon_2^*}\cdot\v{q_1})\right)
+ V_3\left(q_2^{\mu} - q_1^{\mu}\frac{(q_1.q_2)}{q_1^2}\right)
(\v{\epsilon_2^*}\cdot\v{q_1}) \\
&+ V_4\left( p^{\mu} - q_1^{\mu}\frac{(p.q_1)}{q_1^2}\right)
(\v{\epsilon_2^*}\cdot\v{q_1}) \,.
\end{split}
\end{align}
Therefore there are only three independent form factors when the
outgoing photon is real.
Finally if the ingoing photon is also real, $q_1^2=0$, one can
contract the amplitude $A^{\mu}_{\text{VCS}}$ with the polarization
$\epsilon_1$ and impose similarly the conditions $\epsilon_1\cdot q_1
= \epsilon_1\cdot p = 0$. Hence the real Compton amplitude has in
general two independent form factors,
\begin{align}
\label{RCS}
A_{\text{RCS}} = \epsilon_1^{\mu}\,T_{\mu\nu}\,\epsilon_2^{\star\nu}
= V_1\,\v{\epsilon_1}\cdot\v{\epsilon_2}^{\star}
+ V_3\,(\v{\epsilon_1}\cdot \v{q_2})(\v{\epsilon_2}^{\star}\cdot\v{q_1}) \,.
\end{align}
\section{The soft-wall model}
\label{MODEL}
The AdS/CFT correspondence is based upon the fact that the isometry
group of the five-dimensional (5D) anti-de Sitter space is the same as
the four-dimensional conformal group $SO(4,2)$. In Poincar\'e coordinates,
the AdS$_5$ metric reads
\begin{align}
\label{ADS}
ds^2 &= \frac{R^2}{z^2}\left(\eta_{\mu\nu}dx^{\mu}dx^{\nu}+dz^2\right)\,,
\qquad \sqrt{-g} = \frac{R^5}{z^5}\,,
\end{align}
where $\eta_{\mu\nu}\equiv(-1,1,1,1)$ is the four-dimensional
Minkowski metric. We shall set the curvature radius $R$ to 1 from now on.
Following \cite{PS03} we introduce a massless 5D vector field
$A_m(x,z)$ with a $U(1)$ gauge invariance, which is dual to the
electromagnetic current. The free field $A_m(x,z)$ must satisfy the
Maxwell equations in the bulk (with no dilaton coupling, which would
break the conformal invariance of the electromagnetic field),
\begin{align}
\label{MAXWELL}
(\nabla_m F)^{mn} = \frac{1}{\sqrt{-g}}\,
\p_m\left(\sqrt{-g}F^{mn}\right) = 0\,.
\end{align}
Choosing the linear gauge fixing condition,
\begin{align}
\p^{\mu}A_{\mu} + z\partial_z\left(z^{-1}A_z\right) = 0\,,
\end{align}
the general plane-wave solution of \eqref{MAXWELL} in the space-like
region, $Q^2=q\cdot q>0$, reads
\begin{align}
\begin{split}
A_{\mu}(x,z) &= \epsilon_{\mu}e^{iq\cdot x}QzK_1(Qz)\,,\\
A_z(x,z) &= -i\frac{(\epsilon\cdot q)}{q^2}e^{iq\cdot x}
\p_z\left(QzK_1(Qz)\right)\,,
\end{split}
\end{align}
where $\epsilon_{\mu}$ is a polarization vector. The boundary
condition is chosen such that the solution becomes a plane-wave on the
Minkowski slice at $z=0$,
\begin{align}
\label{BC}
\lim_{z\rightarrow 0} A_{\mu}(x,z) = \epsilon_{\mu}\,e^{iq\cdot x}\,.
\end{align}
The boundary condition in the timelike region, $q^2<0$, can be
obtained by analytic continuation in the $q^2$ variable. A crucial
property of the boundary condition \eqref{BC} is that the vector field
$A_{\mu}(x,z)$ is in fact a constant plane-wave throughout the bulk of
the AdS space when $q^2=0$.
We introduce a massive 5D scalar field $\Phi(x,z)$ which will be the
dual of an operator which creates the spinless target. Following
\cite{KKSS} the bulk scalar field is coupled to a background dilaton
field $\chi(z)$ which deforms the AdS$_5$ metric. The action which
describes the propagation of $\Phi$ in this background reads
\begin{align}
S_{\Phi} = \frac{1}{2}\int
d^4x\,dz\sqrt{-g}e^{-\chi}\left(g^{ij}\partial_i\Phi\partial_j\Phi +
m^2_S\Phi^2\right)\,,
\end{align}
where $g$ is the AdS$_5$ metric. The classical field equation reads
\begin{align}
\label{laplace}
\Delta_g\Phi \equiv \frac{e^{\chi}}{\sqrt{-g}}\partial_i
\left(e^{-\chi}\sqrt{-g}g^{ij}\p_j\Phi\right)
= m_S^2\Phi\,.
\end{align}
In Poincar\'e coordinates, the Laplacian equation becomes
\begin{align}
\label{scalar}
z^2\square\Phi + z^5e^{\chi}\partial_z
\left(z^{-3}e^{-\chi}\partial_z\Phi\right)
= m_S^2\Phi\,.
\end{align}
Looking for a solution that is a plane-wave in Minkowski space and setting
\begin{align}
\Phi(x,z) = e^{ip\cdot x}e^{\chi(z)/2}z^{3/2}\psi(z) \equiv
e^{ip\cdot x}\,\h{\Phi}(z) \,,
\end{align}
the Laplacian equation is transformed into a Schr\"odinger-like equation
\begin{gather}
\frac{d^2\psi}{dz^2} - V(z)\psi = p^2\psi\,, \\
V(z) = \frac{m_S^2+15/4}{z^2} + \frac{3}{2z}\partial_z\chi
+ \frac{1}{4}(\partial_z\chi)^2 - \frac{1}{2}\partial_z^2\chi\,.
\end{gather}
If the potential $V(z)$ has the right properties, and with appropriate
boundary conditions, this equation has a complete
set of solutions $\{\psi_n(z)\,,\ n\in\NB\}$ which form an
orthonormal basis of the Hilbert space $\H$ defined by the inner
product
\begin{gather}
\langle \phi|\psi\rangle_{\H} = \int_0^{\infty}dz\,\phi^{\star}(z)\psi(z) \,.
\end{gather}
We shall assume that the dilaton background is such that $\H$ is
well-defined. Otherwise we let the dilaton profile unspecified to be
as generic as possible, except that conformal invariance in the
ultraviolet requires that $\chi(z)\rightarrow 0$ for $z\rightarrow 0$.
In terms of the plane-wave solutions of the Laplacian equation
\eqref{scalar}, with the corresponding boundary conditions, the
completeness relation takes the form
\begin{align}
\delta(z-z') = \sum_n z'^{-3/2}e^{-\chi(z')/2}\,\h{\Phi}_n^{\star}(z')\,
\h{\Phi}_n(z)\,z^{-3/2}e^{-\chi(z)/2}\,.
\end{align}
Therefore the set of classical solutions $\{\h{\Phi}_n(z)\,,\
n\in\NB\}$ form a complete orthonormal basis of the Hilbert space
$\H_S$ spanned by the solutions of the Laplacian equation and defined
by the inner product,
\begin{align}
\label{HS}
\left\langle\h{\Phi}\right|\left.\h{\Phi}'\right\rangle_{\H_S} =
\int_0^{\infty}dz\,z^{-3}e^{-\chi(z)}\,
\h{\Phi}^{\star}(z)\,\h{\Phi}'(z) \,.
\end{align}
Each $\h{\Phi}_n$ is a normalized eigenfunction, with appropriate
boundary conditions, of the operator
\begin{align}
\label{H}
\h{H}_S\,\h{\Phi}_n = \left(z^3\,e^{\chi}\partial_z
\left(z^{-3}e^{-\chi}\partial_z\right) - m_S^2z^{-2}\right)\h{\Phi}_n
= -m_n^2\h{\Phi}_n\,.
\end{align}
The scalar Green function $G(x,z;x',z')$ is defined by the
inhomogeneous equation
\begin{align}
\label{Green}
\left(\Delta_g - m_S^2\right) G(x,z;x',z') = \frac{e^{\chi}}{\sqrt{-g}}
\delta^{4}(x-x')\delta(z-z')\,.
\end{align}
Its four-dimensional Fourier transform $\h{G}(z,z';p)$
\begin{align}
\label{FT}
G(x,z;x',z') =
\frac{1}{(2\pi)^4}\int_{-\infty}^{+\infty}d^4p
\,e^{ip(x'-x)}\h{G}(z,z';p)
\end{align}
satisfies the equation
\begin{align}
\label{green}
\h{H}_S\,\h{G} = p^2\h{G} + z^3e^{\chi}\,\delta(z-z') \,.
\end{align}
$\h{G}$ has an expansion in terms of the normalized
eigenfunctions satisfying the same boundary conditions,
\begin{align}
\label{expansion}
\h{G}(z,z';p) &= -\sum_{n=0}^{\infty}
\frac{\h{\Phi}^{\star}_n(z)\h{\Phi}_n(z')}{p^2+m^2_n-i\epsilon} \,.
\end{align}
The handling of the singularities at $p^2=-m_n^2$ is done with the
standard Feynman prescription.
In order to complete the definition of the model we still need to
specify the interaction between the bulk scalar field and the bulk
$U(1)$ field. Since we are interested in describing the
electromagnetic interactions of a charged spinless hadron, we shall
take a $U(1)$ covariant coupling, \DS{D_n\Phi = \p_n\Phi -ie A_n\Phi}.
Hence we shall consider the full anti-de Sitter action,
\begin{equation}
\label{AdS}
S_{\text{AdS}}[\Phi,\Phi^*,A^m]=\int d^4xdz\ \sqrt{-g}
\left(-\frac{1}{4}F^{mn}F_{mn}
+ e^{-\chi}\left((D^m\Phi)^*D_m\Phi+m_S^2\Phi^*\Phi\right)\right)\ .
\end{equation}
\section{Calculation of four-point functions in AdS/QCD}
\label{4PT}
The gauge/gravity correspondence relates generating functions in a
strongly-coupled gauge theory to the classical supergravity
partition function in the following way:
\begin{align}
\label{corres}
\begin{split}
Z_{CFT}(c,\bar{c},n+\bar{n}) &= \left\langle\exp\left(\int d^4x\
(n_\mu+\bar{n}_\mu)J^\mu+\bar{c}\,O+c\,O^\dagger\right)\right\rangle_{CFT} \\
&= \exp\left(-S^{cl}_{\text{AdS}}
[\Phi(c),\Phi^*(\bar{c}),A^m(n_\mu+\bar{n}_\mu)]
\right) \,,
\end{split}
\end{align}
where the 4-dimensional sources for the CFT appear as boundary
conditions for the 5-dimensional classical supergravity fields.
Correlation functions of CFT operators can be obtained by expanding
\eqref{corres} to linear-order with respect to the sources. We shall
use the prescription \eqref{corres} as a recipe for the AdS/QCD model.
The correlation function we are interested in can be
obtained from the coefficient of $\bar{c}n_\mu\bar{n}_\nu c.$
In the contracted Compton amplitude, $\epsilon_{1\mu}
T^{\mu\nu}\epsilon_{2\nu}^*,$ the QCD operators are coupled to
asymptotic states, therefore these will serve as boundary conditions
for the free bulk fields $\Phi^{(0)},$ $\Phi^{*(0)},$ and
$A_\mu^{(0)}.$ After we express $S^{\text{cl}}_{\text{AdS}}$ in terms
of free fields, it will be easy to read off the
$\bar{c}n_\mu\bar{n}_\nu c$ coefficient.
With the notations of the previous section, the equations of motion
for the interacting classical bulk fields read
\begin{align}
\frac{1}{\sqrt{-g}}\partial_n\left(\sqrt{-g}F^{mn}\right)
&= ie\,e^{-\chi}\left(\Phi^* D^m\Phi-(D^m\Phi^*)\Phi\right) \,, \\
(\Delta_g-m_S^2)\Phi &= V(A)\Phi=ieV_1(A)\Phi+e^2V_2(A)\Phi \,, \\
(\Delta_g-m_S^2)\Phi^* &= \o{V}(A)\Phi^*=-ieV_1(A)\Phi^*+e^2V_2(A)\Phi^* \,,
\end{align}
where the linear operators $V_1(A)$ and $V_2(A)$ act on the right and read
\begin{align}
V_1(A)&=\frac{e^\chi}{\sqrt{-g}}\,
\partial_m\left(\sqrt{-g}e^{-\chi}A^m\right)+2A^m \partial_m \,,
\\ V_2(A)&=A^m A_m \,.
\end{align}
The function $V$ can be also used to write the interaction term in
$S_{\text{AdS}}:$
\begin{align}
S_{int}&= \int d^4xdz\ \sqrt{-g}e^{-\chi}
\left(ieA^m(\Phi^*\partial_m\Phi-\Phi\partial_m\Phi^*)
+e^2A^mA_m\Phi^*\Phi\right) \,,
\nonumber\\
&= \frac{1}{2}\int d^4xdz\ \sqrt{-g}e^{-\chi}
\left(\Phi^*V(A)\Phi+\Phi\o{V}(A)\Phi^*\right) \,.
\end{align}
The solutions for $\Phi$ and $\Phi^*$ can be written as
\begin{align}
\Phi^{(*)}(y)&=\Phi^{(*)}_{(0)}(y)+\int dy'\sqrt{-g'}e^{-\chi(z')}G(y;y')\,
\overset{(-)}{V}\left(A(y')\right)\Phi^{(*)}(y') \,,
\end{align}
where the free bulk fields $\Phi_{(0)}$, $\Phi^*_{(0)}$ and the Green
function $G$ are respectively solutions of \eqref{laplace} and
\eqref{green}. We use the shorthand notations $y=(x,z)$ and
$dy=d^4xdz$. We can now write $\Phi^*V(A)\Phi$ in terms of the free
fields:
\begin{align}
\label{phivphi}
\begin{split}
\Phi^*V(A)\Phi &= \left(\Phi^*_{(0)}(y) -
ie\int dy'\sqrt{-g'}e^{-\chi(y')}\,G(y;y')
\,V_1\left(A_{(0)}(y')\right)\Phi^*_{(0)}(y')\right) \\
&\times\left(ieV_1\left(A_{(0)}(y)\right) + e^2 V_1\left(A_{(1)}(y)\right)
+e^2 V_2\left(A_{(0)}(y)\right)\right) \\
&\times\left(\Phi_{(0)}(y) + ie\int dy''\sqrt{-g''}e^{-\chi(y'')}\,G(y;y'')
\,V_1\left(A_{(0)}(y'')\right)\Phi_{(0)}(y'')\right) + {\cal O}(e^3) \,,
\end{split}
\end{align}
and similarly for $\Phi\o{V}(A)\Phi^*.$ In $S^{cl}_{int},$ the
contribution involving $A_{(0)} A_{(0)} \Phi^*_{(0)}\Phi_{(0)}$
appears at order $e^2:$
\begin{align}
\begin{split}
S^{cl}_{int} &= \int dy\,\sqrt{-g}e^{-\chi}
\left(ieA^m(\Phi^*\partial_m\Phi-\Phi\partial_m\Phi^*)
+e^2A^m A_m\Phi^*\Phi\right) \\
&+ e^2\int dydy'\,\sqrt{-g}e^{-\chi(y)}\sqrt{-g'}e^{-\chi(y')} A^m(y)
\left\{\left(G(y;y')\partial_m\Phi^*(y)\right.\right.
\\ &\left.\left.\qquad\qquad-\Phi^*(y)\partial_m G(y;y')\right)
V_1(A(y'))\Phi(y')+\left(\Phi\leftrightarrow\Phi^*\right)\right\}
+{\cal O}(e^3) \,,
\end{split}
\end{align}
where we have dropped the ${}_{(0)}$ notation for clarity, and we are
now dealing only with free fields. Note that the term
$V_1(A_{(1)})$ in (\ref{phivphi}) contributes also at order
$e^2$ but we have dropped it since it does not contribute to
$T^{\mu\nu},$ but rather to a $\langle\Phi^{*2}\Phi^2\rangle$
correlator. Finally, after integrating by parts over $y'$ one writes:
\begin{align}
S^{cl}_{int} &= ie\int dy\,\sqrt{-g}e^{-\chi}
A^m(\Phi^*\partial_m\Phi-\Phi\partial_m\Phi^*)
+ e^2\int dy\,\sqrt{-g}e^{-\chi} A^m A_m\Phi^*\Phi \nonumber\\
&+ e^2\int dydy'\,\sqrt{-g}e^{-\chi} \sqrt{-g'}e^{-\chi'}
A^m(y)A^n(y') \nonumber\\
&\qquad\quad\left\{
\Bigl(\Phi^*(y)\partial_m-(\partial_m\Phi^*(y))\Bigl)
\Big(\Phi(y')\partial'_n-(\partial'_n\Phi(y'))\Bigl)\right.\nonumber\\
&\qquad\quad\left.+\Bigl(\Phi(y)\partial_m-(\partial_m\Phi(y))\Bigl)
\Bigl(\Phi^*(y')\partial'_n-(\partial'_n\Phi^*(y'))\Bigl)
\right\}G(y;y')\,.
\end{align}
In this expression, the boundary conditions at $z=0$ of the classical
fields $A^m,$ $\Phi,$ and $\Phi^*$ are respectively
$n_{\mu}+\bar{n}_{\mu},$ $c,$ and $\bar{c}.$ These enter in a linear
way in the fields, therefore the $\bar{c}n_\mu\bar{n}_\nu c$
coefficient in the expansion of \eqref{corres} is simply obtained.
After contractions, one can write
\begin{align}
\epsilon_\mu T^{\mu\nu}\epsilon^*_\nu
&=2 e^2\int dy\,\sqrt{-g}e^{-\chi}\
A^m(y)A^*_m(y)\Phi^*(y)\Phi(y)\nonumber\\
&+e^2\int dydy'\,\sqrt{-g}e^{-\chi} \sqrt{-g'}e^{-\chi'}
\left(A^m(y)A^{*n}(y')+A^{*m}(y)A^n(y')\right)
\nonumber\\&\quad\times
\Big(\Phi(y)\partial_m-(\partial_m\Phi(y))\Big)\
\Big(\Phi^*(y')\partial'_n-(\partial'_n\Phi^*(y'))\Big)
G(y;y')\,,
\label{fullTmunu}
\end{align}
where the fields $\Phi(x,z)$ and $A(x,z)$ are the plane-wave solutions
defined in section \ref{MODEL}.
In a diagrammatic representation, the first contribution in
(\ref{fullTmunu}) is a contact interaction, while the second
contribution contains an $s$-channel diagram (with the term
$A^m(x,z)A^{*n}(x',z')$) and a $u$-channel diagram (with the term
$A^{*m}(x,z)A^{n}(x',z')$) \cite{GX09}. This is not surprising, since
it is well-known that taking the classical limit of a quantum field
theory is equivalent to keeping only tree diagrams in the perturbative
expansion.
\section{On-shell Compton amplitude}
\label{SCALAR}
The $s$-channel contribution in eq.\,\eqref{fullTmunu} can be written as
\begin{align}
\begin{split}
\epsilon^{\mu}T_{\mu\nu}^s\epsilon^{\star\nu} &=
(ie)^2\int d^4x\,d^4x'\,dz\,dz'\,z^{-3}e^{-\chi(z)}A_k(x,z)
A^{\star}_l(x',z')z'^{-3}e^{-\chi(z')} \\
&\qquad\qquad\quad\times
\left(\Phi_{\text{in}}(x,z)\op{\p^k}_{(x,z)}G(x,z;x',z')
\op{\p^l}_{(x',z')}\Phi^{\star}_{\text{out}}(x',z')\right) \,,
\end{split}
\end{align}
where the initial and final wave-functions of the bulk scalar fields,
$\Phi_{\text{in}}$ and $\Phi_{\text{out}}$, are normalized plane-wave
solutions of the Laplacian equation \eqref{scalar}, whereas the bulk
vector field $A_k$ is a normalized plane-wave solution of the Maxwell
equations. The Green function $G(x,z;x',z')$ is defined by
eqs.\,\eqref{FT} and \eqref{green}. We shall introduce the shorthand
notations,
\begin{align}
\label{notations}
\begin{split}
&\Phi_{\text{in}}(x,z) = e^{ip_1\cdot x}\h{\Phi}_i(z)\,,\quad
\Phi_{\text{out}}(x',z') = e^{ip_2\cdot x'}\h{\Phi}_f(z')\,, \\
&A_{\mu}(x,z) = \epsilon_{\mu}e^{iq_1\cdot x}A_1(z)\,,\quad
A_z(x,z) = -i\frac{\epsilon\cdot q_1}{q_1^2}e^{iq_1\cdot x}\p_zA_1(z)\,, \\
&A_{\nu}(x',z') = \epsilon_{\nu}e^{iq_2\cdot x'}A_2(z')\,,\quad
A_z'(x',z') = -i\frac{\epsilon\cdot q_2}{q_2^2}
e^{iq_2\cdot x'}\p_{z'}A_2(z')\,,\\
&A_1(z) = Q_1z\,K_1(Q_1z) \,,\quad A_2(z) = Q_2z\,K_1(Q_2z) \,.
\end{split}
\end{align}
A straightforward calculation yields,
\begin{align}
\begin{split}
T_{\mu\nu}^s &= (2\pi)^4\delta^{(4)}(p_1+q_1-p_2-q_2)\,e^2\times\biggl(
(p_1+k)_{\mu}(p_2+k)_{\nu}\,\F_1\left(q^2_1,q_2^2,s\right) \\
&\quad
- \frac{(p_1+k)_{\mu}q_{2\nu}}{q_2^2}\,\F_2\left(q^2_1,q_2^2,s\right)
+ \frac{q_{1\mu}(p_2+k)_{\nu}}{q_1^2}\,\F_3\left(q^2_1,q_2^2,s\right)
- \frac{q_{1\mu}q_{2\nu}}{q_1^2q_2^2}\,\F_4\left(q^2_1,q_2^2,s\right)
\biggr)\,.
\end{split}
\end{align}
with $k=p_1+q_1=p_2+q_2$, $s=k^2$, and
\begin{align}
\begin{split}
\F_1\left(q^2_1,q_2^2,k^2\right) &=
\iint dz\,dz'\,\,z^{-3}e^{-\chi(z)}A_1(z)\h{\Phi}_i(z)\h{G}(z,z',k^2)
\h{\Phi}_f^{\star}(z')A_2^{\star}(z')z'^{-3}e^{-\chi(z')}\,, \\
\F_2\left(q^2_1,q_2^2,k^2\right) &=
\iint dz\,dz'\,\,z^{-3}e^{-\chi(z)}A_1(z)\h{\Phi}_i(z)\left(\h{G}(z,z',k^2)
\op{\p}_{z'}\h{\Phi}_f^{\star}(z')\right)
\p_{z'}A_2^{\star}(z')z'^{-3}e^{-\chi(z')}\,,\\
\F_3\left(q^2_1,q_2^2,k^2\right) &=
\iint dz\,dz'\,\,z^{-3}e^{-\chi(z)}\p_zA_1(z)
\left(\h{\Phi}_i(z)\op{\p}_z\h{G}(z,z',k^2)\right)
\h{\Phi}_f^{\star}(z')A_2^{\star}(z')z'^{-3}e^{-\chi(z')}\,, \\
\F_4\left(q^2_1,q_2^2,k^2\right) &=
\iint dz\,dz'\,\,z^{-3}e^{-\chi(z)}\p_zA_1(z)
\left(\h{\Phi}_i(z)\op{\p}_z\h{G}(z,z',k^2)
\op{\p}_{z'}\h{\Phi}_f^{\star}(z')\right)
\p_{z'}A_2^{\star}(z')z'^{-3}e^{-\chi(z')}\,.
\end{split}
\end{align}
The $u$-channel amplitude is related to the $s$-channel amplitude
according to the crossing symmetry by interchanging $\mu\lr\nu$,
$q_1\lr -q_2$,\ $\epsilon\lr \epsilon^{\star}$ and
$A_1\lr A_2^{\star}$. Hence it reads
\begin{align}
\begin{split}
T_{\mu\nu}^u &= (2\pi)^4\delta^{(4)}(p_1+q_1-p_2-q_2)\,e^2
\times\biggl( (p_2+k')_{\mu}(p_1+k')_{\nu}\,\F_1(q_2^2,q_1^2,u) \\
&\quad
+ \frac{q_{1\mu}(p_1+k')_{\nu}}{q_1^2}\,\F_2(q_2^2,q_1^2,u)
- \frac{(p_2+k')_{\mu}q_{2\nu}}{q_2^2}\,\F_3(q_2^2,q_1^2,u)
- \frac{q_{1\mu}q_{2\nu}}{q_1^2q_2^2}\,\F_4(q_2^2,q_1^2,u) \biggr)\,,
\end{split}
\end{align}
with $k' = p_1-q_2 = p_2-q_1$ and $u=k'^2$.
Integrating by parts the partial derivative of $A_1$,
\begin{align}
\F_3(q_1^2,q_2^2,k^2) &= -\iint dz\,dz'\,\,z'^{-3}e^{-\chi(z')}
A^{\star}_2(z')\h{\Phi}_f^{\star}(z')A_1(z)\p_{z}
\left(z^{-3}e^{-\chi(z)}\left(\h{\Phi}_i(z)\op{\p}_z\h{G}(z,z',k^2)\right)
\right)\,,
\end{align}
and using equations \eqref{scalar} and \eqref{green}, one gets
\begin{align}
\nonumber
&\p_{z}\left(z^{-3}e^{-\chi(z)}\left(\h{\Phi}_i(z)\op{\p}_z\h{G}(z,z',k^2)
\right)\right) = \\
\nonumber
&\h{\Phi}_i(z)\p_{z}\left(z^{-3}e^{-\chi(z)}\p_z\h{G}(z,z',k^2)\right)
-\p_{z}\left(z^{-3}e^{-\chi(z)}\p_z\h{\Phi}_i(z)\right)\h{G}(z,z',k^2) = \\
\label{rel}
& z^{-3}e^{-\chi}\left(k^2-p_1^2\right)\h{\Phi}_i(z)\h{G}(z,z',k^2)
+ \delta(z-z')\h{\Phi}_i(z) \,.
\end{align}
Hence
\begin{align}
\F_3(q_1^2,q_2^2,k^2) = (p_1^2-k^2)\F_1(q_1^2,q_2^2,k^2) -
\int dz\, z^{-3}e^{-\chi(z)}A_1(z)A_2^{\star}(z)
\h{\Phi}_i(z)\h{\Phi}_f^{\star}(z)\,.
\end{align}
Similarly,
\begin{align}
\F_2(q_1^2,q_2^2,k^2) &=
-\iint dz\,dz'\,z^{-3}e^{-\chi(z)}A_1(z)\h{\Phi}_i(z)A_2^{\star}(z')
\p_{z'}\left(z'^{-3}e^{-\chi(z')}\left(\h{G}(z,z',k^2)
\op{\p}_{z'}\h{\Phi}_f^{\star}(z')\right)\right)\,, \nonumber\\
&= (k^2-p_2^2)\F_1(q_1^2,q_2^2,k^2) + \int dz\, z^{-3}e^{-\chi(z)}
A_1(z)A_2^{\star}(z)\h{\Phi}_i(z)\h{\Phi}_f^{\star}(z)\,.
\end{align}
The four-point interaction amplitude reads
\begin{align}
\label{CONTACT}
\epsilon^{\mu}T^c_{\mu\nu}\epsilon^{\star\nu} &=
-2(ie)^2\int d^4x\,dz\,\sqrt{-g}e^{-\chi}
\h{\Phi}_i(x,z)g^{mn}A_m(x,z)A_n^{\star}(x,z)\h{\Phi}_f^{\star}(x,z)\,,
\nonumber\\
\begin{split}
T^c_{\mu\nu} &= 2 e^2(2\pi)^4\delta^{(4)}(p_1+q_1-p_2-q_2) \times \\
&\quad\int dz z^{-3}e^{-\chi}\,
\h{\Phi}_i(z)\left(g_{\mu\nu}A_1(z)A_2^{\star}(z) +
\frac{q_{1\mu}q_{2\nu}}{q_1^2q_2^2}
\p_zA_1(z)\p_zA_2^{\star}(z)\right)\h{\Phi}_f^{\star}(z) \,.
\end{split}
\end{align}
Let
\begin{align}
\C_1(q_1^2,q_2^2) &= \C_1(q_2^2,q_1^2) = \int dz\,z^{-3}e^{-\chi(z)}
\,A_1(z)\,A_2^{\star}(z)\,\h{\Phi}_i(z)\,\h{\Phi}_f^{\star}(z) \,, \\
\C_0(q_1^2,q_2^2) &= \C_0(q_2^2,q_1^2) =
\int dz\, z^{-3}e^{-\chi(z)}\,\p_zA_1(z)\,\p_zA_2^{\star}(z)
\,\h{\Phi}_i(z)\,\h{\Phi}_f^{\star}(z) \,.
\end{align}
Setting $p=p_1+p_2$, the total amplitude can be written as
\begin{align*}
T_{\mu\nu} &= e^2(2\pi)^4\delta^{(4)}(p_1+q_1-p_2-q_2) \times \biggl(\\
&\quad
\left(\F_1\left(q_1^2,q_2^2,s\right) +
\F_1\left(q_2^2,q_1^2,u\right)\right)\times
\biggl(\left(p_{\mu}p_{\nu} - \frac{p\cdot q_2}{q_2^2}p_{\mu}q_{2\nu}
- \frac{p\cdot q_1}{q_1^2}p_{\nu}q_{1\mu}\right) + \\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
\left(q_{1\nu}q_{2\mu} - \frac{q_1\cdot q_2}{q_1^2}q_{1\mu}q_{1\nu}
- \frac{q_1\cdot q_2}{q_2^2}q_{2\mu}q_{2\nu}\right)\biggr) \\
&+ \left(\F_1\left(q_1^2,q_2^2,s\right)-\F_1\left(q_2^2,q_1^2,u\right)\right)
\times\biggl(
\left(p_{\mu}q_{1\nu} - \frac{q_1\cdot q_2}{q_2^2}p_{\mu}q_{2\nu}
- \frac{p\cdot q_1}{q_1^2} q_{1\mu}q_{1\nu}\right) + \\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
\left(p_{\nu}q_{2\mu} - \frac{q_1\cdot q_2}{q_1^2}p_{\nu}q_{1\mu} -
\frac{p\cdot q_2}{q_2^2}q_{2\mu}q_{2\nu}\right)\biggr) \\
&\ + 2\C_1\left(q_1^2,q_2^2\right)\left(g_{\mu\nu}
- \frac{q_{1\mu}q_{1\nu}}{q_1^2}
- \frac{q_{2\mu}q_{2\nu}}{q_2^2}\right)
+ \left(2\C_0(q_1^2,q_2^2) - \F_4\left(q_1^2,q_2^2,s\right) -
\F_4\left(q_2^2,q_1^2,u\right)\right)\frac{q_{1\mu}q_{2\nu}}{q_1^2q_2^2}
\biggr) \,.
\end{align*}
Comparing with the gauge-invariant tensor basis \eqref{Vmunu}, gauge
invariance holds true if, and only if,
\begin{align}
\begin{split}
2\C_0(q_1^2,q_2^2) &= \F_4\left(q_1^2,q_2^2,s\right)
+\F_4\left(q_2^2,q_1^2,u\right) \\
&+ \left(\F_1\left(q_1^2,q_2^2,s\right)+\F_1\left(q_2^2,q_1^2,u\right)\right)
\left((p\cdot q_1)(p\cdot q_2)+(q_1\cdot q_2)^2\right) \\
&+ \left(\F_1\left(q_1^2,q_2^2,s\right)-\F_1\left(q_2^2,q_1^2,u\right)\right)
\left(p\cdot q_1 + p\cdot q_2\right)(q_1\cdot q_2) \\
&+ 2\C_1(q_1^2,q_2^2)(q_1\cdot q_2)\,.
\end{split}
\end{align}
Expanding the bidirectional derivatives, the form factor $\F_4$ reads
\begin{align}
\begin{split}
\F_4 &= \iint dz\,dz'\,z^{-3}e^{-\chi(z)}\,\p_zA_1(z)\times \\
&\qquad\qquad\biggl(
\left(\h{\Phi}_i(z)\,\p_z\h{G} - \h{G}\,\p_z\h{\Phi}_i(z)\right)
\p_{z'}\h{\Phi}_f^{\star}(z')
- \p_{z'}\left(\h{\Phi}_i(z)\,\p_z\h{G} - \p_z\h{\Phi}_i(z)\,\h{G}\right)
\h{\Phi}_f^{\star}(z') \biggr) \\
&\qquad\quad\times \p_{z'}A_2^{\star}(z')z'^{-3}e^{-\chi(z')} \,.
\end{split}
\end{align}
We first integrate by parts over $\p_zA_1(z)$ and use equation
\eqref{rel} in the $s$-channel,
\begin{align}
\nonumber
\begin{split}
\F_4(s) &= -\iint dz\,dz'\,A_1(z)\times \\
&\qquad\qquad\biggl(
\left(z^{-3}e^{-\chi(z)}\left(s-p_1^2\right)\h{\Phi}_i(z)\h{G}(z,z',s)
+ \delta(z-z')\,\h{\Phi}_i(z)\right)\p_{z'}\h{\Phi}_f^{\star}(z') \\
&\qquad\qquad
-\p_{z'}\left(z^{-3}e^{-\chi}\left(s-p_1^2\right)\h{\Phi}_i(z)\h{G}(z,z',s)
+ \delta(z-z')\,\h{\Phi}_i(z)\right)\h{\Phi}_f^{\star}(z') \biggr) \\
&\qquad\quad\times \p_{z'}A_2^{\star}(z')z'^{-3}e^{-\chi(z')} \,,
\end{split}
\\
\begin{split}
&= (p_1^2-s)\iint dz\,dz'\,z^{-3}e^{-\chi(z)}\,A_1(z)\,\h{\Phi}_i(z)\times \\
& \qquad\qquad\qquad\quad\left(\h{G}\,\p_{z'}\h{\Phi}_f^{\star}(z') -
\h{\Phi}_f^{\star}(z')\,\p_{z'}\h{G}\right)
\p_{z'}A_2^{\star}(z')\,z'^{-3}e^{-\chi(z')} \\
&\quad-
\int dz\,z^{-3}e^{-\chi(z)}\,A_1(z)\,\h{\Phi}_i(z)\,\p_z\h{\Phi}_f^{\star}(z)
\,\p_zA_2^{\star}(z) \\
&\quad- \int dz\,A_1(z)\,\h{\Phi}_i(z)\,\p_z
\left(\h{\Phi}_f^{\star}(z)\,\p_zA_2^{\star}(z)\,z^{-3}e^{-\chi(z)}\right)\,.
\end{split}
\end{align}
We now integrate the first term by parts over $\p_{z'}A^{\star}_2(z')$,
use again equation \eqref{rel}, and integrate by parts the third term,
\begin{align}
\begin{split}
\F_4(s) &= (p_1^2-s)
\iint dz\,dz'\,z^{-3}e^{-\chi(z)}\,A_1(z)\,\h{\Phi}_i(z)\times
\\ & \qquad\qquad\qquad\quad
\left(z'^{-3}e^{-\chi(z')}\left(s-p_2^2\right)\h{G}(z,z',s)
+ \delta(z-z')\right)\h{\Phi}_f^{\star}(z')\,A_2^{\star}(z') \\
&\quad-
\int dz\,z^{-3}e^{-\chi(z)}\,A_1(z)\,\h{\Phi}_i(z)\,\p_z\h{\Phi}_f^{\star}(z)
\,\p_zA_2^{\star}(z) \\
&\quad+ \int dz\,z^{-3}e^{-\chi(z)}\,\p_z\left(A_1(z)\,\h{\Phi}_i(z)\right)
\,\h{\Phi}_f^{\star}(z)\,\p_zA_2^{\star}(z) \,.
\end{split}
\end{align}
Hence
\begin{align}
\label{T4}
\begin{split}
\F_4(s) &= -(p_1^2-s)(p_2^2-s)\F_1(s) + (p_1^2-s)\C_1 +\C_0 \\
&\quad- \int dz\,z^{-3}e^{-\chi(z)}\,A_1(z)
\left(\h{\Phi}_i(z)\,\op{\p}_z\h{\Phi}_f^{\star}(z)\right)
\,\p_zA_2^{\star}(z)\,.
\end{split}
\end{align}
A similar identity holds in the $u$-channel. Noting that
\begin{gather}
\nonumber
\begin{split}
(p\cdot q_1)(p\cdot q_2)+(q_1\cdot q_2)^2 \pm
(p\cdot q_1 + p\cdot q_2)(q_1\cdot q_2) =
(p\pm q_1)\cdot q_2 \times (p\pm q_2)\cdot q_1 \,,
\end{split}
\\
\begin{split}
(p+q_2)\cdot q_1\times(p+q_1)\cdot q_2 = (p_1^2-s)(p_2^2-s) \,, \\
(p-q_2)\cdot q_1\times(p-q_1)\cdot q_2 = (p_1^2-u)(p_2^2-u) \,,
\end{split}
\end{gather}
gauge invariance is recovered when $p_1^2=p_2^2$. Indeed, then
$\h{\Phi}_i(z)=\h{\Phi}_f(z)$ since they satisfy the same equation, so the
last term in eq.\,\eqref{T4} vanishes.
To summarize, the on-shell Compton amplitude reads, with
$p_1^2=p_2^2=-m^2$,
\begin{align}
\label{SCA}
\begin{split}
T^{\mu\nu} &= e^2\left(2\C_1\,\V_1^{\mu\nu}
+ \C_+\left(\V_2^{\mu\nu} + \V_3^{\mu\nu}\right)
+ \C_-\left(\V_4^{\mu\nu} + \V_5^{\mu\nu}\right)\right) \,,
\end{split}
\end{align}
where the tensors $\V_i^{\mu\nu},\ i=1,\cdots 5$, are defined in
eqs.~\eqref{Vmunu} and
\begin{align}
C_{\pm}(m^2,q_1^2,q_2^2,s,u) = \F_1(m^2,q_1^2,q_2^2,s) \pm
\F_1(m^2,q_2^2,q_1^2,u) \,.
\end{align}
The $\delta$ factor expressing energy-momentum conservation is
implicitly understood from now on in all formulas for the Compton
amplitude. The form factors $\F_1(m^2,q_1^2,q_2^2,k^2)$ and
$\C_1(m^2,q_1^2,q_2^2)$ are defined by
\begin{align}
\label{SVCF}
\begin{split}
\F_1\left(m^2,q^2_1,q_2^2,k^2\right) &=
\iint dz_1\,dz_2\,\,z_1^{-3}e^{-\chi(z_1)}\,A_1(z_1)\,\h{\Phi}_i(z_1)\,
\h{G}(z_1,z_2,k^2)
\,\h{\Phi}_f^{\star}(z_2)\,A_2^{\star}(z_2)\,z_2^{-3}e^{-\chi(z_2)}\,, \\
\C_1(m^2,q_1^2,q_2^2) &= \int dz\,z^{-3}e^{-\chi(z)}
\,A_1(z)\,A_2^{\star}(z)\,\h{\Phi}_i(z)\,\h{\Phi}_f^{\star}(z) \,.
\end{split}
\end{align}
These formulas are valid for any dilaton background $\chi(z)$ on the
5D AdS space which yields a well-defined inner product on the vector
space of solutions of the classical field equations. If $\chi\equiv
0$, our formulas coincide with the Compton amplitude off a dilaton
calculated in \cite{GX09}, once we identify the normalized functions
$\h{\Phi}_i$ and $\h{\Phi}_f$ with the Bessel solutions of the
hard-wall model. We have shown quite generally that the
gauge-invariant structure of the Compton amplitude is similar for
hard-wall and soft-wall models. Note that we only get three
independent Compton form factors out of a possible five. These generic
properties do not depend upon the special way of breaking the
conformal invariance for extending the AdS/CFT correspondence to QCD,
nor upon detailed relations between Bessel functions or other special
functions. It depends only upon the general structure of the classical
equations satisfied by the bulk scalar fields and their Green
functions in AdS space with a dilaton background.
Moreover we have shown explicitly that the off-shell Compton
amplitudes ($p_1^2 \ne p_2^2$) calculated in these simplest AdS/QCD
models are not gauge-invariant. This result is also not obvious and is
in fact very specific to single scalar intermediate states. For
example, had we allowed for non-minimal couplings with only single
vector intermediate states, we would get gauge-invariant off-shell
Compton amplitudes. One should perhaps emphasize that the non-gauge
invariance of the off-shell four-point Compton amplitude has nothing
to do with the (trivial) non-gauge invariance of the three-point
function when the conformal dimensions of the initial and final scalar
states are different. In our case, the initial, final, and
intermediate scalar states are solutions of the same classical field
equation in AdS space and have the same conformal dimension. The
non-gauge invariance, when the initial and final scalar states are not
the same mass eigenstates in Minkowski space, is due to the particular
form of the contact term \eqref{CONTACT}.
\section{Deeply inelastic scattering}
\label{DIS}
The Compton amplitude \eqref{SCA} has a rather simple structure which
it is enlightening to unravel. Expanding the Green function over the
orthonormal eigenstates, we can write the form factor $\F_1$ in the
doubly spacelike region as
\begin{align}
\label{IF}
\F_1\left(m^2,q^2_1,q_2^2,k^2\right) &= -\sum_{n=0}^{\infty}
\frac{\Gamma(m^2,m_n^2,q_1^2)\,\Gamma^*(m^2,m_n^2,q_2^2)}
{k^2+m^2_n-i\epsilon} \,, \\
\label{IV}
\Gamma(m^2,m_n^2,Q^2) &= Q\int dz\,z^{-2}e^{-\chi(z)}
K_1(Qz)\,\h{\Phi}(z)\,\h{\Phi}^*_n(z) \,.
\end{align}
It is immediate to show from the AdS/QCD dictionary that the vertex
function $\Gamma$ is just the unique form factor which parametrizes
the most general matrix element of the conserved electromagnetic
current between two (pseudo)scalar states,
\begin{align}
\label{FF}
\langle p_2|J^{\mu}(0)|p_1\rangle = \Gamma(p_1^2,p_2^2,k^2)
\left(p^{\mu} - \frac{p_2^2-p_1^2}{k^2}\,k^{\mu}\right)
\,,\qquad p =p_1 + p_2\,,\quad k = p_2-p_1\,.
\end{align}
The electromagnetic form factor of the spinless target is defined as
the elastic limit of $\Gamma$,
\begin{align}
\label{FFEM}
F_{\gamma}(Q^2) = \Gamma(m^2,m^2,Q^2) = \C_1(m^2,Q^2,0)
= Q\int dz\,z^{-2}e^{-\chi(z)}
K_1(Qz)\,\left|\h{\Phi}_m(z)\right|^2 \,.
\end{align}
Taking into account the tensorial identities
\begin{align}
\V_2^{\mu\nu} + \V_3^{\mu\nu} \pm \V_4^{\mu\nu} \pm \V_5^{\mu\nu}
= \left((p\pm q_2)^{\mu} - q_1^{\mu}\frac{(p\pm q_2)\cdot q_1}{q_1^2}\right)
\left((p\pm q_1)^{\nu} - q_2^{\nu}\frac{(p\pm q_1)\cdot q_2}{q_2^2}\right) \,,
\end{align}
the amplitude \eqref{SCA} can be written in a form which exhibits the
tensorial structure used in \cite{GX09}
\begin{align}
\label{SCA2}
\begin{split}
T^{\mu\nu} &= e^2\biggl\{2\C_1(m^2,q_1^2,q_2^2)\,\V_1^{\mu\nu} \\
&\qquad\quad-
\left(2p_1^{\mu} - q_1^{\mu}\frac{2p_1\cdot q_1}{q_1^2}\right)
\left(2p_2^{\nu} - q_2^{\nu}\frac{2p_2\cdot q_2}{q_2^2}\right)
\sum_{n=0}^{\infty}
\frac{\Gamma(m^2,m_n^2,q_1^2)\,\Gamma^*(m^2,m_n^2,q_2^2)}{s+m^2_n-i\epsilon}
\\ &\qquad\quad-
\left(2p_2^{\mu} - q_1^{\mu}\frac{2p_2\cdot q_1}{q_1^2}\right)
\left(2p_1^{\nu} - q_2^{\nu}\frac{2p_1\cdot q_2}{q_2^2}\right)
\sum_{n=0}^{\infty}
\frac{\Gamma(m^2,m_n^2,q_2^2)\,\Gamma^*(m^2,m_n^2,q_1^2)}{u+m^2_n-i\epsilon}
\biggr\} \,.
\end{split}
\end{align}
As a by-product, the absorptive part of the forward Compton scattering
amplitude reads
\begin{align}
\label{IM}
\nonumber
&\text{Im}\,T^{\mu\nu}(q^2,s) = e^2
\left(p^{\mu} + \frac{1}{x}q^{\mu}\right)
\left(p^{\nu} + \frac{1}{x}q^{\nu}\right)
\sum_{n=0}^{\infty} \delta(s+m_n^2)\left|\Gamma(m^2,m_n^2,q^2)\right|^2 \,,\\
&\qquad\qquad\quad\ \approx
e^2\left.\left(\px{m_n^2}{n}\right)^{-1}\right|_{m_n^2=-s}
\left|\Gamma(m^2,-s,q^2)\right|^2
\left(p^{\mu} + \frac{1}{x}q^{\mu}\right)
\left(p^{\nu} + \frac{1}{x}q^{\nu}\right) \,,\\
\nonumber
&\qquad\qquad
q_1=q_2=q\,,\qquad p=2p_1=2p_2\,,\qquad x = -\frac{q^2}{p\cdot q} \,.
\end{align}
and do yield, in the hard-wall model, the same structure functions
$F_1=0$ and $F_2$ as found in \cite{PS03}, in the large-$x$ region and
in the one-particle approximation for intermediate states.
More generally, eq.\,\eqref{IM} relates the scaling properties of the
vertex function $\Gamma$ and of the structure function $F_2$ in a
generic soft-wall model. Indeed the function $Qz\,K_1(Qz)$ decreases
monotonically from 1 to 0 and is exponentially small at large
$Qz$. Hence the $z$-dependence of the electromagnetic current can be
roughly approximated as a step function of width $\O(1/Q)$ and
\begin{align}
\label{overlap}
\Gamma(m^2,m_n^2,Q^2) \approx \int_0^{1/Q}\frac{dz}{z^3}\,e^{-\chi(z)}
\,\h{\Phi}(z)\,\h{\Phi}^*_n(z) \,.
\end{align}
One can differentiate three kinematical regimes for the evaluation of
the overlap integral \eqref{overlap}:
\begin{itemize}
\item $Q^2\gg m^2,\ Q^2\gtrsim m_n^2.\quad$ For $z\lesssim 1/Q$ and
$Q$ large enough at fixed $\Q^2/m_n^2$, $\h{\Phi}(z)$ has the
asymptotic behavior $\h{\Phi}(z)\sim z^{\Delta}$, where the
conformal dimension $\Delta$ is the same as in the hard-wall model
as long as $\chi(z)\rightarrow 0$ when $z\rightarrow 0$. On the
other hand, since $n$ is a highly excited state, it is legitimate to
use a WKB approximation for $\h{\Phi}_n(z)$ \cite{PRSW},
\begin{align}
\Gamma(m^2,m_n^2,Q^2) \propto C(m_n)
\left(\frac{1}{Q}\right)^{\Delta}\,F\left(\frac{Q}{m_n}\right) \,.
\end{align}
The squared normalization constant $C^2(m_n)$ of $\h{\Phi}_n(z)$ and
the semiclassical density of states $\frac{\p m_n^2}{\p n}$ have the
same dependence upon $m_n$. Therefore the structure function
$F_2(Q^2,x)$ has the power-law scaling,
\begin{align}
\frac{x}{Q^2}F_2(Q^2,x) \propto
\left(\frac{1}{Q^2}\right)^{\Delta} F^2(x)
\,,\qquad x = \frac{Q^2}{Q^2-s} \,.
\end{align}
\item $Q^2\gg m^2,\ Q^2\gg m_n^2.\quad$ Then both $\h{\Phi}(z)$ and
$\h{\Phi}_n(z)$ behave as $z^{\Delta}$ for $z\lesssim 1/Q$. It
follows that the electromagnetic form factor has the power law
scaling \cite{BRO07},
\begin{align}
F_{\gamma}(Q^2) \propto \left(\frac{1}{Q^2}\right)^{\Delta-1} \,.
\end{align}
The identity of the asymptotic scaling behavior for the
electromagnetic form factor and for the structure functions is a
generic property of the AdS/QCD models that we consider. Such a
property, which does not even depend upon whatever value of $\Delta$ is
picked, is certainly difficult to reconcile with a partonic picture.
\item $Q^2\rightarrow 0.\quad$ In that limit the overlap integral
reduces to the scalar product of $\h{\Phi}$ and $\h{\Phi}_n$. Since
the asymptotic states $\Phi_{\text{in}}(x,z)$ and
$\Phi_{\text{out}}(x,z)$ must be eigenstates of the operator
\eqref{H} with the same mass eigenvalue, all inelastic channels
decouple when $Q^2=0$ and only the elastic channel
remains opened. We shall work out some of the consequences in the
next sections.
However we can already observe that there is a violation of elastic
unitarity in the Compton amplitude \eqref{SCA2} which is inherent to
the $N_c\rightarrow\infty$ approximation involved in the AdS/QCD
recipes. Indeed the total elastic Compton cross-section is of order
$e^4$. Hence the imaginary part of the forward amplitude must vanish
at order $e^2$ in the elastic limit to comply with the optical
theorem. An absorptive part of the elastic amplitude can be
generated only by loop effects which are at present beyond the reach
of the AdS/QCD correspondence.
\end{itemize}
\section{Virtual Compton Scattering}
\label{DVCS}
When the outgoing photon is real, $q_2^2=0$, a mere inspection of
eq.\,\eqref{Vmunu} shows that, in order to cancel the poles in
$q_2^2$, we must have the following relations between the form factors,
\begin{gather}
\label{DVCSR}
\begin{split}
V_1 + (q_1\cdot q_2)V_3 + (p\cdot q_2)V_5 = 0 \,, \\
(p\cdot q_2)V_2 + (q_1\cdot q_2) V_4 = 0\,.
\end{split}
\end{gather}
Hence only three Compton form factors remain independent as already
observed in section \ref{COMPTON}. However eqs.\,\eqref{DVCSR} are not
manifestly satisfied by eqs.\,\eqref{SCA} and \eqref{SVCF}. We should
have
\begin{gather}
\label{DVCSR1}
2\C_1 + (q_1\cdot q_2)\,\C_+ + (p\cdot q_2)\,\C_- = 0 \,, \\
\label{DVCSR2}
(p\cdot q_2)\,\C_+ + (q_1\cdot q_2)\,\C_- = 0\,.
\end{gather}
The second relation \eqref{DVCSR2} reads explicitly, with the
notations \eqref{notations},
\begin{align}
\begin{split}
&(p+q_1)\cdot q_2\times \F_1(m^2,q_1^2,0,s)
+ (p-q_1)\cdot q_2\times \F_1(m^2,0,q_1^2,u) = \\
&\iint dz_1\,dz_2\,\,z_1^{-3}e^{-\chi(z_1)}\,\h{\Phi}_i(z_1)\,\times \\
&\qquad\left(
-(p_1^2-s) A_1(z_1)\,\h{G}(z_1,z_2,s)\,A_2^{\star}(z_2) +
(p_2^2-u) A_2^{\star}(z_1)\,\h{G}(z_1,z_2,u)\,A_1(z_2)
\right) \\
&\qquad \times\h{\Phi}_f^{\star}(z_2)\,z_2^{-3}e^{-\chi(z_2)}
= 0 \,.
\end{split}
\end{align}
We now use the completeness relation \eqref{expansion} satisfied by
the Green function $\h{G}$, and take into account the fact that for
the electromagnetic field,
\begin{align}
\lim_{q_2^2\rightarrow 0} A_2(z) = 1 \,,\qquad
\lim_{q_2^2\rightarrow 0} \p_{z}A_2(z) = 0 \,.
\end{align}
We integrate each term over $z_2$ and $z_1$ respectively,
\begin{align}
\begin{split}
\int dz_2\,z_2^{-3}e^{-\chi(z_2)}\,
\h{\Phi}_f^{\star}(z_2)\,\h{\Phi}_n(z_2)
&= C^{\star}(p_2^2,m_n^2) \,, \\
\int dz_1\,z_1^{-3}e^{-\chi(z_1)}\,
\h{\Phi}_i(z_1)\,\h{\Phi}_n^{\star}(z_1)
&= C(p_1^2,m_n^2) \,.
\end{split}
\end{align}
Hence eq.\,\eqref{DVCSR2} reads
\begin{align}
\begin{split}
&-(p_1^2-s)\sum_n\frac{C^{\star}(p_2^2,m_n^2)}{m_n^2+s-i\epsilon}
\int dz\,z^{-3}e^{-\chi(z)}\,\h{\Phi}_i(z)\,
\h{\Phi}^{\star}_n(z)\,A_1(z) \\
&+
(p_2^2-u)\sum_n\frac{C(p_1^2,m_n^2)}{m_n^2+u-i\epsilon}
\int dz\,z^{-3}e^{-\chi(z)}\,\h{\Phi}_f^{\star}(z)\,
\h{\Phi}_n(z)\,A_1(z) = 0 \,.
\end{split}
\end{align}
Using the orthogonality relations, the identity \eqref{DVCSR2} holds
true exactly only if the virtual Compton scattering is on-shell,
\begin{align}
p_1^2 = p_2^2 = -m_{n_0}^2\,,\quad\text{for some }n_0\,.
\end{align}
By the same token eq.\,\eqref{DVCSR1} reduces to the definition of
$\C_1$ in \eqref{SVCF},
\begin{align}
\begin{split}
&2\C_1(m^2,q_1^2,0) + (p+q_1)\cdot q_2\times \F_1(m^2,q_1^2,0,s)
- (p-q_1)\cdot q_2\times \F_1(m^2,0,q_1^2,u) = \\
&2\C_1(m^2,q_1^2,0)
- 2\int dz\,z^{-3}e^{-\chi(z)}\left|\h{\Phi}_m(z)\right|^2\,A_1(z)
= 0 \,.
\end{split}
\end{align}
We can solve for $\C_{\pm}$ in terms of $\C_1$:
\begin{gather}
\begin{split}
\C_+ = -\frac{2(q_1\cdot q_2)}{(q_1\cdot q_2)^2 - (p\cdot q_2)^2}\,\C_1\,,
\qquad
\C_- = \frac{2(p\cdot q_2)}{(q_1\cdot q_2)^2 - (p\cdot q_2)^2}\,\C_1\,, \\
(q_1\cdot q_2)^2 - (p\cdot q_2)^2 = (q_1-p)\cdot q_2 \times
(q_1+p)\cdot q_2 = (m^2+s)(m^2+u) \,,
\end{split}
\end{gather}
The VCS amplitude has no absorptive part since $(m^2+s)(m^2+u)$ vanish
only when $q_2=0$ owing to a non-vanishing mass $m$. Therefore the exact
VCS amplitude with $q_2^2=0$ can be written as
\begin{align}
\label{TDVCS}
\begin{split}
T^{VCS}_{\mu\nu} &= e^2\,\C_1(m^2,q_1^2,0)\biggl(2\V_{1,\mu\nu} -
\frac{2m^2+s+u}{(m^2+s)(m^2+u)}\left(\V_{2,\mu\nu}+\V_{3,\mu\nu}\right) \\
&\qquad\qquad\qquad\qquad+
\frac{s-u}{(m^2+s)(m^2+u)}
\left(\V_{4,\mu\nu}+\V_{5,\mu\nu}\right)\biggr) \,.
\end{split}
\end{align}
The unique Compton form factor reads
\begin{align}
\label{FF1}
\C_1(m^2,q_1^2,0) &=
\int_0^{\infty} dz\,z^{-3}e^{-\chi(z)}
\left|\h{\Phi}_m(z)\right|^2\,A_1(z) \,,
\end{align}
and is nothing but the electromagnetic form factor of the spinless
target. Note that this formula holds in principle (for on-shell
amplitudes) for any $q_1^2$, spacelike or timelike, if one can make an
analytic continuation in the photon momentum.
The tensorial structure of the amplitude \eqref{TDVCS} is identical to
point-like scalar electrodynamics in the tree level approximation
except for the electromagnetic form factor which encodes all the
internal structure of a spinless particle in this kind of AdS/QCD
models. The threshold theorem imposes that
\begin{align}
\label{NORM}
\C_1(m^2,0,0) &= \int_0^{\infty} dz\,z^{-3}e^{-\chi(z)}
\left|\h{\Phi}_m(z)\right|^2 = 1 \,.
\end{align}
The normalization is completely fixed by electromagnetic gauge
invariance and the Hilbert space structure of the classical solutions
in AdS space with appropriate dilaton background.
\section{Bjorken scaling of the DVCS amplitude}
\label{GPD}
It is instructive to understand the consequences of the simplistic form
of the DVCS amplitude \eqref{TDVCS} for a would-be dual picture in
terms of partonic constituents in these kinds of AdS/QCD models.
The Bjorken scaling of the virtual Compton form factors on a
(pseudo)scalar target is usually analyzed in terms of independent
gauge-invariant tensors expressed in the momenta $q=(q_1+q_2)/2$,
$p=p_1+p_2$ and $\Delta=p_2-p_1=q_1-q_2$ with $p_1^2=p_2^2=-M^2$. The
four independent scalar invariant $s$, $u$, $q_1^2$ and $q_2^2$ are
ordinarily traded for $Q^2$, $\Delta^2$, and the scaling variables
$\xi$ and $\eta$ defined by
\begin{align}
\label{BJ}
Q^2 = q^2 \,,\qquad
\xi = -\frac{Q^2}{p\cdot q}\,,\qquad
\eta=-\frac{\Delta\cdot q}{p\cdot q} \,.
\end{align}
The large $Q^2$ expansion of the virtual Compton scattering amplitude
on any target can be described up to twist-three, and with $q_1^2$ and
$q_2^2$ arbitrary, by the tensorial structure \cite{BMKS}
\begin{align}
\label{TWIST3}
\begin{split}
T^{TW3}_{\mu\nu}(q,p,\Delta) &= -\P_{\mu\sigma}g^{\sigma\tau}\P_{\tau\nu}
\frac{q\cdot W_1}{p\cdot q}
+ \left(\P_{\mu\sigma}p^{\sigma}\P_{\rho\nu}
+ \P_{\mu\rho}p^{\sigma}\P_{\sigma\nu}\right)
\frac{W^{\rho}_2}{p\cdot q} \\
&\quad- \P_{\mu\sigma}i\epsilon^{\sigma\tau q\rho}\P_{\tau\nu}
\frac{A_{1\rho}}{p\cdot q}\,.
\end{split}
\end{align}
where current conservation is ensured by means of the projector
\begin{align}
\P_{\mu\nu} = g_{\mu\nu} - \frac{q_{2\mu}q_{1\nu}}{q_1\cdot q_2} \,,
\end{align}
and the transverse component of the momentum transfer is defined by
\begin{align}
\Delta^{\perp}_{\mu} = \Delta_{\mu} + \eta\,p_{\mu} \,.
\end{align}
The vector $W_{2\rho}$ depends on $W_{1\rho}$ and $A_{1\rho}$ by the
relation ($\epsilon_{0123}=1$),
\begin{align}
W_{2\rho} &= \xi W_{1\rho} - \frac{\xi}{2}\frac{q\cdot W_1}{p\cdot q} p_{\rho}
+ \frac{i}{2}\frac{\epsilon_{\rho\sigma\Delta q}}{p\cdot q} A_{1\sigma} \,.
\end{align}
For a spinless target, the vectors $W_{1\rho}$ and $A_{1\rho}$ are
defined in terms of three generalized form factors
$\H(\xi,\eta,\Delta^2,Q^2)$, $\H_3(\xi,\eta,\Delta^2,Q^2)$ and
$\w{\H}_3(\xi,\eta,\Delta^2,Q^2)$ by the relations
\begin{align}
\begin{split}
W_1 &= \H\,p + \H_3\,\Delta_{\perp}\,,\quad
A_{1\rho} = \frac{i\epsilon_{\rho\Delta p q}}{p\cdot q}\w{H}_3\,.
\end{split}
\end{align}
When $q_2^2=0$, the VCS relations \eqref{DVCSR} are satisfied and the
generalized form factors $\H$'s are related to the independent form
factors $V_1$, $V_3$ and $V_4$ in \eqref{Vmu} as follows,
\begin{align}
\begin{split}
V_1 &= -\H \,, \\
V_3 &= \frac{1}{\left(2-\frac{\eta}{\xi}\right)}\frac{1}{Q^2}\left(
\left(1-\frac{1}{2-\frac{\eta}{\xi}}\right)\H
- \xi\w{H}_3\left(2-3\frac{\eta}{\xi}\right)\right) \,, \\
V_4 &= \frac{-\xi}{2-\frac{\eta}{\xi}}\frac{1}{Q^2}
\left(\H + 2\eta\H_3
+ \xi\w{\H}_3\left(2-\frac{\eta}{\xi}\right)^2\right) \,\\
Q^2 &= \frac{q_1^2}{2}\left(1-\frac{\Delta^2}{2q_1^2}\right) \,.
\end{split}
\end{align}
In perturbative QCD, these form factors can in principle be related,
through factorization, to generalized parton distributions (GPDs).
However, the absence of an absorptive part in the DVCS amplitude
\eqref{TDVCS} is difficult to accommodate with a partonic
interpretation which is based on the convolution of real GPDs with
coefficient functions which contain both a real and an imaginary part.
Specializing to the Bjorken limit of the DVCS amplitude \eqref{TDVCS},
\begin{align}
\Delta^2 = 0\,,\qquad \xi=\eta=\frac{x_B}{2-x_B}\,,\qquad
x_B = -\frac{q_1^2}{2p_1\cdot q_1} \,,
\end{align}
one gets for the generalized form factors
\begin{align}
\begin{split}
\H &= -2\,\C_1(m^2,2Q^2,0) \,, \\
\H_3 &= -\w{\H}_3 = \frac{\xi}{1-\xi^2}\,\H
= \frac{x_B(2-x_B)}{4(1-x_B)}\,\H\,.
\end{split}
\end{align}
Therefore the asymptotic behavior in $Q^2$ of the DVCS cross-section
integrated over $t$ and over the azimuthal angle is governed by the
power-law behavior of the electromagnetic form factor. A power-law
behavior in accordance with the dimensional counting rules, e.g. a
scaling dimension $\Delta=2$ for the pion, cannot be consistent with a
partonic interpretation of the DVCS amplitude for spinless hadronic
targets.
\section{Polarizabilities}
\label{POL}
The structure of the VCS amplitude \eqref{TDVCS} and the threshold
theorem, eq.\,\eqref{NORM}, have a still more drastic
consequence. Real Compton scattering on a scalar target in AdS/QCD
models with minimal coupling to the photon is exactly the same as in
point-like scalar electrodynamics in the tree level approximation, a
fact which was observed in the hard-wall model of \cite{GX09}. We
elaborate on the implications for AdS/QCD in this section.
The first consequence is that the static polarizabilities of the scalar
target vanish. Polarizabilities give the corrections to Thompson
scattering which are quadratic in the energy of the photons
\cite{POL}. The amplitude for real Compton scattering off a spinless
particle like the pion,
\begin{align*}
\gamma(q_1)\,\pi(p_1)\ \longrightarrow\ \gamma(q_2)\,\pi(p_2) \,,
\end{align*}
can be expanded in powers of the energies of the photons near
threshold and reads, in the non-relativistic limit, $\omega_i^2\ll
m_{\pi}^2$, in the laboratory frame and in the Coulomb gauge,
\begin{align}
A(\gamma\pi\rightarrow\gamma\pi) = 2e^2\v{\epsilon_1}\cdot\v{\epsilon_2}
+ 8\pi m_{\pi}\,\omega_1\omega_2
\bigl(\alpha_{E}\v{\epsilon_1}\cdot\v{\epsilon_2} + \beta_{M}\,
(\v{\epsilon_1}\times \h{q}_1)\cdot(\v{\epsilon_2}\times \h{q}_2)\bigr)
+ \cdots \,
\end{align}
where $q_i=\omega_i(1,\h{q}_i)$ and $\v{\epsilon_i}$ are the momentum
and polarization vector of each photon (with
$\h{q}_i^2=\v{\epsilon_i}^2=1$). $\alpha_{E}$ and $\beta_{M}$ are the
electric and magnetic polarizabilities respectively. They measure the
linear response of a particle with an internal structure to a small
external electromagnetic perturbation.
The cancellations of the poles in $q_1^2=0$ and $q_2^2=0$ in the
Compton tensor \eqref{Vmunu} impose the following relations between
the Compton form factors,
\begin{align}
V_1 + (q_1\cdot q_2) V_3 = -(p\cdot q_1)V_4 = -(p\cdot q_2)V_5
= \frac{(p\cdot q_1)(p\cdot q_2)}{q_1\cdot q_2}\,V_2\,.
\end{align}
The most general gauge-invariant real Compton tensor can be written in
terms of the two independent form factors $V_1$ and $V_2$,
\begin{align}
T^{\mu\nu}_{RCS} &= V_1
\left(g^{\mu\nu}-\frac{q_1^{\nu}q_2^{\mu}}{q_1\cdot q_2}\right)
+ V_2\left(p^{\mu}-\frac{p\cdot q_1}{q_1\cdot q_2}q_2^{\mu}\right)
\left(p^{\nu}-\frac{p\cdot q_2}{q_1\cdot q_2}q_1^{\nu}\right) \,.
\end{align}
Since the static polarizabilties $\alpha_E$ and $\beta_M$ are defined
in the laboratory frame, $\v{p_1}= \b{0}$, it is convenient to choose
the Coulomb gauge and impose the conditions $\epsilon_1\cdot p_1 =
\epsilon_2^*\cdot p_1 = 0$. Therefore the contracted real Compton
amplitude can be written as
\begin{align}
A_{\text{RCS}} = \epsilon_1^{\mu}\,T_{\mu\nu}\,\epsilon_2^{\star\nu}
= V_1\,\v{\epsilon_1}\cdot\v{\epsilon_2}^{\star}
+\left(V_3-V_2\right)
(\v{\epsilon_1}\cdot \v{q_2})(\v{\epsilon_2}^{\star}\cdot\v{q_1}) \,.
\end{align}
We can use the identity,
\begin{align}
(\v{\epsilon_1}\times\v{q_1})\cdot
(\v{\epsilon_2}\times\v{q_2})
&= (\v{\epsilon_1}\cdot\v{\epsilon_2})
(\v{q_1}\cdot\v{q_2}) -
(\v{\epsilon_1}\cdot\v{q_2})(\v{\epsilon_2}\cdot\v{q_1}) \,,
\end{align}
and relate the electric and magnetic polarizabilities in the lab-frame
to the Compton form factors,
\begin{align}
\label{EM}
\begin{split}
8\pi m\,\alpha_E &= \left.\pxx{}{\omega_1}{\omega_2}
\left(V_1+(V_3-V_2)\v{q_1}\cdot\v{q_2}\right)\right|_{\omega_1=\omega_2=0}\,,
\\ 8\pi m\,\beta_M &= \left.(V_2-V_3)\right|_{\omega_1=\omega_2=0} \,.
\end{split}
\end{align}
In the particular case of \eqref{TDVCS}, in the limit $q_1^2=0$, we
have $V_1 = 2$ and $V_2=V_3$. Hence $\alpha_E=\beta_M=0$ as expected
for real Compton scattering on a structureless particle.
The point-like nature of the real Compton scattering is a direct
consequence of the minimal coupling between the bulk vector field and
the bulk scalar field in AdS space together with the boundary
condition \eqref{BC}. In order to get non-vanishing polarizabilities
we need to introduce non-minimal couplings between the bulk vector
field and the bulk scalar field in anti-de Sitter space.
It is well-known that the same problematics is encountered in the
calculation of the pion polarizabilities in chiral perturbation theory
($\chi$PT). At lowest-order in the momentum expansion, the pion is
coupled minimally to the electromagnetic field $A_{\mu}$ and the
polarizabilities vanish. The chiral Lagrangian at tree evel can only
predict the $\pi$-$\pi$ scattering lengths. Only the phenomenological
chiral couplings of order 4 can produce non-zero polarizabilities. It
can be shown \cite{DH89} that the predictions at order $p^4$ in the
chiral limit for the electric and magnetic polarizabilities of the
charged pion,
\begin{align}
\alpha_E = \frac{4\alpha}{m_{\pi}F_{\pi}^2}(L_9^r+L_{10}^r)\,,\qquad
\alpha_E + \beta_M = 0 \,,
\end{align}
are generated by the four-dimensional effective Lagrangian,
\begin{gather}
-i\,L_9\,F_{\mu\nu}\tr\left(Q\,D^{\mu}U(D^{\nu}U)^{\dagger}
+ Q\,(D^{\mu}U)^{\dagger}D^{\nu}U\right)
+ L_{10}\,F_{\mu\nu}F^{\mu\nu}\,
\tr\left(Q\,U\,Q\,U^{\dagger}\right) \,, \\ \nonumber
D_{\mu}U = \p{_\mu}U + ie\,A_{\mu}\,[Q,U] \,.
\end{gather}
It is therefore very easy to write down a covariant and
gauge-invariant effective action in the 5D AdS space that can be added
to the minimal action \eqref{AdS} to generate non-vanishing
polarizabilities at the classical level for a charged (pseudo)scalar
particle, e.g.,
\begin{align}
\label{XPT}
\begin{split}
S'_{\text{AdS}}[\Phi,\Phi^*,A^m] &=
\int d^4xdz\,\sqrt{-g}e^{-\chi}\biggl(
ig_1\frac{e}{2}\,F_{mn}(D^m\Phi(D^n\Phi)^* - (D^m\Phi)^*D^n\Phi) \\
&\qquad\qquad\qquad\qquad\qquad
+ g_2\frac{e^2}{4} F_{mn}F^{mn}\Phi^*\Phi\biggr) \,.
\end{split}
\end{align}
Of course one could even go one step further in phenomenology and
introduce non-minimal couplings between bulk fields of various spin
and parity, in the spirit of the effective Lagrangian approach.
\section{Conclusion}
\label{END}
We have worked within the bottom-up approach to the AdS/QCD
correspondence and calculated the Compton amplitude with an arbitrary
dilaton background. There is a very recent study \cite{GX09}, within
the approach of \cite{PS03}, which overlaps with ours. There
are however significant differences which make the two papers
complementary. Working in a generic soft-wall model has helped us to
clarify the Lorentz-invariant and gauge-invariant structure of the
Compton amplitude predicted by AdS/QCD. Moreover the structure of the
Compton amplitude does not depend upon the infrared cutoff
parametrized by the dilaton background.
We have found that the minimal coupling of a bulk (pseudo)scalar field
to the electromagnetic current cannot reproduce the expected
low-energy behavior of the Compton amplitude off a spinless composite
charged particle, and produces a too simple structure in the DVCS
kinematical region for a partonic interpretation.
We have pointed out an obvious signature of this failure, namely the
vanishing of the electric and magnetic polarizabilities of the scalar
target. The experimental situation is rather unsatisfactory since the
extraction of the experimental values is model dependent. For instance
the most recent experimental values for the polarizabilities of the
charged pion are,
\begin{align}
\begin{split}
(\alpha_E-\beta_M)_{\pi^+} &=
(11.6\pm 1.5_{stat}\pm 3.0_{syst}\pm0.5_{mod})\times 10^{-4}\text{fm}^3
\quad\text{\cite{MAMI}}\,, \\
(\alpha_E = -\beta_M)_{\pi^+} &= (2.5 \pm 1.7_{stat}\pm 0.6_{syst})
\times 10^{-4}\text{fm}^3\quad\text{\cite{COMPASS}}\,.
\end{split}
\end{align}
Even if these values are still imprecise, the inclusion of non-minimal
couplings to the photon is certainly required to obtain a realistic
description of Compton scattering in AdS/QCD at the classical level.
For instance, we note that non-minimal couplings to vector fields
generate five independent Compton form factors, (as opposed to only
three with the minimal coupling we considered in this paper), as
allowed by gauge and Lorentz invariance.
Such couplings appear naturally in chiral perturbation theory.
Besides, an algebra of currents based on chiral symmetry is the
standard framework to describe the hadronic electromagnetic
current. The AdS/QCD models we have examined do not incorporate chiral
flavor symmetry nor vector meson dominance. There are several
variants of chiral AdS/QCD models \cite{SS04,EKSS05,RP05,HS05} and it
is not the purpose of the present work to commit to one of
them. Nethertheless we have identified a bare-bones effective anti-de
Sitter action that can contribute to the polarizabilities in many
chiral models.
In any case it should now be clear that the calculation, and the
precise measurement, of the hadronic polarizabilities is a selective
testing ground for the AdS/QCD correspondence.
\subsection*{Acknowledgements}
We wish especially to thank B.~Pire for inspiring discussions and very
useful comments about the manuscript. We also wish to thank
V.~Bernard, J.P.~Lansberg, B.~Moussallam, T.N.~Pham, U.~Reinosa,
L.~Szymanowski and B.~Xiao for very interesting discussions. C.M. is
supported by the European Commission under the FP6 program, contract
No. MOIF-CT-2006-039860. This work is partly supported by the
ANR-06-JCJC-0084.
\subsection*{Appendix:\quad Explicit formulas in the hard-wall model}
The hard-wall model is defined by the absence of dilaton background,
$\chi(z)=0$, and by imposing a Dirichlet boundary condition on the
massive fields at some finite cutoff in $z$. Then the plane-wave
solution for a massive scalar field reads (in the timelike region),
\begin{align}
\Phi(x,z) &= C_{\Delta-1}(m)\,e^{ip\cdot x}z^2\,
J_{\Delta-2}\left(m\,z\right)
\,,\quad \Delta=2\pm\sqrt{m_S^2+4} \geq 1\,,\quad p^2=-m^2<0 \,,
\end{align}
(this is not the most general admissible solution for $1\leq\Delta\leq
3$). The normalization constants,
\begin{align}
\label{normalization}
C_{\Delta-1}(m) = \sqrt{2}\,\Lambda\,
J_{\Delta-1}^{-1}\left(\frac{m}{\Lambda}\right)\,,
\end{align}
are defined by requiring,
\begin{align}
\int_0^{1/\Lambda}dz\,z^{-3}|\Phi(x,z)|^2 &= 1\,,
\quad J_{\Delta-2}\left(\frac{m}{\Lambda}\right) = 0 \,.
\end{align}
The scalar propagator reads,
\begin{gather}
\h{G}(z_1,z_2,-m^2) = \sum_n C^2_{\Delta-1}(m_n)
\frac{z_1^2J_{\Delta-2}(m_nz_1)\,z_2^2J_{\Delta-2}(m_nz_2)}
{m^2-m_n^2+i\epsilon}\,,\\
m_n = \zeta_{\Delta-2,n}\Lambda \,,
\end{gather}
where $\zeta_{\nu,n}$ are the zeroes of the Bessel function
$J_{\nu}(z)$. When $\Lambda\rightarrow 0$, the scalar propagator
becomes,
\begin{align}
\h{G}(z_1,z_2,-m^2) \underset{\Lambda\rightarrow 0}{\approx}
(z_1z_2)^2\int_0^{\infty}d\mu\,\mu\,
\frac{J_{\Delta-2}(\mu z_1)\,J_{\Delta-2}(\mu z_2)}
{m^2-\mu^2+i\epsilon} + \O(\Lambda) \,.
\end{align}
\vskip 0.2cm
Plugging the explicit wave-functions into \eqref{FF1} one
gets for the scalar DVCS form factor,
\begin{align}
\C_1(m^2,Q^2,0) = C^2_{\Delta-1}(m)Q\int_0^{1/\Lambda} dz\, z^2\,
J^2_{\Delta-2}(mz)\,K_1(Qz) \,.
\end{align}
As long as $\Delta>1$ and $Q\gg\Lambda$, we can set $\Lambda=0$ in the
integration domain and use the integral formula,
\begin{align}
\C_1(m^2,Q^2,0) &\simeq 2(\Delta-1)\frac{C^2_{\Delta-1}(m)}{m^2}\times
\left(\frac{m^2}{Q^2}\right)^{\Delta-1}\times
\frac{(1-w)^{2\Delta}}{(1-w^2)^2}
\left(1 + \frac{1}{\Delta-1}\frac{2w^2}{1-w^2}\right) \,,
\end{align}
where $w$ is defined by
\begin{align}
w = 1+\frac{Q^2}{2m^2} -
\sqrt{\left(1+\frac{Q^2}{2m^2}\right)^2-1}\,.
\end{align}
Noting that
\begin{gather}
w \underset{Q^2\rightarrow\infty}{\approx}
\frac{m^2}{Q^2}\left(1+\O\left(\frac{m^2}{Q^2}\right)\right)
\longrightarrow 0\,,
\end{gather}
we obtain the leading large $Q^2$ behavior of the DVCS form factor
found in \cite{GX09},
\begin{align}
\C_1(m^2,Q^2,0) &= 2(\Delta-1)\frac{C^2_{\Delta-1}(m)}{m^2}\times
\left(\frac{m^2}{Q^2}\right)^{\Delta-1} \,.
\end{align}
|
\section{Introduction}
\label{sec:intro}
Models with extra spatial dimensions allow us to confront some of the
outstanding issues of the Standard Model (SM) (see \cite{ADD, RS,ED6}).
In particular, the Universal Extra Dimensions (UED) scenario \cite{Appelquist:2000nn}
leads to an interesting dark
matter candidate \cite{Servant:2002aq,Cheng:2002ej,Kong:2005hn,Burnell:2005hm,Arrenberg:2008wy}
as well as a foil for searches for Supersymmetry
at colliders \cite{Rizzo:2001sd,Cheng:2002ab,Datta:2005zs,Battaglia:2005zf}. In the original minimal UED picture (MUED), all of the
SM fields live in a 5-dimensional $S^1/Z_2$ orbifolded bulk with a
compactification radius $R$. Due to the breaking of 5D Lorentz invariance,
Kaluza-Klein (KK) number is no longer conserved although a $Z_2$ symmetry,
KK-parity, remains. This being the case, the tree-level wave functions for the
various KK states are either sines or cosines in the coordinate
of the extra dimension. Allowing for radiative loop
corrections to the tree-level particle masses, the physics of MUED is then
described by only two parameters beyond those of the SM \cite{Cheng:2002iz}:
$R$ and a cutoff
scale, $\Lambda$, used to define these loop corrections, which is usually
taken such that $\Lambda R \sim {\cal O}(10-100)$ but with only logarithmic sensitivity
to this particular choice.
In MUED and its extension to higher dimensions
\cite{Burdman:2006gy,Dobrescu:2007xf,Freitas:2007rh,Dobrescu:2007ec},
the bulk masses of the
SM fermions are taken to be zero. However, this is no longer the case in
Split-UED (SUED) \cite{sued1,sued2,sued3,sued4}. Indeed this `bulk mass' term is naturally included in the effective Lagrangian as the term is compatible with 5D Lorentz invariance as well as gauge invariance of the model. Here one notes that in order to maintain the KK parity the `coefficient' of the $\bar \Psi \Psi$ fermion bilinear term in the action must be an odd function of the 5D coordinate,
$y$, defined on an interval, $y \in (-L, L)$ where $L=\frac{\pi R}{2}$.
The simplest choice to make in
this case, as is similarly done in the Randall-Sundrum (RS) model,
is to write this coefficient as $\mu \theta(y)$, where $\mu$ is
a dimensionful parameter whose value is, in general, dependent upon which
SM field is being considered and $\theta(y)=1(-1)$ for $y>(<)0$. Naturally,
one might expect that the values of $\mu$ can be of either sign and be of
order $\sim 1/R$. The effects of including a non-zero value for $\mu$ are
two-fold: First, depending upon its sign, the fermion zero-modes, which are
identified with the known SM fermions, no longer have flat wave functions in
the extra dimension. These are now found to be either peaked near $y=0$ or at
the orbifold boundaries; this leads to potentially large differences in the
various couplings of these fermions to the KK gauge fields from those expected
in MUED. In particular, the zero mode fermions now have tree-level couplings
to the KK-number even gauge modes. Second, the KK fermion wave functions and
masses (which are given by $\sim n/R$ at tree-level) are now somewhat more
complicated and are explicitly dependent upon the specific value
of the $\mu$ parameter. In particular, the expressions for the KK fermion
masses are {\it different} depending upon whether the relevant
KK-number is even or odd.
The purpose of this paper is ($i$) to explore in some of the detailed
implications of non-zero values for fermion mass parameter $\mu$ leading
to alterations from the conventional MUED
phenomenology and ($ii$) to investigate the regions in the $R-\mu$ plane
which are accessible to current and future collider experiments.
To these ends, in Section \ref{sec:SUED}, we provide a basic overview of the masses,
wave functions and couplings of the fermion KK states in
split-UED model and display their
explicit dependence upon the parameter $\mu$ pointing out important
differences with the MUED case. Here we will assume that the $\mu$ parameter
takes on a universal value for all fermions for simplicity of the analysis so
that there is only one new parameter to consider beyond that of MUED. In
Section \ref{sec:collider} we will discuss the collider phenomenology of split-UED
and, in particular, the properties of the KK states and the potential for their
discovery at the LHC. Further, we obtain the regions of the $R-\mu$ plane
which are allowed by current experimental data and show the regions which will
be made accessible by searches at the LHC. Our conclusions can be found in
Section \ref{sec:conclusion}.
Appendix \ref{app:spectrum} contains detailed information of KK decomposition and mass spectrum.
\section{Split Universal Extra Dimensions}
\label{sec:SUED}
\subsection{Model}
\label{sec:model}
Universal extra dimensions postulates that all of the Standard Model particles are propagating in a small extra dimension(s). Orbifold compactification makes it possible to construct a chiral four dimensional effective theory.
In contrast to the brane world scenarios \cite{ADD,RS}, the translational symmetry along the extra dimension leads to a remnant discrete symmetry, dubbed KK parity, so that the lightest Kaluza-Klein particle can be a good dark matter candidate. Also this parity mimics the R-parity in supersymmetric theory so that UED phenomenology shares several common features with MSSM \cite{Cheng:2002ab}. On the other hand, it has been often overlooked in UED models that the bulk Dirac masses are generically allowed and are not in conflict with higher dimensional
Lorentz symmetry or gauge invariance. In this section, we review split-UED model where these bulk Dirac masses are generically allowed in a way that KK parity is intact.
In split-UED, quarks $(Q, U^c, D^c)$ and leptons $(L, E^c)$ are all promoted to fields in five dimensional spacetime on $S^1/Z_2 \times M^4$ orbifold with two fixed points $y=- L$ and $y=L$, respectively, where $y$ is the coordinate along extra dimension with the half length $L=\pi R/2$. In the minimal setup, the gauge group is the same as in the Standard Model: $SU(3)_c\times SU(2)_W \times U(1)_Y$ under which charges are assigned as follows
\begin{eqnarray}
\Psi_i(x,y)=(Q_i, U^c_i, D^c_i, L_i, E^c_i)^c=((3,2)_{1/6},( \bar{3},1)_{-2/3},(\bar{3},1)_{1/6},(1,2)_{-1/2},(1,1)_{1})^c \, ,
\end{eqnarray}
where the index $i$ runs for three generations of fermions.
Allowing bulk mass term in split-UED
the generic action $S=\int d^4 x \int_{-L}^L dy {\cal L}_5$ is given by
\begin{eqnarray}
{\cal L}_5= \sum_{i,j=1}^3 \frac{i}{2} (D_M \bar{\Psi}_i \Gamma^M \Psi_j -
\bar{\Psi}_i \Gamma^M D_M \Psi_j) -m_{ij}(y) \bar{\Psi}_i \Psi_j \, ,
\label{Eq:action}\end{eqnarray}
where the covariant derivative is $D_M =\partial_M + i g_3 \frac{\lambda^\alpha}{2} G^\alpha_M + i g_2 \frac{T^a}{2} W^a_M + i g_1Y B_M $ with the usual Gell-Mann and Pauli matrices $\lambda$ and $T$.
The $g_1$, $g_2$, $g_3$ and $G$, $W$, $B$ are the gauge coupling constants and
the gauge fields of the corresponding gauge groups, respectively.
Without loss of generality we can diagonalize the action in Eq. (\ref{Eq:action})
by unitary transformations. Therefore the mass term $m_{ij}$ can be taken as
$ m_{ij} = m_i \delta_{ij}$ and there is no kinetic mixing between different flavors (for $i\neq j$).
In general, we may have dimensionful parameters $(m_Q, m_{U^c}, m_{D^c}, m_L, m_{E^c})$ for each generation. Now imposing Dirichlet boundary conditions for unnecessary chiral component of fermions,
we can finally get exactly the same spectra in the SM for the lowest Kaluza-Klein modes.
All the details of derivation to get the Kaluza-Klein spectra are described in the Appendix.
The most prominent feature of split-UED is that the fermion profile in the extra dimension is either localized near the origin or at boundaries depending the sign of bulk mass parameter $m_i(y)=\mu_i \theta(y)$ in a way that Kaluza-Klein parity is respected. Having a non-zero bulk mass, $m$, a field still has a massless zero mode which satisfies Neumann boundary conditions. However its Kaluza-Klein excitation states get additional contributions and the mass is given by
$m_n = \sqrt{k_n^2 + \mu^2}$ where $k_n$ is the momentum to the extra dimension which is determined by $\mu= \pm k_n \cot k_n L$ for $n\in Z_{\rm odd}$ or $k_n = \pi n/L$ for $n\in Z_{\rm even}$. Here we impose Dirichlet boundary conditions for $\Psi_L$ modes so that $\Psi_R$ contains the SM fermions in our
convention. We assume that gauge sector and Higgs sector remain the same as in the conventional UED models. Therefore the zero modes have flat profiles and Kaluza-Klein modes have cosine wave functions satisfying Neumann boundary conditions.
In summary, in split-UED, there are new $15$ dimensionful parameters $\mu_\Psi$ for three generations,
the cutoff scale ($\Lambda$) and one length parameter $L$ ($=\pi R/2$) given by the size of extra dimension in addition to the SM parameters.
In this study we will consider all bulk mass parameters are the same for simplicity,
and study $\mu L \geq -1$ region
\footnote{For $\mu L < -1$, the KK spectra contain the unacceptable light modes
below KK scale $\sim$ TeV.}.
\subsection{Behavior of couplings}
\label{sec:couplings}
Having the explicit wave functions of fields as given in the Appendix, we can calculate the explicit Lagrangian for interactions among those fields. Essentially the overlap integral of the wave functions gives the effective couplings. For a gauge boson $V=(G, W, B)$ after choosing a simplifying gauge to get rid of the fifth component of
gauge multiplet, $V_5$, by the orbifold condition, we find
\begin{eqnarray}
-{\cal L}_{\rm int} &\ni& g_V \int_{-L}^L d y \, \bar{\Psi}\Gamma^\mu \Psi V_\mu \\
&=& g_V \sum_{\ell mn}\int_{-L}^L d y \,
\Big [ \bar{\psi}^\ell {f^\ell_\Psi}^* (y) \Big ] \gamma^\mu
\Big [ \psi^m f_\Psi^m(y) \Big ]
\Big [ V_\mu^n f_V^n(y) \Big ] \\
&=& \sum_{\ell mn} g^{eff}_{\ell mn} \bar{\psi}^\ell \gamma_\mu \psi^m V_\mu^n \, ,
\end{eqnarray}
where the effective coupling is obtained by the integration of the wave function overlap
with a convenient dimensionless variable $x_\Psi=\mu_\Psi L$:
\begin{eqnarray}
g^{eff}_{\ell mn} &&\equiv g_V \int_{-L}^L d y \, {f^\ell _\Psi}^*(y) f_\Psi^m(y) f_V^n(y) \\
&&\equiv g_V {\cal F}_{\ell mn}(x_\Psi) \, .
\end{eqnarray}
As the profiles of gauge bosons are universal and the profile of fermions depend on
the bulk mass parameter $\mu_\Psi$, the overlap integral ${\cal F}_{\ell mn}$ is the same for all gauge bosons
but depends on $\mu_\Psi$. The suppressed gauge group indices should be understood.
Let us now see the coupling between KK bosons ($G, W, B$) of $SU(3)_c$, $SU(2)_W$ and $U(1)_Y$ and the zero mode SM fermion pair for the definiteness \footnote{As it is clear in minimal UED, the weak mixing angles for KK gauge bosons are suppressed by $m_W/m_{\rm KK} \ll 1$. Thus essentially gauge eigenstates are well aligned by mass eigenstates.}.
The zero mode wave function profile for SM fermion $\Psi_i=(Q,U^c, D^c, L,E^c)^c$ is given by
\begin{eqnarray}
f_i^{(0)}(y) = \sqrt{\frac{\mu_i}{1-e^{-2\mu_i L}}} e^{-\mu_i |y|} \, .
\label{Eq:zeromode}
\end{eqnarray}
If $\mu_i>0 (<0)$, the profile is exponentially localized near the center (at the boundaries). The zero mode is massless in the absence of the electroweak symmetry breaking even though its KK modes get additional mass from the bulk mass $\mu_i$.
\FIGURE[t]{
\centerline{
\epsfig{file=FIGURES/fermi_new.ps, width=7.6cm} \hspace{0.0cm}
\epsfig{file=FIGURES/vec_new.ps, width=7.6cm}}
\caption{\sl Profiles of zero mode fermion with various $\mu$'s in (a)
and first three KK gauge bosons $n=0,1,2$ in (b). Kaluza-Klein parity
is obviously respected by the zero mode profile and the localization depends on
the sign of the bulk mass $\mu$. KK-odd mode ($n=1, 3, 5, \cdots$) are odd and even modes ($n=0,2,\cdots$) are even under KK parity.
}
\label{fig:profile}}
The KK gauge bosons commonly have the same profiles as in MUED
\begin{eqnarray}
f_{V=G, W, B}^{(n>0)}(y)=\frac{1}{\sqrt{L}} \cos \frac{ n \pi (y+L) }{2L} \, ,
\label{eq:vec}
\end{eqnarray}
and the zero mode profile, $f_V^{(0)}=1/\sqrt{2L}$, is flat, as shown in Fig.~\ref{fig:profile}.
Note that $\int_{-L}^L dy \, \big (f_V^{(n)} \big )^2=1$.
The coupling of level-$n$ bosons to SM fermion pair is now written as
\begin{equation}
{\cal L}_{\rm eff}\ni -\sum_\psi \sum_n \frac{C_{n}(\mu_\psi)}{\sqrt{2L}} \left[\bar{\psi_0}\gamma^\mu \left(g^{5D}_3\frac{\lambda^a}{2} G^{a,(n)}_\mu +g^{5D}_2\frac{T^i_\psi}{2} W_\mu^{i,(n)} + g^{5D}_1 Y_\psi B_\mu^{(n)}\right)
\psi_0\right] \, , \,
\label{Eq:int}
\end{equation}
where $C_n$ is a dimensionless parameter measuring the overlap of wave functions
between two SM fermions and a KK gauge boson defined as
\begin{eqnarray}
C_{n}(\mu_\psi L)&\equiv& \sqrt{2L}\int_{-L}^{L} dy (f_\psi^{(0)})^2 f_V^{(n)} \\
&=& \left\{ \begin{array}{ll}
0 , & \hspace{1cm}\mbox{$n =1, 3, 5, 7, \cdots$};\\
{\cal F}_{00n}(x_\psi) , & \hspace{1cm}\mbox{$n = 0, 2, 4, 6, 8, \cdots$} ,\end{array} \right.
\end{eqnarray}
where $x_\psi = \mu_\psi L$ and ${\cal F}$ is explicitly calculated to be
\begin{equation}
{\cal F}_{002m} (x)
=\frac{x^2 (-1+(-1)^m e^{2x})(\coth x-1)}{\sqrt{2(1+\delta_{m0})}(x^2 + m^2\pi^2/4)} \, ,
~~~ m=0, \, 1, \, 2, \,3 , \, \cdots \, .
\end{equation}
From the KK-parity conservation, $C_{odd}=0$ is easily understood. The Standard Model coupling constants are obtained as
\begin{eqnarray}
g^{SM}=g_{000}^{eff}=\frac{g^{5D}}{\sqrt{2L}} C_0 (x_\psi) = \frac{g^{5D}}{\sqrt{2L}} \, ,
\label{eq:gsm}
\end{eqnarray}
as $C_0(x_\psi)={\cal F}_{000}(x_\psi)=1$ independent of $x_\psi$. Here $g_{\ell nm}^{eff}$
denotes the effective coupling constant for the $\psi_\ell -\psi_m-V_n$ interaction.
Finally for the even $n$'s we get the coupling between the Standard Model fermions and even-KK excitation states of gauge bosons, as
\begin{eqnarray}
g_{002n}^{eff}=g^{SM} {\cal F}_{002n} (x_\psi) \, .
\end{eqnarray}
\FIGURE[t]{
\centerline{
\epsfig{file=FIGURES/couplings.ps, width=7.5cm}
\epsfig{file=FIGURES/F2n.ps, width=7.5cm}}
\caption{\sl The ratio of tree level couplings in SUED to the corresponding SM couplings.
Couplings involving level-2 (level-1) KK bosons are shown in red (blue) in (a). (b) contains zero mode fermion couplings to KK-even gauge bosons $f_0-f_0-V_{2n}$.
The MUED limit ($\mu=0$) is denoted by the vertical solid line (in magenta).
}
\label{fig:couplings}}
The various couplings associated with one vector boson and two fermions are
shown in Fig.~\ref{fig:couplings}.
We find that there are two interesting regions.
One is the MUED limit, {\it i.e.,} $\mu \to 0$, which is
shown as the vertical solid line (in magenta) and the other is large positive $\mu$ limit.
For $\mu \to \infty$, the zero mode fermions are well localized near the center ($y=0$)
so that their couplings to KK gauge bosons asymptotically approach to the well known value $(-1)^n \sqrt{2}$ as one can see the red
curve for $f_0-f_0-V_2$ (Fig.~\ref{fig:couplings}(a))
as well as the curves for $f_0-f_0-V_{2n}$ (Fig.~\ref{fig:couplings}(b)).
The alternating sign can be understood as the $2n$-th KK gauge boson wave function in Eq. (\ref{eq:vec}) is proportional to $\cos n\pi = (-1)^n $ at $y=0$ where the fermion wave function is mostly localized. The $\sqrt{2}$ is from the zero mode normalization in Eq. (\ref{eq:gsm}). These vertices are all vanishing in the limit of $\mu\to 0$ because of KK number conservation in MUED.
For collider phenomenology, we are mostly interested in interactions for
low lying Kaluza-Klein modes, $n=0,1,2$ as heavier modes are too massive and
easily decouple from the low energy phenomenology.
The most relevant couplings in our study are the interactions
and decays of the second Kaluza-Klein gauge boson.
The coupling $f_n$-$f_n$-$V_0$ remains the same for all $\mu$ due to the normalization condition of
wave functions for $n$-th fermion profile, while all other couplings now change
for non-vanishing bulk masses.
The $f_2$-$f_0$-$V_0$ coupling remains zero in SUED
but in principle this coupling can be generated
by the unknown physics at the cutoff scale ($\Lambda$), and the lowest order coupling may take
the form \cite{Cheng:2002iz}
\begin{equation}
\bar{f}_2 \sigma^{\mu\nu} T^a P_{L/R} f_0 F_{0\mu\nu}^a \, .
\end{equation}
However, being higher dimensional, we expect it to be suppressed at least
by one power of $1/\Lambda$, hence we shall neglect it in the discussion that follows.
It is interesting to notice that the $SU(3)_c$ coupling for the KK gluon can be {\it chiral}.
Let us examine the level-$2n$ gluon couplings with quarks:
\begin{eqnarray}
-{\cal L}_{\rm eff}=g_s\sum_{n \geq 0} [
\bar{u}\gamma^\mu \left({\cal F}_{002n}(x_Q)P_L + {\cal F}_{002n}(x_U)P_R \right)u\nonumber \\
+\bar{d}\gamma^\mu \left({\cal F}_{002n}(x_Q)P_L + {\cal F}_{002n}(x_D)P_R \right)d]G_\mu^{(2n)} \, .
\end{eqnarray}
All the KK-parity violating interactions are forbidden.
Now it is obvious that KK-gluon has chiral interactions with the SM quarks,
if $\mu_Q \neq \mu_U$ or $\mu_Q \neq \mu_D$, in general.
Finally the vector (V) and axial-vector (A) couplings of KK gluons
with an up-type quark and a down-type quark,
\begin{eqnarray}
-{\cal L}_{\rm eff}=g_s \sum_{q=u,d}\sum_{n\geq 0}\bar{q}\gamma^\mu (V^q_{2n} -A^q_{2n}\gamma_5) q G_\mu^{(2n)} \, ,
\end{eqnarray}
are determined as
\begin{eqnarray}
V_{2n}^{u/d} =\frac{1}{2}\left({\cal F}_{002n}(x_Q)+{\cal F}_{002n}(x_{U/D})\right) \\
A_{2n}^{u/d} =\frac{1}{2}\left({\cal F}_{002n}(x_Q)-{\cal F}_{002n}(x_{U/D})\right) \, .
\end{eqnarray}
The same is similarly true for all other gauge bosons as well. When $x_Q=x_{U/D}$, only the vectorial coupling is non-vanishing. However, in general, $x_Q \neq x_{U/D}$ and the non-vanishing axial couplings are allowed. For instance, if $x_Q=0$ and $x_{U/D}\neq 0$, the vectorial and axial couplings have opposite signs but have the same size: $V_{2n}^{u/d}=-A_{2n}^{u/d}={\cal F}_{002n}(x_{U/D})$. With non-vanishing axial couplings in the even KK gauge boson interactions, one might expect, for instance, additional contribution to the forward-backward asymmetry of the top quark pair production ($A_{FB}^t$) via the quark pair annihilation channel. The cross-section for $q\bar{q}$ annihilation into top quarks of mass $m_t$
through the $2n$-th KK gluons reads
\begin{eqnarray}
\frac{d\sigma (q\bar{q}\rightarrow g_{2n}^* \to t \bar{t})}{d\cos \hat{\theta}} &=&
\frac{\pi \beta \alpha_S^2}{9 \hat{s}} \Big \{ 1+c^2 \beta^2+ \frac{4 m_t^2}{\hat s} \nonumber \\
&& \hspace*{-3cm} +
\sum_{n \geq 1} \frac{2 \hat{s} (\hat{s}-m_{2n}^2)} {(\hat{s}-m_{2n}^2)^2+m_{2n}^2 \Gamma_{2n}^2}
\Big [ V_{2n}^q \, V_{2n}^t \, \big (1+c^2 \beta^2+ \frac{4 m_t^2}{\hat s} \big )
+ 2 \, A_{2n}^q \, A_{2n}^t \, c \beta \Big ] \nonumber \\
&& \hspace*{-3cm}+
\sum_{n,\ell \geq 1} \hat{s}^2 \frac{(\hat{s}-m_{2n}^2)(\hat{s}-m_{2\ell}^2) + m_{2n} m_{2\ell} \Gamma_{2n} \Gamma_{2\ell}}
{ [ (\hat{s}-m_{2n}^2)^2+m_{2n}^2 \Gamma_{2n}^2 ] [ (\hat{s}-m_{2\ell}^2)^2+m_{2\ell}^2 \Gamma_{2\ell}^2 ] } \\
&& \hspace{-2.5cm}\times \Big [ \Big ( V_{2n}^q V_{2\ell}^q + A_{2n}^q A_{2\ell}^q \Big )
\Big ( V_{2n}^t V_{2\ell}^t \big (1+c^2 \beta^2+ \frac{4 m_t^2}{\hat s} \big )
+ A_{2n}^t A_{2\ell}^t \beta^2 (1+c^2) \Big ) \nonumber \\
&& \hspace{-1cm} + 2 \, c \beta \, \Big ( V_{2n}^q A_{2\ell}^q + V_{2\ell}^q A_{2n}^q \Big )
\Big ( V_{2n}^t A_{2\ell}^t + V_{2\ell}^t A_{2n}^t \Big ) \Big ] \nonumber
\Big\} \, , \nonumber
\label{eq:bornqq}
\end{eqnarray}
where $\hat{\theta}$ is the polar angle of the top quark with respect
to the incoming quark in the center of mass rest frame,
$\hat{s}$ is the squared partonic invariant mass,
$\beta = \sqrt{1-\frac{4 m_t^2}{\hat s}}$ is the velocity of the top quark,
with $c = \cos \hat{\theta}$.
The parameters $V_{2n}^q (V_{2n}^t)$ and $A_{2n}^q(A_{2n}^t)$ represent,
respectively, the vector and axial-vector couplings of the
KK gluons to the light quarks (top quarks).
Considering experiments at Tevatron, the parton level energy $\hat{s}$ is typically
much less than the KK gluon mass so that the interference term (the second term) is
dominant over the pure new physics term (the third term).
And the leading contribution in the second term is
the interference between two diagrams the SM gluon and the level-2 KK gluon.
As the tree level SM contribution (the first term) does not produce
the forward-backward asymmetry after integrating over
$-1< \cos \hat{\theta}<1$, the main contribution is from the linear term of cosine
in the second term for $n=1$:
\begin{eqnarray}
A_{FB}^t
&\propto & - \frac{A_{2}^q A_{2}^t}{m_{2}^2}.
\end{eqnarray}
When $x_t = -1$, $x_{U/D} \to \infty$ and $x_Q=0$,
\begin{eqnarray}
A_{2}^q \to \frac{1}{\sqrt{2}}, \,\, A_{2}^t \to -\frac{1}{4},
\end{eqnarray}
thus the forward-backward asymmetry is positive, which is consistent with the recent measurements at Tevatron
\cite{newcdf, cdf, d0}, but we find that its size is not large enough to explain the current anomaly
for $R^{-1} \sim 1$ TeV.
\subsection{Mass spectrum}
\label{sec:masses}
The mass spectrum of fermions gets tree level modifications from the bulk parameters $\mu_\Psi$ as well as the loop induced mass correction from the RG running effect for a given boundary condition at some high scale $\Lambda$ just as in the case in conventional UED. Taking the vanishing boundary condition at $\Lambda$, it is known that the one-loop induced mass correction is minor ($\sim \%$ level for electroweak particles). This is due to lack of long RG running from $\Lambda$ which is argued to be less than $100$ TeV based on naive dimensional analysis (see e.g. \cite{Cacciapaglia:2005da}). Thus we may neglect the loop-induced mass correction for fermions as long as the bulk mass parameter is sufficiently large $\mu_\Psi > 0.1/L$.
The mass of KK fermon ($M_n$) gets contributions from the bulk mass at tree level
as follows
\begin{eqnarray}
M^2_{n} = k_n^2 + \mu^2 \, ~~~ {~\rm for ~} n\ge 1 \, ,
\end{eqnarray}
where
\begin{eqnarray}
&& k_n {\rm ~is~the~}\frac{n+1}{2}-{\rm th~solution~of~} \mu=-k \, cot(k L) \, ,
{\rm ~if~} n=2m-1 \, ,\label{eq:oddmodes}\\
&& k_n = \frac{n}{R}\, , \hspace*{6.75cm}{\rm ~if~} n=2m \, . \label{eq:evenmodes}
\end{eqnarray}
In the MUED limit, $\mu \to 0$, Eqs.~(\ref{eq:oddmodes}-\ref{eq:evenmodes}) both
reduce to $k_n = \frac{n}{R}$. On the other hand, all KK boson masses remain the same,
$\frac{n}{R}$, and show no $\mu$ dependence.
Including EW symmetry breaking and the radiative corrections, a naive estimate gives
\begin{eqnarray}
M_n &\approx& M_n^{tree} \left ( 1 + {\rm ~ loop ~corrections} \right ) \, , \\
M_n^{tree} &=& \sqrt{ k_n^2 + \mu^2 + m_0^2} \, ,
\end{eqnarray}
where $m_0$ is expected from the electroweak symmetry breaking.
\subsection{Constraints from contact interactions}
\label{sec:constraints}
One of the most prominent features of SUED having non-vanishing bulk mass parameters is
the existence of tree level KK number violating interactions.
From $W^3_{2n},B_{2n}$ exchange diagrams we can effectively obtain the contact interaction Lagrangian ${\cal L}_{\rm eff}$ which is stringently constrained by electroweak precision measurements
\cite{Alcaraz:2006mx,Amsler:2008zzb}
\begin{eqnarray}
{\cal L}_{\rm eff}=\sum_{i,j=L,R}\sum_{f}\frac{4\pi}{(\Lambda_{AB}^{ef})^2}
[\bar{e_i}\gamma_\mu e_i][ \bar{f}_j \gamma^\mu f_j] \, .
\label{Eq:eff}
\end{eqnarray}
\TABLE[t]{
\centerline{
\begin{tabular}{c|c|c|c|c}
{}&u&d&$\mu^+\mu^-$ &$\tau^+\tau^-$\\
\hline
$LL$ (TeV) & 10.2 & 6.0 & 12.5 & 8.6 \\
$RR$ (TeV) & 8.3 & 4.3 & 11.9 & 8.2
\end{tabular}
}
\label{table:contact}
\caption{Bounds for contact interaction \cite{Alcaraz:2006mx,Amsler:2008zzb}}.}
Assuming a universal bulk mass $\mu$, the $B_{2n}$ and $W^3_{2n}$ mediated interaction
effective Lagrangian is obtained.
The most stringent bound comes from the contact interaction for $ee\mu\mu$:
\begin{equation}
\bar{e}_L \gamma_\mu e_L \sum_{n} \frac{({\cal F}_{002n})^2}{4}
\left(\frac{g_1^2}{m_{B_{2n}}^2} +\frac{g_2^2}{m_{W^3_{2n}}^2}\right)\bar{\mu}_L \gamma^\mu \mu_L
+
\bar{e}_R \gamma_\mu e_R \sum_{n} ({\cal F}_{002n})^2
\left(\frac{g_1^2}{m_{B_{2n}}^2}\right)\bar{\mu}_R \gamma^\mu \mu_R \, .
\label{Eq:eff2}
\end{equation}
Taking Eqs. (\ref{Eq:eff}-\ref{Eq:eff2}) into account with $m_{B_{2n}}\simeq m_{W^3_{2n}}\simeq (2n)/R$, we obtain the following relations:
\begin{eqnarray}
\frac{1}{\Lambda_{LL}^2}&=&\frac{g_1^2+g_2^2}{64\pi} R^2 \sum_n \frac{({\cal F}_{002n}(\mu L))^2}{n^2},
\\
\frac{1}{\Lambda_{RR}^2}&=&\frac{g_1^2}{16\pi} R^2 \sum_n \frac{({\cal F}_{00,2n}(\mu L))^2}{n^2},
\end{eqnarray}
where the bounds for $\Lambda_{LL}$ and $\Lambda_{RR}$ are given in Table \ref{table:contact}.
We also consider the constraints arising from the dilepton resonance searches at Tevatron \cite{CDFdimuon}
and find that those for $\gamma_2$ give a slightly better constraint on $R^{-1}$
than those for $Z_2$, while $W_2^\pm$ gives a similar limit to that for $Z_2$ \cite{:2007bs}.
In the next Section we include these as well as constraints from contact interactions.
\section{Collider phenomenology}
\label{sec:collider}
A large amount of effort has given into examining the collider aspects of
Universal Extra Dimensions \cite{Appelquist:2000nn}
at LHC \cite{Rizzo:2001sd,Cheng:2002ab,Datta:2005zs,Burdman:2006gy,Dobrescu:2007xf} and
ILC \cite{Battaglia:2005zf,Freitas:2007rh},
as well as its astrophysical implications \cite{Servant:2002aq,Cheng:2002ej,Kong:2005hn,Burnell:2005hm,Dobrescu:2007ec,Arrenberg:2008wy}.
In this Section we would like to investigate the implications of non-vanishing bulk mass in SUED.
\subsection{Level-1 modes}
\label{sec:level1}
We start our discussion with the level-1 KK modes.
Their phenomenology depends on the precise value of the bulk mass and
the radiative corrections to KK masses. Therefore here we would like discuss only
generic features.
A small value for the bulk mass ($0 \leq|\mu L| \ll 1$) would give the similar decay patterns
as in the MUED case. The dominant production is provided by the strong interaction
at a hadron collider, {\it i.e.,} KK quark production ($Q_1Q_1$, $q_1q_1$ and $Q_1 q1$),
KK gluon production ($g_1g_1$) and associated production ($g_1Q_1$, $g_1q_1$).
The $SU(2)_W$-doublet KK quarks ($Q_1$) dominantly decay into
$SU(2)_W$ KK gauge bosons ($Z_1$ and $W_1^\pm$) while
the $SU(2)_W$-singlet KK quarks ($q_1$) decay into KK photon.
SM leptons are obtained from the decay of EW gauge bosons ($Z_1$ and $W_1^\pm$)
to KK leptons.
Here the difference with MUED would be the mass splitting between each mode.
A bulk mass term would increase the mass of the KK fermion,
making the decay product of KK bosons softer than that in MUED.
However the other decay products from the KK fermions to KK bosons become more energetic
due to the increased splitting.
For instance, KK quarks can have mass just below the KK gluon and
the jet from the decay of KK gluon ($g_1 \to Q_1 q$) would be softer
while the jet from the decay of KK quark ($Q_1 \to q Z_1$ or $Q_1 \to q' W_1^\pm$)
becomes harder than in MUED. The same is true for KK leptons and KK gauge bosons.
However we do not expect a dramatic change in the reach
for this model, as long as the decay patterns are the same and the mass splitting is not too small.
The other extreme limit is the case of very large $\mu$, $\mu L \gg 1$
\footnote{The opposite limit (a large negative $\mu L$)
is also interesting as shown in the Appendix.
In this case, taking a very large $R^{-1}$, the masses of the level-1 KK fermions remain at EW scale
while all other KK fermions at higher level are decoupled from the theory due to the large mass splitting
between the level-1 and the level-2 KK fermions. All KK bosons are also very heavy due to the large $R^{-1}$.
Therefore in this limit, {\it the only} available KK modes are the level-1 KK leptons and quarks.
This study will be elsewhere \cite{ongoing}.}.
In this case, all KK fermions become much heavier than KK bosons and
they may not be within the reach of the LHC. KK gauge bosons go through 3 body decays
to the KK photon ($g_1 \to j j \gamma_1$ and $Z_1,W_1^\pm \to f\bar{f}' \gamma_1$) and
production would be via the KK gluon ($g_1g_1$) and EW gauge bosons
($Z_1 W_1^\pm$)
\footnote{$Z_1Z_1$, $\gamma_1\gamma_1$ and $Z_1\gamma_1$ involve KK fermions
in the $t$- and $u$-channels and their production cross sections are negligible
for heavy KK quarks.}.
It is interesting to notice that this situation is similar to
{\it the focus point region of supersymmetry}.
For the moderate ranges of the bulk mass ($|\mu L| \sim 1$),
the gauge bosons may still go through 3-body decays
while the LHC will be able to produce KK quarks and KK leptons.
Unlike MUED, now all KK quarks dominantly decay to the KK gluon since it has the largest coupling.
Therefore the collider signature would be quite jetty.
An interesting possibility is that all KK fermions are very heavy with a large $\mu$
so that they are unobservable but still the KK bosons are within the reach of the LHC. Even in this case, we might expect to observe e.g. the level-2 gluon through its interaction with the level zero Standard Model quarks with a sizable coupling ($g\simeq \sqrt{2}g_s$). Dilepton production through $Z_2$ and $\gamma_2$ is also sizable and provides a golden search channel for split-UED. Detection of a dark matter (DM) particle, on the other hand, is quite challenging because the DM-SM coupling through the level-1 fermion will be highly suppressed by the large KK fermion mass.
We plot cross sections for the gauge boson production at the LHC, as function of mass
in Fig.~\ref{fig:xsection1}, assuming $\mu L \gg 1$; here the curve is
dotted for $g_1 g_1$ (in black), solid for $Z_1W_1^+$ (in red) and dashed for $Z_1W_1^-$ (in blue).
For KK gluon pair production from the $gg$ initial state,
there are $s$-, $t$-, $u$- and four point interaction diagrams, and
all couplings are fixed by $SU(3)_c$ gauge invariance.
There is a contribution from the $q\bar{q}$ initial state but it is smaller than that from $gg$ at the LHC
for the mass range shown.
$Z_1 W_1^\pm$ is produced by $W^\pm$ exchange in the $s$-channel and
contribution from KK quarks in the $t$- and $u$-channels.
Having these 3 diagrams is also necessary by $SU(2)_W$ gauge invariance and
neglecting any of them gives inconsistent results.
However in our case, considering the limit where the KK quarks are much heavier than KK gauge bosons,
their exchanges with a few TeV in mass barely affects the production cross sections.
\FIGURE[t]{
\centerline{
\epsfig{file=FIGURES/xsection.ps, width=8.5cm} }
\caption{\sl Cross section for gauge boson pair production as a function of mass at the LHC (for $\mu L \gg 1$).}
\label{fig:xsection1}}
From Fig.~\ref{fig:couplings} we can immediately read off the patterns of KK particle decay branching fractions.
The level-1 gauge boson only couples to $f_0-f_1$ thanks to KK parity. This coupling becomes less significant as $\mu$ becomes larger in which case the level-1 fermion becomes significantly heavier than
the level-1 gauge boson. Due to the large mass gap, the decay products of the level-1 fermion are
reasonably energetic.
\subsection{Level-2 modes}
\label{sec:level2}
Now we turn to discussion of the level-2 KK modes.
In general level-2 KK fermions ($f_2$) can decay into
either two level-1 KK states $f_1V_1$
or one level-2 and one SM mode $f_0 V_2$
(the branching fraction of $f_2$ to $f_2^\prime V_0$ is suppressed by the small mass splitting between
$f_2$ and $f_2^\prime$.).
In the limit of $\mu L \gg 1$ in split-UED, all KK fermions masses are raised, and
level-2 KK quarks ($Q_2$ and $q_2$) decay to $q g_2 $ and $Q_1g_1$.
Then $g_1$ gives rise to a missing energy signature through a 3 body decay
while $g_2$ can directly decay into two jets and may appear as dijet resonance.
In MUED, the coupling of level-2 resonances to the SM fermions is suppressed by 1-loop.
The branching fractions of electroweak level-2 gauge bosons
into dilepton final states are small partly due to the competing decay modes into
other level-2 and level-1 KK states, and partly due to difference between
the strong and electroweak couplings.
Therefore one has to rely on indirect production of level-2 KK gauge bosons from
the KK gluon and KK quarks to enhance the production cross sections.
The corresponding reach has been estimated in Ref. \cite{Datta:2005zs}.
In SUED, however, this coupling exists at tree level
due to the fermion bulk mass term and
it could be as large as $\sqrt{2}$ times the corresponding SM coupling strength, which makes
dilepton searches in this model promising.
At the same time, the bulk mass increases the mass of KK fermions,
thus reducing the branching fraction of level-2 bosons into other KK states.
The decay width of level-2 KK bosons into SM fermion final states is given by
\begin{eqnarray}
\label{eq:width1} \Gamma &=& \frac{N_c M}{24 \pi} \left [
\Big ( g_L^2 + g_R^2 \Big ) \Big ( 1- \frac{m^2}{M^2} \Big )
+ 6 g_L g_R \frac{m^2}{M^2} \right ] \sqrt{1 - \frac{4m^2}{M^2}} \\
&=& \label{eq:width2}
\frac{N_c M }{ 24 \pi} \Big ( g_L^2 + g_R^2 \Big ) ~~~~{\rm for ~ M \gg m} \, .
\end{eqnarray}
\FIGURE[t]{
\centerline{
\epsfig{file=FIGURES/widths.ps, width=7.5cm} \hspace*{0.1cm}
\epsfig{file=FIGURES/BRZ2.ps,width=7.5cm} }
\caption{\sl (a) The ratio of widths of level-2 KK bosons to their masses and
(b) branching fractions of 1 TeV $Z_2$, as a function of $\mu L$.}
\label{fig:width_br}}
The 1-loop correction is expected to be the smallest for $\gamma_n$
among all other KK states at the same level, and in fact, KK photons receive
negligible correction from RG running, making the lightest KK photon
a viable dark matter candidate.
Therefore the decay channels of $\gamma_2$ into $f_1$-$f_1$ or $f_0$-$f_2$ are closed
and $\gamma_2$ always can appear as a resonance.
As in Eqs. (\ref{eq:width1})-(\ref{eq:width2}), the width dependence on the SM fermion mass
is negligible even for the top quark, if the resonance is heavy enough.
In this case, the ratio of the total width to its mass becomes
mass-independent.
The total width of $Z_2$ ($\gamma_2$) is then $\sim$ 7\% (3.5\%) of its mass
as $\mu$ increases, as shown in Fig.~\ref{fig:width_br}(a),
while in MUED, the widths of level-2 KK bosons are
much less than 1\% \cite{Datta:2005zs}.
This makes it challenging to probe double resonances which are separated from each other in mass
by the small 1-loop corrections (7\% or so).
The branching fractions of the $\gamma_2$ are $\mu$-independent for a universal bulk mass,
which is the case that we consider. They are 25\% in dilepton, 36.7\% in dijet,
4.2 in $b\bar{b}$, 14\% $t\bar{t}$ and 12.5\% in $\tau \bar\tau$.
The $\gamma_2$ decays invisible through SM neutrinos 7.5\% of the time.
Notice that the branching fraction into dilepton channel is about 20 times larger than
in case of MUED.
The decay of $Z_2$ is somewhat more complicated than $\gamma_2$ due to the slightly larger
1-loop correction, which we assume to be about 7\% as in the MUED.
In this case the decay modes to other KK states remain open.
Without knowing the exact 1-loop mass corrections for all KK particles,
it is impossible to compute its total width and branching fractions.
For a rough estimate (only for this purpose), we assume that
KK fermions only gets corrections from the bulk mass
while $Z_2$ gets heavier by 7\% from RG running.
This is certainly an inconsistent setup.
However, 1-loop corrections to KK fermion masses are known to be merely a few percents
(1\% for singlet KK fermions and 3\% for doublet KK fermions), and
for a large $\mu$, bulk mass enhances its mass by quite a large amount and
this 1-loop contribution becomes negligible.
This approximation is valid in a broad range of $\mu$.
Given that, one can compute the partial widths of $Z_2$ in three different channels and
the results are shown in Fig.~\ref{fig:width_br}(b).
The level-2 KK fermion does not get correction from the bulk mass, but the
$f_0$-$f_2$-$Z_2$ coupling becomes smaller, as shown in Fig.~\ref{fig:couplings},
making the relevant branching fraction smaller for large $\mu$.
The same is true for $f_1$-$f_1$-$Z_2$ while $f_0$-$f_0$-$Z_2$ coupling behaves
in the opposite manner.
Moreover, unlike $f_2$, the level-1 KK modes get heavier as the $\mu L$ increases, and
at some value of $\mu$, the $Z_2$ decay to $f_0f_2$ gets closed.
In Fig.~\ref{fig:width_br}(b), this transition value of $\mu L$ is about 0.6 for a 1 TeV $Z_2$.
Having $f_0$-$f_0$-$Z_2$ as a dominant channel, it is straightforward to compute
the relevant branching fractions. Since $Z_2$ ($W_2^3$) couples to SM pair with the same strength,
one needs to count relevant degrees of freedom.
The branching fractions are 1/24 in $\tau\bar\tau$, 1/12 in dilepton,
1/2 in dijet and 1/8 in either $b\bar{b}$ and $t \bar{t}$.
$Z_2$ also can decay invisibly 1/24 of the time.
\subsection{The LHC reach for $\gamma_2$ and $Z_2$ in dilepton channel}
\label{sec:resonances}
We simulate dilepton resonances in the Split-UED at the LHC with
$\sqrt{s}$=10 TeV, using a private Monte-Carlo generator.
We assume the mass splitting between two bosons ($\gamma_2$ and $Z_2$) is given by
$M_{Z_2}=1.07 M_{\gamma_2}$ and $M_{\gamma_2}\approx \frac{2}{R}$ as in MUED.
We include both $\gamma_2$ and $Z_2$ in the dilepton signal and
use the CTEQ6.6 PDF with NLO K-factor.
The leptons from the decay of these KK bosons are highly energetic and
can easily pass triggers.
For heavy resonances, the energy resolution is better in electron final states
than in muon final states, and hence we consider electron final state
with 1\% mass resolution smearing.
$|\eta| < 2.5$ and $M_{\ell\ell} > M_{\gamma_2} - 500$ GeV are imposed as cuts and
we only count events with dielectron masses greater than 0.8 of $M_{\gamma_2}$.
The dominant background is Drell-Yan, and $t\bar{t}$ and fakes are expected to
be significantly smaller.
In all cases the background is smaller than the signal by a factor of $\sim$100.
Fig.~\ref{fig:lhc1}(a) shows the required luminosity to observe at least 10
signal events as a function of $\mu L$ for several values of masses.
The LHC should able to cover the large parameter space
(up to $M_{V_2} \sim 1.5$ TeV for $\mu L \ge 1$)
even with the early data at the level of $\sim$100 pb$^{-1}$ or less.
With the integrated luminosity of $\sim$100 fb$^{-1}$, the most of parameter space
would be probed, setting limit on the bulk mass and the mass of the KK gauge boson.
The expected number of signal events is plotted in the $\mu L$ versus $R^{-1}$ plane
in Fig.~\ref{fig:lhc1}(b), for ${\cal L}=1$ fb$^{-1}$.
The shaded region in the left side (in yellow) is a projected Tevatron exclusion
at 95\% C.L. assuming 10 fb$^{-1}$ \cite{CDFdimuon}.
The limit on $R^{-1}$ from $\gamma_2$ gives the best exclusion since it is lighter
than $Z_2$ and $W_2^\pm$ by 7\%, and
constrains on $Z_2$ and $W_2^\pm$ are comparable, and hidden below that from $\gamma_2$.
The other shaded area in the left upper corner (in green) is EW constraint from LEP II
considering contact interaction in SUED, as discussed in Section~\ref{sec:constraints}.
\FIGURE[t]{
\centerline{
\epsfig{file=FIGURES/Lumi10.ps, width=7.6cm} \hspace*{-0.2cm}
\epsfig{file=FIGURES/muL_Rinv.ps,width=7.6cm} }
\caption{\sl The luminosity required to obtain 10 events as a function of $\mu L$
for several values of masses in (a) and the number of signal events
in the $\mu L$ versus $R^{-1}$ plane in (b),
for $\sqrt{s}$=10 TeV, ${\cal L}=1$ fb$^{-1}$, $M_{Z_2}=1.07 M_{\gamma_2}$.
In all cases the background is smaller by a factor of $\sim$100.
We used the CTEQ6.6 with NLO K-factor and 1\% mass resolution smearing.
In obtaining the result we only count events with dilepton masses greater than 0.8 $\times M_{\gamma_2}$.
}
\label{fig:lhc1}}
In Fig.~\ref{fig:invmass} invariant mass distributions are shown for (a)
$R^{-1}=1$ TeV, $\sqrt{s}= 14$ TeV and ${\cal L}=100$ fb$^{-1}$
and (b) $R^{-1}=0.75$ TeV, $\sqrt{s}= 10$ TeV and ${\cal L}=1$ fb$^{-1}$.
For both cases, we assume $\mu L \gg 1$.
The yellow histogram is the SM background while
the red histogram includes both signal and backgrounds.
At the early phase of LHC, one may able to see a bump and get to resolve it into
double resonance structure as more data gets accumulated.
Notice the negative interference between the SM background and the KK signal,
which implies the relative sign difference in the couplings.
\FIGURE[t]{
\centerline{
\epsfig{file=FIGURES/split.ps, width=7.5cm}\hspace*{0.2cm}
\epsfig{file=FIGURES/split4.ps,width=7.5cm} }
\caption{\sl Invariant mass distributions at the LHC for (a)
$R^{-1}=1$ TeV, $\sqrt{s}=14$ TeV and ${\cal L}=100$ fb$^{-1}$
and (b) $R^{-1}=0.75$ TeV, $\sqrt{s}=10$ TeV and ${\cal L}=1$ fb$^{-1}$.
The yellow histogram is the SM background while
the red histogram includes both signal and backgrounds.}
\label{fig:invmass}}
\section{Conclusions}
\label{sec:conclusion}
The Minimal Universal Extra Dimensions scenario has received great attention.
Recently non-vanishing bulk fermion masses have been introduced
without spoiling the virtue of KK-parity.
The fermion profiles are no longer simple sine/cosine functions and depend upon
the specific values of bulk parameters.
The profiles of fermions are split along the extra dimensions
while the wave functions of the bosons remain the same as in UED.
A simple introduction of a KK-parity conserving bulk fermion mass has significant influences
on collider aspects as well as astrophysical implications of UED.
For instance, the DM annihilation fraction into certain SM fermion pairs is
either enhanced or reduced (compared to the MUED case) so that one can perhaps explain
the PAMELA positron excess while suppressing the anti-proton flux.
In this paper, we have concentrated on collider phenomenology of
Split Universal Extra Dimensions.
We have revisited the KK decomposition in detail and
analyzed wave function overlaps to compute relevant couplings for collider studies.
We have discussed general collider implication for level-1 KK modes and
level-2 KK with non-zero bulk mass and have computed LHC reach for the
EW level-2 KK bosons, $\gamma_2$ and $Z_2$, in the dilepton channel.
The LHC should able to cover the large parameter space
(up to $M_{V_2} \sim 1.5$ TeV for $\mu L \ge 1$)
even with early data assuming $\sim$100 pb$^{-1}$ or less.
The existence of double resonances is one essential feature arising from extra dimensional models.
Whether or not one can see double resonances depends
both on how degenerate the two resonances are and on the mass resolution of the detector.
The very high $P_T$ from the decay makes resolution in dimuon channel worse than in
dielectron final state. This is because one can reconstruct electron from ECAL but
muon momentum reconstruction relies on its track, which is barely curved in this case.
Further indication for SUED might be the discovery of
$W^\prime$-like signature of mass close to $Z_2$.
The MUED predicts a somewhat lower event rate due to 1-loop suppressed
coupling of level-2 bosons to SM fermion pair, while it exists at tree level in SUED.
Therefore in UED, one has to rely on indirect production of level-2 bosons, whose
collider study requires complete knowledge of the model:
{\it the mass spectrum and all the couplings}.
On the other hand, in the large $\mu$ limit of SUED,
the dependence on mass spectrum is diminished since level-2 KK bosons decay
only into SM fermion pairs.
This allows us to estimate the signal rate from their direct production,
so that they can be discovered at the early phase of the LHC.
The indirect production mechanism only increases production cross sections,
improving our results.
Once a discovery has been made, one should try to reconstruct events and
do further measurements such as spin and coupling determination,
with more accumulated data \cite{Li:2009xh,Petriello:2008zr,Rizzo:2009pu},
which might discriminate KK resonances from other $Z^\prime$ models.
The coupling measurement is directly related to the determination of the bulk masses.
A challenging issue might be the existence of two resonances which are rather close to each other.
\bigskip
\acknowledgments
We thank J. Shu and K. Wang for discussion on the forward-backward asymmetry and also
thank C. Cs\'aki, J. Heinonen and J. Hubisz for helpful discussion.
S. Park is supported by the World Premier International Research Center Initiative
(WPI initiative) by MEXT and also supported by the Grant-in-Aid for scientific
research (Young Scientists (B) 21740172) from JSPS, Japan.
K. Kong and T. G. Rizzo are supported in part by the DOE under contract DE-AC02-76SF00515.
|
\section{Introduction}
At the beginning of 90's, L.Tartar \cite{Tar} and P.Gerard
\cite{Ger} independently introduced the $H$-measures (microlocal
defect measures). The $H$-measures appeared to be very powerful tool
in many fields of mathematics and physics (see randomly chosen
\cite{?7, ?5, ?2, ?9, ?3, JMSpa, ?4, ?1}). They are given by the
following theorem:
\begin{theorem}\cite{Tar}
\label{tbasic1} If $(u_n)=((u_n^1,\dots, u_n^r))$, is a sequence in
$L^2(I\!\!R^d;I\!\!R^r)$ such that $u_n\rightharpoonup 0$ in
$L^2(I\!\!R^d;I\!\!R^r)$, then there exists its subsequence $(u_{n'})$ and a
positive definite matrix of complex Radon measures
$\mu=\{\mu^{ij}\}_{i,j=1,\dots,d}$ on $I\!\!R^d\times S^{d-1}$ such that
for all $\varphi_1,\varphi_2\in C_0(I\!\!R^d)$ and $\psi\in C(S^{d-1})$:
\begin{equation}
\label{basic1}
\begin{split}
\lim\limits_{n'\to \infty}\int_{I\!\!R^d}&(\varphi_1
u^i_{n'})(x)\overline{{\mathcal A}_\psi(\varphi_2
u^j_{n'})(x)}dx=\langle\mu^{ij},\varphi_1\overline{\varphi_2}\psi
\rangle\\&= \int_{I\!\!R^d\times
S^{d-1}}\varphi_1(x)\overline{\varphi_2(x)}\psi(\xi)d\mu^{ij}(x,\xi),
\ \ i,j=1,\dots,d,
\end{split}
\end{equation}where ${\mathcal A}_\psi$ is a multiplier operator with the symbol $\psi\in C^\kappa(S^{d-1})$.
\end{theorem}
The complex matrix Radon measure $\{\mu^{ij}\}_{i,j=1,\dots r}$
defined in the previous theorem we call the {\em $H$-measure}
corresponding to the subsequence $(u_{n'})\in L^2(I\!\!R^d;I\!\!R^r)$.
The $H$-measures describe a loss of strong $L^2$ precompactness for
the corresponding sequence $(u_n)\in L^2(I\!\!R^d;I\!\!R^r)$. In order to
describe loss of $L^1_{loc}$ precompactness for a sequence $(u_n)\in
L_{loc}^p(I\!\!R^d)$, $p>1$, we have extended the notion of the
$H$-measures in \cite{MPB} as follows.
\begin{theorem} \cite{MPB}
\label{tbasic2} Assume that $(u_n)$, is a sequence in
$L_{loc}^{p}({I\!\!R}^{d})$, $p>1$, such that $u_n\rightharpoonup 0$,
$n\to \infty$, in $L_{loc}^{p}(I\!\!R^{d})$, $\beta>0$. Assume that
$(v_n)$ is a bounded sequence in $L^\infty(I\!\!R^{d})$.
Then, there exist subsequences $(u_{n'})$ and $(v_{n'})$ of the
sequences $(u_{n})$ and $(v_n)$, respectively, such that there
exists a complex valued distribution $\mu\in {\mathcal D}'(I\!\!R^{d'}
\times S^{d-1})$, such that for every $\varphi_1, \varphi_2\in
C_0(I\!\!R^{d})$ and $\psi\in C^{\kappa}(S^{d-1})$, $\kappa>d/2$,
$\kappa\in I\!\!N$:
\begin{equation}
\label{basic2}
\begin{split}
\lim\limits_{n'\to \infty}\int_{I\!\!R^{d}}(\varphi_1 u_{n'})(x)&{{\mathcal
A}_{\psi}(\varphi_2 v_{n'})(x)}dx=\langle\mu ,\varphi_1\varphi_2\psi
\rangle,
\end{split}
\end{equation}where ${\mathcal A}_{\psi}:L^p(I\!\!R^d)\to L^p(I\!\!R^d)$ is a multiplier operator with the symbol $\psi\in C^\kappa(S^{d-1})$.
\end{theorem}
The first commutation lemma was one of the key point in the proof
of Theorem \ref{tbasic1} as well as Theorem \ref{tbasic2} (with a
simple modification; see \cite[Lemma 14]{MPB}). It is stated as
follows:
\begin{lemma}\cite[Lemma 14]{MPB}
\label{scl} (First commutation lemma) Let $a\in C(S^{d-1})$ and
$b\in C_0(I\!\!R^d)$. Let ${\mathcal A}$ be a multiplier operator with the
symbol $a$, and $B$ be an operator of multiplication given by the
formulae:
\begin{align*}
&{\mathcal F}({\mathcal A}u)(\xi)=a(\frac{\xi}{|\xi|}){\mathcal F}(u)(\xi) \ \ a.e. \ \ \xi\in I\!\!R^d,\\
&Bu(x)=b(x)u(x) \ \ a.e. \ \ x\in I\!\!R^d,
\end{align*} where ${cal F}$ is the Fourier transform.
Then $C={\mathcal A}B-B{\mathcal A}$ is a
compact operator from $L^p(I\!\!R^d)$ into $L^p(I\!\!R^d)$, $p>1$.
\end{lemma}
The proof of the lemma heavily relies on the fact that the function
$a$ is actually defined on the unit sphere. Recently, two new
variants of the $H$-measures were introduced -- the parabolic
$H$-measures \cite{Ant2} and ultra-parabolic $H$-measures
\cite{JMSpa}. In both cases a variant of the first commutation lemma
is needed, and in both cases its proof is based on the fact that a
symbol $a$ of appropriate multiplier ${\mathcal A}$ is defined on a
smooth, bounded, simply connected manifold of codimension one.
In order to motivate our variant of the first commutation lemma,
notice that from the proof of Theorem \ref{tbasic1} (see
\cite{Tar}), it follows that we need to "commute" ${\mathcal
A}(\varphi_2 u_n)$ by $\varphi_2 {\mathcal A}(u_n)$, where $(u_n)$ is
the sequence bounded in $L^2(\Omega)$, which was exactly done in
Lemma \ref{scl}. Similarly, in Theorem \ref{tbasic2} (see
\cite{MPB}), we need to "commute" ${\mathcal A}(\varphi_2 v_n)$ by
$\varphi_2 {\mathcal A}(v_n \chi_{{\rm supp \varphi_2}})$, where
$\chi_{V}$ is the characteristic function of the set $V$, and
$(v_n)$ is the sequence bounded in $L^\infty(\Omega)$. Therefore, it
is enough to prove that the commutator $C$ is compact operator from
$L^\infty_0(\Omega)$ into $L^p_{loc}(\Omega)$, $p>1$. We state:
\begin{lemma}
\label{scl'} Let $a\in C^{\kappa}(I\!\!R^d)$, $\kappa=\lfloor
d/2\rfloor+1$, and $b\in C_0(I\!\!R^d)$. Suppose that for some constant
$k>0$ and for any real number $r>0$
\begin{equation}
\label{cond_1'} \int_{\frac{r}{2}\leq \|\xi\|\leq
r}|D_{\xi}^{\alpha}a(\xi)|^2d\xi\leq k^2r^{d-2n(\alpha)}
\end{equation} holds for every $\alpha=(\alpha_1,\dots, \alpha_d)\in I\!\!N_0^d$ satisfying $n(\alpha)=\sum\limits_{i=1}^d\alpha_i\leq \kappa$.
Let ${\mathcal A}$ be a multiplier operator with the symbol $a$, and $B$
be an operator of multiplication given by the formulae:
\begin{align}
\label{oper_1}
&{\mathcal F}({\mathcal A}u)(\xi)=a(\xi){\mathcal F}(u)(\xi) \ \ a.e. \ \ \xi\in I\!\!R^d,\\
\label{oper_2} &Bu(x)=b(x)u(x) \ \ a.e. \ \ x\in I\!\!R^d,
\end{align} where ${\mathcal F}$ is the Fourier transform.
Then $C={\mathcal A}B-B{\mathcal A}$ is a compact operator from
$L_0^\infty(I\!\!R^d)$ into $L_{loc}^{p_0}(I\!\!R^d)$ for every
$1<p_0<\infty$.
\end{lemma}
\begin{remark}
We hope that the lemma could serve for defining variants of the
$H$-measures adapted to equations which change type (such as
non-strictly parabolic equations).
\end{remark}
\begin{proof}
On the first step notice that $a$ satisfies conditions of the
H\"{o}rmander-Mikhlin theorem (see \cite{Oki, Gra}). Therefore, for
every $p>1$ there exists a constant $k_p$ such that $\|{\mathcal
A}\|_{L^p\to L^p}\leq k_p$. Thus,
$$
\|C\|\leq 2 k_p \|b\|_{L^\infty(I\!\!R^d)}.
$$
Then, notice that we can assume $b\in C^1_0(I\!\!R^d)$. Indeed, if we
assume merely $b\in C_0(I\!\!R^d)$ then we can uniformly approach the
function $b$ by a sequence $(b_n)\in C^1_0(I\!\!R^d)$. The corresponding
sequence of commutators $C_n={\mathcal A} B_n-B_n {\mathcal A}$, where
$B_n(u)=b_n u$, converges in norm toward $C$. So, if we prove that
$C_n$ are compact for each $n$, the same will hold for $C$ as well.
Then, fix a real non-negative function $\omega$ with a compact
support and total mass one. Take the characteristic function
$\chi_{B(0,2)}$ of the ball $B(0,2)\subset I\!\!R^d$ and denote:
$$
\chi(x)=\chi_{B(0,2)}\star
\frac{1}{\varepsilon^d}\Pim\limits_{i=1}^d\omega(\frac{x_i}{\varepsilon})
$$for an $\varepsilon>0$ small enough so that we have
$\chi(x)=1$ for $x\in B(0,1)$, and $(1-\chi)\equiv 1$ out of the
ball $B(0,3)$.
Next, notice that $ {\mathcal A}={\mathcal A}_{a \chi}+{\mathcal A}_{a
(1-\chi)}$, where ${\mathcal A}_{a \chi}$ is a multiplier operator with
the symbol $a \chi$, and ${\mathcal A}_{a (1-\chi)}$ is a multiplier
operator with the symbol $a (1-\chi)$. Accordingly,
\begin{align*}
C={\mathcal A}B-B{\mathcal A}&={\mathcal A}_{a \chi}B-B{\mathcal A}_{a \chi}
\\&+{\mathcal A}_{a (1-\chi)} B-B {\mathcal A}_{a (1-\chi)}=
C_{a \chi}+C_{a (1-\chi)},
\end{align*}where
$C_{a \chi}= {\mathcal A}_{a \chi}B-B{\mathcal A}_{a \chi}$ and $C_{a
(1-\chi)}={\mathcal A}_{a (1-\chi)} B-B {\mathcal A}_{a (1-\chi)}$.
First, consider the commutator $C_{a \chi}$. Notice that since $a
\chi_{B(0,2)}$ has a compact support, the multiplier ${\mathcal A}_{a
\chi}$ is actually the convolution operator with the kernel
$\psi=\bar{\mathcal F}(a\chi)\in L^2(I\!\!R^d)$, where $\bar{\mathcal F}$ is the
inverse of the Fourier transform ${\mathcal F}$:
\begin{equation}
\label{oper_1'} {\mathcal A}_{a \chi}(u)=\psi\star u, \ \ u\in
L^\infty_0(I\!\!R^d).
\end{equation} Therefore, we can state that
\begin{equation}
\label{comm_1}
\begin{split}
&\qquad\qquad C_{a \chi}u(x)=\int_{I\!\!R^d}\left(b(x)-b(y)\right)\psi(x-y)u(y)dy,\\
&\text{is a compact operator from $L_0^\infty(I\!\!R^d)$ into
$L_{loc}^{p_0}(I\!\!R^d)$, $p_0\geq 2$.}
\end{split}
\end{equation}
Indeed, take an arbitrary bounded sequence $(u_n)\in
L_0^\infty(I\!\!R^d)$ such that $u_n\rightharpoonup 0$ weak-$\star$ in
$L^\infty(I\!\!R^d)$ and ${\rm supp}u_n\subset \hat{V}\subset\subset
I\!\!R^d$, for a relatively compact set $\hat{V}$. In order to prove
that $C_{a \chi}$ is compact, it is enough to prove that $C_{a
\chi}u_n$ strongly converges to zero in $L_{loc}^{p_0}(I\!\!R^d)$,
$p_0>1$.
Since $\psi\in L^2(I\!\!R^d)$ we also have $\psi\in L^1_{loc}(I\!\!R^d)$.
Thus, it holds for every fixed $x\in I\!\!R^d$
\begin{equation}
\label{comm_1-c1} C_{a
\chi}u_n(x)=\int_{I\!\!R^d}\left(b(x)-b(y)\right)\psi(x-y)u_n(y)dy\to 0,
\ \ n\to 0.
\end{equation} Next, since the sequence $(u_n)$ has compact support, we also have:
\begin{equation}
\label{comm_1-c2} |C_{a \chi}u_n(x)|\leq \hat{C},
\end{equation}for a constant $\hat{C}$ depending on the support of the sequence $(u_n)$ as well as $L^2$ norm of the kernel $\psi$.
Combining \eqref{comm_1-c1} and \eqref{comm_1-c2} with the Lebesgue
dominated convergence theorem, we see that for an arbitrary
relatively compact $V\subset\subset I\!\!R^d$ and every $p_0>0$, it
holds:
\begin{equation}
\label{comm_1-c3} \int_V|C_{a \chi}u_n(x)|^{p_0}dx\to 0, \ \ n\to
\infty,
\end{equation}proving \eqref{comm_1}.
In order to prove that $C_{a (1-\chi)}$ is compact, we need more
subtle arguments basically involving techniques from the proof of
the Hormander-Mikhlin theorem from e.g. \cite{Oki}.
So, let $\Theta$ be a non-negative infinitely differentiable
function supported by $\{\xi\in R^n:\frac{1}{2}\leq\|\xi\|\leq2\}$.
Also let $ \theta(\xi)=\Theta(\xi)\bigg/
\sum_{j=-\infty}^{\infty}\Theta(2^{-j}\xi)$, $\theta(0):=0$. Then,
$\theta$ is non-negative, it is supported by $\{\xi\in
R^n:\frac{1}{2}\leq \|\xi\|\leq2\}$, it is infinitely
differentiable, and is such that if $\xi\neq 0$, then $
\sum_{j=-\infty}^{\infty}\theta(2^{-j}\xi)=1. $
Now, let $a_j(\xi)=a(\xi)(1-\chi(\xi))\theta(2^{-j}\xi)$, $j>0$.
Then, ${a}_j$ has support in the set $ \{\xi \in R^n:
2^{j-1}\leq\|\xi\|\leq 2^{j+1}\}$, $j>0,$ and $
a(\xi)(1-\chi(\xi))=\sum_{j=0}^\infty {a}_j(\xi)$. Furthermore, it
holds for any $p\in \{1,\dots,d\}$, $\alpha_p\in I\!\!N_0$ that
$$
\frac{\partial^{\alpha_p}}{\partial^{\alpha_p}\xi_p}(a(\xi)(1-\chi(\xi))\theta(2^{-j}\xi))=\sum_{l=0}^{\alpha_p}{\alpha_p\choose
l}\frac{\partial^l a(\xi)}{\partial^l \xi_p}
\frac{\partial^{\alpha_p-l}(1-\chi(\xi))\theta(2^{-j}\xi)}{\partial^{\alpha_p-l}\xi},
$$ so that, with suitable bounded functions $a_{\beta\gamma}$,
$\beta+\gamma=\alpha$, $\alpha\in I\!\!N_0^d$, we have $
D_\xi^{\alpha}{\phi}_j(\xi)=\sum_{\beta+\gamma=
\alpha}a_{\beta\gamma}2^{-jn(\beta)}D_\xi^{\gamma}{a}(\xi)$.
From here, on applying hypothesis \eqref{cond_1'} with $r=2^j$, it
follows from the Minkowski inequality that
\begin{align}
\label{corr_ns1} \int_{R^d}|D^{\alpha}_\xi{a}_j(\xi)|^2d\xi &\leq
p^2_0\sum_{\beta+\gamma=\alpha}a_{\beta\gamma}2^{-jn(\beta)}\int_{2^{j-1}
\leq\|\xi\|\leq 2^{j+1}}|D^{\gamma}_\xi{a}(\xi)|^2d\xi
\\&\leq C k^2
2^{j(d-2n(\alpha))}, \nonumber
\end{align}
where $C$ is a constant independent on $k$.
Denote by $\bar{a}_j=\bar{{\mathcal F}}(a_j)(x)$, $x\in I\!\!R^d$, the inverse
Fourier transform of the function $a_j$. From \eqref{corr_ns1}, the
Cauchy-Schwartz inequality, Plancharel's theorem and the well known
properties of the Fourier transform, for every $s>0$ it holds (see
also the proof of \cite[Theorem 7.5.13]{Oki}):
\begin{align}
\label{aux_5}
\int_{\|x\|>s}|\bar{a}_j(x)|dx&\leq\bigg(\int_{\|x\|\geq s}
\|x\|^{-2\kappa}dx\bigg)^{{1}/{2}} \bigg(\int_{\|x\|\geq s}
\|x\|^{2\kappa}|\bar{a}_j(x)|^2d\xi\bigg)^{{1}/{2}}
\\&\leq \left(\frac{2\pi^{d-1}s^{d-2\kappa}}{2\kappa-d} \right)^{1/2} \left(d^\kappa \sum\limits_{i=1}^d
\int_{I\!\!R^d}|x_i|^{2\kappa}|\bar{a}_j(x)|^2dx\right)^{1/2}
\nonumber\\
&= \left(\frac{2\pi^{d-1}s^{d-2\kappa}}{2\kappa-d} \right)^{1/2}
\left(d^\kappa \sum\limits_{i=1}^d
\int_{I\!\!R^d}|D^{\kappa}_{\xi_i}{a}_j(\xi)|^2d\xi\right)^{1/2}
\nonumber\\
&\leq C_1 k (2^j s)^{(\frac{1}{2}d-\kappa)}, \nonumber
\end{align}where $C_1$ depends only on the functions $\theta$ and $\chi$.
Next, consider $ \bar{A}_n(x)=\sum_{j=0}^n\bar{a}_j(x)$, $x\in
I\!\!R^d$. For an arbitrary fixed $s>0$, the series
$\sum_{j=0}^n\bar{a}_j(x)$ is absolutely convergent in
$L^1(I\!\!R^d\backslash B(0,s))$. Indeed,
\begin{align}
\label{sep138} \|\bar{A}_n(x)\|_{L^1(I\!\!R^d\backslash B(0,s))}&\leq
\sum_{j=0}^n\|\bar{a}_j\|_{L^1(I\!\!R^d\backslash B(0,s))}\\&\leq C_1 k
s^{(\frac{1}{2}d-\kappa)}\sum\limits_{j=0}^n
2^{j(\frac{1}{2}d-\kappa)}\leq C_3 <\infty, \nonumber
\end{align}for a constant $C_3>0$, since $\frac{1}{2}d-\kappa < 0$.
Thus, for every $s>0$ there exists $\bar{A}_s\in L^1(I\!\!R^d\backslash
B(0,s))$ such that
\begin{equation}
\label{sep148} \sum_{j=0}^\infty\bar{a}_j(x)=\bar{A}_s(x), \ \ x\in
I\!\!R^d\backslash B(0,s).
\end{equation}
Furthermore, for an odd $d$, from \eqref{aux_5} we have
\begin{equation}
\label{sep128} \int_{\|x\|<s}\|x\|\cdot|\bar{A}_n(x)|dx \to 0, \ \
s\to 0,
\end{equation}while for an even $d$ conclusion \eqref{sep128}
follows from:
\begin{align*}
\int_{\|x\|<s}\|x\|\cdot|\bar{A}_n(x)|dx&\leq
\left(\int_{\|x\|<s}\|x\|^{1-2\kappa}
\right)^{1/2}\left(\int_{\|x\|<s}\|x\|^{2\kappa-2}|\bar{a}_j(x)|^2dx\right)^{1/4}\times\\&\times
\left(\int_{\|x\|<s}\|x\|^{2\kappa}|\bar{a}_j(x)|^2dx\right)^{1/4}
\leq C_4 s^{1/2}.
\end{align*}
Now, take the convolution operator:
$$
{\mathcal A}_n(u)=\bar{A}_n\star u, \ \ u\in L^\infty_0(I\!\!R^d).
$$ and consider the commutator
$ C_n={\mathcal A}_n B-B {\mathcal A}_n$. It holds:
\begin{equation*}
C_n(u)(x)=-\int_{I\!\!R^d}\bar{A}_n(x-y)(b(x)-b(y))u(y)dy.
\end{equation*}Given a fixed $s>0$, rewrite $C_n(u)$ in the following way:
\begin{align*}
&C_n(u)(x)=\int_{I\!\!R^d}\bar{A}_n(x-y)(b(x)-b(y))u(y)dy\\
&=\int_{\|x-y\|>s}\bar{A}_n(x-y)(b(x)-b(y))u(y)dy+\int_{\|x-y\|\leq
s}\bar{A}_n(x-y)(b(x)-b(y))u(y)dy.
\end{align*} From here, combining $b\in C^1_0(I\!\!R^d)$ with \eqref{sep148} and \eqref{sep128}, we conclude for an arbitrary relatively compact
$V\subset\subset I\!\!R^d$:
\begin{equation}
\limsup\limits_{n\to \infty} \| C_n(u)(x)\|_{L^{p_0}(V)}\leq
\|\int_{\|x-y\|>s}
\bar{A}_s(x-y)(b(x)-b(y))u(y)dy\|_{L^{p_0}(V)}+o_s(1),
\end{equation}where
\begin{align*}
o_s(1)&=\int_{\|x-y\|\leq s}\bar{A}_n(x-y)(b(x)-b(y))u(y)dy\\&
=\int_{\|x-y\|\leq
s}\|x-y\|\bar{A}_n(x-y)\frac{(b(x)-b(y))}{\|x-y\|}u(y)dy
\to^{\eqref{sep128}} 0, \ \ \varepsilon\to 0.
\end{align*} Furthermore, arguing as for \eqref{comm_1}, we infer that the operator
\begin{equation}
\label{nk1} u\mapsto \int_{\|x-y\|>s}\bar{A}_s(x-y)(b(x)-b(y))u(y)dy
\end{equation}is a compact operator from $L^\infty_0(I\!\!R^d)$ to
$L^{p_0}_{loc}(I\!\!R^d)$ for an arbitrary $p_0>1$.
Next, notice that we have for any $u\in L^2$:
\begin{align*}
&\|(C_n-C_{\phi (1-\chi)})(u)\|_{L^2}\\&\leq \|({\mathcal A}_n-{\mathcal
A}_{a (1-\chi)}(bu)\|_{L^2}+\|b\|_{\infty} \|({\mathcal A}_n-{\mathcal
A}_{\phi
(1-\chi)}(u)\|_{L^2}\\
&=\|(A_n-a(1-\chi)){\mathcal F}(bu)\|_{L^2}+\|b\|_{\infty}\|(A_n-a(1-\chi)){\mathcal F}(u)\|_{L^2}
\to 0, \ \ n\to\infty,
\end{align*}
and from here and the Fatou lemma:
\begin{align*}
&\|C_{a (1-\chi)}(u)\|_{L^{p_0}(V)}\leq
\limsup\limits_{n\to\infty}\bigg(\|\int_{\|x-y\|\geq
s}\bar{A}_n(x-y)(b(x)-b(y))u(y)dy\\&+\int_{\|x-y\|<
s}|x-y|\bar{A}_n(x-y)\frac{(b(x)-b(y))}{|x-y|}u(y)dy\|_{L^{p_0}(V)}
\bigg)\\&\leq \|\int_{\|x-y\|\geq
s}\bar{A}_s(x-y)(b(x)-b(y))u(y)dy\|_{L^{p_0}(V)}+ o_s(1),
\end{align*} where $o_s(1)$ denotes a quantity tending to zero as $s\to 0$, and appears here due to
\eqref{sep128}. From here and \eqref{nk1}, it follows that
$C_{a(1-\chi)}$ is a compact operator since it can be estimated by a
sum of a compact operator, and an operator bounded by an arbitrary
small constant.
Thus, we see that $C$ can be represented as the sum of two compact
operators $C_{a \chi}$ and $C_{a(1-\chi)}$, which means that $C$ is
a compact operator itself.
This concludes the proof.
\end{proof}
|
\section*{Introduction}\label{Intro}
The random sequential adsorption model has been used as a paradigm to study irreversible phenomena like deposition and adsorption of particles onto a substrate in systems where adsorbed particles do not significantly diffuse within the experimental time scale \cite{Bartelt91b,Privman00a,Evans93,Privman94,Privman97,Cadilhe07,Araujo08}.
Nonetheless, the random sequential adsorption model has also been applied to the modeling of apparently dissimilar systems like the cases of the classical works on intramolecular reactions taking place in polymers \cite{Flory39,Gonzalez74} and the famous car parking problem \cite{Renyi58,Renyi63}.
The random sequential adsorption model has been also applied to study the irreversible adsorption of colloids \cite{Onoda86,Privman91b} and, more generally, the deposition of extended objects \cite{Ballani06,Gromenko09a,Gromenko09b,Subashiev07a,Subashiev07b}.
Recently, interest has shifted towards the study of the effect of adsorbing particles of different sizes, with research becoming focused on the study of binary mixtures of particles undergoing either {\it competitive} adsorption or {\it preadsorption}.
Regarding the latter case, particles adsorb during two subsequent stages with each of the particle species attempting adsorption at a time \cite{Lee96,Cadilhe04,Dorsogna05,Weronski05,Kondrat06}, while during {\it competitive} adsorption both particle species attempt adsorption simultaneously \cite{Subashiev07a,Subashiev07b,Cadilhe04,Bartelt91a,Bonnier01a,Hassan01,Hassan02,Araujo06}.
In general, several physical properties are related to the interparticle distance, so proper characterization of the corresponding probability distribution function is fundamental.
The interest stems from the theoretical perspective in gaining insight on the functional dependence of the interparticle distribution even when an analytical solution is not available.
From the experimental perspective, the availability of the interparticle distribution function provides for controlled analysis.
In the present work we focus on the {\it competitive} adsorption of segments of two different sizes on a line and characterize various interparticle distribution functions with moments up to the fourth order.
We propose a novel approach to characterize the gap-size distribution functions\footnote{This is a term that will be interchangeably used to represent the interparticle distribution function.} based on the systematic use of their cumulants.
Comparing the semi-analytical results using an {\it ansatz} with recently reported Monte Carlo simulation results \cite{Araujo06}, we find good qualitative agreement.
The paper is organized as follows: In the next section, we start introducing the motivation, some terminology, and theory.
In Sec.~\ref{Distrib} we introduce the {\it ansatz} and we compute the various cumulants.
We present our results and final remarks in Secs.~\ref{Res}~and~\ref{concl}.
\section{Motivation and theory}\label{Model}
We consider the competitive adsorption of two size segments on the line, namely short segments, which we denote as $A$ segments, and long ones as $B$.
The model corresponds to an extension of the continuum random sequential adsorption model \cite{Privman00a,Evans93,Cadilhe07,Bonnier01a,Hassan02,Privman00b}, where a segment successfully adsorbs if it does not overlap a previously adsorbed one, therefore mimicking an excluded volume, short-range interaction.
The model has been studied both analytically \cite{Bonnier01a,Hassan02} and by Monte Carlo simulations \cite{Cadilhe07,Hassan02,Araujo06}.
Without loss of generality, we rescale the size of the system by the length of short segments, so that the length of the smaller segments becomes unity while that of the longer segments is $R$.
We term as the size ratio the quotient of the lengths of the larger to the smaller segments, which equals $R$ (the length of the larger segments in dimensionless units).
In general, one can have both segment types to arrive at different rates per unit length and per unit time, which represents the incident particle flux in the one-dimensional case.
In the present work, we take the particular case of equal incident fluxes for both types of segments, i.e., on average, equal number of segments attempt deposition per unit time and per unit length.
Since adsorption of segments is irreversible, i.e., in physical systems where surface mobility and detachment upon adsorption can be safely neglected, it leads to a jammed state, where no segment can fit into any of the available interstitial spaces.
The present model has been studied in the literature in various limiting cases.
For example, for $R= 1$ it corresponds to the R{\'e}nyi or car problem result \cite{Bartelt91b,Privman00a,Evans93,Privman94,Privman97,Cadilhe07,Araujo08,Flory39,Gonzalez74,Renyi58,Renyi63,Bonnier01a,Hassan02,Bartelt91c}.
The case of $R \to \infty$ has been previously discussed \cite{Araujo06}.
The limit of $R \to 0$ leads to kinetics of pointlike particles or fragmentation kinetics \cite{Bonnier01a,Hassan01,Bartelt91c}.
Though we limit ourselves to size ratios $R \ge 1$, the pointlike kinetics corresponds to the limit of $R \to \infty$, followed by a rescaling of the linear dimensions by the length of the large particle.
To properly characterize the jammed state, we start by introducing some definitions and relations.
The basic quantity to characterize is the probability distribution of the interparticle distance defined as
\begin{equation}
\hspace*{-1.0cm}P_{\emptyset}(x)\mbox{d}x= \frac{\ \!\mbox{Number of empty intervals of size }x\mbox{ within }]x, x + \mbox{d}x[\ \!}{\mbox{Total number of empty intervals}}\mbox{d}x.
\end{equation}
Thus, $P_{\emptyset}(x)\mbox{d}x$ at the jammed state has the property
\begin{equation}\label{normalization}
\int_0^1 P_{\emptyset}(x)\mbox{d}x=1.
\end{equation}
Discriminating all gap types in terms of pairs of consecutively adsorbed segments provides four distinct cases, namely $AA$, $AB$, $BA$, and $BB$.
Density distribution functions $P_b(x)$, with $b\in \{AA, AB, BA, BB\}$, i.e., for each of the gap types, are defined, from which the following relation is obeyed
\begin{equation}\label{discrimination}
P_{\emptyset} (x)=P_{AA} (x) + 2P_{AB} (x) + P_{BB} (x),
\end{equation}
where the equality $P_{AB} (x)= P_{BA} (x)$ was used.
Note from equations (\ref{normalization}) and (\ref{discrimination}) that the $P_b (x)$ are not normalized.
Keeping in mind the above definitions, it is now straightforward to reckon moments of any order of the gap-size distribution functions, defined by
\begin{equation}\label{moments}
\langle x^n\rangle _a= \frac{\int_0^1 x^nP_a (x)\mbox{d}x}{\int_0^1P_a (x)\mbox{d}x},
\end{equation}
with $a\in \{\emptyset, AA, AB, BA, BB\}$ and $n= 0$, $1$, $2$, \dots, though in the present work, we restrict ourselves to moments with $n \le 4$.
Using equation (\ref{discrimination}) one can relate the moments of the gap-size distributions with the corresponding moment of the global gap-size distribution function given by equation (\ref{moments}) yielding
\begin{eqnarray}
{\langle x^n\rangle}_{\emptyset}&=&{\langle x^n \rangle}_{AA}\int_0^1P_{AA} (x)\mbox{d}x \nonumber \\
&&\quad+2{\langle x^n\rangle}_{AB}\int_0^1P_{AB} (x)\mbox{d}x \nonumber \\
&&\quad+{\langle x^n\rangle}_{BB}\int_0^1P_{BB} (x)\mbox{d}x,
\end{eqnarray}
again, defined for all values of $n=$ $0$, $1$, $2$, $\dots$.
We also compute the cumulants, $\kappa_m^a$, of the gap-size distributions defined as,
\begin{equation}\label{cumulants}
\ln G_a (k)= \sum_{m= 1}^\infty\frac{(ik)^m}{m!}\kappa_m^a,
\end{equation}
where $G_a (k)$ is the so called characteristic function defined by \cite{Kampen92},
\begin{equation}\label{characteristic}
G_a (k)\equiv {\langle e^{ikx}\rangle}_a= \frac{\int_0^1e^{ikx}P_a (x)\mbox{d}x}{\int_0^1P_a (x)\mbox{d}x}.
\end{equation}
Therefore, from equations (\ref{moments}), (\ref{cumulants}), and (\ref{characteristic}), one derives the first four cumulants as
\begin{equation}\label{cumulant_k1}
\kappa_1^a= {\langle x\rangle}_a,
\end{equation}
\begin{equation}\label{cumulant_k2}
\kappa_2^a= \langle x^2\rangle _a-\langle x\rangle _a^2,
\end{equation}
\begin{equation}\label{cumulant_k3}
\kappa_3^a= \langle x^3\rangle _a-3\langle x^2\rangle _a \langle x\rangle _a+2\langle x\rangle _a^3,
\end{equation}
\begin{equation}\label{cumulant_k4}
\kappa_4^a= \langle x^4\rangle _a-4\langle x^3\rangle _a\langle x\rangle _a-3\langle x^2\rangle _a^2+12\langle x^2\rangle _a\langle x\rangle _a^2-6\langle x\rangle _a^4,
\end{equation}
where $\kappa_1$ is just the mean value and $\kappa_2$ is the variance.
While the first and second cumulants have well-established meanings, those of the third and fourth order cumulants are put in terms of the skewness,
\begin{equation}\label{skewness}
S_a= \frac{\kappa_3^a}{\left(\kappa_2^a\right)^{3/2}},
\end{equation}
and kurtosis,
\begin{equation}\label{kurtosis}
K_a= \frac{\kappa_4^a}{\left(\kappa_2^a\right)^2},
\end{equation}
respectively \cite{Kampen92}, in order to facilitate their interpretation.
For later convenience, we also take the opportunity to define the dispersion, $\sigma_a\equiv \sqrt{\kappa_2^a}$, used in the analysis of the results provided in Sec.~\ref{Res}.
\section{The gap-size distribution functions}\label{Distrib}
Since the probability that a segment is deposited in a given gap is proportional to the {\it effective} size of the gap, the probability distribution describing the interparticle distance should be reasonably approximated by an exponential distribution.
The {\it effective} gap size is defined as the difference between the actual size of the gap and the size of the segment attempting adsorption, so it represents the actual length available for deposition.
In the jamming state, the state where no more particles can be deposited, the probability of having a gap of length larger than one is zero.
So, we propose as the {\it ansatz} for the probability distribution function of the interparticle distance in the jammed state, an exponential function truncated to the interval $[0, 1[$,
\begin{equation}\label{Prob}
P_a (x)= \left\{
\begin{array}{ll}
\!\!A_a (\alpha_a) e^{-\alpha_a x}, & 0 \le x < 1 \\
\!\!0, & x > 1 \\
\end{array}
\right.,
\end{equation}
where $\alpha_a$ is the single parameter used to characterize the distribution function of the interparticle distance.
Since $P_{\emptyset}(x)$ is normalized (eq.~(\ref{normalization})), one has
\begin{equation}\label{norm.factor}
A^{-1}_{\emptyset}(\alpha_{\emptyset})= \frac{1 - e^{-\alpha_{\emptyset}}}{\alpha_{\emptyset}}= 1.
\end{equation}
Some considerations are due before proceeding to the derivation of the various cumulants.
The idea of using the proposed {\it ansatz} (eq.~(\ref{Prob})) is to describe the jammed state.
Though a truncated-exponential {\it ansatz} best describes uncorrelated deposition events, it allows us to assess the presence of correlated deposition events at the jammed state.
The description of the full kinetics towards the jammed state would not be possible with the proposed {\it ansatz}.
In particular, during the adsorption process, the upper limit of the integral in eq.~(\ref{normalization}) would have to be extended to infinity.
Moreover, at early times, the kinetics of deposition is uncorrelated and an exponential distribution represents a better fit \cite{Cadilhe07}.
Thus, we are interested in the jammed state when systematic elimination, by the adsorption of particles, of all gaps of length $x \ge 1$ have taken place.
From eq.~(\ref{moments}) the various moments of the gap-size distributions in eq.~(\ref{Prob}) are calculated.
The first moment equals the first-order cumulant (eq.~(\ref{cumulant_k1})),
\begin{equation}\label{moment_m1}
\kappa_1^a\equiv\langle x\rangle_a= \frac{1}{1-e^{\alpha_a}}+\frac{1}{\alpha_a}.
\end{equation}
The second moment is
\begin{equation}\label{moment_m2}
{\langle x^2\rangle}_a= \alpha_a^{-2}\left[2+\frac{\alpha_a (2+\alpha_a)}{1-e^{\alpha_a}}\right],
\end{equation}
and the dispersion is
\begin{equation}\label{cumulant_teo_k2}
\sigma_a= \sqrt{\frac{1}{\alpha_a^2}+\frac{1}{2-2 \cosh{\left(\alpha_a\right)}}}.
\end{equation}
\begin{figure}[floatfix]
\includegraphics[width=8.0cm]{fig1.eps}
\caption
{ \label{fig:1}Plot of $\alpha_a$, with $a\in\left\{AA, AB, BB\right\}$, as a function of size ratio computed from first moment of Monte Carlo simulations \cite{Araujo06}.
Each of the $AA\mbox{-}$, $AB\mbox{-}$, and $BB$-gap types are represented by circles, squares, and triangles, respectively.
}
\end{figure}
To better characterize the distribution function, as previously referred to, one resorts to the calculation of higher moments.
The third- and fourth-order cumulants are reckoned in terms of the moments and used for the definition of the skewness and kurtosis, which are quantities with a more intuitive meaning.
The skewness yields information about the asymmetry of the distribution, while the kurtosis provides information about the peakness or flatness of the distribution function.
For completeness, we provide the relations for the third and fourth moments,
\begin{equation}\label{moment_m3}
{\langle x^3\rangle}_a= {\alpha_a}^{-3}\left[6 + {\frac{\alpha_a(6 + \alpha_a(3 + \alpha_a))}{1 - e^{\alpha_a}}}\right]
\end{equation}
and
\begin{equation}\label{moment_m4}
{\langle x^4\rangle}_a= {\alpha_a}^{-4}\left[24 + \frac{\alpha_a(24 + \alpha_a(12 + \alpha_a (4 + \alpha_a)))}{1 - e^{\alpha_a}}\right].
\end{equation}
After some tedious, though straightforward, algebra we obtain the third and fourth cumulants,
\begin{equation}\label{cumulant_teo_k3}
\kappa_3^a= \frac{2}{\alpha_a^3} - \frac{1}{4} \coth{\left(\frac{\alpha_a}{2}\right)}\mbox{csch}^2{\left(\frac{\alpha_a}{2}\right)}
\end{equation}
and
\begin{equation}\label{cumulant_teo_k4}
\kappa_4^a= \frac{6}{\alpha_a^4} - \frac{1}{8} (2 + \cosh\left(\alpha_a\right)) \mbox{csch}^4{\left(\frac{\alpha_a}{2}\right)}.
\end{equation}
\begin{figure}[floatfix]
\includegraphics[width=11.5cm]{fig2.eps}
\caption{\label{fig:2}Plots involving cumulants up to the fourth order of the gap-size distribution functions for each gap type as a function of the size ratio, according to equations (\ref{moment_m1}), (\ref{cumulant_teo_k2}), (\ref{skewness_teo}), and~(\ref{kurtosis_teo}).
a) Average and
b) dispersion of the distance between pairs of segments.
c) Skewness.
d) Kurtosis.
The legend convention follows that presented in Fig.~\ref{fig:1}.
}
\end{figure}
From eqs.~(\ref{skewness}) and~(\ref{kurtosis}), and using eqs.~(\ref{cumulant_teo_k2}), (\ref{cumulant_teo_k3}), and~(\ref{cumulant_teo_k4}), we obtain the skewness
\begin{equation}\label{skewness_teo}
S_a=\frac{4-4 \cosh\left(\alpha_a\right)+\alpha_a^3 \coth\left(\frac{\alpha_a}{2}\right)}{\alpha_a \sqrt{\frac{1}{\alpha_a^2}+\frac{1}{2-2 \cosh\left(\alpha_a\right)}} (2+\alpha_a^2-2 \cosh\left(\alpha_a\right))},
\end{equation}
and the kurtosis
\begin{equation}\label{kurtosis_teo}
K_a=-2\frac{(24+\alpha_a^4) \cosh\left(\alpha_a\right)+2 (-9+\alpha_a^4-3 \cosh\left(2 \alpha_a\right))}{(2+\alpha_a^2-2 \cosh\left(\alpha_a\right))^2},
\end{equation}
for the truncated exponential distribution function defined in eq.~(\ref{Prob}).
\section{Results and Discussion}\label{Res}
In a previous work \cite{Araujo06}, we have performed an extensive Monte Carlo simulation study of the present random sequential model.
Here, we utilize results from that work, specifically from the first moment as defined in eq.~(\ref{moment_m1}), to determine the values of $\alpha$ for all gap types and size ratios ranging from $1$ to $20$.
Figure~\ref{fig:1} shows the computed values of $\alpha$ as a function of the size ratio.
Heretoforth, the values of the size ratio match those of the Monte Carlo simulations to simplify comparison~\cite{Araujo06}.
For the $BB$ gap type a monotonic decrease of $\alpha$ with the size ratio is observed.
The $AA$ and $AB$ gap types are characterized by maxima for values of the size ratio at $1.30\pm0.05$ and $1.20\pm0.05$, respectively.
To characterize the system, in Fig.~\ref{fig:2} the average distance, dispersion, skewness, and kurtosis are computed, respectively, with eqs.~(\ref{moment_m1}), (\ref{cumulant_teo_k2}), (\ref{skewness_teo}), and~(\ref{kurtosis_teo}) as a function of the size ratio.
For the $AA$ and $AB$ gap types, a non-monotonic behavior for all these quantities is observed.
For the BB gap, the average distance and dispersion increase with the size ratio as shown in Fig.~\ref{fig:2}(a) and~(b), while the skewness and kurtosis decrease monotonically as shown in parts~(c) and~(d) of the same figure.
For the average distance and dispersion (Fig.~\ref{fig:2}(a)~and~(b)) there is a minimum for the $AA$ gap at a value of the size ratio of $1.30\pm0.05$ and of $1.20\pm0.05$ for the $AB$ gap.
For the skewness and kurtosis (Fig.~\ref{fig:2}(c)~and~(d)) there is a maximum for the $AA$ gap at $1.30\pm0.05$ and at $1.20\pm0.05$ for $AB$ gap.
The values of the maxima and minima for the $AA$ and $AB$ gap types coincide with the minimum values of $\alpha$ as a function of the size ratio.
\begin{figure}[floatfix]
\includegraphics[width=12.0cm]{fig3.eps}
\caption{\label{fig:3}Plots involving the ratio of cumulants of the gap-size distribution functions of the same order.
The plots provide a direct comparison of the results from the truncated exponential {\it ansatz} to those from Monte Carlo simulations as explained in the text.
a) Dispersion.
b) Skewness.
c) Kurtosis.
The legend convention follows that of Fig.~\ref{fig:1}.
}
\end{figure}
We compare results using the truncated exponential {\it ansatz} with simulational ones found in Ref.~\cite{Araujo06}.
All computed quantities possess the same qualitative behavior as that obtained from an extensive Monte Carlo study.
The exception regards the values of the dispersion at sizes ratios near unit.
In this situation, streaks of alternating $A$ and $B$ segments are common, since two consecutive $BB$ segments might still have a gap of length larger than unit, where a single $A$ segment can fit in.
In fact, at size ratios lower than $1.55$, the population of $BB$ gaps remains significant, as compared with the $AA$ gap one.
However such streaks leading to gaps of size near unit, which we denoted as {\it snug fits} in Ref.~\cite{Araujo06}, are not possible to be described by a simple truncated exponential as the latter does not account for influences due to the history of several segment deposition events.
The ratios of the values of the cumulants of the same order from the truncated exponential {\it ansatz} to those from Monte Carlo simulations are show in Fig.~\ref{fig:3}.
The $BB$ gap shows better compliance with proposed {\it ansatz}.
In a sense, this result is not entirely surprising, since the larger segments are deposited at early times.
Consequently, there will be less interaction with previously deposited segments.
At late times, i.e., in the asymptotic regime, the $A$ segments will fill the gaps left out by the $B$ segments, particularly as the size ratio increases.
Improved compliance to the truncated exponential is also observed for the $AA$ and $AB$ gap types with increasing values of the size ratio for similar reasons to those of the $BB$ gaps.
As the size ratio increases, the $B$ segments have a sharper decay in the rate of adsorption due to single $A$ segments blocking their adsorption.
Consecutive $B$ segments also leave larger spacing between them, on average.
Consequently, the adsorption of $A$ segments follows more closely the argument in favor of a truncated exponential presented above.
At size ratios close to one, the results provided by the {\it ansatz} deviate more, since the populations of the various gap types are nearly equilibrated and one expects the equations describing these gap types to be more {\it coupled}, i.e., to be more dependent on multisegment deposition events.
\section{Final remarks}\label{concl}
In short, we introduced a novel approach with potential for a semi-quantitative analysis of models where the interparticle distribution function of the distances is of relevance.
Specifically, the method simply uses a single moment to obtain information for the other three.
However, it requires some prior knowledge of what can constitute a good {\it ansatz}.
We present a coherent study of the dependence on the size ratio up to the fourth moment.
We propose a truncated exponential function for the gap-size distribution and compute the mean gap size, dispersion, skewness, and kurtosis.
Qualitatively, the results for the proposed {\it ansatz} are in good agreement with the results obtained from simulational work.
However, the dispersion of AB gap type at small values of size ratio, the exponential function does not reproduce the behavior of the distribution function, which {\it can be explained by {\it snug fit} events}.
Finally, we point out that the present approach can be applied to a broader range of models involving the irreversible deposition of particles on a line.
\begin{ack}
This research has been funded by a Funda{\c c}{\~a}o para a Ci{\^e}ncia e a Tecnologia research grant and Search (Services and Advanced Research Computing with HTC/HPC clusters) (under contract CONC-REEQ/443/2001).
One of us, N.~A., thanks Funda{\c c}{\~a}o para a Ci{\^e}ncia e a Tecnologia for a Ph.D.\ fellowship (SFRH/BD/17467/2004).
A.~C. wants also to thank both Funda{\c c}{\~a}o para a Ci{\^e}ncia e a Tecnologia (SFRH/BPD/34375/2007) and Funda{\c c}{\~a}o Calouste Gulbenkian for fellowships to visit Los Alamos National Laboratory.
A.~C. also acknowledges the warm hospitality of the T-1~Group at Los Alamos National Laboratory.
Finally, we want to thank suggestions and comments on earlier versions of the manuscript by N.~Henson, V.~Privman, and C.~Reichhardt.
\end{ack}
\section*{References}
|
\section{Introduction}
It is well known that the usual model for the quantized single modes of the electromagnetic field is
the harmonic oscillator with an infinity of states. The infinite-dimensional character of the
representation space of the corresponding oscillator algebra constitutes a drawback to define a phase
operator in a consistent way \cite{Louisell63}-\cite{Susskind68}. In order to get rid of this difficulty,
Pegg and Barnett suggested to truncate to some finite (but arbitrarily large) order the infinite-dimensional
representation space of the oscillator algebra \cite{Pegg}. Their approach also provided a valid way
for calculating the so-called phase states (the eigenvectors of the phase operator). In the same vein,
Vourdas proposed a definition of a phase operator for $su(2)$ and calculated its eigenstates without a truncation
procedure since $su(2)$ admits finite-dimensional unitary irreducible representations \cite{Vourdas90}. He also
constructed a phase operator and its eigenstates for $su(1,1)$, without a truncation procedure although $su(1,1)$
admits infinite-dimensional unitary irreducible representations \cite{Vourdas90}.
The main aim of the present work is to develop a method to build unitary phase operators\footnote{We deal
here with {\it unitary} rather than {\it Hermitian} phase operators. The two kinds of operators are related
via an exponentiation trick.} and temporally stable phase states for some exactly solvable quantum systems. Various
algebraic structures were used to construct (temporaly stable or not) coherent states in connection with some quantum
systems \cite{BarutGi}-\cite{Kinani}. The construction of temporally stable
phase states to be developed in this work is based on a generalized oscillator algebra which takes its
root in \cite{DaoKib12, DaoKibPLAJMP}. This algebra was introduced to construct isospectral
shape invariant potentials in the framework of fractional supersymmetry.
A second facet of this work is to show that the obtained temporally stable phase states can be used to generate mutually
unbiased bases (MUBs). Such bases are of considerable interest in quantum information and were recently investigated from
an angular momentum approach \cite{KibKPAlbKib, KibJPhysA0809}. It is not the purpose of this paper to deal with unsolved
problems concerning MUBs but to give a way to construct MUBs from temporally stable states associated with some exactly
solvable systems.
The paper is organized as follows. Section 2 is devoted to the generalized oscillator algebra ${\cal A}_{\kappa}$. Temporally
stable phase states associated with ${\cal A}_{\kappa}$ are studied in section 3. Section 4 deals with the truncated oscillator
algebra ${\cal A}_{\kappa, s}$ and the correponding phase states. As a first application, the derivation of MUBs
from phase states is developed in section 5. A second application is made in section 6 to some exactly solvable quantum systems.
The notations are standard. Let us simply mention that: $\delta_{a , b}$ stands for the Kronecker symbol of $a$ and $b$, $I$
for the identity operator, $A^{\dagger}$ for the adjoint of the operator $A$, and $[A , B]$ and $\{ A, B \}$ for respectively
the commutator and the anticommutator of the operators $A$ and $B$. We use a notation of type $\vert \psi \rangle$ for a vector
in an Hilbert space and we denote $\langle \phi \vert \psi \rangle$ and $\vert \phi \rangle \langle \psi \vert$ respectively
the inner and outer products of the vectors $\vert \psi \rangle$ and $\vert \phi \rangle$.
\section{Generalized oscillator algebra}
\subsection{The algebra ${\cal A}_{\kappa}$}
Let ${\cal A}_{\kappa}$ be the algebra spanned by the three linear operators $a^-$, $a^+$ and $N$
satisfying the following relations
\begin{eqnarray}
[a^- , a^+] = I + 2 \kappa N \qquad
[N , a^{\pm}] = \pm a^{\pm} \qquad
\left( a^- \right)^{\dagger} = a^+ \qquad
N^{\dagger} = N,
\label{thealgebra}
\end{eqnarray}
where $\kappa$ is a real parameter. Note that, for $\kappa = 0$, the algebra ${\cal A}_{0}$ is nothing
but the usual harmonic oscillator algebra. The operators $a^-$, $a^+$ and $N$ in (\ref{thealgebra})
generalize the annihilation, creation and number operators used for the harmonic oscillator.
Therefore, the algebra ${\cal A}_{\kappa}$ shall be called generalized oscillator algebra. This algebra
turns out to be a particular case of the generalized Weyl-Heisenberg algebra $W_k$ introduced in
\cite{DaoKib12, DaoKibPLAJMP} and not to be confused with the Lie algebra of the Heisenberg-Weyl group $HW(\mathbb{R})$
used in quantum information \cite{KibJPhysA0809}. In fact, ${\cal A}_{\kappa}$ is identical to $W_k$ with
\begin{eqnarray}
k = 1 \qquad f_0(N) = aN + b \qquad \frac{1}{\sqrt{b}} X_{\pm} = a^{\pm} \qquad \kappa = \frac{1}{2} \frac{a}{b},
\label{lien avec Wk}
\end{eqnarray}
where the operators $f_0(N)$ and $X_{\pm}$, and the parameters $k$, $a$ and $b$ are defined in \cite{DaoKibPLAJMP}. It
should be noted that the $C_{\lambda}$-extended oscillator algebra worked out in \cite{MD33LesQuene} is a particular case
of $W_k$ (for $\lambda = k$).
\subsection{Hilbertian representation of ${\cal A}_{\kappa}$}
We denote by ${\cal F}_{\kappa}$ the finite- or infinite-dimensional Hilbert space on which the
operators $a^-$, $a^+$ and $N$ are defined. Let
\begin{eqnarray}
\{ \vert n \rangle : n = 0, 1, \ldots, d({\kappa}) \}
\end{eqnarray}
(with $d({\kappa})$ finite or infinite) be an orthonormal basis, with respect to the inner product
$\langle n \vert n' \rangle = \delta_{n,n'}$, of the space ${\cal F}_{\kappa}$. It is easy to check
that the actions
\begin{eqnarray}
& & a^+ \vert n \rangle = \sqrt{F(n+1)} e^{{-i [F(n+1)- F(n)] \varphi }} \vert n+1 \rangle, \nonumber \\
& & a^- \vert n \rangle = \sqrt{F(n)} e^{{ i [F(n) - F(n-1)] \varphi }} \vert n-1 \rangle, \label{action sur les n} \\
& & a^- \vert 0 \rangle = 0 \qquad
N \vert n \rangle = n \vert n \rangle
\nonumber
\end{eqnarray}
provide an Hilbertian representation of the algebra ${\cal A}_{\kappa}$ defined by (\ref{thealgebra}). In
equation (\ref{action sur les n}), the real parameter $\varphi$ is arbitrary and the positively-valued function
$F : \mathbb{N} \to \mathbb{R}_+$ satisfies the recurrence relation
\begin{eqnarray}
F(n+1) - F(n) = 1 + 2 \kappa n \qquad F(0) = 0.
\label{recurrence}
\end{eqnarray}
The iteration of (\ref{recurrence}) yields
\begin{eqnarray}
F(n) = n [1 + \kappa (n - 1)],
\label{eigenvalue}
\end{eqnarray}
which is linear in $n$ only for $\kappa = 0$. Since $F(n) \in \mathbb{R}_+$, we must have the following condition
\begin{eqnarray}
1 + \kappa (n - 1) > 0
\label{condition}
\end{eqnarray}
for $n > 0$. The condition (\ref{condition}) determines the value of $d({\kappa})$ and then the
dimension of ${\cal F}_{\kappa}$. The finiteness or infiniteness of ${\cal F}_{\kappa}$ depends on the
sign of the parameter $\kappa$. For $\kappa \geq 0$, the space ${\cal F}_{\kappa}$ is infinite-dimensional. In
fact, for $\kappa = 0$, the space ${\cal F}_{0}$ coincides with the usual Hilbert-Foch space for the harmonic oscillator. For
$\kappa < 0$, there exists a finite number of states satisfying the condition (\ref{condition}). As a matter of fact, for
$\kappa < 0$, $n$ can take the values
\begin{eqnarray}
n = 0, 1, \ldots, E(-\frac{1}{\kappa}) \equiv d-1,
\end{eqnarray}
where $E(x)$ stands for the integer part of $x$. The finiteness of the space ${\cal F}_{\kappa}$ induces properties
of the operators $a^-$ and $a^+$ which differ from those corresponding to an infinite-dimensional space. In particular,
the trace of any commutator in the finite-dimensional space must be zero. This implies that the parameter $\kappa$ is
related to the dimension $d$ of the space ${\cal F}_{\kappa}$ by
\begin{eqnarray}
d = 1 - \frac{1}{\kappa}.
\label{dimension}
\end{eqnarray}
Equation (\ref{dimension}) requires that $-1/\kappa$ be a positive integer. In the following, we shall assume that
$-1/\kappa \in \mathbb{N}^*$ when $\kappa < 0$.
\subsection{A generalized oscillator Hamiltonian}
We are now in a position to define an operator which generalizes (up to an additive constant) the
Hamiltonian $a^+ a^- + 1/2$ for the one-dimensional harmonic oscillator. Starting from
\begin{eqnarray}
a^+ a^- \vert n \rangle = F(n) \vert n \rangle \Rightarrow F(N) = a^+ a^-,
\end{eqnarray}
we refer $F(N)$ to as an Hamiltonian associated with the generalized oscillator algebra ${\cal A}_{\kappa}$. The
eigenvalue equation
\begin{eqnarray}
F(N) \vert n \rangle = n [1 + \kappa (n-1)] \vert n \rangle
\end{eqnarray}
gives the energies (\ref{eigenvalue}) of a quantum dynamical system described by the Hamiltonian operator $F(N)$. Let us
discuss the degeneracies of the levels $F(n)$ given by (\ref{eigenvalue}).
(i) In the case $\kappa \geq 0$, the spectrum of $F(N)$ is nondegenerate.
(ii) In the case $\kappa < 0$, the eigenvalues of $F(N)$ can be rewritten as
\begin{eqnarray}
F(n) = n \frac{d-n}{d-1},
\label{eigenvalue avec d}
\end{eqnarray}
so that
\begin{eqnarray}
F(n) = F(d-n) \qquad n = 1, 2, \ldots, d-1.
\end{eqnarray}
Thus, for $d$ even the levels are doublets except the fundamental level $n=0$ and the level
$n = d/2$ which are nondegenerate. For $d$ odd the levels are two-fold degenerate except the
fundamental level $n=0$ which is a singlet.
In both cases ($\kappa \geq 0$ and $\kappa < 0$), we note that the Perron-Frobenius theorem \cite{Reed} is satisfied, namely,
the fundamental level is nondegenerate.
It is known that one-dimensional quantum dynamical systems (on the real line) correspond to
nondegenerate spectra. Therefore, the representation obtained for ${\cal A}_{\kappa}$ with
$\kappa < 0$ cannot be used to describe a particle evolving in some nonrelativistic potential
on the real line. However, a modification of the generalized oscillator algebra ${\cal A}_{\kappa}$
can be achieved in orded to avoid the degeneracies of $F(N)$. This will be done in section 4
by means of a truncation procedure which will prove also useful in the case $\kappa \geq 0$
to define in a consistent way the phase operator for some exactly solvable systems.
\section{Temporally stable phase states for ${\cal A}_{\kappa}$}
We shall treat separately the cases $\kappa \geq 0$ and $\kappa < 0$ associated with the
infinite- and the finite-dimensional representation of the generalized oscillator algebra
${\cal A}_{\kappa}$, respectively.
\subsection{The infinite-dimensional case}
In the case $\kappa \geq 0$, we decompose $a^-$ and $a^+$ as
\begin{eqnarray}
a^- = E_{\infty} \sqrt{F(N)} \qquad a^+ = \sqrt{F(N)} \left( E_{\infty} \right)^{\dagger},
\label{decompo cas infini}
\end{eqnarray}
where
\begin{eqnarray}
E_{\infty} := \sum_{n=0}^{\infty} e^{i [F(n+1)- F(n)] \varphi } \vert n \rangle \langle n+1 \vert.
\end{eqnarray}
It is important to emphasize that
\begin{eqnarray}
E_{\infty}\left( E_{\infty} \right)^{\dagger} = \sum_{n=0}^{\infty} \vert n\rangle\langle n\vert = I \qquad
\left( E_{\infty} \right)^{\dagger} E_{\infty}= \sum_{n=1}^{\infty} \vert n\rangle\langle n\vert = I - \vert 0 \rangle \langle 0 \vert,
\end{eqnarray}
a result which means that $E_{\infty}$ is not a unitary operator.
To find the phase states corresponding to $\kappa \geq 0$,
let us consider the eigenvalue equation
\begin{eqnarray}
E_{\infty} \vert z \rangle = z \vert z \rangle \qquad z \in \mathbb{C}.
\end{eqnarray}
By expanding the vector $\vert z \rangle$ of ${\cal F}_{\kappa}$ as
\begin{eqnarray}
\vert z \rangle = \sum_{n=0}^{\infty} C_n z^n \vert n \rangle,
\end{eqnarray}
it is easy to see that the complex coefficients $C_n$ satisfy the relation
\begin{eqnarray}
C_{n+1} = e^{-i [F(n+1) - F(n)] \varphi } C_n \qquad n \in \mathbb{N}.
\end{eqnarray}
It follows that
\begin{eqnarray}
C_n = e^{-i F(n) \varphi} C_0 \qquad n \in \mathbb{N},
\end{eqnarray}
where the coefficient $C_0$ can be determined from the normalization condition of the states $\vert z \rangle$. As a result,
we can take (up to a phase factor)
\begin{eqnarray}
\vert z \rangle = \sqrt{1 - |z|^2} \sum_{n=0}^{\infty} z^n e^{- i F(n) \varphi} \vert n \rangle
\end{eqnarray}
on the domain $\{ z \in \mathbb{C}, |z| < 1 \}$.
Following the method developed in \cite{VoudasBM} for the Lie algebra $su(1,1)$, we define the states $\vert \theta , \varphi \rangle$ by
\begin{eqnarray}
\vert \theta , \varphi \rangle := \lim_{z \rightarrow e^{i\theta}} \frac{1}{\sqrt{1 - |z|^2}} \vert z \rangle,
\end{eqnarray}
where $\theta \in [-\pi , +\pi]$ (see also \cite{VourdasLimit} where a limit of type
$z \rightarrow e^{i\theta} \Rightarrow \vert z \vert \rightarrow 1$ is used in a similar way). We
thus get the states
\begin{eqnarray}
\vert \theta , \varphi \rangle = \sum_{n=0}^{\infty} e^{i n \theta} e^{- i F(n) \varphi} \vert n \rangle.
\end{eqnarray}
These states, defined on the unit circle $S^1$, turn out to be phase states. Indeed, we have
\begin{eqnarray}
E_{\infty} \vert \theta , \varphi \rangle = e^{i \theta} \vert \theta , \varphi \rangle.
\end{eqnarray}
Hence, the operator $E_{\infty}$ is a (nonunitary) phase operator.
The main properties of the states $\vert \theta , \varphi \rangle$ are the following.
(i) They are temporally stable in the sense that the relation
\begin{eqnarray}
e^{-i F(N) t} \vert \theta , \varphi \rangle = \vert \theta , \varphi + t)
\end{eqnarray}
is satisfied for any value of the real parameter $t$. This property is due to
the presence of the parameter $\varphi$ in the phase operator $E_{\infty}$.
(ii) They are not normalized and not orthogonal. However, for fixed $\varphi$, they satisfy the closure relation
\begin{eqnarray}
\frac{1}{2\pi} \int_{-\pi}^{+\pi} d\theta \vert \theta , \varphi \rangle \langle \theta , \varphi \vert = I.
\end{eqnarray}
Finally, observe that for $\varphi = 0$ the states $\vert \theta , 0 \rangle$ have the same form than those derived
in \cite{VoudasBM} for $su(1,1)$.
\subsection{The finite-dimensional case}
For $\kappa < 0$ with $-1/\kappa \in \mathbb{N}^*$, the Hilbert space
${\cal F}_{\kappa}$ is $d$-dimensional with $d = 1 - 1/\kappa$. The action of $a^-$ and $a^+$ on
${\cal F}_{\kappa}$ is given by (\ref{action sur les n}) supplemented by
\begin{eqnarray}
a^+ \vert d-1 \rangle = 0,
\end{eqnarray}
which easily follows from the calculation of $\langle d-1 \vert a^- a^+ \vert d-1 \rangle$.
Let us look for a decomposition of the creation $a^+$ and annihilation $a^-$ operators similar to
(\ref{decompo cas infini}) for the case $\kappa \geq 0$. Thus, let us put
\begin{eqnarray}
a^- = E_{d} \sqrt{F(N)} \Leftrightarrow a^+ = \sqrt{F(N)} \left( E_{d} \right)^{\dagger}.
\label{decompo cas fini}
\end{eqnarray}
The operator $E_{d}$ can be seen to satisfy
\begin{eqnarray}
E_d \vert n \rangle = e^{i [F(n) - F(n-1)] \varphi } \vert n-1 \rangle
\label{action de Ed}
\end{eqnarray}
for $n = 1, 2, \ldots, d-1$. For $n=0$, we shall assume that
\begin{eqnarray}
E_d \vert 0 \rangle = e^{i [F(0)- F(d-1)] \varphi} \vert d-1 \rangle
\end{eqnarray}
so that (\ref{action de Ed}) is valid modulo $d$. (Note that, in view of (\ref{decompo cas fini}),
$a^- \vert 0 \rangle = 0$ does not imply that $E_d \vert 0 \rangle = 0$.) It follows that we have
\begin{eqnarray}
\left( E_d \right)^{\dagger} \vert n \rangle = e^{-i [F(n+1) - F(n)] \varphi } \vert n+1 \rangle,
\end{eqnarray}
where $n+1$ should be understood modulo $d$. As an important result (to be contrasted with the situtation where $\kappa \geq 0$),
the operator $E_d$ is unitary. Therefore, equation (\ref{decompo cas fini}) constitutes a polar decomposition
of $a^-$ and $a^+$.
We are now ready to derive the eigenstates of the operator $E_d$. Let us consider the eigenvalue equation
\begin{eqnarray}
E_d \vert z \rangle = z \vert z \rangle \qquad \vert z \rangle = \sum_{n = 0}^{d-1} C_n z^n \vert n \rangle
\end{eqnarray}
with $z \in \mathbb{C}$. Here again (as in the case $\kappa \geq 0$), we obtain a recurrence relation for
the coefficients $C_n$, viz.,
\begin{eqnarray}
C_{n} = e^{-i [F(n) - F(n-1)] \varphi} C_{n-1} \qquad n = 1, 2, \ldots, d-1
\end{eqnarray}
with the cyclic condition
\begin{eqnarray}
C_{0} = z^d e^{-i [F(0) - F(d-1)] \varphi} C_{d-1}.
\end{eqnarray}
Therefore, we get
\begin{eqnarray}
C_n = e^{-i F(n) \varphi} C_0 \qquad n = 0, 1, \ldots, d-1,
\end{eqnarray}
with the discretization condition
\begin{eqnarray}
z^d = 1.
\end{eqnarray}
As a consequence, the complex variable $z$ is a root of unity
given by
\begin{eqnarray}
z = q^m \qquad m = 0, 1, \ldots, d-1,
\end{eqnarray}
where
\begin{eqnarray}
q := e^{2 \pi i / d}
\label{definition of q}
\end{eqnarray}
is reminiscent of the parameter used in the theory of quantum groups. The constant $C_0$
can be calculated from the normalization condition $\langle z \vert z \rangle = 1$ to be
\begin{eqnarray}
C_0 = \frac{1}{\sqrt d}
\end{eqnarray}
up to a phase factor. Finally, we arrive at the following
eigenstates $\vert z \rangle \equiv \vert m , \varphi \rangle$ of $E_d$
\begin{eqnarray}
\vert m , \varphi \rangle = \frac{1}{\sqrt d} \sum_{n=0}^{d-1} e^{-i F(n) \varphi} q^{mn} \vert n \rangle.
\label{coherentstatemvarphi}
\end{eqnarray}
The states $\vert m , \varphi \rangle$, labeled by the parameters $m \in \mathbb{Z}/d\mathbb{Z}$ and $\varphi \in \mathbb{R}$, satisfy
\begin{eqnarray}
E_d \vert m , \varphi \rangle = e^{i\theta_m} \vert m , \varphi \rangle \qquad \theta_m = m \frac{2 \pi}{d},
\end{eqnarray}
which shows that $E_d$ is indeed a phase operator. In the particular case $\varphi = 0$, the states $\vert m , 0 \rangle$
are similar to those derived in \cite{VoudasBM} for the Lie algebra $su(2)$. In this case, the states $\vert m , 0 \rangle$
correspond to an ordinary discrete Fourier transform of the basis $\{ \vert n \rangle : n = 0, 1, \ldots, d-1 \}$ of the
$d$-dimensional space ${\cal F}_{\kappa}$.
The phase states $\vert m , \varphi \rangle$ have remarkable properties (to be
compared to those for the states $\vert \theta , \varphi \rangle$ of the case $\kappa \geq 0$).
(i) They are temporally stable under ``time evolution''. In other words, they satisfy
\begin{eqnarray}
e^{-i F(N) t} \vert m , \varphi \rangle = \vert m , \varphi + t \rangle.
\end{eqnarray}
for any value of the real parameter $t$. We note here the major role of the parameter
$\varphi$ in ensuing the temporal stability of the states $\vert m , \varphi \rangle$.
(ii) For fixed $\varphi$, they satisfy the equiprobability relation
\begin{eqnarray}
| \langle n \vert m , \varphi \rangle | = \frac{1}{\sqrt{d}} \qquad n, m \in \mathbb{Z}/d\mathbb{Z}.
\label{computational et MUB}
\end{eqnarray}
(iii) For fixed $\varphi$, they satisfy the orthonormality relation
\begin{eqnarray}
\langle m , \varphi \vert m' , \varphi \rangle = \delta_{m,m'} \qquad m, m' \in \mathbb{Z}/d\mathbb{Z}
\end{eqnarray}
and the closure property
\begin{eqnarray}
\sum_{m = 0}^{d-1} \vert m , \varphi \rangle \langle m , \varphi \vert = I.
\end{eqnarray}
(iv) The overlap between two phase states $\vert m' , \varphi' \rangle$ and $\vert m , \varphi \rangle$ reads
\begin{eqnarray}
\langle m , \varphi \vert m' , \varphi' ) = \frac{1}{d} \sum_{n=0}^{d-1} q^{\rho(m-m', \varphi - \varphi', n)},
\label{overlap}
\end{eqnarray}
where
\begin{eqnarray}
\rho(m-m', \varphi - \varphi', n) = - (m - m')n + \frac{d}{2\pi} (\varphi - \varphi') F(n)
\end{eqnarray}
and $q$ is defined in (\ref{definition of q}). Therefore, the temporally stable phase states are not all
orthogonal.
\section{{Truncated generalized oscillator algebra and phase states}}
As discussed in section 2, in the case $\kappa \geq 0$ the Hilbert space ${\cal F}_{\kappa}$ associated with ${\cal A}_{\kappa}$
is infinite-dimensional. It is then impossible to define a unitary phase operator (see section 3). On the other hand, in the case $\kappa < 0$
with $- 1/ \kappa \in \mathbb{N}^*$ the space ${\cal F}_{\kappa}$ is finite-dimensional and there is no problem to define a
unitary phase operator. However, the spectrum of the Hamiltonian $F(N)$ associated with ${\cal A}_{\kappa}$ for
$- 1/ \kappa \in \mathbb{N}^*$ exhibits degeneracies. Therefore, it is appropriate to truncate the space ${\cal F}_{\kappa}$
for both $\kappa \geq 0$ and $\kappa < 0$ in order to get a subspace ${\cal F}_{\kappa, s}$ of dimension $s$ with the basis
$\{ |n\rangle : n = 0, 1, \ldots, s-1\}$. For $\kappa \geq 0$, the truncation is done at $s$ sufficiently large (note that the
difference $F(n+1) - F(n)$ between two consecutive states increases with $n$ for $\kappa > 0$ so that we can ignore, in a perturbative
scheme, the states with $n$ large). For $\kappa < 0$, the truncation can be done at $s = (d+2)/2$ for $d$ even and at $s = (d+1)/2$ for
$d$ odd (with $d$ given by (\ref{dimension})) in order to avoid the degeneracies of $F(N)$.
\subsection{{The truncated algebra ${\cal A}_{\kappa, s}$}}
Inspired by the work of Pegg and Barnett \cite{Pegg}, we
define the truncated generalized oscillator algebra ${\cal A}_{\kappa, s}$ through the three linear operators $b^-$, $b^+$ and $N$
satisfying the following relations
\begin{eqnarray}
[b^- , b^+] = I + 2 \kappa N - F(s) \vert s-1 \rangle \langle s-1 \vert \quad
[N , b^{\pm}] = \pm b^{\pm} \quad
\left( b^- \right)^{\dagger} = b^+ \quad
N^{\dagger} = N.
\label{thetruncatedalgebra}
\end{eqnarray}
The algebra ${\cal A}_{\kappa, s}$ generalizes the one introduced by Pegg and Barnett for the harmonic oscillator in their discussion
of the phase operator for the single modes of the electromagnetic field \cite{Pegg}. Indeed, the algebra
${\cal A}_{0, s}$, for $\kappa = 0$, is identical to the truncated oscillator algebra considered in \cite{Pegg}.
Following the same approach as in subsection 2.2, we define a $s$-dimensional representation of ${\cal A}_{\kappa, s}$ (whatever
the sign of $\kappa$ is) via the actions
\begin{eqnarray}
& & b^+ \vert n \rangle = \sqrt{F(n+1)} e^{{-i [F(n+1)- F(n)] \varphi }} \vert n+1 \rangle, \nonumber \\
& & b^- \vert n \rangle = \sqrt{F(n)} e^{{ i [F(n) - F(n-1)] \varphi }} \vert n-1 \rangle, \label{action des b sur les n} \\
& & b^- \vert 0 \rangle = 0 \qquad b^+ \vert s-1 \rangle = 0 \qquad
N \vert n \rangle = n \vert n \rangle
\nonumber
\end{eqnarray}
for $n = 0, 1, \ldots, s-1$. Note that a further condition is necessary here, namely, the upper limit condition
$b^+ \vert s-1 \rangle = 0$. It can be checked that the recurrence relation (\ref{recurrence}) is equally valid
for ${\cal A}_{\kappa, s}$. Therefore, equations (\ref{eigenvalue}) and (\ref{eigenvalue avec d}) can be applied
with $n = 0, 1, \ldots, s-1$.
It is interesting to note that the creation and annihilation operators $b^-$ and $b^+$
satisfy (in the representation under consideration) the nilpotency relations
\begin{eqnarray}
(b^-)^s = (b^+)^s = 0,
\end{eqnarray}
which are similar (for $s = k \in \mathbb{N} \setminus {0,1})$ to those
describing the so-called $k$-fermions that are objects interpolating between
fermions (for $k = 2$) and bosons (for $k \to \infty$) \cite{DaoHasKib}.
\subsection{{Phase states for ${\cal A}_{\kappa, s}$}}
For the truncated algebra ${\cal A}_{\kappa, s}$ (corresponding to $d(\kappa)$
finite or infinite), the analog of the phase operator $E_d$ is the unitary operator
\begin{eqnarray}
E_s := e^{i [F(0) - F(s-1)] \varphi } \vert s-1 \rangle \langle 0 \vert +
\sum_{n=1}^{s-1} e^{i [F(n) - F(n-1)] \varphi } \vert n-1 \rangle \langle n \vert.
\end{eqnarray}
By using the same reasoning as in subsection 3.2, we obtain
\begin{eqnarray}
E_s \vert m , \varphi \rangle = e^{i\theta_m} \vert m , \varphi \rangle \qquad \theta_m = m \frac{2 \pi}{s},
\end{eqnarray}
where
\begin{eqnarray}
\vert m , \varphi \rangle = \frac{1}{\sqrt s} \sum_{n=0}^{s-1} e^{-i F(n) \varphi} (q_s)^{mn} \vert n \rangle,
\label{coherentstatemvarphi pour Akappas}
\end{eqnarray}
with $m \in \mathbb{Z} / s \mathbb{Z}$, $\varphi \in \mathbb{R}$ and $q_s$ given by
\begin{eqnarray}
q_s : = e^{2 \pi i / s}.
\label{definition de qs}
\end{eqnarray}
We are thus left with phase states $\vert m , \varphi \rangle$ associated with the phase operator $E_s$. These states
satisfy the same properties as those for $E_d$ (see section 3.2) except that $d$ is replaced by $s$ in some places.
\subsection{{A new type of discrete phase states}}
It is well known that, for quantum systems with a finite spectrum (like the Morse system) or for Lie algebras with
finite-dimensional unitary representations (as for instance $su(2)$), the construction of coherent states cannot be
achieved by looking for the eigenstates of an annihilation operator \cite{GazeauK} or of a compact shift operator
\cite{BarutGi}.
For the algebra ${\cal A}_{\kappa, s}$ the difficulty inherent to the finiteness of the representation can be overcome
as follows. We define the operator
\begin{eqnarray}
V_s := b^- + \frac{(b^+)^{s-1}}{E(s-1)},
\label{operatorVs}
\end{eqnarray}
where the function $E$ is defined via
\begin{eqnarray}
E(0) := 1 \qquad E(n) := F(1) F(2) \ldots F(n) \qquad n = 1, 2, \ldots, s-1.
\end{eqnarray}
The operator $V_s$ is an idempotent operator of order $s$ since
\begin{eqnarray}
(V_s)^s = I.
\label{idempotencyVs}
\end{eqnarray}
Let us consider the eigenvalue equation
\begin{eqnarray}
V_s \vert z \rangle = z \vert z \rangle \qquad
\vert z \rangle = \sum_{n=0}^{s-1} C_n z^n \vert n \rangle \qquad z \in \mathbb{C}.
\end{eqnarray}
By using (\ref{idempotencyVs}), we obtain that $z$ is discretized as
\begin{eqnarray}
z = (q_s)^\mu \qquad \mu \in \mathbb{Z}/s\mathbb{Z},
\end{eqnarray}
with $q_s$ defined by (\ref{definition de qs}). Then, it is a simple matter to calculate the coefficients $C_n$ and to normalize
the $\mu$- and $\varphi$-dependent states $| z \rangle \equiv | \mu , \varphi \rangle$.
This leads to
\begin{eqnarray}
\vert \mu , \varphi \rangle = C_0 \sum_{n=0}^{s-1} \frac{1}{\sqrt{E(n)}} (q_s)^{n \mu} e^{-i F(n) \varphi} \vert n \rangle,
\label{phase states for Akappas}
\end{eqnarray}
where the normalization factor $C_0$ is such that (up to a phase factor)
\begin{eqnarray}
{\it C_0}^{-2} = \sum_{n=0}^{s-1} \frac{1}{E(n)}.
\end{eqnarray}
The states $| \mu , \varphi \rangle$ are temporally stable and are similar to the coherent
states introduced by Gazeau and Klauder \cite{GazeauK} except that their labeling includes an integer and they
correspond to the eigenvectors of a polynomial in terms of generalized creation and annhilation operators. They satisfy
\begin{eqnarray}
\langle \mu , \varphi \vert \mu' , \varphi' \rangle = C_0^2 \sum_{n=0}^{s-1} \frac{1}{E(n)} (q_s)^{n(\mu'-\mu)} e^{-i F(n) (\varphi' - \varphi)}
\end{eqnarray}
and
\begin{eqnarray}
\frac{1}{s} \sum_{m=0}^{s-1} \vert \mu , \varphi \rangle \langle \mu , \varphi \vert =
C_0^2 \sum_{n=0}^{s-1} \frac{1}{E(n)} \vert n \rangle \langle n \vert.
\end{eqnarray}
Consequently, they are not orthogonal.
We close this subsection with a remark concerning the unitary operator
\begin{eqnarray}
U_s := (q_s)^N
\end{eqnarray}
that is a companion of $V_s$ in the following sense. This operator satisfies the cyclicity relation
\begin{eqnarray}
(U_s)^s = I.
\label{idempotencyUs}
\end{eqnarray}
Furthermore, we have the $s$-commutation relation
\begin{eqnarray}
V_sU_s - q_s U_sV_s = 0.
\label{scommutation}
\end{eqnarray}
Equations (\ref{idempotencyVs}), (\ref{idempotencyUs}) and (\ref{scommutation}) are necessary conditions
for the pair ($U_s, V_s$) be a pair of Weyl (see \cite{KibJPhysA0809}). However, this is not the case
because $V_s$ is not unitary.
\section{Application to mutually unbiased bases}
As an {\it a priori} unexpected connection, the approach in subsection 3.2 and 4.2 for the finite-dimensional
cases (for ${\cal A}_{\kappa}$ and ${\cal A}_{\kappa, s}$) can be further developed for deriving MUBs. Let us
recall that two orthonormal bases $\{ \vert a \alpha \rangle : \alpha = 0, 1, \ldots, d-1 \}$
and $\{ \vert b \beta \rangle : \beta = 0, 1, \ldots, d-1 \}$ in a $d$-dimensional Hilbert space
(with an inner product $\langle \, \vert \, \rangle$) are said to be mutually unbiased iff
\begin{eqnarray}
\vert \langle a \alpha \vert b \beta \rangle \vert = \delta_{a,b} \delta_{\alpha,\beta} + \frac{1}{\sqrt{d}} (1 - \delta_{a,b}).
\end{eqnarray}
For fixed $d$, it is known that the number ${\cal N}$ of MUBs is such that ${\cal N} \leq d+1$ and that the limit ${\cal N}=d+1$
is reached when $d$ is the power of a prime number \cite{Ivanovic, WoottersFields}.
\subsection{MUBs from phase states for ${\cal A}_{\kappa}$}
In order to generate MUBs along the line of the developments of subsection 3.2,
let us further examine some properties of the phase operator $E_d$ for ${\cal A}_{\kappa}$
with $\kappa < 0$. This operator can be written in a compact form as
\begin{eqnarray}
E_d = \sum_{n=0}^{d-1} e^{i [F(n) - F(n-1)] \varphi } \vert n-1 \rangle \langle n \vert
\end{eqnarray}
(in this section, the summations on $n$ are understood modulo $d$). It is easy to check that
\begin{eqnarray}
\left( E_d \right)^d = I,
\end{eqnarray}
so that $E_d$ is idempotent. The operator $E_d$ can be decomposed as
\begin{eqnarray}
E_d = U_{\varphi} V,
\end{eqnarray}
where the operators $U_{\varphi}$ and $V$ are defined by
\begin{eqnarray}
U_{\varphi} := e^{i [F(N+1) - F(N)] \varphi} \qquad V := \sum_{n=0}^{d-1} \vert n-1 \rangle \langle n \vert.
\end{eqnarray}
The operators $U_{\varphi}$ and $V$ are unitary and satisfy the pseudo-commutation relation
\begin{eqnarray}
U_{\varphi} V = e^{2 i \varphi / (d-1)} V U_{\varphi}.
\label{pseudocommutation}
\end{eqnarray}
In addition, the operator $V$ satisfies the idempotency relation
\begin{eqnarray}
V^d = I
\end{eqnarray}
and, when the parameter $\varphi$ is quantized as
\begin{eqnarray}
\varphi = - \pi \frac{d-1}{d} p \qquad p \in \mathbb{Z}/d\mathbb{Z},
\label{phidiscrete}
\end{eqnarray}
we have
\begin{eqnarray}
\left( U_{\varphi} \right)^d = e^{i \pi (d-1)p} I.
\end{eqnarray}
In view of (\ref{phidiscrete}), equation (\ref{pseudocommutation}) can be rewritten as
\begin{eqnarray}
VU_{\varphi} = q^p U_{\varphi}V
\end{eqnarray}
(see (\ref{definition of q}) for the definition of $q$). For the discrete values of $\varphi$ afforded
by (\ref{phidiscrete}), equation (\ref{coherentstatemvarphi}) yields the phase states
$\vert m , \varphi \rangle \equiv \vert m , p \rangle$ given by
\begin{eqnarray}
\vert m , p \rangle = \frac{1}{\sqrt{d}} \sum_{n = 0}^{d-1} q^{n(d-n)p/2 + nm} \vert n \rangle
\qquad p, m \in \mathbb{Z}/d\mathbb{Z},
\label{MUBpm}
\end{eqnarray}
which coincides with the vector $\vert a \alpha \rangle$, with $a \equiv p$ and $\alpha \equiv m$,
obtained in \cite{KibKPAlbKib} in an $SU(2)$ approach to MUBs. Alternatively, by putting
\begin{eqnarray}
k := d - n - 1 \qquad \vert n \rangle = \vert d - k - 1 \rangle \equiv \vert k \rangle,
\end{eqnarray}
equation (\ref{MUBpm}) becomes
\begin{eqnarray}
\vert m , p \rangle = \frac{1}{\sqrt{d}} \sum_{k = 0}^{d-1} q^{(k+1)(d-k-1)p/2 - (k+1)m} \vert k \rangle,
\qquad p, m \in \mathbb{Z}/d\mathbb{Z},
\label{soixante dix neuf}
\end{eqnarray}
which coincides with the vector $\vert a \alpha \rangle$, with $a \equiv p$ and $\alpha \equiv m$,
derived in \cite{KibJPhysA0809} in an angular momentum approach to MUBs. It is to be observed that
(\ref{MUBpm}) and (\ref{soixante dix neuf}) correspond to quadratic discrete Fourier transforms.
To make a further contact with \cite{KibKPAlbKib, KibJPhysA0809}, let us note that when $\varphi$
is discretized according to (\ref{phidiscrete}), the inner product
$\langle m , \varphi \vert m' , \varphi' \rangle \equiv \langle m , p \vert m' , p' \rangle$
(see equation (\ref{overlap})) can be rewritten as
\begin{eqnarray}
\langle m , p \vert m' , p' \rangle =
\frac{1}{d} S(u, v, w)
\label{overlapMUB}
\end{eqnarray}
with
\begin{eqnarray}
u := p-p' \qquad v := - (p-p')d + 2(m' - m) \qquad w := d.
\end{eqnarray}
In equation (\ref{overlapMUB}), the factor $S(u, v, w)$ denotes a generalized quadratic Gauss sum defined
by \cite{les2Berndt}
\begin{eqnarray}
S(u, v, w) := \sum_{k=0}^{|w|-1} e^{i \pi (uk^2 + vk) / w},
\label{Gauss sum}
\end{eqnarray}
where $u$, $v$ and $w$ are integers (the nonvanishing of $S(u,v,w)$ requires $uw + v$ even). In the special case where $d$ is a prime integer and $p' \not= p$, the calculation
of $S(u, v, w)$ in (\ref{overlapMUB}) through the methods developed in \cite{les2Berndt, lesquatre} (see also \cite{KibKPAlbKib}) leads to
\begin{eqnarray}
\vert \langle m , p \vert m' , p' \rangle \vert = \frac{1}{\sqrt{d}}.
\end{eqnarray}
This result shows that the $d$ bases
\begin{eqnarray}
B_p := \{ | m , p \rangle : m = 0, 1, \ldots, d-1 \} \qquad p = 0, 1, \ldots, d-1
\end{eqnarray}
of the $d$-dimensional space ${\cal F}_{\kappa}$, with $d$ given by (\ref{dimension}), are mutualy unbiased. On
the other hand, in view of (\ref{computational et MUB}), it is clear that any basis $B_p$ and the basis
\begin{eqnarray}
B_d := \{ | n \rangle : n = 0, 1, \ldots, d-1 \},
\end{eqnarray}
known as the computational basis in quantum information and quantum computation, are mutually unbiased. As a conclusion, for
$d$ prime, the $d$ bases $B_p$ with $p = 0, 1, \ldots, d-1$ and the computational basis $B_d$ constitute a complete set of
$d+1$ MUBs. This result, in agreement with the one derived in \cite{KibKPAlbKib, KibJPhysA0809}, is the starting point for
constructing MUBs in power prime dimension.
\subsection{MUBs from phase states for ${\cal A}_{\kappa, s}$}
By applying a discretization procedure similar to the one introduced in subsection 5.1, we
can construct MUBs from the phase states (\ref{coherentstatemvarphi pour Akappas}) for
the truncated algebra ${\cal A}_{\kappa , s}$ with $\kappa \not= 0$. Let us quantize the
parameter $\varphi$ by putting
\begin{eqnarray}
\varphi = \pi \frac{2}{s \kappa} p \qquad p \in \mathbb{Z} / s \mathbb{Z}.
\end{eqnarray}
Then, equation (\ref{coherentstatemvarphi pour Akappas}) leads to the states $| m , \varphi \rangle \equiv | m , p \rangle$
given by
\begin{eqnarray}
\vert m , p \rangle = \frac{1}{\sqrt{s}} \sum_{n = 0}^{s-1} (q_s)^{n(\delta - n)p + nm} \vert n \rangle
\qquad p, m \in \mathbb{Z}/s\mathbb{Z},
\label{MUBmp pour Akappas}
\end{eqnarray}
where
\begin{eqnarray}
\delta := 1 - \frac{1}{\kappa}.
\end{eqnarray}
In this subsection, we shall assume that $1 / \kappa \in \mathbb{Z}$ so that $\delta \in \mathbb{Z}$ (note that $\delta = d$ for $\kappa < 0$).
The overlap $\langle m , p \vert m' , p' \rangle$ can be written in terms of the generalized quadratic Gauss sum (\ref{Gauss sum}). Indeed, we have
\begin{eqnarray}
\langle m , p \vert m' , p' \rangle = \frac{1}{s} \sum_{n=0}^{s-1} (q_s)^{n(\delta - n) (p' - p) + n(m' - m)} = \frac{1}{s} S(u, v, w),
\end{eqnarray}
where
\begin{eqnarray}
u := 2(p-p') \qquad v := 2 \delta (p'-p) + 2(m'-m) \qquad w := s.
\end{eqnarray}
We can proceed as in subsection 5.1 in order to show that the various states $\vert m , p \rangle$ generate, together
with the $s$-dimensional basis $\{ | n \rangle : n = 0, 1, \ldots, s-1 \}$, $s+1$ MUBs when $s$ is a prime integer.
\section{Application to exactly solvable potentials}
The main goal of this section is to show how the generalized oscillator algebra ${\cal A}_{\kappa}$ is relevant
for the study of one-dimensional exactly solvable potentials in the context of supersymmetric quantum mechanics
and how MUBs can be derived from the temporally stable phase states for some quantum mechanical systems.
\subsection{Creation, annihilation and transfer operators}
Ordinary supersymmetric quantum mechanics can be presented in differnt ways
(e.g., see \cite{29W81}-\cite{MD26fresh}). We
adopt here the approach according to which a supersymmetric dynamical system is defined by a triplet $(H, Q_+, Q_-)_2$ of
linear operators acting on a $\mathbb{Z}_2$-graded Hilbert space ${\cal H}$ and satisfying the following
relations
\begin{eqnarray}
H = H^{\dagger} \qquad Q_- = Q_+^{\dagger} \qquad Q_{\pm}^2=0 \qquad
\{ Q_- , Q_+ \} = H \qquad [ H , Q_{\pm}] = 0.
\label{relations triplet}
\end{eqnarray}
(In this approach, ordinary supersymmetric quantum mechanics is a particular case, corresponding to $k=2$, of fractional
supersymmetric quantum mechanics of order $k$ dealing with triplets $(H, Q_+, Q_-)_k$ which satisfy relations
generalizing (\ref{relations triplet}) and which correspond to a $\mathbb{Z}_k$ grading \cite{DaoKibPLAJMP}.) The
operators $Q_+$ and $Q_-$ are the supercharges of the one-dimensional system. We suppose that the spectrum of the
self-adjoint operator $H$, the supersymmetric Hamiltonian of the system, is discrete. The
Hamiltonian $H$ can be written
\begin{eqnarray}
H = H_0 + H_1,
\end{eqnarray}
where $H_0$ and $H_1$ act on the states $\vert \Psi_n , 0 \rangle$ and $\vert \Psi_n , 1 \rangle$ of even and odd grading,
respectively. In other words, the Hilbert space ${\cal H}$ is decomposed as
\begin{eqnarray}
{\cal H} = {\cal H}_0 \oplus {\cal H}_1 \qquad
{\cal H}_0 := \{\vert \Psi_n , 0 \rangle : n \ {\rm ranging} \} \qquad
{\cal H}_1 := \{\vert \Psi_n , 1 \rangle : n \ {\rm ranging} \},
\end{eqnarray}
which reflects the ${\mathbb{Z}}_2$ grading. We shall assume that there is no supersymmetry breaking. In this case,
the Hamiltonians $H_0$ and $H_1$ are isospectral except that the ground state of $H_0$ has no supersymmetric
partner in the spectrum of $H_1$.
By combining the above-mentioned considerations on supersymmetry with the Infeld and Hull factorization method
\cite{27Infeld}, we can construct creation, annihilation and transfer operators for an exactly solvable
Hamiltonian in one dimension \cite{29W81}-\cite{russes}. For this purpose, let us consider a one-dimensional quantum system embedded
in a real potential $v_0 : x \mapsto v_0(x)$. The corresponding Hamiltonian is
\begin{eqnarray}
H_0 := - \frac{1}{2} \frac{d^2}{dx^2} + v_0.
\end{eqnarray}
Let us suppose that the Hamiltonian $H_0$ is exactly solvable and admits the discrete spectrum
\begin{eqnarray}
e_0 = 0 < e_1 < e_2 < \ldots < e_n < e_{n+1} < \ldots,
\end{eqnarray}
with a finite or infinite number of levels. We know that the Hamiltonian $H_0$ of this system can be factorized as
\cite{31CKS9501, 30J96, 27Infeld, MD27Mielnik, MD26fresh}
\begin{eqnarray}
H_0 = x^+ x^- \qquad
x^+ := \frac{1}{\sqrt{2}} \left( -\frac{d}{dx} + w \right) \qquad
x^- := \frac{1}{\sqrt{2}} \left( \frac{d}{dx} + w \right).
\end{eqnarray}
The superpotential $w : x \mapsto w(x)$ satisfies the Ricatti equation
\begin{eqnarray}
v_0 = \frac{1}{2} \left( w^2 - \frac{dw}{dx} \right).
\end{eqnarray}
Since the ground state energy is assumed to be zero, it is easy to see that the potential
$v_0$ and the superpotential $w$ can be expressed in terms of the ground state wavefunction.
It is important to stress that the operators $x^+$ and $x^-$ are not in general creation and annihilation operators for $H_0$
\cite{31CKS9501, 30J96, MD27Mielnik, MD26fresh, russes}. They
are indeed transfer operators from the spectrum of $H_0$ to the one of $H_1$ and vice-versa. To identify them, we start by
representing the supercharge operators and the supersymmetric Hamiltonian by $2 \times 2$ matrices
\cite{31CKS9501, 30J96, 32CGK04, MD26fresh}
\begin{eqnarray}
Q_- =
\pmatrix{
0 & x^- \cr
0 & 0 \cr
}
\qquad
Q_+ =
\pmatrix{
0 & 0 \cr
x^+ & 0 \cr
}
\qquad
H =
\pmatrix{
H_1 & 0 \cr
0 & H_0 \cr
},
\end{eqnarray}
where the operator
\begin{eqnarray}
H_1 = x^-x^+ = - \frac{1}{2} \frac{d^2}{dx^2} + v_1
\end{eqnarray}
is the supersymmetric partner of $H_0$ and corresponds to a new potential $v_1 : x \mapsto v_1(x)$. The potential
\begin{eqnarray}
v_1 = \frac{1}{2} \left( w^2 + \frac{dw}{dx} \right)
\end{eqnarray}
is the supersymmetric partner of the potential $v_0$. The Hamiltonian $H_1$ is also exactly solvable and isospectral to $H_0$ (except for the ground state). Indeed,
\begin{eqnarray}
H_0 \vert \Psi_n , 0 \rangle = e^0_n \vert \Psi_n , 0 \rangle \Rightarrow
H_1 (x^- \vert \Psi_n , 0 \rangle) = e^0_n (x^- \vert \Psi_n , 0 \rangle),
\label{action de H1}
\end{eqnarray}
where $e^0_n := e_n$. Similarly,
\begin{eqnarray}
H_1 \vert \Psi_n , 1 \rangle = e^1_n \vert \Psi_n , 1 \rangle \Rightarrow
H_0 (x^+ \vert \Psi_n , 1 \rangle) = e^1_n (x^+ \vert \Psi_n , 1 \rangle).
\label{action de H0}
\end{eqnarray}
(For more details see \cite{31CKS9501, 30J96, MD27Mielnik} and the recent topical review \cite{MD26fresh}.)
From equations (\ref{action de H1}) and (\ref{action de H0}), it is clear that we can take
\begin{eqnarray}
x^- \vert \Psi_{n+1} , 0 \rangle = \sqrt{e_{n+1}}e^{i (e_{n+1}- e_{n}) \varphi} \vert \Psi_n , 1 \rangle
\label{action de xmoins}
\end{eqnarray}
\begin{eqnarray}
x^+ \vert \Psi_n , 1 \rangle = \sqrt{e_{n+1}}e^{-i (e_{n+1}- e_{n}) \varphi} \vert \Psi_{n+1} , 0 \rangle,
\label{action de xplus}
\end{eqnarray}
where $\varphi$ is a real number, and that the energies of the supersymmetric partners $H_0$ and $H_1$ are related by
\begin{eqnarray}
e^1_n = e^0_{n+1}.
\end{eqnarray}
Note that the operator $x^-$ (respectively $x^+$) converts an eigenfunction of $H_0$ (respectively $H_1$) into an
eigenfunction of $H_1$ (respectively $H_0$) with the same energy. Thus, the operators $x^-$ and $x^+$ transfer the
states from one spectrum to its partner spectrum. To introduce the ladder operators inside a given spectrum, we
first consider the unitary operator $U$ relating the states $\vert \Psi_n , 0 \rangle$ and $\vert \Psi_n , 1 \rangle$
through (cf \cite{MD28Faddev}-\cite{MD29Samsonov})
\begin{eqnarray}
U := \sum_{n} \vert \Psi_n , 1 \rangle \langle \Psi_n , 0 \vert \Rightarrow \vert \Psi_n , 1 \rangle = U \vert \Psi_n , 0 \rangle.
\end{eqnarray}
Operators similar to $U$ were already considered for continuous spectra \cite{MD28Faddev, MD29Samsonov} and for discrete
spectra \cite{MD30Pursey, MD31Kumar}. Then, we define the operators \cite{MD30Pursey}-\cite{MD31Kumar}
\begin{eqnarray}
a^+ := x^+ U \qquad a^- := U^{\dagger}x^-.
\end{eqnarray}
By using equations (\ref{action de xmoins}) et (\ref{action de xplus}), we obtain
\begin{eqnarray}
a^- \vert \Psi_n , 0 \rangle = \sqrt{e_{n}} e^{ i (e_{n}- e_{n-1}) \varphi } \vert \Psi_{n-1} , 0 \rangle
\label{action de amoins sur Psin0}
\end{eqnarray}
\begin{eqnarray}
a^+ \vert \Psi_n , 0 \rangle = \sqrt{e_{n+1}} e^{-i (e_{n+1}- e_{n}) \varphi } \vert \Psi_{n+1} , 0 \rangle.
\label{action de aplus sur Psin0}
\end{eqnarray}
Consequently, $a^+$ and $a^-$ are creation and annihilation operators for the Hamiltonian $H_0$. Furthermore,
it is easily seen that
\begin{eqnarray}
a^+ a^- = x^+ x^- = H_0.
\label{a x H0}
\end{eqnarray}
Ladder operators for the Hamiltonian $H_1$ can be introduced in a similar way.
\subsection{Physical realizations of the generalized oscillator algebra}
To simplify the notation, we set $\vert \Psi_n \rangle := \vert \Psi_n , 0 \rangle$.
From equations (\ref{action de amoins sur Psin0}) et (\ref{action de aplus sur Psin0}), we get
\begin{eqnarray}
[ a^- , a^+] \vert \Psi_n \rangle = (e_{n+1} - e_n) \vert \Psi_n \rangle.
\end{eqnarray}
The number operator $N$ defined by
\begin{eqnarray}
N \vert \Psi_n \rangle = n \vert \Psi_n \rangle
\end{eqnarray}
is in general (for an arbitrary quantum system) different from the product $a^+a^-$.
Let us consider the situation where the creation
and annihilation operators satisfy the commutation relation
\begin{eqnarray}
[ a^- , a^+] = a N + b,
\end{eqnarray}
a relation used in the study of the so-called polynomial Heisenberg algebra introduced in \cite{MD32Fernandez}. In
other words, we assume that the energy gap $e_{n+1} - e_n$ between two succussive levels is linear in $n$, i.e.
\begin{eqnarray}
e_{n+1} - e_n = a n + b,
\end{eqnarray}
where $a$ and $b$ are two real parameters. We also assume that the eigenvalues of the operator $aN+b$
are positive. With these choices, the
algebra generated by the operators $a^+$, $a^-$ and $N$ is identical to the generalized oscillator algebra
${\cal A}_{\kappa}$ modulo the replacements
\begin{eqnarray}
a^{\pm} \rightarrow \frac{a^{\pm}}{\sqrt{b}} \qquad \kappa \rightarrow \frac{1}{2} \frac{a}{b}
\end{eqnarray}
in equation (\ref{thealgebra}). Thus, from equations (\ref{action de amoins sur Psin0}-\ref{a x H0}), we have
\begin{eqnarray}
H_0 = a^+ a^- = \frac{1}{2} a N(N-1) + b N.
\end{eqnarray}
For $a \not= 0$, the spectrum of $H_0$ is non-linear and is given by
\begin{eqnarray}
H_0 \vert \Psi_n \rangle = e_n\vert \Psi_n \rangle \qquad e_n = \frac{1}{2} a n(n-1) + b n.
\label{add124}
\end{eqnarray}
Particular realizations of (\ref{add124}) in terms of one-dimensional solvable potentials were previously considered
in \cite{Antoine, Kinani, DaoKibPLAJMP, MD34Quesne99, MD35Angelova}. Following the developments in \cite{DaoKibPLAJMP},
we consider the following remarkable cases.
(i) For ($a = 0$, $b > 0$), the spectrum of $H_0$
is infinite-dimensional ($n \in \mathbb{N}$) and does not present degeneracies.
(ii) For ($a > 0$, $b \geq 0$), the spectrum of $H_0$
is infinite-dimensional ($n \in \mathbb{N}$) and does not present degeneracies.
(iii) For ($a < 0$, $b \geq 0$), the spectrum of $H_0$ is finite-dimensional with $n = 0, 1, \ldots, s-1$ where
\begin{eqnarray}
s & = & -\frac{b}{a} + \frac{3}{2} \ {\rm for} \ - 2\frac{b}{a} \ {\rm odd} \\
s & = & -\frac{b}{a} + 1 \ {\rm for} \ - 2\frac{b}{a} \ {\rm even},
\end{eqnarray}
and all the states are nondegenerate.
It is possible to find a realization of each of the three cases above in terms of exactly solvable dynamical systems in one dimension. We give below the corresponding potential $v_0$ and transfer operators.
(i) The case ($a = 0$, $b = 1$) corresponds to the harmonic oscillator
(for which $n \in \mathbb{N}$) with
\begin{eqnarray}
v_0(x) = \frac{1}{2} (x^2 - 1)
\end{eqnarray}
and
\begin{eqnarray}
x^{\pm} \equiv a^{\pm} = \frac{1}{\sqrt{2}} \left( \mp \frac{d}{dx} + x \right).
\end{eqnarray}
(For the harmonic oscillator, $U$ reduces to the identity operator.)
(ii) The case ($a = 1$, $2b = u+v+1$), with $u > 1$ and $v >1 $, corresponds to the P\" oschl-Teller potential
(for which $n \in \mathbb{N}$) with
\begin{eqnarray}
v_0(x) = \frac{1}{8} \bigg[\frac{u(u-1)}{\sin^2 \frac{x}{2}} + \frac{v(v-1)}{\cos^2 \frac{x}{2}}\bigg] -\frac{1}{8}(u+v)^2
\end{eqnarray}
and
\begin{eqnarray}
x^{\pm} = \frac{1}{\sqrt{2}} \left[ \mp \frac{d}{dx} + \frac{1}{2}\bigg(u \cot\frac{x}{2} - v \tan\frac{x}{2}\bigg) \right].
\end{eqnarray}
(iii) The case ($a = -1$, $2b = 2l-1$), with $l \in \mathbb{N}^*$, corresponds to the Morse potential
(for which $n = 0, 1, \ldots, l$) with
\begin{eqnarray}
v_0(x) = \frac{1}{2} \left[ e^{-2x} - (2l+1)e^{-x} + l^2 \right],
\end{eqnarray}
and
\begin{eqnarray}
x^{\pm} = \frac{1}{\sqrt{2}} \left( \mp \frac{d}{dx} + l - e^{-x} \right).
\end{eqnarray}
\subsection{{Phase states and MUB for exactly solvable systems}}
From equation (\ref{coherentstatemvarphi pour Akappas}), we can obtain the phase states for a general
quantum system described by a truncated generalized oscillator algebra ${\cal A}_{\kappa , s}$. We get
\begin{eqnarray}
\vert m , \varphi \rangle = \frac{1}{\sqrt s} \sum_{n=0}^{s-1} e^{-i e_n \varphi} (q_s)^{nm} \vert \Psi_n \rangle,
\label{coherentstatemvarphi pour Akappasbis}
\end{eqnarray}
with $s$ sufficiently large for the harmonic oscillator and the
P\" oschl-Teller systems and $s = l+1$ for the Morse system. Furthermore,
equation (\ref{MUBmp pour Akappas}) provides with a mean to generate MUBs associated with
the cases (i), (ii) and (iii) of subsection 6.2.
On the other hand, the discrete phase state (\ref{phase states for Akappas}) reads here
\begin{eqnarray}
\vert \mu , \varphi \rangle = C_0 \sum_{n=0}^{s-1} \frac{1}{\sqrt{E(n)}} e^{-i e_n \varphi} (q_s)^{n \mu} \vert \Psi_n \rangle,
\label{phase states for Akappas avec exactly s s}
\end{eqnarray}
where the factor $E(n)$ can be calculated in the different cases (i), (ii) and (iii). A simple calculation
gives the following results in term of the $\Gamma$ function.
(i) For the harmonic oscillator potential:
\begin{eqnarray}
E(n) = \Gamma(n+1).
\end{eqnarray}
(ii) For the P\" oschl-Teller potential:
\begin{eqnarray}
E(n)= \frac{\Gamma(n+1)\Gamma(n+u+v+1)}{2^n \Gamma(u+v+1)}.
\end{eqnarray}
(iii) For the Morse potential:
\begin{eqnarray}
E(n) = \frac{\Gamma(n+1)\Gamma(2l)}{2^n \Gamma(2l-n)}.
\end{eqnarray}
It should be mentioned that the discrete phase states given by (\ref{phase states for Akappas avec exactly s s})
differ from the coherent states for exactly sovable potentials derived in \cite{Kinani, MD31Kumar, MD32Fernandez, MD35Angelova, MD36Bagrov}
from supersymmetric quantum mechanics techniques. The noticeable difference comes from the fact that the states
(\ref{phase states for Akappas avec exactly s s}) are temporally stable and are labeled by an integer instead of
a continuous complex variable as in the coherent states derived in \cite{Kinani, MD31Kumar, MD32Fernandez, MD35Angelova, MD36Bagrov}.
The states (\ref{phase states for Akappas avec exactly s s}) are eigenstates of the operator (\ref{operatorVs}) whereas the coherent states
in \cite{Kinani, MD31Kumar, MD32Fernandez, MD35Angelova, MD36Bagrov} are obtained from the three standard definitions (involving
annihilation operator, displacement operator, and uncertainty relation).
\section{Concluding remarks}
The starting point of this article is based on the definition of a generalized oscillator algebra ${\cal A}_{\kappa}$. This
algebra is interesting in two respects. First, it describes in an unified way some exactly solvable one-dimensional
systems having a nonlinear spectrum (for $\kappa \not= 0$) or a linear spectrum (for $\kappa = 0$). As typical examples, these quantum systems correspond to the P\" oschl-Teller potential (for $\kappa > 0$), the Morse potential (for $\kappa
< 0$) and the infinite square well potential (for $\kappa = 1/3$) in addition to the harmonic oscillator potential (for
$\kappa = 0$). Second, the algebra ${\cal A}_{\kappa}$ can take into account some nonlinear effects that may occur in
the quantum description of quantized modes of the electromagnetic field (cf. \cite{Walls}).
In connection with the algebra ${\cal A}_{\kappa}$, the present work adresses three problems: the construction of a
phase operator, the determination of its temporally stable eigenstates (the so-called phase states) and the derivation
of MUBs from the obtained phase states. This is the first time that a connection between MUBs and dynamical systems is
established. In this regard, the character "temporally stable" of the eigenstates of the phase operator is essential for the derivation of MUBs. The main results of this paper are as follows.
For the case $\kappa \geq 0$ (which corresponds to an infinite representation of ${\cal A}_{\kappa}$), the phase operator is not unitary. We note in passing that the corresponding phase states are similar to those derived in \cite{VoudasBM} except that our states are temporally stable. However for
$\kappa \geq 0$, by making a ({\it \`a la} Pegg and Barnett) truncation, which gives rise to a truncated generalized
oscillator algebra ${\cal A}_{\kappa , s}$, we can define a unitary phase operator whose eigenstates lead to MUBs.
For the case $\kappa < 0$ (which corresponds to a finite representation of ${\cal A}_{\kappa}$), it is possible to
construct a unitary phase operator whose eigenstates are temporally stable. MUBs can be derived as a subset of
these states. For $\kappa < 0$, the consideration of a truncated generalized oscillator algebra ${\cal A}_{\kappa , s}$
is nevertheless necessary in order to establish a connection with the Morse system and to derive associated MUBs.
As a conclusion, in both cases ($\kappa \geq 0$ and $\kappa < 0$), the truncation procedure makes it possible to
define a unitary phase operator for exactly solvable systems and to generate temporally stable phase states from
which MUBs can be derived.
Another result of this paper concerns a new type of phase states. These temporally stable phase states, namely
the states (\ref{phase states for Akappas}), are associated with the truncated algebra ${\cal A}_{\kappa , s}$.
They are eigenstates of an operator defined in the enveloping algebra of ${\cal A}_{\kappa , s}$ and constitute discrete
analogs of the coherent states derived in \cite{GazeauK}. More generally, this result shows that it is possible,
for a finite spectrum, to derive new phase states similar to the coherent states of \cite{GazeauK} constructed,
for an infinite spectrum, as eigenstates of an annihilation operator. The key of the derivation of the new states
(for a finite spectrum) is to add a power of the creation operator to the annihilation operator.
To close this paper, let us mention that the concept of MUBs was recently extended to infinite-dimensional
Hilbert spaces \cite{Weigert}. In this vein, it is hoped that the temporally stable phase states derived in this
work for the infinite-dimensional case could serve as a hint for deriving MUBs for continuous variables, a
difficult challenge.
\section*{Acknowledgments}
One of the authors (M D) would like to thank the hospitality and kindness extended to him by the {\it Groupe de physique
th\'eorique de l'Institut de Physique Nucl\'eaire de Lyon} where this work was done. The other author (M R K) is grateful
to Michel Capdequi-Peyran\`ere for useful comments. Thanks are due to one Referee and to the Adjudicator for constructive
suggestions.
\newpage
\baselineskip=14pt
|
\section{Introduction}
Interferometry is the method of choice in achieving the most precise measurements, or when trying to detect the most feeble effects or signals. Interferometry with matter waves \cite{cronin:09} became a very versatile tool with many applications ranging from precision experiments to fundamental studies.
In an interferometer, an incoming \textit{'beam'} (light or matter ensemble) is split into two parts (pathways), which can be separated in either internal state space or real space. The splitting process prepares the two paths with a well-defined relative phase $\delta \theta(t=0) = \theta_1(t=0) - \theta_2(t=0)$. After the splitting they evolve separately, and can accumulate different phases $\theta_1(t)$ and $\theta_2(t)$ due to the different physical settings they evolve in. Finally, in the recombination process after time $T$ the relative phase $\delta \theta (T) = \theta_1(T) - \theta_2(T)$ accumulated in the two paths can be read out.
The sensitivity of an interferometer measurement depends now on two distinct points: How good can the phase difference $\delta \theta$ be measured, and how long can one accumulate a phase difference in the split paths. For perfect read-out contrast and standard (binomial) splitting and recombination procedures, the uncertainty in determining $\Delta \theta$ is given by the standard quantum limit $\Delta \theta = 1/\sqrt{N}$, where $N$ is the number of registered counts (e.g. atom detections). The second point concerns the question of how long the beams can be kept in the 'interaction region' of the distinguishable interferometer arms, and for how long the paths stay coherent. Ultra cold (degenerate) atoms can be held and manipulated in well controlled traps and guides, and therefore promise ultimate precision and sensitivity for interferometry. Both dipole traps \cite{grimm:00} and atom chips \cite{folman:00,folman:02,Reichel:2002,fortagh:07} have been used to analyze different interferometer geometries \cite{hinds:01,haensel:01,andersson:02,Kreutzmann:04} and employed for experimental demonstration of splitting \cite{Houde:00,cassettari:00,cassettari:00b,Dumke:02} and interference \cite{shin:04,schumm:05,wang:05,albiez:05,hofferberth:07,jo:07,jo2:07,hofferberth:08,esteve:08,bohi:09} with trapped or guided ultra cold atoms.
The power of interferometry lies in the precision and robustness of the phase evolution, which provides the measurement stick. This robustness of the phase evolution is based on the linearity of time propagation in the different paths, which is the case for most light interferometers, atom interferometers \cite{cronin:09} with weak and dilute beams, or neutron interferometers \cite{rauch:74}. Measurements loose precision when this robustness of the phase evolution cannot be guaranteed. This is the case if the phase evolution in the paths depends on the intensity (density), that is, when the time evolution becomes non-linear. Atom optics is fundamentally non-linear, the non-linearity being created by the interaction between atoms. For ultra cold (degenerate) trapped Bose gases this mean field energy associated with the atom-atom interaction can even dominate the time evolution. Consequently, in many interferometer experiments with trapped atoms the atom-atom interaction creates a non-linearity in the time propagation, and the accumulated phase depends on the local atomic densities. Thus, number fluctuations induced by the splitting cause phase diffusion \cite{lewenstein:96,javanainen:97}, which currently limits the coherence and sensitivity of interferometers with trapped atoms much more then decoherence coming from other sources like the surface \cite{henkel:03,skagerstam:06,hohenester.pra2:07}.
In the present manuscript we will discuss the physics that leads to degradation of performance of an atom interferometer with trapped atoms, and how one can counteract it by using optimal input states. We first discuss the performance of a trapped atom interferometer in the simplest two-mode model. This will allow us to illustrate the basic physics. We then investigate how this simple two-mode model has to be modified when taking into account the many-body structure of the wave function. Optimal control techniques are applied to prepare the desired input states \cite{hohenester.pra:07,grond.pra:09,grond.pra:09b}.
In our calculations we always assume zero temperature. The effects of additional dephasing and decoherence due to fundamental quantum noise and due to thermal excitations at finite temperature will be discussed in the last section.
\section{Two-mode model description of atom interferometry}\label{sec:2mode}
When atom interferometry is performed with a trapped Bose Einstein condensate (BEC), we consider the following key stages: The splitting stage (with the time duration $T_{\rm split}$), where the condensate wave function is split into two parts, the phase accumulation stage ($T_{\rm phase}$), where the atoms in one arm of the interferometer experience an interaction with some weak (classical) field, and finally the read-out stage ($T_{\rm tof}$), where the phase accumulated is measured after the condensates have expanded in time-of-flight (TOF).
In our theoretical description, we start by introducing a simple but generic description scheme of an interferometer in terms of a two-mode model for the split condensate. Such a model has been also proven successful for the description of interference with spin squeezed states \cite{oblak:05,fernholz:08} and of condensates in double wells \cite{albiez:05,esteve:08}. To properly account for the many-boson wave function, we introduce the field operator in second-quantized form \cite{leggett:01}
\begin{equation}\label{eq:fieldop}
\hat\Psi(x)=\hat a_L\,\phi_L(x)+\hat a_R\,\phi_R(x)\,.
\end{equation}
Here, $a_{L,R}^\dagger$ are bosonic field operators that create an atom in the left or right well, with wavefunctions $\phi_{L,R}(x)$, respectively. In many cases, we can properly describe the dynamics of the interacting many-boson system by means of a generic \textit{two-mode hamiltonian} \footnote{From now on we use conveniently scaled time, length, and energy units \cite{grond.pra:09b}, unless stated differently. First we set $\hbar=1$ and scale the energy and time according to a harmonic oscillator with confinement length $a_{ho} = \sqrt{\hbar/(m\omega_{ho})}=1$ $\mu m$ and energy $\hbar \omega_{ho}$. Our considerations deal with Rb atoms, we therefore measure mass in units of the $^{87}$Rb atom, and then time is measured in units of $1/\omega_{ho}=1.37$ ms and energy in units of $\hbar \omega_{ho}=2\pi\cdot 116.26$ Hz.} \cite{milburn:97,javanainen:99,grond.pra:09b}
\begin{equation}\label{eq:hamtwomode}
\hat H=-\frac{\Omega}2\left(\hat a_L^\dagger \hat a_R^{\phantom\dagger}+
\hat a_R^\dagger \hat a_L^{\phantom\dagger}\right)+
\kappa\left(
\hat a_L^\dagger\hat a_L^\dagger\hat a_L^{\phantom\dagger}\hat a_L^{\phantom\dagger}+
\hat a_R^\dagger\hat a_R^\dagger\hat a_R^{\phantom\dagger}\hat a_R^{\phantom\dagger}
\right)\,.
\end{equation}
Here $\Omega$ describes a tunnelling process, that allows the atoms to hop between the two wells and $\kappa$ is the non-linear atom-atom interaction, that energetically penalizes states with a high atom-number imbalance between the left and right well. We treat $\Omega$ as a free parameter, but choose a fixed $\kappa\approx U_0/2$. $U_0$ is the effective one-dimensioanl (1D) interaction strength along the direction where the potential is split into a double well.
A more accurate way of how to relate $\kappa$ and $U_0$ will be given in section~\ref{sec:MCTDHB}.
In the example we discuss in this manuscript the initial state of the interferometry sequence is prepared by deforming a confinement potential from a single to a double well. This corresponds to a situation where one starts with the two-mode hamiltonian of equation~\eqref{eq:hamtwomode} for a large tunnel coupling, $\Omega\gg\kappa$, and then turns off $\Omega$, as a consequence of the reduced spatial overlap of the wavefunctions $\phi_{L,R}(x)$.
In the beginning of the splitting sequence tunnelling dominates over the non-linear interaction, and all atoms reside in the bonding orbital $\phi_L(x)+\phi_R(x)$. This results in a binomial atom number distribution. When $\Omega$ is turned off sufficiently fast, and the dynamics due to the non-linear coupling plays no significant role, the orbitals $\phi_{L,R}(x)$ become spatially separated, but the atom number distribution remains binomial. In contrast, when $\Omega$ is turned off sufficiently slowly, the system can adiabatically follow the groundstate of the Hamiltonian \eqref{eq:hamtwomode} and ends up approximately in a Fock state. As we will discuss below, such states with reduced atom number fluctuations are appealing for the purpose of atom interferometry.
\subsection{Pseudospin operators and Bloch sphere}
A convenient representation of the two-mode model for a many-boson system is in terms of angular-momentum operators \cite{milburn:97,javanainen:99,hohenester.fdp:09}. Quite generally, the internal state of an ensemble of atoms which are allowed to occupy two states (here left and right) can be described as a collective (pseudo)spin ${\hat{\bm J}} = \sum_{i=1}^N {\hat{\bm j}_i}$, which is the sum of the individual spins of all atoms. Here the total angular momentum is $N/2$, and the projection $m$ on the $z$-axis corresponds to states where, starting from a state where the left and right well are each populated with $N/2$ atoms, $m$ atoms are promoted from the right to the left well. Through the Schwinger boson representation
\begin{equation}\label{eq:pseudospin}
\mbox{\hspace*{-2cm}}
\hat J_x=\frac 12\left(
\hat a_L^\dagger \hat a_R^{\phantom\dagger}+
\hat a_R^\dagger \hat a_L^{\phantom\dagger}\right)\,,\quad
\hat J_y=-\frac i2\left(
\hat a_L^\dagger \hat a_R^{\phantom\dagger}-
\hat a_R^\dagger \hat a_L^{\phantom\dagger}\right)\,,\quad
\hat J_z = \frac 12\left(
\hat a_L^\dagger \hat a_L^{\phantom\dagger}-
\hat a_R^\dagger \hat a_R^{\phantom\dagger}\right)
\end{equation}
we can establish a link between the field operators $\hat a_{L,R}$ and the pseudo-spin operators. $\hat J_x$ promotes an atom from the left to the right well, or vice versa, and $\hat J_z$ gives half the atom number difference between the two wells.
\begin{figure}
\center\includegraphics[width=0.5\columnwidth]{figure1}
\caption{Bloch sphere representation of a number squeezed state with squeezing factor (defined in text) $\xi_N\approx0.2$. $\Delta J_z$ corresponds to number squeezing, and $\Delta J_y$ is proportional to phase squeezing. In the ring below the sphere we show the polarization of the state along the x-axis, which is proportional to the coherence factor $\alpha=2\langle \hat J_x\rangle/N$. For squeezed states there is also a noise $\Delta J_x$ in the polarization of the state on the equator. }\label{fig:blochintro}
\end{figure}
One can map the two-mode wave function onto the Bloch sphere \cite{arecchi:72}, which provides an extremely useful visualization tool for the purpose of atom interferometry (see figure~\ref{fig:blochintro}). A state on the north pole corresponds to all atoms residing in the left well and a state on the south pole to all atoms in the right well. All atoms in the bonding orbital corresponds to a state localized around $x=N/2$. This is a product state where the atoms are totally uncorrelated. In this state the quantum noise is evenly distributed among $ \Delta J_y = \Delta J_z = {\sqrt{N}}/{2}$, i.e., it has equal uncertainty in number difference (measured along the $z$-axis) and in the conjugate phase observable (measured around the equator of the sphere). Similar to optics with photons, or as discussed in the context of spin squeezing, quantum correlations can reduce the variance of one spin quadrature, for a given angle $\phi$, $\hat J_{\phi}=\cos\phi\hat J_z+\sin\phi\hat J_y$ at the cost of increasing the variance of the orthogonal quadrature: at the angle $\phi$ the variance $\Delta J_\phi ^2$ becomes minimal, whereas the orthogonal variance $\Delta J_{\phi+\pi/2} ^2$ becomes maximal \cite{kitagawa:93}. For example, the squeezed state shown in figure~\ref{fig:blochintro} has reduced number fluctuations, as described by the normalized number squeezing factor $\xi_N=\Delta J_z/(\sqrt{N}/2)$, and enhanced phase fluctuations, as described by the normalized phase squeezing factor $\xi_{\rm phase}=\Delta J_y/(\sqrt{N}/2)$.
Within the pseudo-spin framework, the two-mode Hamiltonian becomes
\begin{equation}\label{eq:hamtwomode.pseudospin}
\hat H=-\Omega\hat J_x+2\kappa \hat J_z^2
\end{equation}
which is completely analogous to the Josephson Hamiltonian of superconductivity. $\Omega$ is associated with the (time-dependent) Josephson energy and the non-linearity $\kappa$ with the charging energy \cite{barone:82}. For a given state on the Bloch sphere, the tunnel coupling rotates the state around the $x$-axis, whereas the non-linear part distorts the state such that the components above and below the equator become twisted to the right- and left-hand side, respectively. The twist rate due to $\hat J_z^2$ increases with distance from the equator (see figure \ref{fig:blochtwist}).
\begin{figure}
\center\includegraphics[width=0.75\columnwidth]{figure2a_2c}
\caption{Time evolution of states on Bloch sphere. The different panels report results for a (a) binomial, (b) number squeezed [$\xi_N\approx0.2$], and (c) phase squeezed state [$\xi_{\phi}\approx0.2$]. The time interval is $T=12.8$ for the phase squeezed state, and $T=32$ otherwise. Due to the non-linear atom-atom interactions the states become distorted, thus spoiling the interferometer performance.}\label{fig:blochtwist}
\end{figure}
\subsection{Readout noise in the interference pattern}
We next address the question which states would be the best for the purpose of \textit{reading} out an atom interferometer. For this purpose we neglect the non-linear atom-atom interaction that distorts the wave function during the phase accumulation time and postpone the question of how the specific states could be actually prepared in experiment. Our discussion (which closely follows Ref.~\cite{jaaskelainen:04}, with some extensions) is primarily intended to set the stage for the later discussion of the full atom interferometer sequence in presence of atom-atom interactions.
We assume that in the phase accumulation stage the wave function has acquired a phase $\theta$. Instead of describing the interaction process dynamically, which would correspond to a Bloch-sphere rotation of the state around the $z$-axis, we directly assign the phase to the single-particle wave function, such that the field operator reads \cite{jaaskelainen:04}
\begin{equation}\label{eq:fieldoptheta}
\hat\Psi(x)=\hat a_L\,\phi_L(x)+\hat a_R\,e^{-i\theta}\phi_R(x)\,.
\end{equation}
To read out $\theta$, one usually turns off the double well potential and lets the two atom clouds overlap in time of flight (TOF). The accumulated phase is then determined from the interference pattern \cite{schumm:05}. After release, the wave function evolves under the free Hamiltonian. Here and in the rest of this paper we ignore the influence of atom-atom interactions during TOF which is justified in low dimensional systems \cite{imambekov:09}, but might be problematic under other circumstances \cite{xiong:06,masiello:07,cederbaum:08}.
As a representative example, we consider for the dispersing wavefunctions $\phi_L(x)$ and $\phi_R(x)$ two Gaussians with variance $\sigma^2$, which are initially separated by the interwell distance $d$. The density operator $\hat n(x)=\hat\Psi^\dagger(x)\hat\Psi(x)$ associated with this TOF measurement can then be computed by using the pseudospin operators of equation~\eqref{eq:pseudospin} as
\begin{equation}\label{eq:densityop}
\mbox{\hspace*{-2cm}}
\hat n(x)=\frac N2(n_L+n_R)+2\sqrt{n_Ln_R}\Bigl(
\cos(kx+\theta)\hat J_x-\sin(kx+\theta)\hat J_y\Bigr)+(n_L-n_R)\hat J_z\,.
\end{equation}
In the following we analyze the mean (ensemble-averaged) atomic density
and its fluctuations around the mean value, such that for a single-shot mea-
surement the outcome is with a high probability within the error bonds
defined by the variance if the probability distribution is Gaussian.
If the initial state preparation was for a symmetric double well potential, then the mean atom number difference vanishes $\langle \hat J_z\rangle=0$, and for the real-valued tunnel coupling we can set $\langle \hat J_y\rangle=0$. The mean density thus becomes
\begin{equation}\label{eq:denop}
\mbox{\hspace*{-2cm}}
n(x)=\left<\hat n(x)\right>=
\frac N2\Bigl(n_L(x)+n_R(x)\Bigr)+2\sqrt{n_L(x)n_R(x)}\cos(kx+\theta)\left<\hat J_x\right>\,,
\end{equation}
which is shown for representative examples in figure~\ref{fig:den}. Here, the visibility of the interference fringes [equation~\eqref{eq:denop}] is determined by the \textit{coherence factor}
\begin{equation}\label{eq:coherence}
\alpha=2\left<\hat J_x\right>/N\,.
\end{equation}
It is one for a coherent state with perfect polarization, $\langle \hat J_x \rangle = N/2$, where all atoms reside in the bonding orbital $\phi_L(x)+e^{-i\theta}\phi_R(x)$, while $\langle \hat J_y \rangle = \langle \hat J_z \rangle = 0$. In contrast, a Fock-type state, where half of the atoms reside in the left well and the other half in the right well, has no defined phase relation between the orbitals: it is completely delocalized around the equator of the Bloch sphere, and the coherence $\alpha=0$ vanishes. In general, squeezed states have $0<\alpha<1$.
From equation~\eqref{eq:densityop} we can also obtain the density fluctuations
\begin{eqnarray}\label{eq:denopfluc}
&&\mbox{\hspace*{-2.0cm}}
\Delta n(x)^2=\langle \bigl(\hat n(x)-n(x)\bigr)^2\rangle \\
&&\mbox{\hspace*{-0.7cm}}=
4n_Ln_R\left(\left<\Delta\hat J_x^2\right>\cos^2(kx+\theta)+
\left<\hat J_y^2\right>\sin^2(kx+\theta)\right)
+(n_L-n_R)^2\left<\hat J_z^2\right>\,.\nonumber
\end{eqnarray}
Noise contributions from all pseudospin operators contribute, differently weighted by the time and space dependent probability distributions $n_{L,R}(x,t)$. The $x$-contribution, proportional to $\langle\Delta\hat J_x^2\rangle$, accounts for an uncertainty in the polarization of the state on the equator (see figure~\ref{fig:blochintro}). It is zero only for a coherent state. We will discuss the physical meaning of this quantity later in context of MCTDHB, section~\ref{sec:MCTDHB}. The $y$-contribution accounts for the intrinsic phase width of the quantum state, which provides a fundamental limit for the phase measurement, and the $z$-contribution for the number fluctuations between the two wells.
\subsection{Phase sensitivity}
The ideal state for detecting small variations $\Delta\theta$ is one where $n(x)$ as a function of $\theta$ has a sufficiently large derivative, and the fluctuations $\Delta n(x)$ are sufficiently small. In order to resolve $\Delta\theta$ the inequality
\begin{equation}\label{eq:sigtonoise}
\Delta n(x)\le\left|\frac{\partial n(x)}{\partial\theta}\right|\Delta\theta
\end{equation}
has to be fulfilled. The explicit expression in terms of $n_{L,R}$ follows immediately from equations~\eqref{eq:denop} and \eqref{eq:denopfluc}. A particularly simple expression is obtained if we keep in equation~\eqref{eq:sigtonoise} only the dominant contribution from the phase noise, which is proportional to $\langle\hat J_y^2\rangle$ (other contributions depend on $n_{L,R}(x)$ and are expected to be less important), which leads to the estimate for the phase sensitivity in terms of shot noise by Kitagawa and Ueda \cite{kitagawa:93} (see also table \ref{tab:squeezing})
\begin{equation}\label{eq:phasesqueezing}
\xi_R=\frac{\sqrt{N\langle\Delta\hat J_y^2\rangle}}{\left|\langle\hat J_x\rangle\right|}\,.
\end{equation}
This expression will serve us as a guiding principle for optimizing atom interferometry. We will consider in the following the phase squeezing at the optimal working point. The normalization by $\langle \hat J_x \rangle$ takes into account that improving interferometric sensitivity requires not only to reduce noise but also to maintain a high interferometer contrast $\alpha$ (which determines the difference to $\xi_{\rm phase}$ via $\xi_R=\xi_{\rm phase}/\alpha$) . Equation~\eqref{eq:phasesqueezing} provides a simple way to estimate the sensitivity obtainable by a given initial state. For a coherent state with a binomial atom number distribution, the mean value and the variance are proportional to the total atom number $N$. Thus, $\xi_R=1$ and the phase sensitivity $\Delta\theta=1/\sqrt N$ is {shot-noise} limited. We refer to this limit as the \textit{standard quantum limit}.
\begin{table}
\begin{center}
\begin{tabular}{l|l}
Squeezing factor & Definition \\
\hline
\hline
Useful squeezing & $\xi_R=\Delta J_y/(\alpha\sqrt{N}/2)$ \\
Phase squeezing & $\xi_{\rm phase}=\Delta J_y/(\sqrt{N}/2)=\xi_R\alpha$ \\
Number squeezing & $\xi_N=\Delta J_z/(\sqrt{N}/2)$ \\
\end{tabular}
\end{center}
\caption{Definition of squeezing factors used in this work. $\xi_R$ is the useful squeezing that determines the sensitivity of an interferometer. It differs from the phase squeezing through the coherence factor $\alpha=2\langle \hat J_x \rangle/N$. $\xi_N$ is the number squeezing that depends on the number fluctuations between the two wells.
\label{tab:squeezing}}
\end{table}
For the purpose of reading the interference pattern, phase squeezed states are ideal because they allow to reduce the sensitivity below shot noise. $\xi_R$ provides a measure of \emph{useful squeezing} for metrology \cite{wineland:94}: a state with $\xi_R < 1$ allows to overcome the standard quantum limit by a factor $\xi_R$. The lower bound of the sensitivity is provided by the Heisenberg limit. From the commutation relation $[\hat J_y,\hat J_z]=i\hat J_x$ one obtains the uncertainty relation $\Delta J_y\Delta J_z\ge N/2$. Thus, for a state with a maximal number uncertainty, $\Delta J_z=N/2$, the standard deviation $\Delta J_y$ is on the order of unity, and the Heisenberg limit becomes $\xi_R=\sqrt{2/N}$.
\begin{figure}
\includegraphics[width=\columnwidth]{figure3a_3c}
\caption{Density and density noise for (a) binomial, (b) phase squeezed, and (c) number squeezed state, and for a phase $\theta=0$ (green areas) and $\theta=1/\sqrt{N}$ (red areas), corresponding to shot noise. We use $N=100$, width $\sigma=0.05$, a read-out time $T_{\rm tof}=10$ and an interwell separation of $d=5$. The degree of squeezing is characterized by a squeezing factor of $\xi_N=0.18$ ($\xi_{\phi}=0.18$). The dashed lines show the density profile of the condensates without interference. The insets magnify the region where the noise is least. The phase sensitivity is best for the phase squeezed state, although in certain regions noise is enhanced due to $\Delta J_x$. \label{fig:den}}
\end{figure}
In figure~\ref{fig:den} we show density and density fluctuations for a (a) binomial, (b) phase-squeezed, and (c) number squeezed state. One observes that the fluctuations are smallest for the phase-squeezed state, and become larger for the coherent and number squeezed state. To appreciate the sensitivity of the interferometer, we also plot the density profile for a state that has acquired a phase $\theta$ of the order of shot noise. As apparent from the insets, which magnify the regions of smallest noise, the smallest $\theta$ variations can be resolved with the phase-squeezed state.
In a typical double well interferometer experiment, where the interference is read out in TOF, the exact number of atoms $N$ is not known, and one has to obtain $N$ and the accumulated phase $\theta$ from a suitable fitting procedure. To benefit from the regions of reduced density noise, the density distribution $n(x)$ has to be weighted appropriately by $\Delta n(x)$.
\begin{figure}[b]
\includegraphics[width=\columnwidth]{figure4a_4b}
\caption{(a) Best achievable $\xi_R$ versus read-out time $T_{\rm tof}$ for atom number $N=100$, $\sigma=0.05$, and $d=5$ (as in Fig.~\ref{fig:den}). For given $T_{\rm tof}$, the optimal phase squeezing $\xi_{\rm phase}$ (dashed line) is found such that $\xi_R$ (solid black line) is minimized. $\xi_R$ approaches the Heisenberg limit (gray line) for large $T_{\rm tof}$. The difference between $\xi_R$ and $\xi_{\rm phase}$ is due to $\alpha < 1$ (b) Polarization noise, $\Delta J_x$, of the corresponding phase squeezed states of (a).
\label{fig:den2}}
\end{figure}
The dependence on the duration of the read-out stage is shown in figure~\ref{fig:den2}. For short $T_{\rm tof }$ the overlap between the condensates is small, and the contribution from $\Delta J_z$ (the third term in equation~\eqref{eq:denopfluc}) important. Since small phase fluctuations are accompanied by large number fluctuations, there exists an optimal degree of phase squeezing for the interferometer input state. For large enough $T_{\rm tof}$ this effect is less important and $\xi_R$ approaches the Heisenberg limit. We optimize the degree of phase squeezing of the initial state (i.e., at $T_{\rm tof}=0$) for given $T_{\rm tof}$, such that $\xi_R$ is minimized. This is shown in (a) versus $T_{\rm tof}$ (solid line), together with the corresponding optimal phase squeezing $\xi_{\rm phase}$ (dashed line). For small $\xi_{\rm phase}$, polarization noise $\Delta J_x$ [shown in panel (b)] is introduced. This noise reduces the spatial region where the sensitivity is high [see also figure~\ref{fig:den}(b)]. In conclusion, the expansion in TOF has to be sufficiently long, such that the atom clouds can expand sufficiently far, and the phase sensitivity is no longer limited by number fluctuations between the clouds.
\subsection{Interferometry in presence of atom-atom interactions during the phase accumulation stage \label{subsec:interactions}}
In the following we discuss how the atom-atom interactions affect the phase distribution of the split condensate during the phase accumulation stage and spoil atom interferometry, and what could be done to minimize those effects.
For the purpose of atom interferometry, it is convenient to introduce the \textit{phase eigenstates} \cite{menotti:01}
\begin{equation}
|\phi\rangle=\frac{1}{\sqrt{2\pi}}\sum_m e^{i\phi m}|m\rangle\,,
\end{equation}
where $|m\rangle$ is a state with atom number imbalance $m$ between the left and right well. The relation between $|m\rangle$ and $|\phi\rangle$ corresponds to a Fourier transformation. We can now project any state on the phase eigenstates and obtain the phase representation of the state. From this we obtain the
phase width $\Delta \phi$ (see also \cite{menotti:01}).
The noise $\Delta J_y$, which enters the phase sensitivity of the TOF-interferometer according to equation~\eqref{eq:phasesqueezing}, has a similar time evolution as the phase width . However, while the phase width is bound to values below $2\pi$, the upper bound of $\Delta J_y$ is given by the total atom number $N$, and therefore $\Delta J_y=\sqrt{N}\Delta \phi$.
Phase diffusion due to atom-atom interactions is best illustrated in the Bloch sphere representation (see figure~\ref{fig:blochtwist}): The non-linear coupling $\kappa\hat J_z^2$ from equation~\eqref{eq:hamtwomode.pseudospin} twists the state, and the larger the number fluctuations $\Delta J_z$ the faster the state winds around the Bloch sphere. In the notion of phase eigenstates, a state with a well defined phase has a broad atom-number distribution. As each atom-number eigenstate $|m\rangle$ evolves with a different frequency, the time evolution of a superposition state will suffer phase diffusion. The phase width broadens with rate \cite{javanainen:97}
\begin{equation}\label{eq:diffrate}
R= 8U_0\Delta n=8U_0\xi_N(\sqrt N/2)\,.
\end{equation}
As a result, states with small number fluctuations ($\xi_N <1$) are more stable during the phase accumulation time.
One immediately sees a conflict of requirements. For best readout of the interference pattern we want phase squeezed states, but those are very fragile and result in a fast phase diffusion and a short measurement time. On the other hand, number-squeezed states allow for longer measurement times but have a rather poor readout performance. In the remaining part of the paper we will discuss the optimal strategy for interference experiments, and how one can implement them in realistic settings.
To demonstrate the above reasoning, we show in figure~\ref{fig:demo} the achievable phase sensitivity as a function of phase accumulation time $T_{\rm phase}$ in presence of atom-atom interactions for a binomial (solid), a number squeezed (dashed line), and a phase squeezed (dashed-dotted line) state with squeezing factors $\xi_N\approx 0.22$ and $\xi_{\phi}\approx 0.22$, respectively. The phase squeezed state has initially sub-shot noise phase sensitivity. For longer hold times, the number squeezed state outperforms the phase squeezed and binomial ones due to its smaller phase diffusion rate.
\begin{figure}
\center\includegraphics[width=0.7\columnwidth]{figure5}
\caption{Example for the phase sensitivity of a binomial (solid line), a number squeezed (dashed line), and a phase squeezed (dashed-dotted line) state with squeezing factors $\xi_N\approx 0.22$ and $\xi_{\phi}\approx 0.22$, respectively. Parameters are $N=100$ and $U_0 N=1$. The binomial and phase squeezed states have a better initial phase sensitivity, whereas the number squeezed is much more stable against phase diffusion, and has a much better sensitivity at later times. The gray line corresponds to shot noise ($\xi_R=1$).
\label{fig:demo}}
\end{figure}
\section{Optimizing atom interferometry\label{sec:opt}}
When designing a trapped atom interferometer for measurements, one has to consider the conflicting requirements from phase diffusion and readout, the one asking for number squeezing, the other for phase squeezing. In the following we will outline a few strategies of how to optimize interferometer performance for realistic double well settings.
\subsection{A very simple estimate}\label{sec:opt.simple}
We first analyze the impacts of trap geometry and initial state preparation. We start by employing a simple model to illustrate the effects of the non-linearity in the evolution of the split trapped BEC. Let us first look at the interaction energy of a trapped cloud, and how it changes with the number of trapped particles. Without loosing generality, we discuss in the context of a harmonic trap characterized by the mean confinement $\omega_0$ and the length scale $a_0$, which are defined through
\begin{equation}
\omega_0 = \sqrt[3]{\omega_x \omega_y \omega_z}\,, \qquad
a_0=\sqrt{\frac{\hbar}{m \omega_0}} \, .
\end{equation}
With $N$ atoms in the trap the chemical potential $\mu$ in Thomas-Fermi approximation \cite{dalfovo:99} is given by
\begin{equation}
\mu = \frac{\hbar \omega_0}{2} \left( \frac{15 \, N \, a_{s}}{a_0} \right)^{\frac{2}{5}}\,,
\end{equation}
where $a_s$ is the $s$-wave scattering length.
Adding a single atom to the trap, changes $\mu$ by
\begin{equation}
\frac{\partial \mu}{\partial N} = \frac{2}{5} \frac{\mu}{N}\,.
\end{equation}
This quantity corresponds to the effective 1D interaction parameter $U_0$ discussed earlier. We can now estimate the scaling of the phase diffusion rate $R$ caused by a number distribution with fluctuations $\Delta N = \xi_N \sqrt{N}/2$ after splitting ($\xi_N=1$ corresponds to a binomial number distribution, whereas $\xi_N < 1$ to number squeezing),
\begin{equation}
R \propto \xi_N \, N^{-\frac{1}{10}} \,
\omega_0 ^{\frac{6}{5}} \, a_{s} ^{\frac{2}{5}} \, m^{\frac{1}{5}} \,.
\end{equation}
For $^{87}$Rb atoms ($a_{s}=5.2$ nm, $m=87$) one obtains $R \approx 0.022 \, \xi_N \, N^{-\frac{1}{10}} \, \omega_0 ^{\frac{6}{5}}$ s$^{-1}$, or in scaled units $R \approx 0.29 \, \xi_N \, N^{-\frac{1}{10}} \, \omega_0 ^{\frac{6}{5}}$.
If we now set the phase diffusion rate $R$ equal to the phase accumulation rate $\frac{1}{\hbar}\Delta E$ (the signal we want to measure), we find the limit $\frac{1}{\hbar} \Delta E^{\rm min} = R$ for the sensitivity of a single-shot interferometer measurement, even for perfect readout. It is interesting to note that this sensitivity limit is only very weakly dependent on the atom number $N$.
If the interferometer measurement is not limited by readout, we can identify the following strategies for improving the interferometer performance (or, equivalently, reducing the effect of phase diffusion).
\begin{description}
\item[Minimize the scattering length.] The best is to set $a_{s}=0$, which can in principle be achieved by employing Feshbach resonances \cite{chin:10}. Drastic reduction of phase diffusion when bringing the scattering length close to zero was recently demonstrated in two experiments \cite{gustavsson:08,Fattori:08}. A disadvantage thereby is that using Feshbach resonances requires specific atoms and specific atomic states. These states need to be tunable, and are therefore \textit{not} the '\textit{clock}' states usually used in precision experiments which are immune to external disturbances like magnetic fields.
\item[Choose a trap with weak confinement.] This route seems problematic, since the timescale in splitting and manipulating the trapped atoms scales with the trap confinement. Optimal control techniques, like discussed in Refs.~\cite{hohenester.pra:07,grond.pra:09,grond.pra:09b}, will be needed to allow splitting much faster than the phase diffusion time scale. It is interesting to note that one needs a strong confinement only in the splitting direction. In the other two space directions the confinement can be considerably weaker. This suggests to work with strongly anisotropic traps. \\
For an \textit{elongated cigar-shaped} trap (1D geometry) with confinement ratio $C_{1D}=\frac{\omega_z}{\omega_\perp}$ (strong confinement $\omega_\perp$ in radial directions, weak confinement $\omega_z$ in axial direction) one finds
\begin{equation}
\omega_0 ^{(1D)}=\omega_\perp \sqrt[3]{C_{1D}} \, .
\end{equation}
For a \textit{flat pan cake-shaped} trap (2D geometry) with confinement ratio $C_{2D}=\frac{\omega_{\rm plane}}{\omega_\perp}$ (strong transverse confinement $\omega_\perp$ and weak in-plane confinement $\omega_{\rm plane}$) one finds
\begin{equation}
\omega_0 ^{(2D)}=\omega_\perp \sqrt[3]{C_{2D} ^2} \,.
\end{equation}
With a confinement ratio $C \sim 1/1000$, which is easily obtainable in experiments the phase diffusion is reduced by a factor 10 in a 1D geometry and a factor 100 in a 2D geometry.
\item[Increase number squeezing in the splitting process.] This directly reduces the phase diffusion rate and hence leads to a better limit for the minimal detectable signal. Number squeezing can be achieved during the splitting process. It is mediated by the atom interactions and one has to achieve a careful balance between the interactions necessary to obtain sizable number squeezing and the decremental effect of the interactions during the phase accumulation time. This will be one of the central parts in our optimization discussed below.
\end{description}
In an ideal interferometer one would like to use clock states, create strong squeezing during the splitting process, exploiting the non-linearity in the time evolution, and then turn off the interactions (by setting the scattering length to $a_s = 0$) after splitting. All together might, however, be difficult or even impossible to achieve. In the remainder of the manuscript we will discuss the different contributions to the precision of an atom interferometer, and investigate how the performance can be optimized.
\subsection{Optimization of the many-boson states\label{subsec:results}}
\begin{figure}
\center\includegraphics[width=0.7\columnwidth]{figure6a_6c}
\caption{(a) Optimal phase sensitivity versus phase accumulation time for $N=100$ (solid lines), $N=500$ (dashed lines), $N=2000$ (dashed-dotted lines), and $N=8000$ (dots). Interaction strength is such that $U_0 N=1$. The black lines show results for number-squeezed states, and the gray lines result for slightly tilted number-squeezed states. (b) The corresponding optimal number squeezing. (c) The total tilt angle $\phi_{\rm tilt}=\Omega_{\rm tilt}T_{\rm pulse}$ is determined by the pulse duration (here $T_{\rm pulse}=2$) and strength $\Omega_{\rm tilt}$ (very similar for all $N$). The diamond symbols in (a) and (b) show estimates from a simple model.
\label{fig:figxi}}
\end{figure}
\subsubsection{Optimized Number Squeezing \label{subsubsec:optsplit}}
First we shortly discuss how number-squeezed states can be prepared during the splitting stage. Number squeezed states are created in a double well potential when the interaction energy starts to dominate over the tunnel coupling, the latter being controlled by the barrier height and the double well separation. A natural way to achieve high number squeezing is dynamic splitting of a BEC \cite{menotti:01,streltsov:07}, such that the wavefunction can adiabatically follow the ground state \cite{javanainen:99}. However, this may take very long, possibly longer than the phase diffusion time of the split condensate. In our earlier work \cite{grond.pra:09,grond.pra:09b,hohenester.fdp:09} we employed optimal control theory (OCT) to find splitting protocols which allow for high number squeezing on a fast timescale, at least one order of magnitude shorter than for the quasi-adiabatic splitting. These protocols can be viewed as the continuous transformation of a (close to) harmonic potential into a double well. Thereby, a barrier is ramped up at the center, and simultaneously the two emerging wells are separated. In many cases this splitting process can be parameterized by a single parameter, whose time variation is obtained within optimal control theory such that sizeable squeezing is created, condensate oscillations are prevented after splitting, and phase coherence is better preserved at the end of the splitting process \cite{grond.pra:09,grond.pra:09b,hohenester.fdp:09}.
Unless stated otherwise, in the following discussion of the dynamics during the phase accumulation stage we use number-squeezed states as initial states which are obtained as the ground states of equation~\eqref{eq:hamtwomode.pseudospin} for finite values of tunneling $\Omega$. They are very similar to those obtained by OCT the splitting. Similar initial states obtained by exponential splitting have a smaller degree of coherence. During the phase accumulation stage, we set $\Omega=0$.
How much number squeezing is ideal for a pre-determined phase accumulation time of an interferometer? In figure~\ref{fig:figxi} we show results where we optimize the degree of number squeezing for the interferometer input states at $T_{\rm phase}=0$, in order to achieve the best phase sensitivity at a given time $T_{\rm phase}$. The top panel reports the best achievable phase sensitivity for an initially number squeezed state (black lines), and the middle panel reports the corresponding number squeezing. For short phase accumulation times, less initial number squeezing is better. With increasing accumulation time more number squeezing becomes favorable. This is due to the competition between phase fluctuations $\Delta J_y$, which increase with number squeezing, and the decrease of phase diffusion for states with high number squeezing.
The results can be well explained by a simple model. Neglecting the effects of reduced phase coherence, we have initially $\xi_R=\xi_{\rm phase}^0$, the initial phase squeezing. Phase diffusion with rate $R$ then results in $\xi_R=\sqrt{(\xi_{\rm phase}^0)^2+R^2 T_{\rm phase}^2}$. We next use that $\xi_N \xi_{\rm phase}^0\approx 1$, which is in the spirit of the Heisenberg uncertainty principle and agrees well with our OCT results. Putting in all the constants, we have
\begin{equation}
\xi_R= \sqrt{\frac{1}{\xi_N^2 N}+16 N\xi_N^2U_0^2 T_{\rm phase}^2}.
\end{equation}
The minimum with respect to $\xi_N$ is found as $\xi_N^{\rm min}=1/\bigl(2\sqrt{U_0 N T_{\rm phase}}\bigr)$, which yields a best phase sensitivity $\xi_R^{\rm min}=2\sqrt{2U_0 N T_{\rm phase}}$. For a given final $\xi_R$ we see that $T_{\rm phase}$ is indirectly proportional to the interaction parameter $U_0$, which allows rescaling of $T_{\rm phase}$ in case of a different $U_0$.
Predictions of the simple model are shown in figure~\ref{fig:figxi} by the diamond symbols. The agreement with the exact results is very good in (a), and gives the right scaling in (b). Indeed, $\xi_R^{\rm min}$ in figure~\ref{fig:figxi} is independent of $N$ for long times, as long as $U_0 N$ is constant. This is not true for $\xi_N^{\rm min}$. The reason is that neglecting phase coherence makes the minima with respect to $\xi_N$ much more shallow. For longer times, however, phase coherence becomes more important and the present approximations are no longer valid for small N.
\subsubsection{Optimized specialized initial states \label{subsubsec:spec}}
A different strategy for improving the interferometry performance is to prepare the system at the beginning of the phase accumulation stage (i.e., at $T_{\rm phase}=0$) in a special state, which evolves under the influence of the non-linear interaction after some pre-determined time into a state with high intrinsic sensitivity. The ideal initial state would be the time reversal of a phase squeezed state. We denote this strategy as \textit{refocusing}. When the condensate is released at the optimal time, and expands in absence of interactions to form the interference pattern, interferometry can be performed with a sensitivity determined by the properties of the refocused state. This can be achieved because the phase accumulation (rotation around $z$-axis on the Bloch sphere) and the non-linear coupling do not interfere.
Such a refocusing strategy is related to spin echo techniques, which where investigated by turning the scattering length $a_s$ from repulsive to attractive \cite{widera:08}. However, the latter has given a rather poor improvement, because it does not lead to a perfect time reversal of the many-body dynamics \cite{sakmann2:09}. Artificial preparation of the desired time reversed states, seems to be very difficult. We did not succeed in this task using optimal control techniques.
\begin{figure}
\begin{tabular}{ll}
\includegraphics[height=0.37\columnwidth]{figure7a}&\includegraphics[height=0.37\columnwidth]{figure7b}
\end{tabular}
\caption{ Refocusing control sequence for $N=100$, $U_0 N=1$, $T_{\rm pulse}=2$, and (a) $T_{\rm phase}=1$ as well as (b) $T_{\rm phase}=4$. The lower panel shows $\xi_R$, while the upper panel the coherence factor $\alpha$. In the first stage, the tunnel pulse is applied and \emph{tilting} is achieved. In the second stage, the state refocuses to a state with a good phase sensitivity, in (a) below shot-noise. The Bloch spheres visualize the time evolution.
\label{fig:tiltdetails}}
\end{figure}
One state that leads to very good refocusing can be prepared by \textit{tilting} the initial number squeezed state on the Bloch sphere slightly \textit{against} the direction of the twist originating from $\hat J_z^2$. The tilt can be achieved by applying a short tunnel pulse within a time interval $T_{\rm pulse}$ that rotates the number squeezed state. In real space, this operation corresponds to lowering the barrier for a suitable amount of time, which cannot be done arbitrarily fast because of condensate oscillations. Appropriate controls of the barrier will be discussed in detail in context of a realistic modeling in section~\ref{sec:MCTDHB}. Within the generic model we consider for simplicity square-$\Omega$ pulses, which is the best possible pulse in presence of interactions \cite{pezze2:06}. Examples are shown in figure~\ref{fig:tiltdetails} for different $T_{\rm phase}$. During the \emph{tilting} pulse sequence and the phase accumulation stage, 'rephasing' happens and the phase fluctuations decrease. Simultaneously, the phase coherence is restored to a value close to one. This significantly improves the phase sensitivity, for short times one can even reach below shot noise. However, this cannot be done perfectly, the degree of phase squeezing $\xi_{\rm phase}$ achieved after refocusing is always less than the degree of number squeezing $\xi_N$ of the original state.
We next optimize systematically both parameters of initial number squeezing and tilt angle. The lowest panel of figure~\ref{fig:figxi} shows the optimal tilt angles. From the upper panels (bright lines) we find a clear improvement of phase sensitivity for a given phase accumulation time. The dependence on $N$ is very distinct now for small atom numbers, and saturates for large $N$. We find that for small $N$ the improvement is roughly a factor of three in time, and for large $N$ approximately an order of magnitude.
\subsection{Optimize trapping potential and atom number. esults of generic two-mode model\label{subsec:opt_trap}}
We next proceed to a more detailed analysis of the ideal trap parameters. Considering our previous discussion of section~\ref{sec:opt.simple}, we expect an improvement of the interferometer performance when increasing the anisotropy of the trap. In the following calculations we choose a fixed $N=100$ (the results are not expected to depend decisively on $N$).
\begin{figure}
\begin{tabular}{l l}
\includegraphics[width=0.44\columnwidth]{figure8a}&\includegraphics[width=0.44\columnwidth]{figure8b}
\end{tabular}
\caption{ Best achievable phase sensitivity versus interaction strength for various pulse durations for (a) $T_{\rm phase}=1$ and (b) $T_{\rm phase}=20$. Atom number is $N=100$. We optimize for the initial number squeezing and the tunnel pulse.
\label{fig:Tpulse}}
\end{figure}
We first investigate the role of the interaction $U_0 N$, and the pulse duration $T_{\rm pulse}$ for the refocusing strategy. In figure~\ref{fig:Tpulse} we plot the best phase sensitivity using refocusing versus interaction strength for (a) $T_{\rm phase}=1$ and (b) $T_{\rm phase}=20$. Sub-shot noise phase sensitivity is clearly achievable for short $T_{\rm pulse}$, or for $U_0 N \ll 1$. Short pulses are favorable and give better phase sensitivity. This is in particular important for short phase accumulation times $T_{\rm phase}$ (see figure~\ref{fig:figxi}). Optimizing the pulse form and duration within a realistic modeling of \emph{tilting} on the Bloch sphere will be discussed in detail in section~\ref{sec:MCTDHB}.
Quite generally, we can expect that for a reduced interaction parameter it is more difficult to obtain high number squeezing in the splitting process. To estimate the time scale for achieving a certain degree of number squeezing, we consider splitting protocols derived in a previous work within the framework of optimal control theory \cite{grond.pra:09b}, and discussed already in section~\ref{subsubsec:optsplit}. They directly provide us optimized number squeezing for a given splitting time $T_{\rm split}$. For a given phase accumulation time $T_{\rm phase}$, we optimize then the tunnel pulse which tilts the number squeezed state, similar to the analysis of section~\ref{subsubsec:spec}.
\begin{figure}
\begin{tabular}{l l}
\includegraphics[width=0.49\columnwidth]{figure9a}& \includegraphics[width=0.49\columnwidth]{figure9b}\\
(a)&(b)
\end{tabular}
\caption{(a,b) Phase sensitivity for a sequence including splitting and phase accumulation, versus interaction strength $U_0 N$ and $T_{\rm split}$. Parameters are $N=100$, (a) $T_{\rm phase}=1$, (b) $T_{\rm phase}=20$, and $T_{\rm pulse}=2$. For a given $T_{\rm split}$ we take the best number squeezing achieved by OCT. Number squeezing is higher for larger $T_{\rm split}$ and $U_0$ values. We also optimize for a tunnel pulse which rotates the number squeezed state. The minimum phase sensitivity decreases very slowly with $T_{\rm split}$. The contour lines are at (a) $\xi_R=0.35$ and (b) $\xi_R=0.6$.
\label{fig:kappaturnoff}}
\end{figure}
The best possible phase sensitivity for various $T_{\rm split}$ and $U_0 N$ values is shown in figure~\ref{fig:kappaturnoff} for (a) $T_{\rm phase}=1$ and (b) $T_{\rm phase}=20$. For both cases sensitivity distinctly sub-shot noise sensitivity can be achieved. To achieve the squeezing needed to boost interferometer performance, a finite $U_0$ is needed, and for a given $T_{\rm split}$ there exists an optimal value of $U_0$. This value decreases for longer splitting times.
In order to analyze the dependence on $N$, we consider a realistic 3D cigar-shaped trap with transverse trapping frequency $\omega_{\perp}=2\pi\times 2$ kHz as typically realized in atom chip interference experiments. The effective 1D interaction strength in the splitting direction $U_0$ is then approximately proportional to $C_{1D}^{2/5}$. A more rigorous estimate, that is used in the calculations, is given in Ref.~\cite{grond.pra:09b}. For $N=100$, $U_0 N\sim1(0.1)$ corresponds to an aspect ratio of $C_{1D}\sim1/100(1/1000)$. These values change for $N=1000$ to $C_{1D}\sim1/1000(1/10000)$. For the pan cake-shaped trap we have $C_{2D}\sim \sqrt{C_{1D}}$, and thus $U_0 N=0.01$ is within reach for $C_{2D}\sim 1/1000$ and $N=100$, $C_{2D}\sim 1/10000$ and $N=1000$.
Let us first consider the case without refocusing. We find approximately $\xi_R \sim 2\sqrt{2U_0 N T_{\rm phase}}$, and, considering the dependence of $U_0$ on the trapping potential, we obtain
\begin{equation}\label{eq:estxiN}
\xi_R \sim a_s^{1/5} \omega_{\perp}^{3/5} C_{1D}^{1/5}N^{1/5}T_{\rm phase}^{1/2}\;.
\end{equation}
In order to reach sub-shot noise phase sensitivity we need the confinement ratio $C_{1D}$, atom number $N$, and $T_{\rm phase}$ all to be small. In figure~\ref{fig:Vmin} (a) $\xi_R$ is plotted for $C_{1D}=1/100,1/300$ and $1/1000$ (solid, dashed, and dashed-dotted line, respectively).
As we have seen in section~\ref{subsec:results}, equation~\eqref{eq:estxiN} is valid only if the coherence is well preserved. We estimate breakdown of this approximation when $\alpha\approx1-\frac{\xi_N^2}{2N}\approx 0.6$. From this we can obtain the time after which $\xi_R$ is expected to grow rapidly because the coherence factor tends to zero. It is given as $T_{\rm coh}=\frac{N^{3/5}}{2\cdot15^{2/5} a_s^{2/5} \omega_{\perp}^{6/5}C_{1D}^{6/15}}$, and shown in figure~\ref{fig:Vmin} (f) for different $C_{1D}$.
We now turn to refocusing. In the 1D elongated trapping geometry we find a more moderate increase of $\xi_R$ with N compared to the case without refocusing, see figures~\ref{fig:Vmin} (a) and (b). This illustrates that the refocusing works better for large N, as long as $U_0 N$ is constant, see also figure~\ref{fig:figxi}.
\begin{figure}
\begin{tabular}{l l}
\includegraphics[width=0.67\columnwidth]{figure10a_10d}& \includegraphics[width=0.33\columnwidth]{figure10e_10f}
\end{tabular}
\caption{Optimization for a realistic cigar shaped trapping potential with transverse frequency $\omega_{\perp}=2\pi\cdot 2$ kHz ($\omega_{\perp}\approx 17 $ in scaled units) and aspect ratio $C_{1D}=1/100$ (solid lines), $C_{1D}=1/300$ (dashed lines), and $C_{1D}=1/1000$ (dashed-dotted lines). (a), (b) $\xi_R$ versus $N$ for (a) $T_{\rm phase}=1$ and (b) $T_{\rm phase}=20$. (c), (d) Minimal detectable potential $\Delta E^{\rm min}$ versus $N$ for (c) $T_{\rm phase}=1$ and (d) $T_{\rm phase}=20$. (e) $\Delta E^{\rm min}$ versus time for $N=1000$ (the vertical dashed lines in (a)-(d) denote cuts at $N=1000$). The black lines show results for the case when only the initial number squeezing is optimized. The coloured lines show results with refocusing, i.e., where both initial number squeezing and tunnel pulse are optimized. (f) Time $T_{\rm coh}$ after which $\xi_R$ is expected to grow rapidly due to loss of coherence, if refocusing is not applied.
\label{fig:Vmin}}
\end{figure}
The accumulated phase for a potential $\Delta E$ is given as $\theta=\Delta E T_{\rm phase}$, and the smallest detectable potential difference in time $T_{\rm phase}$ becomes $\Delta E^{\rm min}={\xi_R}/({\sqrt{N} T_{\rm phase}})$. Without refocusing, we find an improvement with N and $T_{\rm phase}$, $\Delta E^{\rm min}\sim N^ {-1/3}T_{\rm phase}^{-1/2}$, figure~\ref{fig:Vmin} (c), (d) and (e). Similar scalings also holds for the case with refocusing. We expect that for 2D traps, $\Delta E^{\rm min}$ somewhat below $10^{-3}$ is within reach.
This confirms, in agreement with the scaling analysis of section~\ref{sec:opt.simple}, that a stronger trap anisotropy appreciably helps to reduce phase diffusion and, in turn, to improve the phase sensitivity of the interferometer. Possible limitations of strongly anisotropic systems are discussed in Sec.~\ref{sec:temp}.
The atom number $N$ helps to improve absolute sensitivity, but makes it more difficult to demonstrate measurements with a sensitivity below shot noise.
\section{Interferometer performance within MCTDHB\label{sec:MCTDHB}}
Until now we have used a generic two-mode model to describe the interferometer, which captures the basic processes and physics, but ignores many details of the condensate dynamics in realistic microtraps. More specifically, the modeling of the splitting process and of the rotation pulses requires in many cases a more complete dynamical description in terms of the \emph{multi-configurational time dependent Hartree for Bosons} method (MCTDHB) \cite{alon:08}. In this section we discuss first the MCTDHB details relevant for our analysis, and its relation to the generic model. The main part will be concerned with the simulation and optimization of tunnel pulses for achieving tilted squeezed states, as discussed in section~\ref{subsubsec:spec} in context of refocusing. An exhaustive discussion of MCTDHB, as well as optimal condensate splitting can be found elsewhere \cite{alon:08,grond.pra:09b}.
In the two mode Hamiltonian of equation~\eqref{eq:hamtwomode} we did not explicitly consider the shape of the two {\textit{orbitals} $\phi_L$ and $\phi_R$, but lumped them into the effective parameters $\Omega$ and $\kappa$. The dynamics is then completely governed by the wavefunction accounting for the atom number dynamics. Within MCTDHB, both the orbitals and the number distribution are determined self-consistently from a set of coupled equations, which are obtained from a variational principle. This leads us to a more complete description, accounting for the full condensates' motion in the trap. The state of the system is then given by a superposition of symmetrized states (permanents), which comprise the time dependent orbitals. Instead of left and right orbitals, such as used in the two-mode model, we employ for the symmetric confinement potential of our present concern orbitals with \emph{gerade} and \emph{ungerade} symmetry. The time dependent orbitals then obey non-linear equations, which depend on the one- and two-particle reduced densities \cite{sakmann:08} describing the mean value and variances of the number distributions \cite{streltsov:06}.
The atom number part of the wavefunction obeys a Schr\"odinger equation with the Hamiltonian
\begin{equation}\label{eq:TMMCham}
\mathcal{H}=\Omega\hat{J}_x+\frac{1}{2}\sum_{k,q,l,m}\hat a_k^{\dagger}\hat a_q^{\dagger}\hat a_l\hat a_m W_{kqlm}\,,
\end{equation}
where the indices are either $g$ (\emph{gerade}) or $e$ (\emph{ungerade} or excited). We observe that, in contrast to the two-mode Hamiltonian of equation~\eqref{eq:hamtwomode}, the atom-atom interaction elements
\begin{equation}\label{eq:mvs}
W_{kqlm}=U_0\int dx \phi_k^*(x,t)\phi_q^*(x,t)\phi_l(x,t)\phi_m(x,t)\,,
\end{equation}
as well as the tunnel coupling $\Omega=\int dx \phi_e^*(x)\hat{h}\phi_e(x)-\int dx \phi_g^*(x)\hat{h}\phi_g(x)$ are governed by the orbitals. The only input parameter of the MCTDHB approach is the trapping potential $V_{\lambda}(x)$, which enters in the single-particle Hamiltonian $\hat h(x)=-(\nabla^2/2)+V_{\lambda}(x)$. We note that $\Omega$ obtained within MCTDHB cannot be directly interpreted as tunnel rate, but has to be renormalized if the two-body matrix elements differ from each other \cite{javanainen:99,ananikian:06}. Thus, there is in general no direct correspondence between the two-mode model and the MCTDHB approach. MCTDHB, which relies on time-dependent orbitals, captures a large class of excitations not included in a two-mode model. If the calculations converge when using more modes, MCTDHB reproduces the exact quantum dynamics, as discussed in \cite{sakmann:09,streltsov:09,grondscrinzi:10}.
In our MCTDHB calculations we consider a cigar-shaped magnetic confinement potential prototypical for atom chips \cite{folman:02}. Splitting is assumed to be along a transverse direction and is accomplished using rf dressing \cite{lesanovsky:06,lesanovsky:06b,hofferberth:06,hofferberth:07b}. For illustration purpose, the trapping potential in splitting direction can to very high accuracy be described by a quartic potential of the form:
\begin{equation}
V(x,t) = a(t) x^2 + b(t) x^4\;,
\end{equation}
where for most cases $b(t)$ varies very slowly and can be assumed as constant. This potential grasps the essential features of the initial and final potential and the time evolution $a(t)$ describes how the potential is split and the barrier is ramped up. $a(t)$ large and positive characterize the initial single well, $a(t)$ large and negative the split double well, the constant $b(t)$ the confinement during the splitting. In the calculations we use the exact form of the potential used in atom chip double well experiments \cite{lesanovsky:06}. To describe the transformation of the potential we introduce a control parameter $\lambda(t)$ connected to the amplitude (phase) of the RF field. Thereby, values of $-2/3<\lambda<0$ translate to a single well, and $\lambda\sim 1$ to a double well.
Within MCTDHB the pseudospin operator $\hat J_x$ has to be rewritten in terms of the \emph{gerade} and \emph{ungerade} orbitals. In the new basis it measures the atom number difference with respect to the two states, $\hat{J}_x=\frac 1 2 (\hat a_g^{\dagger}\hat a_g -\hat a_e^{\dagger}\hat a_e)$, as discussed in more detail in the appendix. It is important to note that the \emph{gerade} and \emph{ungerade} orbitals are \emph{natural} orbitals, i.e., they diagonalize the one-body reduced density of the system \cite{sakmann:08}. Therefore, if both of them are macroscopically populated one obtains a \emph{fragmented} condensate \cite{leggett:01}. We can thus interpret the coherence factor $\alpha$ as the degree of fragmentation. The system has maximal coherence, if its state is \emph{not} fragmented but forms a single condensate. Coherence is lost if the condensate fragments into two independent condensates. In between, we have a finite, but reduced coherence. Similarly, we can interpret $\Delta J_x$ as the number uncertainty between the fragmented parts.
Optimal control theory (OCT) \cite{peirce:88,hohenester.pra:07}, is a very powerful tool to find a path which optimizes for a certain control target. In our earlier work \cite{grond.pra:09,grond.pra:09b} we implemented and optimized condensate splitting within MCTDHB \cite{peirce:88,hohenester.pra:07}. We found that, although the generic two-mode model describes qualitatively the splitting dynamics, the more complete MCTDHB description is needed for a realistic modeling. This is because one needs to control condensate oscillations during splitting, and to ensure a proper decoupling of the condensates at the end of the control sequence. In this section we employ OCT to find the appropriate paths in varying the trapping potential to achieve the desired \emph{tilting} on the Bloch sphere in short time $T_{\rm pulse}$ and prevent excitation of condensate oscillations.
\subsection{Parameter correspondence between the models \label{subsec:Corr}}
\begin{figure}
\center\includegraphics[width=0.75\columnwidth]{figure11}
\caption{ Two-body matrix elements for the ground states of a magnetic trap \cite{lesanovsky:06} versus splitting distance for $N=100$ and $U_0 N=1$. The splitting is parametrized by the parameter $\lambda$. $W_{gege}$ and $W_{ggee}$ coincide for the ground states, however not in general, see the examples of figures~\ref{fig:demoOCT1} and \ref{fig:demoOCT2}. \label{fig:mv}}
\end{figure}
For the calculation of time-dependent condensate dynamics including oscillations, a self-consistent approach like MCTDHB is mandatory. However, we expect the generic model (equation~\eqref{eq:hamtwomode}) to properly describe the phase accumulation stage ($\Omega=0$), provided the condensates are at rest. The two-body overlap integrals of the orbitals from MCHB (time-independent version of MCTDHB \cite{streltsov:06}), given in equations~\eqref{eq:mvs}, are then constant and coincide. This is because $\phi_g$ and $\phi_e$ have degenerate moduli for split condensates. When comparing equations~\eqref{eq:hamtwomode} and \eqref{eq:TMMCham}, we find the value of $\kappa$ to be used in the generic model. Similarly, in the context of optimized splitting protocols in our earlier work \cite{grond.pra:09,grond.pra:09b}, we have found that both optimizing $\Omega(t)$ in the generic model and optimizing $\lambda(t)$ within MCTDHB yield the same amount of number squeezing for a given time interval.
In figure~\ref{fig:mv} we show how the two-body overlap integrals from MCHB vary with the control parameter that determines the shape of the confinement potential. During the transition from a single-well to a double well, they drop by roughly a factor of two. This is because in the final state the \emph{gerade} and \emph{ungerade} orbitals are delocalized over both wells. After reaching a minimum around $\lambda\sim0.7$, the overlap integrals start to increase again slightly. In context of splitting we found that it is reasonable to assume $\kappa=U_0/2$ throughout the splitting process \cite{grond.pra:09b}, i.e., to take the value in the most relevant regime during condensate breakup ($\lambda\sim0.7-0.8$). Atom interferometry has to be performed with split condensates. This requires a splitting distance of some micrometers, corresponding to $\lambda \gtrsim 1$. With the corresponding value of $\kappa\approx 0.65 U_0$, phase diffusion in the phase accumulation stage can be well described using the generic model. This has in particular the advantage that we can translate our findings for the optimal states for atom interferometry from section \ref{subsubsec:spec} to MCTDHB calculations of tunnel pulses.
\subsection{Pulse optimization}
A key ingredient in interferometer performance is the preparation of an optimized initial state, as discussed in section~\ref{subsubsec:spec}. This can be achieved by a 'tunnel pulses' facilitating the \emph{tilting} of the initial number squeezed state. This operation has been analyzed by Pezz\'e \emph{et al.} in the context of a cold atom beam splitter. They used the generic model \cite{pezze2:06} and studied to which extent the creation of phase squeezed states from number-squeezed states is spoiled by atom-atom interactions. The real space dynamics has been neglected.
In real space, the confining potential has to be modified to bring the condensates together for the tunnel pulse which accomplishes the desired 'tilt' on the Bloch sphere. As has been discussed in section~\ref{subsec:opt_trap}, the duration of the tunnel pulse ($T_{\rm pulse}$) is very critical for the interferometer performance, the pulse should be as short as possible. This has to be done without significantly disturbing the many-body wave function. To design appropriate control schemes with the shortest possible time duration we will have to take the real space dynamics of the BEC into account using MCTDHB.
To find a control sequence for the preparation of the optimal initial states at $T_{\rm phase}=0$ for a given interferometer sequence within MCTDHB, we first start, following section~\ref{subsec:Corr}, with the optimized initial number squeezing and the tilt angle required for rephasing at a given $T_{\rm phase}$ as calculated in section~\ref{subsubsec:spec} within the generic model. We then choose the top of a Gaussian shaped initial guess
\begin{equation}\label{eq:pulseguess}
\lambda(t)=\lambda_0+(\lambda_{\rm end}-\lambda_0)\cdot t/T_{\rm pulse}-A\Bigl[e^{-\frac{(t-T_{\rm pulse}/2)^2}{2 B^2}}-e^{-\frac{(T_{\rm pulse}/2)^2}{2 B^2}}\Bigr]\,.
\end{equation}
We start with a state of the double well potential $V_{\lambda_0}$ with the required initial number squeezing. Then, as $\lambda$ decreases, the barrier is ramped down, the condensates approach each other. The desired $\Delta J_y^d$ to be reached at $T_{\rm pulse}$ is fixed by the required tilt angle. It can be tuned by the parameters $A$ and $B$, corresponding to the depth and the width of the control parameter deformation, respectively. Finally, a double well $V_{\lambda_{end}}$ is re-established, which completely suppresses tunneling of the two final condensates, at least if they are in the ground state.
Results of our MCTDHB calculations are shown in figures~\ref{fig:demoOCT1} and \ref{fig:demoOCT2} for interactions $U_0 N=0.1$ and $U_0 N=1$, respectively. The pulse achieves the desired $\xi_R$ [dashed lines in (d)], as we expected from the generic model (dashed-dotted lines). However, it not only affects the atom number distribution, but also leads to an oscillation of the condensates in the microtrap. This can be seen in the density, which is depicted in figures~\ref{fig:demoOCT1}(b) and \ref{fig:demoOCT2}(b). Condensate oscillations during the phase accumulation and release stage are expected to substantially degrade the interferometer performance, and may even lead to unwanted condensate excitations \cite{grondscrinzi:10}. These oscillations can be avoided by using more refined \emph{tilting} pulses, which can be obtained within the framework of OCT, where we now optimize for phase squeezing and a desired $\Delta J_y^d$ at the final time of the control interval $T_{\rm pulse}$, corresponding to $T_{\rm phase}=0$. Some details of our approach are given in the appendix.
\begin{figure}
\center\includegraphics[width=0.9\columnwidth]{figure12a_12f}
\caption{(a) Typical OCT-control (green solid line) for weak interactions $U_0 N=0.1$, atom number $N=100$, $T_{\rm pulse}=2$ and $T_{\rm phase}=1$, compared to a simpler control (blue dashed line). The dashed vertical line separates the control sequence from the phase accumulation time thereafter. The corresponding condensates density are shown (b) for the initial guess and (c) for the OCT-solution, and we compare (c) useful squeezing $\xi_R$ and (d) tunnel coupling $\Omega$. Additionally we compare to results from the generic two-mode model, where a square-$\Omega$ pulse is used (red dashed-dotted lines). (f) The two-body matrix elements of the OCT-solution are shown, similar as in figure~\ref{fig:mv}. \label{fig:demoOCT1}}
\end{figure}
For weak interactions a short tunnel pulse $T_{\rm pulse}=2$ can be easily found which properly tilts the atom number distribution and brings the condensates to a stationary state at the end of the process. A typical control sequence for $U_0 N=0.1$ and optimized for a short $T_{\rm phase}=1$ is shown in figure~\ref{fig:demoOCT1}(a) [green solid line], the corresponding density given in (c). In panel (d) we depict the phase sensitivity $\xi_R$ (solid line), which compares in the phase accumulation stage very well with the desired behavior given by the generic model (dashed-dotted line).
The gain of our OCT solution with respect to an `adiabatic' control of Gaussian type [equation~\eqref{eq:pulseguess}], where $\lambda$ is modified sufficiently slowly in order to suppress condensate oscillations, depends on the chosen value of $T_{\rm phase}$. Smaller $T_{\rm phase}$ values require larger tunnel pulses [see figure~\ref{fig:figxi}(c)]. In our example, the OCT pulses can be at least one order of magnitude faster, which means an improvement of up to $30 \%$ in $\xi_R$ [compare also with figure~\ref{fig:Tpulse}]. In figure~\ref{fig:demoOCT1}, panels (e) and (f), we report the tunnel coupling and two-body matrix elements. For the initial guess, which has a wildly oscillating density, also the tunnel coupling oscillates strongly and takes on a finite value after the control sequence. We interpret this as a signature of condensate excitations, which go beyond the two-mode MCTDHB model. In contrast, a smooth tunnel pulse and stationary final condensates are achieved for the optimal control. MCTDHB simulations with a higher number of modes indicate that the two-mode approach provides a very accurate level of description \cite{grondscrinzi:10}. The two-body matrix elements of the optimized solutions show a complex behaviour and deviate from each other quite appreciably, which demonstrates that the dynamics depends critically on the orbitals [see equation~\eqref{eq:mvs}].
\begin{figure}
\center\includegraphics[width=0.9\columnwidth]{figure13a_13f}
\caption{Same as figure~\ref{fig:demoOCT1}, but for $U_0 N=1$, $T_{\rm pulse}=4$ and $T_{\rm phase}=5$. (a) The inset magnifies the controls. (b)-(f) describe the same quantities as in figure~\ref{fig:demoOCT1}. \label{fig:demoOCT2}}
\end{figure}
For stronger interactions, we find that it becomes more difficult to decouple the condensate at the end of the \emph{tilting} pulse, as shown in figure~\ref{fig:demoOCT2} for $U_0 N=1$ and $T_{\rm phase}=5$. Although we easily achieve trapping of the orbitals in a stationary state, also for shorter $T_{\rm pulse}$ or $T_{\rm phase}$, we find that number fluctuations are not constant after the control sequence (results are not shown), which is a signature of additional unwanted condensate excitations \cite{grondscrinzi:10}. These excitations are most pronounced when the coherence factor tends to one, and the \emph{ungerade} orbital becomes depopulated. In particular, we could not find pulses shorter than $T_{\rm pulse}=4$ which lead to a final decoupling of the condensates.
The same holds for the initial guess (blue dashed lines in figure~\ref{fig:demoOCT2}). Although the condensate oscillations after the pulse sequence are moderate, they do not uncouple, as indicated by the growing tunnel coupling at later times. In contrast, the OCT control [green solid line in panel (a), magnified in the inset] achieves decoupling and a complete suppression of any tunnelling. Also the two-body matrix elements are stationary to a good degree.
In conclusion, we find that optimized \emph{tilting} pulses are crucial for avoiding condensate oscillations in the phase accumulation stage, and for reducing unwanted condensate oscillations. Adiabatic pulses might be orders of magnitude longer, and thus appreciably reduce the interferometer performance.
\section{Influence of temperature on the coherence of interferometer measurements \label{sec:temp}}
Up to now we considered the atom cloud (BEC) at zero temperature. In realistic experiments the quantum system will be at some finite temperature, and especially for the favorable low dimensional geometries temperature effects and decoherence due to thermal or even quantum fluctuations might become important \cite{Petrov2004,Cazalilla2004,Imambekov2007,Bistritzer2007}. We now turn to look at the effects of temperature on the decoherence of a split BEC interferometer.
\subsection{One-dimensional systems}
In one dimensional quantum systems ($k_B T$ and $\mu \; < \; \hbar \omega_\perp$) fundamental quantum fluctuations prevent the establishment of phase coherence in an infinitely long system even at zero temperature \cite{Cazalilla2004}. The coherence between two points in the longitudinal direction $z_1$ and $z_2$ decays as $ \left|z_1 - z_2 \right|^{-1/2{\cal K}} $, where ${\cal K} = \pi \hbar \sqrt{{n_{1D}}/{(g_{1D} m)}}= \pi n_{1D} \zeta_h$ is the Luttinger parameter. Thereby $\zeta_h = \hbar / \sqrt{m n_{1D} g_{1D}}$ is the healing length, $n_{1D}$ is the 1D density, and $g_{1D}=2 \hbar \omega_\perp a_s$ is the 1D coupling strength in a system with transversal confinement $\omega_\perp$ and scattering length $a_s$. For a weakly interacting 1D system ${\cal K} \gg 1$, and the length scale where quantum fluctuations start to destroy phase coherence is
\begin{equation}
l_\Phi ^{quant} \approx \zeta_h e^{2 {\cal K}}\;.
\end{equation}
For weakly interacting one dimensional Bose gases $\zeta_h$ is in the order of 0.1 to 1 $\mu$m. With $\mathcal K > 10$ one can safely neglect quantum fluctuations (see also \cite{Bistritzer2007}).
These thermal phase fluctuations are also present in a very elongated 3D Bose gas. In such a finite 1D system one can achieve phase coherence (i.e., $l_\Phi ^{therm}$ becomes larger than the longitudinal extension of the atomic cloud) if the temperature is below $T_\Phi = T_{1D} \frac{\hbar \omega_z}{\mu} \approx n_{1D} \zeta_h \hbar \omega_z$. To achieve a practically homogeneous phase along the BEC $T < T_\Phi/10$ is desirable \cite{Petrov2004}.
For interferometry only the relative phase between the two interfering systems and its evolution are important. After a coherent splitting process, even a phase fluctuating condensate is split into two copies with a uniform relative phase. For interferometer measurements decoherence of this definite relative phase is adverse.
The loss of coherence in 1D systems due to thermal excitations was considered theoretically by Burkov, Lukin and Demler \cite{burkov:07}, and by Mazets and Schmiedmayer \cite{Mazets2009}, and probed in an experiment by Hofferberth \emph{et al.} \cite{hofferberth:07} and Jo \emph{et al.} \cite{jo:07}. The coherence in the system left is characterized by the coherence factor
\begin{equation}
\Psi(t) = \left\langle \frac{1}{L} \left| \int_0 ^L dz \; e^{i \theta(z,t)}\right| \right\rangle\;,
\end{equation}
where $\theta(z,t)$ is the relative phase between the two condensates. The angular brackets denote an ensemble average. The key feature in both calculations is that for 1D systems the coherence factor $\Psi(t)$ decays non-exponentially :
\begin{equation}
\Psi(t) \propto \exp\left[{-(t/t_0)^{2/3}}\right]\;.
\end{equation}
Burkov, Lukin and Demler \cite{burkov:07} give for the characteristic time scale
\begin{equation}
t_0 = 2.61 \pi \frac{\hbar \mu}{T^2} \; {\cal K}\;,
\end{equation}
whereas Mazets and Schmiedmayer \cite{Mazets2009} find:
\begin{equation}
t_0 = 3.2 \frac{\hbar \mu }{T^2} \left(\frac{{\cal K}}{\pi}\right)^2 = 3.2 \frac{\hbar^3 n_{1D} ^2}{m T^2}\;.
\end{equation}
Even though the two show a different scaling with the Luttinger parameter $\cal K$, both are consistent with Hofferberth \emph{et al.} \cite{hofferberth:07} within the experimental error bars in the probed range of ${\cal K} \sim 30$.
There are two strategies to get long coherence times:
\begin{itemize}
\item The characteristic timescale for decoherence scales like $T^{-2}$ indicating that very long coherence times can be reached for low temperatures.
\item The time scale of decoherence scales with the 1D-density as $t_0 \propto n_{1D} ^2$ in \cite{Mazets2009}, and $t_0 \propto n_{1D} ^{3/2}$ in \cite{burkov:07}. Increasing the 1D density will enhance the coherence time available for the measurement.
\end{itemize}
Putting all together coherence times $t_0$ in the order of $10^4 \hbar /\mu$ can be achieved for $T \approx 20$ nK and $n_{1D} \approx~ 60$ atoms/$\mu m$.
The above estimates have also to be taken with care, especially since the 1D calculations are only strictly valid for $T < \mu$, but we believe that they are a reasonable estimate also for elongated 3-d systems.
The question is if the required low temperatures ($T \sim \mu/5$ in the above example) can be reached in 1D systems where two-body collisions are frozen out? In the weakly interacting regime, that is for small scattering length and relatively high density, this should be possible, because thermalization can be facilitated by virtual three body collisions \cite{Mazets2008,Mazets2009b,Mazets2010}. In the experiments by Hofferberth \emph{et al.} $T \approx 30$ nK and $n_{1D} \approx~ 60$ atoms/$\mu m$ were achieved \cite{hofferberth:08}
Another point are dynamic excitations in the weakly confined direction. The interaction parameter $U_0$ in general varies when $\lambda$ is changed and this can lead to longitudinal excitations. We find numerically that this effect is very small for the trapping potentials considered. In addition one can in principle account for the effects of the change in interaction energy by controlling the longitudinal confinement.
\subsection{Two-dimensional systems}
Quantum fluctuations are not important in 2D systems since a repulsive bosonic gas exhibits true condensate at $T=0$.
In a two dimensional Bose gas at finite temperature the phase fluctuations scale logarithmically with distance ($r\gg \lambda_T$) \cite{Petrov2004}:
\begin{equation}
\left\langle \left[ \varphi(r)-\varphi(0) \right]\right\rangle_T \approx \frac{2 T}{T_{2D}} \ln \left(\frac{r}{\lambda_T}\right)\;,
\end{equation}
where $T_{2D}=2 \pi \hbar^2 n_{2D}/m$ is the degeneracy temperature for a 2D system, $n_{2D}$ being the peak 2-d density. For $T \gg \mu$ the length scale is $\lambda_T \approx \zeta_h$, where $\zeta_h = \hbar / \sqrt{m n_{2D} g_{2D}}$ is the healing length in the 2D system with the effective coupling $g_{2D}$. For weak interactions it is given by $g_{2D}=\frac{\sqrt{8 \pi} \hbar^2 a_s}{m \, l_\perp}$ with $l_\perp=\sqrt{\frac{\hbar}{m \, \omega_\perp}}$ \cite{Petrov2004}. For $T \ll \mu$ one finds the length scale $\lambda_T \approx \frac{\hbar \, c_s}{T}$, which is equal to the wavelength of thermal phonons. The phase coherence length $l_\Phi$ is given by the distance where the mean square phase fluctuations become of the order of unity.
\begin{equation}
l_\Phi \approx \lambda_T \exp \left(\frac{T_{2D}}{2 T}\right)\; .
\end{equation}
The decoherence in 2D systems was considered by Burkov, Lukin and Demler \cite{burkov:07}, where they find a power law decay of the coherence factor:
\begin{equation}
\Psi(t) \propto t^{-T/8 T_{KT}}\;,
\end{equation}
for times larger then
\begin{equation}
t_0 = \frac{3 \sqrt{3}}{4 \pi} \frac{\mu T_{KT}}{T^3}\;,
\end{equation}
where $\mu = n_{2D} g_{2D}$ is the chemical potential of the 2D system.
The temperature for the Kosterlitz-Thouless transition is given by
\begin{equation}
T_{KT} = \frac{\pi}{2} \frac{n_{2D} ^{sf}}{m} \leq \frac{1}{4}T_{2D}\;,
\end{equation}
where $n_{2D} ^{sf} \leq n_{2D}$ is the super fluid density. Again at sufficiently low temperatures the decoherence due to thermal phase fluctuations is smaller than the phase diffusion due to the non-linear interactions during the phase accumulation stage.
\section{Summary and conclusions}
From our analysis of a double well interferometer for trapped atoms, it becomes evident that the main limiting factor to measurements with atom interferometers is the phase diffusion caused by the non linearity created by the atom-atom interactions. Consequently many of the recent experiments used interferometry to study the intriguing quantum many-body effects caused by interactions \cite{albiez:05,hofferberth:07,jo:07,hofferberth:08,esteve:08}.
Optimal control techniques can help improving interferometer performance significantly by designing optimized splitting ramps and rephasing pulses, but the overall performance of the interferometer is still limited by atom interactions, and not by the readout, except for experiments with very small atom numbers. In general low dimensional confinement of the trapped atomic cloud is better for interferometry. Nevertheless we found it difficult to get a performance for the minimal detectable shift $\Delta E_{\rm min} < 10^{-4} \mu$ even for an optimized setting with a 1D elongated trap. In addition we would like to point out that even though we did our analysis for a generic double well the same will hold for trapped atomic clocks \cite{treutlein:04}, where the signal comes from Ramsey interference of internal states. For the internal state interferometers the difference of the interaction energies is the relevant quantity to compare.
The most direct way to achieve a much improved performance is to decrease the atom atom interaction. The best is to cancel it completely by either putting the atoms in an optical lattice, where on each site the maximal occupancy is 1, or by tuning the scattering length $a_{s}=0$, which can in principle be achieved by employing Feshbach resonances \cite{chin:10}. Drastic reduction of phase diffusion when bringing the scattering length close to zero was recently demonstrated in two experiments in Innsbruck \cite{gustavsson:08} and Firenze \cite{Fattori:08}. The big disadvantage thereby is that using Feshbach resonances requires specific atoms and specific atomic states. These states need to be tunable, and are therefore \textit{not} the '\textit{clock}' states which are insensitive to external fields and disturbances.
In an ideal interferometer one would like to use clock states, create strong squeezing during the splitting process by exploiting the non-linearity in the time evolution due to atom-atom interactions, and then, after the splitting turn off the interactions (by setting the scattering length to $a_s = 0$). All together might be difficult or even impossible to achieve.
For the interferometers considered here one can always reach low enough temperature to neglect decoherence due to thermal excitations even for 1D and 2D systems.
In addition to the interferometer scheme considered here, there exist other ideas of how atom interferometry could be improved. An interesting route will be to exploit in interferometry the correlations of the many boson states, and to establish a readout procedure which is immune to phase diffusion. One approach is to use Bayesian phase estimation schemes for the analysis of the phase sensitivity \cite{pezze:07,pezze:09}. Other proposed schemes include monitoring the coherence and revival dynamics of the condensates \cite{dunningham:04}, the measurement of a phase gradient along the double well potential containing tunnel-coupled condensates by means of a contrast resonance \cite{jae:06}, or inhibition of phase diffusion by quantum noise \cite{khodorkovsky:08}. For those more advanced ways to read out the interference patterns, however, the requirements for the temperature will become more stringent the better the new read-out schemes can compensate for the adverse effects of phase diffusion.
\section*{Acknowledgments}
We thank J. Chwede\'nczuk, F. Piazza, A. Smerzi, G. von Winckel, O. Alon, and T. Schumm for helpful discussions. This work has been supported in part by NAWI GASS, the FWF and the ESF Euroscores program: EuroQuaser project QuDeGPM.
|
\section{Introduction}
The interaction of hadrons in nuclei is a problem that goes to the very core of
nuclear physics. In fact, our present knowledge of the nuclear force in free space is, in itself,
the result of decades of struggle,\cite{Mac89} which will not be reviewed in this article.
The nature of the nuclear force in the medium is of course a much more complex problem,
as it involves aspects of the force that cannot be constrained
through free-space nucleon-nucleon (NN) scattering. Predictions of properties of nuclei are
the ultimate test for many-body theories.
Nuclear matter is an alternative and convenient theoretical laboratory for many-body theories. By ``nuclear matter" we mean an infinite system
of nucleons acted on by their mutual strong forces and no electromagnetic interactions. Nuclear matter
is characterized by its energy/particle as a function of density and other thermodynamic quantities, if
appropriate (e.g.~temperature). Such relation is known as the nuclear matter equation of state (EoS).
The translational invariance of the system facilitates theoretical calculations. At the same time, adopting
what is known as
``local density approximation", one can use the EoS directly in calculations of finite systems. This procedure
is applied,
for instance, in Thomas-Fermi calculations within the liquid drop model, where an appropriate energy functional is
written in terms of the EoS.\cite{Oya98,Furn,SL09}
Symmetric nuclear matter (that is, matter with equal densities of protons and
neutrons) has been studied extensively. The so-called conventional
approach goes back to earlier works by Brueckner and others\cite{5,6,7,8} and is known
as the Brueckner-Hartree-Fock (BHF) theory.
The Brueckner theory is based on a linked-cluster perturbation series of the ground state energy
of a many-body system.\cite{5,6,7,8,HT70,10} Such series was shown to converge when the cluster diagrams are
regrouped according to the number of hole lines.
The variational approach was also pursued as an
alternative method\cite{11,12} and yielded predictions in close agreement with those from Brueckner theory
when realistic NN potentials were employed.\cite{13}
During the 1980's, the Dirac-Brueckner-Hartree-Fock (DBHF) approach
was developed.\cite{14,15,BM84} The main
break-through came with the observation that the DBHF theory, unlike
conventional Brueckner theory,
could describe successfully the saturation properties of nuclear matter, that
is, saturation energy and density of the equation of state.
The DBHF method adopts realistic NN interactions and contains important relativistic features.
It describes the nuclear mean field in terms of
strong, competing scalar and vector fields that, together, account for the binding
of nucleons as well as the large spin-orbit splitting seen in nuclear states.
Regardless of the chosen many-body theory, a quantitative NN potential should be part of its input.
Recently, chiral effective theories of the nuclear force\cite{chi} have become popular as a mean to respect
the symmetries of Quantum Chromo Dynamics (QCD) while retaining the basic degrees of freedom typical of low-energy nuclear physics.
Chiral effective theories provide a well-defined scheme to determine the appropriate many-body
diagrams to be included at each order of the perturbation. However, being based on a low-momentum
expansion, interactions derived from chiral perturbation theory are not suitable for applications in dense nuclear/neutron
matter, where high Fermi momenta are involved. Instead,
meson-theoretic or phenomenological NN potentials are typically employed
as input to the many-body theory.
Mean-field models (relativistic and non-relativistic)
are a popular, although non-microscopic, alternative to methods based on the in-medium reaction matrix, such
as BHF and DBHF.
It is the purpose of this article to review the status of microscopic studies of nuclear and
neutron-rich matter, with particular emphasis on the latter, as it relates
to present empirical investigations within terrestrial laboratories
(heavy-ion collisions) or astrophysical observations (neutron stars). At the same time, we will review the
present status concerning available empirical information which can be utilized to guide and constrain
theories.
We also wish to provide a self-contained account of the recent work with asymmetric matter done by the Idaho group,
including extensive numerical tables for the interested user.
Clearly, the goal
to describe the properties of (dense) many-body systems
consistently from the underlying forces {\it and} including all potentially important mechanisms,
is a most ambitious program and far from having been completed. The importance of pursuing a microscopic approach towards the accomplishment of
this goal is a theme
that will surface repeatedly throughout this article.
\section{Asymmetric Nuclear Matter: Some Facts and Phenomenology}
Isospin-asymmetric nuclear matter (IANM) simulates the interior of a ``nucleus" with unequal densities of protons and neutrons.
The equation of state of (cold) IANM is then a function of density as well as the relative concentrations
of protons and neutrons.
The recent and fast-growing interest in IANM stems from its close connection to neutron-rich nuclei, or,
more generally, asymmetric nuclei, including the very ``exotic" ones known as ``halo" nuclei.
At this time, the boundaries of the nuclear chart are uncertain, with a few hundreds stable nuclides
known to exist and perhaps a few thousands believed to exist.
The Facility for Rare Isotope Beams (FRIB) has recently been approved for design and construction at
Michigan State University (MSU).
The facility will deliver intense beams of rare isotopes, the study of which can provide crucial
information on short-lived elements normally not found on earth.\cite{FRIB}
Thus, this new experimental program will have widespread impact, ranging from the origin of elements to the
evolution of the cosmos.
It is estimated that
the design and construction of FRIB will take ten years.\cite{FRIB}
In the meantime, systematic investigations to determine
the properties of asymmetric nuclear matter are proliferating at existing facilities.
From the theoretical side, some older studies of IANM can be found in Refs.\cite{BC68,Siem}
Interactions adjusted to fit properties of finite nuclei, such as those based
on the non-relativistic Skyrme Hartree-Fock theory\cite{B+75} or the relativistic mean field
theory,\cite{ST94} have been used to extract
phenomenological EoS.
Variational calculations of asymmetric matter have also been reported.\cite{12,APR}
Fuchs {\it et al.}\cite{FLW} defined a Lorentz invariant functional of the baryon field operators
to project Dirac-Brueckner nuclear matter results onto the meson-nucleon vertices of an effective
density-dependent field theory. This was then applied to asymmetric matter and finite nuclei
in Hartree calculations.\cite{HKL}
Extensive work with IANM has also been reported by Lombardo and collaborators.\cite{Catania1,Catania2}
Dirac-Brueckner-Hartree-Fock
calculations of IANM properties were performed by the Oslo group,\cite{Oslo}
the Idaho group,\cite{AS03} and by Fuchs and collaborators.\cite{Fuchs}
Typically, considerable model dependence is observed among the different EoS of IANM, especially in the
high-density region.
Asymmetric nuclear matter can be characterized by the neutron density,
$\rho_n$, and the proton density, $\rho_p$, defined as the number of neutrons or protons per unit of volume.
In infinite matter, they are obtained by summing the neutron or proton states per volume (up to their respective
Fermi momenta, $k^{n}_{F}$ or $k^{p}_{F}$) and applying the appropriate degeneracy factor. The result is
\begin{equation}
\rho_i =\frac{ (k^{i}_{F})^3}{3 \pi ^2} , \label{rhonp}
\end{equation}
with $i=n$ or $p$.
It is more convenient to refer to the total density
$\rho = \rho_n + \rho_p$ and the asymmetry (or neutron excess) parameter
$\alpha = \frac{ \rho_n - \rho_p}{\rho}$.
Clearly, $\alpha$=0 corresponds to symmetric matter and
$\alpha$=1 to neutron matter.
In terms of $\alpha$ and the average Fermi momentum, $k_F$, related to the total density in the usual way,
\begin{equation}
\rho =\frac{2 k_F^3}{3 \pi ^2} , \label{rho}
\end{equation}
the neutron and proton Fermi momenta can be expressed as
\begin{equation}
k^{n}_{F} = k_F{(1 + \alpha)}^{1/3} \label{kfn}
\end{equation}
and
\begin{equation}
k^{p}_{F} = k_F{(1 - \alpha)}^{1/3} , \label{kfp}
\end{equation}
respectively.
The energy/particle in IANM can, to a very good degree of approximation, be written as
\begin{equation}
e(\rho, \alpha) \approx e_0({\rho}) + e_{sym}(\rho)\alpha ^2, \label{e}
\end{equation}
where the first term is the energy/particle in symmetric matter and
$e_{sym}$ is known as the symmetry energy. In the Bethe-Weizs{\" a}cker formula for the nuclear binding energy, it represents the amount of binding a nucleus has
to lose when the numbers of protons and neutrons are unequal.
The symmetry energy is also closely related to
the neutron $\beta$-decay in dense matter, whose threshold depends on the proton fraction.
A typical value for $e_{sym}$
at nuclear matter density ($\rho_0$) is 30 MeV,
with theoretical predictions spreading approximately between 26 and 35 MeV.
The effect of a term of fourth order in the asymmetry parameter (${\cal O}(\alpha ^4)$) on the bulk properties of neutron stars
is very small, although it may impact the proton fraction at high density.
More generally,
non-quadratic terms are usually associated with isovector pairing, which is a surface effect and thus vanishes
in infinite matter.\cite{Steiner}
Equation~(\ref{e}) displays a convenient separation between the symmetric and aymmetric parts of the EoS,
which facilitates the identification of observables that may be sensitive, for instance, mainly to the
symmetry energy. At this time, groups from GSI,\cite{GSI} MSU,\cite{Tsang} Italy,\cite{Greco} France,\cite{IPN} and
China\cite{China2,China3,China4}
are investigating the density dependence of the symmetry energy through heavy-ion collisions.
From recent results, such as those reported at the 2009 ``International Workshop on Nuclear Dynamics
in Heavy-ion Reactions and the Symmetry Energy"
(Shanghai, China, August 22-25, 2009), these investigations appear to agree reasonably well on the following parametrization
of the symmetry energy:
\begin{equation}
e_{sym}(\rho) = 12.5 \, MeV \Big (\frac{\rho}{\rho_0}\Big )^{2/3} +
17.5 \, MeV \Big (\frac{\rho}{\rho_0}\Big )^{\gamma_i}, \label{es}
\end{equation}
where the first term is the kinetic contribution and
$\gamma_i$ (the exponent appearing in the potential energy part) is found to be between 0.4 and 1.0.
Naturally, there are uncertainties associated with all transport models.
Recent constraints from MSU\cite{Tsang} were extracted from simulations of $^{112}$Sn
and $^{124}$Sn collisions with an Improved Quantum Molecular Dynamics transport model and are
consistent with isospin diffusion data and the ratio of neutron and proton spectra.
Typically, parametrizations like the one given in Eq.~(\ref{es}) are valid
at or below the saturation density, $\rho_0$. Efforts to constrain the behavior of the symmetry energy
at higher densities
are presently being pursued through observables such as $\pi ^-/\pi^+$ ratio,
$K ^+/K^0$ ratio, neutron/proton differential transverse flow, or nucleon elliptic flow.\cite{Ko09}
A more detailed discussion of the symmetry energy and the status of its theoretical predictions will be presented in
Sec.~3.2.3.
\section{The Microscopic Approach}
In this section, we present a discussion of the microscopic approach to nuclear matter, in general, and the DBHF
method, in particular.
\subsection{The two-body potential}
By {\it ab initio} we mean that the starting point of the many-body calculation is a realistic NN interaction which is then applied in the
nuclear medium without any additional free parameters.
Thus the first question to be confronted concerns the choice of the ``best" NN interaction.
As already mentioned in the Introduction,
after the development of QCD and the understanding of its symmetries,
chiral effective theories\cite{chi} were developed as a way to respect the
symmetries of QCD while keeping the degrees of freedom (nucleons and pions) typical of low-energy nuclear physics. However,
chiral perturbation theory (ChPT)
has definite limitations as far as the range of allowed momenta is concerned.
For the purpose of applications in dense matter, where higher and higher momenta become involved
with increasing Fermi momentum, NN potentials based on ChPT are unsuitable.
Relativistic meson theory is an appropriate framework to deal with the high momenta encountered in dense
matter. In particular,
the one-boson-exchange (OBE) model has proven very successful in describing NN data in free space
and has a good theoretical foundation.
Among the many available OBE potentials, some being part of the ``high-precision generation",\cite{pot1,pot2,pot3}
we seek a momentum-space potential developed within a relativistic scattering equation, such as the
one obtained through the Thompson\cite{Thom} three-dimensional reduction of the Bethe-Salpeter equation.\cite{BS}
Furthermore, we require a potential that uses
the pseudovector coupling for the interaction of nucleons with pseudoscalar mesons.
With these constraints in mind,
as well as the requirement of a good description of the NN data,
Bonn B\cite{Mac89} is a reasonable choice. As is well known, the NN potential model dependence
of nuclear matter predictions is not negligible. The saturation points obtained with different NN potentials
move along the famous ``Coester band" depending on the strength of the tensor force, with the weakest tensor
force yielding the largest attraction. This can be understood in terms of medium effects (particularly
Pauli blocking) reducing the (attractive) second-order term in the expansion of the reaction matrix.
A large second-order term will undergo a large reduction in the medium. Therefore, noticing that the second-order term
is dominated by the tensor component of the force, nuclear potentials with a strong tensor component will
yield less attraction in the medium.
For the same reason (that is, the role of the tensor force in
nuclear matter),
the potential model dependence is strongly reduced in pure (or nearly pure) neutron matter, due to the
absence of isospin-zero partial waves.
Already when QCD (and its symmetries) were unknown, it was observed that the contribution from the
nucleon-antinucleon pair diagram, Fig.~\ref{2b}, becomes unreasonably large if the pseudoscalar (ps) coupling is used,
leading to very large pion-nucleon scattering lengths.\cite{GB79}
We recall that the Lagrangian density for pseudoscalar coupling of the nucleon field ($\psi$) with the pseudoscalar meson
field ($\phi$) is
\begin{equation}
{\cal L}_{ps} = -ig_{ps}\bar {\psi} \gamma _5 \psi \phi. \label{ps}
\end{equation}
On the other hand, the same contribution (Fig.~\ref{2b})
is heavily reduced by the pseudovector (pv) coupling (a mechanism which
became known as ``pair suppression"). The reason for the suppression is the presence of the
covariant derivative
at the pseudovector vertex,
\begin{equation}
{\cal L}_{pv} = \frac{f_{ps}}{m_{ps}}{\bar \psi} \gamma _5 \gamma^{\mu}\psi \partial_{\mu} \phi,
\label{pv}
\end{equation}
which reduces the contribution of the vertex for low momenta and, thus,
explains the small value of the pion-nucleon
scattering length at threshold.\cite{GB79}
Considerations based on chiral symmetry\cite{GB79} can further motivate
the choice of the pseudovector coupling.
\begin{figure}
\centering
\vspace*{-3.2cm}
\hspace*{-2.0cm}
\scalebox{1.0}{\includegraphics{ijmpe1.ps}}
\vspace*{-21.0cm}
\caption{Contribution to the NN interaction from virtual pair excitation.
Upward- and downward-pointing arrows represent nucleons and antinucleons, respectively.
Dashed lines denote mesons.
}
\label{2b}
\end{figure}
In closing this section, we wish to highlight
the most important aspect of the ``{\it ab initio}" approach: namely, the only free parameters of the
model (the parameters of the NN potential)
are determined by fitting the free-space NN data and never readjusted in the medium. In other
words, the model parameters are tightly constrained and the calculation in the medium is
parameter free.
The presence of free parameters in the medium would generate effects and sensitivities which are hard to
control and interfere with the predictive power of the theory.
\subsection{The Dirac-Brueckner-Hartree-Fock approach to symmetric and asymmetric nuclear matter}
\subsubsection{Formalism}
The main strength of the DBHF approach is in its inherent ability to account for important three-body forces
through its density dependence.
In Fig.~\ref{3b} we show a three-body force (TBF) originating from virtual excitation of a nucleon-antinucleon pair,
known as ``Z-diagram". Notice that the observations from the previous section ensure that the corresponding diagram
at the two-body level, Fig.~\ref{2b}, is moderate in size when the pv coupling is used.
The main feature of
the DBHF method turns out to be closely related to
the TBF depicted in Fig.~\ref{3b}, as we will argue next. In the DBHF approach, one describes the positive energy solutions
of the Dirac equation in the medium as
\begin{equation}
u^*(p,\lambda) = \left (\frac{E^*_p+m^*}{2m^*}\right )^{1/2}
\left( \begin{array}{c}
{\bf 1} \\
\frac{\sigma \cdot \vec {p}}{E^*_p+m^*}
\end{array}
\right) \;
\chi_{\lambda},
\label{ustar}
\end{equation}
where the effective mass, $m^*$, is defined as $m^* = m+U_S$, with $U_S$ an attractive scalar potential.
(This will be derived below.)
It can be shown that both the description of a single-nucleon via Eq.~(\ref{ustar}) and the evaluation of the
Z-diagram, Fig.~\ref{3b}, generate a repulsive effect on the energy/particle in symmetric nuclear matter which depends on the density approximately
as
\begin{equation}
\Delta E \propto \left (\frac{\rho}{\rho_0}\right )^{8/3} \, ,
\label{delE}
\end {equation}
and provides the saturating mechanism missing from conventional Brueckner calculations.
(Alternatively, explicit TBF are used along with the BHF method in order to achieve a similar result.)
Brown showed that the bulk of the desired effect can be obtained as a lowest order (in $p^2/m$) relativistic correction
to the single-particle propagation.\cite{GB87}
With the in-medium spinor as in Eq.~(\ref{ustar}), the correction to the free-space spinor can be written
approximately as
\begin{equation}
u^*(p,\lambda) -u(p,\lambda)\approx
\left( \begin{array}{c}
{\bf 0} \\
-\frac{\sigma \cdot \vec {p}}{2 m^2}U_S
\end{array}
\right) \;
\chi_{\lambda},
\label{delu}
\end{equation}
where for simplicity the spinor normalization factor has been set equal to 1, in which case it
is clearly seen that the entire effect originates from the modification of the spinor's lower component.
By expanding the single-particle energy to order $U_S^2$, Brown showed that the correction to the
energy consistent with Eq.~(\ref{delu}) can be written as $\frac{p^2}{2m}(\frac{U_S}{m})^2$. He then proceeded to
estimate the correction to the energy/particle and found it to be approximately as given in Eq.~(\ref{delE}).
\begin{figure}[!t]
\centering
\vspace*{-3.2cm}
\hspace*{-2.0cm}
\scalebox{0.9}{\includegraphics{ijmpe2.ps}}
\vspace*{-19.0cm}
\caption{Three-body force due to virtual pair excitation. Conventions as in the previous figure.
}
\label{3b}
\end{figure}
The approximate equivalence of the effective-mass description of Dirac states and the contribution from the Z-diagram
has a simple intuitive explanation in the observation
that Eq.~(\ref{ustar}), like any other solution of the Dirac equation,
can be written as a superposition of positive and negative energy solutions. On the other hand, the ``nucleon" in the
middle of the Z-diagram, Fig.~\ref{3b}, is precisely a superposition of positive and negative energy states.
In summary, the DBHF method effectively takes into account a particular class of
TBF, which are crucial for nuclear matter saturation.
Of course,
other, more popular, three-body forces (not included in DBHF) need to be addressed as well.
Figure~\ref{Delta} shows the TBF that is included in essentially all TBF models, regardless
other components; it is the Fujita-Miyazawa TBF.\cite{FM}
With the addition of contributions from $\pi N$ S-waves, one ends up with the
well-known Tucson-Melbourne TBF.\cite{TM} The microscopic TBF of Ref.\cite{Catania3} includes
contributions from excitation of the Roper resonance (P$_{11}$ isobar) as well.
Now, if diagrams such as the one shown on the left-hand side of Fig.~\ref{Delta} are taken into account,
consistency requires that medium modifications at the corresponding two-body level are also included,
that is, the diagram on the right-hand side of Fig.~\ref{Delta} should be present and properly medium modified.
Large cancellations then take place, a fact that was brought up a long time ago.\cite{DMF}
When the two-body sector is handled via OBE diagrams, the two-pion exchange is
effectively incorporated through the $\sigma$ ``meson", which
cannot generate the (large) medium effects (dispersion and Pauli blocking on $\Delta$
intermediate states) required by the consistency arguments presented above. Thus, caution needs to
be exercised when applying TBF in a particular space (of nucleons only, or nucleons and $\Delta$).
\begin{figure}
\begin{center}
\vspace*{-3.0cm}
\hspace*{-2.0cm}
\scalebox{0.9}{\includegraphics{ijmpe3.ps}}
\vspace*{-19.0cm}
\caption{Left: three-body force arising from $\Delta$-isobar excitation (thick line).
Right: two-meson exchange contribution to the NN interaction involving
$\Delta$-isobar excitation.
}
\label{Delta}
\end{center}
\end{figure}
Having first summarized the main DBHF philosophy,
we now proceed to describe the DBHF calculation of IANM.\cite{AS03}
In the end, this will take us back to the crucial point of the DBHF approximation, Eq.~(\ref{ustar}).
We start from the Thompson\cite{Thom} relativistic three-dimensional reduction
of the Bethe-Salpeter equation.\cite{BS} The Thompson equation is applied to nuclear matter in
strict analogy to free-space scattering and reads, in the nuclear matter rest frame,
\begin{eqnarray}
&& g_{ij}(\vec q',\vec q,\vec P,(\epsilon ^*_{ij})_0) = v_{ij}^*(\vec q',\vec q) \nonumber \\
&& + \int \frac{d^3K}{(2\pi)^3}v^*_{ij}(\vec q',\vec K)\frac{m^*_i m^*_j}{E^*_i E^*_j}
\frac{Q_{ij}(\vec K,\vec P)}{(\epsilon ^*_{ij})_0 -\epsilon ^*_{ij}(\vec P,\vec K)}
g_{ij}(\vec K,\vec q,\vec P,(\epsilon^*_{ij})_0) \, ,
\label{gij}
\end{eqnarray}
where $g_{ij}$ is the in-medium reaction matrix
($ij$=$nn$, $pp$, or $np$), and the
asterix signifies that medium effects are applied to those quantities. Thus the NN potential,
$v_{ij}^*$, is constructed in terms of effective Dirac states (in-medium spinors) as explained above.
In Eq.~(\ref{gij}),
$\vec q$, $\vec q'$, and $\vec K$ are the initial, final, and intermediate
relative momenta, and $E^*_i = \sqrt{(m^*_i)^2 + K^2}$.
The momenta of the two interacting particles in the nuclear matter rest frame have been expressed in terms of their
relative momentum and the center-of-mass momentum, $\vec P$, through
\begin{equation}
\vec P = \vec k_{1} + \vec k_{2} \, , \label{P}
\end{equation}
\begin{equation}
\vec K = \frac{\vec k_{1} - \vec k_{2}}{2} \, . \label{K}
\end{equation}
The energy of the two-particle system is
\begin{equation}
\epsilon ^*_{ij}(\vec P, \vec K) =
e^*_{i}(\vec P, \vec K)+
e^*_{j}(\vec P, \vec K)
\label{eij}
\end{equation}
and $(\epsilon ^*_{ij})_0$ is the starting energy.
The single-particle energy $e_i^*$ includes kinetic energy and potential
energy contributions (see Eq.~(\ref{spe}) below).
The Pauli operator, $Q_{ij}$, prevents scattering to occupied $nn$, $pp$, or $np$ states.
To eliminate the angular
dependence from the kernel of Eq.~(\ref{gij}), it is customary to replace the exact
Pauli operator with its angle-average.
Detailed expressions for the Pauli operator
and the average center-of-mass momentum in the case of two different Fermi seas
can be found in Ref.\cite{AS03}.
With the definitions
\begin{equation}
G_{ij} = \frac{m^*_i}{E_i^*(\vec{q'})}g_{ij}
\frac{m^*_j}{E_j^*(\vec{q})}
\label{Gij}
\end{equation}
and
\begin{equation}
V_{ij}^* = \frac{m^*_i}{E_i^*(\vec{q'})}v_{ij}^*
\frac{m^*_j}{E_j^*(\vec{q})} \, ,
\label{Vij}
\end{equation}
one can rewrite Eq.~(\ref{gij}) as
\begin{eqnarray}
&& G_{ij}(\vec q',\vec q,\vec P,(\epsilon ^*_{ij})_0) = V_{ij}^*(\vec q',\vec q) \nonumber \\[4pt]
&& + \int \frac{d^3K}{(2\pi)^3}V^*_{ij}(\vec q',\vec K)
\frac{Q_{ij}(\vec K,\vec P)}{(\epsilon ^*_{ij})_0 -\epsilon ^*_{ij}(\vec P,\vec K)}
G_{ij}(\vec K,\vec q,\vec P,(\epsilon^*_{ij})_0) \, ,
\label{Geq}
\end{eqnarray}
which is formally identical to its non-relativistic counterpart.
The goal is to determine self-consistently the nuclear matter single-particle potential
which, for IANM, will be different for neutrons and protons.
To facilitate the description of the procedure, we will use a schematic
notation for the neutron/proton potential.
We write, for neutrons,
\begin{equation}
U_n = U_{np} + U_{nn} \; ,
\label{un}
\end{equation}
and for protons
\begin{equation}
U_p = U_{pn} + U_{pp} \, ,
\label{up}
\end{equation}
where each of the four pieces on the right-hand-side of Eqs.~(\ref{un}-\ref{up}) signifies an integral of the appropriate
$G$-matrix ($nn$, $pp$, or $np$) obtained from Eq.~(\ref{Geq}).
Clearly, the two equations above are coupled through
the $np$ component and so they must be solved simultaneously. Furthermore,
the $G$-matrix equation and Eqs.~(\ref{un}-\ref{up})
are coupled through the single-particle energy (which includes the single-particle
potential, itself defined in terms of the $G$-matrix). So we have a coupled system to be solved self-consistently.
Before proceeding with the self-consistency,
one needs an {\it ansatz} for the single-particle potential. The latter is suggested by
the most general structure of the nucleon self-energy operator consistent with
all symmetry requirements. That is:
\begin{equation}
{\cal U}_i({\vec p}) = U_{S,i}(p) + \gamma_0 U_{V,i}^{0}(p) - {\bf \gamma}\cdot {\vec p} U_{V,i}(p) \, ,
\label{Ui1}
\end{equation}
where $U_{S,i}$ and
$U_{V,i}$ are an attractive scalar field and a repulsive vector field, respectively, with
$ U_{V,i}^{0}$ the timelike component of the vector field. These fields are in general density and momentum dependent.
We take
\begin{equation}
{\cal U}_i({\vec p}) \approx U_{S,i}(p) + \gamma_0 U_{V,i}^{0}(p) \, ,
\label{Ui2}
\end{equation}
which amounts to assuming that the spacelike component of the vector field is much smaller than
both $U_{S,i}$ and $U_{V,i}^0$. Furthermore, neglecting the momentum dependence of the scalar and
vector fields and inserting Eq.~(\ref{Ui2}) in the Dirac equation for neutrons/protons propagating in
nuclear matter,
\begin{equation}
(\gamma _{\mu}p^{\mu} - m_i - {\cal U}_i({\vec p})) u_i({\vec p},\lambda) = 0 \, ,
\label{Dirac1}
\end{equation}
naturally leads to rewriting the Dirac equation in the form
\begin{equation}
(\gamma _{\mu}p^{\mu *} - m_i^*) u_i({\vec p},\lambda) = 0 \, ,
\label{Dirac2}
\end{equation}
with positive energy solutions as in Eq.~(\ref{ustar}), $m_i^* = m + U_{S,i}$, and
\begin{equation}
(p^0)^* = p^0 - U_{V,i}^0 (p) \, .
\label{p0}
\end{equation}
The subscript ``$i$'' signifies that these parameters are different for protons and
neutrons.
As in the symmetric matter case,\cite{BM84} evaluating the expectation value of Eq.~(\ref{Ui2})
leads to a parametrization of
the single particle potential for protons and neutrons (Eqs.(\ref{un}-\ref{up})) in terms of the
constants $U_{S,i}$ and $U_{V,i}^0$ which is given by
\begin{equation}
U_i(p) = \frac{m^*_i}{E^*_i}<{\vec p}|{\cal U}_i({\vec p})|{\bf p}> =
\frac{m^*_i}{E^*_i}U_{S,i} + U_{V,i}^0 \; .
\label{Ui3}
\end{equation}
Also,
\begin{equation}
U_i(p) =
\sum_{p'_j \le k_F^i} G_{ij}({\vec p} _i,{\vec p}'_j) \; ,
\label{Ui4}
\end{equation}
which, along with Eq.~(\ref{Ui3}), allows the self-consistent determination of the single-particle
potential as explained below.
The kinetic contribution to the single-particle energy is
\begin{equation}
T_i(p) = \frac{m^*_i}{E^*_i}<{\vec p}|\gamma \cdot {\vec p} + m|{\vec p}> =
\frac{m_i m^*_i + {\vec p}^2}{E^*_i} \; ,
\label{KE}
\end{equation}
and the single-particle energy is
\begin{equation}
e^*_i(p) = T_i(p) + U_i(p) = E^*_i + U^0_{V,i} \; .
\label{spe}
\end{equation}
The constants $m_i^*$ and
\begin{equation}
U_{0,i} = U_{S,i} + U_{V,i}^0
\label{U0i}
\end{equation}
are convenient to work with as they
facilitate
the connection with the usual non-relativistic framework.\cite{HT70}
Starting from some initial values of $m^*_i$ and $U_{0,i}$, the $G$-matrix equation is
solved and a first approximation for $U_{i}(p)$ is obtained by integrating the $G$-matrix
over the appropriate Fermi sea, see Eq.~(\ref{Ui4}). This solution is
again parametrized in terms of a new set of constants, determined by fitting the parametrized $U_i$,
Eq.~(\ref{Ui3}),
to its values calculated at two momenta, a procedure known as the ``reference spectrum approximation".
The iterative procedure is repeated until convergence is reached.
Finally, the energy per neutron or proton in nuclear matter is calculated from
\begin{equation}
\bar{e}_{i} = \frac{1}{A}<T_{i}> + \frac{1}{2A}<U_{i}> -m \; .
\label{ei}
\end{equation}
The EoS, or energy per nucleon as a function of density, is then written as
\begin{equation}
\bar{e}(\rho_n,\rho_p) = \frac{\rho_n \bar{e}_n + \rho_p \bar{e}_p}{\rho} \, ,
\label{enp}
\end{equation}
or
\begin{equation}
\bar{e}(k_F,\alpha) = \frac{(1 + \alpha) \bar{e}_n + (1-\alpha) \bar{e}_p}{2} \, .
\label{eav}
\end{equation}
Clearly, symmetric nuclear matter is obtained as a by-product of the calculation described above
by setting $\alpha$=0.
In the DBHF calculation of Ref.\cite{Fuchs},
a similar scheme is applied to obtain the self-consistent $G$-matrix and spinor basis. At each
step of the iterative procedure, the nucleon self-energy is first calculated by integrating the $G$-matrix
elements over the Fermi sea.
The self-energy components are then obtained from
the appropriate traces\cite{Fuchs}, i.~e.
\begin{equation}
\Sigma _S = \frac{1}{4} tr[\Sigma] \, , \, \, \,
\Sigma _0 = -\frac{1}{4} tr[\gamma _0\Sigma] \, , \, \, \,
\Sigma _V = -\frac{1}{4|{\vec p}|^2} tr[\gamma \cdot {\vec p}\Sigma] \, .
\label{Sig}
\end{equation}
A covariant representation of the $G$-matrix is used to facilitate the transitions between the two-nucleon
center-of-mass frame and the nuclear matter rest frame.
Again, the presence of the medium naturally leads to define an effective mass, which
is given by\cite{Fuchs}
\begin{equation}
m^*(p,k_F) = (m + Re \Sigma _S(p,k_F))/(1 + \Sigma _V(p,k_F)) \, .
\label{mstar}
\end{equation}
Although the effective mass is density and momentum dependent, ultimately it is defined as the
value of the expression given in Eq.~(\ref{mstar}) at $p=k_F$. Thus, the ``reference spectrum
approximation" is also employed in Ref.\cite{Fuchs}, although differently
in some of the technical aspects as compared to the Idaho method.
These technical aspects appear to impact the predictions of some isovector quantities, as it will be
discussed in Sec.~3.2.4.
\subsubsection{EoS predictions with the DBHF approach}
In Fig.~\ref{eos}, we show EoS predictions for symmetric matter (solid red) and neutron matter (dashed black)
as obtained from the Idaho calculation described in the previous section.
Equation~(4) then gives the EoS values for any $\alpha$, a behavior which can be verified to be
approximately true.\cite{AS03}
The EoS from DBHF can be characterized as being moderately soft at low to medium density
and fairly ``stiff" at
high densities.
The predicted saturation density and energy for the symmetric matter EoS in Fig.~\ref{eos} are equal to 0.185 fm$^{-3}$ and -16.14 MeV, respectively,
and the compression modulus is 252 MeV. For comparison, the same saturation obserbables as
predicted by the other DBHF model presently on the market\cite{Fuchs} are 0.181 fm$^{-3}$ and -16.15 MeV
for saturation density and energy, and 230 MeV for the incompressibility.
The increased stiffness featured by the DBHF EoS at the higher densities
originates from the strongly density-dependent repulsion inherent to the
Dirac-Brueckner-Hartee-Fock method.
In Ref.\cite{Fuchs2}, it is pointed out
that constraints from neutron star phenomenology together with flow data from heavy-ion
reactions suggest that such EoS behavior may be desirable.
We will come back to this point later, in conjunction with neutron star predictions.
\begin{figure}
\begin{center}
\vspace*{-1.0cm}
\hspace*{-2.0cm}
\scalebox{0.3}{\includegraphics{ijmpe4.ps}}
\vspace*{-2.5cm}
\caption{DBHF predictions for the EoS of symmetric matter (solid red) and neutron matter (dashed black).
}
\label{eos}
\end{center}
\end{figure}
At this point, it it useful
to take a closer look at
some of the predictions included in the analysis of Ref.\cite{Fuchs2}, such as
relativistic mean field (RMF) models. Examples are those from
Refs.\cite{DD,D3C}, which use density-dependent (``DD") meson couplings and are
fitted to the properties of nuclei up to about 0.15 fm$^{-3}$. They generate the steepest
EoS and thus the largest pressures, see Fig.~\ref{Klaen}.
An improvement to the traditional RMF description of nuclear matter
can be obtained through the introduction of non-linear (``NL") self-interactions of the $\sigma$ meson, such as done
in the models of Refs.\cite{NL1,NL2}, with the parametrization of Ref.\cite{NL2} including
both the $\delta$ and the $\rho$ mesons in the isovector channel. The corresponding EoS are much less repulsive
than those of ``DD" models (although the symmetry energy becomes very large at high
density, possibly due to the absence of non-linearity and density dependence at the
isovector level).
Clearly, the pressure as a function of density plays the crucial role in building the structure of
a neutron star.
In Fig.~\ref{psm} we show the pressure in symmetric matter as predicted by the Idaho calculation compared with constraints obtained
from flow data.\cite{MSU}
The predictions are seen to fall just on the high side of the constraints and grow
rather steep at high density.
Comparing with Fig.~\ref{Klaen},
we see that
the Idaho predictions
are well below those of DD-RMF models at low to
moderate density but nearly catch up with them at very high density, a description that
would also be appropriate for the
DBHF predictions of Ref.\cite{Fuchs} (red curve in Fig.~\ref{Klaen}).
Of all the cases studied in
Ref.\cite{Fuchs2}, DD-RMF models predict the largest maximum masses and radii and the
lowest central densities. Thus,
an equation of state where high pressure is sustained for a longer radial distance (moving away
from the center of the star)
will allow the maximum mass star to be heavier, larger, and more ``diffuse" at the center.
On the other hand,
microscopic relativistic models, (such as the DBHF calculation of Ref.\cite{Fuchs}
or the present one, which are in fair agreement with each other), display a rather different
density dependence of the pressure and produce smaller and more compact maximum mass stars.
(All other EoS considered in Ref.\cite{Fuchs2} are softer and generate smaller maximum masses
with smaller radii and larger central densities.)
To conclude this section, we show in Fig.~\ref{pnm} the pressure in neutron matter (red curve)
and $\beta$-equilibrated matter (green) as predicted by DBHF calculations. The pressure contour is again from Ref.\cite{MSU}
and was obtained from flow data together with the assumption of strong density dependence in
the asymmetry term (indicated as ``Asy\_ stiff'' in Ref.\cite{MSU}).
\begin{figure}
\begin{center}
\scalebox{0.28}{\includegraphics[angle=-90]{ijmpe6.ps}}
\vspace*{0.6cm}
\caption{Pressure in symmetric matter predicetd by various models. The shaded area corresponds to the region
of pressure consistent with the flow data analysed in Ref.$^{56}$. [Figure reprinted with permission
from T. Kl{\"a}hn.$^{51}$ Copyright (2006) by the American Physical Society.
http://prc.aps.org/abstract/PRC/v74/i3/e035802]
}
\label{Klaen}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\hspace*{-2.0cm}
\scalebox{0.4}{\includegraphics{ijmpe5.ps}}
\vspace*{-2.5cm}
\caption{Pressure in symmetric matter from the Idaho DBHF calculation. The shaded area corresponds to the region
of pressure consistent with the flow data analysed in Ref.$^{56}$.}
\label{psm}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\hspace*{-2.0cm}
\scalebox{0.4}{\includegraphics{ijmpe7.ps}}
\vspace*{-2.5cm}
\caption{Pressure in neutron (red curve) and baryon-lepton (green curve) matter from the Idaho DBHF calculation. The
shaded area corresponds to the region
of pressure consistent with flow data and the inclusion of strong
density dependence in the asymmetry terms.$^{56}$}
\label{pnm}
\end{center}
\end{figure}
\subsubsection{The symmetry energy and related observables}
Here we will focus specifically on the symmetry energy and its impact on
the structure of neutron-rich nuclei, in particular the neutron skin thickness.
In Fig.~\ref{esym}, we display the Idaho DBHF prediction for the symmetry energy by the solid red curve.
The curve is seen to grow at a lesser rate with increasing density,
an indication that, at large density,
repulsion in the symmetric matter EoS increases more rapidly than
in the neutron matter EoS.
This can be understood in terms of increased repulsion in isospin zero partial waves (absent
from neutron matter) as a function of density.
Our predicted value for the symmetry pressure, $L$, (see Eq.~(\ref{L}) below), is 69.6 MeV.
The various black dashed curves in Fig.~\ref{eos} are obtained with the simple parametrization
\begin{equation}
e_{sym}=C(\rho/\rho_0)^{\gamma} \, ,
\label{esymm}
\end{equation}
with $\gamma$ increasing from 0.7 to 1.0 in steps of 0.1, and $C \approx 32$ MeV.
It seems that a value of $\gamma$ close to 0.8 gives a reasonable description of the DBHF predictions,
although the use of different functions in different density regions would be best for an
optimal fit.
Considering that all of the dashed curves are
commonly used parametrizations
suggested by heavy-ion data,\cite{BA}
Fig.~\ref{eos} clearly reflects our limited knowledge of the symmetry energy,
particularly, but not exclusively, at the larger densities.
\begin{figure}
\begin{center}
\vspace*{-1.0cm}
\hspace*{-2.0cm}
\scalebox{0.4}{\includegraphics{ijmpe8.ps}}
\vspace*{-3.0cm}
\caption{DBHF prediction for the symmetry energy (solid red) compared with various
phenomenological parametrizations (dashed black). See text for details.
}
\label{esym}
\end{center}
\end{figure}
As already mentioned in Sec.~2,
from the experimental side intense effort is going on to obtain reliable empirical information for the less
known aspects of the EoS. Heavy-ion reactions are a popular way to seek constraints on the symmetry
energy, through analyses of observables that are sensitive to the pressure gradient between
nuclear and neutron matter.
Isospin diffusion data from heavy-ion collisions, together with analyses based on isospin-dependent transport
models, can provide information on the slope of the symmetry energy.
Concerning the lower densities,
isospin-sensitive observables can also be identified among the properties of normal nuclei.
The neutron skin of neutron-rich nuclei is a powerful isovector observable, being sensitive to the
slope of the symmetry energy, which determines to which extent neutrons will tend to spread outwards
to form the skin.
It is the purpose of this section to systematically examine and discuss the symmetry energy properties
in microscopic models and the corresponding neutron skin predictions. We will take the skin of
$^{208}$Pb as our representative isovector ``observable".
Parity-violating electron scattering experiments are now a realistic option
to determine neutron distributions with unprecedented accuracy. The neutron radius of
$^{208}$Pb is expected to be measured within 0.05 fm thanks to the electroweak program
at the Jefferson Laboratory.\cite{Piek06} This level of accuracy could not be achieved with hadronic scattering.
Parity-violating electron scattering at low momentum transfer is especially suitable to probe neutron densities, as the
$Z^0$ boson couples primarily to neutrons.
With the success of this program,
reliable empirical information on neutron skins will be able to provide, in turn, much needed {\it independent} constraint on the
density dependence of the symmetry energy.
We form an energy functional based on the semi-empirical mass formula, where the volume and
symmetry terms are contained in the isospin-asymmetric equation of state. Thus, we write the
energy of a (spherical) nucleus as
\begin{eqnarray}
E(Z,A) = \int d^3 r~ e(\rho(r),\alpha(r))\rho(r) + \nonumber \\
+ \int d^3 r f_0(|\nabla \rho(r)|^2 +
\beta|\nabla \rho_I(r)|^2) + Coulomb \, \, term \, .
\label{edrop}
\end{eqnarray}
In the above equation,
$\rho$ and $\rho_I$ are the usual isoscalar and isovector densities, given by $\rho_n +\rho_p$ and
$\rho_n -\rho_p$, respectively, in terms of neutron and proton densities. $\alpha$ is the neutron asymmetry
parameter, $\alpha=\rho_I/\rho$, and $e(\rho,\alpha)$ is the energy per particle in
isospin-asymmetric nuclear matter.
The latest Idaho EoS\cite{Sam0806} will be used in Eq.~(\ref{edrop}), along with others from different models.
From fits to nuclear binding energies,
the constant $f_0$ in Eq.~(\ref{edrop}) is approximately 70 $MeV$ $fm^5$, whereas the contribution of the term proportional to $\beta$
was found to be minor.\cite{Oya98} Thus it is reasonable to neglect it.
(The magnitude of $\beta$ was estimated to be about 1/4 in Ref.\cite{Furn}, where it was observed that,
even with variations of $\beta$ between -1 and +1, the effect of the $\beta$ term on the neutron skin
was negligibly small.)
The parameters of the proton and neutron densities are obtained by minimizing the value
of the energy functional, Eq.~(\ref{edrop}), assuming Thomas-Fermi distribution functions.
Although simple, this method has the advantage of allowing a
very direct connection between the EoS and the properties of finite nuclei. (It could be used, for
instance, to determine a semi-phenomenological EoS by fitting it to both binding energies and
charge radii of closed-shell nuclei.)
In Fig.~\ref{lead}, we show the proton and neutron Thomas-Fermi distributions for $^{208}$Pb as obtained with the method described above and the DBHF model for the
EoS. The predicted proton and neutron root-mean-square radii are
5.39 fm and 5.56 fm, respectively.
\begin{figure}[!t]
\centering
\includegraphics[totalheight=2.0in]{ijmpe9.ps}
\vspace*{0.2cm}
\caption{(color online) Neutron (red) and proton (black) point densities as obtained from the
DBHF model.
}
\label{lead}
\end{figure}
Typically,
predictions of the symmetry energy at saturation density encountered in the literature are in reasonable agreement with
one another, ranging approximately from 26 to 35 MeV. On the other hand,
the slope of the symmetry energy is more model dependent and less constrained.
This is seen through the symmetry pressure, defined as
\begin{equation}
L = 3 \rho_0 \Big (\frac{\partial e_{sym}(\rho)}{\partial \rho}\Big )_{\rho_0} \approx
3 \rho_0 \Big (\frac{\partial e_{n.m.}(\rho)}{\partial \rho}\Big )_{\rho_0} \, .
\label{L}
\end{equation}
Thus, $L$ is sensitive to the gradient of the energy per particle in neutron matter ($e_{n.m.}$).
As to be expected on physical grounds, the neutron skin, given by
\begin{equation}
S = \sqrt{<r_n^2>} - \sqrt{<r_p^2>} \, \, ,
\label{S}
\end{equation}
is highly sensitive to the same gradient.
Values of $L$ are reported to range
from -50 to 100 MeV as seen, for instance, through the numerous
parametrizations of Skyrme interactions (see Ref.\cite{BA} and references therein),
all chosen to fit the binding energies and the
charge radii of a large number of nuclei.
Heavy-ion data impose boundaries for $L$ at $85 \pm 25$ MeV,\cite{Chen07,Dan07} with more
stringent constraints being presently extracted.\cite{Tsang}
Other reports\cite{Ko09} state the constraints as
$85 \pm 55$ MeV.
Also, a nearly linear
correlation is observed between the neutron skin $S$ and the $L$
parameter.\cite{Li05}
Such phenomenological studies are extremely useful, but,
ultimately, they must be compared with {\it ab initio} approaches in order to get true
physical insight.
The direct connection with the underlying nuclear forces will then facilitate the physical understanding,
when combined with reliable constraints.
As explained in Sec.~3.2.1,
the DBHF model does not include three-body forces explicitely, but effectively incorporates the
class of TBF originating from the presence of nucleons and antinucleons (the
``Z-diagrams" in Fig.~\ref{3b}),
see previous discussion.
As the other main input of our comparison, we will take the EoS from the microscopic approach
of Ref.\cite{Catania3}. There (and in previous work by the same authors),
the Brueckner-Hartree-Fock (BHF) formalism is employed along with microscopic three-body forces.
In particular, in Ref.\cite{Catania4}
the meson-exchange TBF are constructed applying the same parameters
as used in the corresponding nucleon-nucleon (NN) potentials, which are: Argonne V18\cite{V18} (V18), Bonn B\cite{Mac89} (BOB), Nijmegen 93\cite{N93} (N93).
The popular (but phenomenological) Urbana TBF\cite{UIX} (UIX) is also utilized in Ref.\cite{Catania4}. Convenient
parametrizations in terms of simple analytic functions are given in all cases and we
will use those to generate the various EoS. We will refer to this approach, generally, as ``BHF + TBF".
This comparison will bring up the discussion on microscopic models and explicit TBF
outlined in Sec.~3.2.
In Fig.~\ref{esmicro}, we display Idaho DBHF predictions for the symmetry energy, solid black curve,
along with those from V18, BOB, UIX, and N93.
All values of the symmetry energy at the respective saturation densities are
between 29 and 34 MeV. A larger spreading is seen in the $L$ parameter,
see left panel of Fig.~\ref{SLK}, where the values range from about 70 to 106 MeV. The respective neutron skin predictions are shown on the vertical axis.
We notice that all BHF+TBF models predict larger $L$, and thus larger neutron skins, as compared with DBHF, corresponding to
a faster growth of the energy per particle in neutron matter relative to symmetric matter.
This can be seen from Fig.~\ref{esmicro}, especially for the higher
densities.
The present calculations reveal that there are more subtle, but significant differences at low to medium densities as well.
Some comments are in place concerning
the nature of the $L$ {\it vs.} $S$ correlation.
If a family
of models is constructed so that they differ {\it mostly} in the slope of the symmetry energy, a linear
correlation may be expected among $L$ and the skin.\cite{Li05} The models we are considering, however,
differ from one another in more than just the slope of the symmetry energy. Thus, although the general
pattern is that a larger $L$ corresponds to a larger skin, the relation is more complex, as the r.m.s.
radius of the neutron distribution will, to some extent, receive feedback from the smaller or larger
degree of attraction that binds neutrons and protons, and which depends on the symmetric matter EoS.
\begin{figure}[!t]
\centering
\includegraphics[totalheight=2.0in]{ijmpe10.ps}
\vspace*{-0.3cm}
\caption{(color online) Predictions for the symmetry energy from DBHF and various
``BHF + TBF" models discussed in the text.
}
\label{esmicro}
\end{figure}
\begin{figure}
\centering
\includegraphics[totalheight=2.5in]{ijmpe11.ps}
\vspace*{-0.3cm}
\caption{Left panel: neutron skin of $^{208}$Pb {\it vs.} the symmetry pressure for the models
considered in the text; Right panel:
neutron skin of $^{208}$Pb {\it vs.} the curvature of the symmetry energy, $K_{sym}$, for the same models.
}
\label{SLK}
\end{figure}
The next term in the expansion of the symmetry energy is the
$K_{sym}$ parameter,
\begin{equation}
K_{sym} = 9 \rho_0^2 \Big (\frac{\partial^2 e_{sym}(\rho)}{\partial \rho ^2}\Big )_{\rho_0} \, ,
\label{Ksym}
\end{equation}
which is a measure for the curvature of the symmetry energy. The neutron skin {\it vs.\ } $K_{sym}$
is shown in the right panel of Fig.~\ref{SLK} for the various models. Although the
values of $K_{sym}$ appear more spread out, the large negative values obtained with
some of the parametrizations of the Skyrme model are not present.
Those
large negative values (as low as -600 MeV) produced by Skyrme models indicate a strongly downward
curvature of the symmetry energy already at low to medium densities.
We also notice from Fig.~\ref{SLK} (right panel)
that no clear correlation can be identified between $K_{sym}$ and the neutron skin.
We conclude this section by showing in Fig.~\ref{RS} the relation between the neutron skin of $^{208}$Pb and the radius
of a 1.4M$_{\odot}$ neutron star for the models we are considering. (A more detailed discussion of neutron stars and the EoS will be presented in Sec.~4.) Stellar matter contains neutrons in $\beta$
equilibrium with protons, electrons, and muons.
Tabulated values for the latest Idaho DBHF EoS can be found later in this article. We have applied $\beta$-stability in the same way to all the
various models of
Ref.\cite{Catania4} starting from the given parametrized versions of the respective symmetric matter and neutron matter EoS.
For the case of leptons, typically treated as a gas of non-interacting fermions, the equations of $\beta$-stability
(which amount to imposing energy conservation and charge neutrality),
are elementary to solve (see Sec.~4.2).
At subnuclear densities, all the EoS considered here are joined with the crustal equations of state from the work of Harrison and Wheeler\cite{HW} (for energy densities between 10 and 10$^{11}$ g~cm$^{-3}$) and the work of Negele and
Vautherin\cite{NV} (for energy densities less than 1.7$\times$10$^{13}$g~cm$^{-3}$).
At first, Fig.~\ref{RS} can appear surprising, since it shows that
larger skin does not necessarily imply larger radius. On the other hand, one must keep in mind that,
as we argued previously, these
models differ from one another in more than just the slope of the symmetry energy. Furthermore, the star
radius probes higher densities than the skin does.
Generally, if a family
of models is constructed so that they differ {\it mostly} in the slope of the symmetry energy, a linear
correlation is expected among $L$, the skin, and $R$.\cite{Moust07}
\begin{figure}[!t]
\centering
\includegraphics[totalheight=2.5in]{ijmpe12.ps}
\vspace*{-0.2cm}
\caption{Left panel: neutron skin of $^{208}$Pb {\it vs.} the radius of a 1.4M$_{\odot}$ neutron star for the models
considered in the text. Right panel: star radius {\it vs.} the central energy density for the same models.
}
\label{RS}
\end{figure}
In summary:
Most models agree on the value of the symmetry energy around the saturation point within a few MeV,
but we are far from a reasonable agreement on its derivative
and even farther from agreement on its curvature.
Although microscopic models do not display as much spread as phenomenological ones,
a point that comes out clearly from the present study is that
a measurement of the neutron skin of $^{208}$Pb with an accuracy of 0.05 fm, as it has been announced,\cite{Hor} should definitely
be able to discriminate among the EoS from these microscopic models.
\subsubsection{Neutron and proton single-particle properties and the symmetry potential}
In this section we will concentrate specifically on predictions of
isovector single-particle properties, such as the neutron and proton single-particle
potentials and the closely related symmetry potential. These ``observables"
depend
sensitively on the difference between neutron and proton properties in asymmetric
matter and play an important role in simulations of heavy-ion collisions
with neutron-rich nuclei.
In terms of the $G$-matrix, calculated as described in Sec.~3.2.1,
we write the single-nucleon potential
(in the case of unequal Fermi levels for protons
and neutrons), as
\begin{equation}
U_i(k) = Re[\sum_{q<k_F^n} <kq|G_{in}|kq-qk>
+ \sum_{q<k_F^p} <kq|G_{ip}|kq-qk>] \, ,
\label{Uik}
\end{equation}
where $i$ = $n/p$ for neutron/proton, and
$k$ refers to states below and above the Fermi momentum (consistent with the ``continuous choice" for the
single-particle spectrum).
\begin{figure}[!t]
\centering
\vspace*{-1.0cm}
\includegraphics[totalheight=3.5in]{ijmpe13.ps}
\vspace*{-2.2cm}
\caption{ The neutron (dashed line) and proton (dotted) single-particle potentials as a function of the
momentum at fixed average density and
$\alpha$=0.4. The solid curve in the middle is the single-nucleon potential at $\alpha$=0.
}
\label{unp}
\end{figure}
We begin by examining the momentum dependence of
$U_{n/p}$, the single neutron/proton potential in neutron-rich matter.
In Fig.~\ref{unp}, we show
$U_{n/p}$ as a
function of the momentum and a fixed value of the asymmetry parameter,
$\alpha=$ 0.4. The middle curve shows $U_n$=$U_p$ at $\alpha=$0.
The total nucleon density considered in the figure is equal to 0.185 fm$^{-3}$ and
corresponds to a Fermi momentum of 1.4 fm$^{-1}$, which is very close to our predicted
saturation density.
For increasing values of
$\alpha$ (obviously, $U_n=U_p$ for $\alpha$=0), the proton potential becomes more attractive while the opposite
tendency is observed in $U_n$. This reflects the fact that the proton-neutron interaction, the
one predominantly felt by the single proton as the proton density is depleted, is more
attractive than the one between identical nucleons.
Also, as it appears reasonable, the dependence on $\alpha$ becomes weaker at larger momenta.
The role of the momentum dependence of the symmetry potential in heavy-ion collisions was
examined and found to be important. Symmetry potentials with and without
momentum dependence can lead to significantly
different predictions of collision observables.\cite{DAS04} More recently, these issues have been
revisited in hot asymmetric matter.\cite{Moust207,Moust08}
Regarding
$U_{n/p}$ as functions of
the asymmetry parameter $\alpha$, one can easily
verify that the following approximate relation applies:
\begin{equation}
U_{n/p}(k,k_F,\alpha) \approx U_{n/p}(k,k_F,\alpha=0) \pm U_{sym}(k,k_F)\alpha \, ,
\label{Unp}
\end{equation}
with the $\pm$ referring to neutron/proton, respectively.
Figure~\ref{ualpha} displays the left-hand side of Eq.~(\ref{Unp}) for fixed density and nucleon momentum and
clearly reveals the linear
behaviour of
$U_{n/p}$ as a function of $\alpha$. Thus,
one can expect isospin splitting of the single-particle potentials to be effective in separating
the collision dynamics of neutrons and protons.
Although the main focus of Fig.~\ref{ualpha} is the $\alpha$ dependence,
predictions are displayed for three different potentials, Bonn A,\cite{Mac89} B,\cite{Mac89} and C.\cite{Mac89}
These three models differ mainly in the strength of the tensor force, which
is mostly carried by partial waves with isospin equal to 0 and thus should fade away
in the single-neutron potential
as the neutron fraction increases. In fact, the figure demonstrates reduced differences among the values
of $U_n$ predicted with the three potentials at large $\alpha$.
\begin{figure}[!t]
\centering
\vspace*{1.0cm}
\includegraphics[totalheight=3.0in]{ijmpe14.ps}
\vspace*{0.2cm}
\caption{ The neutron and proton single-particle potentials as a function of the
asymmetry parameter at fixed average density and momentum equal to the average
Fermi momentum.
}
\label{ualpha}
\end{figure}
Already several decades ago, it was pointed out that the real part of the nuclear
optical potential depends on the asymmetry parameter as in Eq.~(\ref{Unp}).\cite{Lane}
Then,
the quantity
\begin{equation}
\frac{U_{n} + U_p}{2} = U_0 ,
\label{U0}
\end{equation}
which is obviously the single-nucleon potential in absence of asymmetry,
should be a reasonable approximation to the isoscalar part of the optical
potential in the interior of a nucleus. The momentum dependence of $U_0$ (which is shown in Fig.~\ref{unp} as the
$\alpha$=0 curve), is important for extracting information about the symmetric matter EoS
and is reasonably agreed upon
.\cite{u02,u03,u04,u05,u06,u07,u08,u09}
On the other hand,
\begin{equation}
\frac{U_{n} - U_p}{2\alpha} = U_{sym}
\label{Usym}
\end{equation}
should be comparable with the Lane potential,\cite{Lane} or the isovector
part of the nuclear optical potential. (In the two equations
above the dependence upon density, momentum, and asymmetry has been suppressed for
simplicity.)
We have calculated $U_{sym}$ as a function of
the momentum, or rather the corresponding kinetic energy.
The predictions obtained with Bonn B are shown in Fig.~\ref{usym}.
They are compared with the phenomenological expression\cite{Lane}
\begin{equation}
U_{Lane} = a -b T \, ,
\label{Lane}
\end{equation}
where $T$ is the kinetic energy, $a \approx 22-34 MeV$, $b\approx 0.1-0.2 MeV$.
We observe that the strength of the (DBHF) predicted symmetry potential decreases
with energy, a behavior which is consistent with the empirical information.
The same comparison is done in
Ref.\cite{BAL04} starting from a
phenomenological formalism for the single-nucleon potential.\cite{Bomb01,Rizzo}
There, it is shown that it is
possible to choose two sets of parameters which lead to similar values of the
symmetry energy but exactly opposite tendencies in the energy dependence of the symmetry
potential as well as
opposite sign of the proton-neutron mass splitting.
As a consequence of that, these two sets of parameters lead to very different
predictions for observables in heavy-ion collisions induced by neutron-rich nuclei.\cite{Rizzo}
This fact suggests that
constraints from ``differential" observables, namely those specifically sensitive to the difference
between proton and neutron properties in asymmetric matter, are very much needed.
\begin{figure}[!t]
\centering
\vspace*{-2.5cm}
\includegraphics[totalheight=5.5in]{ijmpe15.ps}
\vspace*{-4.5cm}
\caption{ The symmetry potential close to saturation density.
}
\label{usym}
\end{figure}
\begin{figure}[!t]
\centering
\vspace*{3.0cm}
\includegraphics[totalheight=2.0in]{ijmpe16.ps}
\vspace*{-2.0cm}
\caption{ The neutron and proton effective mass as a function of the
asymmetry parameter at fixed average density.
}
\label{mnp}
\end{figure}
The effective masses for proton and neutron corresponding to the single-nucleon potentials
of Fig.~\ref{unp}
are shown in Fig.~\ref{mnp} as a function
of $\alpha$.
The predicted effective mass of the neutron being larger than the
proton's is a trend shared with
microscopic non-relativistic
calculations.\cite{BL91} In the non-relativistic case, one can show from elementary arguments
based on the curvature of the (parabolic) single-particle potential that a more attractive
potential, as the one of the proton, leads to a smaller effective mass.
In the DBHF quasi-particle approximation described in Sec.~3.2.1, one
assumes momentum-independent nucleon self-energies, $U_S$ and $U_V$, with a vanishing
spacial component of the vector part. In such limit, the ``Dirac mass" and the non-relativistic or ``Landau mass"
should coincide to leading order in the expansion of the relativistic single-particle potential.\cite{Fuchs}
It is therefore not surprising that
our proton and neutron effective masses, although defined as Dirac masses ($m^*_i = m+U_{S,i}$), display
a trend that is similar to the one generally expected in the non-relativistic case. Most likely, the main reason for the
discrepancy in the predicted isospin mass splitting between the Idaho DBHF calculation and the DBHF model from Ref.\cite{Fuchs}
is in the way the scalar and vector fields are treated, particularly their momentum dependence.
Closely related to the effective mass is the concept of in-medium effective cross section.
Transport equations, such as the Boltzmann-Uehling-Uhlenbeck equation,
describe the evolution of a non-equilibrium gas of
strongly interacting hadrons drifting
in the presence of the mean field while undergoing
two-body collisions. Thus heavy ion collision simulations require the
knowledge of in-medium two-body cross sections as well as the mean field.
Both should be calculated microscopically and
self-consistently starting from the basic nuclear forces\cite{sig2}
and are closely related to the EoS.
\begin{table}[pt]
\tbl
{$np$ total effective cross sections in symmetric matter calculated with
various many-body models, see text for details.
}
{\begin{tabular}{@{}ccccc@{}} \toprule
$k_{F}(fm^{-1})$ & $q_0(MeV)$ & $\sigma _{np}^{DBHF}(mb)$ &
$\sigma _{np}^{BHF+TBF}(mb)$ & $\sigma _{np}^{BHF}(mb)$
\\ \colrule
1.1 & 250 & 34.38& 44.65 &51.74 \\
1.1 & 300 & 23.14& 29.01 &31.85 \\
1.1 & 350 & 20.63& 23.56 &25.62 \\
1.4 & 250 & 26.74& 31.25 &39.82 \\
1.4 & 300 & 17.26& 25.28 &30.96 \\
1.4 & 350 & 16.77& 21.17 &25.23 \\
1.7 & 250 & 17.20& 19.03 &29.12 \\
1.7 & 300 & 15.06& 17.59 &25.02 \\
1.7 & 350 & 12.33& 13.99 &21.04 \\
\botrule
\end{tabular}}
\end{table}
\begin{table}[pt]
\tbl
{As in the previous Table, for identical particles.}
{\begin{tabular}{@{}ccccc@{}} \toprule
$k_{F}(fm^{-1})$ & $q_0(MeV)$ & $\sigma _{NN}^{DBHF}(mb)$ &
$\sigma _{NN}^{BHF+TBF}(mb)$ & $\sigma _{NN}^{BHF}(mb)$ \\
\colrule
1.1 & 250 & 18.00& 18.15 &22.98 \\
1.1 & 300 & 16.41& 17.47 &19.48 \\
1.1 & 350 & 17.08& 18.55 &19.51 \\
1.4 & 250 & 15.72& 13.74 &19.76 \\
1.4 & 300 & 13.70& 14.43 &17.23 \\
1.4 & 350 & 16.31& 16.89 &19.46 \\
1.7 & 250 & 18.05& 7.87 &13.36 \\
1.7 & 300 & 17.93& 9.98 &13.98 \\
1.7 & 350 & 13.96& 11.48 &16.43 \\
\botrule
\end{tabular}}
\end{table}
To be fully applicable throughout the evolution of the colliding system, where high densities may be involved,
in-medium cross sections must go beyond the conventional BHF.
Although an extensive discussion of in-medium cross sections will not be presented here,
we show in Tables 1 and 2 a comparison of the latest Idaho (DBHF) in-medium cross sections and those
obtained from the BHF+TBF model of Lombardo {\it et al.}.\cite{sig2}
We recall that the relativistic elastic differential cross section is calculated from
\begin{equation}
\sigma(\theta) = \frac{(m^*)^4}{4 \pi ^2 \hbar ^4 s^*} |G(\theta)|^2 \, ,
\label{sigrel}
\end{equation}
where $s^* = 4((m^*)^2 + {\vec q}^2)$.
Notice that the relativistic amplitude, $G$, is related to the non-relativistic one through
$G = \frac{E^*}{m^*}G_{n.r.}$, with
$E^* = ((m^*)^2 + {\vec q}^2)^{1/2}$.
In Tables 1 and 2, $q_0$ signifies the momentum of either nucleon in their center-of-mass frame, for simplicity assumed to coincide with
the nuclear matter rest frame. The 4th and 5th columns show the results obtained by Lombardo {\it et al.}
with BHF calculations implemented or not with TBF. There is generally a fair amount of agreement between
the DBHF and the BHF+TBF results. For
the $np$ case, energy and density dependence appear consistent among the two sets of results displayed
in columns 3 and 4, although the BHF+TBF results are generally larger. For the case of identical nucleons,
the DBHF values are in good agreement with those from BHF+TBF, with the exception of the
highest densities.
This is to be expected in view of the larger differences existing between the DBHF and the BHF+TBF predictions of the EoS at high density.
Besides being a crucial part of the input for transport models, in-medium effective
cross sections are important in their own right as they allow to establish an immediate connection
with the nucleon mean free path, $\lambda$, one of the most fundamental properties characterizing the
propagation of nucleons through matter.\cite{mfp} It can be written as
\begin{equation}
\lambda = \frac{1}{\sigma _{pp} \rho _p + \sigma _{pn} \rho _n} \, ,
\label{lambda}
\end{equation}
for a proton propagating through matter with proton and neutron densities equal to $\rho _p$ and
$\rho _n$, respectively.
The mean free path enters the calculation of the nuclear
transparency function, which is the probability that the projectile will pass through the target without interacting.
This is closely related to the total reaction
cross section of a nucleus, which can be used to extract
nuclear r.m.s. radii within Glauber models.\cite{Glauber} Therefore, microscopic in-medium
isospin-dependent NN cross sections can ultimately help obtain
information about the properties of exotic, neutron-rich nuclei, such as, for instance, the target density.
The latter is a very useful information, especially
for exotic nuclei or nuclei with large neutron skins (where densities cannot be probed with conventional
charged lepton scattering).
\section{Neutron Star Properties}
Proceeding in our review of the various systems and phenomena where
the EoS plays a chief role, we will now take a look at
some of the most exotic systems in the universe, compact stars.
\subsection{A brief review of neutron star structure and available constraints}
Neutron stars are stable configurations containing the most dense form of matter found in the universe.
They are therefore unique laboratories to study the properties of highly compressed (cold) matter.
Furthermore, the possibility of exploring the structure of neutron stars via gravitational waves\cite{GW}
makes these exotic objects even more exciting.
The densities found in neutron stars range from the density of iron to several times normal nuclear density.
In the low-density region, matter is highly compressed and fully ionized and consists of electrons and ions of iron.
As density increases,
charge neutrality requires matter to become more neutron rich. In this density range
($10^7 \leq \rho \leq 10^{11}$~g~cm$^{-3}$), neutron-rich nuclei appear, such as isotopes of
$^{62}$Ni,
$^{86}$Kr,
$^{84}$Se,
$^{80}$Zn,
$^{124}$Mo,
$^{122}$Zr.
Above densities of approximately $ 10^{11}$~g~cm$^{-3}$, free neutrons begin to form a continuum of states.
The inner crust is then a compressed solid with a fluid of neutrons.
At densities equal to approximately 1/2 of saturation density, clusters begin to merge into a continuum. In this phase, matter is a uniform fluid
of neutrons, protons, and leptons. Protons and neutrons can condense in superfluids and superconducting states
below temperatures of about 10$^9$~K.\cite{Sedr06}
Above a few times nuclear matter density, the actual composition of stellar matter is not known.
Strange baryons can appear when the nucleon chemical potential is of the order of their rest mass.
We will address the issue of strange matter later in this article.
At even higher densities, transitions to other phases are speculated, such as a deconfined, rather than hadronic, phase. The critical density for
such transition cannot be predicted reliably because it lies in a range where QCD in non perturbative.\cite{Sedr06}
Recently, a semi-algebraic expression for the EoS of cold quark matter has been derived within the
Dyson-Schwinger formalism.\cite{Klaen09}
The Dyson-Schwinger framework can address both confinement and dynamical chiral symmetry breaking,\cite{Klaen09}
unlike other models commonly used in conjunction with cold quark matter, such as the Nambu-Jona-Lasinio model or
bag-type models. It will be interesting to see applications of the model described in Ref.\cite{Klaen09}
to dense astrophysical systems. For that purpose, the quark matter EoS will need to be supplemented by
the nuclear matter EoS. In the meantime, it is encouraging that the model predicts coincident
deconfinement and chiral symmetry restoring transitions.
The possibility has been speculated that the most stable state at zero pressure may be
$u$, $d$, $s$ quark matter instead of iron. This would imply that strange quark matter is the most
stable (in fact, the absolutely stable) state of strongly interacting matter, as originally proposed
by Bodmer,\cite{Bodmer} Witten,\cite{Witten} and Terazawa.\cite{Teraz}
In such case, hyperonic and hybrid stars would have to be metastable with respect to stars composed
of stable three-flavor strange quark matter,\cite{Weber} which is lower in energy
than two-flavor quark matter due to the extra Fermi levels open to strange quarks.
Whether or not strange quark stars can give rise to pulsar glitches, (which are observed sudden small changes
in the rotational frequency of a pulsar),
may be a decisive test of the strange quark matter hypothesis.\cite{Weber}
The maximum mass and the radius of a neutron star are sensitive to different aspects of the EoS. The maximum mass
is mostly determined by the stiffness of the EoS at densities greater than a few times saturation density.
Non-nucleonic degrees of freedom, which typically make their appearance at those densities, are
known to have a considerable impact on the maximum mass of the star.
Causality imposes a limit of about 3 solar masses to the maximum mass, the existence of which is inherent to general relativity
and is predicted by the equation of hydrostatic equilibrium,
\begin{equation}
\frac{d P(r)}{dr} = -\frac{G}{c^2}\frac{(P(r)+\epsilon(r))(M(r)+4\pi r^3 P(r)/c^2)}{r(r-2GM(r)/c^2)} \; ,
\label{GR1}
\end{equation}
with
\begin{equation}
\frac{d M(r)}{dr} = 4\pi ^2 \rho(r) \, ,
\label{GR2}
\end{equation}
where $\epsilon$ is the total mass-energy density.
The pressure is related to the energy/particle through
\begin{equation}
P(\rho) = \rho ^2 \frac{\partial e(\rho)}{\partial \rho} \, .
\label{Pr}
\end{equation}
It's worth recalling that no mass limit exists in Newtonian gravitation.
The star radius is mainly sensitive to the slope of the symmetry energy. In particular, it is closely connected
to the internal pressure (that is, the energy gradient) of matter at
densities between about 1.5$\rho_0$ and
2-3$\rho_0$.\cite{Latt07}
Lattimer and Prakash investigated the maximum central density question.\cite{Latt07}
Combining the causality limit ($R\geq$2.87$GM/c^2$) with the central density-mass relation implied by the
Tolman VII solution,\cite{Tolman}
\begin{equation}
\rho= \rho _{c} (1 - r^2/R^2) \, ,
\label{Tol}
\end{equation}
they obtained the relation
\begin{equation}
\rho _{c} M^2 \leq 15.3 \times 10^{15} M_{\odot}^2 \, g\,cm^{-3} \, .
\label{ecmax}
\end{equation}
The maximum gravitational mass and the corresponding radius are the typical observables
used to constraint the EoS.
Much more stringent constraints could be imposed on EoS if these two quantities could be
determined independently from each other. At this time, the heaviest neutron star (with
accurately known mass) has a mass of 1.671 $\pm$ 0.008 $M_{\odot}$,\cite{Champ08} but neutron stars
with masses above 2 solar masses are expected to exist.\cite{Fre09}
Mass limits are obtained from observations of binary systems, either a two-neutron star system
or a neutron star and a massive companion, such as a white dwarf. The pulsar in the Hulse-Taylor
binary system\cite{Fre09} has a mass of
1.4408$\pm$ 0.0003 $M_{\odot}$, to date the best mass determination.
Measurements of the radius are considerably less precise than mass measurements.\cite{Latt07}
The radiation or photospheric radius, $R_{\infty}$, is related to the actual stellar radius by
$R_{\infty} = R(1-R/R_S)^{-1/2}$, where $R_S=2GM/R$ is the Schwarzschild radius.
Estimates are usually based on
thermal emission of cooling stars, including redshifts,
and the properties of sources with bursts or thermonuclear explosions at the surface.
Neutrinos are also used as a probe of the EoS.
Neutrino luminosity is controlled by several factors including the total mass of the (proton-neutron)
star and the opacity of neutrinos at high densities, which of course is sensitive to the
EoS of dense matter.
Gravitational waves\cite{GW} are a less conventional way to probe neutron star properties.
Emission of gravitational waves causes orbital decay, the observation of which would allow
an estimate of the moment of inertia and thus, together with an accurate measurement of the
mass, impose stronger boundaries on the EoS.
A measurement of the moment of inertia within 10\%, together with the information on the mass, would
be able to discriminate among various EoS.\cite{Latt07}
To date, the best determination of the moment of inertia is the one for the Crab pulsar\cite{Crab}
which would rule out only very soft EoS.\cite{Latt07}
The minumum mass of a neutron star is also a parameter of interest. For a cold, stable system,
the minimum mass is about
0.09 $M_{\odot}$.\cite{Latt07}
The smallest reliably estimated neutron star mass is the companion of the binary pulsar
J1756-2251, which has a mass of
1.18$\pm$ 0.02 $M_{\odot}$.\cite{Faulk}
\subsection{Comparing predictions of microscopic models}
As constraints from nuclear physics and/or astrophysics promise to become more stringent, it is important to understand and compare how the nature of the various predictions
is related to the features of each model. In microscopic approaches, the
tight connection with the underlying forces facilitates physical interpretation of the predictions
in terms of the characteristics of the nuclear force and its behavior in the medium.
Motivated by these considerations, in this section
we calculate several neutron star properties, for static and/or rotating stars, using
equations of state based on different microscopic models.
These will be the same as those used in Sec.~3.2.3
to examine symmetry energy and neutron skin predictions, namely BOB, V18, N93, and UIX,
along with Idaho DBHF.
In Sec.~3.2.3,
we compared the
predictions of the neutron skin in $^{208}$Pb by these models and correlated them
with differences in the slope of the symmetry energy.
Model differences become larger at high-density and will naturally impact neutron star predictions.
\begin{table}[pt]
\tbl{DBHF equation of state of pure neutron matter.}
{\begin{tabular}{@{}ccc@{}} \toprule
Baryon density(1/$cm^3$) & Energy density($g/cm^3$) & Pressure($dyne/cm^2$) \\
\colrule
0.145902E+38 & 0.245257E+14 & 0.434992E+32 \\
0.231688E+38 & 0.389881E+14 & 0.874118E+32 \\
0.345843E+38 & 0.582634E+14 & 0.156362E+33 \\
0.492421E+38 & 0.830558E+14 & 0.257664E+33 \\
0.675475E+38 & 0.114075E+15 & 0.468635E+33 \\
0.781946E+38 & 0.132151E+15 & 0.621198E+33 \\
0.899057E+38 & 0.152061E+15 & 0.816935E+33 \\
0.102731E+39 & 0.173905E+15 & 0.111487E+34 \\
0.116722E+39 & 0.197791E+15 & 0.156721E+34 \\
0.131929E+39 & 0.223833E+15 & 0.223210E+34 \\
0.148402E+39 & 0.252158E+15 & 0.319807E+34 \\
0.166192E+39 & 0.282905E+15 & 0.463315E+34 \\
0.185350E+39 & 0.316247E+15 & 0.681545E+34 \\
0.205927E+39 & 0.352381E+15 & 0.993009E+34 \\
0.227973E+39 & 0.391534E+15 & 0.141505E+35 \\
0.251538E+39 & 0.433969E+15 & 0.203050E+35 \\
0.276674E+39 & 0.480078E+15 & 0.293314E+35 \\
0.303432E+39 & 0.530247E+15 & 0.409031E+35 \\
0.331861E+39 & 0.584918E+15 & 0.558832E+35 \\
0.362012E+39 & 0.644589E+15 & 0.746722E+35 \\
0.393937E+39 & 0.709762E+15 & 0.965629E+35 \\
0.427685E+39 & 0.780907E+15 & 0.122694E+36 \\
0.463308E+39 & 0.858621E+15 & 0.152765E+36 \\
0.500856E+39 & 0.943372E+15 & 0.185795E+36 \\
0.540380E+39 & 0.103570E+16 & 0.223317E+36 \\
0.581930E+39 & 0.113630E+16 & 0.268106E+36 \\
0.625557E+39 & 0.124615E+16 & 0.324163E+36 \\
0.671312E+39 & 0.136650E+16 & 0.395072E+36 \\
0.719245E+39 & 0.149890E+16 & 0.485417E+36 \\
0.769408E+39 & 0.164502E+16 & 0.589746E+36 \\
0.821850E+39 & 0.180610E+16 & 0.707244E+36 \\
0.876622E+39 & 0.198388E+16 & 0.846053E+36 \\
0.933776E+39 & 0.218030E+16 & 0.100924E+37 \\
0.993361E+39 & 0.239765E+16 & 0.120127E+37 \\
0.105543E+40 & 0.263842E+16 & 0.142307E+37 \\
0.112003E+40 & 0.290521E+16 & 0.167838E+37 \\
0.118721E+40 & 0.320102E+16 & 0.197388E+37 \\
0.125703E+40 & 0.352921E+16 & 0.231442E+37 \\
0.132954E+40 & 0.389345E+16 & 0.270588E+37 \\
0.140478E+40 & 0.429785E+16 & 0.315475E+37 \\
0.148280E+40 & 0.474691E+16 & 0.366826E+37 \\
0.156366E+40 & 0.524560E+16 & 0.425495E+37 \\
0.164741E+40 & 0.579940E+16 & 0.492059E+37 \\
0.173410E+40 & 0.641418E+16 & 0.567978E+37 \\
0.182378E+40 & 0.709679E+16 & 0.654219E+37 \\
\botrule
\end{tabular}}
\end{table}
\begin{table}[pt]
\tbl{As in the previous Table, for $\beta$-equilibrated matter.}
{\begin{tabular}{@{}ccc@{}} \toprule
Baryon density(1/$cm^3$) & Energy density($g/cm^3$) & Pressure($dyne/cm^2$) \\
\colrule
0.145902E+38 & 0.245236E+14 & 0.416322E+32 \\
0.231688E+38 & 0.389827E+14 & 0.822804E+32 \\
0.345843E+38 & 0.582501E+14 & 0.142104E+33 \\
0.492421E+38 & 0.830269E+14 & 0.230061E+33 \\
0.675475E+38 & 0.114019E+15 & 0.409814E+33 \\
0.781946E+38 & 0.132073E+15 & 0.540126E+33 \\
0.899057E+38 & 0.151955E+15 & 0.711724E+33 \\
0.102731E+39 & 0.173765E+15 & 0.970498E+33 \\
0.116722E+39 & 0.197605E+15 & 0.137046E+34 \\
0.131929E+39 & 0.223589E+15 & 0.201627E+34 \\
0.148402E+39 & 0.251837E+15 & 0.261928E+34 \\
0.166192E+39 & 0.282481E+15 & 0.330297E+34 \\
0.185350E+39 & 0.315683E+15 & 0.517205E+34 \\
0.205927E+39 & 0.351619E+15 & 0.757028E+34 \\
0.227973E+39 & 0.390495E+15 & 0.110037E+35 \\
0.251538E+39 & 0.432555E+15 & 0.160843E+35 \\
0.276674E+39 & 0.478138E+15 & 0.235092E+35 \\
0.303432E+39 & 0.527601E+15 & 0.334557E+35 \\
0.331861E+39 & 0.581355E+15 & 0.465834E+35 \\
0.362012E+39 & 0.639861E+15 & 0.632143E+35 \\
0.393937E+39 & 0.703594E+15 & 0.832228E+35 \\
0.427685E+39 & 0.773030E+15 & 0.107139E+36 \\
0.463308E+39 & 0.848703E+15 & 0.134775E+36 \\
0.500856E+39 & 0.931088E+15 & 0.165606E+36 \\
0.540380E+39 & 0.102069E+16 & 0.200271E+36 \\
0.581930E+39 & 0.111814E+16 & 0.241417E+36 \\
0.625557E+39 & 0.122434E+16 & 0.293125E+36 \\
0.671312E+39 & 0.134047E+16 & 0.358691E+36 \\
0.719245E+39 & 0.146797E+16 & 0.442173E+36 \\
0.769408E+39 & 0.160841E+16 & 0.539098E+36 \\
0.821850E+39 & 0.176298E+16 & 0.649283E+36 \\
0.876622E+39 & 0.193337E+16 & 0.779667E+36 \\
0.933776E+39 & 0.212141E+16 & 0.932987E+36 \\
0.993361E+39 & 0.232930E+16 & 0.111395E+37 \\
0.105543E+40 & 0.255941E+16 & 0.132452E+37 \\
0.112003E+40 & 0.281437E+16 & 0.156916E+37 \\
0.118721E+40 & 0.309721E+16 & 0.185697E+37 \\
0.125703E+40 & 0.341144E+16 & 0.219046E+37 \\
0.132954E+40 & 0.376060E+16 & 0.257227E+37 \\
0.140478E+40 & 0.414881E+16 & 0.301377E+37 \\
0.148280E+40 & 0.458067E+16 & 0.352202E+37 \\
0.156366E+40 & 0.506130E+16 & 0.410655E+37 \\
0.164741E+40 & 0.559636E+16 & 0.477543E+37 \\
0.173410E+40 & 0.619202E+16 & 0.554149E+37 \\
0.182378E+40 & 0.685522E+16 & 0.641233E+37 \\
\botrule
\end{tabular}}
\end{table}
\begin{table}[pt]
\tbl{Equation of state used for the crust. See text for details.}
{\begin{tabular}{@{}ccc@{}} \toprule
Baryon density(1/$cm^3$) & Energy density($g/cm^3$) & Pressure($dyne/cm^2$) \\
\colrule
0.59701000E+25 & 0.99998600E+01 & 0.40721410E+12\\
0.19099900E+26 & 0.31992190E+02 & 0.47174340E+14\\
0.38376300E+26 & 0.64279760E+02 & 0.33550870E+15\\
0.15492600E+27 & 0.25949850E+03 & 0.88776630E+16\\
0.31128200E+27 & 0.52139300E+03 & 0.38114990E+17\\
0.62543700E+27 & 0.10476010E+04 & 0.15215920E+18\\
0.25249000E+28 & 0.42291870E+04 & 0.20992040E+19\\
0.50731100E+28 & 0.84974150E+04 & 0.74261220E+19\\
0.20480200E+29 & 0.34304300E+05 & 0.87378890E+20\\
0.41149300E+29 & 0.68925200E+05 & 0.29327530E+21\\
0.13164600E+30 & 0.22050990E+06 & 0.21592540E+22\\
0.26450500E+30 & 0.44305570E+06 & 0.70850350E+22\\
0.53144400E+30 & 0.89020260E+06 & 0.23123900E+23\\
0.21453100E+31 & 0.35937570E+07 & 0.19523170E+24\\
0.43102100E+31 & 0.72207080E+07 & 0.54232090E+24\\
0.13787700E+32 & 0.23100980E+08 & 0.28166750E+25\\
0.34951600E+32 & 0.58569360E+08 & 0.10057550E+26\\
0.70215700E+32 & 0.11767960E+09 & 0.25560490E+26\\
0.28334500E+33 & 0.47507410E+09 & 0.15831900E+27\\
0.56915000E+33 & 0.95453350E+09 & 0.38815460E+27\\
0.22959600E+34 & 0.38534730E+10 & 0.22901680E+28\\
0.46107400E+34 & 0.77425110E+10 & 0.55287600E+28\\
0.14734400E+35 & 0.24770450E+11 & 0.23888140E+29\\
0.29580200E+35 & 0.49769610E+11 & 0.57354730E+29\\
0.59374600E+35 & 0.99998600E+11 & 0.13755860E+30\\
0.30685800E+36 & 0.51813430E+12 & 0.66738040E+30\\
0.60373600E+36 & 0.10203670E+13 & 0.12773240E+31\\
0.10985200E+37 & 0.18580330E+13 & 0.20087290E+31\\
0.15931600E+37 & 0.26960450E+13 & 0.26615840E+31\\
0.19888300E+37 & 0.33666280E+13 & 0.32106350E+31\\
0.23844100E+37 & 0.40373370E+13 & 0.38075420E+31\\
0.26810600E+37 & 0.45404400E+13 & 0.42918160E+31\\
0.29776600E+37 & 0.50436140E+13 & 0.48089980E+31\\
0.32742200E+37 & 0.55468420E+13 & 0.53594420E+31\\
0.35707300E+37 & 0.60501410E+13 & 0.59429390E+31\\
0.38672000E+37 & 0.65534760E+13 & 0.65589600E+31\\
0.41636500E+37 & 0.70568810E+13 & 0.72068480E+31\\
0.44600500E+37 & 0.75603230E+13 & 0.78857870E+31\\
0.48552100E+37 & 0.82316740E+13 & 0.88379610E+31\\
0.51515000E+37 & 0.87352400E+13 & 0.95862580E+31\\
0.54477600E+37 & 0.92388600E+13 & 0.10362880E+32\\
0.57439900E+37 & 0.97425150E+13 & 0.11167080E+32\\
0.60401800E+37 & 0.10246220E+14 & 0.11998160E+32\\
0.65337700E+37 & 0.11085860E+14 & 0.13440860E+32\\
0.69285800E+37 & 0.11757640E+14 & 0.14645200E+32\\
0.73233400E+37 & 0.12429500E+14 & 0.15892440E+32\\
0.77180500E+37 & 0.13101440E+14 & 0.17181270E+32\\
0.82113400E+37 & 0.13941490E+14 & 0.18848650E+32\\
0.86059200E+37 & 0.14613590E+14 & 0.20226200E+32\\
0.89018100E+37 & 0.15117700E+14 & 0.21284120E+32\\
0.93948300E+37 & 0.15958020E+14 & 0.23093300E+32\\
0.98877600E+37 & 0.16798420E+14 & 0.24958070E+32\\
\botrule
\end{tabular}}
\end{table}
As mentioned in Sec.~3.2.3,
at subnuclear densities all the EoS considered here are joined with the crustal equations of state from Harrison and Wheeler\cite{HW} and Negele and Vautherin.\cite{NV}
The composition of the crust is crystalline, with light\cite{HW} or heavy\cite{NV} metals and electron gas.
The DBHF equations of state for neutron matter and for $\beta$-equilibrated matter are given in Table~3 and
Table~4, respectively. Some points of the EoS used for the crust are displayed in Table~5.
The proton fraction in $\beta$-stable matter is calculated by imposing energy conservation and charge
neutrality. The resulting algebraic equations can be found in standard literature.\cite{Glen}
The contribution to the energy density from the electrons is written as
\begin{equation}
e_e= \frac{\hbar c}{4 \pi ^2}(3 \pi ^2 \rho _e)^{4/3} \, ,
\label{el}
\end{equation}
whereas for muons we write
\begin{equation}
e_{\mu}= \rho _{\mu} m_{\mu}c^2 + (\hbar c)^2\frac{(3 \pi ^2 \rho_{\mu})^{5/3}}{10 \pi ^2 m_{\mu}c^2} \, .
\label{emu}
\end{equation}
These contributions are added to the baryonic part to give the total energy density.
The derivative of the total energy/particle with respect to the fraction of a particular particle species
is the chemical potential of that species. The conditions
\begin{equation}
\mu _p + \mu _e = \mu _n \, \, ; \, \,
\mu _{\mu} = \mu _e \, \, ; \, \,
\rho _p = \rho _{\mu} + \rho _e \; ,
\label{beta}
\end{equation}
allow to solve for the densities (or fraction) of protons, electrons, and muons.
Near the saturation density, when the muon fraction is close to zero, one can estimate the equilibrium
proton fraction, $x_p$, to be\cite{Latt07}
\begin{equation}
x_p \approx \Big (\frac{4 e_{sym}(\rho_0)}{\hbar c} \Big )^3/(3 \pi ^2 \rho _0).
\label{xp}
\end{equation}
The fractions of protons, electrons, and muons as predicted by DBHF are shown in Fig.~\ref{pfrac}.
The critical density for the proton fraction to exceed approximately $1/9$ and, thus, allow cooling through the direct Urca processes,
\begin{equation}
n \rightarrow p + e + \bar{\nu}_e \, \, \, \, \, and \, \, \, \, \,
p +e \rightarrow n + {\nu}_e \, ,
\label{Urca}
\end{equation}
is about
$0.36-0.39$ fm$^{-3}$.
\begin{figure}[!t]
\centering
\vspace*{-1.2cm}
\includegraphics[totalheight=3.5in]{ijmpe17.ps}
\vspace*{-2.2cm}
\caption{(color online) Proton (solid black), electron (dashed red), and muon (dotted blue) fractions
in $\beta$-stable matter as a function of total baryon density as predicted by the DBHF model.
}
\label{pfrac}
\end{figure}
In Fig.~\ref{MR}, we show the mass-radius relation for a sequence of static neutron stars as predicted
by the various models considered previously, see Fig.~\ref{esmicro}.
\footnote{All neutron star properties are
calculated from public software downloaded from the website {\it http://www.gravity.phys.uwm.edu/rns}.}
All models besides DBHF share the same many-body
approach (BHF+TBF) but differ in the two-body potential and TBF employed. The resulting differences can be
much larger than those originating from the use of different many-body approaches. This can be seen by comparing
the DBHF and BOB curves, both employing the Bonn B interaction (although in the latter case the non-reltivistic, r-space version of the potential is adopted). Overall, the maximum masses range from 1.8$M_{\odot}$
(UIX) to
2.5$M_{\odot}$ (BOB). Radii are less sensitive to the EoS and range between 10 and 12 km for all models under
consideration, DBHF or BHF+TBF.
Concerning consistency with present constraints, an initial observation of a neutron-star-white dwarf binary system
suggested a neutron star mass (PSR J0751+1807) of 2.1$\pm$0.2$M_{\odot}$.\cite{Nice} Such observation
would imply a considerable constraint on the high-density behavior of the EoS. On the other hand,
a dramatically reduced value
of 1.3$\pm$0.2$M_{\odot}$
was reported\cite{Piek08} later, which does not invalidate any of the theoretical models under
consideration.
The model dependence is shown in Fig.~\ref{MRrot} for the case of rapidly rotating stars.
The 716 Hz frequency corresponds to the most rapidly rotating pulsar, PSR J1748-2446,\cite{Hessels}
although recently an X-ray burst oscillation at a frequency of 1122 Hz has been reported\cite{Kaaret}
which may be due to the spin rate of a neutron star.
As expected, the maximum mass and the
(equatorial) radius become larger with increasing rotational frequency.
Another bulk property of neutron stars is the moment of inertia, $I$.
In Fig.~\ref{IM}, we show the moment of inertia at different rotational speeds (again, for all models), whereas in
Fig.~\ref{IF} we display the moment of inertia corresponding to the maximum mass at different
rotational frequencies.
These values are not in contradiction with observations of the
Crab nebula luminosity. From that, a lower bound on the moment of inertia was inferred to be
$I \geq $4-8 $\times$ 10$^{44}$ g cm$^2$, see Ref.\cite{Weber} and references therein.
The size of $I_{M_{max}}$ changes from model to model in line with the size of the maximum mass,
see Fig.~\ref{IF}.
\begin{figure}[!t]
\centering
\includegraphics[totalheight=2.5in]{ijmpe18.ps}
\vspace*{0.2cm}
\caption{(color online) Static neutron star mass-radius relation for the models
considered in the text.
}
\label{MR}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[totalheight=1.8in]{ijmpe19.ps}
\includegraphics[totalheight=1.8in]{ijmpe20.ps}
\includegraphics[totalheight=1.8in]{ijmpe21.ps}
\vspace*{0.2cm}
\caption{(color online) Mass-radius relation for the models
considered in the text and for different rotational frequencies.
}
\label{MRrot}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[totalheight=1.8in]{ijmpe22.ps}
\includegraphics[totalheight=1.8in]{ijmpe23.ps}
\includegraphics[totalheight=1.8in]{ijmpe24.ps}
\vspace*{0.2cm}
\caption{(color online) Moment of inertia for the models
considered in the text and for different rotational frequencies.
}
\label{IM}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[totalheight=2.5in]{ijmpe25.ps}
\vspace*{-0.5cm}
\caption{(color online) Moment of inertia corresponding to the maximum mass for the models
considered in the text as a function of the rotational frequency.
}
\label{IF}
\end{figure}
Lastly, we calculate the gravitational redshift predicted by each model.
The redshift is given by
\begin{equation}
z = \Big (1 - \frac{2M}{R} \Big ) ^{-1/2} -1 \; .
\label{red}
\end{equation}
This simple formula can be derived considering a photon emitted at the surface of a neutron star and moving
towards a detector located at large distance.\cite{Weber} The photon frequency at the emitter (receiver) is
the inverse of the proper time between two wave crests in the frame of the emitter (receiver).
Assuming a static gravitational field, and writing $g_{00}$ as the metric tensor component
at the surface of a nonrotating star yield the equation above.
Naturally the rotation of the star modifies the metric, and in that case different considerations need to
be applied which result in a frequency dependence of the redshift. We will not consider the general case
here.
From Fig.~\ref{Z}, it appears that the gravitational redshift is not very EoS-dependent (compare, for instance,
the values at the maximum mass for each model), an indication that the gravitational profile at
the surface of the star is similar in all models.
Thus measurements of $z$ may not be the best way to discriminate among different EoS.
\begin{figure}[!t]
\centering
\includegraphics[totalheight=2.5in]{ijmpe26.ps}
\vspace*{-0.5cm}
\caption{(color online) Gravitational redshift for all models.
For each model, the corresponding sequence of static stars is considered.
}
\label{Z}
\end{figure}
Clearly, at the densities
probed by the interior of neutron stars the model dependence is large, but presently available constraints are
still insufficient to discriminate among these EoS.
At very high density ($\rho$ between five and ten times normal density), the most repulsive {\it
symmetric matter} energies are produced with BOB, V18, DBHF, N93 and UIX, in that order.
This explains the maximum mass predictions, which depend mostly on the absolute repulsion
present in the symmetric matter EoS.
In pure neutron matter, N93 follows right after BOB (again, from largest to smallest repulsion). This indicates
a somewhat different balance of attraction/repulsion when only T=1 contributions are
included.
The symmetry energy, see Fig.~\ref{esmicro}, which depends entirely on the repulsion of neutron matter {\it relative}
to symmetric matter and
whose density dependence controls observables such as the neutron skin,
is largest in N93, followed by BOB, V18, UIX, and DBHF.
The model dependence we observe comes from two sources, the two-body potential and the
many-body approach, specifically the presence of explicit TBF or Dirac effects.
The dependence on the two-body potential is very large. Typically, the main source of model dependence
among NN potentials is found in the strength of the tensor force.
Of course, differences at the two-body level impact the TBF as well, whether they are microscopic
or phenomenological (as in the case of UIX).
On the other hand,
when comparing DBHF and BOB, we are looking at differences stemming from the many-body scheme, as
the two models share the same NN potential.
In the BOB model, repulsion grows at a much faster rate than in DBHF, and more strongly so in neutron
matter. (Hence, the much larger symmetry energy with BOB). As an example, at about 6 times
normal density the DBHF energy of symmetric matter is 67\% of the BOB energy and only 51\% at
ten times normal densities. In pure neutron matter, those ratios become 49\% and 30\%, respectively.
Thus, the inclusion of the microscopic TBF in the BHF model introduces considerable more
repulsion than the Dirac effects
through highly non-linear
terms.\cite{Catania3}
Both attractive and repulsive TBF are required for a realistic description of the
saturation point. The density dependence of the repulsive terms is obviously stronger and thus
dominates at high density.
Furthermore, it appears that this is especially true in neutron matter.
In Ref.\cite{Catania3} it is shown that the largest contribution to the net TBF originates from Z-diagrams such
as shown in Fig.~2, a fact which confirms the validity of the DBHF approximation. The actual amount of repulsion,
however, seems to depend sensitively upon the TBF parametrization. One of the advantages of the DBHF method
is the ability to avoid possible inconsistencies between the parameters of the two- and the three-body systems.
At this time, available constraints cannot pin down the high-density behavior of the EoS.
Nevertheless, we argue again that microscopic models allow for a deeper insight into the
origin of the observed physical effects,
and should be pursued along with improved constraints.
\section{Non-Nucleonic Degrees of Freedom}
\subsection{Introduction}
At densities close to normal nuclear density, protons and neutrons are the only baryonic degrees
of freedom. As density increases, other baryons begin to appear, such as strange baryons
or isospin 3/2 nucleon resonances.
Hyperonic states can be classified according to the irreducible representation of the
$SU(3)$ group. The octet of baryons that can appear in neutron matter includes nucleons,
$\Lambda$, $\Sigma ^{0,\pm}$, and $\Xi ^{0,-}$.
Neglecting the nucleon-hyperon interaction, the threshold for stable hyperons to exist in matter is
determined by comparing the hyperon mass with the neutron Fermi energy, which is the largest available energy scale
in neutron-rich matter.
We consider cold neutron stars, after neutrinos have escaped. Strange baryons appear at about 2-3 times
normal density,\cite{nycatania2} an estimate which is essentially model independent, through the
processes $n + n \rightarrow p + \Sigma ^-$ and
$n + n \rightarrow n + \Lambda$.
The equilibrium conditions for these reactions are
\begin{equation}
2 \mu _n = \mu _p + \mu _{\Sigma^-} \, \, ;
\, \, \mu _n = \mu _{\Lambda}\, .
\label{nybeta1}
\end{equation}
Also, we have
\begin{equation}
\mu _e = \mu _{\mu} \, \, ;
\, \, \mu _n = \mu _p + \mu _e \, ,
\label{nybeta1b}
\end{equation}
the equations above being special cases of
\begin{equation}
\mu = b\mu _{n} - q \mu _e \, ,
\label{nybeta1c}
\end{equation}
where $b$ and $q$ are the baryon number and the charge (in units of the electron charge)
of the particular species with chemical potential $\mu$.
Together with the charge neutrality condition and baryon number conservation,
\begin{equation}
\rho _p = \rho _e + \rho _{\mu} + \rho _{\Sigma ^-} \, \, ;
\, \, \rho = \rho _n + \rho _p + \rho _{\Sigma ^-}
+ \rho _{\Lambda}\, ,
\label{nybeta2}
\end{equation}
the above system allows to determine the various particle fractions.
Clearly,
the composition of matter at supra-nuclear densities determines the behavior of stellar matter.
It is also speculated that a transition to a quark phase may take place at very high densities,
the occurrence of which depends sensitively on the properties of the EoS in the hadronic (confined)
phase.
The presence of hyperons in the interior of neutron stars is reported to
soften the equation of state, with the consequence that the predicted
neutron star maximum masses become considerably smaller.\cite{Sch+06} With recent
constraints allowing maximum masses larger than previously accepted limits (see
previous section),
accurate microscopic calculations which include strangeness (in addition to other
important effects, such as those originating from relativity),
become especially timely.
Thus,
there are strong motivations for including strange baryons in nuclear matter.
Moreover, as far as terrestrial nuclear physics is concerned, studies
of hyperon energies in nuclear matter naturally complement our knowledge
of hypernuclei (see, for instance, Refs.\cite{hyper1,hyper2,hyper3,hyper4}).
For example, the EoS of hypermatter is
useful in the development of generalized mass formulas depending on both density and hyperon fraction.\cite{hyper4,nycatania1}
From the experimental side, additional data are very much needed, especially
on $\Lambda \Lambda$ hypernuclei, which would provide information on the
$\Lambda- \Lambda$ interaction. Concerning single hypernuclei, analyses of
data on $\Lambda$ binding energies constrain
the depth of the single-$\Lambda$ potential to be 27-30 MeV.\cite{Dover}
The status of $\Sigma$ hypernuclei and the $\Sigma$-nucleus potential is more controversial
(see Ref.\cite{Saha04} and references therein).
\subsection{Hyperons and neutron star matter}
Interacting hypernuclear matter was initially studied within variational approaches.\cite{Pan,BJ}
Relativistic mean field models were also extended to include hyperons.\cite {yn1,yn2,yn3,yn4}
Microscopic models require
inclusion of realistic nucleon-hyperon and hyperon-hyperon interactions,\cite{NY89,Nij89,Rij99,Stocks99}
but the experimental information on these interactions is still scarce. The nucleon-hyperon
potentials are fitted to $\Lambda N$ or $\Sigma N$ scattering data. The information on the
hyperon-hyperon interaction is limited to the ground state of double-$\Lambda$ hypernuclei.\cite{Gib94}
Additional constraints can be derived from $SU(3)$ symmetry arguments.
The common denominator among microscopic
calculations of the EoS with hyperons is that they are typically conducted within a
non-relativistic framework together with r-space local nucleon-hyperon (NY) potentials.
Microscopic calculations of nuclear matter properties including hyperons have been reported earlier
within the non-relativistic Brueckner-Hartree-Fock framework (BHF) (see, for instance,
Refs.\cite{nycatania2,nycatania1}), using
the Nijmegen\cite{Nij89}
nucleon-hyperon (NY) meson-exchange potential.
Extensive microscopic work on hyperonic nuclear matter, again within the non-relativistic BHF framework, has also been published by
the Barcelona group (see, for instance, Refs.\cite{Vid+00,Vid+00b,Vid+04,Rios+05}).
The issue of interacting hyperons and their impact on the $\beta$-stable EoS and neutron star structure
has been confronted, for instance, in Ref.\cite{nycatania3} using the Paris\cite{Paris} and the Argonne V18\cite{V18} NN potentials together with the Nijmegen soft-core
(NSC89) NY potential.\cite{Nij89}
A remarkable conclusion from that work is that, in the presence of hyperons, the inclusion of nucleonic
TBF does not alter the EoS appreciably. With nucleons only,
TBF bring in considerable repulsion at high density which result in a much stiffer EoS. On the other hand,
when hyperons are present the nucleonic TBF increase the strange baryon population due to the
increased nucleon chemical potential. In turn, this decreases the nucleon population with the final net effect
on the EoS found to be very small. In other words, it would be necessary for the NY interaction to become
very repulsive at high density to compensate for the loss of nucleons and gain a substantial increase of the star maximum
mass.\cite{nycatania3}
Results for the neutron star maximum gravitational mass, radius, and central
density are shown in Table~6.
The use of different NN potentials seem to produce only small variations.
\begin{table}[pt]
\centering \caption
{Neutron star limiting values for different EoS with (Y) and without (no Y) hyperons as
from Ref.$^{132}$.
}
\vspace{5mm}
\begin{tabular}{ccccccc}
\hline
\multicolumn{1}{c}{EoS} &
\multicolumn{2}{c}{$M_{max}/M_{\odot}$} &\multicolumn{2}{c}{$R (km)$} & \multicolumn{2}{c}
{$\rho _c (fm ^{-3})$} \\
& no Y & Y& no Y& Y& no Y& Y \\ \colrule
V18 & 1.64 & 1.26 & 9.10 & 8.70 & 1.53 & 1.86 \\
Paris & 1.67 & 1.31 & 8.90 & 8.62 & 1.59 & 1.84 \\
V18+TBF & 2.00 & 1.22 & 10.54 & 10.46 & 1.11 & 1.25 \\
Paris+TBF & 2.06 & 1.26 & 10.50 & 10.46 & 1.10 & 1.25 \\
\hline
\end{tabular}
\end{table}
More recently,
neutron star structure results
have been revisited by the Catania group
with the microscopic models we considered earlier in this article (BHF+TBF models from Ref.\cite{Catania4})
along with the NSC89 nucleon-hyperon interaction.\cite{Nij89}
There, the nucleonic energy densities obtained with the different NN interactions are used together with
the same NY interaction to calculate stellar matter. It is found that the inclusion of hyperons reduces dramatically the maximum mass.
It also reduces substantially the maximum mass range covered by the models, which goes from 1.8 to
2.5~$M_{\odot}$ with nucleonic energy densities only, see Fig.~\ref{MR}, to the much narrower intervals of 1.3-1.4~$M_{\odot}$.
Furthermore, the stiffer the original EoS, the larger the softening effect from the presence of strangeness.
Thus, in this approach, one may conclude that hybrid stars (containing hadronic and quark matter) are
necessary to reproduce the larger mass values consistent with recent observations.
Overall, there seems to be a consensus among microscopic calculations of the BHF type (with or
without TBF) that the inclusion of hyperons yields a remarkably low value of a neutron star maximum
mass. On the other hand, predictions from different versions of relativistic mean field models are quite
different from one another.\cite{RMF1,RMF2}
The alternative approach to the EoS with strange baryons presented in the next section
will suggest that
present NY interactions are not yet sufficiently constrained to allow for definite conclusions
concerning the occurrence of a hybrid phase.
\subsection{A first DBHF calculation of the EoS with nucleons and $\Lambda$-hyperons}
It is one purpose of this section to bring in the new aspect of Dirac effects
on the $\Lambda$ hyperon as well as the nucleon. By ``Dirac effects" we mean that the single-baryon
wavefunction is calculated self-consistently with the appropriate effective interaction.
The origin and nature of these effects on the nucleonic equation of state was discussed previously,
see Sec.~3.2.1.
A previous calculation from the Idaho group\cite{Sam08} of the binding energy of a
$\Lambda$ impurity in nuclear matter showed that Dirac effects on the $\Lambda$ hyperon
yield a moderate reduction of the binding energy. In that calculation, we used the most recent
nucleon-hyperon (NY) potential reported in Ref.\cite{NY05} (thereafter referred to as NY05), and observed
that $G$-matrix predictions obtained with NY05 are
significantly differerent from calculations using the previous
(energy independent) version
of the J{\"u}lich NY potential\cite{NY94}.
Therefore, in this work we will use both potentials, for comparison.
As usual, the Bonn B potential \cite{Mac89} is used throughout for the nucleon-nucleon (NN) part.
Previous calculations of the EoS with hyperons have typically been conducted within a
non-relativistic framework together with r-space local NY potentials. The DBHF calculation we describe here uses
non-local relativistic momentum-space (NN and NY) potentials and a relativistic many-body
method,
and is therefore fundamentally different.
Next we give a brief review of the formalism.
The single-nucleon and single-$\Lambda$ potentials are obtained as
\begin{equation}
U_N({\vec k_N}) = U_{N\Lambda} ({\vec k_N}) + U_{NN}({\vec k_N}),
\label{UN}
\end{equation}
and
\begin{equation}
U_{\Lambda}({\vec k}_{\Lambda}) = U_{\Lambda N} ({\vec k}_{\Lambda}),
\label{ULAM}
\end{equation}
i.e., the $\Lambda \Lambda$ interaction is neglected.
In the equations above, the various terms, $U_{B_1B_2}$, are the contributions to the potential
of baryon $B_1$ from its interation with all baryons of type $B_2$. They are given by
\begin{equation}
U_{N\Lambda} ({\vec k_N}) = \sum_{T,L,S,J} \frac{(2T+1)(2J+1)}{(2t_N+1)(2s_N+1)}
\int _0 ^{k_F^{\Lambda}} G^{T,L,S,J}_{N \Lambda}(k({\vec k_N},{\vec k_{\Lambda}}),
P({\vec k_N},{\vec k_{\Lambda}})) d^3 k_{\Lambda},
\label{UNL}
\end{equation}
\begin{equation}
U_{N N} ({\vec k_N}) =
\sum_{T,L,S,J} \frac{(2T+1)(2J+1)}{(2t_N+1)(2s_N+1)}
\int _0 ^{k_F^{N}} G^{T,L,S,J}_{N N}(k({\vec k_N},{\vec k_{N}'}),
P({\vec k_N},{\vec k_{N}'})) d^3 k_{N}',
\label{UNN}
\end{equation}
and
\begin{equation}
U_{\Lambda N} ({\vec k}_{\Lambda}) =
\sum_{T,L,S,J} \frac{(2T+1)(2J+1)}{(2t_{\Lambda}+1)(2s_{\Lambda}+1)}
\int _0 ^{k_F^{N}} G^{T,L,S,J}_{\Lambda N}(k({\vec k_N},{\vec k_{\Lambda}}),
P({\vec k_N},{\vec k_{\Lambda}})) d^3 k_{N},
\label{ULN}
\end{equation}
where the channel isospin $T$ can be 0 or 1 for the NN case and is equal to 1/2 for the
$N\Lambda$ case, and $s_i,t_i$ ($i=N,\Lambda$) are the spin and isospin of the nucleon or
$\Lambda$.
Notice that
\begin{equation}
\frac{U_{N\Lambda}}{U_{\Lambda N}} \approx \frac{\rho _{\Lambda}}{\rho _N},
\label{ny10}
\end{equation}
an approximation often used in mean-field approaches.
The average potential energies of nucleons and $\Lambda$'s are determined from
\begin{equation}
<U_N> =\frac{1}{\rho_N}\frac{1}{(2 \pi)^3}4\frac{1}{2}\int_0^{k_F^N} U_{N} ({\vec k_N})
dk^3_N,
\label{ny11}
\end{equation}
and
\begin{equation}
<U_{\Lambda}> =\frac{1}{\rho_{\Lambda}}\frac{1}{(2 \pi)^3}2\frac{1}{2}\int_0^{k_F^{\Lambda}} U_{\Lambda } ({\vec k}_{\Lambda})
dk^3_{\Lambda},
\label{ny12}
\end{equation}
where the factors of 4 and 2 in Eqs.~(\ref{ny11}) and Eq.~(\ref{ny12}), respectively, account for protons and neutrons in both spin states
or $\Lambda$'s in both spin states.
Finally the average potential energy per baryon is obtained as
\begin{equation}
<U> = \frac{\rho _N <U_N> + \rho _{\Lambda}<U_{\Lambda}>}{\rho _{tot}},
\label{ny13}
\end{equation}
from which, together
with a similar expression for the kinetic energy, one obtains the average energy per
baryon.
The N$\Lambda$ $G$-matrix is obtained from the Bethe-Goldstone equation
\begin{eqnarray}
<N\Lambda|G_{N \Lambda}(E_0)|N\Lambda> & = & <N\Lambda|V|N\Lambda> \\
& & \mbox{} +\sum_{Y=\Lambda,\Sigma} <N\Lambda|V|NY> \frac{Q}{E_0 - E}
<NY|G_{N \Lambda}(E_0)|N\Lambda> , \nonumber
\label{ny14}
\end{eqnarray}
where $E_0$ and $E$ are the starting energy and the energy of the intermediate NY state,
respectively, and $V$ is an energy-independent NY potential.
For two particles with masses $M_N$ and $M_{\Lambda}$ and Fermi momenta
$k_F^N$ and $k_F^{\Lambda}$, Pauli blocking requires
\begin{equation}
Q({\vec k},{\vec P}) =
\left\{
\begin{array}{l l}
1 & \quad \mbox{$|\beta {\vec P}+{\vec k}|>k_F^{\Lambda}$ and
$|\alpha {\vec P}-{\vec k}|>k_F^{N}$ } \\
0 & \quad \mbox{otherwise.}
\end{array}
\right.
\label{ny15}
\end{equation}
The above condition implies the restriction
\begin{equation}
\frac{(\frac{M_N}{M}P)^2 +k^2-(k_F^N)^2}{2Pk\frac{M_N}{M}} > cos \theta >
-\frac{(\frac{M_{\Lambda}}{M}P)^2 +k^2-(k_F^{\Lambda})^2}{2Pk\frac{M_{\Lambda}}{M}},
\label{ny16}
\end{equation}
where $\theta$ is the angle between the total (${\vec P}$) and the relative
(${\vec k}$) momenta of the two particles, and $M=M_{\Lambda} + M_N$. Angle-averaging is then applied in the
usual way.
In the present calculation we consider a non-vanishing density of $\Lambda$'s but
do not allow for the presence of real $\Sigma$'s in the medium
(although both $\Lambda$ and $\Sigma$
are included in the coupled-channel calculation of the NY $G$-matrix, see Eq.~(\ref{ny14})).
Essentially we are considering a scenario where a small fraction of nucleons is replaced
with $\Lambda$'s, as could be accomplished by an experiment aimed at producing multi-$\Lambda$ hypernuclei.
Multistrange systems, such as those produced in heavy-ion collisions, may of course contain other hyperons
on the outset.
For small $\Lambda$ densities, though, as those we consider here, the cascade ($\Xi$) and the $\Sigma$ hyperon
are expected to decay quickly through the strong processes
$N+\Xi \rightarrow \Lambda+\Lambda$ and
$N+\Sigma \rightarrow N+\Lambda$. Under these conditions, a mixture of nucleons and $\Lambda$'s
can be considered ``metastable", in the sense of being equilibrated
over a time scale which is long relative to strong processes.
(In fact, the strong reactions mentioned above would have to be
Pauli blocked in order to produce a metastable multistrange system \cite{SG00}.)
We neglect the $YY'$ interaction, as very little is known
about it. Furthermore, non-local momentum-space $YY$ potentials, appropriate for our DBHF framework,
are not available at this time.
For these reasons, we keep the $\Lambda$ concentration
relatively low.
We have incorporated DBHF effects in the $\Lambda$ matter calculation, which amounts to involving the
$\Lambda$ single-particle Dirac wave function in the self-consistent calculation through the
$\Lambda$ effective mass, $M^*_{\Lambda}$.\cite{nysam}
However, a problem that needs to be addressed with the
J{\"u}lich NY potential in conjunction with DBHF calculations
is the use of the pseudoscalar coupling for the interactions of
pseudoscalar mesons (pions and kaons) with nucleons and hyperons. For the reasons
described in Sec.~3.2.1 (that is, the close relationship between Dirac effects and ``Z-diagram"
contributions),
this relativistic correction is known to become unreasonably large when applied to
a vertex involving pseudoscalar coupling. On the other hand, the gradient (pseudovector)
coupling (also supported by chiral symmetry arguments) largely suppresses antiparticle contributions.
To resolve this problem, one can make use of the on-shell equivalence between the pseudoscalar and
the pseudovector coupling, which amounts to
relating the coupling constants as follows:
\begin{equation}
g_{ps} = f_{pv}\frac{M_i + M_j}{m_{ps}},
\label{ny19}
\end{equation}
where $g_{ps}$ denotes the pseudoscalar coupling constant and $f_{pv}$ the pseudovector one;
$m_{ps}$, $M_{i}$, and $M_{j}$ are the masses of the
pseudoscalar meson and the two baryons involved
in the vertex, respectively.
This procedure can be made plausible by
writing down the appropriate
one-boson-exchange amplitudes and observing that, redefining the coupling constants as above, we have
(see Ref.\cite{Mac89} for the two-nucleon case)
\begin{equation}
V_{pv} = V_{ps} + .....
\label{ny20}
\end{equation}
where the ellipsis stands for off-shell contributions.
Thus, the pseudoscalar coupling can be interpreted as pseudovector coupling where the
off-shell terms are ignored. This is what we apply in our DBHF calculations.
More concretely, from a given pseudoscalar potential, first we extract the corresponding $f_{pv}$ from
Eq.~(\ref{ny19}). In the medium, $V_{ps}^*$ will contain the coupling constant
$g=\frac{M_i^* + M_j^*}{m_{ps}}f_{pv}$. When the baryon masses are reduced through the effective mass
prescription, the reduction of the masses appearing at the denominator of the momentum-dependent part of the potential will
be ``balanced" be the equally reduced masses at the numerator of the coupling constant. This
prescription prevents the OBE diagram from growing to unreasonably large values when effective masses are employed.
Next, we will be showing results for both the NY05 potential \cite{NY05} and the previous version of the
J{\"u}lich NY potential, NY94, specifically the model referred to as ${\tilde A}$
in Ref.\cite{NY94}.
In Fig.~\ref{NY1} we show the energy per particle as a function of density for different $\Lambda$
concentrations as obtained from DBHF calculations along with the NY94 potential.
As more nucleons are replaced with
$\Lambda$'s, generally less binding energy per particle is generated. This is due to the weaker nature of the
$N\Lambda$ interaction relative to the NN one.
Furthermore, the $\Lambda$ Fermi momentum grows rather quickly with $\rho _{\Lambda}$, since only
two $\Lambda$'s can occupy each state, rather than four, which implies a fast rise of the
hyperon kinetic energy. (Although, for small hyperon densities,
there is at first some reduction of the kinetic energy, due to the fact that
$\Lambda$'s have larger mass and can occupy lower energy levels.)
We notice that the saturation density remains essentially unchanged
with increasing hyperon concentrations.
As density grows, though, larger $\Lambda$ concentrations start to yield increased attraction.
One must keep in mind that, especially at the higher densities, the NN interaction become less and less
attractive due to medium effects, in particular repulsive Dirac effects. Thus, removing nucleons
from the system can actually amount to increased binding.
Moving now to Fig.~\ref{NY2}, where the NY05 potential is adopted,
we see a very different scenario.
We recall that the NY05 model is considerably more attractive, yielding
about 50 MeV for the $\Lambda$ binding energy \cite{Sam08,Pol} whereas a value close to 30 MeV was found with NY94
\cite{NY94}.
Naturally, we expect these differences to reflect onto the respective EoS predictions.
Here, increasing the hyperon population yields more binding,
a trend opposite to the one seen in the previous figure.
Again, the final balance is the result of a combination of effects.
The fact that the NN component is repulsive at the higher densities,
together with the more attractive nature of the NY05 potential, determines here a net increase in attraction
with decreasing {\it nucleon} density.
However, the additional binding becomes smaller and
smaller with increasing hyperon concentration, indicating that, at sufficiently large $\Lambda$
densities, the net balance may turn repulsive.
Notice also that in Fig.~\ref{NY2} the minimum moves towards higher densities, signifying that
baryon pairs favor a smaller interparticle distance.
To summarize, one must keep in mind that the energy/particle is the result of a delicate balance
of both the kinetic energies and the potential energies of the two baryon species, being weighed by
the respective densities.
Thus, although the NN interaction is generally more attractive than the N$\Lambda$ one,
the net effect of replacing nucleons with $\Lambda$'s will depend
sensitively on the nature of the NY interaction that's being put into the system, as well as the
``stiffness" of the original, nucleonic, EoS.
This is further confirmed in
Fig.~\ref{NY3}, where we show predictions from conventional BHF calculations (i.e., no ``Dirac" effects).
As in the previous figures,
the red curve is the nucleonic EoS. The solid(dashed) curves are the predictions with NY94(NY05).
Clearly, the effects are opposite depending on the NY interaction. Qualitatively, the trend seen
in each group of curves (solid or dash) is approximately consistent with the one observed previously in the
corresponding (i.~e. same NY potential) DBHF predictions. However, the ``cross over" of the curves
seen in Fig.~\ref{NY1} at about twice saturation density is due to a large extent to the more repulsive nature of the
nucleonic EoS in the DBHF calculation (see comments above).
In conclusion, the effect of hyperons on the EoS is strongly dependent upon the baseline
(nucleonic) EoS as well as the NY potential model.
With regard to the first issue, we stress the importance of starting from a realistic EoS, such as the one
predicted in DBHF calculations or obtained with the inclusion of three-body forces.
The second observation confirms the conclusions of Ref.\cite{vlowk}
concerning the large uncertainties originating from the bare NY potentials.
Unfortunately, the existing data do not set sufficient constrains on the potentials, as demonstrated by
the fact that different potentials can fit the available scattering data equally accurately but
produce very different scattering lengths \cite{vlowk}.
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\hspace*{-2.0cm}
\scalebox{0.4}{\includegraphics{ijmpe27.ps}}
\vspace*{-2.6cm}
\caption{(Color online) Energy/particle as a function of density in symmetric nuclear matter for various
$\Lambda$ concentrations $Y_{\Lambda}$. Predictions obtained from DBHF calculations with the NY94 potential.
}
\label{NY1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\hspace*{-2.0cm}
\scalebox{0.4}{\includegraphics{ijmpe28.ps}}
\vspace*{-3.0cm}
\caption{(Color online) Energy/particle as a function of density in symmetric nuclear matter for various
$\Lambda$ concentrations $Y_{\Lambda}$. Predictions obtained from DBHF calculations with the NY05 potential.
}
\label{NY2}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2.5cm}
\hspace*{-2.0cm}
\scalebox{0.4}{\includegraphics{ijmpe29.ps}}
\vspace*{-2.0cm}
\caption{(Color online) Energy/particle as a function of density in symmetric nuclear matter for various
$\Lambda$ concentrations. Predictions obtained from BHF calculations. The solid(dashed) lines
are obtained with the NY94(NY05) potentials.
}
\label{NY3}
\end{center}
\end{figure}
Of course,
the actual fraction of hyperons present in star matter must be determined by the equations
of $\beta$ stability and charge neutrality for highly asymmetric matter containing neutrons, protons, hyperons,
and leptons as described in Sec.~5.1.
What we have learnt at this time is
that the predicted energy/particle in symmetric matter is very sensitive to the
chosen NY interaction.
The uncertainties due to the model dependence discussed in this
paper are likely to impact any conclusions on the properties of strange neutron stars,
which therefore must be interpreted with caution. These include considerations of deconfinement
and possible transition from hadronic to quark matter, which depend sensitively on the equation
of state in the hadronic phase.
\begin{figure}[!t]
\centering
\vspace*{-3.2cm}
\includegraphics[totalheight=5.5in]{ijmpe30.ps}
\vspace*{-3.5cm}
\caption{(color online) Various particle fractions in $\beta$-stable matter.
Neutron: long-dashed black; proton: solid black; $\Sigma ^-$: dash-dotted green;
$\Lambda$: dash-double-dotted blue; electrons: short-dashed red; muons: dotted purple.
}
\label{allfrac}
\end{figure}
\subsection{Neutron star matter with non-interacting hyperons}
As anticipated at the end of Sec.~5.2,
the uncertainties that we have demonstrated in Sec.~5.3
suggest that
present NY interactions are not sufficiently constrained to allow for robust conclusions
concerning the occurrence of a hybrid phase. Therefore, to avoid those uncertainties,
we will now try to {\it estimate} the impact of hyperons on the $\beta$-stable
DBHF EoS
considering {\it non-interacting} hyperons.
Equations~(\ref{nybeta1},\ref{nybeta1b},\ref{nybeta2})
are solved (for fixed total baryon density) treating $\Lambda$ and $\Sigma ^-$ as free
(non-relativistic) fermions. The nucleonic energies are taken from our DBHF EoS for IANM, evaluated at
some nucleon density, $\rho_n + \rho_p$, with
$\rho_n$ and $ \rho_p$ to be determined along with the other particle fractions. (To allow for the solution of
the 6$\times$6 (including the normalization condition) system of algebraic equations for the six unknown
particle concentrations, our nucleonic EoS is first parametrized in terms of analytic functions of density.)
Those fractions are shown in Fig.~\ref{allfrac}. The trend is very similar to the one shown in Ref.\cite{nycatania3}
with free hyperons, which, in turn, is qualitatively consistent with the one seen when the NY interaction
is turned on.\cite{nycatania3} The onset of the $\Sigma^-$ baryon occurs at a density of about 0.3$fm^{-3}$. Leptons start
to disappear from the system after hyperon formation, an indication that charge neutrality through
$\Sigma^-$ production is energetically more favorable than through $\beta$ decay.
The onset of $\Lambda$ production occurs at a density of approximately 0.6$fm^{-3}$. These thresholds are in
good agreement with those reported in Ref.\cite{nycatania3} for both free and interacting hyperons, when
the BHF+TBF predictions are considered (the TBF shifts down the threshold densities for $\Lambda$ and
$\Sigma^-$ formation as compared with the case of two-body forces only).
In Fig.~\ref{Ypr}, we show the pressure in $\beta$-stable matter with the Idaho DBHF EoS with and
without (free) hyperons. The differences are dramatic but consistent with previous observations.\cite{nycatania3}
The effect of including the NY interaction is quite small compared with the differences seen in
Fig.~\ref{Ypr}. Thus, one conclusion seems appropriate at this point: Regardless the chosen NY interaction,
the effect of adding strange baryons is a dramatic softening of the EoS, (consistent with a large reduction of the
neutron star maximum mass). {\it Furthermore, given the uncertainties arising from the NY potential
demonstrated in the previous section, the effect of free hyperons may very well provide a realistic, average estimate
of the impact of including strangeness.}
In Fig.~\ref{Yedens} we compare energy densities in $\beta$-stable matter as obtained from the Idaho
DBHF EoS with free hyperons and from the two models (AV18 + TBF, Paris + TBF) used in Ref.\cite{nycatania3}
along with the NSC NY potential.\cite{Nij89} The similarity between the curves indicates that
the effect of including the NY interaction in star matter is relatively small and/or that differences in the respective
nucleonic EoS may compensate those which originate from the treatment of hyperons.
The neutron star maximum mass we obtain with the DBHF EoS and free hyperons is close to 1.2 $M_{\odot}$.
Considering that inclusion of the NY interaction would increase the pressure at high density, this result is
not inconsistent with those shown in Table 6.
In summary, our analysis confirms a large softening of the $\beta$-stable EoS as a consequence of including strange
baryons, accompanied by a strong reduction of the maximum mass, a conclusion that appears to be nearly model independent.
The softening is
mostly caused by conversion of the kinetic energy of the existing particle species into masses of the newly
produced particles.
The resulting maximum mass reduction
appears in contraddiction with present constraints, and thus
leaves intriguing questions open concerning phase transitions and degrees
of freedom appropriate for high-density stellar matter.
\begin{figure}[!t]
\centering
\includegraphics[totalheight=4.0in]{ijmpe31.ps}
\vspace*{-2.5cm}
\caption{(color online) Pressure in $\beta$-stable matter from the DBHF calculation
with no hyperons (black dash) and
with free hyperons (solid red).
}
\label{Ypr}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[totalheight=4.0in]{ijmpe32.ps}
\vspace*{-2.5cm}
\caption{(color online) Energy density in $\beta$-stable matter from various models: The black dashed and
dotted lines (essentially identical) are the results of BHF calculations with the Argonne V18 and the Paris potentials,
respectively, including TBF and hyperons interacting via the Nijmegen soft-core NY potential (see
Ref.$^{132}$). The red solid line is obtained with the DBHF calculation for the nucleonic EoS and
free hyperons.}
\label{Yedens}
\end{figure}
\section{Conclusions and Outlook}
The EoS of hadronic matter enters in a variety of systems, from atomic nuclei to the most compact
form of matter found in the universe, namely the matter in the interior of neutron stars.
EoS-related properties can and should be calculated {\it ab initio} and consistently from the same
basic nuclear forces. They include: the nuclear matter ``optical potential", effective in-medium cross sections,
nucleon mean free path in nuclear matter, neutron skins, and neutron star properties.
We have reviewed recent calculations of the EoS,
over a large range of densities and isospin asymmetries,
with a particular eye on the microscopic approach.
The obvious advantage of the latter lies in the opportunity of interpreting the predictions in terms
of the input nuclear forces and their behavior in the many-body environment.
In turn, this facilitates the identification of potentially missing dynamics.
We have discussed differences among microscopic models which do or do not include explicit
TBF.
It is quite clear that model dependence amongst predictions can be quite large.
Naturally, this is especially the case at those densities where constraints are the weakest.
Rich and diverse effort is presently going on
to improve the available constraints on the EoS or find new ones.
These constraints are usually extracted through the analysis of selected heavy-ion collision
observables. At the same time,
partnership between nuclear physics and astrophysics is becoming increasingly important towards
advancing our understanding of exotic matter.
Perhaps the best prospects to discriminate among families of EoS, especially at high density,
are in more accurate measurements of neutron stars radii and/or
moments of inertia.
Other EoS-related issues of contemporary interest which we have not reviewed here include
temperature dependence and polarizability of nuclear/neutron matter.
Knowledge of the finite-temperature EoS plays a crucial role in the final stages of a supernova evolution.
Microscopic models which can successfully describe the ground state of nuclear matter
should be the starting point to move on to its excited states.
As a final note,
the emergence of FRIB\cite{FRIB} will mark a turning point in the progress of experimental nuclear physics,
possibly unveiling unknown
areas on the nuclear chart.
It is thus imperative that such large-scale projects
be constantly supported by theoretical calculations with predictive power.
\section*{Acknowledgements}
Support from the U.S. Department of Energy under grant DE-FG02-03ER41270
is gratefully acknowledged.
I like to thank D. Alonso and P. Liu for their help in various phases of the
Idaho project.
I am grateful to M. Baldo, U. Lombardo, and
H.-J. Schulze for useful communications, to F. Weber for providing the
crustal EoS, and to J. Heidenbauer for providing the nucleon-hyperon potential
code.
|
\section{Summary}
In summary, the one-pion exchange current corrections to the
isoscalar and isovector magnetic moments have been studied in the
RMF theory with PK1 effective interaction and compared with previous
relativistic and non-relativistic results. It has been found that
the one-pion exchange current gives a negligible contribution to the
isoscalar magnetic moments but a significant correction to the
isovector ones. However, the one-pion exchange current doesn't
improve the description of nuclear isovector magnetic moments for
the concerned nuclei. In the future investigation, similar as the
non-relativistic cases~\cite{Towner1987,Arima1987}, the second-order
core polarization effects, the $\Delta$ isobar current, crossing
term between MEC and core polarization, and exchange current
corrections due to other mesons should be taken into account. In
addition, the correction due to the restoration of the rotational
symmetry~\cite{Yao2009} may play a role as well. The investigation
towards these directions is in progress.
\begin{acknowledgments}
We would like to thank W. Bentz for his careful reading of the
manuscript and comments. This work is partly supported by Major
State Basic Research Developing Program 2007CB815000, the National
Natural Science Foundation of China under Grant Nos. 10775004,
10720003, 10947013, 10975008, and 10975007, as well as the Southwest
University Initial Research Foundation Grant to Doctor No.
SWU109011.
\end{acknowledgments}
|
\section{Introduction}
An important class of the $\sim 400$ extra-solar planets known to date are the
so-called ``hot Jupiters'', orbiting a fraction of an AU from their host star.
Many of these have significantly larger radii than planets of similar mass in
our Solar System. This has been a challenge for theoretical
models of their structure. The intense radiation and tidal forces from the host
star are expected to play a significant role
\citep[e.g.][]{Bodenheimer2001,Burrows2007}. For planets in highly eccentric
orbits, both of these factors vary greatly over the course of the orbit. Thus,
such planets provide interesting tests for models of the structure and dynamics
of planetary atmospheres \citep[e.g.][and references therein]{Irwin2008}.
A planet in a 111-day orbit around the star HD~80606 was first detected in
radial velocity observations \citep{Naef2001}. The minimum mass (\mbox{$M_p \sin i$}) of the
planet is 3.9 times the mass of Jupiter. HD~80606~b has the most eccentric
orbit of all the extra-solar planets known to date ($e = 0.93$). Infrared ($8
\mu$m) observations clearly show the rapid heating of the planet's atmosphere
during periastron passage \cite[herein referred to as
L09]{Laughlin2009}. Laughlin et al. also reported detection of a secondary
eclipse, implying the orbital inclination is close to 90 degrees, and
motivating efforts to observe a transit of the planet in front of the star.
In early 2009, several groups observed a transit egress in photometry
\citep{Moutou2009,Garcia-Melendo2009,Fossey2009} and spectroscopy
\citep{Moutou2009}. Analyses of these data, together with old and new
radial-velocity measurements provided the first constraints on the planet's
radius and actual mass, as well as its orbital parameters (\citealt{Pont2009},
herein referred to as P09; \citealt{Gillon2009}, G09). These analyses are
limited by the fact that the duration of the transit is not constrained,
which in turn increases uncertainties in the system
parameters. \citet[W09]{Winn2009} report multiple observations of the transit
ingress in June 2009, and combine their data with the earlier ingress
observations and Keck radial-velocity data to constrain the transit duration.
In this paper we present previously unpublished multi-site photometric
observations of a complete transit of HD~80606~b on 14 February 2009, and an
ingress on 25 October 2008. We describe the observations and the resulting data
sets in section~\ref{sec:obsphot}. In section~\ref{sec:analysis} we describe
the methods we used to analyse our data, and in section~\ref{sec:results} we
report and discuss the results.
\section{Observations}
\label{sec:obsphot}
We observed HD~80606 with four telescopes, during predicted transit windows
around 25 October 2008 (herein referred to as Oct08), and 14 February 2009
(Feb09).\footnote{We obtained the predicted transit windows from {\it
http://transitsearch.org}.} As no transit or secondary eclipse had been
reported at the time, these windows were $\sim24$~hours long.
Where available, we observed through Bessell $B$ and $R$ filters, alternating
between them every few minutes. The one exception to this is noted below. In
all cases we deliberately defocused the telescope (often to the point where
the point-spread function became an annulus, but taking care that the target
and its binary companion do not blend). This allowed longer exposure times
without saturating the detector. More importantly, it meant that each star was
measured by a large number of pixels, reducing the effects of flat-fielding
errors and pointing drift on the photometry.
Besides those mentioned below, we also attempted observations at Bisei
Astronomical Observatory in Japan, Kryonerion Astronomical Station in Greece,
and an amateur observatory in Athens, Greece, but did not yield any useful data
due to adverse weather.
\subsection{Faulkes Telescope North, Hawaii}
The 2~m Faulkes Telescope North (FTN) --- part of the Las Cumbres Observatory
Global Telescope Network (LCOGT.net) --- is located on Haleakala in Hawaii. We
used FTN and a Merope camera with a 2k$\times$2k e2v CCD (binned $2\times
2$). Observations were conducted by the telescope's Robotic Control System,
alternating between sets of 37-sec exposures in $B$ and 20-sec exposures in
$R$. We only obtained useful data for the Feb09 transit. Images were processed
by our automated pipeline to remove the effects of bias, dark current and CCD
sensitivity variations.
\subsection{IUCAA Girawali Observatory, India}
Observations were carried out from IUCAA Girawali Observatory
\citep[IGO;][]{Das1999} and we obtained useful data during the Feb09 transit window.
We used the IUCAA Faint Object Spectrograph Camera \citep[IFOSC]{Gupta2002}
mounted on the direct Cassegrain focus of the IGO 2~m telescope. Exposure times
were 15--30~sec in $B$ and 4--10~sec in $R$. Images were corrected using
overscan strips and master bias frames followed by flat-fielding, using the
IFOSC photometry pipeline package developed under IRAF.
\subsection{Oskar-L\"{u}hning Telescope, Germany}
We observed HD~80606 on three nights in Feb09 with the 1.2~m
Oskar-L\"{u}hning Telescope (OLT) at Hamburg Observatory. We changed filters
after every 5 exposures, with exposure times of 55~s in $B$ and 35~s in $R$.
Images were bias-/dark-subtracted and flat-fielded using the {\it Figaro}
subroutine in the {\it Starlink} software package.
\subsection{Liverpool Telescope, Canary Islands}
Liverpool Telescope \citep[LT;][]{Steele2004} data were obtained on 2008
October 25 and 26 using the RISE frame transfer wide field camera
\citep{Steele2008} in $2\times 2$ binned mode, giving a pixel scale of
1.07~arcsec/pixel. The filter employed consisted of 2~mm Schott KG5 bonded to
3~mm Schott OG515, giving half-maximum transmission from 515~nm to 695~nm
(i.e. roughly equivalent to a combined $V+R$ filter). An exposure time of
1~sec was used, with effectively no dead time between exposures. All data were
dark-subtracted and flat-fielded using the ULTRACAM pipeline, which is
optimised for time series photometry. A master twilight flat was constructed
from a median of several hundred frames.
\begin{figure*}
\includegraphics[]{fit/hd80606_lc.eps}
\caption{The complete data set analysed in this paper: A transit ingress of
HD~80706~b observed with the Liverpool telescope (LT, top panel), and the
following transit observed by the Faulkes Telescope North (FTN), the IUCAA
Girawali Observatory's 2m telescope (IGO), and the Oskar-L\"{u}hning
Telescope (OLT). The middle and bottom panels show the same transit in the
Bessell $R$ and $B$ bands, respectively. The horizontal bar across the top
shows the section of data used in the fit. The error bars indicate the
weights used in the fit, i.e. they include the statistical uncertainty as
well as the red noise estimate ($\sigma_i$ and $\sigma_{r,s}$ in
equation~\ref{equ:merit}). \label{fig:lc}}
\end{figure*}
\subsection{Photometry} \label{subsec:photometry}
Starting with the calibrated images, we performed aperture photometry on all
data sets in the same manner, with the exception of the LT data.
For each instrument, a single master image was selected, and used as a
reference for all images obtained with that instrument. The coordinates of the
target and reference stars in the master image were obtained by convolving it
with a copy of the PSF from the same image, then running DAOFIND in IRAF on the
result.
All other images were registered to the master using direct correlation with
the XREGISTER task. The master list of coordinates was then transformed into
the reference frame of each image. DAOPHOT was used to obtain photometry in a
range of apertures (slightly different for each instrument). The photometric
scatter was smallest in the largest aperture used for all data sets (radius
11'' for FTN, 8'' for IGO, and 10'' for OLT).
For the LT data, we also performed aperture photometry, trying out different
aperture radii, choosing the one that minimised the noise (8.6~arcsec).
Apertures were centred in each frame by cross-correlation of the image with a
Gaussian.
We measured the flux of HD~80606 relative to its binary companion HD~80607, as
it was the only reference star available in all data sets. Additional stars in
the fields of view of some of the instruments were significantly fainter and of
different colours, and were only used as check stars.
We prepared each light curve for fitting by taking the following steps.
\begin{enumerate}
\item We removed points more than $5\sigma$ from the median of the time series.
\item We normalised the flux values to a median of 1 for points well outside
the transit.
\item We binned the light curve into 10-minute bins in order to reduce the
number of points to fit. We estimated uncertainties as the standard deviation
of the values within each bin, divided by the square root of the number of
points in the bin.\footnote{For bins containing only a single original point,
we estimated the uncertainty based on the nearest two bins with at least two
points.}
\end{enumerate}
\section{Analysis}
\label{sec:analysis}
The complete set of useful data we obtained is shown in Figure~\ref{fig:lc}. It
includes a transit ingress observed by the LT in October 2008, and --- between
the FTN, IGO and OLT data sets --- nearly continuous coverage of the transit in
February 2009 in two filters.
\subsection{Light-curve model}
We modelled the HD~80606 system as a planet of radius \mbox{$R_p$}, orbiting a star of
radius \mbox{$R_*$}, with orbital period \mbox{$P$}, semi-major axis $a$, eccentricity $e$,
and inclination $i$ (relative to our line of sight). The argument of periastron
is \mbox{$\omega$}, and the planet passes there at time \mbox{$T_p$}.
To obtain light curves for a given set of system parameters, we first solved
Kepler's equation using Newton-Raphson iteration, then calculated the planet's
position in its orbit, and its distance (in the plane of the sky) from the
centre of the star. Taking this as input, the analytical transit models of
\citet{MandelAgol2002} provided the relative fluxes expected for our physical
model. We assumed a quadratic limb-darkening law, with coefficients
interpolated from the tables of \citet{Claret2000}.\footnote{Based on the
stellar parameters reported by \citet{Naef2001}, we obtained the limb-darkening
coefficients $u_1=0.5043$, $u_2=0.2535$ in $R$ band, and $u_1=0.8827$,
$u_2=-0.0166$ in $B$ band. For the LT data set ($V+R$), we used the mean of the
coefficients for $V$ and $R$: $u_1=0.6117$, $u_2=0.1853$.}
\subsection{Fitting}\label{subsec:fitting}
We determined the parameters of the model using a Markov-Chain Monte Carlo
(MCMC) approach similar to that of \citet{TorresWinnHolman2008} and references
therein. Briefly, from a random starting point, small jumps in parameter space
are generated, and each new point is evaluated using a merit function
($\chi^2$). If the jump results in a lower $\chi^2$, it is executed (i.e. the
new point becomes the next ``link'' in the chain). Otherwise,
it is only executed with a probability $\exp(-\Delta\chi^2/2)$. If a jump is
not executed, the previous point is repeated in the chain. After a large
number of jumps, the result is a distribution of points in parameter space
that approximates the joint posterior probability density of all the
parameters given our data.
We used a merit function of the form:
\begin{equation}
\label{equ:merit}
\chi^2 = \sum_{i=1}^{N}{ \frac{(n_s f_i - m_i)^2}{\sigma_i^2+\sigma_{r,s}^2} }
+ \sum_{j=1}^{M}{ \frac{(v_j - c_j)^2}{\sigma_{c,j}^2} }
\end{equation}
The first sum is the usual chi-squared statistic and is related to the
probability of the data given the model. For
each of $N$ data points, $f_i$ is the measured relative flux, $m_i$ is the
model at that point, $\sigma_i$ is the statistical uncertainty in the
measurement, and $\sigma_{r,s}$ is an estimate of the correlated noise in
each data set (see below). We also included a normalisation factor
$n_s$ (constant within each data set $s$) to allow the model to match the
out-of-transit relative flux.
The additional terms in the merit function represent constraints from
previously published measurements, as detailed in
Table~\ref{tab:constraints}. In each case a value calculated from our
model ($v_j$) is being compared to a measured value ($c_j$) with
corresponding uncertainty ($\sigma_{c,j}$). We included the time and
duration of the secondary eclipse from L09, as well as the period,
eccentricity and argument of periastron they derived from Keck
radial-velocity data. W09 note that the uncertainty in \mbox{$\omega$}\ quoted by
L09 is likely to be underestimated. Indeed, both W09 and P09 analyse
larger data sets and find uncertainties in \mbox{$\omega$}\ of
$\sim0.2$\degr. Therefore we adopt the value of \mbox{$\omega$}\ from L09, but only
give it a weight corresponding to $\sigma_{c,\mbox{$\omega$}} = 0.2$. We did not
include constraints from any other photometric observations in order
to obtain results as independent as possible from other analyses.
\begin{table}
\caption{Measurements by \citet{Laughlin2009} included as constraints in the
MCMC merit function. The three columns correspond to $v_j$, $c_j$ and
$\sigma_{c,j}$, respectively, in equation~\ref{equ:merit}. The last two rows
refer to the epoch and duration ($2^{nd}$ to $3^{rd}$ contact) of the
secondary eclipse.} \label{tab:constraints}
\begin{tabular}{@{}lrl}
\hline
Parameter & Value & Uncertainty \\
\hline
Orbital period, \mbox{$P$}\ (days) & 111.4277 & 0.0032 \\
Orbital eccentricity, $e$ & 0.9327 & 0.0023 \\
Argument of periastron, \mbox{$\omega$}\ (deg) & 300.5 & 0.2\,$^a$ \\
Epoch of mid-eclipse (HJD) & 2454424.736 & 0.003 \\
Total-eclipse duration (days) & 0.070 & 0.009 \\
\hline
\end{tabular}
$^a$ This is our conservative estimate of the uncertainty (see section~\ref{subsec:fitting}). The value reported by \citet{Laughlin2009} is \mbox{$\omega$}$= 300.4977 \pm 0.0045$\degr.
\end{table}
During the MCMC simulation we varied the following parameters: $a/\mbox{$R_*$}$, \mbox{$P$},
\mbox{$T_p$}, $e'$ (see below), \mbox{$\omega$}, $b = a/\mbox{$R_*$} \cos i$ (proportional to the transit
impact parameter), $\mbox{$R_p$}/\mbox{$R_*$}$, and a normalisation factor for each
instrument/filter combination ($n_s$). A total of 14 parameters were varied. In
order to avoid strong correlations which restrict the movement of the MCMC
chain through parameter space, we used combinations of some of the physical
parameters that are more directly constrained by the data (e.g. the radius
ratio instead of the absolute radii). When eccentricity was included as one of
the varied parameters, correlations caused different MCMC chains to converge to
different locations in (\mbox{$P$}, $e$, \mbox{$\omega$}) space. We found that most of the
variation in $e$ over all chains could be described as a linear function of
\mbox{$P$} and \mbox{$\omega$}. Thus, we replaced $e$ with $e' = e - k_0\mbox{$P$} - k_1\mbox{$\omega$}$, where $k_0
= 0.035$ and $k_1 = -0.0035$ were determined from the previous fit. This
resulted in the MCMC simulation moving more freely through parameter space and
a more continuous distribution of points when the chains were combined.
We generated jumps by adding a Gaussian-distributed
random value to one of the varied parameters (randomly chosen) at each
iteration. We tuned the jump sizes for each parameter so that approximately
25--50\% of jumps were executed.
\subsection{Correlated noise}
Each of our data sets contains some noise that is correlated in time due to
gradual variations in atmospheric or instrumental factors. The {\it internal}
uncertainties we estimated for each data point (as described in
section~\ref{subsec:photometry}) do not measure this {\it red noise}, yet it
can significantly affect the results of our analysis \citep{Pont2006}. To
account for the effect of red noise, we reduced the weight of each point in the
merit function (equation~\ref{equ:merit}) with an additional noise term,
$\sigma_{r,s}$, which we estimated as follows. We first performed an initial
fit setting $\sigma_{r,s}$ to an optimistic 0.0005 for all data sets. For each
data set $s$ (of $N_s$ points), we then found the value of $\sigma_{r,s}$ for
which the reduced chi-squared
\begin{equation}
\label{equ:rchisq}
\chi^2_r = \frac{1}{N_s-8}
\sum_{i=1}^{N_s}{ \frac{(f_i-m_i)^2}{\sigma_i^2+\sigma_{r,s}^2} }
\end{equation}
was equal to unity.\footnote{The effective number of parameters for this
chi-squared is 8, consisting of the 7 physical parameters and one normalisation
factor $n_s$ for the given data set.} We then repeated the fit, and checked
that the values of $\sigma_{r,s}$ were self-consistent. The final values used
were
LT: 0.001,
FTN($R$): 0.0016,
IGO($R$): 0.0015,
OLT($R$): 0.0012,
FTN($B$): 0.0019,
IGO($B$): 0.0017,
OLT($B$): 0.001.
The average internal uncertainty of data used in the fit was only 0.0006, so
--- as is generally the case in such analyses --- red noise is the main factor
limiting the accuracy of our results.
\begin{figure*}
\includegraphics[]{fit/hd80606_fit.eps}
\caption{Close-up of the data used in the fit, along with the best-fit model
obtained from MCMC simulations. Symbols and error bars are the same as in
Figure~\ref{fig:lc}, only the horizontal scale is different.
\label{fig:fit}}
\end{figure*}
\begin{table*}
\begin{minipage}{126mm}
\caption{Physical parameters of the HD~80606~b system obtained from the MCMC fit.} \label{tab:results}
\begin{tabular}{lcc}
\hline
Parameter & Median & Uncertainty \\
\hline
Scaled semi-major axis, $a/\mbox{$R_*$}$ & 101.2 & -5.5, +6.9 \\
Orbital period, \mbox{$P$}\ (days) & 111.4273 & $\pm$0.0031 \\
Orbital eccentricity, $e$ & 0.93369 & $\pm$0.00068 \\
Epoch of periastron, \mbox{$T_p$}\ (HJD) &2454424.8575& $\pm$0.0040 \\
Argument of periastron, \mbox{$\omega$}\ (deg)& 300.53 & $\pm$0.19 \\
Orbital inclination, $i$ (deg) & 89.341 & -0.063, +0.073\\
Planet/star radius ratio,$\mbox{$R_p$}/\mbox{$R_*$}$ & 0.0967 & -0.0035, +0.0032 \\
Planet radius, $\mbox{$R_p$}$\,$^a$ ($R_{Jup}$) & 0.921 & $\pm$0.036 \\
\hline
Epoch of mid-transit (HJD) & 2454876.3173 & $\pm$0.0036 \\
Complete transit duration ($t_4-t_1$, hours) & 12.14 & $\pm$0.36 \\
Ingress/egress duration ($t_2-t_1$, hours) & 2.33 & $\pm$0.37 \\
\hline
\end{tabular}
\\
$^a$ Using $\mbox{$R_*$} = 0.978 \pm 0.015 R_{\sun}$ \citep{Pont2009}.
\end{minipage}
\end{table*}
\section{Results and Discussion}
\label{sec:results}
For our final fit, we created 25
independent chains of 42,000
steps each, and discarded the first 2,000
steps of each chain to minimise the effect of the (randomly chosen) starting
points on our results. Combining the remaining steps of all the chains, the
resulting distribution of parameter values approximates their joint probability
density. For each parameter we report the median of this distribution as the
``best-fit'' value,\footnote{As \citet{Pont2009} point out, the set of
parameters thus obtained does not actually correspond to the MCMC step with the
lowest $\chi^2$ value. This is because each parameter's median value comes from
a different step in the chain.} and the boundaries of the central 68\% as the
1-$\sigma$ uncertainties.
The results thus obtained are shown in Table~\ref{tab:results} and the fit
itself is compared to the data in Figure~\ref{fig:fit}. The reduced chi-squared
for this fit is 0.63, which suggests we have slightly overestimated the
uncertainties in our data. Thus the uncertainties in our results are somewhat
conservative. We also note that there are unavoidable correlations between some
parameters. Besides the \mbox{$P$}-$e$-\mbox{$\omega$}\ correlation mentioned in
section~\ref{subsec:fitting}, the strongest of these are between $a/\mbox{$R_*$}$, $b$
and $\mbox{$R_p$}/\mbox{$R_*$}$. There are also significant correlations between \mbox{$T_p$}\ and \mbox{$\omega$}, and
also between $\mbox{$R_p$}/\mbox{$R_*$}$ and the normalisation factor for the IGO data. The
uncertainties in results include the effects of all these correlations.
Since we used the values of the period (\mbox{$P$}), eccentricity ($e$) and argument
of periastron (\mbox{$\omega$}) from L09 as direct constraints in the fit, we obtained
results that agree with those values. We did find a larger uncertainty for \mbox{$\omega$},
but a more precise value for the eccentricity.
Our results also agree well with the parameters reported by P09, G09 and
W09. The only significant exception to this is the period, where our result is
$3\sigma$ smaller than the other analyses. We note, however, that we have
essentially forced our simulations to agree with the period reported by L09. If
we repeat our fit without the direct constraints on \mbox{$P$}, $e$ and \mbox{$\omega$}\ (see
section~\ref{subsec:fitting}), the best-fit period is almost unchanged, but the
uncertainty in its value is four times greater. This means our photometric data
are in fact consistent with the period reported by P09, G09 and W09.
We also find a planet-to-star radius ratio that is marginally smaller (by
$1.8\sigma$) than the values reported by W09 and G09. A possible cause for this
is the nature of our data from the IGO, which contains the majority of the
in-transit points, but no out-of-transit points on the same night. This means
the normalisation of those data is not strongly constrained, and leads to the
correlation noted above. In addition, the IGO data appear to show a significant
systematic trend in the middle of the transit, which we were unable to remove
by de-correlation against any physical parametrers of the observations (such as
airmass or position on the CCD).
\section{Conclusion}
\label{sec:conclusion}
We have obtained multi-site observations of a transit ingress and a complete
transit of HD~80606~b across its host star. We analysed these data
independently of any other photometric data, and found system parameters
consistent with previously reported values.
These observations were made using four telescopes at different sites.
This allowed us to obtain near-continuous coverage of this 12-hour
event. However, the differences between the instruments, telescopes
and time-allocation procedures were, to an extent, limitations on our
ability to obtain a uniform data set. In the near future, the
completion of LCOGT's network of 1m robotic telescopes \citep{Brown2010} will greatly facilitate observations of this kind, providing near-identical
instrumentation at a number of sites, under the control of a flexible,
central scheduling system.
\section*{Acknowledgments}
We are grateful to Greg Laughlin for pointing out the opportunity to observe
this transit. We thank Tom Marsh for the use of the ULTRACAM pipeline.
|
\section{Introduction}
Self-consistent mean-field (SCMF) models have proven to be reliable
tools for describing nuclear structure and dynamics, for reviews see
\cite{Ben03aR,Vre05aR,Stone_rew}. They are steadily developing to
improve the quality of the predictions and to accommodate more and
more observables. One of the very demanding observables are fission
life-times. Fission barriers have been discussed already in the early
stages of SCMF models and even been used as benchmark for calibration
\cite{skms}. Fission life-times are much more involved as their
computation requires not only the potential energy surface along the
fission path, but also the corresponding collective masses and a safe
estimate of the collective ground state correlations for the initial
state. There are thus not so many self-consistent calculations of
fission life-times\cite{Ber01a,War06} --
{mostly using still approximate masses } { and
quantum corrections (=zero point energies \cite{Rei87aR})}.
The vast majority of calculations employs the microscopic-macroscopic
method which combines shell corrections with a macroscopic liquid-drop
model background, see e.g. \cite{Mol87a,Smo95a}. It is the aim of
this contribution to explore the systematics of fission life-times all
over the landscape of super-heavy elements. This is done using the
Skyrme-Hartree-Fock method as one widely used nuclear SCMF model
\cite{Ben03aR}.
The paper is outlined as follows: In section \ref{sec:formal} we
briefly review the theoretical background and explain how we compute
fission life-times. Results are presented in section
\ref{sec:results} covering tests with experimentally known life-times,
comparison with $\alpha$-decay life-times and a large scale
systematics.
\section{Formal framework}
\label{sec:formal}
The starting point for the self-consistent microscopic description
is the SHF energy functional
\begin{equation}
E
=
E_\mathrm{Skyrme}(\rho,\tau,\mathbf{J};\mathbf{j},
\mbox{\boldmath$\sigma$},\mbox{\boldmath$\tau$};\chi)
\quad
\end{equation}
which is expressed in terms of a few local densities and currents
obtained as sums over single-particle wave functions: density $\rho$,
kinetic density $\tau$, spin-orbit density ${\bf J}$, current ${\bf
j}$, spin density ${\bf\sigma}$, and pair density $\chi$ where each
occurs twice, once for protons and once for neutrons. The zero-range
pairing functional is handled in a stabilized BCS ansatz to achieve
smooth transitions of the occupation numbers. For details see \cite{Erl08}.
There exist various parameterizations for the Skyrme functional.
In order to explore the possible sensitivity of the fission life-times
to the parameterization, we
{confine the survey}
to three
sufficiently different parameterizations:
SkP as a force with effective nucleon mass $m^*/m=1$
\cite{skp}, SkI3 as a fit which
{has very low mass $m^*/m=0.6$}
and maps the relativistic iso-vector
structure of the spin-orbit force \cite{ski3}, and Sly6 as a fit which
{has $m^*/m=0.7$}
and
includes information on isotopic trends and neutron matter
\cite{sly46}.
The computation of the fission path and the ingredients of the
corresponding collective Hamiltonian is detailed in
\cite{Fle05a,Klu08a}. We give a brief summary. The mean-field
equations are derived variationally from the given energy
functional. They are complemented by a quadrupole constraint to
generate the fission path thus reading
\begin{subequations}
\begin{eqnarray}
&&
\big[\hat{h}-\lambda\hat{Q}_{20}\big]|\Phi_q\rangle
=
\mathcal{E}|\Phi_q\rangle
\quad,\hspace{0.5cm} q =\langle\Phi_q|\hat{Q}_{20}|\Phi_q\rangle
\\
&&
\mathcal{V}(q)
=
E_\mathrm{Skyrme}(\rho_q,\tau_q,
\mathbf{J}_q,\chi_q)
\quad,
\label{eq:rawpot}
\end{eqnarray}
\end{subequations}
where $\hat{h}$ is the mean-field Hamiltonian and $\rho_q$
is the local density for the state $|\Phi_{q}\rangle$, similarly for
$\tau_q$, $\mathbf{J}_q$, and
$\chi_q$. The optimal fission path should, in fact, be
generated by adiabatic time-dependent Hartree-Fock
\cite{Rei87aR,Ska08a}. The quadrupole constraint is a plausible
and generally used approximation to that.
To compute the collective mass along the fission path, we need to
explore the dynamical response of the system to changing deformation,
commonly called self-consistent cranking.
To that end we use the collective-momentum operator $\hat{P}_{20}$
(and its collective momentum p) as
additional constraint. The $\hat{P}_{20}$ is deduced as generator
of deformation, the response to $\hat{P}_{20}$ creates a
momentum-dependent path and subsequently total energy, from which
the mass is finally obtained as second order term in collective momentum,
altogether
\begin{subequations}
\begin{eqnarray}
&&
\hat{P}_{20}:\qquad
\hat{P}_{20}|\Phi_q\rangle
\propto
\mathrm{i}\partial_q|\Phi_q\rangle
\quad,
\\
&&
\big[\hat{h}-\lambda\hat{Q}_{20}-\mu\hat{P}_{20}\big]
|\Phi_{qp}\rangle
=
\mathcal{E}|\Phi_{qp}\rangle
\quad,
\\
&&
B_{20}
=
\frac{1}{2}\partial_{p_{20}}^2E(\rho_{qp},...)\big|_{p=0}
\quad.
\end{eqnarray}
\end{subequations}
Analogously, the momentum of inertia $\Theta_x$ for rotations about
the $x$- and $y$-axis is computed by self-consistent cranking.
The potential $\mathcal{V}$ as given in eq. (\ref{eq:rawpot}) has to be
augmented by quantum corrections to account for angular-momentum
projection and spurious zero-point quadrupole motion \cite{Rei87aR}
yielding the net potential
\begin{eqnarray}
\mathcal{V}
\quad\longrightarrow\quad
V
&=&
\mathcal{V}
-
B_{20}\langle\hat{P}_{20}^2\rangle
-
\frac{1}{8\langle\hat{P}_{20}^2\rangle}\partial_q^2\mathcal{V}
-
\frac{\langle\hat{J}_x^2+\hat{J}_y^2\rangle}{2\Theta_x}
\quad.
\label{eq:zpecor}
\end{eqnarray}
Having determined the properly corrected collective potential
$V(\alpha_{20})$ and inverse collective mass $B(q)$, we
compute the fission life-time in the standard semi-classical fashion
using the WKB expressions for tunneling probability $W$ and repetition
time $T$:
\begin{eqnarray}
W
=
\exp{\Big(\!-2\!\int_b^cdq\sqrt{\frac{V(q)\!-\!E}{B_{20}}}\Big)}
\quad,\quad
\label{equ:period-koll}
T
&=&
\hbar\int_a^b\!dq\frac{1}{\sqrt{B_{20}(E\!-\!V(q))}}
\quad.
\end{eqnarray}
The integrals are evaluated by trapezoidal rule with separate handling
at the divergent $\sqrt{E-V}$ at the classical turning points. A
quantity which enters very critically these expressions is the
tunneling energy E, which is also the ground state energy of the
nucleus before fission. That is computed with great care in a fully
quantum-mechanical treatment of the collective quadrupole oscillations
using all five quadrupole degrees-of-freedom, for details of that part
of the calculations see \cite{Fle05a,Klu08a}.
{
It is to be noted that the fission path employs only axially symmetric
shapes. One knows from actinides that triaxial shapes can produce
barriers which are 0--2~MeV lower. The present results are thus to be
understood as an upper limit on barriers and
lifetimes.
Actinides show the famous double-humped barrier. The outer barrier
dissappears for SHE leaving only one, the formerly inner barrier
\cite{Bur04}. This barrier is related to still symmetric shapes while
asymmetry develops quickly when stepping down further towards fission.
The fact that we have only one barrier simplifies the survey of
fission properties of SHE.
}
\section{Results and discussion}
\label{sec:results}
\begin{SCfigure}[0.5]
\hspace*{0.5cm}{\epsfig{file=barriers-mass2-bw.eps,width=6.5cm}}
\caption{\label{fig:barriers-mass2-bw}
Example of potential energy (lower) and collective mass
(upper) along the axially symmetric fission path of
{$^{260}$Rf}
.
For the potential energy, we show the raw expectation value
$\mathcal{V}$ and the corrected one $V$, see eq.~(\ref{eq:zpecor}).
}
\end{SCfigure}
Fig.~\ref{fig:barriers-mass2-bw} shows collective potential and mass
along the fission path of
{$^{260}$Rf}
. The quantum corrections on the
potential (lower panel) reduce the barrier by an important amount
making 3 to 6 orders lower life-times than would be obtained with
the raw potential \cite{Rei87aR}.
{
One recognizes a steep increase of the energy at deformation
$\alpha_{20}\approx 0.6$ connected with a quick change of
hexadecapole momentum $\alpha_{40}$. The system moves from
one ridge (with high $\alpha_{40}$) to another ridge
(with low $\alpha_{40}$). The segregation of the fission landscape
into ridges is well known for actinides, see e.g. \cite{Moe01a},
still visible for that smaller SHE, and much smoothened with
further increasing system size.
}
The quadrupole mass (upper panel)
shows large fluctuations. High values are related to regions of
level crossings. The detailed pattern are quantitatively important and
can hardly be simulated by some constant collective mass.
\begin{SCfigure}[0.5]
{\epsfig{file=compare-half-lives-2-bw.eps,width=9.5cm}}
\caption{\label{fig:compare-half-lives-2-bw}
Fission life-times computed with three different Skyrme
parameterizations, as indicated, and compared with
data from
\protect\cite{expRf254,expRf256,expRf258,expSg262,expSg266,expHs266}.
}
\end{SCfigure}
As a benchmark, we compare in fig.~\ref{fig:compare-half-lives-2-bw}
with experimental results available at the lower edge of
SHE. There
are large differences in the predictions from the various forces. But
the experimental values are nicely bracketed by the band of
predictions and the trends within each force compare well with
experiment. The deviations of $\pm 4$ orders of magnitude look
huge. But they seem bearable in view of the extremely delicate balance
in computing the tunneling rates.
\begin{figure}
\centerline{\epsfig{file=system-barriers-bw.eps,width=\linewidth}}
\caption{\label{fig:system-barriers-bw}
Fission barriers for a broad variety of super-heavy elements
and for three Skyrme parameterizations as indicated.
}
\end{figure}
\begin{figure}
\centerline{\epsfig{file=system-forces-bw.eps,width=\linewidth}}
\caption{\label{fig:system-lifetimes-bw}
Fission life-times
for a broad variety of super-heavy elements
and for three Skyrme parameterizations as indicated.
}
\end{figure}
We are now going to present systematics of results all over the
landscape of SHE where we consider all elements which are found to be
stable against immediate nucleon emission at ground state and along
the whole fission path. Fig.~\ref{fig:system-barriers-bw} shows the
fission barriers. {They agree} in value and systematics with
{the previous survey of barriers \cite{Bur04}}.
One sees two islands of enhanced fission
stability (high barriers), one of deformed SHE around Z/N=104/152 and the other
of spherical SHE around 120/184. The magic neutron number N=184 is
also clearly marked while no unambiguous sign of a magic proton number
can be found in that upper island \cite{Ben01a}. All three forces
agree in predicting a ridge of very low barriers between the upper and
the lower stability island.
{Barriers however, are mainly indications of stability.
The full information on path, {mass, and quantum corrections}
is required to calculate life-times.}
Fig.~\ref{fig:system-lifetimes-bw} shows the corresponding fission
life-times. The difference in barrier heights from 0 to 12 MeV
translates to life-times from almost immediate decay to $10^{16}$ s,
demonstrating again the enormous sensitivity of fission life-times to
any ingredient in its computation. The long lived SHE are found in the
two islands of stability. Practically immediate decay appears in the
ridge of instability between the islands. The variance of the
predictions is moderate in the islands (see also
fig.~\ref{fig:compare-half-lives-2-bw}) and significantly larger in
the unstable region. The latter region is not only unstable but also
very hard to control theoretically.
\begin{figure}
\centerline{\epsfig{file=system-alpha-bw.eps,width=\linewidth}}
\caption{\label{fig:system-alphalifetimes-bw}
$\alpha$-decay lifetime
for a broad variety of super-heavy elements
and for three Skyrme parameterizations as indicated.
}
\end{figure}
The competing decay channel for many SHE is $\alpha$-decay. We have
computed the $\alpha$ life-times from the $Q_\alpha$ reaction energies
using a recently improved recipe \cite{Sam07} based on the Viola
systematics. Fig.~\ref{fig:system-alphalifetimes-bw} shows the results. The
$\alpha$ life-times change much more smoothly. There are no shell
effects visible and the variance of life-times all over the landscape
is very much smaller. Comparing with
fig.~\ref{fig:system-lifetimes-bw}, one sees that $\alpha$-decay is
indeed the dominating decay channel in the islands of
stability. Stepping down from the upper island, one sees a crossover
of $\alpha$-decay to spontaneous fission at about Z=112 or 110, much
in accordance with the experimental findings so far, see
e.g. \cite{exp116}.
The relation between $Q_\alpha$ value and $\alpha$ life-time is
comparatively simple
{when adopting the Viola systematics \cite{Sam07}}.
It is interesting to ponder whether one could
hope for a similarly simple direct connection between fission barrier
and fission life-time. A comparison of figure
\ref{fig:system-barriers-bw} with \ref{fig:system-lifetimes-bw} shows
that the very gross trends may look similar.
{
We have checked whether one could establish a simple relation between
barriers and lifetimes. This turned out to be impossible.
}
Fission life-times are very subtle quantities which depend on
{more ingredients than just barriers alone.}
\begin{SCfigure}[0.5]
{\epsfig{file=trends-halflifes-bw.eps,width=7.5cm}}
\caption{\label{fig:trends-halflifes-bw}
Fission life-times and $\alpha$-decay life-times
for $\quad$
elements along the decay chain from $Z/N=118/176$
computed with three different Skyrme
parameterizations, as indicated, and compared with experimental
data from \protect\cite{exp118}.
}
\end{SCfigure}
There exist a few fission life-times along the decay channel from the
upper island of stability. Fig.~\ref{fig:trends-halflifes-bw} compares
computed life-times with data for fission and $\alpha$-decay. The
$\alpha$ life-times are well reproduced by the calculations, but the
fission life-times are dramatically underestimated towards the lower
end of the chain. The reason for that is still unclear. The mismatch
calls for further investigations.
\section{Conclusions}
Fission life-times have been computed using the Skyrme-Hartree-Fock
method. The fission path has been generated with a quadrupole
constraint where only axially symmetric deformations are considered.
For the corresponding collective mass, we use self-consistent
cranking. The quantum corrections to the collective potential
(angular momentum projection, vibrational zero-point energy) are
properly taken into account. The fission life-time is computed for
thus given potential and mass by the WKB approximation, while the
ground state energy, which is at the same time the entrance energy for
fission, is computed quantum mechanically. Results have been produced
for three different Skyrme forces, SkP, SLy6 and SkI3, to explore the
sensitivity to the parameterization. For comparison, we have also
computed the $\alpha$-decay life-times using the Viola systematics.
A first test was performed by comparing with known fission life-times
in the lower region of super-heavy elements, Z=104-108. The
theoretical results gather around the experimental values with
deviations of $\pm 2$ orders of magnitude. That can be called a
promising agreement, the more so, as the isotropic trends are well
reproduced. The second test went for fission and $\alpha$-decay
life-times in the $\alpha$-decay chain from 118/176. The
$\alpha$-decay life-times are well reproduced. But it turned out that
the fission life-times are grossly underestimated by all three forces
towards the lower end of the chain at 112/170. The reasons for that
mismatch are not yet clear. It is to be reminded that these
super-heavy nuclei have extremely soft deformation energy surfaces
with several shape isomers which complicates the determination of the
ground state.
Notwithstanding that open problem, we have produced a systematic
survey of fission barriers, fission life-times and $\alpha$-decay
life-times for all conceivable super-heavy elements in the range
Z=104-124 and N=150-190. Barriers and fission life-times show nicely
the islands of shell stabilization around 104/152 and 120/184. All
forces predict a band of fission instability being orthogonal to the
line connecting the two islands and crossing that around
112/166. There seems to be no path which could connect the 120/184
region with the lower island while avoiding fission.
While the fission life-times show dramatic variation over the chart of
super-heavy elements (from instability to 10$^{16}$s), the
$\alpha$-decay times vary gently with small overall changes and
without visible shell effects. The general crossover from
$\alpha$-decay to fission along the decay chains from the upper island
is qualitatively reproduced by all three forces in the survey.
Altogether, the results are promising and challenging at the same
time. They call for further investigations exploring more
systematically the sensitivity to the Skyrme parameterization and
improving the description in the very soft transitional region around
112/166.
\section*{Acknowledgments}
We thank the regional computing center of the university
Erlangen-N\"urnberg for generous supply of computer time for the
demanding calculations. The work was supported by the BMBF under
contracts 06 ER 808.
|
\section{Introduction}
Recently, the question of when a pseudo-Riemannain manifold can be locally
characterized by its scalar polynomial curvature invariants
constructed from the Riemann tensor and its covariant derivatives has been addressed.
In \cite{inv} it was shown that in four dimensions (4d) a Lorentzian spacetime
metric is either $\mathcal{I}$-non-degenerate, and hence locally
characterized by its scalar polynomial curvature invariants,
or is a degenerate Kundt spacetime \cite{Kundt}.
Therefore, the degenerate Kundt spacetimes are the only spacetimes
in 4d that are {\em not} $\mathcal{I}$-non-degenerate, and their
metrics are the only metrics not uniquely determined by their curvature
invariants.
In the proof of the $\mathcal{I}$-non-degenerate theorem in
\cite{inv} it was necessary, for example, to determine for which Segre types
for the Ricci tensor the spacetime is
$\mathcal{I}$-non-degenerate. By analogy, in higher dimensions it is useful to
utilize the bivector
formalism for the Weyl tensor.
Indeed, by defining the Weyl bivector operator
in higher dimensions \cite{bivect}
and making use of the alignment theory \cite{class},
it is possible to algebraically classify any tensor (including the
Weyl tensor utilizing the eigenbivector problem) in a Lorentzian
spacetime of arbitrary dimensions, which has proven very useful
in contemporary theoretical physics.
The link between a metric (or rather, when two metrics can be distinguished up to diffeomorphisms)
and the curvature tensors is provided through \emph{Cartan's equivalence principle}.
An exact statement of this is given in \cite{exsol}; however, the idea is that knowing the
Riemann curvature tensor, and its covariant derivatives, with respect to a fixed
frame
determines the metric up to isometry. It should be pointed out that it may be difficult to actually \emph{construct} the metric, however, the metric is determined in principle. More precisely, given a point, $p$, a frame $E_\alpha$, and the corresponding Riemann tensor $R$ and its covariant derivatives $\nabla R, \dots, \nabla^{(k)}R,\dots $, then if there exists a map such that:
\beq
(p,E_\alpha)&\mapsto & (\bar{p},\bar{E}_\alpha), \quad \text{and} \nonumber \\
(R_{\alpha\beta\gamma\delta},R_{\alpha\beta\gamma\delta;\mu_1},...,R_{\alpha\beta\gamma\delta;\mu_1...\mu_k}) & \mapsto & (\bar{R}_{\alpha\beta\gamma\delta},\bar{R}_{\alpha\beta\gamma\delta;\mu_1},...,\bar{R}_{\alpha\beta\gamma\delta;\mu_1...\mu_k}), \nonumber
\eeq
then there exists an isometry $\phi$ such that $\phi(p)=\bar{p}$ (which induces the abovementioned
map). Therefore, if we know the components of the Riemann tensor and its covariant derivatives
with respect to a fixed frame, then the geometry and, in principle, the metric is determined.
The equivalence principle consequently manifests the connection between the curvature tensors
and the spacetime metric. The question of whether we can, at least in principle, reconstruct the metric
from the invariants thus hinges on the question whether we can reconstruct the curvature
tensors from its scalar polynomial curvature invariants, indicated with a question mark in the following figure:
\[ \begin{array}{cl} \boxed{\text{Metric}, g_{\mu\nu}} \\
{\Updownarrow} & \hspace{-1cm} \leftarrow\text{Equivalence principle} \\
\boxed{\text{Curvature tensors}} \\
\Downarrow \quad \Uparrow ? \\
\boxed{\text{Invariants}}
\end{array}
\]
It is to address this question, the curvature operators are particularly useful.
In this paper we continue the study of pseudo-Riemannain manifolds and their local
scalar polynomial curvature invariants. In principle, we
are interested
in the case of arbitrary dimensions and all signatures, although we are primarily interested
from a physical point of view in the Lorentzian case and most illustrations will be done in
4d. We also briefly
discuss the neutral
signature case.
First, we shall discuss the concept of a differentiable manifold being characterised by
its scalar invariants in more detail.
Suppose we consider the set of invariants as a function of the metric
and its derivatives: $\mf{I}:g_{\mu\nu}\mapsto \mathcal{I}$
(the $\mf{I}$-map);
we are then interested in under what circumstances, given a set
of invariants $\mathcal{I}_0$,
this function has an inverse
$\mf{I}^{-1}(\mathcal{I}_0)$.
For a metric which is
$\mathcal{I}$-non-degenerate the invariants characterize the
spacetime uniquely, at least locally, in the space of (Lorentzian)
metrics, which means that we can therefore distinguish such metrics
using their curvature invariants and, hence, $\mf{I}^{-1}(\mathcal{I}_0)$ is unique
(up to diffeomorphisms).
We then further discuss curvature operators (and introduce an operator calculus)
and their relationship to spacetime invariants.
Any even-ranked tensor can be
considered as an operator, and there is a natural matrix representation of the
operator ${\sf T}$.
In particular, for a curvature operator
(such as the zeroth order curvature operator, $\mathcal{R}_0$)
we can consider an eigenvector ${\sf
v}$ with eigenvalue $\lambda$; i.e., ${\sf T}{\sf v}=\lambda{\sf
v}$. This allows us to introduce the important notion of {\emph {diagonalisability}}.
In addition, we then
introduce the notion of {\emph {analytic metric extension}},
which is a complex analytic continuation of a
real metric
under more general coordinate transformations than the real diffeomorphisms
which results in a real bilinear form but which does not necessarily
preserve the metric
signature.
We shall show that if a space, $(\mathcal{M},g_{\mu\nu})$, of any signature can be analytically
continued (in this sense) to a Riemannian space (of Euclidean signature), then that
spacetime is characterised by its invariants. Moreover, if a spacetime is not characterised by its
invariants, then there exists no such analytical continuation of it to a Riemannian space.
Finally, we present a summary of the Lorentzian signature case.
All of the results presented here are
illustrated in the 4d case. We also
discuss the 4d pseudo-Riemannian case of neutral
signature $(--++)$ (NS space).
\subsection{Metrics characterised by their invariants}
Consider the continuous metric deformations defined as follows \cite{inv}.
\begin{defn}
\noindent For a spacetime $(\mathcal{M},g)$, a (one-parameter) \emph{metric deformation}, $\hat{g}_\tau$, $\tau\in [0,\epsilon)$, is a family of smooth metrics on $\mathcal{M}$ such that
\begin{enumerate}
\item{} $\hat{g}_\tau$ is continuous in $\tau$,
\item{} $\hat{g}_0=g$; and
\item{} $\hat{g}_\tau$ for $\tau>0$, is not diffeomorphic to $g$.
\end{enumerate}
\end{defn}
\noindent For any given spacetime $(\mathcal{M},g)$ we define the set of all scalar polynomial curvature invariants
\begin{equation}
\mathcal{I}\equiv\{R,R_{\mu\nu}R^{\mu\nu},C_{\mu\nu\alpha\beta}C^{\mu\nu\alpha\beta}, R_{\mu\nu\alpha\beta;\gamma}R^{\mu\nu\alpha\beta;\gamma}, R_{\mu\nu\alpha\beta;\gamma\delta}R^{\mu\nu\alpha\beta;\gamma\delta},\dots\} \,. \nonumber
\end{equation}
\noindent Therefore, we can consider the set of invariants as a function of the metric
and its derivatives. However, we are interested in to what extent,
or under what circumstances, this function has an inverse.
More precisely, define the function $\mf{I}:g_{\mu\nu}\mapsto \mathcal{I}$
(the $\mf{I}$-map) to be the function that calculates the set of
invariants $\mathcal{I}$ from a metric $g_{\mu\nu}$. Given a set
of invariants, $\mathcal{I}_0$, what is the nature of the (inverse)
set $\mf{I}^{-1}(\mathcal{I}_0)$?
In order to address this question, we first need to introduce some terminology \cite{inv}.
\begin{defn}
Consider a spacetime $(\mathcal{M},g)$ with a set of invariants
$\mathcal{I}$. Then, if there does not exist a metric deformation
of $g$ having the same set of invariants as $g$, then we will call
the set of invariants \emph{non-degenerate}. Furthermore, the
spacetime metric $g$, will be called
\emph{$\mathcal{I}$-non-degenerate}.
\end{defn}
This implies that for a metric which is
$\mathcal{I}$-non-degenerate the invariants characterize the
spacetime uniquely, at least locally, in the space of (Lorentzian)
metrics. This means that these metrics are characterized by their
curvature invariants and therefore we can distinguish such metrics
using their invariants. Since scalar curvature invariants are
manifestly diffeomorphism-invariant we can thereby avoid the
difficult issue whether a diffeomorphism exists connecting two
spacetimes.
In the above sense of $\mathcal{I}$-non-degeneracy,
a spacetime is completely characterized by its
invariants and there is only one (non-isomorphic) spacetime with this set of
invariants (at least locally in the space of metrics, in the sense above). However, we may also want to use the notion of characterization by
scalar invariants in a different (weaker) sense
(see below).
In order to emphasise the difference between such cases, we will say that
$\mathcal{I}$-non-degenerate metrics are characterised by their invariants \emph{in the strong
sense}, or in short, \emph{strongly} characterised by their invariants
(this is the general sense, and if the clarifier
`strong sense' is omitted, this meaning is implied).
However, in the definition \ref{def:char} (below)
it is useful to regard a pseudo-Rie\-mannian manifold as being
characterised by its invariants if the curvature tensors
(i.e., the Riemann tensor and all of its covariant derivatives)
can be reconstructed from a knowledge all of the scalar
invariants.
It is clear that $\mathcal{I}$-non-degenerate metrics fall into this class, but
another important class of metrics will also be included. For example, (anti-)de Sitter space is
maximally symmetric and consequently has only one independent curvature component, and also
only one independent curvature invariant (namely the Ricci scalar). Therefore, we can determine
the curvature tensors of (anti-)de Sitter space by knowing its invariants. In our definition,
(anti-)de Sitter space is not $\mathcal{I}$-non-degenerate but it will be
regarded as being characterised by its
invariants.
Therefore, we shall say that metrics that are
characterised by their invariants but are not $\mathcal{I}$-non-degenerate, are \emph{weakly}
characterised by their invariants. A common factor for metrics being characterised by their
invariants is that the curvature tensors, and hence the spacetime, will inherit certain
properties of the invariants. In particular, as we will see later, all CSI spacetimes being
characterised by their invariants will be automatically be locally homogeneous; however, not
all of them are $\mathcal{I}$-non-degenerate, like (anti-)de Sitter space.
The definition we will adopt is closely related to the familiar problem of diagonalising
matrices. We will use curvature operator to relate our problem to linear algebra, and
essentially, we will say that a spacetime is characterised by its invariants if we can
diagonalise its curvature operators. In order to make this definition a bit more rigourous, we
need to introduce some formal definitons.
\section{Definition}
Consider an even-ranked tensor $T$. By raising or lowering the indices appropriately, we can construct the tensor with components $T^{\alpha_1...\alpha_k}_{\phantom{\alpha_1...\alpha_k}\beta_1...\beta_k}$. This tensor can be considered as an operator (or an endomorphism) mapping contravariant tensors onto contravariant tensors
\[ {\sf T}: V\mapsto V,\]
where $V$ is the vectorspace $V=(T_pM)^{\otimes k}(\equiv\bigotimes_{i=1}^kT_pM)$, or, if the tensor possess
index symmetries, $V\subset(T_pM)^{\otimes k}$. Here, since $V$ is a vectorspace, there exists a set of basis vectors ${\bf e}_{I}$ so that $V=\mathrm{span}\{e_{I}\}$. Expressing ${\sf T}$ in this basis, we can write the operator in component form: $T^I_{~J}$. This means that we have a natural matrix representation of the operator ${\sf T}$.
Note that we can similarly define a dual operator ${\sf T}^*:V^*\mapsto V^*$ where
$V^*=(T^*_pM)^{\otimes k}$, mapping covariant tensors onto covariant tensors.
In component form the dual operator is $T^{*I}_{\phantom{*}~J}$ which in matrix
form can be seen as the transpose of ${\sf T}$; i.e., ${\sf T}^*={\sf T}^T$. Consequently,
there is a natural isomorphism between the operator ${\sf T}$ and its dual. We can also consider operators mapping mixed tensors onto mixed tensors.
For an odd-ranked tensor, $S$, we can also construct an operator: however, this time we need to
consider the tensor product of $S$ with itself, $T=S\otimes S$, which is of even rank. We
can also construct an operator with two odd-ranked tensors $S$ and $S'$: $T=S\otimes S'$. In
this way $ T$ is even-ranked and we can lower/raise components appropriately.
Now, consider an operator. If $T^I_{~J}$ are the components in a certain basis, then this as
a natural matrix representation. This is an advantage because now we can use standard
results from linear algebra to estabish some related operators.
The utility of these operators when it comes to invariants is obvious since any
curvature invariant is always an invariant constructed from some operator. For example, the Kretchmann
invariant, $R^{\alpha\beta\mu\nu}R_{\alpha\beta\mu\nu}$, is trivially (proportional to) the
trace of the operator ${\sf
T}=(T^A_{~B})=\left(R^{\alpha_1\alpha_2\alpha_3\alpha_4}R_{\beta_1\beta_2\beta_3\beta_4}\right)$.
Similarly, an arbitrary invariant,
$T^{\alpha_1\alpha_2\cdots\alpha_k}_{\phantom{\alpha_1\alpha_2\cdots\alpha_k}\alpha_1\alpha_2\cdots\alpha_k}$,
is the trace of the operator ${\sf
T}=(T^A_{~B})=\left(T^{\alpha_1\alpha_2\cdots\alpha_k}_{\phantom{\alpha_1\alpha_2\cdots\alpha_k}\beta_1\beta_2\cdots\beta_k}\right)$.
Therefore, it is clear that we can study the polynomial curvature invariants by studing the
invariants of curvature operators.
The archetypical example of a curvature operator is the Ricci operator, ${\sf R}=(R^\mu_{~\nu})$, which maps vectors onto vectors; i.e.,
\[ {\sf R}: T_pM\mapsto T_pM.\]
All of the Ricci invariants can be constructed from the invariants of this operator, for example, $\Tr({\sf R})$ is the Ricci scalar and $\Tr({\sf R}^2)=R_{\mu\nu}R^{\mu\nu}$.
Another commonly used operator is the Weyl operator, ${\sf C}$, which maps bivectors onto bivectors:
\[ {\sf C}:\wedge^2T_pM\mapsto \wedge^2T_pM.\]
The Weyl invariants can also be constructed from this operator by considering traces of powers of ${\sf C}$: $\Tr({\sf C}^n)$.
\section{Eigenvalues and projectors}
For a curvature operator, ${\sf T}$, consider an eigenvector ${\sf
v}$ with eigenvalue $\lambda$; i.e., ${\sf T}{\sf v}=\lambda{\sf
v}$. Note that the symmetry group
$SO(d,n)$ of the spacetime (of signature $(d,n)$), is naturally imbedded through the tensor
products $(T_pM)^{\otimes k}$. By considering the eigenvalues of $\T$ as solutions of the
characteristic equation: \[ {\mathrm{det}}(\T -\lambda{\sf 1})=0, \] which are
$GL(k^n,\mb{C})$-invariants. Since the orthogonal group $SO(d,n-d)$, using the tensor
product, acts via a representation $\Gamma: SO(d,n-d)\mapsto GL(k^n)\subset GL(k^n,\mb{C})$,
Hence, \emph{the eigenvalue of a curvature operator is an $O(d,n-d)$-invariant curvature scalar}.
Therefore, curvature operators naturally provide us with a set of
curvature invariants (not necessarily polynomial invariants but derivable from them)
corresponding to the set of distinct eigenvalues: $\{\lambda_A
\}$. Furthermore, the set of eigenvalues are uniquely determined
by the polynomial invariants of ${\sf T}$ via its characteristic
equation. The characteristic equation, when solved, gives us the
set of eigenvalues, and hence these are consequently determined by
the invariants.
We can now define a number of associated curvature operators. For
example, for an eigenvector ${\sf v}_A$ so that ${\sf T}{\sf
v}_A=\lambda_A{\sf v}_{A}$, we can construct the annihilator
operator:
\[ {\sf P}_A\equiv ({\sf T}-\lambda_{A}{\sf 1}).\]
Considering the Jordan block form of ${\sf T}$, the eigenvalue ${\lambda_A}$ corresponds to a set of Jordan blocks. These blocks are of the form:
\[ {\sf B}_A=\begin{bmatrix}
\lambda_A & 0 & 0& \cdots & 0 \\
1 & \lambda_A & 0& \ddots & \vdots \\
0 & 1 &\lambda_A& \ddots & 0 \\
\vdots & \ddots &\ddots& \ddots & 0 \\
0 & \hdots &0 & 1 & \lambda_A
\end{bmatrix}.\]
There might be several such blocks corresponding to an eigenvalue
$\lambda_A$; however, they are all such that $({\sf
B}_A-\lambda_A{\sf 1})$ is nilpotent and hence there exists an
$n_{A}\in \mathbb{N}$ such that ${\sf P}_A^{n_A}$ annihilates the
whole vector space associated with the eigenvalue $\lambda_A$.
This implies that we can define a set of operators $\widetilde{\bot}_A$ with eigenvalues $0$ or $1$ by considering the products
\[ \prod_{B\neq A}{\sf P}^{n_B}_B=\Lambda_A\widetilde{\bot}_A,\]
where $\Lambda_A=\prod_{B\neq A}(\lambda_A-\lambda_B)^{n_B}\neq 0$
(as long as $\lambda_B\neq \lambda_A$ for all $B$). Furthermore, we can now define
\[ \bot_A\equiv {\sf 1}-\left({\sf 1}-\widetilde{\bot}_A\right)^{n_A} \]
where $\bot_A$
is a \emph{curvature projector}. The set of all such curvature
projectors obeys:
\beq {\sf 1}=\bot_1+\bot_2+\cdots+\bot_A+\cdots,
\quad \bot_A\bot_B=\delta_{AB}\bot_A.
\eeq We can use these
curvature projectors to decompose the operator ${\sf T}$:
\beq
{\sf T}={\sf N}+\sum_A\lambda_A\bot_A. \label{decomp}
\eeq
The
operator ${\sf N}$ thus contains all the information not
encapsulated in the eigenvalues $\lambda_A$. From the Jordan form
we can see that ${\sf N}$ is nilpotent; i.e., there exists an
$n\in\mathbb{N}$ such that ${\sf N}^n={\sf 0}$. In particular, if
${\sf N}\neq 0$, then ${\sf N}$ is a negative/positive
boost weight operator which can be used to lower/raise the
boost weight of a tensor \cite{class,bivect}.
We also note that
\[ {\sf N}=\sum_A {\sf N}_A, \quad {\sf N}_A\equiv \bot_A {\sf N}\bot_A.\]
Consequently, we have the orthogonal decomposition:
\beq
{\sf T}=\sum_A\left({\sf N}_A+\lambda_A\bot_A\right). \label{decomp2}
\eeq
Note that we can achieve this decomposition by just using operators and their invariants.
In linear algebra we are accustomed to the concept of diagonalisable matrices.
Similarly, we will call an operator \emph{diagonalisable} if
\beq
{\sf T}=\sum_A\lambda_A\bot_A.
\label{Tdiag}\eeq
In differential geometry we are particularly interested in quantities like the Riemann
curvature tensor, $R$. Now, the Riemann tensor naturally defines two curvature operators:
namely, the Ricci operator ${\sf R}$, and the Weyl operator ${\sf C}$.
Let us, at a point $p$, define the tensor algebra (or tensor concomitants) of 0th order curvature operators,
$\mathcal{R}_0$,\footnote{Here, $\mathcal{R}$ refers to the Riemann tensor and index $0$ refers to the number of covariant derivatives. } as follows:
\begin{defn}
The set ${\mathcal R}_0$ is a $\mathbb{C}$-linear space of operators (endomorphisms) which satisfies the following properties:
\begin{enumerate}
\item{} Identity: the metric tensor, as an operator, is in $\mathcal{R}_0$.
\item{} the Riemann tensor, as an operator, is in $\mathcal{R}_0$.
\item{} if ${\sf T}\in\mathcal{R}_0$, then any tensor contraction and any eigenvalue of ${\sf T}$ is also in $\mathcal{R}_0$.
\item{} ``Square roots'': if $\tilde{{\sf T}}\in\mathcal{R}_0$ and $\tilde{ T}={ T}\otimes{ T}$ (as tensors) where $T$ is of even rank, then ${\sf T}\in \mathcal{R}_0$
\item{} any operator that can be considered as a tensor product or expressible as functions of elements in $\mathcal{R}_0$ is also in $\mathcal{R}_0$.
\end{enumerate}
An element of $\mathcal{R}_0$, will be referred to as a \emph{curvature operator of order 0}.
\end{defn}
Similarly, allowing for contractions of the covariant derivative of the Riemann tensor, $\nabla R$,
we can define the tensor algebra of 1st order curvature operators, $\mathcal{R}_1$, etc. If
we allow for arbitrary number of derivatives, we will simply call it $\mathcal{R}$. Hence\footnote{As a $\mathbb{C}$-linear vector space, $\mathcal{R}_k$ will be finite dimensional, while $\mathcal{R}$ is infinite dimensional. However, in practice, it is sufficient to consider $\mathcal{R}_k$ for sufficiently large $k$ due to the fact that a finite number of derivatives of the Riemann tensor are sufficient to determine the spacetime up to isometry.}:
\[ \mathcal{R}_0\subset\mathcal{R}_1\subset \cdots \subset \mathcal{R}_k\subset \mathcal{R}.\]
For example, the following tensor:
\[ R^{\mu\nu\alpha\beta}R_{\alpha\beta\rho\sigma}+45R^{\alpha\beta\gamma\delta;\mu\nu}R_{\alpha\beta\epsilon\eta}R^{\epsilon\eta}_{~\phantom{\epsilon\eta}\gamma\delta;\rho\sigma},\]
is a curvature tensor of order 2 and is consequently an element of $\mathcal{R}_2$.
Of particular interest is when all curvature operators can be determined using their invariants.
This means that there is preferred set of operators (namely the projectors) which acts as
a basis for the curvature operators. This is related to diagonalisability of operators, and hence:
\begin{defn}
$\mathcal{R}_k$ is called diagonalisable if there exists a set of projectors $\bot_A\in \mathcal{R}_k$ forming a tensor basis for $\mathcal{R}_k$; i.e., for every ${\sf T}\in \mathcal{R}_k$,
\[{\sf T}=T^{AB...D}\bot_A\otimes\bot_B\otimes\cdots\otimes\bot_D.\]
\end{defn}
Note that the components may be complex. If an operator is diagonalisable, we can, by solving the characteristic equation,
determine the expansion eq. (\ref{Tdiag}). This enables us to reconstruct the operator itself.
In this sense, the operator would be characterised by its invariants. Hence, we have the following definition:
\begin{defn} \label{def:char}
A space $(\mathcal{M},g_{\mu\nu})$ is said to be \emph{characterised by its invariants} iff, for every $p\in \mathcal{M}$, the set of curvature operators, $\mathcal{R}$, is diagonalisable (over $\mb{C}$).
\end{defn}
As pointed out earlier we can further subdivide this category into spaces that are characterised by their invariants in a strong or weak sense. If, in addition, the space $(\mathcal{M},g_{\mu\nu})$ is $\mathcal{I}$-non-degenerate, then the space is characterised by its invariants in the \emph{strong} sense; otherwise its said to be characterised by its invariants in the \emph{weak} sense.
A word of caution: note that even though we know all the operators of a spacetime that is characterised by
its invariants, we do not necessarily know the frame. In particular, by raising an index of the metric
tensor, $g_{\mu\nu}$, we get, irrespective of the signature, $g^{\mu}_{~\nu}=\delta^{\mu}_{~\nu}$. This
means that for a Lorentzian spacetime we would lose the information of which direction is time. However,
there might still be ways of determining which is time by inspection of the invariants (but this is no
guarantee). If, say, an invariant can be written as $I=r^{\alpha}r_{\alpha}$, and $I<0$, then clearly, $r^\alpha$ is timelike and, consequently, can be used as time. On the other hand, an example of when this cannot be done is the following case, in $d$ dimensions, when the only non-zero invariants
are the 0th order Ricci invariants:
\[ \Tr({\sf R})=R^{\mu}_{~\mu}=(d-1)\lambda, \quad \Tr({\sf
R}^n)=R^{\mu_1}_{~\mu_2}R^{\mu_2}_{~\mu_3}\cdots R^{\mu_{n}}_{~\mu_1}=(d-1)\lambda^n.\] Then, if this
spacetime is characterised by its invariants, we get: \[ {\sf R}=\diag(0,\lambda,\lambda,\cdots,\lambda)\]
However, there is no information in the invariants if the direction associated with the $0$ eigenvalue is a
space-like or time-like direction. Consequently, in the definition above we have to keep in mind that
these spacetimes are characterised by their invariants, up to a possible ambiguity in which direction is
associated with time, and which is space\footnote{Note that in the Riemannian signature case, there is no
such ambiguity since all directions are necessarily space-like.}.
In the following we will reserve the word ``spacetime'' to the Lorentzian-signature case,
while ``Riemannian'' (space) will correspond to the Euclidean signature case. For the Riemannian case, the operators provide us with an extremely simple proof of the following theorem:
\begin{thm}\label{thm:riemannian}
A Riemannian space is always characterised by its invariants.
\end{thm}
\begin{proof}
In an orthonormal frame, the Riemannian metric is $g_{\mu\nu}=\delta_{\mu\nu}$. Therefore, the components
of an operator $T^I_{~J}$ will be the same as those of $T_{IJ}$, possibly up to an overall constant. Therefore, let us decompose into the symmetric,
$S_{IJ}$, and antisymmetric part $A_{IJ}$, of $T_{IJ}$:
\[ T_{IJ}=S_{IJ}+A_{IJ}, \quad S_{IJ}\equiv T_{(IJ)}, \quad A_{IJ}\equiv T_{[IJ]}.\]
Since the signature is Euclidean, the operators $S^{I}_{~J}$ and $A^{I}_{~J}$ will also be symmetric and
antisymmetric (as matrices), respectively. A standard result is that we can consequently always diagonalise ${\sf S}$:
\beq
{\sf S}=\diag(\lambda_1,\lambda_2,\cdots,\lambda_N);
\eeq
i.e., the operator ${\sf S}$ is diagonalisable. For ${\sf A}$:
\beq
{\sf A}=\mathrm{blockdiag}\left(\begin{bmatrix} 0 & a_1 \\ -a_1 & 0 \end{bmatrix},\begin{bmatrix} 0 & a_2 \\ -a_2 & 0 \end{bmatrix},\cdots,\begin{bmatrix} 0 & a_k \\ -a_k & 0 \end{bmatrix}, 0,\cdots, 0 \right)
\eeq
Consequently, ${\sf A}$ is diagonalisable also (with purely imaginary eigenvalues).
So for ${\sf T}\in\mathcal{R}$, then also ${\sf S}\in\mathcal{R}$ and ${\sf A}\in\mathcal{R}$, and therefore any curvature operator is characterised by its invariants. This proves the theorem.
\end{proof}
We note that this theorem also follows from a group-theoretical perspective
\cite{PTV,Procesi}; however, the
operators provide an alternative (and fairly straight-forward) proof.
Of course, it is known that a Lorentzian spacetime is not necessarily characterised by its invariants
in any dimension.\footnote{In the proof of theorem \ref{thm:riemannian} it can be seen that what goes wrong is that $S_{IJ}$ and $A_{IJ}$ are not necessarily symmetric or antisymmetric as operators $S^I_{~J}$ and $A^I_{~J}$. Therefore, the Jacobi canonical forms are really needed in this case.}
However, the spacetimes being characterised by their curvature invariants play a special role for certain
classes of metrics. For example, we have that:
\begin{prop}[VSI spacetimes]
Among the spacetimes of vanishing scalar curvature invariants (VSI), the only spacetimes being characterised by their curvature invariants (weakly or strongly) are locally isometric to flat space.
\end{prop}
Note that flat space is a characterised by its invariants only in the weak sense.
\begin{prop}[CSI spacetimes]
If a spacetime has constant scalar curvature invariants (CSI) and is characterised by its invariants (weakly or strongly), then it is a locally homogeneous space.
\end{prop}
These locally homogeneous spacetimes can be characterised by their invariants in either a weak or strong sense.
Regarding the question of which spacetimes are characterised by its invariants, in the sense defined above, we have that the following conjecture is true:
\begin{con}
If a spacetime is characterised by its invariants (weakly or strongly), then it is either
$\mathcal{I}$-non-degenerate, or of type D to all orders; i.e., type D$^k$.
\end{con}
There are several results that support this conjecture. First, it seems reasonable that all
$\mathcal{I}$-non-degenerate metrics are (strongly) characterised by their invariants. In 4
dimensions, this seems clear from the results of \cite{inv}. Furthermore, that type D$^k$ spacetimes
are also characterised by its invariants (in a weak sense necessarily) follows from the results of \cite{shortinv}. However, it
remains to prove that these are all.
We should also point out a sublety in the definition of what we mean by characterised by the invariants.
Now a class of metrics that, in general, does not seem to be characterised by their curvature invariants
is the subclass of Kundt metrics \cite{inv,Kundt}:
\[ \d s^2=2\d u[\d v+v^2H(u,x^k)\d u+vW_i(u,x^k)\d x^i]+g_{ij}(u,x^k)\d x^i \d x^j. \]
For sufficiently general functions $H$, $W_i$, and $g_{ij}$, these metrics
will not be characterised by their invariants as defined above because they will be of type II
to all orders (in three dimensions, see the example below). However, the invariants still enable
us to reconstruct the metric. In some sense, the metric above is the simplest metric with this set of
invariants (however, it is not characterised by its invariants as defined above!).
\paragraph{Example: 3D degenerate Kundt metrics.} Let use study the details of the above example
in three dimensions, but generalising to the full class of degenerate Kundt metrics, which can be written:
\[ \d s^2=2\d u[\d v+H(v,u,x)\d u+W(v,u,x)\d x]+\d x^2, \]
where $H=v^2H^{(2)}(u,x)+vH^{(1)}(u,x)+H^{(0)}(u,x)$ and $W=vW^{(1)}(u,x)+W^{(0)}(u,x)$.
In 3D the Weyl tensor is zero and so we only need to consider the Ricci tensor which, using
the coordinate basis $(u,v,x)$, can be written in operator form:
\beq
{\sf R}=\begin{bmatrix}
\lambda_1 & 0 & 0 \\
R_{uu} & \lambda_1 & R_{ux} \\
R_{ux} & 0 & \lambda_3
\end{bmatrix}.
\eeq
The eigenvalues are $\lambda_1, \lambda_1, \lambda_3$; consequently,
if $\lambda_1\neq \lambda_3$ the Segre type is $\{21\}$ or $\{(1,1)1\}$. Unfortunately,
the coordinate basis is not the canonical Segre basis so the form is not manifest. However,
the projection operators are frame independent operators so let us find the various projectors
in this case. Since $\lambda_1$ has multiplicity 2, (while $\lambda_3$ has multiplicity 1) the
operator $({\sf R}-\lambda_1{\sf 1})^2$ must be proportional to the projection operator $\bot_3$. By ordinary matrix multiplication we get:
\beq
({\sf R}-\lambda_1{\sf 1})^2=\begin{bmatrix}
0 & 0 & 0 \\
R_{ux}^2 & 0 & (\lambda_3-\lambda_1)R_{ux} \\
(\lambda_3-\lambda_1)R_{ux} & 0 &(\lambda_3-\lambda_1)^2
\end{bmatrix}.
\eeq
The proportionality constant is $(\lambda_3-\lambda_1)^2$. Consequently:
\[ \bot_3=\begin{bmatrix}
0 & 0 & 0 \\
\tfrac{R_{ux}^2}{(\lambda_3-\lambda_1)^2} & 0 & \tfrac{R_{ux}}{(\lambda_3-\lambda_1)} \\
\tfrac{R_{ux}}{(\lambda_3-\lambda_1)} & 0 & 1
\end{bmatrix}.
\]
We note that the projection operator is not diagonal in the coordinate basis: however, we can easily verify
that $\bot_3^2=\bot_3$.
There are only two projection operators in this case, so $\bot_1$ can be calculated
using $\bot_1={\sf 1}-\bot_3$. Alternatively, we can define
$\widetilde{\bot}_1\equiv(\lambda_1-\lambda_3)^{-1}({\sf R}-\lambda_3{\sf 1})$ so that the projection
operator is given by $\bot_1={\sf 1}-({\sf 1}-\widetilde{\bot}_1)^2$. Using either
way to calculate $\bot_1$, we obtain:
\[ \bot_1=\begin{bmatrix}
1 & 0 & 0 \\
-\tfrac{R_{ux}^2}{(\lambda_3-\lambda_1)^2} & 1 & -\tfrac{R_{ux}}{(\lambda_3-\lambda_1)} \\
-\tfrac{R_{ux}}{(\lambda_3-\lambda_1)} & 0 & 0
\end{bmatrix}.
\]
To check which Segre type the Ricci tensor has we calculate the nilpotent operator given by the
expansion eq. (\ref{decomp}):
\beq
{\sf N}={\sf R}-\lambda_1\bot_1-\lambda_3\bot_3=\begin{bmatrix}
0 & 0 & 0 \\
R_{uu}-\tfrac{R_{ux}^2}{(\lambda_3-\lambda_1)} & 0 & 0 \\
0 & 0 & 0
\end{bmatrix}.
\eeq
Therefore, the metric is Segre type $\{21\}$ in general, while if $R_{uu}=\tfrac{R_{ux}^2}{(\lambda_3-\lambda_1)}$ then it is Segre type $\{(1,1)1\}$.
For the above metric we have
\beq
\lambda_1&=& 2H^{(2)}+\frac{1}{2}\frac{\partial W^{(1)}}{\partial x}-\frac 12 (W^{(1)})^2, \\
\lambda_3&=& \frac{\partial W^{(1)}}{\partial x}-\frac 12 (W^{(1)})^2.
\eeq
The invariants will only depend on $H^{(2)}$ and $W^{(1)}$,
and the invariants will determine these functions up to diffeomorphisms. However,
in general, this metric will be of Segre type $\{2 1\}$; therefore, even if $W^{(0)}=H^{(1)}=H^{(0)}$,
this metric will not be determined by the curvature invariants, as defined above.
\section{The operator calculus}
Since all of the results so far are entirely point-wise, it is also an advantage to consider
the operator calculus; i.e. derivatives of operators. Most useful for our purposes is the Lie derivative.
The Lie derivative preserves the order and type of tensors and is thus particularly useful.
Consider a vector field ${\mbold\xi}$ defined on a neighbourhood $U$. We must assume that \emph{the operator decomposition, (\ref{decomp2}), does not change over $U$.} This assumption is essential in what follows.
Consider the one-parameter group of diffeomorphisms, $\phi_t$, generated by the vector field ${\mbold\xi}$. Assume that two points $p$ and $\hat{p}$ (both in $U$) are connected via $\phi_t$; i.e., $\hat{p}=\phi_t(p)$. Since the eigenvalues are scalar functions over $U$, then $\phi_t^*(\lambda)=\hat{\lambda}$. Furthermore, by assumption, the eigenvalue structure of ${\sf T}$ does not change over $U$; consequently, for a eigenvector ${\sf v}$ with eigenvalue $\lambda$ we have:
\[ \phi^*_t({\sf T}{\sf v}-\lambda{\sf v})=\hat{\sf T}\hat{\sf v}-\hat{\lambda}\hat{\sf v}={\sf T}{\sf v}-\lambda{\sf v}=0.\]
This implies that eigenvectors are mapped onto eigenvectors. Therefore, if the eigenvalue
$\lambda$ is mapped onto $\hat{\lambda}$ (these are scalar functions determined by the characteristic equation), then the eigenvector ${\sf v}$ is mapped onto the corresponding eigenvector $\hat{\sf v}$. Then if ${\bf e}_I$ spans the eigenvectors of eigenvalue $\lambda$, and $\hat{\bf e}_I$ spans the eigenvectors with eigenvalue $\hat{\lambda}$, then there must exist, since $\phi_t$ also preserves the norm, an invertible matrix $M^I_{~J}$ so that:
\[ \hat{\bf e}_J=M^I_{~J}{\bf e}_I.\]
If ${\mbold\omega}^I$ and $\hat{\mbold\omega}^I$ are the corresponding one-forms, then $\hat{\mbold\omega}_J=(M^{-1})^I_{~J}{\mbold\omega}_I$.
For a projector, the eigenvalues are $\lambda = 0,1$; therefore, we can write
\[ \bot = \delta^{J}_{~I}{\bf e}_J\otimes {\mbold\omega}^I\]
where the eigenvectors ${\bf e}_J$ have eigenvalue 1.
Therefore:
\[ \phi^*_t(\bot)\equiv\hat{\bot}_t=\bot.\]
Consequently,
\[ \pounds_{\mbold\xi}\bot = 0.\]
Therefore, we have the remarkable property that the projectors are Lie transported with respect to
any vector field ${\mbold\xi}$.
For the Lie derivative we thus have the following result:
\begin{thm}
Assume that the operator, ${\sf T}$, has the decomposition (\ref{decomp2}). Then, the Lie derivative with respect to ${\mbold\xi}$ is
\[ \pounds_{\mbold\xi}{\sf T}=\sum_A\left[\pounds_{\mbold\xi}{\sf N}_A+{\mbold\xi}\left(\lambda_A\right)\bot_A\right], \quad \text{where}\quad \pounds_{\mbold\xi}{\sf N}_A=\bot_A(\pounds_{\mbold\xi}{\sf N})\bot_A. \]
\end{thm}
\begin{proof}
We have
\[ \pounds_{\mbold\xi}\left(\sum_A \lambda_A\bot_A\right)=\sum_A\left[\left(\pounds_{\mbold\xi}\lambda_A\right)\bot_A+\lambda_A\pounds_{\mbold\xi}(\bot_A)\right]=\sum_A{\mbold\xi}(\lambda_A)\bot_A, \]
and
\beq \pounds_{\mbold\xi}\left(\bot_A{\sf N}\bot_B\right)& =&\left(\pounds_{\mbold\xi}\bot_A\right){\sf N}\bot_B+\bot_A\left(\pounds_{\mbold\xi}{\sf N}\right)\bot_B+\bot_A{\sf N}\left(\pounds_{\mbold\xi}{\bot_B}\right)\nonumber \\ &=&\bot_A\left(\pounds_{\mbold\xi}{\sf N}\right)\bot_B\nonumber.
\eeq
Since $\bot_A{\sf N}\bot_B=\delta_{AB}{\sf N}_A$, the theorem now follows.
\end{proof}
Now clearly, this has some interesting consequences. If, for example, the operator ${\sf T}$ is diagonalisable, i.e., ${\sf N}=0$, then ${\sf T}$ is Lie transported if and only if the eigenvalues are Lie transported. This can formulated as follows:
\begin{cor}
If an operator, {\sf T}, is diagonalisable, then, for a vector field, ${\mbold\xi}$,
\[ \pounds_{\mbold\xi}{\sf T}=0\quad \Leftrightarrow\quad {\mbold\xi}(\lambda_A)=0\]
\end{cor}
\begin{cor}
If a spacetime $(\mathcal{M},g_{\mu\nu})$ is characterised by its invariants (weakly or strongly), then if there exists
a vector field, ${\mbold\xi}$, such that
\[ {\mbold\xi}(I_i)=0. \quad \forall I_i\in \mathcal{I},\]
then there exists a set, $\mathcal{K}$, of Killing vector fields such that at any point the vector field ${\mbold\xi}$ coinsides with a Killing vector field $\tilde{\mbold\xi}\in \mathcal{K}$.
\end{cor}
The last corollary enables us to reduce the question of Killing vectors down to the existence of vectors annihilating all the curvature invariants $I_i$. This may be easier in some cases if the form of the metric is totally unmanageable.
\subsection*{Examples}
\paragraph{CSI spacetimes:}
Consider now a spacetime which has all constant scalar curvature invariants \cite{CSI,3CSI,4CSI}. Assume also that the spacetime is characterised by its invariants, weakly or strongly. Let us now see how we can give an alternative proof that this must be locally homogeneous.
We choose, at any point $p$, a local coordinate system $\{x^k\}$. Using the coordinate frame, we can define ${\mbold\xi}_k=\partial_k$ we get:
\[ {\mbold\xi}_k(I_i)=0, \quad \forall I_i\in\mathcal{I}.\]
Therefore, for each ${\mbold\xi}_k$ (which are linearly independent at $p$), there would be a Killing vector field $\tilde{\mbold\xi}_k$ coinsiding with ${\mbold\xi}_k$ at $p$. Therefore, the Killing vectors are also linearly independent and span the tangent space, consequently, the spacetime is \emph{locally homogeneous}.
\paragraph{Kinnersley class I vacuum metrics:}
Let us consider the Kinnersley class I Petrov type D vacuum metric \cite{Kinnersley} which is a $\mathcal{I}$-non-degenerate metric \cite{typeD}. The Cartan invariants can all be reconstructed from the 4 (complex) scalar polynomial invariants:
\begin{eqnarray}
I &=& \frac 12\Psi_{abcd}\Psi^{abcd}=3(2Cil+m)^2 z^{-6}, \nonumber \\
C^{\alpha\beta\gamma\delta}C_{\alpha\beta\gamma\delta}&=& 24(2Cil+m)^2 z^{-6} +24(-2Cil+m)^2 z^{*-6}\nonumber \\
{\Psi}^{(abcd;e)f'} {\Psi}_{(abcd;e)f'} &=& 180(2Cil+m)^2 Sz^{-8}\nonumber \\
C^{\alpha\beta\gamma\delta;\mu}C_{\alpha\beta\gamma\delta;\mu}&=& 720(2Cil+m)^2 Sz^{-8} +720(-2Cil+m)^2 Sz^{*-8}\nonumber,
\end{eqnarray}
where $\Psi_{abcd}$ is the Weyl spinor, $a,b,..$ are spinor indices, and $\alpha, \beta,...$ are
frame indices. Here, $C$, $m$ and $l$ are all constants while all of the functions ($S$ etc.)
depend only on the coordinate $r$.
Since the invariants only depends on one variable, namely $r$, this spacetime possesses (at least) 3 transitive Killing vectors at any point.
\section{Analytic metric continuation}
The operators also have another property that seems to be very useful. The metric $g_{\mu\nu}$ does not
appear explicity in the analysis above and only appears after raising an index: $g^{\mu}_{~\nu}=\delta^{\mu}_{~\nu}$.
Consequently, this analysis is independent of the signature of the metric. In the eigenvalue equation the identity
operator ${\sf 1}$ would therefore be independent of the signature also. We can thus consider what happens
under more general coordinate transformations than the real diffeomorphisms preserving the metric
signature. Let us therefore consider complex analytic continuations of the real metric of this form.
Consider a point $p$ and a neighbourhood, $U$ at $p$, and we will assume this nighbourhood is an analytic
neighbourhood and that $x^{\mu}$ are coordinates on $U$ so that $x^{\mu}\in \mathbb{R}^n$. We will adapt
the coordinates so that the point $p$ is at the origin of this coordinate system. Consider now the
complexification of $x^{\mu}\mapsto x^{\mu}+iy^{\mu}=z^{\mu}\in \mathbb{C}^n$. This complexification
enables us to consider the complex analytic neighbourhood $U^{\mathbb C}$ of $p$.
Furthermore, let $g_{\mu\nu}^{\mathbb C}$ be a complex bilinear form induced by the analytic extension of the metric:
\[ g_{\mu\nu}(x^{\rho})\d x^\mu\d x^{\nu}\mapsto g^{\mathbb C}_{\mu\nu}(z^{\rho})\d z^\mu\d z^{\nu}.\]
Consider now a real analytic submanifold containing $p$: $\bar{U}\subset U^{\mathbb C}$ with coordinates $\bar{x}^\mu\in \mathbb{R}^n$. The imbedding $\iota:\bar{U}\mapsto U^{\mathbb C}$ enables us to pull back the complexified metric $g^{\mathbb C}$ onto $\bar{U}$:
\beq
\bar{g}\equiv \iota^*{g}^{\mathbb{C}}.
\eeq
In terms of the coordinates $\bar{x}^\mu$: $\bar{g}=\bar{g}_{\mu\nu}(\bar{x}^\rho)\d \bar{x}^\mu\d
\bar{x}^{\nu}$. This bilinear form may or may not be real. However, \emph{if the bilinear form
$\bar{g}_{\mu\nu}(\bar{x}^\rho)\d \bar{x}^\mu\d \bar{x}^{\nu}$ is real (and non-degenerate) then we will
call it an analytic metric extension of $g_{\mu\nu}(x^{\rho})\d x^\mu\d x^{\nu}$ with respect to $p$.}
In the following, let us call the analytic metric extension for $\bar{\phi}$; i.e., $\bar{\phi}:U\mapsto \bar{U}$. We note that this transformation is complex, and we can assume, since $U$ is real analytic, that $\bar{\phi}$ is analytic.
The analytic metric continuation leaves the point $p$ stationary. Therefore, it induces a linear
transformation, $M$, between the tangent spaces $T_pU$ and $T_p\bar{U}$. The transformation $M$ is
complex and therefore may change the metric signature; consequently, even if the metric
$\bar{g}_{\mu\nu}$ is real, it does not necessarily need to have the same signature as $g_{\mu\nu}$.
Consider now the curvature tensors, $R$ and $\nabla^{(k)}R$ for $g_{\mu\nu}$, and $\bar{R}$ and
$\bar{\nabla}^{(k)}\bar{R}$ for $\bar{g}_{\mu\nu}$. Since both metrics are real, their curvature tensors
also have to be real. The analytic metric continuation induces a linear transformation of
the tangent spaces; consequently, this would relate the Riemann tensors $R$ and $\bar{R}$ through a complex
linear transformation. However, how are these related?
By using $\bar{\phi}$ we can relate the metrics $g=\bar{\phi}^*\bar{g}$. Since the map is analytic (albeit
complex), the curvature tensors are also related via $\bar\phi$. For the operators this has a very useful
consequence. First, we note that \emph{scalar polynomial invariants are invariant under
$GL(\mathbb{C},n)$}; therefore, the eigenvalues of ${\sf R}$ and $\bar{\sf R}$ are identical at $p$.
Consequently, the eigenspace decomposition is identical also. (This seems almost too remarkable to be
true). Over the neigbourhoods $\bar{U}$ and $U$ we can, in general, relate the eigenvalues:
$\lambda_{A}=\bar\phi^*\bar\lambda_A$.
Since, a Riemannian space is always characterised by its invariants, we immediately have the following result:
\begin{thm} \label{analytic}
Assume that a space, $(\mathcal{M},g_{\mu\nu})$, of any signature can be analytically continued, in the sense above, to a Riemannian space (of Euclidean signature). Then the spacetime is characterised by its invariants in either a weak or strong sense.
\end{thm}
Note that the result is actually stronger than this because not only is it diagonalisable, if we have a
symmetric tensor giving rise to an operator (for example ${\sf R}$) then, since the Riemannian space must
have a symmetric operator $\bar{\sf R}$, the eigenvalues are real. Consequently, the eigenvalues at $p$
must also be real.
Interestingly, the reverse of the above theorem is just as useful:
\begin{cor}
If a spacetime is not characterised by its invariants (weakly or strongly), then there exists no analytical continuation of it to a Riemannian space.
\end{cor}
In the above we have restricted to analytic continuations $\bar{\phi}$ leaving a point $p$ fixed. In general, one can consider that does not necessarily leave a point fixed. However, for such complex mappings, we need to be a bit careful with the radius of convergence. On the other hand, assuming convergence, the invariants, $I_i$, can still be related via $\bar{\phi}^*\bar{I}_i=I_i$. Recall that we are restricting ourselves to neighbourhoods where operators do not change algebraic form.
For spacetime metrics of algebraically special types there are also some useful 'no-go' theorems for Lorentzian manifolds:\footnote{Note that type D is by definition excluded from this Corollary.}
\begin{cor}
If a spacetime is of (proper) Weyl, Ricci or Riemann type N, III or II, then there exists no analytical continuation of it to a Riemannian space.
\end{cor}
\begin{proof}
This follows from the fact that \emph{any} symmetric operator of a Riemannian space is diagonalisable. Since the Weyl, Ricci and Riemann operator of type N, III and II are not diagonalisable, the corollary follows.
\end{proof}
There are also a bundle of similar results that follows from a similar analysis.
\subsection*{Signature and convention}
Since analytic metric continuation can change the signature, we may be in a
situation in which we are comparing invariants of spacetimes of different signatures.
Let us note a word of
caution in this regard.
Here, we have (implicitly) assumed that the Riemann tensor $R^{\alpha}_{~\beta\gamma\delta}$ is given in a
coordinate basis with the standard formula involving the metric and the Christoffel symbols. This formula
is invariant under an overall change of signature; for example, $(++\cdots +)$ is the same as $(--\cdots
-)$. Consequently, the Ricci tensor, $R_{\mu\nu}=\bar{R}_{\mu\nu}$ also. However, the sign matters for
the Ricci scalar defined as $R=g^{\mu\nu}R_{\mu\nu}$ where we get $R=-\bar{R}$. Therefore, even if the
invariants may change with a sign, the overall signature is just a convention. This is also evident from
entities constructed from a non-zero curvature ``vector'' $r^{\alpha}$. The norm of this vector is an
invariant $I=r^{\alpha}r_{\alpha}$; however, in signature $(++\cdots +)$ this is necessarily positive,
$I>0$; while signature $(--\cdots -)$ this is necessarily negative $I<0$.
It is therefore crucial when comparing invariants through an analytic continuation that we specify the signature. Albeit it is conventional to say that metrics with an overall change of signature are indentical. However, the invariants of such metrics may change by a sign.
Let us be more precise. Consider an overall change of signature: $g_{\mu\nu}\mapsto \epsilon g_{\mu\nu}$, where $\epsilon =\pm 1$. Using a coordinate basis, we see that the Christoffel symbols $\Gamma^{\alpha}_{~\mu\nu}$ do not depend on $\epsilon$. Consequently, the Riemann tensor, $R^{\alpha}_{~\beta\mu\nu}$ does not depend on $\epsilon$ either. Moveover, we can also see that the definition of the covariant derivatives, $R^{\alpha}_{~\beta\mu\nu;\lambda_1...\lambda_k}$, do not depend on $\epsilon$. However, \emph{raising} or \emph{lowering} indices introduce $\epsilon$. Therefore, depending on whether you raise/lower an odd or even number of indicies to create an invariant, the invariant will depend (if odd) or not depend (if even) on $\epsilon$. We can therefore split the invariants into invariants that depend on $\epsilon$, denoted $ I^\epsilon_i$, and those that do not, denoted $I_j$. Since \emph{the choice of overall sign is merely a convention} we make the following identification:
\beq
(g_{\mu\nu},~I^\epsilon_i, ~I_j)\sim (-g_{\mu\nu},~-I^\epsilon_i, ~I_j).
\eeq
Note that in the neutral signature case changing the overall sign does not change the signature; consequently, the above map gives a non-trivial equivalence relation between invariants of neutral signature metrics.
\subsection*{Examples}
\paragraph{Kasner universe:}
In 1921 E. Kasner \cite{Kasner} wrote down a Ricci flat metric of Riemannian signature (in dimension 4 but we
present it in $n$ dimension):
\beq
\d s^2=\d t^2+\sum_{i=1}^nt^{2p_i}(\d x^i)^2, \qquad \sum_{i=1}^{n}p_i=\sum_{i=1}^np_i^2=1.
\label{Kasner}\eeq
Clearly, this has $n$ Killing vectors, $\frac{\partial}{\partial x^i}$, and therefore the
invariants can only depend on $t$: $I=I(t)$. Furthermore, since this is a Riemannian space,
this space is characterised by its invariants.
There is a number of analytic metric continuations possible using $\bar{x}^j=ix^j$, where $i$ is the imaginary unit and $x^j$ is any of the coordinates except $t$. The so-called Kasner Universe is the Lorentzian version obtained by $(\bar{x}^1,\cdots,\bar{x}^n)=(ix^1,\cdots,ix^n)$ of signature $(+--\cdots -)$. Consequently, the invariants of this spacetime are identical to those of Riemannian version of signature $(+++\cdots+)$. In particular, this means that curvature invariants constructed from the norm of the gradient of an invariant, must be non-negative; i.e., $(\nabla_{\alpha}I)(\nabla^\alpha I)\geq 0$.
Usually, the Lorentzian version is written with signature $(-++\cdots+)$, which can be obtained from
the $(---\cdots-)$ version; consequently, $(\nabla_{\alpha}I)(\nabla^\alpha I)\leq 0$.
\paragraph{Schwarzschild spacetime:} The 4D Schwarzschild spacetime is given by
\beq
\d s^2=-\left(1-\frac{2m}r\right)\d t^2+\frac{\d r^2}{\left(1-\frac{2m}r\right)}+r^2(\d \theta^2+\sin^2\theta \d \phi^2).
\eeq
There exists an analytic metric continuation (not necessarily unique) to a Euclidean signature $(++++)$ given by
\[ (\bar{t},r,{\theta},\phi)=(it,r,\theta,\phi).\]
This shows that the Schwarzschild spacetime is characterised by its invariants. Furthermore, the invariants are idential to the invariants given by the Euclidean version. Note that there is also a Lorentzian $(+---)$ version obtained by $\bar{\theta}=i\theta$ with identical invariants. This illustrates the subtleties in the difference of signatures and the invariants.
In 5D the Schwarzschild spacetime is given by
\beq
\d s^2=-\left(1-\frac{2m}{r^2}\right)\d t^2+\frac{\d r^2}{\left(1-\frac{2m}{r^2}\right)}+r^2(\d \theta^2+\sin^2\theta \d \Omega^2).
\eeq
Here, there is a similar continuation, $\bar{t}=it$ that relates the $(-++++)$ version with the
Euclidean $(+++++)$. Interestingly, in 5D we can also do the continuation $\bar{r}=ir$ which turns it into a Euclidean space of signature $(-----)$:
\[
\d s^2=-\left(1+\frac{2m}{\bar{r}^2}\right)\d t^2-\frac{\d \bar{r}^2}{\left(1+\frac{2m}{\bar{r}^2}\right)}-\bar{r}^2(\d \theta^2+\sin^2\theta \d \Omega^2).
\]
Note that this mapping does not leave a point invariant ($r=0$ is a singularity and therefore has to be omitted); therefore, the invariants are related via $\bar{I}(\bar{r})=I(r=-i\bar{r})$.
\paragraph{A pair of metrics:} Let us consider the pair of metrics:
\beq\label{weakly1}
\d s^2_1&=&\frac{1}{z^2}\left(\d x^2+\d y^2+\d z^2\right)-\d \tau^2, \\
\d s^2_2&=&\frac{1}{z^2}\left(-\d x^2+\d y^2+\d z^2\right)+\d \tau^2. \label{weakly2}
\eeq
Now, we can easily see that these can be analytically continued into eachother. This shows that these metrics have indentical invariants. Furthermore, it is also clear that we can analytically continue them into a Euclidean metric of signature $(++++)$. Consequently, these metrics are characterised by their invariants.
This pair of metrics illustrate two things. First,
the metric (\ref{weakly1}) is of Segre type $\{1,(111)\}$ and is consequently (weakly) $\mathcal{I}$-non-degenerate\footnote{Recall that strongly $\mathcal{I}$-non-degenerate is defined to be the case where the inverse of the $\mathfrak{I}$-map consists of one point, and one point only ($\mathfrak{I}^{-1}(\mathcal{I}_0)\cong\redpoint$ in the notation in section \ref{sect:summary}), while weakly $\mathcal{I}$-non-degenerate is defined to be the case where the inverse of the $\mathfrak{I}$-map may consist of several isolated points (e.g., $\mathfrak{I}^{-1}(\mathcal{I}_0)\cong 3\redpoint$) \cite{inv}.}. The metric (\ref{weakly2}), on the other hand, is actually a Kundt degenerate metric. There are therefore continuous metric deformations of the latter metric to other metrics with identical invariants.
Second, the metric (\ref{weakly1}) is weakly $\mathcal{I}$-non-degenerate but not strongly $\mathcal{I}$-non-degenerate. The reason for this is the existence of an analytic continuation between two Lorentzian metrics. We might wonder if such spacetimes (which are weakly but not strongly $\mathcal{I}$-non-degenerate) always have an associated analytic continuation.
\section{Summary: the 4D Lorentzian case}
\label{sect:summary}
Let us present a summary of the Lorentzian signature case. All of the results presented here are
based on the 4-dimensional (4D) case, but it is believed that they are also true in higher-dimensions.
Let us consider the space of 4D Lorentzian metrics $\mathfrak{M}$ over an open neighbourhood $U$.
Consider also, $\mathcal{I}=\{I_i\}$, as the set of scalar polynomial invariants over $U$. The set of invariants
can be considered as an element of the Cartesian product of smooth functions over $U$,
$\left[C^{\infty}(U)\right]^{\times N}$ for some $N$. We can thus consider the
calculation of invariants, $I_i$, from a metric $g_{\mu\nu}\in \mf{M}$ as a map
$\mf{I}:\mf{M}\mapsto \left[C^{\infty}(U)\right]^{\times N}$, given by the $\mf{I}$-map:
\[ \mf{I}: g_{\mu\nu}\mapsto I_i.\]
This map is clearly not surjective, so for a given
$\mathcal{I}\in \left[C^{\infty}(U)\right]^{\times N}$, the
inverse image $\mf{I}^{-1}(\mathcal{I})$, may or may not be empty. Let us henceforth assume that we consider only points in the image, $\mathcal{B}\equiv\mf{I}(\mf{M})\subset \left[C^{\infty}(U)\right]^{\times N}$ of $\mf{I}$, and let us consider the sets of metrics with identical invariants; i.e., for a $\mathcal{I}_0\in \mathcal{B}$, $\mf{I}^{-1}(\mathcal{I})$.
The connected components (in the set of metrics) of this inverse image can be of three kinds:
\begin{enumerate}[i{)}]
\item{} An isolated point. This would correspond to a metric being $\mathcal{I}$-non-degenerate. We will use the symbol $\redpoint$ for such a point.
\item{} A generic Kundt-wedge (or Kundt-tree; it is not entirely clear what the topology of this set is).
This corresponds to a set of degenerate Kundt metrics \cite{Kundt}, none of which is
characterised by its invariants, strongly or weakly. This set would be Kundt metrics connected via metric deformations
with identical invariants. We will use the symbol $\tree$ to illustrate a set of this kind.
\item{} A special Kundt-wedge containing a metric weakly characterised by its invariants. This set
contains degenerate Kundt metrics: however, one of the members of this set is a special Kundt metric which is characterised by its invariants in the weak sense. We will use the symbol, $\redtree$, of a set of this kind. The symbol $\redpoint$ corresponds to the special point which is (weakly) characterised by its invariants, at the ``top'' of the Kundt-wedge.
\end{enumerate}
The various sets $\mf{I}^{-1}(\mathcal{I}_0)$, will consist of these components. In all of
the examples known to the authors, the points $\redpoint$ are actually connected
via analytic continuations. Figure \ref{fig1} presents a summary of the various cases that are known.
\begin{figure}[tbp]
\centering \includegraphics[width=10cm]{trees.png}
\caption{Figures showing the sets of metrics of identical invariants.}
\label{fig1}
\end{figure}
Let us consider each figure in turn and comment on the various cases.
\begin{enumerate}
\item{} The generic case: Strongly $\mathcal{I}$-non-degenerate. This case consists of only one isolated point $\redpoint$. There is a unique metric with this set of invariants.
\item{} Two or more isolated points $\redpoint$, all of which are (weakly) $\mathcal{I}$-non-degenerate.
The Kasner metrics, eq. (\ref{Kasner}), are examples of this kind.
\item{} Two or more components of types $\redpoint$ and $\redtree$. The pair of metrics, eqs. (\ref{weakly1}) and (\ref{weakly2}), are examples of this type and the metrics given corresponds to the two points $\redpoint$.
\item{} One, two or more components of type $\redtree$. As an example of this kind are the two metrics $AdS_2\times S^2$ and $\mathbb{H}^2\times dS_2$.
\item{} A generic Kundt-wedge, $\tree$. Examples of this case are generic degenerate Kundt metrics.
\end{enumerate}
It is not clear whether there might exist sets of type $\tree$ combined with any of the other
components (like $\tree+\redpoint$) but no such examples are known to date.
If $\mf{I}^{-1}(\mathcal{I}_0)$ consists of several components, then in all of the examples known
to the authors, there is always an analytic metric continuation connecting points in them (not necessarily all points of the components).
\section{Neutral Signature}
Let us also briefly discuss the 4D pseudo-Riemannian case of neutral
signature $(--++)$ (NS space). Here, little has previously been done with regards to the connection between the invariants and the NS space.
Let us first consider the Segre types of the Ricci tensor. This is essentially the canonical forms of of the Ricci operator and these will also be given for the non-standard types. Note that the canonical forms are obtained using real $SO(2,2)$ transformations. The possible types are \cite{Petrov}:
\begin{itemize}
\item{} $\{11,11\}$: All real eigenvalues. Degenerate cases are: $\{(11),11\}$, $\{11,(11)\}$, $\{(11),(11)\}$, $\{(1,1)(1,1)\}$, $\{1(1,11)\}$, $(11,1)1\}$, $\{(11,11)\}$.
\item{} $\{211\}$.
\item{} $\{31\}$.
\item{} $\{4\}$: 4 equal eigenvalues. Canonical forms for Ricci operator, and metric, $g_{\mu\nu}$:
\beq
{\sf R}=\begin{bmatrix}
\lambda & 1 & 0 & 0 \\
0 & \lambda & 1 & 0 \\
0 & 0 & \lambda & 1 \\
0 & 0 & 0 & \lambda
\end{bmatrix}
, \qquad
(g_{\mu\nu})=
\begin{bmatrix}
0 & 0 & 0 & 1 \\
0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
1 & 0 & 0 & 0
\end{bmatrix}.
\eeq
\item{} $\{22\}$: Two non-diagonal Jordan blocks with 2 distinct eigenvalues. Canonical form:
\[ {\sf R}=\diag({\sf B}_1,{\sf B}_2), \qquad (g_{\mu\nu})=\diag({\sf g}_1,{\sf g}_2),\]
where
\[ {\sf B}_A=\begin{bmatrix}
\lambda_A & \pm 1 \\
0 & \lambda_A
\end{bmatrix}, \qquad
{\sf g}_A=\begin{bmatrix}
0 & 1 \\ 1 & 0
\end{bmatrix} \]
\item{} $\{1_{\mathbb C}1,1\}$: 2 real and 2 complex conjugate eigenvalues. The block $1_{\mathbb C}$ has canonical form:
\[ {\sf B}_A=\begin{bmatrix}
\alpha_A & \beta_A \\
-\beta_A & \alpha_A
\end{bmatrix}, \qquad
{\sf g}_A=\begin{bmatrix}
1 & 1 \\ 1 & -1
\end{bmatrix} \]
\item{} $\{1_{\mathbb C}1_{\mathbb C}\}$: 2 pairs of complex conjugate eigenvalues. Canonical form is \[ {\sf R}=\diag({\sf B}_1,{\sf B}_2), \qquad (g_{\mu\nu})=\diag({\sf g}_1,{\sf g}_2),\]
where
\[ {\sf B}_A=\begin{bmatrix}
\alpha_A & \beta_A \\
-\beta_A & \alpha_A
\end{bmatrix}, \qquad
{\sf g}_A=\begin{bmatrix}
1 & 1 \\ 1 & -1
\end{bmatrix} \]
\item{} $\{2_{\mathbb{C}}\}$: 2 equal pairs of complex conjugate eigenvalues with a non-diagonal Jordan block. Canonical form:
\[ {\sf R}=\begin{bmatrix}
\alpha & \beta & 1 & 0 \\
-\beta & \alpha & 0 & 1\\
0 & 0 & \alpha & \beta \\
0 & 0 & -\beta & \alpha
\end{bmatrix}
, \qquad (g_{\mu\nu})=\begin{bmatrix}
0 & 0 & 1 & 1 \\
0 & 0 & 1 & -1\\
1 & 1 & 0 & 0 \\
1 & -1 & 0 & 0
\end{bmatrix}.\]
\end{itemize}
\begin{cor}
Neutral signature metrics of Segre types $\{11,11\}$, $\{(11),11\}$, $\{11,(11)\}$, $\{(11),(11)\}$, $\{1_{\mathbb C}1,1\}$ and $\{1_{\mathbb C}1_{\mathbb C}\}$ are $\mathcal{I}$-non-degenerate, and are, consequently, characterised by the invariants in the strong sense.
\end{cor}
\begin{proof}
The proof of this corollary is, except for types $\{(11),(11)\}$ and $\{1_{\mathbb C}1_{\mathbb C}\}$, identical to the Lorentzian case \cite{inv}. For type $\{(11),(11)\}$, we obtain two projectors, each with symmetry $SO(2)$. We can therefore project onto $SO(2)$-tensors and since this is a compact group, the invariants split orbits. Consequently, this is case is $\mathcal{I}$-non-degenerate. The case $\{1_{\mathbb C}1_{\mathbb C}\}$ is analogous to case $\{1_{\mathbb C}1,1\}$, but we need to perform a complex transformation for each of the two complex blocks. Since all eigenvalues must be different, we get that this case is also $\mathcal{I}$-non-degenerate.
\end{proof}
For the Weyl operator ${\sf C}$, this splits into a self-dual and anti-self-dual part: ${\sf C}={\sf W}^+\oplus {\sf W}^-$. In \cite{Law} Law classified the Weyl tensor of NS metrics using the Weyl operator (or endomorphism). Using the Hodge star operator, $\star$, which commutes with the Weyl operator -- i.e., $\star \circ{\sf C}={\sf C}\circ\star$ -- the (anti-)self-dual operators can be defined as:
\[ {\sf W}^\pm=\tfrac 12\left({\sf 1}\pm\star\right){\sf C}.\]
Each of the parts can be considered to be symmetric and tracefree with respect to the 3-dimensional Lorentzian metric with signature $(+--)$. Consquently, each of the operators ${\sf W}^{\pm}$ can be classified according to ``Segre type''(the ``Type'' refers to Law's enumeration):
\begin{itemize}
\item{} Type Ia: $\{1,11\}$
\item{} Type Ib: $\{z\bar{z}1\}$
\item{} Type II: $\{21\}$
\item{} Type III:$\{3\}$.
\end{itemize}
It is also advantageous to refine Law's enumeration for the special cases:
\begin{itemize}
\item{} Type D: $\{(1,1)1\}$
\item{} Type N: $\{(21)\}$
\end{itemize}
Based on this classification, we get the following result:
\begin{thm}
If a neutral signature 4D metric has Weyl operators ${\sf W}^{\pm}$ both of type I, then the metric is $\mathcal{I}$-non-degenerate.
\end{thm}
\begin{proof}
The proof utilises the fact that each type I operator ${\sf W}^{\pm}$, can pick up the $+$-direction of the fictitious 3D space with signature $(+--)$. In the real case, this gives rise to the eigenbivector $F_{\mu\nu}$ where, in the orthonormal frame,
\[
{\sf F}^{\pm}\equiv (F^{\mu}_{~\nu})=\diag({\sf F}_1,\pm {\sf F}_1), \quad
{\sf F}_1=\begin{bmatrix} 0 & 1 \\
-1 & 0
\end{bmatrix}
\]
where the signs $\pm$ correspond to ${\sf W}^{\pm}$. Now by using these eigenvectors, we can construct the operator:
\[ {\sf F}^+{\sf F}^-=\diag(-1,-1,1,1), \]
which is of type $\{(11),(11)\}$. Consequently, using the argument from the Ricci type $\{(11),(11)\}$, this NS space is $\mathcal{I}$-non-degenerate.
\end{proof}
Note that this is the generic case; that is,
the general NS space is characterised by its invariants in the strong sense.
Let us also discuss the NS case in light of theorem \ref{analytic}. As pointed out,
the theorem is valid for any signature and thus applies to the NS case also. Consider the Euclidean 4D Schwarzschild spacetime as an example:
\beq
\d s^2=\left(1-\frac{2m}r\right)\d\tau^2+\frac{\d r^2}{\left(1-\frac{2m}r\right)}+r^2(\d \theta^2+\sin^2\theta \d \phi^2).
\eeq
Here, there are (at least) three complex metric extensions yielding an NS space, namely $(\bar{\tau},r,\theta,\bar{\phi})=(i{\tau},r,\theta,i{\phi})$, $({\tau},r,\bar\theta,{\phi})=({\tau},r,i\theta,{\phi})$ and $(\bar{\tau},r,\bar\theta,\bar{\phi})=(i{\tau},r,i\theta,i{\phi})$, giving the triple of metrics:
\beq
\d s^2_1&=&-\left(1-\frac{2m}r\right)\d\bar{\tau}^2+\frac{\d r^2}{\left(1-\frac{2m}r\right)}+r^2(\d \theta^2-\sin^2\theta \d \bar{\phi}^2), \label{NS1} \\
\d s^2_2&=&\phantom{-}\left(1-\frac{2m}r\right)\d\tau^2+\frac{\d r^2}{\left(1-\frac{2m}r\right)}-r^2(\d \bar{\theta}^2+\sinh^2\bar\theta \d \phi^2), \label{NS2}\\
\d s^2_3&=&-\left(1-\frac{2m}r\right)\d\bar\tau^2+\frac{\d r^2}{\left(1-\frac{2m}r\right)}-r^2(\d \bar{\theta}^2-\sinh^2\bar\theta \d \bar\phi^2). \label{NS3}
\eeq
Therefore, these three metrics are characterised by its invariants. In order to determine whether they are strongly or weakly characterised by their invariants, a more thorough investigation of the NS case is needed. As a preliminary investigation we can try to determine the Weyl type of these metrics. By calculating ${\sf W}^\pm$ using the definitions, we get the following types:
\begin{enumerate}
\item{} Metric (\ref{NS1}): Type $\{(1,1)1\}\times \{(1,1)1\}$ or type D$\times$D.
\item{} Metric (\ref{NS2}): Type $\{1,(11)\}\times \{1,(11)\}$ or type Ia$\times$Ia.
\item{} Metric (\ref{NS3}): Type $\{(1,1)1\}\times \{(1,1)1\}$ or type D$\times$D.
\end{enumerate}
We can therefore conclude that metric (\ref{NS2}) is $\mathcal{I}$-non-degenerate, while the other metrics require a more thorough study.\footnote{The two metrics (\ref{NS1}) and (\ref{NS3}) are actually locally diffeomorphic, however, these two different forms are useful if we wish to extend these to degenerate metrics.}
In the Lorentzian case the spacetimes not characterised by their scalar invariants
are Kundt spacetimes. We may therefore wonder what are the NS spaces that are not characterised by their invariants. A hint may be provided by utilising an analytic metric continuation of the following Lorentzian case.
Consider the vacuum plane wave spacetimes \cite{exsol}:
\beq
\d s^2=2\d u(\d v+H(u,x,y)\d u)+\d x^2+\d y^2,
\eeq
where $H(u,x,y)=f(u,x+iy)+\bar{f}(u,x-iy)$ for an analytic function $f(u,z)$. This is a well-known vacuum Kundt-VSI spacetime.
Assume that we consider the special case $\bar{f}(u,\bar{z})=f(u,\bar{z})$. Then, we can perform the analytic metric continuation $\tilde{y}=iy$ which gives the NS metric:
\beq
\d s^2=2\d u(\d v+\tilde{H}(u,x,\tilde{y})\d u)+\d x^2-\d \tilde{y}^2,
\eeq
where $\tilde{H}(u,x,\tilde{y})=f(u,x+\tilde{y})+{f}(u,x-\tilde{y})$.
Since the analytic continutation preserves the invariants and the structure,
this metric must be VSI and a solution to the vacuum equations. Moreover,
it cannot be characterised by its invariants and therefore represents a ``Kundt analogue'' of the NS case. Interestingly, this metric is a special 4D Walker metric \cite{Walker}.
The Walker metrics can be written as
\beq\label{Walker}
\d s^2=2\d u(\d v+H_1\d u+W\d x)+2\d x(\d y+H_2\d x+W\d u),
\eeq
where $H_1=H_1(u,v,x,y)$, $H_2=H_2(u,v,x,y)$ and $W=W(u,v,x,y)$. Such metrics admit a field of parallell null 2-planes. These metrics seem to possess some of the right curvature properties to be NS candidates for degenerate NS metrics \cite{curvW}.
Let us introduce the double null-frame:
\beq
{\mbold\omega}^1 =\d u && {\mbold\omega}^2=\d v+H_1\d u+W\d x\nonumber \\
{\mbold\omega}^3=\d x && {\mbold\omega}^4=\d y+H_2\d x+W\d u.
\label{doublenull}\eeq
For neutral metrics we can introduce two independent boost weights, $(b_1,b_2)$, corresponding to the two boosts:
\beq
{\mbold\omega}^1 \mapsto e^{\lambda_1}{\mbold\omega}^1, && {\mbold\omega}^2 \mapsto e^{-\lambda_1}{\mbold\omega}^2,\nonumber \\
{\mbold\omega}^3 \mapsto e^{\lambda_2}{\mbold\omega}^3, && {\mbold\omega}^4 \mapsto e^{-\lambda_2}{\mbold\omega}^4.
\eeq
Note that any invariant must have boost weight $(0,0)$, in addition, denoting $(T)_{(b_1,b_2)}$ as the projection of the tensor $T$ onto the boost weight $(b_1,b_2)$, then $(T)_{(b_1,b_2)}\otimes(\tilde{T})_{(\tilde{b}_1,\tilde{b}_2)}$ is of boost weight $(b_1+\tilde{b}_1,b_2+\tilde{b_2})$. Moreover,
\[ (T\otimes\tilde{T})_{(\hat{b}_1,\hat{b}_2)}=\sum_{(b_1+\tilde{b}_1,b_2+\tilde{b}_2)=(\hat{b}_1,\hat{b}_2)}(T)_{(b_1,b_2)}\otimes(\tilde{T})_{(\tilde{b}_1,\tilde{b}_2)}.\]
Further, we will say that a tensor $T$, possesses the ${\bf N}$-property, iff:
\beq
(T)_{(b_1,b_2)}=0,\quad\text{for}\quad
\begin{cases}
b_1>0, & b_2~\text{arbitrary},\\
b_1=0, & b_2>0, \\
b_1=0, & b_2=0.
\end{cases}
\eeq
\begin{figure}[tbp]
\centering \includegraphics[width=6cm]{Nprop.png}
\caption{The tensors that fulfill the ${\bf N}$-property have components as indicated in the figure. The allowable boost weights, $(b_1,b_2)$, fill a semi-infinite grid in ${\mb Z}^2$ (which is closed under addition) indicated by black dots. }
\label{fig2}
\end{figure}
Figure \ref{fig2} illustrates the allowable components in terms of their boost weight.
We note that if both $T$ and $S$ possess the ${\bf N}$-property, then so does $T\otimes S$ since the resulting boost weights can be considered as ``vector addition'' in ${\mb Z}^2$.
\begin{prop}
The eigenvalues of an operator, ${\sf T}$, possessing the ${\bf N}$-property (as a tensor) are all zero; consequently, ${\sf T}$ is nilpotent.
\end{prop}
\begin{proof}
We note that the metric $g$ is of boost weight $(0,0)$; thus, if a tensor, ${\sf T}$ possesses the ${\bf N}$-property, so will any full contraction. Therefore, since an invariant must be of boost weight $(0,0)$, $\Tr( {\sf T})=0$. Furthermore, since ${\sf T}^n$ must possess the ${\bf N}$-property also, $\Tr ({\sf T}^n)=0$. The eigenvalues are therefore all zero.
\end{proof}
Let us now consider the Walker metrics. In the basis (\ref{doublenull}) we note that by counting the type of index (when all are downstairs!) we can get the boost weight as follows: $b_1=\#(1)-\#(2)$ and $b_2=\#(3)-\#(4)$. We are now in a position to state:
\begin{prop}
The Walker metrics eq.(\ref{Walker}) where
\beq
H_1&=& vH_1^{(1)}(u,x,y)+H_1^{(0)}(u,x,y)\nonumber \\
H_2&=& vH_2^{(10)}(u,x)+yH_2^{(01)}(u,x)+H_2^{(0)}(u,x)\nonumber \\
W&=& vW^{(1)}(u,x)+W^{(0)}(u,x,y),
\eeq
are NS metrics having the following properties:
\begin{enumerate}
\item{} If $H_2^{(10)}(u,x)\neq 0$, then all 0th and 1st order invariants vanish; i.e., they are VSI$_1$ metrics.
\item{} If $H_2^{(10)}(u,x)=0$ then \emph{all} polynomial curvature invariants vanish; i.e., they are VSI metrics.
\end{enumerate}
\end{prop}
\begin{proof}
Using the functions as defined in the theorem, then we get by direct calculation (using, for example, GRTensorII):
\begin{itemize}
\item{} The Riemann tensor, $R$, possesses the ${\bf N}$-property.
\item{} The covariant derivative of the Riemann tensor, $\nabla R$, possesses the ${\bf N}$-property.
\end{itemize}
Consequently, it is a VSI$_1$ space, and the first part of the theorem follows.
Concerning the last part, we note that for $H^{(10)}_2=0$, then the rotation coefficients, $\Gamma^\alpha_{~\mu\nu}$, possess the ${\bf N}$-property also. The covariant derivative of a tensor $T$ can symbolically be written
\[ \nabla T=\partial T-\sum \Gamma*T,\]
Therefore, if both $T$ and $\Gamma$ possess the ${\bf N}$-property, then the only term that can prevent $\nabla T$ possessing the ${\bf N}$-property also is the partial derivative term $\partial T$.
We therefore need to check the terms that raise the boost weight in such a way that it could potentially violate the ${\bf N}$-property.
For the Riemann tensor we can use the Bianchi identity,
\beq
R_{\alpha\beta\mu\nu;\rho}=-R_{\alpha\beta\rho\mu;\nu}+R_{\alpha\beta\rho\nu;\mu},
\eeq
and for general tensors, the generalised Ricci identity,
\beq
[\nabla_{\mu},\nabla_{\nu}]T_{\alpha_1...\alpha_k}=\sum_{i=1}^kT_{\alpha_1...\lambda...\alpha_k}R^{\lambda}_{~\alpha_i\mu\nu}.
\eeq
We need to check the components $\nabla_{1}R_{\alpha\beta\mu\nu;\rho_1...\rho_k}$ and $\nabla_{3}R_{\alpha\beta\mu\nu;\rho_1...\rho_k}$ that can potentially violate the ${\bf N}$-property. We will do this by induction in $n$. From the first part of the theorem (which we have already shown), we know it is true for $n=0$ and $n=1$. Assume therefore its true for $n=k$ and $n=k-1$. Then we need to check for $n=k+1$. Let us first consider $\nabla_{1}R_{\alpha\beta\mu\nu;\rho_1...\rho_k}=R_{\alpha\beta\mu\nu;\rho_1...\rho_k 1}$. We can now use the generalised Ricci identity (and possibly the Bianchi identity) to rewrite this component as a derivative w.r.t. $2$-, $3$-, or $4$-components. Of these, the only potentially dangerous term is the $R_{\alpha\beta\mu\nu;\rho_1...\rho_k 3}$, of boost weight $(0,0)$. If $\rho_k=1$ or $\rho_k=3$, then we use generalised Ricci to permute the indices so that $\rho_k$ is $2$ or $4$ (we can always do this). Then we use the generalised Ricci again to permute the last two indices $\rho_n$ and $3$. We now see that the $(0,0)$ component can be replaced with $\nabla_{2}R_{\alpha\beta\mu\nu;\rho_1...\rho_k}$ or $\nabla_{4}R_{\alpha\beta\mu\nu;\rho_1...\rho_k}$; consequently, these are zero. Hence, the tensor $\nabla_{\rho_{k+1}}R_{\alpha\beta\mu\nu;\rho_1...\rho_k}$ fulfills the ${\bf N}$-property also.
The theorem now follows from these results.
\end{proof}
This clearly indicates that the Walker metrics are indeed ``Kundt analogues'' for NS metrics, however, its not clear whether all degenerate NS metrics are Walker metrics. The investigation of the degenerate NS spaces will be left for future work.
\section{Discussion}
In this paper we have further studied the question of when a pseudo-Riemannain manifold can be locally
characterised by its scalar polynomial curvature invariants.
In particular, we have introduced the new concepts of diagonalisability
and analytic metric extension and proven some important theorems.
In the final two sections we have discussed some applications of these results. First,
we considered the 4d Lorentzian case, in part to illustrate the explicit theorems
that are known \cite{inv}. Second, we have initiated an investigation of the
neutral signature case, which has not been widely studied previously. We intend to continue the
study of the neutral signature case in future work.
\newpage
{\em Acknowledgements}.
This work was in part supported by NSERC of Canada.
|
\section{Introduction}
Warped geometries have attracted considerable attentions from diverse areas of modern theoretical physics ranging e.g. from braneworld models to warped compactification of string theory. Most studies on warped geometries are based on a specific ansatz on the warp factor, which may be subject to some additional discrete symmetry (e.g. in braneworld theories the discrete symmetry is often taken to be $Z_2$), and the number of warped extra dimensions is often limited to 1.
Meanwhile, Einstein manifolds are a class of spacetime manifolds which play a very important role in General Relativity. Such spacetime manifolds correspond to vacuum solutions of Einstein's theory of gravity either with or without a cosmological constant. In many cases, black hole and cosmological solutions are either a specific example of or intimately related to this class of spacetimes. Under the assumption that extra spacetime dimensions exist it is interesting to ask whether an Einstein manifold of lower dimension can be embedded into another Einstein manifold of higher dimension.
In this article, we shall consider the problem of embeddings between Einstein manifolds of different spacetime dimensions. From the string theory or experimental detection perspectives, there is no reason to restrict the number of extra dimensions to 1, and no additional discrete symmetry is required {\it a priori}. To be more concrete, what we shall study is the the following $D$-dimensional spacetime (bulk) with metric $\hat{g}_{MN}$ having a warped form
\begin{equation}
ds^{2}=\hat{g}_{MN}dX^{M}dX^{N}=e^{a(y)}g_{\mu \nu}(x)dx^{\mu}dx^{\nu}%
+h_{ij}(y)dy^{i}dy^{j}, \label{eq:1}
\end{equation}
where $X^{M}=(x^{\mu},y^{i})$ are bulk coordinates, $a(y)$ is a scalar function of $y^{i}$ but not of $x^{\mu}$,
with $M=0,1,...,D-1$, $\mu=0,1,...,d-1$ and $i=d,...,D-1$. Quantities associated with the bulk are labeled by a hat. The difference $n=D-d$ between the dimensions of the bulk and the submanifold with metric $g_{\mu\nu}$ will be referred to as the codimension of the embedding.
The metric $\hat{g}_{MN}$ of the bulk and
$g_{\mu \nu}$ of the $d$-dimensional submanifold are both required to be Einstein manifolds, i.e.
\begin{align}
\hat{R}_{MN} = \frac{2\hat{\Lambda}}{D-2} \hat{g}_{MN} ,\quad
R_{\mu\nu} &= \frac{2\Lambda}{d-2} g_{\mu\nu} .\label{r-eq}
\end{align}
Notice that simply identifying a cosmological constant is not sufficient to justify the spacetime to be a de Sitter/anti-de Sitter or Minkowski spacetime, because in the usual notion of de Sitter/anti-de Sitter and Minkowski spacetimes a maximal symmetry is implicitly assumed. However, in order to identify the nature of the bulk and the submanifold we are going to deal with, we shall abuse terminologies and denote Einstein manifolds with negative, zero and positive cosmological constants respectively with $AdS_{D}$, $M_{D}$ and $dS_{D}$, leaving the maximal symmetry ansatz aside\footnote{The spacetime manifolds that are referred to as of Minkowski in this paper may be properly called Ricci-flat Einstein manifolds.}.
Thus we are facing up with 9 different types of embeddings, i.e.
\begin{equation}
AdS_{d} \subset \left\{
\begin{array}{l}
AdS_{D} \\
M_{D} \\
dS_{D}
\end{array}
\right. ,\qquad
M_{d} \subset \left\{
\begin{array}{l}
AdS_{D} \\
M_{D} \\
dS_{D}
\end{array}
\right. ,\qquad
dS_{d} \subset \left\{
\begin{array}{l}
AdS_{D} \\
M_{D} \\
dS_{D}
\end{array}
\right. . \label{9cases}
\end{equation}
The question to be asked can be clearly described as: for each of the 9 cases listed in (\ref{9cases}), can we find exact, analytic solutions of the embedding metric, especially, can we find $e^{a(y)}$ and $h_{ij}(y)$ for each cases without specifying the details of $g_{\mu\nu}$? Partial answer to this question will be given in the main context of this article, and for each cases for which we can give an explicit answer we shall point out some of its intriguing features and make brief discussions on the potential applications of the specific answer.
By straightforward calculations we can see that
the Ricci tensors for the bulk and
the submanifold are linked as follows:
\begin{align}
\hat{R}_{\mu \nu} & =R_{\mu \nu}(g)
+\frac{1}{2}\left[e^{-a} \square ^{(h)} e^a +\frac{1}{2}(d-2) h^{ij}\nabla^{(h)}_i a \nabla^{(h)}_j a\right] \hat{g}_{\mu \nu} ,
\label{munu-D}\\
\hat{R}_{ij} &=R_{ij}(h)+\frac{d}{2}\left[
\nabla^{(h)}_i \nabla^{(h)}_j a
+\frac{1}{2}\nabla^{(h)}_i a \nabla^{(h)}_j a \right], \label{ij-D}
\end{align}%
where $\square^{(h)}$ and $\nabla^{(h)}_i$ respectively denote the D'Alembertian and covariant derivatives associated to
the metric $h_{ij}$.
Similar calculations can also be found in \cite{Leblond}, but with different notations.
It is better to parametrize the cosmological constants $\hat{\Lambda}$ and $\Lambda$ in a suitable way so that the solutions look more elegant. We choose the parametrization
\begin{align}
\hat{\Lambda}= \mathrm{sign}(\hat{\Lambda})\frac{(D-1)(D-2)}{2} \hat{k}^2, \quad
\Lambda = \mathrm{sign}({\Lambda})\frac{(d-1)(d-2)}{2} k^2, \label{lambda}
\end{align}
where $\mathrm{sign}(\hat{\Lambda})$ denotes the signature of
the cosmological constant $\hat{\Lambda}$ (which takes one of
the ``values'' $(-, 0, +)$), and $\hat{k}$ and $k$ are some
positive numbers. In section 4, we also write
\begin{align}
\hat{\kappa}^2= (D-1) \hat{k}^2, \quad
\kappa^2 = (d-1)k^2. \label{lambda-2}
\end{align}
\section{Codimension 1 cases} \label{S2}
For $n=1$ we can fix $h_{D-1,D-1}=-1$ without loss of generality. These are the most familiar cases because all braneworld models with codimension 1 belong to this class. Some of the 9 types of embeddings were already known from the studies of braneworld theories \cite{Pope}, and we include them here both for completeness and for triggering the ansatz for the higher codimension cases.
Inserting equations (\ref{munu-D}) and (\ref{ij-D}) into (\ref{r-eq}) and taking
the parametrization (\ref{lambda}), we get the following equations for $a(y)$,
\begin{align}
&\frac{1}{2}\left[ a^{\prime \prime}(y)+\frac{D-1}{2}a^{\prime}(y)^{2}\right]
-\mathrm{sign}(\Lambda) (D-2) k^2e^{-a(y)}
+\mathrm{sign}(\hat{\Lambda}) (D-1) \hat{k}^2 =0 , \label{embed-1}\\
&\frac{1}{2}\left[ a^{\prime \prime}(y)+\frac{1}{2}a^{\prime}(y)^{2}\right]
+\mathrm{sign}(\hat{\Lambda}) \hat{k}^2 = 0. \label{embed-2}
\end{align}
Notice that (\ref{embed-1}) and (\ref{embed-2}) is a system of two equations for a single unknown function $a(y)$, so
the existence of solutions is not guaranteed.
Different choices of $\mathrm{sign}(\hat{\Lambda})$ and $\mathrm{sign}({\Lambda})$ correspond to different types of embeddings mentioned in (\ref{9cases}), and the corresponding solutions varies drastically.
We list the results here in a case by case manner.
\begin{enumerate}
\item {$AdS_{D-1}\subset AdS_{D}$}.
In this case both $\hat{\Lambda}$ and $\Lambda$ are negative. The common solution to (\ref{embed-1}) and (\ref{embed-2}) is given by
\begin{equation}
\exp(a(y))=\left(\frac{k}{\hat{k}}\right)^2 \cosh^2(\hat{k}y).
\label{C1}
\end{equation}
Here and below we shall always absorb the unimportant integration constant by a shift in the coordinate $y$.
\item {$AdS_{D-1}\subset M_{D}$}.
Now we have $\hat{\Lambda}=0$ and $\Lambda < 0$. The solution reads
\begin{equation}
\exp(a(y))=-\left(ky\right)^2. \label{ads-m}
\end{equation}
The minus sign on the right hand side of (\ref{ads-m}) is unusual. It implies flipping the roles of spacelike and timelike coordinates in $x^\mu$. So, if originally there were only one timelike coordinate in $x^\mu$, the minus sign would results in a metric with multiple timelike coordinates unless $D=3$.
\item {$AdS_{D-1}\subset dS_{D}$}, i.e.
$\mathrm{sign}(\hat{\Lambda})=+1$ and $\mathrm{sign}({\Lambda})=-1$. The solution is
\begin{equation}
\exp(a(y))=-\left(\frac{k}{\hat{k}}\right)^2\cos^2(\hat{k}y).
\end{equation}
Like in the previous case, this solution involves a bulk geometry with more than one timelike coordinates unless $D=3$.
\item {$M_{D-1}\subset AdS_{D}$}, i.e. $\hat{\Lambda}<0$ and $\Lambda=0$. The solution reads
\begin{equation}
\exp(a(y))=\exp(\pm 2\hat{k}y).
\end{equation}
This is the case on which the original Randall-Sundrum (RS) braneworld scenario \cite{Randall1,Randall2} was built. The difference between the RS braneworld model and the present embedding lies in that RS assumed an extra $Z_2$ symmetry which brought about a $\delta$-function-like discontinuity in the second derivatives of the metric, which is explained as the contribution from brane tension.
\item {$M_{D-1}\subset M_{D}$}.
Now both $\hat{\Lambda}$ and $\Lambda$ are zero. This is equivalent to saying that both
$\hat{k}$ and $k$ are zero. The solution reads
\begin{equation}
a(y)=const.
\end{equation}
This is perhaps the most uninteresting of all possible cases, because it implies no warping in the bulk geometry at all.
\item {$M_{D-1}\subset dS_{D}$}.
Now we have $\hat{\Lambda}>0$ and $\Lambda=0$. The only common solution to (\ref{embed-1}) and (\ref{embed-2}) is essentially complex,
\begin{equation}
\exp(a(y))=\exp(\pm 2i\hat{k} y),
\end{equation}
implying that the embedding $M_{D-1}\subset dS_{D}$ is impossible in principle if the bulk is a real manifold. This explains why there is no RS like braneworld model with a de Sitter bulk.
\item {$dS_{D-1}\subset AdS_{D}$}:
Here we have $\hat{\Lambda}<0$ and $\Lambda>0$. The solution is
\begin{equation}
\exp(a(y))=-\left(\frac{k}{\hat{k}}\right)^{2}\cosh^{2}(\hat{k}y).
\end{equation}
One sees that this type of embedding must also involve a bulk geometry with more than one timelike dimensions unless $D=3$.
\item {$dS_{D-1}\subset M_{D}$}, i.e. $\mathrm{sign}(\hat{\Lambda})=0$ and $\mathrm{sign}(\Lambda)=+1$. The solution
reads
\begin{equation}
\exp(a(y))=(ky)^{2}.
\end{equation}
This kind of embedding has already been used in the study of black rings and black strings \cite{Chu, Zhao}.
\item {$dS_{D-1}\subset dS_{D}$}, i.e. $\hat{\Lambda}>0$ and $\Lambda>0$. The corresponding solution is
\begin{equation}
\exp(a(y))=\left(\frac{k}{\hat{k}}\right)^{2} \cos^{2}(\hat{k}y).
\end{equation}
This is another interesting case which may found applications in the future.
\end{enumerate}
To summarize, not all of the 9 different types of embeddings yield physically interesting solutions. In some case the embedding is even impossible for real bulk manifold. We arrange the 9 different cases into the following classes:
\begin{itemize}
\item Real emdedings exist and are physically interesting. This class contains the embeddings $AdS_{D-1}\subset AdS_{D}$, $dS_{D-1}\subset dS_{D}$, $dS_{D-1}\subset M_{D}$ and $M_{D-1}\subset AdS_{D}$;
\item Real embedding exists but is trivial. This class contains only one case, i.e. $M_{D-1}\subset M_{D}$;
\item Real embeddings exist, but may involve bulk geometries with multiple timelike directions. Not knowing of any physical interpretations of spacetimes with multiple timelike directions, we regard this class of embeddings as containing some physical illness, however if Wick rotations are taken into account, such embeddings may still yield interesting results. This class contains the cases $AdS_{D-1}\subset M_{D}$, $AdS_{D-1}\subset dS_{D}$ and $dS_{D-1}\subset AdS_{D}$;
\item Real embedding is impossible. This class contains only one case, i.e. $M_{D-1}\subset dS_{D}$.
\end{itemize}
We remark here that the above classification makes sense only for codimension 1 embeddings. If more extra dimensions were allowed, some of the impossibilities or illnesses might be resolved. For instance, although the embedding $dS_{D-1} \subset AdS_{D}$ is ill in the sense that it requires more than one timelike directions in the bulk, the embedding $dS_{D-1} \subset AdS_{D+1}$ can be achieved without such illness, e.g. via the chain of embeddings $dS_{D-1} \subset M_{D}\subset AdS_{D+1}$. In the next section we shall consider some of the codimension $>1$ embeddings, but we will focus exclusively on the one step embeddings, i.e. excluding the above mentioned chain-like embeddings.
\section{Codimension $n>1$}
Comparing to the codimension 1 cases the biggest difference of codimension $n>1$ cases lies in that the geometry of the extra dimensional subspace (i.e. that described by the metric $h_{ij}$) is no longer trivial. Thus it is basically impossible to list all types of warped embeddings as we did for the codimension 1 cases, because each types of embeddings listed in (\ref{9cases}) would be subdivided into many different cases according to the choice of geometry $h_{ij}$. In this section, we shall only take the simplest choices of $h_{ij}$, so that it possesses an $SO(n)$ symmetry, i.e. contains an $(n-1)$-sphere as a subspace. Other possible choices will not be considered. Meanwhile, due to the fact that already in the codimension 1 cases some of the one-step embeddings yield solutions with physical illness or even give rise to no real solution, we shall only consider some of the 9 different types of embeddings listed in (\ref{9cases}), rather than make a complete case by case study. The equations (\ref{munu-D})-(\ref{ij-D}) are now too complicated to be solved directly, so we shall adopt a trial-and-check approach.
\subsection{$AdS_{d} \subset AdS_{D}$ with $n=D-d>1$} \label{cod-higher}
We need to make an assumption for the bulk metric which reduces to the solution (\ref{C1}) in the limit $n=1$. A mathematically viable ansatz for the metric of this kind reads
\begin{equation}
ds^{2}=\left(\frac{k}{\hat{k}}\right)^{2}\cosh^{2}(\hat{k}\rho)\left(g_{\mu\nu}dx^{\mu}dx^{\nu}\right)-d\rho^{2}-A^{2}(\rho)d\Omega_{n-1},
\label{assump}
\end{equation}
where $A(\rho)$ is to be determined by the embedding equations and $d\Omega_{n-1}$ represents the metric of a unit $(n-1)$-sphere. Using the embedding equations we can get
\begin{displaymath}
A(\rho)=\hat{k}^{-1}\sinh(\hat{k}\rho),
\end{displaymath}
i.e. the final bulk metric takes the form
\begin{equation}
ds^{2}=\left(\frac{k}{\hat{k}}\right)^{2}\cosh^{2}(\hat{k}\rho)\left(g_{\mu\nu}dx^{\mu}dx^{\nu}\right)-d\rho^{2}-\hat{k}^{-2}\sinh^{2}(\hat{k}\rho)d\Omega_{n-1}. \label{adsads-n}
\end{equation}
The solution (\ref{adsads-n}) has some interesting features. If we omit the part of the $d$-dimensional submanifold and look only at the extra $n$-dimensional subspace, the metric turns out to be that of an $n$-dimensional cone, with an $(n-1)$-spheric foliation along the $\rho$ axis. The tip of the cone is located at $\rho=0$, where the size of the $(n-1)$-sphere shrinks to zero. That is to say, if we come close to $\rho=0$ in the $D$-dimensional metric (\ref{adsads-n}), the spacetime would appear to be $d$-dimensional. So there is a spontaneous compactification effect near the tip of the cone. This whole picture is very similar to the Klebanov-Strassler model of superstring cosmology \cite{Klebanov}, in which our 4-dimensional universe ``dances at the tip of a pin'' (the end of Klebanov-Strassler throat). However the problem we are considering here is much simpler: we only consider standard General Relativity and no supersymmetry is needed in this picture.
\subsection{$dS_{d} \subset dS_{D}$ with $n=D-d>1$}
Similar considerations can be carried out for the case $dS_{d} \subset dS_{D}$. The solution turns out to be
\begin{equation}
ds^{2}=\left(\frac{k}{\hat{k}}\right)^{2}\cos^{2}(\hat{k}\rho)\left(g_{\mu\nu}dx^{\mu}dx^{\nu}\right)-d\rho^{2}-\hat{k}^{-2}\sin^{2}(\hat{k}\rho)d\Omega_{n-1}. \label{dsds-n}
\end{equation}
Now the geometry of the extra $n$-dimensional subspace is no longer a cone but rather an $n$-sphere. It is interesting to note that there is a see-saw mechanism between the factors $\cos^{2}(\hat{k}\rho)$ in front of the $d$-dimensional submanifold and $\sin^{2}(\hat{k}\rho)$ in front of the $(n-1)$-sphere. Explicitly, near the poles $\hat{k}\rho =0, \pi$ of the $n$-dimensional sphere, the whole spacetime looks $d$-dimensional, while near the equator $\hat{k}\rho =\pi/2$ of the $n$-dimensional sphere, the whole spacetime looks $(n-1)$-dimensional. In either cases the effective dimension of the bulk spacetime is drastically decreased, so this can be understood as another type of spontaneous compactification.
\subsection{$M_{d} \subset M_{D}$ with $n=D-d>1$}
Already in the codimension 1 case the embedding of $M_{D-1}$ into $M_{D}$ has shown its trivialness. No surprises occur at higher codimensions. The only solution we get in the spirit of subsection \ref{cod-higher} is the following,
\begin{equation}
ds^{2}=g_{\mu\nu}dx^{\mu}dx^{\nu}-d\rho^{2}-\rho^{2}d\Omega_{n-1}. \label{mm-n}
\end{equation}
This is nothing but a direct sum of the $d$-dimensional flat submanifold with an $n$-dimensional Euclidean space. Such solutions were used in the construction of (unstable) black branes \cite{GL} in the literature.
\subsection{Other cases}
Though it is our hope to analyse every cases listed in (\ref{9cases}) in detail, it turned out that finding explicit solutions for other types of warped embeddings is very difficult. Even the existence of a real analytic solution is not guaranteed. However, we can check by direct calculations that the following is a valid embedding from a 4-dimensional de Sitter spacetime to a 7-dimensional flat spacetime:
\[
ds^{2}=\frac{1}{5}\left[\Lambda g_{\mu\nu}dx^{\mu}dx^{\nu}-d\rho^{2}-\rho^{2}d\Omega_{2}\right],
\]
but we simply do not know of any generalizations to arbitrary dimensions.
\section{Codimension $n>1$: another branch of solutions}
The solutions described in the last section rely on the assumption (\ref{assump}) or the like, which require that in the limit of $n=1$ the codimension 1 solutions should be recovered. Since the codimension $n>1$ cases have much richer structure, it is natural to ask whether there exist solutions which do not approch the codimension 1 solutions in the limit $n=1$. To answer this question, we can make no reference to the results accumulated in section \ref{S2}. So we take another route: fix the geometry of the extra $n$-dimensional subspace as in the last section and see if there exist other solutions for the warp factors. This seems to be a very strange route to take, but, as it turns out, there indeed exist positive answers to the question.
\subsection{$AdS_{d} \subset AdS_{D}$ with $n=D-d>1$}
Instead of (\ref{assump}), we now take the ansatz
\begin{equation}
ds^{2}=B^{2}(\rho)\left(g_{\mu\nu}dx^{\mu}dx^{\nu}\right)-d\rho^{2}-\hat{\kappa}^{-2}\sinh^{2}(\hat{\kappa}\rho)
d\Omega_{n-1}, \label{assump2}
\end{equation}
where $B(\rho)$ is to be determined by the embedding equations. Notice that in writing (\ref{assump2}), we have implicitly used the relation (\ref{lambda-2}).
Inserting (\ref{assump2}) into the embedding equations we find, after straightforward calculations, that the only solution for $B(\rho)$ is a constant, $
B(\rho)=\left(\frac{\kappa}{\hat{\kappa}}\right)^2$.
Therefore the bulk metric in this case looks as follows,
\begin{equation}
ds^{2}=\left(\frac{\kappa}{\hat{\kappa}}\right)^2\left(g_{\mu\nu}dx^{\mu}dx^{\nu}\right)-d\rho^{2}-\hat{\kappa}^{-2}\sinh^{2}(\hat{\kappa}\rho)
d\Omega_{n-1}. \label{adsads-n2}
\end{equation}
This solution is quite similar to the one given by (\ref{adsads-n}) but for one thing: the warp factor in front of $g_{\mu\nu}$ becomes constant. Thus the present solution belongs to a completely different branch of solutions. Note that the existence of two different branches of solutions to the embedding equations is not a supprise: similar phenomena has already been reported in \cite{Kinoshita} in the context of Fruend-Robin compactification. The present solution also contains an $n$-dimensional cone metric as a factor, and at the tip $\rho=0$ of the cone, the spacetime becomes $d$-dimensional without the aid of a Fruend-Robin field. So the spontaneous compactification mechanism is still present in (\ref{adsads-n2}), just like in
(\ref{adsads-n}).
\subsection{$dS_{d} \subset dS_{D}$ with $n=D-d>1$}
We can make analogous assumptions like (\ref{assump2}) in the case of embedding $dS_{d} \subset dS_{D}$. Following similar steps we get the solution
\begin{equation}
ds^{2}=\left(\frac{\kappa}{\hat{\kappa}}\right)^2\left(g_{\mu\nu}dx^{\mu}dx^{\nu}\right)-d\rho^{2}-\hat{\kappa}^{-2}\sin^{2}(\hat{\kappa}\rho)
d\Omega_{n-1}. \label{dsds-n2}
\end{equation}
This solution also looks like (\ref{dsds-n}) but with the warp factor changed into a constant. Similar to (\ref{dsds-n}), the geometry of the extra $n$-dimensional subspace is an $n$-sphere. But unlike (\ref{dsds-n}), the missing warp factor in front of $g_{\mu\nu}$ makes the see-saw mechanism in the bulk metric (\ref{dsds-n2}) disappear. Accordingly, the bulk metric will still reduce to a $d$-dimensional Einstein manifold at the poles $\hat{\kappa}\rho=0,\pi$ of the $n$-sphere, but it will not reduce to an $(n-1)$-sphere at the equator $\hat{\kappa}\rho=\pi/2$ of the $n$-sphere.
\subsection{Other cases}
For the embedding $M_{d}\subset M_{D}$, there is no difference between the two parametrization schemes (\ref{lambda}) and (\ref{lambda-2}). Thus we should get the same answer as in (\ref{mm-n}) even if we start from an ansatz like (\ref{assump2}) and taking the limit $\hat{\kappa}\rightarrow 0$.
We don't have explicit solutions to other cases like $dS_{d} \subset M_{D}$ or $M_{d} \subset AdS_{D}$ at hand.
However we expect that there should be some interesting solutions to such embeddings. We wish to make further study on such cases in later works.
\section{Discussions}
Embeddings between Einstein manifolds of different dimensions have found significant applications in various areas in gravitational physics. In this article, we presented some explicit results of such warped embeddings, focusing on the
codimension $n>1$ cases. Now let us make some discussions on the potential applications of the codimension $n>1$ solutions obtained in this article.
The first potential application is on a possible spontaneous compactification mechanism. As we have already noticed earlier, for the embeddings $AdS_{d} \subset AdS_{D}$, both branches of solutions share the common feature that the bulk metric reduces to that of the $d$-dimensional submanifold at the tips of the extra $n$-dimensional cone, and for the embeddings $dS_{d} \subset dS_{D}$, both branches of solutions share a similar feature: the bulk metric reduces to that of the $d$-dimensional submanifold at the poles of the $n$-sphere.
Therefore it is interesting to ask whether there exists a mechanism such that it makes the $d$-dimensional submanifolds prefer to stay at the special positions mentioned above (tips of the cone or poles of the $n$-spheres). Possible choices for the mechanisms include introducing tension terms for the $d$-dimensional submanifolds or adding matter to these submanifolds.
Another potential application is in the studies of higher dimensional black holes and black branes. The trivial embedding (\ref{mm-n}) and its codimension 1 analogue has already been used in the studies of black branes and strings and was shown to give rise to black brane/string instabilities \cite{GL}. In the last few years many five-dimensional black hole solutions with horizon topology $S^2 \times S^1$ (i.e. ``black rings'') were found in flat spacetime (i.e. with vanishing cosmological constant). Among these black rings the so-called large black rings are thermodynamically more favored and are believed to be more stable. Attempts in finding black ring solutions in the presence of a cosmological constant have since been made consecutively. However no analytic, singularity-free solutions of this kind have been found to this date.
We postulate that the embeddings considered in this article may be helpful in constructing higher dimensional black ring solutions with nonvanishing cosmological constant. However the application of codimension 1 embeddings in constructing black ring solutions was a failure \cite{Chu}: the resulting spacetime contains naked singularity and hence is not a black ring solution in the usual sense. The inclusion of more extra dimensions may help to resolve the naked singularity problem and give rise to physically acceptable black ring solutions in higher dimensions. We shall attempt to make more explorations in this direction in the future.
\section*{Acknowledgement}
This work was initiated when both authors were visiting KITPC. L.Z. is supported by the Natural Science Foundation of China (NSFC) through grant No. 10875059.
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
In recent years intensive investigations of surface plasmon
polaritons (SPPs) have been made in the promising context of
nanophotonics. This research is actually motivated by the
current trends for optical device miniaturisation and by the
possibilities of merging aspects of nanophotonics with those of
electronics. SPPs are electromagnetic waves bounded to
dielectric-metal interface. As surface waves, SPPs are
exponentially damped in the directions perpendicular to the
interface \cite{Raether}. Furthermore, SPPs could be used to
transfer optical information in a two dimensional (2D) environment.
This appealing property can be used for optical addressing of
different 2D optical systems and nanostructures located at a
dielectric/metal interface. Actually several 2D SPP devices
including passive nanostructures including mirrors or beam splitter
and active elements like molecules or quantum dots are currently
under development and investigation. Developments such as these
raise the prospect of a new branch of photonics using SPPs, for
which the term "plasmonics" emerged \cite{Barnes:2003,Ebbesen:2007,Drezet:Micron}. \\
However, for experimental investigations of optical devices an
important characteristic of SPP modes is that their spatial
extent is governed and defined by the geometry of the nanoelements
rather that by the optical wavelength \cite{Krenn:1999}. This
consequently opens possibilities for breaking the diffraction
limit but requires instruments of observation adapted essentially
to the subwalength regime and being capable of imaging the
propagation of SPPs in their 2D environment. Usually the analysis of
the subwalength regime implies necessarily near field optical (NFO)
methods \cite{Pohl,Courjon:2003} able to collect the evanescent (i.~e., non
radiative) components of the electromagnetic fields associated
with SPPs. However, when the metal film on which the 2D optical
elements are built is thin enough (i.~e~., with a thickness below
80-100 nm) and when the subtratum optical constant (usually glass)
is higher than the one of the superstratum medium an other
possibility for analyzing SPP propagation occurs. This possibility
is based on the detection of coherent leaking of SPPs through the
substratum. Such a far-field optical method is called leakage
radiation microscopy (LRM) \cite{Hecht,Bouhelier,Stepanov} and allows indeed a direct quantitative
imaging and analysis of SPP propagation on thin metal films.\\
The aim of this article is to present a short overview of recent
progress in the field of SPP imaging using LRM. In a first part of
this work we will describe the theoretical principles underlying
LRM. In the second part we will discuss modern leakage radiation
methods and illustrate the LRM potentialities by analyzing few
experiments with SPP waves interacting with 2D plasmonic devices.
\section{Leakage radiation and surface plasmon polaritons}
In order to describe the theoretical mechanisms explaining leakage
radiation it will be sufficient for the present purpose to limit
our analysis to the case of a metal film of complex permittivity
$\epsilon_1(\omega)=\epsilon'_1+i \epsilon''_1$ ($\omega=2\pi
c/\lambda$ is the pulsation) sandwiched between two dielectric
media of permittivity $\epsilon_0$ (substrate) and
$\epsilon_2<\epsilon_0$ (superstratum). This system is
theoretically simple and to a good extent experimentally
accessible \cite{Raether,Burke}. In the limiting case where the
film thickness $D$ is much bigger than the SPP penetration length
in the metal (i.~e., $D\gtrsim 70$ nm for gold or silver in the
visible domain) one can treat the problem as two uncoupled single
interfaces. We will consider as an example the interface 0/1 (the
media 0 and 1 are located in the domain $z\geq 0$ and $z\leq 0$
respectively ). Such an interface will be identified in the
following with the plane $z=const.$ in cartesian coordinates. An
elementary harmonic SPP wave is actually a TM electromagnetic mode
characterized by its pulsation $\omega$ and its magnetic field
$\mathbf{H}=[0,H_y,0]$ where the y component can be written
\begin{eqnarray}
H_0=\alpha e^{ik_x x}e^{ik_{z0} z}e^{-i\omega t} & \textrm{in the medium 0} \nonumber \\
H_1=e^{ik_x x}e^{ik_{z1} z}e^{-i\omega t} & \textrm{in the medium
1,}
\end{eqnarray}and where $k_{x}=k'_x+ik''_x$ is the (complex valued) wavevector of the SPP propagating in the x direction along the interface.
$k_{zj}\equiv k_{j}=\pm\sqrt{[(\omega/c)^2\epsilon_i-k_x^2]}$ are
the wave vectors in the medium j =[0 (dielectric), 1 (metal)]
along the direction $z$ normal to the interface. By applying
boundary conditions to Maxwell's equations one deduces
additionally $\alpha=1$ and
\begin{eqnarray}
\frac{k_1}{\epsilon_1}-\frac{k_0}{\epsilon_0}=0,
\end{eqnarray}
which implies
\begin{equation}
k_{x}=\pm(\omega/c)\sqrt{\frac{\epsilon_{0}\epsilon_1}{\epsilon_{0}+\epsilon_1}}
\end{equation}
\begin{equation}
k_{j}=\pm(\omega/c)\sqrt{\frac{\epsilon_{j}^2}{\epsilon_{0}+\epsilon_1}}
\end{equation}
for a SPP wave propagating along the x direction. The choice of
the sign convention connecting the z and x components of the wave
vector is a priori arbitrary and must be done only on a physical
ground.\begin{figure}[hbtp]
\includegraphics[width=8cm]{Fig1.eps}
\caption{Structure of the SPP magnetic field $\textrm{Real}[H_y]$
across an interface air/gold (thick line) and glass/gold (dashed
line). The optical wavelength considered is $\lambda=800$ nm. The
permittivity of glass is taken to be
$\epsilon_{\textrm{glass}}=2.25$. The inset shows the conventions
for the axes x and z. The arrows indicate the direction of the
real part of the wave vector normally and parallely to the
interface.}
\end{figure} Indeed, due to ohmic losses in the metal we expect an exponentially decaying SPP wave
propagating along the interface. This condition implies the
relation $k'_x\cdot k''_x\geq 0$ \cite{Burke}. This inequality is
actually always fulfilled since from Eq.~3 one deduces
\begin{equation}
k'_x\cdot
k''_x=\frac{1}{2}(\omega/c)^2\frac{\epsilon_{0}^2\epsilon''_1}{(\epsilon_{0}+\epsilon'_1)^2+(\epsilon''_1)^2}>
0
\end{equation}
which is indeed positive because $\epsilon''_1>0$. By writing
$k_{zj}=k'_j+ik''_j$ one additionally obtains the relation
\begin{equation}
-k'_0\cdot k''_0=(\omega/c)^2\epsilon''_{1}/2-k'_1\cdot
k''_1=k'_x\cdot k''_x\geq 0.
\end{equation}
This relation fixes the sign conventions since the wave must also
decay exponentially when going away from the interface in both
media. More precisely one gets
\begin{equation}
k'_0\cdot k''_0\leq 0,
\end{equation}
\begin{equation}
k'_1\cdot k''_1=\geq 0. \end{equation}The product $k'_1\cdot
k''_1$ is positive if $\epsilon'_1\geq
-|\epsilon_1|^2/(2\epsilon_0)$, a fact which is indeed true for
silver and gold interfaces with air or glass in most of the
visible optical domain. However small negative values of Eq.~8
occur for silver close to the interband region around $\lambda\sim
350 nm$. Additionally a higher value of $\epsilon_0$ will also
change the sign in Eq.~8. Fig.~1 shows the behavior of the SPP
magnetic field close to an interface gold/air and gold/glass at
the optical wavelength $\lambda=800$ nm. At such a wavelength the
conditions given by Eqs.~5-8 impose the solutions
\begin{eqnarray}
k_{x}=\pm(\omega/c)\sqrt{\frac{\epsilon_{0}\epsilon_1}{\epsilon_{0}+\epsilon_1}},&
k_{j}=-(\omega/c)\sqrt{\frac{\epsilon_{j}^2}{\epsilon_{0}+\epsilon_1}}.
\end{eqnarray}
The real parts of the $k_z$ components of the SPP wave vector are
for both media oriented in the same direction corresponding to a
wave propagating from the air side to the metal side (see inset in
Fig.~1). Furthermore the waves are exponentially damped when going
away from the interface in agreement with Eq.~7, 8 (see Fig.~1).
Most important for us is that the Poynting vector \cite{Jackson}
$\mathbf{S}=\textrm{Real}[\mathbf{E}\times \mathbf{H}^{\ast}]/2$
is defined in the medium j by
\begin{eqnarray}
\mathbf{S}_j=\frac{1}{2}c\textrm{Real}[\frac{k_x
\hat{\mathbf{x}}+k_j\hat{\mathbf{z}}}{\omega\epsilon_j/c}]e^{-2k''_x
x-2k''_z z}.
\end{eqnarray}
On the dielectric side the energy flow is as expected oriented in
the direction of $\textrm{Real}[\mathbf{k}]$. However it can be
shown on the metal side and for wavelengths not too close from the
spectral region associated with the interband transition of gold
or silver that the energy flow in the x direction is oriented
oppositely to the wave vector $\textrm{Real}[k_x]$ since
$\textrm{Real}[k_x/\epsilon_1]=(k'_x\epsilon'_1+k''_x\epsilon''_1)/|\epsilon_1|^{2}$
is dominated by $k'_x\cdot \epsilon'_1$ and since $\epsilon'_1<0$.
However the total energy flow in the x direction $S_x=
\int_0^{+\infty}S_x^{(\textrm{air})}dz+\int_{-\infty}^0
S_x^{(\textrm{metal})}dz$ is oriented along $k'_x$. Additionally
in the z direction
$\textrm{Real}[k_1/\epsilon_1]=(k'_1\epsilon'_1+k''_1\epsilon''_1)/|\epsilon_1|^{2}$
is parallel to $k'_1$ since $k''_1\epsilon''_1$ dominate. This
implies in particular that the energy associated with the SPP wave
is absorbed by the metal during its propagation through the
interface from the air side to the glass side. It should be
observed that already in the case of the ideal plasma model
neglecting losses the wave vector $k_x=k'_x$ is antiparallel to
$S_x$ in the metal since there is no imaginary part and since
$\epsilon_1=1-\omega_p^2/\omega^2<0$
(see also \cite{Zhinzhin,Kats}). \\
We show in Fig.~2 the curves associated with the dispersion
relations of SPPs propagating along a gold/air and gold/glass
interface respectively. Fig.~2A represents the dependencies
$\omega$ versus $k'_x$ corresponding to Eq.~3. Fig.~2B shows the
dependencies $\omega$ versus $L_{SPP}$ where $L_{SPP}=1/(2k''_x)$
is the propagation length of the SPP waves (for the metal optical
constant we used the experimental values given in \cite{Johnson}).
The typical back bending around $\lambda=520 nm$ corresponds to
the resonance associated with the bound electrons and the
interband transition (for a good discussion see
\cite{Novotny:2006}). Far away from the interband the real part of
the dispersion is close to the asymptotic light lines: we speak
about Zenneck surface modes by opposition to Fano and evanescent
modes existing close to the interband \cite{Zhinzhin,Halevi}.
\begin{figure}[hbtp]
\includegraphics[width=8cm]{Fig2.eps}
\caption{Dispersion relations of SPP waves propagating along a
air/gold (thin dashed curve) and glass/gold (thick dashed curve)
interface. (A) Real part of the dispersion relations showing the
evolution of $k'_x$ with $\omega/c=2\pi/\lambda$. The light cones
corresponding to an optical wave propagating in air and glass
(i.~e., in the bulk medium) are represented by a thin and thick
continuous line respectively. The back bending at $\lambda=520$ nm
correspond to the interband resonance. (B) Imaginary part of the
dispersion relations showing the evolution of $L_{SPP}$ with
$\omega/c=2\pi/\lambda$. The damping is higher for the glass/gold
interface.}
\end{figure}
An important feature occurs close to this interband transition
since the slope in Fig.~2A is diverging. This means that the group
velocity defined by $v_g=\partial\omega/\partial k'_x$ is
infinite. One can qualitatively deduce that there is however no paradox with causality by remarking that the
propagation length is approaching zero at such wavelengths.
Actually the SPP life time $\tau_{SPP}$ defined by $v_g \cdot
\tau_{SPP}\approx L_{SPP}$ is tending as well to zero. This is a
clear signature of the absence of significant SPP propagation in the
interband spectral region. This fact actually precludes faster than light information
transfer by SPP waves and is in agreement with relativistic causality.\\
A second important feature concerning Fig.~2 A is that the
air/gold dispersion curve is located inside of the light cone for
glass defined by the equation
$\mathbf{K}^2=(\omega/c)^2\epsilon_{\textrm{glass}}$ where
$\mathbf{K}$ is a real light wave vector and
$\epsilon_{\textrm{glass}}=2.25$. Writing $\mathbf{K}=[K_x,0,K_z]$
the wavevector of a TM (i.~e., p polarized) plane wave propagating
away from the interfaces into the glass side one see that SPPs
propagating at the air/metal interface can radiate into the glass
substrate if the condition
\begin{equation}
K_{x}\approx\pm
\textrm{Real}[(\omega/c)\sqrt{\frac{\epsilon_{0}\epsilon_1}{\epsilon_{0}+\epsilon_1}}]
\end{equation} is approximately fulfilled.
Here we neglected the role of the imaginary part in Eq.~3. Similarly one can deduce that none of the SPPs propagating at the
two interfaces can radiates into the air side. \\
In order to have a more complete analysis one must actually consider the
problem with two coupled interface 0/1 (glass/metal) and 1/2 (metal/air) supporting SPP
waves and separated by a small distance $D$. The two interfaces are coupled by evanescent SPP waves tunnelling through the metal slab.
Such a mathematical problem can only be treated numerically by resolving an implicit equation. As
for the single interface this equation can be defined
directly from Maxwell Equations \cite{Burke}. However it is much easier and convenient for the following
to remark with Raether \cite{Raether} that Eq.~2 and consequently Eq.~3 are
obtained by finding the zeros of the numerator in the Fresnel
reflectivity coefficient for a TM wave coming from the dielectric
side:
\begin{equation}
R_{0,1}^{p}=\frac{(k_{0}/\epsilon_{0}-k_{1}/\epsilon_{1})}{(k_{1}/
\epsilon_{1}+k_{0}/\epsilon_{0})},\end{equation} with
$k_{j}=\pm\sqrt{((\omega/c)^2\epsilon_{j}-k_{x}^{2})}$. Actually
Raether \cite{Raether} reasoned with the denominator of the
Fresnel coefficient due to different conventions for the signs of
the wave vectors $k_j$. However it is remarkable that the result
is the same at the end of the calculations. Identically one can
thus define the Fresnel coefficient for a TM wave reflected by the
slab 0-1-2 \cite{Jackson,Novotny:2006,Raether}:
\begin{equation}
R_{0,1,2}^{p}=\frac{R_{0,1}^{p}+R_{1,2}^{p}e^{2ik_{1}D}}{1+R_{0,1}^{p}R_{1,2}^{p}e^{2ik_{1}D}},\end{equation}
and find the zeros of the numerator, i.~e., one can solve the
implicit equation
\begin{equation}
R_{0,1}^{p}+R_{1,2}^{p}e^{2ik_{1}D}=0,
\end{equation}
in order to define the SPP dispersion relations. From this
equation it follows that
\begin{eqnarray}
(k_{0}/\epsilon_{0}+k_{1}/\epsilon_{1})(k_{2}/\epsilon_{2}-k_{1}/\epsilon_{1})e^{ik_{1}D}\nonumber\\
+(k_{1}/\epsilon_{1}-k_{0}/\epsilon_{0})(k_{2}/\epsilon_{2}+k_{1}/\epsilon_{1})e^{-ik_{1}D}=0.
\end{eqnarray}
As for the single interface one has an important relation between
the real and imaginary parts of the SPP wave vectors in the
different medium:
\begin{equation}
-k'_0\cdot k''_0=(\omega/c)^2\epsilon''_{1}/2-k'_1\cdot
k''_1=-k'_2\cdot k''=k'_x\cdot k''_x.
\end{equation}
Since we are interested only into the solutions which are decaying
along the interface we (in agreement with our previous treatment
of the single interface) suppose the condition $k'_x\cdot
k''_x\geq 0$ satisfied. A second important point is that due to
the arbitrariness in the sign of $k_j$ there are in fact apriori 8
possibilities for writing Eq.~15. However, Eq.~15 is invariant
under the transformation $k_1 \rightarrow -k_1$. This means that
the number of apriori possibilities for the sign of $k_j$ is
reduced from 8 to 4. This multiplicity was studied by Burke
\emph{et al.} \cite{Burke} however since for the present purpose
we are looking for SPP waves leaking from the air/metal interface
into the glass substrate we will consider only the possibility
\begin{eqnarray}
k_{0}=-\sqrt{((\omega/c)^2\epsilon_{0}-k_{x}^{2})}\nonumber\\
k_{1}=+\sqrt{((\omega/c)^2\epsilon_{1}-k_{x}^{2})}\nonumber\\
k_{2}=-\sqrt{((\omega/c)^2\epsilon_{2}-k_{x}^{2})}.
\end{eqnarray}
The sign of $k_1$ is however arbitrary as explained above and we
choose it here positive by definition. In order to define a SPP
wave leaking into the glass substrate one has thus to substitute
Eq.~17 into Eq.15 and find numerically (i.~e., by a minimization
procedure \cite{Hohenau,Dionne}) the zeros of the implicit
equation with variable $k'_x$ and $k''_x$. \begin{figure}[h]
\includegraphics[width=8cm]{Fig3.eps}
\caption{Dispersion relations of SPP waves propagating along an
air/gold interface and leaking into glass. The dispersion is
calculated for four film thickness $D$ : 70 nm (dashed line
curve)es, 50 nm (continuous line with triangular markers), 20 nm
(continuous line with square markers), and 10 nm (continuous line
with circular markers).(A) Real part of the dispersion relations
showing the evolution of $k'_x$ with $\omega/c=2\pi/\lambda$. The
light cone corresponding to optical wave propagating in air
(i.~e., in the bulk medium) is represented by a continuous line.
(B) Imaginary part of the dispersion relations showing the
evolution of $L_{SPP}$ with $\omega/c=2\pi/\lambda$.}
\end{figure}This has to be done only in the
quarter of the complex plane corresponding to $k'_x>0$, $k_x''>0$.
The quarter $k'_x<0$, $k_x''<0$ must be equivalent due to
symmetry and corresponds actually to decaying SPP waves
propagating in the negative x direction. The two other quarters of
the complex plane correspond to growing SPP waves along the
interface and will be rejected on a physical ground (compare
\cite{Burke}).
Figs.~3A and 3B show numerical calculations of dispersion
relations corresponding to a SPP wave leaking through a gold film
from the air side to the glass side. The thickness is taken to be
$D$=70, 50, 20, and 10 nm respectively. For the value $D\geq 70$
nm the dispersion relation is identical to the dispersion for the
single air/gold interface for semi infinite media. However, for
smaller thickness the coupling between the interface increases and
the propagation length decreases as shown on Fig.~3 B.
\begin{figure}[hbtp]
\includegraphics[width=8cm]{Fig4.eps}
\caption{ Structure of the magnetic field $\textrm{Real}[H_y]$
associated with a SPP mode leaking through a gold film of
thickness 50 nm (medium 1) from the air side (medium 2) to the
glass side (medium 0). The evolution is represented along the
normal $z$ to the interfaces. The optical wavelength considered is
$\lambda=800$ nm. The two horizontal lines show the interfaces
separated by 50 nm. The inset shows the conventions for the axis x
and z. The arrows indicate the direction of the real part of the
wave vector normally and parallely to the interface.}
\end{figure}
The magnetic field associated with SPP electromagnetic modes in this layered system is given by
\begin{eqnarray}
H_0=e^{ik_x x}e^{ik_{0} z}e^{-i\omega t} \nonumber \\
H_1=e^{ik_x x}[\alpha\sin{(k_{1}z)}+\beta\cos{(k_{1}z)}]e^{-i\omega t}\nonumber \\
H_2=\gamma e^{ik_x x}e^{ik_{2} (z-D)}e^{-i\omega t}.
\end{eqnarray}
The coefficients $\alpha, \beta, \gamma$ are obtained by
considering the boundary conditions and one finds
\begin{eqnarray}
\alpha=i\frac{k_0\epsilon_1}{k_1\epsilon_0}, \beta=1 \nonumber\\
\gamma=i\frac{k_0\epsilon_1}{k_1\epsilon_0}\sin{(k_{1}D)}+\cos{(k_{1}D)}.
\end{eqnarray}
As an illustration we show in Fig.~4 the evolution of the real
part of the magnetic field across a 50 nm thick gold slab
sandwiched between the glass substratum and the air superstratum
for an optical wavelength $\lambda=800 nm$. As visible the SPP
field is located in the vicinity of the air/gold interface and is
evanescent on the air side. This is clearly reminiscent of our
former analysis of SPPs propagating at the single air/metal. In
addition however the wave is leaking radiatively (i.~e.,
propagatively) into the glass substrate. However from Eq.~18 and
the condition $k'_0\cdot k''_0\leq 0$ it is clear that the leaking
wave is exponentially growing in the -z direction when going away
from the gold slab. This is already the result we obtained when we
considered the limit of the thick slab. An exponentially growing
wave looks non physical and is in particular associated with
infinite radiated energy in the far field. There is now the
question of how to connect a growing wave with the basic reasoning
giving rise to Eq.~11 and the idea of phase matching between the
(real part) of the SPP wave vector with a propagative plane wave
vector in the glass substrate. However such paradoxes disappear
for two reasons: First, an infinite energy occurs only because we
considered an infinite interface or equivalently because we did
not consider how the SPP is locally launched on the metal film.
When such conditions are taken into account this paradox must
disappear \cite{Burke}. Second, the SPP wave defined by Eq.~18 is
actually a wave packet when looked at through the Fourier basis of
propagative TM plane waves. Since in the far field (i.~e., in the
glass substrate) one actually detect such plane waves one must do
a Fourier transform in order to generalize Eq.~11
\cite{Raether,Burke}. Instead of Eq.~11 one obtains consequently a
statistical distribution of (real) wavevectors $K_x$ given by
\begin{equation}
I(K_x)=\frac{\textrm{const.}}{(K_x-k'_x )^2+(k_x'')^2},
\end{equation}
where $2k_x''$ defines the full width at half maximum (FWHM) of
this Lorentz distribution of radiated plane waves. By noting as
usual $\theta$ the angle between the wave vectors $\mathbf{K}$ of
the plane waves refracted into the glass substrate and the normal
$z$ to the interfaces one has by definition
$K_x=2\pi\sqrt{\epsilon_0}/\lambda\sin{\theta}$ and the angular
distribution of radiated power is in the far field given by:
\begin{equation}
I(\theta)=\frac{\textrm{const.}}{(2\pi\sqrt{\epsilon_0}/\lambda\sin{\theta}-k'_x )^2+(k_x'')^2}.
\end{equation}
\section{Leakage radiation microscopy}
Historically the first observations of leakage radiation by SPP
propagating on a thin metal film were reported by analyzing
scattering by rough metal surfaces of light into SPPs
\cite{Raether,Simon}. The possibility of using rough surface to
excite SPPs was extensively studied in the past \cite{Raether} and
is based on the fact that the scattering by small defects on a
flat film can represent a source of evanescent momentum sufficient
for the light waves to match the SPP dispersion relation.
Equivalently the amount of momentum needed can be carried by
grating coupling \cite{Raether}. SPP waves are subsequently
emitted back into the glass substrate as leakage radiation (see
Fig.~5).
\begin{figure}[hbtp]
\includegraphics[width=6cm]{Fig5.eps}
\caption{Observation of leakage radiation (LR) through a rough
thin silver film (wavelength $\lambda=550 nm$). (A) SPPs are
launched on the air/silver interface by using scattering by film
rugosities to produce the amount of momentum necessary to match
the SPP dispersion relation. SPPs are leaking through the glass
substrate and detected using a photographic plate in the
far-field. (B) Photography showing the leakage radiation cone. The
photography is extracted from the work by H.~J Simon and
J.~K.~Guha \cite{Simon} ($\copyright$ Opt.~Comm., Elsevier, 1976).
}
\end{figure}
This light collected on a photographic plate forms a ring-like
distribution in agreement with Eqs.~11, 20, 21. The FWHM of the
SPP wavevector distribution is in direct correspondence with the
radial width of the ring \cite{Simon}. Further progress was
obtained recently with the development of near field scanning
optical microscopy (NSOM) which allows the local optical
excitation of evanescent waves in the vicinity of a metal surface.
Such evanescent waves can carry a sufficient amount of momentum to
couple to SPP waves. Direct observations have indeed confirmed
this principle \cite{Sonnichsen,Brun:2003,Hecht}. As an example we
show in Fig.~6 an experiment in which the NSOM tip launches SPPs
on an aluminum film which after interaction with a hole excites
optically some quantum dots (QDs) located below. The collected
signal shows a specific QD luminescence spectrum \cite{Brun:2002}.
By scanning the sample around the NSOM tip one can realize SPP
mapping since the hole acts a probe structure for the field
emitted by the tip. Quantitative analysis of the total
luninescence of the QDs associated with a given hole show clearly
that the QD excitation is mediated by SPPs propagating on the
aluminum film.
\begin{figure}[hbtp]
\includegraphics[width=6cm]{Fig6.eps}
\caption{(A) Sketch of the experiment using a NSOM tip to excite
SPPs on an aluminum film with nanoholes. The film containing 240
nm diameter holes covers a CdTe/ZnTe system of QDs whose spectral
luminescence induced by SPP excitation is characteristic of a
given hole. The laser excitation wavelength is 514.5 nm. (B)
Radial mapping of the SPP intensity. The SPP intensity is
proportional to the QDs luminescence of a given hole. The curve
represents the evolution of the total luminescence collected
(using a detector in the optical far field) as a function of the
distance $R$ separating the NSOM (source) from the hole (probe).
The experiment is made at 4.2 K \cite{Brun:2003} ($\copyright$
Europhys.~Lett., EDP, 2003).}
\end{figure}
Fig.~6B shows the radial dependence of the collected intensity.
These results agree well with a 2D SPP dipole model supposing an
effective dipole located at the tip apex \cite{Hecht,Brun:2003}
(see also the discussions concerning the Bethe-Bouwkamp
\cite{Bethe,Bouwkamp1,Bouwkamp2} theory of diffraction by a small
aperture in a metal film in
\cite{Karrai1,Karrai2,Drezet:EPL2001,Drezet:PRE2002,Drezet:EPL2004}).
Following this model the SPP wave can be theoretically modelled by
a scalar wave $\Psi(\rho,\theta)$ given by
\begin{equation}
\Psi(\rho,\phi)=\textrm{const.} \frac{e^{ik_{\textrm{SPP}}
\rho}\cos{\phi}}{\sqrt{\rho}},
\end{equation}
where $k_{\textrm{SPP}}=k'_x+ik''_x$ is given by Eq.~3, $\rho,
\phi $ are polar coordinates on the metal film and the origin of
the coordinate is taken at the dipole position. $\phi$ is the
angle between the dipole associated with the NSOM tip (parallel to
the polarization of the laser beam injected in the NSOM tip) and
the the radial vector $\mathbf{\rho}=[x,y]$. This simplified model
can be theoretically justified by using the Green Dyadic Formalism
\cite{Bozhe} and has been applied by many authors successfully
\cite{Hecht,Bouhelier,Brun:2003,Harry1,Harry2,Drezet:APL2005,Laluet}
to SPP waves propagating in various environments.\\
Several authors applied NSOM methods coupled to LRM
\cite{Hecht,Bouhelier}. In particular in \cite{Hecht} Hecht
\emph{et al.} realized an optical setup using an immersion oil
objective to collect the leakage radiation emitted by the NSOM tip
on gold or silver films (see fig.~7).
\begin{figure}[hbtp]
\includegraphics[width=6cm]{Fig7.eps}
\caption{(A) Principle of the experiment to generate SPP leakage
radiation (LR) using a NSOM tip. LR is collected using an
immersion oil objective (not shown here). (B) Map of SPP intensity
using a charge coupled device (CCD) camera to collect the LR
emitted through the glass substrate. SPPs are launched on a 60 nm
thick silver film at the optical laser wavelength $\lambda=633nm$.
The image in B is taken from B.~Hecht \emph{et al.} \cite{Hecht}
($\copyright$ Phys.~Rev.~Lett, American Physical Society, 1996). }
\end{figure}
The system shown in Fig.~7 B is a 60 nm thick silver film
optically excited by a NSOM tip at the laser wavelength
$\lambda=633nm$. It can be shown by analyzing Fig.~7 B that the
radiation pattern is well described by a 2D dipole model in
agreement with Eq.~22. In particular the SPP propagation length
was measured and is in fair agreement with our analysis in
Sect.~1. Additionally it was shown in \cite{Hecht} that one can
also analyze the Fourier distribution of SPP momentum (given by
Eqs.~20, 21)
by defocusing the objective lens. As expected SPP rings similar to the one of Fig.~5 were observed.\\
\begin{figure}[hbtp]
\includegraphics[width=6cm]{Fig8.eps}
\caption{(A) Principle of the experiment to generate SPP leakage
radiation (LR) using a focussed laser beam (objective $O_1$, 10x,
numerical aperture $NA=0.3$). LR is collected using an immersion
oil objective (objective $O_2$, 63x, $NA=1.3$) and refocussed on
the CCD camera by using an auxiliary lens. (B) Map of SPP
intensity using a CCD camera to collect the LR emitted through the
glass substrate. SPPs are launched on a 50 nm gold film at the
optical laser wavelength $\lambda=800nm$ from a gold ridge (50 nm
height, 150 nm width) represented by a white line. The size of the
laser spot with a diameter $\simeq 8 $ $\mu$m is represented by
the white dashed circle.}
\end{figure}
In the same context we developed in recent years a systematic
approach using far-field microscopy to analyze quantitatively the
interaction between SPPs and plasmonic devices by using LRM. The
nanodevices studied were all fabricated by electron beam
lithography (EBL) allowing the precise and reproducible tailoring
of metal and dielectric surfaces on a lateral size dimension down
to 20 nm \cite{EBL}. As an example we show in Fig.~8 B a LRM image
obtained by using a gold ridge (50 nm height, 150 nm width)
lithographed on a 50 nm thick gold film to launch two well
collimated and counter propagating SPP beams. These beams are
launched by focussing a laser beam with a microscope objective
(10x, numerical aperture $NA=0.3$) on the gold ridge. Scattering
by the nanostructure gives rise to evanescent waves supplying the
right amount of momentum necessary for generating a SPP wave. The
optical LRM setup is sketched in Fig.~8 A. Leakage radiation
emitted through the glass substrate is collected by an immersion
oil objective (63x, numerical aperture, $NA=1.3$). Light is
subsequently refocussed on a charge coupled device (CCD) camera.
The direct mapping of the SPP intensity with this method provides
a one-to-one correspondence between the 2D SPP intensity and the
image recorded on the CCD camera. It should be observed that an
incident laser beam with diameter $Dx$ is in the focal plane
(object plane) of the objective $O_1$ focussed into a disc of
diameter $w$ (i.~e., $w=$beam waist) such that
\begin{equation}
\tan{\alpha}=\frac{2\lambda}{\pi w}=\frac{Dx}{2f}
\end{equation}
where $f$ is the focal length of the objective and $\alpha$ the
divergence angle of the laser beam focussed on the sample. The
direct application of this version of the Heisenberg relation
\cite{Teich} implies that the divergence angle $\alpha_{SPP}$ of
the SPP beam launched on the metal film must equal the divergence
angle $\alpha$ of the impinging laser beam. This result is in good
agreement with the experimental case shown in Fig.~8 with $Dx=2$
mm, $w=8.3$ $\mu$m and $\alpha=3.5^{\circ}$. Changing the
objective focal length is a straightforward means to
obtain different divergence angles $\alpha_{SPP}$ (see for example \cite{Drezet:APL2006,Stepanov,Drezet:EPL2006}).\\
\begin{figure}[h]
\includegraphics[width=8cm]{Fig9.eps}
\caption{Experimental scheme for LRM. SPPs are excited by laser
light focused by the microscope objective O1 (50x, numerical
aperture, $NA=0.7$) onto a structured gold thin film on a glass
substrate. LR emitted into the glass substrate from the gold/glass
interface $\Sigma$ is collected by the immersion microscopy
objective O2 (63x, NA=1.25) and imaged by a CCD camera. Depending
on the lateral position of the convex lens L3 either the back
focal plane or the sample plane is imaged. BB beam block, L1,2,3
convex lenses, f focal length of L1 and L2; $\Sigma$, $\Sigma'$,
sample plane and image thereof; F, F' back focal plane and image
thereof \cite{Drezet:APL2006}.}
\end{figure}
As a further improvement it is possible to modify the previous
optical setup in order to image not only the direct space
information but also the momentum corresponding to the Fourier
space. It is indeed a well known fact of Fourier optics
\cite{Teich,Born,Novotny:2006} that such a mapping of the wave
vector distribution (as given by Eq.~20) is in principle always
possible by recording the LR light in the back focal plane $F$ of
the oil immersion objective. In the optical setup shown in Fig.~9
\cite{Drezet:APL2006} we realized a dual microscope able to image
SPP propagation in both the direct and Fourier space. In
particular the back focal plane $F$ of the oil immersion objective
$O_2$
\begin{figure}[h]
\includegraphics[width=8cm]{Fig10.eps}
\caption{(A) Direct space LRM imaging of two SPPs beams launched
from the ridge and propagating to the lower left ($L$) and upper
right to be partly reflected ($R$) and transmitted ($T$) by the
Bragg mirror. The image is taken with no beam block. Ridge and
Bragg mirror are indicated by the dashed line and rectangle,
respectively. (B) Fourier plane LRM imaging corresponding to (A).
C indicates the wave vector components of the directly transmitted
laser beam $C$; $R,L,T$ have the same meaning as in (A). (C)-(D)
are Fourier filtered images after removing the central beam $C$
with a beam block in $F'$ (see text). In (C) the reflected beam R
is removed by blocking the arc of the SPP ring \emph{R} shown in
(B). In (D) the transmitted beam T and the beam incident on the
Bragg mirror are removed by blocking the arc of SPP ring $T$ in
(B). Data from \cite{Drezet:APL2006} ($\copyright$,
Appl.~Phys.~Lett., American Institute of physics, 2006).}
\end{figure}
imaged onto a CCD camera in Fig.~9. With such a microscope it is
furthermore possible to act experimentally in the Fourier space
image plane $F'$. First we can thereby remove the directly
transmitted laser beam by using a beam block located on the
optical axis. As an application of this method of filtering we
consider the reflection of a SPP beam by an in-plane Bragg mirror.
SPPs are launched as before from a gold ridge (50 nm height, 150
nm width) lithographed on a 50 nm thick gold film. The Bragg
mirror \cite{Born} constitutes a one-dimensional lattice of
parallel gold ridges (50 nm height, 140 nm width) separated by a
distance $P$ defining the period of the lattice. The period $P$ is
connected to the SPP wavelength by
$\lambda_{SPP}=2\pi/k_{SPP}<\lambda$ and to the angle of incidence
reflection $\theta_{SPP}$ of the SPP beam relatively to the (in
plane) normal to the lattice by
\begin{equation}
P=\frac{\lambda_{SPP}}{2\cos{\theta_{SPP}}}.
\end{equation}
In the present case shown in Fig.~10 the Bragg mirror is optimized
for $\lambda=800 nm$ (i.~e., $\lambda_{SPP}=785 nm$) and for
$\theta_{SPP}=45^{\circ}$ incidence angle which means
$P\simeq555nm$. The experimental analysis of such a Bragg mirror
when the resonance condition ($\lambda$, $\theta_{SPP}$) is
fulfilled reveals a very high reflectivity of $R\simeq 95$\% (see,
for example, \cite{Harry2} for some earlier results on SPP Bragg
mirrors studied with fluorescence microscopy). However, in the
present experiment we choose an incident angle
$\theta=65^{\circ}$. As a consequence the reflectivity was much
lower (see Fig.~10 A and more details in \cite{Drezet:APL2006}).
\begin{figure}[h]
\includegraphics[width=6cm]{Fig11.eps}
\caption{(A) Fourier space cross-cut along the short solid line in
Fig.~10 B corresponding to the $L$ beam (see inset). Experimental
data (symbols) are compared with a Lorentz fit (solid curve). The
intensity is normalized by setting the maximum intensity of $L$ to
1. (B) Direct space cross-cut along the L beam in Fig. 10 A. Data
points (symbols) are compared to an exponential fit (solid line).
Data from \cite{Drezet:APL2006} ($\copyright$, Appl.~phys.~Lett.,
American Institute of physics, 2006). }
\end{figure}
This configuration reveals SPP interferences in the vicinity of
the mirror (Fig.~10 A). In Fig.~10 B we show the corresponding
Fourier space image. The different observed arcs of LR rings
correspond to the reflected (R), and transmitted plus incident (T)
beams. The L beam is associated with a SPP launched in the
direction to the left, i.~e, away from the mirror. C is the
directly transmitted laser beam distribution. By acting in the
Fourier plane image $F'$ of the LRM microscope we now block the
information associated with the central beam and with the R or T
beams \cite{Drezet:APL2006}. Thereby the according SPP beam images
are erased from the image plane and consequently weak intensity
beams otherwise observed by interference can be directly
analyzed. For further analysis we extracted radial cross-cuts
along the white lines as shown in Fig.~10 B \cite{Drezet:APL2006}.
Results are shown in Fig.~11 A for the cross cut along $L$. The
data points agree very well with a Lorentz fit given by Eqs.~20,
21. The FWHM of the Lorentzian distribution gives us a SPP
propagation length of $L_{SPP}=20$ $\mu$m. This value is in
perfect agreement with the cross-cut made along the beam $L$ in
the direct space image 10 A (see Fig.~11 B).
Both data agree also with values given by the dispersion relations discussed in sect.~1 (see Figs.~2B and 3B).\\
\begin{figure}[h]
\includegraphics[width=8cm]{Fig12.eps}
\caption{(A) Principle of SPP in-plane wavelength demultiplexing
using plasmonic crystal. SPPs launched from a gold ridge (50 nm
height, 150 nm width) are propagating in the direction of a 2D
rectangular lattice (plasmonic crystal) as described in the text.
(B) SEM images of a part of the crystal, Bragg lines are indicated
by dashed lines. Such crystals can be seen as two sets of Bragg
mirrors perpendiculary oriented to each other. (C) Sketch
representing the unit rectangular cell and the direction of the
reflected SPP beam for the two different wavelengths at which the
Bragg reflections occur. (D, E) LRM images of SPP interacting with
the demultiplexer at SPP wavelengths $\lambda_{x}\simeq 730$ nm
and $\lambda_{y}\simeq 784$ nm, respectively. The gold ridge and
plasmonic crystal are indicated by the gray line and the gray box
, respectively. The vector basis $\mathbf{x}, \mathbf{y}$ of the
latices is indicated in A and E. Data from
\cite{Drezet:Nanolett2007} ($\copyright$ Nanoletters, American
Chemical Society, 2007). }
\end{figure}
LRM was subsequently applied by us to many SPP in-plane devices
such as beam splitters \cite{Stepanov}, dielectric lenses, prisms
\cite{Hohenau} and wave-guides \cite{Steinberger}. In particular,
LRM experiments were compared to near field optical experiments (
photon scanning tunnelling microscopy) and showed good agreement
in the cases considered \cite{Hohenau,Steinberger}. LRM appears
thus in this context as a complementary far-field optical method
to NFO such as NSOM. LRM was applied as well for analyzing SPP
Bragg mirrors (with high reflectivity $R\simeq 90-95$\%),
interferometers \cite{Drezet:EPL2006,Drezet:plasmonics2006} and
SPP elliptical cavities \cite{Drezet:APL2005} or 2D SPP
microscopes \cite{submitted}. In this context we observed
\cite{Drezet:EPL2006} stationary SPP waves with very high
visibility
$V=(I_{\textrm{max}}-I_{\textrm{min}})/(I_{\textrm{max}}+I_{\textrm{min}})$
by using LRM. This proves directly that SPP wave coherence is
conserved in LRM and can exploited for quantitative analysis down
to the spatial resolution limit $\lambda_{SPP}/2$. Actually SPP
interferometers such as the ones described in
\cite{Drezet:APL2005,Drezet:plasmonics2006,Harry2} reveal clear
interference pattern and oscillation characteristics of these
devices.
It is thus possible to develop 2D interferometry for SPP waves having all the advantages of current macroscopic interferometry techniques.\\
We also mention the realization of plasmonic crystals (i.~e.,
photonic crystals for SPPs) which were studied using LRM (see
Fig.~12). In such devices \cite{Drezet:Nanolett2007} rectangular
2D latices made of gold nano-protrusions (200 nm diameter, 50 nm
height) deposited on a 50 nm thick gold film (see Fig.~12B) are
used to create photonic band gaps at specific SPP wavelengths
$\lambda_{x}\simeq 730$ nm and $\lambda_{y}\simeq 784$ nm (i.~e.,
laser wavelengths of respectively 750 nm and 800 nm) corresponding
to the two periods of the lattice $P_{x}=\lambda_{x}/\sqrt{2}=
516$ nm and $P_{y}=\lambda_{y}/\sqrt{2}=554$ nm. The existence of
these band gaps implies that SPP plane waves impinging on small
devices build up with such lattice will generate stationary waves
in the crystal. More precisely this implies that SPPs will be
reflected in specific and different directions if their
wavelengths match the values $\lambda_x$ or $\lambda_y$ and if the
angle of incidence relatively to the normal to the Bragg planes of
the crystal (Figs.~12 B, C) equals $45^{\circ}$. Such devices act
consequently as an efficient in-plane wavelength demultiplexer for
SPPs \cite{Drezet:Nanolett2007} as it was indeed observed
experimentally (see Fig.~12D and E). While the results discussed
here were achieved within the visible spectral range, plasmonic
crystal devices are expected to perform even better (e.g., in
terms of spectral selectivity) in the infrared (telecom) spectral
range due to significantly lower ohmic losses
\cite{Nikolajsen:2003}. In general, the use of multiplexers,
splitters and tritters \cite{Drezet:Nanolett2007} in photonic
applications might be specifically appealing due to their small
footprint in the range of $10\times10$ $\mu$m$^2$. Furthermore,
the use as building blocks for classical \cite{Teich} or quantum
\cite{Knill:2001} optical
computing can be envisaged. \\
\section{Conclusion}
In this article we reviewed the field of leakage radiation
microscopy (LRM) theoretically and experimentally. Theoretically
we analyzed how SPP can generate leaky wave in the glass substrate
by tunnelling from the air side through a thin metal film
supporting SPP waves. We showed in particular that for thick film
($D\geq 70$ nm) leakage radiation (LR) does not affect the
dispersion relation on the air/metal interface. Importantly the
angular distribution of LR is located on a cone matching the SPP
dispersion relation. We also reviewed the first experimental
results reporting the observation of LR on rough surface and
using near field optics methods. We analyzed more recent
application of LRM to SPP nano-devices fabricated by electron
beam lithography. From all these results we can conclude that LRM
is a convenient and versatile far field optical method for
analyzing quantitatively SPP propagation on flat film and their
interaction with various nano-devices of direct practical
interest. Such versatility positions LRM as an appealing
alternative to near field optics for studying SPP propagation on a
scale of, or larger than the wavelength.
For financial support the European Union, under project FP6
2002-IST-1-507879 is acknowledged.
|
\section{Metric entropy with bracketing} \label{sec:entropy}
We first state several results about the entropy with bracketing, which will be used to prove Proposition~\ref{prop:entropy}. They are mainly adapted from \citep{GenoveseWasserman2000}, but several are improved or writen here in a more general form. These lemmas can be seen as a toolbox to calculate the metric entropy with bracketing of complex models from the metric entropy of simpler elements.
We consider a measurable space $(A, {\mathcal A})$, and $\mu$ a $\sigma$-finite positive measure on $A$. We consider a model $\mathcal{M}$, which is a set of probability density functions with respect to $\mu$. All functions considered in the following will be positive functions in $\ensuremath{\mathbb{L}}^1(\mu)$.
\begin{lem} \label{lem:maj-u-in-bracket}
Let $\varepsilon>0$. Let $[l,u]$ be a bracket in $\ensuremath{\mathbb{L}}^1(\mu)$, with $\h$-diameter less than $\varepsilon$, and containing $s$, a probability density function with respect to $\mu$. Then
\[ \int l\, d\mu \leq 1 \leq \int u\, d\mu \leq (1+\varepsilon)^2. \]
\end{lem}
\begin{proof}
First two inequalities are immediate, from $l\leq s \leq u$. For the last one, we use triangle inequality in $\ensuremath{\mathbb{L}}^2(\mu)$, and the definition of $\h$:
\begin{align*}
\int u\, d\mu &= \int \left(\sqrt{l} + \left(\sqrt{u} - \sqrt{l}\right) \right)^2 d\mu\\
&\leq \left(\sqrt{\int l\, d\mu} + \h(u,l)\right)^2 \\
&\leq (1+\varepsilon)^2.
\end{align*}
\end{proof}
\begin{lem}[Bracketing entropy of product densities] \label{lem:entropy-prod}
Let $n\geq 2$, and consider a collection $(A_i, {\mathcal A}_i, \mu_i)_{1\leq i \leq n}$ of measured space. For any $1\leq i \leq n$, let $\mathcal{M}_i$ be a collection of probability density functions on $A_i$ fulfilling \ref{cond:M}. Consider the product model
\[ \mathcal{M} = \left\{ s=\otimes_{i=1}^{n} s_i; \forall 1\leq i \leq n, s_i \in \mathcal{M}_i \right\}. \]
$\mathcal{M}$ contains density functions on $A=\prod_{i=1}^{n} A_i$ with respect to $\mu=\otimes_{i=1}^{n} \mu_i$.
$\mathcal{M}$ fulfills \ref{cond:M} and, for any sequence of positive numbers $(\delta_i)_{1\leq i\leq n}$, if $\varepsilon \geq \prod_{i=1}^{n} (1+\delta_i) -1$ then
\[ H_{[\cdot]}\left(\varepsilon, \mathcal{M}, \h\right) \leq \sum_{i=1}^{n} H_{[\cdot]}\left(\delta_i, \mathcal{M}_i, \h\right). \]
\end{lem}
\begin{proof}
Let us consider some $s=\otimes_{i=1}^{n} s_i$ in $\mathcal{M}$. For $1\leq i \leq n$, let $\mathcal{M}_i'$, $A_i'$ and a sequence $(t_{i,k})_{k\geq 1}$ be such as needed for $\mathcal{M}_i$ to verify \ref{cond:M}.
Then, with the choice $t_k = \otimes_{i=1}^{n} t_{i,k}$ and $A' = \prod_{i=1}^{n} A_i'$, \ref{cond:M} is true for $\mathcal{M}$ too.
Let $\delta > 0$. For any $1\leq i \leq n$, let $[l_i, u_i]$ a bracket containing $s_i$, with $\h$-diameter less than $\delta_i$. Let us set $l = \otimes_{i=1}^{n} l_i$, and $u = \otimes_{i=1}^{n} u_i$. Then $s$ belongs to bracket $[l, u]$. We can compute its $\h$-diameter:
\begin{align*}
\h(l,u) &= \sqrt{\int_{A} \bigg( \sum_{j=1}^{n} \bigg(\prod_{i=1}^{j-1} \sqrt{l_i} \prod_{i=j}^{n} \sqrt{u_i} - \prod_{i=1}^{j} \sqrt{l_i} \prod_{i=j+1}^{n} \sqrt{u_i} \bigg) \bigg)^2\,d\mu} \\
&\leq \sum_{j=1}^{n} \prod_{i=1}^{j-1} \sqrt{\int_{A_i} l_i\,d\mu_i} \prod_{i=j+1}^{n} \sqrt{\int_{A_i} u_i\,d\mu_i} \ \h(l_j, u_j) \\
&\leq \sum_{j=1}^{n} \delta_j \prod_{i=j+1}^{n} (1+\delta_i)
= \prod_{j=1}^{n} (1+\delta_j) -1
\end{align*}
thanks to triangle inequality and Lemma~\ref{lem:maj-u-in-bracket} (empty products equal $1$).
Let $\varepsilon \geq \prod_{i=1}^{n} (1+\delta_i) -1$. For any $1\leq i \leq n$ consider a minimal covering of $\mathcal{M}_i$ with brackets of $\h$-diameter less than $\delta_i$. With the previous process we can build a covering of $\mathcal{M}$ with brackets of $\h$-diameter less than $\varepsilon$. So the minimal cardinality of such a covering verifies
\[ N_{[\cdot]}\left(\varepsilon, \mathcal{M}, \h\right) \leq \prod_{i=1}^{n} N_{[\cdot]}\left(\delta_i, \mathcal{M}_i, \h\right). \]
\end{proof}
\begin{lem}[Bracketing entropy of mixture densities] \label{lem:entropy-mixture}
Let $n\geq2$, and for any $1\leq i \leq n$, let $\mathcal{M}_i$ be a set of probability density functions, all on the same measured space $(A, {\mathcal A}, \mu)$ and fulfilling \ref{cond:M}. Let us consider the set of all mixture densities
\[ \mathcal{M} = \left\{ \sum_{i=1}^{n} \pi_i s_i: \mathbf{\pi}=(\pi_i)_{1\leq i \leq n} \in \ensuremath{\mathbb{S}}_{n-1}; \forall 1\leq i \leq n, s_i \in \mathcal{M}_i \right\}. \]
Then $\mathcal{M}$ fulfills \ref{cond:M}, and for any $\delta>0$, $\eta>0$, and $\varepsilon \geq \delta + \eta + \delta \eta$,
\[ H_{[\cdot]}\left(\varepsilon, \mathcal{M}, \h\right) \leq H_{[\cdot]}\left(\delta, \ensuremath{\mathbb{S}}_{n-1}, \h\right) + \sum_{i=1}^{n} H_{[\cdot]}\left(\eta, \mathcal{M}_i, \h\right). \]
\end{lem}
\begin{proof}
First, let us note that $\ensuremath{\mathbb{S}}_{n-1}$ is separable for its usual topology. Then, checking that $\mathcal{M}$ fulfills \ref{cond:M} is easy, and we do not explicit it.
We do not develop either the proof of the last relation, because it is exactly the same as in \citep[proof of Theorem~2]{GenoveseWasserman2000}. Let us just say that at the end we get, using our Lemma~\ref{lem:maj-u-in-bracket} instead of \citep[Lemma~3]{GenoveseWasserman2000},
\begin{align*}
\h^2(l,u) &\leq \eta^2 \, (1+\delta)^2 + \delta^2 + 2 \eta \, \delta\, (1+\delta) \\
&\leq \varepsilon^2.
\end{align*}
\end{proof}
Next result is just Lemma~2 from \citep{GenoveseWasserman2000}:
\begin{lem}[Bracketing entropy of the simplex] \label{lem:entropy-simplex}
Let $n\geq 2$ be an integer. Let $\mu$ be the counting measure on $\{1, \ldots, n\}$. We identify any probability on $\{1, \ldots, n\}$ with its density $s\in\ensuremath{\mathbb{S}}_{n-1}$ with respect to $\mu$. Then, if $0 < \delta \leq 1$,
\[ H_{[\cdot]}\left(\delta, \ensuremath{\mathbb{S}}_{n-1}, \h\right) \leq (n-1) \ln\left(\frac{1}{\delta}\right) + \frac{\ln 2 + \ln(n+1) + n \ln(2\pi e)}{2}. \]
\end{lem}
To deal with Case~\ref{config:double-multinomial}, we also need the metric entropy of the collection of all Hardy-Weinberg genotype distributions for a given variable.
\begin{lem}[Bracketing entropy of Hardy-Weinberg genotype distributions] \label{lem:entropy-genotypes}
Suppose that, for some variable $l$, there exist $A_l\geq2$ different states. Let $\Omega_l$ be the collection of all genotype distributions following Hardy-Weinberg model (\ref{HW_modele}). Then $\Omega_l$ fulfills \ref{cond:M}, and for any $\delta>0$ and $\varepsilon \geq \sqrt{2}\, \delta\, (2+\delta)$,
\[ H_{[\cdot]}\left(\varepsilon, \Omega_l , \h\right) \leq H_{[\cdot]}\left(\delta, \ensuremath{\mathbb{S}}_{A_l-1}, \h\right). \]
\end{lem}
\begin{proof}
(\ref{HW_modele}) permits to associate a parameter $\alpha=(\alpha_1, \ldots, \alpha_{A_l}) \in \ensuremath{\mathbb{S}}_{A_l-1}$ to any density in $\Omega_l$. More generally, for any $\alpha \in [0,1]^{A_l}$, we define a function
\[ d_\alpha(x) = \left(2-\mathbbm{1}_{x_{1}=x_{2}}\right) \alpha_{x_{1}} \alpha_{x_{2}} \]
on the set of all genotypes $x=\{x^1, x^2\}$ on $A_l$ states. Consider some $\delta>0$ and $d_\alpha\in \Omega_l$. Let $[l, u]$ be some bracket containing $\alpha$, with $\h$-diameter less than $\delta$. Then $d_\alpha$ belongs to the bracket $[d_l, d_u]$. Let us calculate its diameter.
\begin{align*}
\h^2(d_l, d_u) &= \sum_{a=1}^{A_l} \left(u_a - l_a\right)^2 + \sum_{1\leq a<b\leq A_l} \left(\sqrt{2u_a u_b} - \sqrt{2l_a l_b} \right)^2 \\
&\leq 2 \sum_{a=1}^{A_l} \sum_{b=1}^{A_l} \left(\sqrt{u_a u_b} - \sqrt{u_a l_b} + \sqrt{u_a l_b} - \sqrt{l_a l_b} \right)^2 \\
&\leq 2 \left( \sqrt{\sum_{a=1}^{A_l} u_a \sum_{b=1}^{A_l} \left(\sqrt{u_b} - \sqrt{l_b}\right)^2} + \sqrt{\sum_{a=1}^{A_l} \left(\sqrt{u_a} - \sqrt{l_a}\right)^2 \sum_{b=1}^{A_l} l_b} \right)^2 \\
&\leq 2 \left( (1+\delta)\, \delta+ \delta \right)^2
\end{align*}
using Lemma~\ref{lem:maj-u-in-bracket}. So $\h(d_l, d_u) \leq \sqrt{2}\, \delta\, (2+\delta)$.
Let $(\alpha^{(k)})_{k\geq 1}$ a sequence of elements of $\ensuremath{\mathbb{S}}_{A_l-1} \cap \ensuremath{\mathbb{Q}}^{A_l}$, which tends to $\alpha$ for the usual topology as $k$ tends to infinity. Then, for any genotype $x=\{x^1, x^2\}$, $\ln d_{\alpha^{(k)}}(x)$ tends to $\ln d_{\alpha}(x)$. Therefore $\Omega_l$ fulfills \ref{cond:M}.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:entropy}]
We build the proof for Case~\ref{config:double-multinomial}. For Case~\ref{config:simple-multinomial} everything is similar, with a simplification: we directly have $\ensuremath{\mathbb{S}}_{A_l-1}$ instead of $\Omega_l$.
Using (\ref{Melange_selection}) we see that a probability $P_{(K,S)}\big(\cdot|\theta\big)$ is the product of a mixture density corresponding to the variables in $S$, and a product density in $\bigotimes_{l\notin S} \Omega_l$ for the other variables. Let us call $\mathcal{M}$ the collection of all mixtures of $K$ densities in $\bigotimes_{l\in S} \Omega_l$.
We first deal with the non clustering variables. Using Lemma~\ref{lem:entropy-prod} and Lemma~\ref{lem:entropy-genotypes}, $\bigotimes_{l\notin S} \Omega_l$ fulfills \ref{cond:M}. For any $\varepsilon \in (0,1)$,
\begin{align*}
H_{[\cdot]}\left((1+2\sqrt{2}\varepsilon+\sqrt{2}\varepsilon^2)^{L-|S|}-1, \bigotimes_{l\notin S} \Omega_l, \h\right) &\leq \sum_{l\notin S} H_{[\cdot]}\left(2\sqrt{2}\varepsilon+\sqrt{2}\varepsilon^2, \Omega_l, \h\right) \\
&\leq \sum_{l\notin S} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{A_l-1}, \h\right).
\end{align*}
On the same way
\[ H_{[\cdot]}\left((1+2\sqrt{2}\varepsilon+\sqrt{2}\varepsilon^2)^{|S|}-1, \bigotimes_{l\in S} \Omega_l, \h\right) \leq \sum_{l\in S} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{A_l-1}, \h\right). \]
We can apply Lemma~\ref{lem:entropy-mixture}, and get that $\mathcal{M}$ fulfills \ref{cond:M} and
\begin{multline*}
H_{[\cdot]}\left((1+2\sqrt{2}\varepsilon+\sqrt{2}\varepsilon^2)^{|S|}(1+\varepsilon)-1, \mathcal{M}, \h\right) \\
\leq \mathbbm{1}_{K\geq 2} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{K-1}, \h\right) + K \sum_{l\in S} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{A_l-1}, \h\right).
\end{multline*}
Lemma~\ref{lem:entropy-prod} again, applied to $\mathcal{M}$ and $\bigotimes_{l\notin S} \Omega_l$, gives that $\mathcal{M}_{(K, S)}$ fulfills \ref{cond:M}, and for any $\varepsilon \in (0,1)$,
\begin{multline*}
H_{[\cdot]}\left(\eta(\varepsilon), \mathcal{M}_{(K, S)}, \h\right) \\
\leq \mathbbm{1}_{K\geq 2} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{K-1}, \h\right) + K \sum_{l\in S} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{A_l-1}, \h\right) + \sum_{l\notin S} H_{[\cdot]}\left(\varepsilon, \ensuremath{\mathbb{S}}_{A_l-1}, \h\right).
\end{multline*}
At this point, it only remains to use Lemma~\ref{lem:entropy-simplex} and to compute the constants.
\end{proof}
\section{Establishing the penalty} \label{sec:phi}
First, we need to establish some properties of function $\eta$.
\begin{lem}[Properties of function $\eta$] \label{lem:prop-eta}
We consider the function $\eta$ defined in Proposition~\ref{prop:entropy}, from $\ensuremath{\mathbb{R}}_+$ into $\ensuremath{\mathbb{R}}_+$. $\eta$ is nonnegative, increasing and convex. $\eta(0)=0$, and $\eta'(0)=L+1$ in Case~\ref{config:simple-multinomial} while $\eta'(0)=2\sqrt{2}\,L+1$ in Case~\ref{config:double-multinomial}.
\end{lem}
\begin{proof}
The proof in Case~\ref{config:simple-multinomial} is immediate, so we develop only Case~\ref{config:double-multinomial}.
Setting $u(x) = 1 + 2 \sqrt{2}\, x + \sqrt{2}\, x^2$, we can write $\eta(x) = (1+x)\, u(x)^L -1$. Then, calculus gives
\[ \eta'(x) = (2L +1)\, u(x)^L + 2 L\, (\sqrt{2} -1)\, u(x)^{L-1}. \]
Since $u$ is positive on $(0, +\infty)$, $\eta$ is increasing. But $\eta(0) = 0$, so $\eta$ is nonnegative on $\ensuremath{\mathbb{R}}_+$. We also have $\eta'(0) = 2 \sqrt{2}\, L +1$. Next,
\[ \eta''(x) = 2\sqrt{2}\, (1+x)\, \left( (2L^2 +L)\, u(x)^{L-1} + 2 L\, (L-1)\, (\sqrt{2} -1)\, u(x)^{L-2} \right) \]
which is positive on $\ensuremath{\mathbb{R}}_+$.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:phi-m}]
Let $0<\sigma\leq\eta(1)$, and $\delta = \eta^{-1}(\sigma)$. Then, for any $u \in \mathcal{M}_{m}$,
\begin{align*}
&\int_{0}^{\sigma} \sqrt{H_{[\cdot]}\left(x, \mathcal{M}_{m}(u,\sigma), \h\right)} d x \\
&\qquad\leq \sum_{j=1}^{\infty} \int_{\eta(2^{-j}\delta)}^{\eta(2^{-j+1}\delta)} \sqrt{H_{[\cdot]}\left(x, \mathcal{M}_{m}, \h\right)} d x \\
&\qquad\leq \sum_{j=1}^{\infty} \left(\eta\left(2^{-j+1}\delta\right) - \eta\left(2^{-j}\delta\right)\right) \sqrt{C_{m} - D_{m} \ln \delta + D_{m} j\ln 2} \\
&\qquad\leq \eta(\delta) \sqrt{C_{m} - D_{m} \ln \delta} \\
&\qquad\quad + \sqrt{D_{m}\ln 2}\, \sum_{j=1}^{\infty} \sqrt{j}\, \left(\eta\left(2^{-j+1}\delta\right) - \eta\left(2^{-j}\delta\right)\right).
\end{align*}
We deal with the last term of this sum in the following way:
\begin{align*}
\sum_{j=1}^{\infty} \sqrt{j}\, \left(\eta\left(2^{-j+1}\delta\right) - \eta\left(2^{-j}\delta\right)\right)
&\leq \sum_{j=1}^{\infty} j\, \left(\eta\left(2^{-j+1}\delta\right) - \eta\left(2^{-j}\delta\right)\right) \\
&= \sum_{k=1}^{\infty} \eta\left(2^{-k+1}\delta\right) \\
&\leq \sum_{k=1}^{\infty} 2^{-k+1} \eta(\delta) = 2 \sigma.
\end{align*}
So
\[ \int_{0}^{\sigma} \sqrt{H_{[\cdot]}\left(x, \mathcal{M}_{m}(u,\sigma), \h\right)} d x \leq \phi_m(\sigma). \]
Since $\eta$ is increasing, $\phi_m(x)/x$ is decreasing. To check that $\phi_m$ is nondecreasing, it is enough to prove that function $f(x)= x \sqrt{b-\ln \eta^{-1}(x)}$ is nondecreasing on $(0, \eta(1)]$, where $b = \frac{C_{m}}{D_{m}}$. From (\ref{eq:def-RKS}), we get $C_{m} > \frac{\ln(2\pi e)}{2} D_{m} > D_{m}$, so $b>1$. Calculus gives
\[ f'(x) = \sqrt{b-\ln \eta^{-1}(x)} - \frac{x}{2 \eta^{-1}(x)\, \eta'\left(\eta^{-1}(x)\right) \sqrt{b-\ln \eta^{-1}(x)}}. \]
Let $y\in (0, 1]$. $\eta$ is convex on $(0, 1]$, and that entails $\frac{\eta(y)}{y\, \eta'(y)} \leq 1$. Thus
\[ \sqrt{b-\ln y}\, f'\left(\eta(y)\right) \geq b - \ln y -1/2 > 0. \]
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem:sigma-less-than-eta}]
Since $\phi_m(x)/x$ is non-increasing,
for any $\sigma>0$ such that $\sqrt{n}\, \sigma^2 > \phi_m(\sigma)$, $\sigma> \sigma_m$. So, we look for situations such that $\sqrt{n} > \dfrac{\phi_m(\eta(1))}{\eta^2(1)}$.
For all $1\leq l \leq L$, $A_l\geq 2$. Since $\frac{1}{2} \ln (1+x) \leq x-1$ for $x\geq 2$, we get the following bounds
\begin{equation} \label{eq:R-K-S}
\frac{1+\ln (2\pi)}{2} D_{m} \leq C_{m} \leq \left( 2+\ln (2\pi) + \frac{\ln 2}{2}\right) D_{m}.
\end{equation}
Therefore
\begin{align*}
\frac{\phi_m(\eta(1))}{\eta^2(1)} &< \frac{4 \sqrt{D_{m}}}{\eta(1)}
\end{align*}
On the other hand, we have
\begin{equation*}
D_{m} \leq K\, L\, A_{\max}.
\end{equation*}
So, since $\phi_m(x)/x^2$ is decreasing, $\sigma_m < \eta(1)$ as soon as $n > \xi^2 K$. This is true when $\xi < 1$, since $K\leq n$: the number of clusters is not bigger than the number of individuals.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem:sum-x-m}]
We define $\delta = 1/2$, from which $e^{-x_m} = \delta^{D_{m}}$. If we consider the collection $\ensuremath{\mathbb{M}}$, we can discern two cases: $K=1$ and $S=\emptyset$, or $K\geq 2$ and $S\neq \emptyset$. So, using (\ref{eq:Dimension}),
\begin{align*}
\sum_{m\in\ensuremath{\mathbb{M}}} e^{-x_m} &= \delta^{\sum_{l=1}^{L} \left(A_l-1\right)} \Bigg(1+ \sum_{S\neq\emptyset} \sum_{K\geq 2} \left(\delta^{1+\sum_{l\in S}\left(A_l-1\right)}\right)^{K-1} \Bigg) \\
&= \delta^{\sum_{l=1}^{L} \left(A_l-1\right)} \Bigg(1+ \sum_{S\neq\emptyset} \frac{\delta^{1+\sum_{l\in S}\left(A_l-1\right)}}{1-\delta^{1+\sum_{l\in S}\left(A_l-1\right)}} \Bigg) \\
&\leq \delta^L \Bigg(1+\frac{\delta}{1-\delta} \sum_{S\neq\emptyset} \delta^{|S|}\Bigg) \\
&= \delta^L (1+\delta)^L.
\end{align*}
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem:pen}]
$\eta^{-1}$ in concave and nondecreasing, $\eta(0)=0$, so for any $0\leq x \leq \eta(1)$,
\[ \eta^{-1}(x) \geq \frac{\eta^{-1}(2)}{2}\, \min (x , 2). \]
On the other hand (\ref{eq:sigma-tilde}) and (\ref{eq:R-K-S}) entail
\begin{equation} \label{eq:bounds-sigma-tilde}
\widetilde{\sigma}_{m} \geq C_1 \sqrt{\frac{D_{m}}{n}} \geq C_1 \sqrt{\frac{L}{n}}
\end{equation}
where $C_1 = 2\sqrt{\ln 2} + \sqrt{\frac{1+\ln (2\pi)}{2}} >2\sqrt{2}$. Therefore
\[ -\ln \eta^{-1}\left(\widetilde{\sigma}_{m}\right) \leq -\ln \left(\frac{\eta^{-1}(2)}{2}\right) -\ln 2 + \max\left(0, \frac{1}{2}\left(\ln n - \ln L - \ln 2\right) \right). \]
Consider Case~\ref{config:simple-multinomial}.
Since $\eta$ is a convex function and $\eta'(0)=L+1$,
\[ \eta^{-1}(2) \leq \frac{2}{L+1}. \]
Now,
\[ \eta\left(\frac{2}{L+1}\right) = \left(1+ \frac{2}{L+1}\right)^{L+1} -1 \leq e^2-1. \]
Then
\[ \frac{\eta^{-1}(2)}{2} \geq \frac{2/(L+1)}{\eta\left(2/(L+1)\right)} \geq \frac{2}{(e^2-1)(L+1)}. \]
Therefore
\[ -\ln \left(\frac{\eta^{-1}(2)}{2}\right) \leq \ln (e^2-1) -\ln 2 + \ln L + \ln (3/2) \]
and
\[ -\ln \eta^{-1}\left(\widetilde{\sigma}_{m}\right) \leq \ln (e^2-1) -\frac{7}{2} \ln 2 + \ln 3 + \max\left(\frac{1}{2} \ln n + \frac{1}{2} \ln L, \frac{\ln 2}{2} + \ln L\right). \]
Using now (\ref{eq:sigma-upper}), we get
\begin{align*}
\sigma_m^2+\frac{x_m}{n} &\leq \frac{D_{m}}{n}\, \left(\frac{1}{2} + \left(2\sqrt{\ln 2} + \sqrt{2+\ln (2\pi) + \frac{\ln 2}{2}-\ln \eta^{-1}\left(\widetilde{\sigma}_{m}\right)} \right)^2\right) \\
&\leq \frac{D_{m}}{n}\, \Bigg(\frac{1}{\sqrt{2}} + 2\sqrt{\ln 2} + \sqrt{2+\ln (2\pi) - 3 \ln 2 + \ln 3 + \ln (e^2-1)} \\
&\qquad + \sqrt{\max \left( \frac{\ln n + \ln L}{2}, \frac{\ln 2}{2} + \ln L\right)} \Bigg)^2 \\
&\leq \frac{D_{m}}{n}\, \left(5 + \sqrt{\max \left( \frac{\ln n + \ln L}{2}, \frac{\ln 2}{2} + \ln L\right)} \right)^2.
\end{align*}
Next, consider Case~\ref{config:double-multinomial}, and follow the same method. Then
\[\eta^{-1}(2) \leq \frac{1}{\sqrt{2}\,L}\]
and
\[\eta\left(\frac{1}{\sqrt{2}\,L}\right) \leq 2 \exp \left(2+\frac{1}{\sqrt{2}}\right).\]
This leads to
\[ -\ln \eta^{-1}(x) \leq 2+\frac{1}{\sqrt{2}} +\frac{3 \ln 2}{2} + \ln L - \ln \min (x , 2) \]
and
\[ -\ln \eta^{-1}\left(\widetilde{\sigma}_{m}\right) \leq 2+\frac{1}{\sqrt{2}} + \max\left(\frac{1}{2} \ln n + \frac{1}{2} \ln L, \frac{\ln 2}{2} + \ln L\right). \]
Now we obtain
\begin{align*}
\sigma_m^2+\frac{x_m}{n} &\leq \frac{D_{m}}{n}\, \Bigg(\frac{1}{\sqrt{2}} + 2\sqrt{\ln 2} + \sqrt{4+\ln (2\pi)+\frac{\sqrt{2}+\ln 2}{2}} \\
&\qquad + \sqrt{\max \left( \frac{\ln n + \ln L}{2}, \frac{\ln 2}{2} + \ln L\right)} \Bigg)^2 \\
&\leq \frac{D_{m}}{n}\, \left(5 + \sqrt{\max \left( \frac{\ln n + \ln L}{2}, \frac{\ln 2}{2} + \ln L\right)} \right)^2.
\end{align*}
\end{proof}
\section{Introduction} \label{Sect:Introduction}
This article is concerned with the unsupervised classification on categorical multivariate data. The model-based clustering, which uses finite mixture models, is an intuitive and rigorous framework for the unsupervised classification.
However there is no clear consensus on the way to gather individuals in general: on the basis of well separated clusters, or on the basis of the components of the mixture distribution? We refer to \cite{BaudryPhD} for a general discussion on this topic. Finite mixture models are specially adapted when each class is supposed to be characterized by a set of parameters, for instance in population genetics: in this case the populations that the biologists look for are characterized by their allelic frequencies and a genetic equilibrium; this corresponds to the notion of population as a reproduction unit, or a group of individuals sharing the same genetic structure.
Finite mixture models are also known in the literature as the latent class models.
The observations are $n$ independent realizations of a random vector, whose number $L$ of coordinates (variables) may be large.
The individuals of the sample are clustered into a certain unknown number $K$ of populations on the basis of the frequencies of apparition of the possible states of each variable. It may happen that only a subset $S$ of the variables are relevant for clustering purposes, and the others are just noise. Thus, in addition to the number $K$ of populations and the frequencies of the different states, we are also interested in the subset $S$, which may have significance in the interpretation of the results.
A number of clustering methods for categorical multivariate data have been proposed in recent years in the context of genomics (see \citep{Pritchard2000,Chen2006,Corander2008}). But the problem of variable selection for clustering using such data was first addressed in \citep{ToussileGassiat2009}, where the question is regarded as a model selection problem in a density estimation framework. First the components of a finite mixture distribution are identified, then the individuals are clustered into these components using the Maximum A Posteriori (MAP) method. \\
Using simulated data, that article shows that the variable selection procedure based on the Bayesian Information Criterion ($\BIC$) significantly improves clustering and prediction capacities in our framework. It also gives a theoretical consistency result: when the true density $P_0$ underlying the observations belongs to one of the competing models, then there exists a smallest model $\mathcal{M}_{\left(K_0,\ S_0\right)}$ containing $P_0$; further, the $\BIC$ type criteria select $\mathcal{M}_{\left(K_0,\ S_0\right)}$ with probability tending to one as the sample size $n$ goes to infinity. This consistency approach requires large sample sizes which may be difficult to obtain. However the knowledge of the true model, aside the frequencies of the states, is an important information for the interpretation of the results.
In the present paper we adopt an oracle approach. We do not aim at choosing the true model underlying the data, even if our procedure performs well also for that. The criteria are rather designed to minimize some risk function of the estimated density with respect to the true density. In this context simpler models can be preferred to $\mathcal{M}_{\left(K_0,\ S_0\right)}$, in which too many parameters can entail estimators which overfit the data. Actually there is no need to assume that $P_0$ belongs to one of the competing models $\mathcal{M}_{(K,S)}$.
$\BIC$ relies on a strong asymptotic assumption, and can thus require large sample sizes to reach its asymptotic behavior; practically $\BIC$ is known to overpenalize, and therefore selects too small models for small or medium values of $n$ (see \citep{Nadif1998}). On the contrary Akaike's Information Criterion ($\AIC$) is known to underpenalize, and selects too large models for large and medium values of $n$. We would like a criterion which gathers the virtues of both $\AIC$ and $\BIC$, and performs well for different values of $n$.
In this article, we propose a non asymptotic penalized criterion based on the metric entropy theory of Massart (in particular \citep{Massart2007}). It leads to a non asymptotic oracle inequality, which compares the risk of the selected estimator to the risk of the estimator associated with the unknown best model (see Theorem~\ref{thm:main-theorem} below). There exists a large literature on model selection via penalization from a non asymptotic perspective. This literature is still in development with the appearance of sophisticated tools of probability such as concentration and deviations inequalities (see \citep{Massart2007} and the references therein). In mixture models the non asymptotic approach is very recent, the first related work being \citep{MaugisMichel2008thm} for the Gaussian mixture model.
However, the obtained penalty function presents drawbacks: it depends on a multiplicative constant for which sharp upper bounds are not available, and it leads in practice to an overpenalization --- even worse than $\BIC$. Therefore our theoretical result mainly suggests the shape of the penalty function:
\begin{equation*}
\pen_n\left(m\right)=\lambda D_m/n,
\end{equation*}
where $D_m$ is the dimension of model $m$, and $\lambda$ an unknown parameter depending on the sample size and the complexity of the collection of models under competition, which has to be calibrated. A calibration of $\lambda$ with the so-called slope heuristics has been proposed in \citep{BirgeMassart2007} in such a case. We propose a modified version based on a sliding window of this calibration method. The resulting criterion does not require an ad-hoc choice of the penalty parameters and adapts automatically to the data.
Although the full theoretical validation of slope heuristics is provided only in the Gaussian homoscedastic and heteroscedastic regression frameworks \citep{BirgeMassart2007, ArlotMassart2009}, they have been implemented in several other frameworks (see \citep{MaugisMichel2008slope,LebarbierPhDThesis,Verzelen2009,Villers2007} for applications in density estimation, genomics, etc.). The simulations performed in Subsection~\ref{sect:Simulated_experiments} illustrate that our criterion behaves well with respect to more classical criteria as $\BIC$ and $\AIC$, both to estimate the density, even when $n$ is relatively small, and to retrieve the true model. It can be seen as a representative of the family of the General Information Criteria (see for instance \citep{BaiRaoWu1999} whose criterion is less intuitive but presents some analogy with the slope heuristics).
The paper is organized as follows. Section~\ref{Sect:Model_methods} is devoted to the presentation of the mixture models framework and to the model selection paradigm. In Section~\ref{Sect:Main_result} we state and prove our main result, the oracle inequality. Section~\ref{Sect:In_practice} is devoted to the practical aspect of our procedure which has been implemented in the stand alone software \texttt{MixMoGenD} (Mixture Model using Genotypic Data) (see \citep{ToussileGassiat2009}). Results on simulated experiments are also presented: we compare our proposed criterion to classical $\BIC$ and $\AIC$, in both points of view of the selection of the true model and of density estimation. Eventually, the Appendices contain several technical results used in the main analysis.
\section{Model and methods} \label{Sect:Model_methods}
\subsection{Framework} \label{Subsect:Framework}
We suppose we deal with independent and identically distributed (iid) realizations of a multivariate random vector $X=(X^l)_{1\leq l\leq L}$. We consider two
main settings:
\begin{enumerate}
\item \label{config:simple-multinomial} Each $X^l$ is a multinomial variable taking values in $\left\{1,\ldots,\ A_l\right\}$.
\item \label{config:double-multinomial} Each $X^l$ consists in a (non ordered) set $\left\{X^{l,1},\ X^{l,2}\right\}$ of two (that may be equal) qualitative variables taking their values in the same set $\left\{1,\ldots,\ A_l\right\}$.
\end{enumerate}
All along this article, these two settings will be referred to as Case~\ref{config:simple-multinomial} and Case~\ref{config:double-multinomial}. In both cases, the numbers $A_l$ of allowed states are supposed to be known, and to verify $A_l\geq 2$.
The first case is a usual latent class model with various applications (psychomatrics, marketing, credit scoring, genomics, etc.), while the last one is more specific to genotypic data. In this context $X=(X^l)_{1\leq l\leq L}$ represents the genotype of an individual at $L$ loci of its DNA. Case~\ref{config:simple-multinomial} corresponds to haploid organisms, with a single representative of each chromosome; at any locus $l$ a single allele $X^l$ is measured. Case~\ref{config:double-multinomial} corresponds to diploid organisms, with two representatives of each chromosome; at any locus $l$, two alleles $X^{l,1}$ and $X^{l,2}$ are observed together.
We consider a model-based clustering, which means that the sample is a finite mixture of an unknown number $K$ of populations (clusters), each being characterized by a set of frequencies of the states. Let denote by $Z$ the (unobserved) population an individual comes from. Variable $Z$ takes its values in the set $\left\{1,\ldots,\ K\right\}$ of the labels of the different clusters. Its distribution is given by the vector $\mathbf{\pi} = (\pi_k)_{1\leq k \leq K}$, where $\pi_k=P(Z=k)$. Conditionally to $Z$, the variables $X^1,\ldots,\ X^L$ are supposed to be independent. In Case~\ref{config:double-multinomial}, the states $X^{l,1}$ and $X^{l,2}$ for the $l^{\text{th}}$ variable are also supposed to be independent conditionally to $Z$. The preceding two assumptions are what biologists respectively call \emph{Linkage Equilibrium} (LE) and \emph{Hardy-Weinberg Equilibrium} (HWE). According to these assumptions, the probability distribution of a genotype $x=\left(x^l\right)_{1\leq l\leq L}$ in a population $k$ is given in the following equations
\begin{align}
P\left(x|\ Z=k\right) &= \prod_{l=1}^L P\left(x^l | Z=k\right) \notag \\
\text{Case~\ref{config:simple-multinomial}: } P\left(x^l | Z=k\right) &= \alpha_{k,l,x^{l}} \notag \\
\text{Case~\ref{config:double-multinomial}: } P\left(x^l | Z=k\right) &= \left(2-\mathbbm{1}_{x^{l,1}=x^{l,2}}\right) \alpha_{k,l,x^{l,1}} \alpha_{k,l,x^{l,2}} \label{HW_modele}
\end{align}
where $\alpha_{k,l,j}$
is the probability of state $j$ associated to variable $X^l$ in population $k$. The mixing proportions $\pi_k$ and the probabilities $\alpha_{k,l,j}$ will be treated as parameters.
In the context of genomics, Hardy-Weinberg and linkage equilibria are based on several simplifying assumptions that can seem unrealistic; however they have still proven to be useful in describing many population genetic attributes and serve as a base model in the development of more realistic models of microevolution. Further, the choice of estimators derived from the maximum likelihood estimator (MLE) responds to the wish of biologists to group the sample into clusters minimizing the Hardy-Weinberg and linkage desequilibria, and this brings some robustness to our modeling (see \citep{Latch2006} and references therein).
Going deeper, the oracle approach emphasizes that we should often prefer simplified and misspecified models. This introduces a modeling bias in order to get more robust estimators and classifiers, and at the end we get a smaller estimation error. This legitimizes also the following simplification.
It may happen that the structure of interest is contained in only a subset $S$ of the $L$ available variables, the others been useless or even harmful to detect a reasonable clustering into statistically different populations. For the variables in $S$, the frequencies of the states in at least two populations are different: we will call them clustering variables. For the other variables, the states are supposed to be equally distributed across the clusters. This approximation is theoretically justified by the oracle heuristics, which is able to take advantage of the misspecification; the simulations performed in \citep{ToussileGassiat2009} illustrate its benefits. \\
We denote by $\beta_{l,j}$ the frequency of state $j$ associated to variable $X^l$ in the whole population:
\begin{equation*}
\beta_{l,j}=\alpha_{1,l,j}=\cdots=\alpha_{k,l,j}\cdots=\alpha_{K,l,j}\ \textrm{for any } l\notin S \textrm{ and } 1\leq j \leq A_l.
\end{equation*}
Obviously, $S=\emptyset$ if $K=1$, otherwise $S$ belongs to $\mathcal{P}^{*}(L)$, the set of all non empty subsets of $\{1, \ldots, L\}$.
Summarizing all these assumptions, we can write down the likelihood of an observation $x=\left(x^l\right)_{1\leq l\leq L}$:
\begin{align}
\begin{split}
\text{Case~\ref{config:simple-multinomial}: } P_{(K, S, \theta)}(x) &=\left[\sum_{k=1}^K \pi_k \prod_{l\in S} \alpha_{k,l,x^{l}}\right] \times\prod_{l\notin S} \beta_{l,x^{l}} \\
\text{Case~\ref{config:double-multinomial}: } P_{(K, S, \theta)}(x) &= \left[\sum_{k=1}^K \pi_k \prod_{l\in S}\left(2-\mathbbm{1}_{x^{l,1}=x^{l,2}}\right)\alpha_{k,l,x^{l,1}}\times\alpha_{k,l,x^{l,2}}\right]\\
&\quad \times\prod_{l\notin S} \left(2-\mathbbm{1}_{x^{l,1}=x^{l,2}}\right)\beta_{l,x^{l,1}}\beta_{l,x^{l,2}}
\end{split} \label{Melange_selection}
\end{align}
where $\theta=\left(\mathbf{\pi}, \mathbf{\alpha}, \mathbf{\beta}\right)$ is a multidimensional parameter, with
\begin{align*}
\mathbf{\alpha} &=\left(\alpha_{k,l,j}\right)_{1\leq k \leq K;\ l\in S;\ 1\leq j \leq A_l}\\
\mathbf{\beta} &=\left(\beta_{l,j}\right)_{l\notin S;\ 1\leq j \leq A_l}.
\end{align*}
For a given $K$ and $S$, $\theta=\theta_{(K, S)}$ ranges in the set
\begin{equation} \label{Theta-KS}
\Theta_{(K, S)} = \ensuremath{\mathbb{S}}_{K-1}\times\bigg[\prod_{l\in S} \ensuremath{\mathbb{S}}_{A_l-1}\bigg]^K \times \prod_{l\notin S}\mathbb{S}_{A_l-1},
\end{equation}
where $\ensuremath{\mathbb{S}}_{r-1}=\left\{p=\left(p_1,\ p_2,\ldots,\ p_r\right)\in\left[0,\ 1\right]^r : \sum_{j=1}^r p_j = 1 \right\}$ is the $(r-1)$-dimensional simplex.
Then we consider the collection of all parametric models
\begin{equation} \label{model_KS}
\mathcal{M}_{(K, S)}=\left\{P_{(K, S, \theta)}: \theta \in\Theta_{(K, S)} \right\}
\end{equation}
with $(K, S) \in \ensuremath{\mathbb{M}} := \{(1, \emptyset)\} \cup (\ensuremath{\mathbb{N}} \backslash \{0, 1\}) \times \mathcal{P}^{*}(L)$. To alleviate notations, we will often use the single index $m\in \ensuremath{\mathbb{M}}$ instead of $(K, S)$.
Each model $\mathcal{M}_{(K, S)}$ corresponds to a particular structure situation with $K$ clusters and a subset $S$ of clustering variables.
Inferring $K$ and $S$ becomes a model selection problem in a density estimation framework. It also leads to a data clustering, via the estimation $\widehat{\theta}$ of the parameter $\theta_{(K, S)}$ and the prediction of the class $z$ of an observation $x$ by the MAP method:
\[ \widehat{z} = \mathop{\mathrm{arg\,max}}_{1\leq k\leq K} P_{(K, S, \widehat{\theta})}\left( Z=k | X=x\right). \]
\subsection{Model selection via penalization} \label{Subsect:Model_selection}
A common method to solve model selection problems consists in the minimization of a penalized maximum likelihood criterion. In each model $\mathcal{M}_{(K,S)}$, consider the maximum likelihood estimator (MLE) $\widehat{P}_{(K,S)} = P_{(K, S, \widehat{\theta})}$, which minimizes the log-likelihood contrast
\begin{equation} \label{eq:log-likelihood-contrast}
\gamma_n\left(P\right)=- \dfrac{1}{n}\sum_{i=1}^n\ln P\left(X_i\right)
\end{equation}
where $X_i$ describes the individual $i$ in the sample. Then a data driven selected model $\mathcal{M}_{\left(\widehat{K}_n,\widehat{S}_n\right)}$ is chosen, where $\left(\widehat{K}_n,\widehat{S}_n\right)$ minimizes a penalized maximum likelihood criterion of the form
\begin{equation*}
\crit(K,S) = \gamma_n\big(\widehat{P}_{(K,S)}\big) + \pen_n(K,S),
\end{equation*}
where $\pen_n:\ \ensuremath{\mathbb{M}}\rightarrow\mathbb{R}_+$ is the penalty function. Eventually the selected estimator is $\widehat{P}_{(\widehat{K}_n,\ \widehat{S}_n)}$.
The penalty function is designed to avoid overfit problems. Classical penalties, such as the ones used in $\AIC$ and $\BIC$ criteria, are based on the dimension of the model. In the following, we will refer to the number of free parameters
\begin{equation} \label{eq:Dimension}
D_{(K, S)} = K-1 + K \sum_{l\in S}\left(A_l-1\right) + \sum_{l\notin S}\left(A_l-1\right)
\end{equation}
as the dimension of the model $\mathcal{M}_{(K, S)}$. The penalty functions of $\AIC$ and $\BIC$ are respectively defined by
\begin{equation} \label{eq:def-BIC-AIC}
\begin{split}
\pen_{\AIC}\left(m\right)&=\frac{1}{n}\cdot D_m;\\
\pen_{\BIC}\left(m\right)&=\frac{\ln n}{2n}\cdot D_m.
\end{split}
\end{equation}
Our work is centered on the MLE estimator $\widehat{P}_{\left(K,\ S\right)}$, but this last one presents a drawback. For the sake of density estimation, we would like to use the Kullback-Leibler divergence $\KL$ as a risk function to measure the quality of an estimator. Unfortunately, when an state is not present in the sample, the MLE estimator assigns to it a zero probability. As a consequence, the Kullback risk $\ensuremath{\mathbb{E}}_{P_0}\left[\KL\left(P_0,\ \widehat{P}_{\left(K,\ S\right)}\right)\right]$ is infinite.
The Hellinger distance offers an alternative to the Kullback-Leibler divergence. Let us consider two probability distribution $P$ and $Q$, admitting respectively $s$ and $t$ as density functions with respect to a common $\sigma$-finite measure $\mu$. We call Hellinger distance between $P$ and $Q$ the quantity $\h(P,Q)$ defined by
\begin{equation} \label{eq:def-hellinger}
\h(P,Q)^2 = \int \left(\sqrt{s(x)} - \sqrt{t(x)}\right)^2 d\mu(x).
\end{equation}
Let $\left(K^*,\ S^*\right)$ be a minimizer in $\left(K,\ S\right)$ of the Hellinger risk of the MLE estimator
\begin{equation} \label{eq:H2-risk}
R_{\left(K,\ S\right)}=\ensuremath{\mathbb{E}}_{P_0}\left[\h^2\left(P_0,\ \widehat{P}_{\left(K,\ S\right)}\right)\right].
\end{equation}
The density $\widehat{P}_{\left(K^*,\ S^*\right)}$ is called oracle for the Hellinger risk. It is not an estimator, since it depends on the true density $P_0$. However it can be used as a benchmark to quantify the quality of our model selection procedure: in the simulation performed in paragraph~\ref{sect:oracle_experiment}, we compare the Hellinger risk of the selected estimator $\widehat{P}_{(\widehat{K}_n,\ \widehat{S}_n)}$ to the oracle risk.
\section{New criteria and non asymptotic risk bounds} \label{Sect:Main_result}
\subsection{Main result} \label{Subsect:Main_result}
Our main theorem provides an oracle inequality for both Case~\ref{config:simple-multinomial} and Case~\ref{config:double-multinomial}. It links the Hellinger risk of the selected estimator to the Kullback-Leibler divergence $\KL$ between the true density and each model in the models collection. Unlike $\KL$ which is not a metric, the Hellinger distance $\h$ permits to take advantage of the metric properties (metric entropy) of the models.
\begin{thm} \label{thm:main-theorem}
We consider the collection $\ensuremath{\mathbb{M}}$ of models defined above, and a corresponding collection of $\rho$-MLEs $\big(\widehat{P}_{(K,S)}\big)_{(K,S) \in \ensuremath{\mathbb{M}}}$, which means that for every $(K, S) \in \ensuremath{\mathbb{M}}$
\[ \gamma_n\big(\widehat{P}_{(K,S)}\big) \leq \inf_{Q\in \mathcal{M}_{(K, S)}} \gamma_n(Q) + \rho. \]
Let $A_{\max}=\sup_{1\leq l\leq L}A_l$, and let $\xi$ be defined by $\xi = \dfrac{4 \sqrt{A_{\max}}\sqrt{L}}{2^{L+1}-1}$ in Case~\ref{config:simple-multinomial} and $\xi = \dfrac{4 \sqrt{A_{\max}}\sqrt{L}}{2 (1+3\sqrt{2})^{L}-1}$ in Case~\ref{config:double-multinomial}. Assume that $\xi< 1$ or $n>\xi^2K$.
There exists absolute constants $\kappa$ and $C$ such that whenever
\begin{equation} \label{eq:pen}
\pen_n(K,S) \geq \kappa\, \left(5 + \sqrt{\max\left(\frac{1}{2} \ln n + \frac{1}{2} \ln L, \frac{\ln 2}{2} + \ln L\right)} \right)^2 \frac{D_{(K, S)}}{n}
\end{equation}
for every $(K, S)\in {\ensuremath{\mathbb{M}}}$, then the model $\mathcal{M}_{\left(\widehat{K}_n,\widehat{S}_n\right)}$ where $\left(\widehat{K}_n,\widehat{S}_n\right)$ minimizes
\[\crit(K,S) = \gamma_n\big(\widehat{P}_{(K,S)}\big) + \pen_n(K,S)\]
over ${\ensuremath{\mathbb{M}}}$ exists and moreover, whatever the underlying probability $P_0$,
\begin{multline*}
\ensuremath{\mathbb{E}}_{P_0}\left[\h^2\left(P_0,\widehat{P}_{\left(\widehat{K}_n,\widehat{S}_n\right)}\right)\right] \\
\leq C\, \left( \inf_{(K, S)\in \ensuremath{\mathbb{M}}} \left(\KL\left(P_0,\mathcal{M}_{(K, S)}\right) + \pen_n(K, S)\right) + \rho + \frac{(3/4)^L}{n}\right)
\end{multline*}
where, for every $(K, S) \in \ensuremath{\mathbb{M}}$, $\KL\left(P_0,\mathcal{M}_{(K, S)}\right) = \inf_{Q\in \mathcal{M}_{(K, S)}} \KL(P_0,Q)$.
\end{thm}
The condition $\xi<1$ is used in the proof to avoid more complicated calculations. In practice, $\xi$ is very likely to be smaller than $1$ for $L$ not too small.
Note that as soon as $n\geq 2L$, (\ref{eq:pen}) is simplified in the following way
\[ \pen_n(K, S) \geq \kappa\, \left(5 + \sqrt{\frac{1}{2} \ln n + \frac{1}{2} \ln L} \right)^2 \frac{D_{(K, S)}}{n}. \]
The leading term for large $n$ is $\kappa \dfrac{\ln n}{2}\, \dfrac{D_{(K, S)}}{n}$, which is a multiple of the penalty function of $\BIC$. As a consequence, we can apply Theorem~2 from \citep{ToussileGassiat2009}: when the underlying distribution $P_0$ belongs to one of the competing models, the smallest model $(K_0, S_0)$ containing $P_0$ is selected with probability tending to $1$ as $n$ goes to infinity.
Such a penalty is not surprising in our context: it is in fact very similar to the one obtained in \citep{MaugisMichel2008thm} in a Gaussian mixture framework.
Sharp estimates of $\kappa$ are not available. Theorem~\ref{thm:main-theorem} is too conservative in practice, and leads to an over-penalized criterion which is outperformed by smaller penalties. So it is mainly used to suggest the shape of the penalty function
\begin{equation} \label{eq:shape.penalty}
\pen_n(K, S) = \lambda\, \frac{D_{(K, S)}}{n}
\end{equation}
where $\lambda$ is a parameter to be chosen depending on $n$ and the collection $\ensuremath{\mathbb{M}}$ --- but not on $(K,S)$. Slope heuristics \citep{BirgeMassart2007,ArlotMassart2009} can be used in practice to calibrate $\lambda$: this is done in Section~\ref{Sect:In_practice}, where we use change-point detection \citep{LebarbierPhDThesis} in relation to slope heuristics.
Since $\h^2$ is upper bounded by $2$, the non-asymptotic feature of Theorem~\ref{thm:main-theorem} is interesting when $n$ is large enough with respect to $D_{(K, S)}$. However, even with small values of $n$, the simulations performed in Subsection~\ref{sect:Simulated_experiments} show that the penalized criterion calibrated using the slope heuristics keep good behaviors.
\subsection{A general tool for model selection} \label{Subsect:General_tool}
Theorem~\ref{thm:main-theorem} is obtained from \citep[Theorem~7.11]{Massart2007}. This last result deals with model selection problems by proposing penalty functions related to geometrical properties of the models, namely metric entropy with bracketing for Hellinger distance.
The framework here is the following. We consider some measurable space $(A, {\mathcal A})$, and $\mu$ a $\sigma$-finite positive measure on $A$. A collection of models $(\mathcal{M}_m)_{m \in \ensuremath{\mathbb{M}}}$ is given, where each model $\mathcal{M}_m$ is a set of probability density functions $s$ with respect to $\mu$. The following relation permits us to extend the definition of $\h$ to positive functions $s$ or $t$ whose integral is finite but not necessary $1$. Denoting $\sqrt{s}$ the function defined by $\sqrt{s}(x) = \sqrt{s(x)}$, and by $\| \cdot \|_2$ the usual norm in $\ensuremath{\mathbb{L}}^2(\mu)$, then
\begin{equation*}
\h(s,t) = \|\sqrt{s} - \sqrt{t} \|_2.
\end{equation*}
Let us now recall the definition of metric entropy with bracketing. Consider some collection $F$ of measurable functions on $A$, and $d$ one of the following metrics on $F$: $\h$, $\|\cdot\|_1$, or $\| \cdot \|_2$. A bracket $[l,u]$ is the collection of all measurable functions $f$ such that $l \leq f \leq u$. Its $d$-diameter is the distance $d(u,l)$. Then, for every positive number $\varepsilon$, we denote by $N_{[\cdot]}(\varepsilon, F, d)$ the minimal number of brackets with $d$-diameter not larger than $\varepsilon$ which are needed to cover $F$. The $d$-entropy with bracketing of $F$ is defined as the logarithm of $N_{[\cdot]}(\varepsilon, F, d)$, and is denoted by $H_{[\cdot]}(\varepsilon, F, d)$.
We assume that for each model $\mathcal{M}_m$ the square entropy with bracketing $\sqrt{H_{[\cdot]}(\varepsilon, \mathcal{M}_m, \h)}$ is integrable at $0$. Let us consider some function $\phi_m$ on $\ensuremath{\mathbb{R}}_{+}$ with the following properties
\begin{enumerate}
\setcounter{enumi}{8}
\renewcommand{\theenumi}{\textup{(\Alph{enumi})}
\item \label{cond:i} $\phi_m$ is nondecreasing, $x\mapsto \phi_m(x)/x$ is non-increasing on $(0,+\infty)$ and for every $\sigma \in \ensuremath{\mathbb{R}}_+$ and every $u \in \mathcal{M}_m$
\[ \int_{0}^{\sigma} \sqrt{H_{[\cdot]}\left(x, S_m(u,\sigma), \h\right)} d x \leq \phi_m(\sigma), \]
where $S_m(u,\sigma) = \left\{ t\in \mathcal{M}_m: \|\sqrt{t} - \sqrt{u}\|_2 \leq \sigma \right\}$.
\end{enumerate}
\ref{cond:i} is verified in particular with $\phi_m(\sigma) = \int_{0}^{\sigma} \sqrt{H_{[\cdot]}\left(x, \mathcal{M}_m, \h\right)} d x$.
In order to avoid measurability problems, we suppose that for each $m\in {\ensuremath{\mathbb{M}}}$, the following separability condition is verified for $\mathcal{M}_m$:
\begin{enumerate}
\setcounter{enumi}{12}
\renewcommand{\theenumi}{\textup{(\Alph{enumi})}
\item \label{cond:M} There exists some countable subset $\mathcal{M}_m'$ of $\mathcal{M}_m$ and a set $A'\subset A$ with $\mu(A')=\mu(A)$ such that for every $t\in \mathcal{M}_m$, there exists some sequence $(t_k)_{k\geq 1}$ of elements of $\mathcal{M}_m'$ such that for every $x\in A'$, $\ln(t_k(x))$ tends to $\ln(t(x))$ as $k$ tends to infinity.
\end{enumerate}
\begin{thm} \label{thm:Massart}
Let $X_1,\ldots,X_n$ be iid random variables with unknown density $s$ with respect to some positive measure $\mu$. Let $\{\mathcal{M}_m\}_{m \in \ensuremath{\mathbb{M}}}$ be some at most countable collection of models, each fulfilling \ref{cond:M}. We consider a corresponding collection of $\rho$-MLEs $(\widehat{s}_m)_{m \in \ensuremath{\mathbb{M}}}$. Let $\{x_m\}_{m \in \ensuremath{\mathbb{M}}}$ be some family of nonnegative numbers such that
\[ \sum_{m \in \ensuremath{\mathbb{M}}} e^{-x_m} = \Sigma < \infty, \]
and for every ${m \in \ensuremath{\mathbb{M}}}$ considering $\phi_m$ with property (i) define $\sigma_m$ as the unique positive solution of the equation
\begin{equation} \label{eq:def-sigma-m}
\phi_m(\sigma) = \sqrt{n} \sigma^2.
\end{equation}
Let $\pen_n: {\ensuremath{\mathbb{M}}} \rightarrow \ensuremath{\mathbb{R}}_+$ and consider the penalized log-likelihood criterion
\[ \crit(m) = \gamma_n\left(\widehat{s}_m\right) + \pen_n(m). \]
Then, there exists some absolute constants $\kappa$ and $C$ such that whenever
\begin{equation*}
\pen_n(m) \geq \kappa\, \left(\sigma_m^2 + \frac{x_m}{n}\right)\ \text{for every}\ m\in {\ensuremath{\mathbb{M}}},
\end{equation*}
some random variable $\widehat{m}$ minimizing $\crit$ over ${\ensuremath{\mathbb{M}}}$ exists and moreover, whatever the density $s$
\begin{equation*}
\ensuremath{\mathbb{E}}_s\left[\h^2\left(s,\widehat{s}_{\widehat{m}}\right)\right] \leq C\, \left( \inf_{m\in \ensuremath{\mathbb{M}}} \left(\KL\left(s,\mathcal{M}_m\right) + \pen_n(m)\right) + \rho + \frac{\Sigma}{n}\right).
\end{equation*}
\end{thm}
In Theorem~\ref{thm:Massart}, $\sigma_m^2$ has the role of a variance term of $\widehat{s}_m$, while the weights $x_m$ take into account the number of models $m$ having the same dimension.
\subsection{Proof of Theorem~\ref{thm:main-theorem}} \label{Subsect:Proof_main_result}
In order to apply Theorem~\ref{thm:Massart}, we need to compute the metric entropy with bracketing of each model $\mathcal{M}_{(K, S)}$. This is done in the following result, which is proved in Appendix~\ref{sec:entropy}.
\begin{prop}[Bracketing entropy of a model] \label{prop:entropy}
Let $\eta:\ensuremath{\mathbb{R}}_+ \rightarrow \ensuremath{\mathbb{R}}_+$ be the increasing convex function defined by
\begin{align*}
\text{Case~\ref{config:simple-multinomial}: } \eta(\varepsilon) &= (1+\varepsilon)^{L+1} - 1, \\
\text{Case~\ref{config:double-multinomial}: } \eta(\varepsilon) &= (1+\varepsilon)\, (1+ \sqrt{2}\, \varepsilon\, (2+\varepsilon))^L - 1.
\end{align*}
For any choice of $K$ and $S$, $\mathcal{M}_{(K, S)}$ fulfills \ref{cond:M}. For any $\varepsilon \in (0,1)$,
\[ H_{[\cdot]}\left(\eta(\varepsilon), \mathcal{M}_{(K, S)}, \h\right) \leq D_{(K, S)} \ln \left(\frac{1}{\varepsilon}\right) + C_{(K, S)},\]
where
\begin{equation}
\begin{split} \label{eq:def-RKS}
C_{(K, S)} &= \frac{1}{2} \bigg( \ln(2\pi e) D_{(K, S)} + \ln(4\pi e) \left(\mathbbm{1}_{K\geq 2} + L + (K-1) |S|\right) \\
&\quad + \mathbbm{1}_{K\geq 2} \ln(K+1) + \sum_{l=1}^{L} \ln(A_l+1) + (K-1) \sum_{l\in S} \ln(A_l+1).\bigg)
\end{split}
\end{equation}
\end{prop}
$C_{(K, S)}$ is a technical quantity measuring the complexity of a model $\mathcal{M}_{(K, S)}$.
In the next step we establish an expression for $\phi_m$. All following results are proved in Appendix~\ref{sec:phi}.
\begin{prop} \label{prop:phi-m}
For any choice of $m=(K, S)$, the function $\phi_m$ defined on $(0,\eta(1)]$ by
\[ \phi_m(\sigma) = \left(2\sqrt{\ln 2}\,\sqrt{D_{(K, S)}} + \sqrt{C_{(K, S)} - D_{(K, S)} \ln \eta^{-1}(\sigma)} \right)\, \sigma \]
fulfills \ref{cond:i}.
\end{prop}
We do not define $\phi_m$ for $\sigma$ bigger than $\eta(1)$, to avoid more complicated expressions. This is why a condition on $\xi$ appears in the following lemma:
\begin{lem} \label{lem:sigma-less-than-eta}
Let $\displaystyle A_{\max}=\sup_{1\leq l\leq L} A_l$, $\xi = \dfrac{4 \sqrt{A_{\max}}\sqrt{L}}{2^{L+1}-1}$ in Case~\ref{config:simple-multinomial}, and $\xi = \dfrac{4 \sqrt{A_{\max}}\sqrt{L}}{2 (1+3\sqrt{2})^{L}-1}$ in Case~\ref{config:double-multinomial}. Then, for all $n \geq 1$ if $\xi < 1$, and for $n > \xi^2 K$ otherwise, the solution $\sigma_m$ of (\ref{eq:def-sigma-m}) verifies $\sigma_m<\eta(1)$.
\end{lem}
From Proposition~\ref{prop:phi-m} we can deduce an upper bound for $\sigma_m$, with a similar reasoning to \citep{MaugisMichel2008thm}. First, $\sigma_m\leq \eta(1)$ entails $\eta^{-1}\left(\sigma_m\right)\leq 1$, and we obtain the lower bound $\sigma_m \geq \widetilde{\sigma}_{m}$, where
\begin{equation} \label{eq:sigma-tilde}
\widetilde{\sigma}_{m} = \frac{1}{\sqrt{n}} \left(2\sqrt{\ln 2}\,\sqrt{D_{m}} + \sqrt{C_{m}} \right).
\end{equation}
This can be used to get an upper bound
\begin{equation} \label{eq:sigma-upper}
\sigma_m \leq \frac{1}{\sqrt{n}} \left(2\sqrt{\ln 2}\,\sqrt{D_{m}} + \sqrt{C_{m} - D_{m} \ln \eta^{-1}\left(\widetilde{\sigma}_{m}\right)} \right).
\end{equation}
Let us now choose the weights $x_m$. If we take something bigger than $n \sigma_m^2$, this will change the shape of the penalty in Theorem~\ref{thm:Massart}. We define
\begin{equation*} \label{eq:x-m}
x_m = (\ln 2) D_{m}.
\end{equation*}
The following Lemma shows that this choice is suitable.
\begin{lem} \label{lem:sum-x-m}
For any model $\mathcal{M}_{m}$, with $m\in \ensuremath{\mathbb{M}}$ as above, let us set $x_m = (\ln 2) D_{m}$. Then
\[ \sum_{m\in\ensuremath{\mathbb{M}}} e^{-x_m} \leq (3/4)^L. \]
\end{lem}
To express the penalty function we have to lower bound $\eta^{-1}\left(\widetilde{\sigma}_{m}\right)$. This is done in the following Lemma.
\begin{lem} \label{lem:pen}
Using the preceding notations,
\[ \sigma_m^2 + \frac{x_m}{n} \leq \frac{D_{(K, S)}}{n}\, \left(5 + \sqrt{\max\left(\frac{1}{2} \ln n + \frac{1}{2} \ln L, \frac{\ln 2}{2} + \ln L\right)} \right)^2. \]
\end{lem}
This ends the proof of Theorem~\ref{thm:main-theorem}.
\section{In practice} \label{Sect:In_practice}
In real datasets the numbers $A_l$ of possible states at each variable $X^l$ are not necessarily known. The numbers $\widehat{A}_l$ of observed states can be used instead. In fact, the MLE estimator select a density with null weight on non-observed states. Then, in each model $\mathcal{M}_{(K, S)}$, an approximated MLE estimator can be computed thanks to the Expectation-Maximization (EM) algorithm (see \citep{Dempster1977}).
The other two points that have to be done before reaching the final estimator $\widehat{P}_{(\widehat{K}_n,\ \widehat{S}_n)}$ are the choice of the penalty function, and the sub-collection of models on which the EM algorithm will be used. These two points are discussed in Subsections~\ref{Subsect:slope_heuristics} and \ref{Subsect:sub-collection}. Then simulations are presented in Subsection~\ref{sect:Simulated_experiments}.
\subsection{Slope heuristics and Dimension jump} \label{Subsect:slope_heuristics}
Theorem \ref{thm:main-theorem} suggests to take a penalty function of the shape (\ref{eq:shape.penalty}), defined modulo a multiplicative parameter $\lambda$ which has to be calibrated. Slope heuristics, as presented in \citep{BirgeMassart2007,ArlotMassart2009}, provide a practical method to find an optimal penalty $\pen_{\mathrm{opt}}(m) = \lambda_{\mathrm{opt}} D_m/n$. These heuristics are based on the conjecture that there exists a minimal penalty $\pen_{\min}(m) =\lambda_{\min} D_m/n$ required for the model selection procedure to work: when the penalty is smaller that $\pen_{\min}$, the selected model is one of the most complex models, and the risk of the selected estimator is large. On the contrary, when the penalty is larger than $\pen_{\min}$, the complexity of the selected model is much smaller. Then the optimal penalty is close to twice the minimal penalty:
\begin{equation*}
\pen_{\mathrm{opt}}\left(m\right)\approx 2\lambda_{\min}D_m/n.
\end{equation*}
The name ``slope heuristics'' comes from $\lambda_{\min}$ being the slope of the linear regression $\gamma_n\left(\widehat{P}_{m}\right)\sim D_m/n$ for a certain sub-collection of the most competing models $m$. For example, on the left panel of Figure \ref{Fig:slope.heuristics.1} below, a slope is visible for the models containing the true model $\mathcal{M}_{\left(K_0,\ S_0\right)}$.
Even if this example is favorable and mainly here for illustration purposes, it shows that the slope heuristics are sensible with the modelings of the present work.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{Rplotdimension.eps}
\includegraphics[width=0.49\textwidth]{Rplotlambda.eps}
\caption{Two ways to compute the slope, on a simulated sample of $1000$ individuals, with $8$ clustering variables among $10$, and $5$ populations. Models have been explored via the modified backward-stepwise described in subsection~\ref{Subsect:sub-collection}, the number $K$ of clusters varying from $1$ to $10$. The size of the sliding window is $0.15$.}
\label{Fig:slope.heuristics.1}
\end{center}
\end{figure}
Instead of estimating $\lambda_{\min}$ by linear regression, another method is jump detection. Suppose we have at hand a reasonable grid $\lambda_1<\ldots<\lambda_r$ of candidate
values of $\lambda_{\min}$, and a sub-collection $\ensuremath{\mathbb{M}}_{explored}$ of the most competitive models. Each $\lambda_i$ leads to a selected model $\widehat{m}_i$ with dimension $D_{\widehat{m}_i}$. If you plot $D_{\widehat{m}_i}$ as a function of $\lambda_i$, $\lambda_{min}$ is expected to lie at the position of the biggest jump. However, the right panel of Figure~\ref{Fig:slope.heuristics.1} illustrates an important point: in that example the biggest jump is at $\lambda\approx 5.1$, but
$\lambda_{\min}$ is around $0.9$, which corresponds to several successive jumps. We propose an improved version of the dimension jump method of \citep{ArlotMassart2009}, based on a sliding window: we consider at a time all jumps in an window of $h\geq1$ following intervals in the grid. Algorithm~\ref{alg:calibration} below describes the procedure.
\begin{algorithm}
\caption{Calibration of Penalty$\left(\ensuremath{\mathbb{M}}_{explored}, \left(\lambda_i\right)_{i=1,\ldots,r}, h\right)$}
\label{alg:calibration}
\begin{algorithmic}
\FOR{$i=1$ to $n_\lambda$}
\STATE $\displaystyle \widehat{m}_i \leftarrow \mathop{\mathrm{arg\,min}}_{m\in\ensuremath{\mathbb{M}}_{explored}}\left\{\gamma_n\big(\widehat{P}_{m}\big)+\lambda_iD_m/n\right\}$
\ENDFOR
\STATE $\displaystyle i_{end} \leftarrow \min \mathop{\mathrm{arg\,max}}_{i\in\left\{h+1,\ldots,r\right\}}\left\{D_{\widehat{m}_{i-h}}-D_{\widehat{m}_{i}}\right\}$
\STATE $i_{init} \leftarrow \max \Big\{ j \in \left[i_{end}-h, i_{end}-1\right], D_{\widehat{m}_{j}}-D_{\widehat{m}_{i_{end}}} = D_{\widehat{m}_{i_{end}-h}}-D_{\widehat{m}_{i_{end}}}\Big\}$
\STATE $\displaystyle \widehat{\lambda}_{\min} \leftarrow \frac{\lambda_{i_{init}}+\lambda_{i_{end}}}{2}$
\RETURN $\widehat{\lambda}_{\min}$
\end{algorithmic}
\end{algorithm}
\subsection{Sub-collection of models for calibration} \label{Subsect:sub-collection}
For a given maximum value $K_{\max}$ of the number of clusters, the number of models under competition is equal to $1+\left(K_{\max}-1\right)*\left(2^L-1\right)$. Since this number is huge in most situations, it is very painful to consider all competing models for calibration of the parameter $\lambda$. On the other hand, we need enough models to ensure that there is a clear jump in the sequence of selected dimension. We consider the modified backward-stepwise algorithm proposed in \citep{ToussileGassiat2009}, which explores of cardinalities of $S$. It enables to gather the most competitive models among all possible $S$ for a given number $K$ of clusters and a given penalty function $\pen_n$. It gives also the choice to add a complementary exploration step based on a similarly modified forward strategy. We will refer to this algorithm as $explorer\left(K,\ \pen_n\right)$.
Since we do not know the final penalty during the exploration step, we consider a reasonable grid $1/2=\lambda_1<\ldots<\lambda_r=\ln n$ containing both penalty functions associated to $\AIC$ and $\BIC$ (\ref{eq:def-BIC-AIC}). To each value $\lambda_i$ of the grid is associated a penalty function $\pen_{\lambda_i}$. We launch $explorer\left(K,\ \pen_{\lambda_i}\right)$ for all values of $K$ in $\left\{1,\ldots,\ K_{\max}\right\}$ and for all values of $\lambda_i$ of the above grid, and we gather the explored models in $\ensuremath{\mathbb{M}}_{explored}$. This sub-collection seemly contains the most competitive models and it is then used to calibrate $\lambda$.
\subsection{Numerical experiments} \label{sect:Simulated_experiments}
Our proposed procedure with a data-driven calibration of the penalty function has been implemented for Case~\ref{config:double-multinomial} in the software \texttt{MixMoGenD} (Mixture Model using Genotypic Data), which already proposed a selection procedure based on asymptotic criteria $\BIC$ and $\AIC$ (see \citep{ToussileGassiat2009}). Here, we conduct numerical experiments on simulated datasets for performances assessment of the new non asymptotic criterion with respect to $\BIC$ and $\AIC$.
We present two experiments, both in Case~\ref{config:double-multinomial}. The first one considers the consistency of the selected model: we study how the procedure retrieves the main features of the true model as the number of individuals in the datasets increases. In the second one, we are rather interested in a validation of the model selection procedure from the oracle point of view: we compare the Hellinger risk of the selected estimator to the oracle risk.
\subsubsection{Consistency performances}
In this experiment we consider a setting with $L=10$ variables of $10$ states each. We chose a parameter with $K_0=5$ populations of equal probability. The frequencies of the states have been chosen such that the genetic differentiation between the populations is decreasing with the variables rank. In the first $6$ variables, the populations are more separated. In the following $2$ variables, the populations are very poorly differentiated. In the last $2$ variables, the states follow the same uniform distribution in all populations. The whole parameter is available at \url{http://www.math.u-psud.fr/~toussile/}.
We considered different values $n$ of the sample size in $\left[50,\ 900\right]$ and for each of them, $10$ datasets have been simulated. The results are summarized in Figure \ref{Fig:S_vs_n-K_vs_n}. The left panel gives the proportion of selecting the subset $\widehat{S}_n$ of clustering variables containing the first $6$ variables, which are the most genetically differentiated variables. The right panel gives the proportion of selected models with $\widehat{K}_n=K_0$.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{S.eps}
\includegraphics[width=0.49\textwidth]{K.eps}
\caption{The figure in the left panel gives the proportion of selected models with $\widehat{S}_n\supseteq \left\{1,\ldots,\ 6\right\}$, and the one in the right gives the proportion of selected models with $\widehat{K}_n=K_0$, versus the sample size.}
\label{Fig:S_vs_n-K_vs_n}
\end{center}
\end{figure}
In this experiment, $\AIC$ seems to be the best criterion for variable selection; however the different between $\AIC$ and the new criterion is not significant.
It also appears that $\AIC$ estimates the number of clusters better than the other criteria for small sample sizes (around $n=100$ and $n=200$), but it overestimates this number from $n=500$. On the contrary, the new criterion perfectly estimates the number of clusters for sample sizes $\geq 300$. $\BIC$ performs poorly for both variables selection and classification on datasets with small sizes. As expected, the data-driven calibration of the penalty function improves globally the performances of the selection procedure, and it gives thus an answer to the question ``Which penalty for which sample size?''.
It may happen that the results obtained on small sample sizes change a little from one run to another. In fact, the EM algorithm can miss the global maximum on such sample sizes, in particular in models of higher dimension. In our experiments, it is probably the case with some datasets of size $n\leq 300$, when the number of free parameters in the simulated model is $\geq 310$.
\subsubsection{Oracle performances of the estimator} \label{sect:oracle_experiment}
Since the new criterion is designed in an oracle perspective, it is interesting to compare the associated estimator to the oracle for Hellinger risk. Recall that the oracle is the estimator associated to the model indexed by the minimizer $\left(K^*,\ S^*\right)$ of the risk $\ensuremath{\mathbb{E}}\left[\h^2\left(P_0,\ \widehat{P}_{\left(K,\ S\right)}\right)\right]$ over the collection of models $\ensuremath{\mathbb{M}}$.
In this experiment, we consider simulated datasets with reduced variability in order to reduce the computation time. The parameter underlying the data admits $L=6$ variables, $3$ states for each variable, and $K_0=3$ populations with equal probability. The frequencies of the states have been chosen in such a way that the genetic differentiation between the population is significant on the first $3$ variables, very small on the $4^\text{th}$ and $5^\text{th}$ variables, while the states of the $6^\text{th}$ variable follow the uniform distribution in all populations. Thus the true model is defined by $K_0=3$ and $S_0=\left\{1,\ 2,\ 3,\ 4,\ 5\right\}$. The whole parameter is available at \url{http://www.math.u-psud.fr/~toussile/}.
We estimated the oracle using a Monte Carlo procedure on $100$ simulated datasets of size $500$ each, and got $\widehat{K^*}=3$ and $\widehat{S^*}=\left\{1,\ 2,\ 3,\ 4\right\}$. The results we obtained are summarized in Figure~\ref{Fig:Percentage.oracle} and Table~\ref{Tab:pairwise.t.test}.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.49\textwidth]{RapportRisqueHellinger.eps}
\includegraphics[width=0.49\textwidth]{Percentageoracle.eps}
\caption{The left panel gives the boxplots, means and their $95\%$ confident intervals, for $\frac{\h^2\big(P_0,\ \widehat{P}_{(\widehat{K}_n,\ \widehat{S}_n)}\big)}{\h^2\big(P_0,\ \widehat{P}_{(\widehat{K^*},\ \widehat{S^*})}\big)}$; the right panel gives the percentages of selection of the estimated oracle $\left(\widehat{K^*},\ \widehat{S^*}\right)$; three criteria have been used: $\AIC$, $\BIC$, and $\CteDim$ which denotes the new criterion with data-driven calibration of the penalty function.}
\label{Fig:Percentage.oracle}
\end{center}
\end{figure}
\begin{table}[ht]
\begin{center}
\begin{tabular}{rcc}
\hline
& $\AIC$ & $\BIC$ \\
\hline
$\AIC$ & - & $<5.40e-05$ \\
$\CteDim$ & $<2.02e-05$ & $<2.20e-16$ \\
\hline
\end{tabular}
\caption{The $p$-values of pairwise student tests comparing the means of the $\h^2\left(P_0,\ \widehat{P}_{\left(\widehat{K}_n,\ \widehat{S}_n\right)}\right)$. The alternative hypothesis is that the mean of the Hellinger distance associated to the criterion in the first column is less than the one associated to the criterion in the first line.}
\label{Tab:pairwise.t.test}
\end{center}
\end{table}
The worst behavior comes from $\BIC$ and it is not a surprise for two main reasons. First $\BIC$ is designed to find the true model which is different to the oracle in our experiments. Second, it is based on asymptotic approximation and therefore requires large samples. In contrary, compared to $\AIC$ and $\BIC$, the new criterion with data-driven calibration of the penalty function is significantly the best in the sense of Hellinger risk and the capacity of selecting the oracle. Recall that both $\AIC$ and the new criterion are designed to find the oracle (see Table \ref{Tab:pairwise.t.test}). But like $\BIC$, $\AIC$ is based on asymptotic approximations. So the advantage of the new criterion over $\AIC$ is probably that it is designed in a non asymptotic perspective.
\section{Conclusion}
In this paper, we have considered a model selection via penalization, which performs simultaneously a variables selection and a detection of the number of populations, in the specific framework of multivariate multinomial mixture. This leads to a clustering in a second time.
Our main result provides an oracle inequality, under the condition of some lower bound on the penalty function. The weakness of such a result is that the associated penalized criterion is not directly usable. Nevertheless, it suggests a shape of the penalty function which is of the form $\pen_n(m)=\lambda D_m/n$, where $\lambda=\lambda\left(n,\ \mathbb{M}\right)$ is a parameter which depends on the data and the collection of the competing models. In practice $\lambda$ is calibrated via the slope heuristics.
In the simulated experiments we conducted, the new criterion with penalty calibration shows good behaviors for density estimation as well as for the selection of the true model. It also performs well both when the number of individuals is large and when it is reasonnably small. This gives an answer to the question ``Which criterion for with sample size?''
In the modeling we considered, the model dimension grows rapidly. In real experiments the number of individuals can be small, so other modeling with reduced dimension may be needed. We currently work on models which cluster the populations differently for each variable, as well as models which allocate the same probability to several states.
\section*{Acknowledgment}
The authors gratefully acknowledge the comments and advice of Elisabeth Gassiat, Pascal Massart, and Gilles Celeux. Many thanks also to Nathalie Akakpo, Nicolas Verzelen, and Cathy Maugis, for the useful discussion we had.
|
\section{Gravity as quantum field theory\label{Sec:Gravity-QFT}}
Even though general relativity has been a very successful classical field theory of gravity, unlike other field theories, {\em e.g.}~electrodynamics, one cannot quantize it in the most straightforward manner, {\em i.e.}~using the canonical quantization or path integral formalism.
A perturbative loop expansion for gravity results in infinitely many ultraviolet (UV) divergent Feynman diagrams, and at each order the theory requires counterterms of ever-increasing degree in curvature. Consequently, it is necessary to fix an infinite number of free parameters to have a well-defined ultraviolet behavior, and we are left with a quantum theory that has no predictive power at small distance scales, where the counterterms are crucial. In this sense gravity is said to be non-renormalizable.
There are in principle two different ways to deal with the situation at hand. The first one is to abandon the whole idea of renormalizability, by assuming that gravity is an effective field theory only valid below the cutoff scale. Here the cutoff is a finite, physically meaningful quantity that acts as a short-distance regulator fixing the divergences in the Feynman diagrams. Beyond that scale the theory requires an ultraviolet completion. A typical example of such a theory is the Fermi theory of the weak nuclear force. The other way to proceed is to insist on the concept of a self-sufficient gravitational theory and further investigate perturbative or non-perturbative tools resulting in a renormalizable (or even finite) quantum theory of gravity.
Ho\v{r}ava--Lifshitz gravity~\cite{Horava:2008ih,Horava:2009uw} is grouped within the second category and investigates extensions to Einstein--Hilbert gravity that explicitly break local Lorentz invariance resulting in a power counting renormalizable $d+1$ dimensional theory of quantum gravity.
In the manuscript at hand key features of Ho\v{r}ava--Lifshitz gravity are reviewed. We then focus on the projectable version of Ho\v{r}ava--Lifshitz gravity~\cite{Horava:2009uw,Sotiriou:2009vn,Sotiriou:2009kx}.
The big appeal of Ho\v{r}ava--Lifshitz gravity as a quantum gravity candidate lies in the fact that --- once the theory is set up --- one has the chance to perform concrete computations and explicitly check its viability. As usual, it would be very difficult to prove the overall working success of the theory. Instead, one tends to focus on the various shortcomings in order to eventually establish whether they can be overcome or whether the theory is ruled out. Our intention here is to overview projectable Ho\v{r}ava--Lifshitz gravity and summarize its shortcomings.
\section{Emergence of local Lorentz symmetry\label{Sec:LIV}}
The key idea behind Ho\v{r}ava--Lifshitz gravity is to sacrifice local Lorentz symmetry in exchange for renormalizability: Local Lorentz symmetry is violated at the non-perturbative level of our field theory, but will hopefully be recovered at long distance\,/\,low energy scales. The particular kind of Lorentz violations we will encounter here have been discussed to some extent within quantum gravity phenomenology~\cite{Mattingly:2005aa,Jacobson:2005aa,Liberati:2006sj}. Let us consider the following class of theories that break Lorentz invariance fundamentally, and that at the kinematical level lead to a dispersion relation which is some function of momentum and mass, $E^2 = m^2 + p^2 \to E^2 = F(p,m)$.
Specifically, the dispersion relations encountered by Ho\v{r}ava--Lifshitz theories are truncated at finite order $2z$,
\begin{equation} \label{Eq:DispRelLIV}
E^2= m^2 + p^2 + (\tilde\eta_{1})_{i} \, p^{i} + (\tilde\eta_{2})_{ij} \, p^{i}p^{j} + \dots + (\tilde\eta_{2 z})_{i \dots j} \, p^{i} \dots p^{j},
\end{equation}
The truncation coefficient $z$ is also referred to as the degree of anisotropy, adopted from condensed-matter jargon referring to the anisotropic scaling between space and time.
As will be demonstrated shortly, local Lorentz symmetry breaking of the above type can act as a quantum field theory regulator in the ultraviolet for appropriate choices of $z$. Next we focus on two particular examples and discuss their (power-counting) renormalizability.
\section{Renormalizability of anisotropic field theories\label{Sec:Power}}
We will lay out in brief a road map to construct field theories in which Lorentz symmetry breaking acts as a quantum field theory regulator. The first example will be a polynomial interacting scalar field, and the second will be gravity. The study of the scalar field case (Lifshitz scalar) is mainly of pedagogical interest, as any detailed renormalization group calculations for the case of gravity are still pending and all assumptions are based on power-counting~\cite{Anselmi:2008bq,Anselmi:2008bs,Anselmi:2008bt,Anselmi:2008ry,Visser:2009ul,Visser:2009ys}. This is in contrast to scalar field theories in this setting, which have been investigated more closely over the last few months~\cite{Anselmi:2008bq,Anselmi:2008bs,Anselmi:2008bt,Anselmi:2008ry,Visser:2009ul,Iengo:2009ix,Visser:2009ys}.
\subsection{Lifshitz scalar field theories\label{Sec:Anisotropic-Scalar}}
The simplest yet insightful field theory to demonstrate the above concepts is a (massive) scalar field with polynomial interactions on a flat background,
\begin{equation}
\label{scalaraction}
S = \int \left\{ \dot \phi^2 - \phi \tilde{D} \phi - m^2 \phi^2 + \sum_{n=1}^{N} g_n \phi^n \right\} d t \; d^d x ,
\end{equation}
where the differential operator $\tilde{D}$ is given by
\begin{equation}
\label{Deq}
\tilde{D}= -c^2 \triangle + \dots + (-\triangle)^z.
\end{equation}
For a relativistic field theory $z=1$ and $\tilde{D}$ is the standard Laplacian, $\triangle$. These kind of theories are well understood, and we know that $\phi^n$\,/\,$\phi^6$\,/\,$\phi^4$\,/\,$\phi^3$ are renormalizable in $1+1$\,/\,$2+1$\,/\,$3+1$\,/\,$5+1$ dimensions respectively.
Suppose now that $z$ is left arbitrary and we impose anisotropic scaling between space and time, using the engineering dimensions
\begin{equation}\label{Eq:AS}
\left[ dx \right] = \left[ \kappa \right]^{-1} \quad \mbox{and} \quad
\left[ dt \right] = \left[ \kappa \right]^{-z},
\end{equation}
where $\kappa$ is a placeholder symbol for some object with dimensions of momentum (this scaling is the most convenient for power counting in this case). Choosing the dimensions of $\phi$ such that $[\phi]^2 [dt \, dx^d]=[1]$,
$\triangle^z$ becomes a marginal operator in eq.~(\ref{Deq}), and the lower-order $\triangle^n$ operators become relevant.
The coupling constants $g_n$ in action (\ref{scalaraction}) have dimensions $[g_n] = [\kappa]^{d+z- n(d-z)/2} = [m]^{[d+z- n(d-z)/2]/z}$, which are non-negative as long as $d+z- n(d-z)/ 2 \geq 0$. For $n \to\infty$ this would require simply $z \geq d$. If we apply our experiences from studies of Lorentz invariant scalar theories, non-negative momentum dimensions are a good indication for renormalizability. In reference~\cite{Visser:2009ul} the superficial degree of divergence was derived and it was shown that for $z=d$ ($z > d$) and with (without) normal ordering the model at hand is perturbatively ultraviolet finite.
Before we move on to gravity we would like to stress that although the ultraviolet behaviour appears to be significantly improved by the inclusion of suitable higher order operators in (\ref{scalaraction}) it has already been shown that the one-loop renormalization and evolution of the coupling constants, for two interacting anisotropic scalar fields, is facing the naturalness problem, see reference~\cite{Collins:2004aa,Iengo:2009ix}: The difference in the speed of light between the two fields runs logarithmically to zero, and thus persists at low energy scales. This is in contradiction to our experimental constraints on Lorentz symmetry breaking, unless we impose some unnatural fine tuning on the bare coupling constants.
\subsection{Ho\v{r}ava--Lifshitz gravity \label{Sec:Anisotropic-Gravity}}
The natural setting in which to apply the concepts discussed above to gravity is within the Hamiltonian formulation of general relativity based on the Arnowitt--Deser--Misner (ADM) decomposition of spacetime,
\begin{equation} \label{Eq:ADM}
\mathrm{d} s^2 = - N^2 c^2 \mathrm{d} t^2 + g_{ij}(\mathrm{d} x^i - N^i \mathrm{d} t) (\mathrm{d} x^j - N^j \mathrm{d} t).
\end{equation}
The ADM formalism foliates spacetime into a family of non-intersecting spacelike hypersurfaces, where the metric field $g_{ij}$ is the induced geometry on the spatial slice, the scalar function $N$ is the lapse, and the three-vector $N^i$ is the shift-vector. The extrinsic curvature is defined as
\begin{equation}
K_{ij}=\frac{1}{2 N} \left\{ -\dot{g}_{ij} + \nabla_{i} N_{j} + \nabla_{j} N_{i} \right\}.
\end{equation}
Within this setting one can write down an action of the form
\begin{equation}
S=\int \left[T - V \right] \sqrt{g} \, N \, d^3x \, dt,
\end{equation}
where $T$ denotes the part that contains time derivatives and is given by
\begin{equation} \label{Eq:ExtCurvature}
T (K) = g_{K} \left\{ (K^{ij}K_{ij} - K^2) + \xi K^2 \right\}.
\end{equation}
The potential $V$ on the other hand does not contain time derivative but it can in principle contain any 3-covariant term constructed with the pieces of the ADM metric. We are abandoning Lorentz invariance and therefore we cannot automatically require $\xi$ to vanish, as is the case in general relativity. Similarly, $V$ can now include higher order spatial derivatives instead of just being equal to the 3-dimensional Ricci scalar, as in general relativity.
An action of the above form is clearly not invariant under general diffeomorphisms. It is, however, invariant under the restricted set of diffeomorphisms that preserve the foliation, {\em i.e.}~$t \to t + \chi^0(t)$, $x^i \to x^i + \chi^i(t,\mathbf{x})$. This symmetry is enough to prevent $N^i$ from appearing in $V$. In the original proposal from Ho\v{r}ava $N$ was also assumed not to enter in $V$ \cite{Horava:2009uw}, however, this constitutes an ad hoc assumption as noticed by \cite{Blas:2009qj} (we will return later to the various variants of the theory).
The way to proceed now is to invoke anisotropic engineering dimensions, as in equation~(\ref{Eq:AS}), and to evaluate the dimensionality of coupling constants of the leading order operators in the UV. Starting from the ADM decomposition~(\ref{Eq:ADM}), we need to have $[N^i]=[\kappa]^{z-1}$ and we are free to choose $[g_{ij}]=[N]=[1]$, resulting in the dimensionality for the line-element $[ds]=[\kappa]^{-1}$, for the speed of light $[c]=[\kappa]^{z-1}$, the volume-element $[dV_{d+1}]=[\kappa]^{-d-z}$, and the extrinsic curvature $[K]=[\kappa]^z$. Under the constraint that the action has to be dimensionless, we obtain for $g_K$ and for the coupling(s) $g_{2z}$ of the operator(s) of order $2z$ in $V$
\begin{equation}
[g_{K}]=[g_{2z}]=[\kappa]^{(d-z)}.
\end{equation}
These operators will dominate at the UV as they are the leading ones in temporal and spatial derivatives respectively. Choosing $z=d$ will make these couplings dimensionless.
(A more detailed treatment of the power-counting and the derivation in general can be found in references~\cite{Sotiriou:2009vn,Sotiriou:2009kx}.)
This implies that by including in $V$ operators of order $2d$, we would end up with a power-counting renormalizable theory, along the lines of the Lifshitz scalar case discussed previously.
What remains to be discussed is which can be the exact form of $V$. As already mentioned, in his initial proposal Ho\v{r}ava assumed that $V$ depends only on $g_{ij}$ and its spatial derivatives. To simplify things further he proposed a specific combination of terms, dubbed detailed balance (see \cite{Horava:2009uw} for more details on the exact form of V in this case), which is inspired by quantum critical systems and non-equilibrium critical phenomena. The hope seemed to be that such a rigid construction would be imposed as a symmetry, simplifying renormalization group calculations. There is, however, no clear physical motivation for detailed balance apart from simplicity, and instead detailed balance appears to be disfavored by the first consistency tests, see for example~\cite{Sotiriou:2009vn,Charmousis:2009mz,Lu:2009fj}.
Once detailed balance is abandoned there is no reason to exclude any possible term.
With or without detailed balance, however, Ho\v{r}ava's model propagates an extra scalar mode with respect to general relativity and appears to be burdened with serious shortcomings , such as instabilities, overconstrained evolution and strong coupling at low energies \cite{Charmousis:2009mz,Blas:2009yd,Li:2009bg,Henneaux:2009zb}. In Ref.~\cite{Blas:2009qj} an extension was proposed, after noticing that terms involving $N$ and its spatial derivatives can be included in $V$ without violating the symmetry of the action. The scalar mode in this model does exhibit improved behavior. However, it still possesses strong coupling at energies orders of magnitude lower than the Planck scale~\cite{Papazoglou:2009fj}. One could do away with the strong coupling if the higher order operators become important at energies below this strong coupling scale, but this requires tuning of their coupling constants (or the introduction of a new scale) \cite{Blas:2009ck,Papazoglou:2009fj}.
We will not review the literature any further here. On what comes next, we will focus instead on a specific version of Ho\v{r}ava--Lifshitz gravity, the so-called projectable version, for reasons that will become clear shortly.
\section{Projectable Ho\v{r}ava--Lifshitz gravity\label{Sec:Complete.Horava}}
We have argued that it is indeed possible to construct a power-counting renormalizable theory of gravity by adding ad-hoc spatial derivatives with twice the order of the spatial dimension of the theory. Therefore, in 3 spatial dimensions there is a plethora of terms which, when added to the Einstein--Hilbert action could lead to a theory with the desirable characteristics. As discussed above, there is no real guiding principle that would allow us to exclude any lower or 6th order operators. On the other hand, taking into account all terms is hardly an easy task from a practical perspective.
Another restriction proposed by Ho\v{r}ava was to impose that the lapse $N$ be just a function of time $t$. The motivation behind this restriction came from the desire to match the reduced symmetry of the action mentioned earlier. There seems to be nothing fundamental in this restriction. On the contrary, it is bound to cause several problems, for instance, it will not allow us to access any gauge choice for which $N$ is not a function of time only.\footnote{On the other hand, it should be noted that the most physically interesting solutions of the Einstein equations, the Schwarzschild, Reissner--Nordstr\"om, Kerr, and Friedmann--Lemaitre--Robertson--Walker spacetimes can by suitable choice of coordinates all be cast into projectable form, in fact with $N=1$~\cite{Wiltshire:2009zz}.} Most importantly, when $N=N(t)$ the Hamiltonian constraint will not be local, as one cannot get rid of the spatial integral after the variation. Taking this shortcomings into account, assuming $N=N(t)$ has also a serious advantage: it allows us to do away with several terms in the action, as all spatial derivatives of $N$ vanish and at the same times many of the possible terms one can write end up differing only by a total divergence. This makes calculations tractable even for the most general action one could write (see also \cite{Sotiriou:2009vn,Sotiriou:2009kx}). In this sense, it is worth exploring this version of the theory, called projectable Ho\v{r}ava--Lifshitz gravity.
Within this setup the most general action can be put together following the lines of Refs.~\cite{Sotiriou:2009vn,Sotiriou:2009kx}: applying integration by parts (discarding surface terms), commutator and Bianchi identities, and special relations arising in $3$ dimensions (vanishing of the Weyl tensor), one obtains, besides the standard Einstein--Hilbert terms,
\begin{equation}
S_\mathrm{EH} = \int \left\{ (K^{ij} K_{ij} - K^2) + \zeta^4 R - g_0\, \zeta^6 \right\} \sqrt{g} \; N\; {\mathrm{d}}^3 x \; {\mathrm{d}} t,
\end{equation}
new Lorentz-violating terms involving in total eight dimensionless coupling constants $(\xi, g_2, \dots, g_8)$
\begin{eqnarray}
S_\mathrm{LIV} &=& \int \Big\{ \xi\, K^2 - g_2 \,\zeta^2\,R^2 - g_3 \, \zeta^2\, R_{ij} R^{ij}
- g_4 \, R^3 - g_5 \, R (R_{ij} R^{ij}) - g_6 \, R^i{}_j R^j{}_k R^k{}_i
\nonumber\\
&&
\qquad
- g_7 \, R \nabla^2 R - g_8 \, \nabla_i R_{jk} \, \nabla^i R^{jk}
\Big\} \sqrt{g} \; N\; {\mathrm{d}}^d x \; {\mathrm{d}} t. \qquad
\end{eqnarray}
Here we have introduced a scale $\zeta$ in order to make all couplings dimensionless.
Instead of explicitly deriving the constraint and field equations by variation with respect to the lapse (Hamiltonian constraint) and the shift vector (super-momentum constraint), and the induced metric (dynamical equations of motion), we refer the reader to Refs.~\cite{Sotiriou:2009vn,Sotiriou:2009kx}. However, we would like to point out the key difference between the projectable and the non-projectable case. As mentioned previously, in the former the Hamiltonian constraint is not a local one, implying that there is no local notion of energy conservation. At best we can say that the integral of the energy at each spatial slice is conserved.
It is worth noting that so far we have been working in units that impose anisotropic scaling between space and time simply because this is convenient for power-counting renormalizability arguments. If instead one needs to work in units where the speed of light $c$ is equal to $1$, then this can be easily achieved by setting $dt\to \zeta^{-2} dt$. After writing the action in these new units, see~\cite{Sotiriou:2009vn,Sotiriou:2009kx}, one can read off the Newton and cosmological constant,
\begin{equation}
\frac{1}{16 \pi G_\mathrm{Newton}}= \zeta^2 \quad \qquad \mbox{and} \quad \qquad \Lambda = {g_0 \, \zeta^2\over2},
\end{equation}
so that $\zeta$ is identified as the Planck scale.
The main result up to now is that it is possible to include all possible terms and to explore --- within the restricted but complete model --- the key features of Ho\v{r}ava--Lifshitz gravity. It should be noted that this result can, in principle, be extended to the non-projectable case through a challenging but straightforward calculation.
\section{Excitations in projectable Ho\v{r}ava--Lifshitz gravity\label{Sec:Excitations}}
One of the most surprising features of general relativity is that, out of the possible fundamental excitations --- spin-$0$, spin-$1$ and spin-$2$ --- only the spin-$2$ excitation, the graviton, is physical. The spin-$1$ and spin-$0$ modes can be gauged away using general covariance. This feature of general relativity is in perfect agreement with observations, which have not so far revealed extra fields mediating gravity, see for example~\cite{Will:2005va}. Therefore, a crucial viability test for any gravity theory is whether it exhibits other physical degrees of freedom which would have already been detected by observations.\footnote{It is possible to construct condensed matter systems that exhibit spin-$2$ excitations, but in these systems all the lower-spin excitations are also physical due to the lack of general coordinate covariance. See for example Ref.~\cite{Gu:2006yq}.} In $3+1$ dimensions it is difficult to answer this question directly. It is much easier to study the degrees of freedom at each order perturbing around some specific background solution.
Within the framework of projectable Ho\v{r}ava--Lifshitz gravity one can assume that $g_0=0$ for simplicity and perturb around flat spacetime geometry, along the lines of Ref.~\cite{Sotiriou:2009kx}. After imposing suitable gauge conditions one can show that there exist a (suitably defined) transverse traceless mode ${ \widetilde H}_{ij}$ satisfying the equation
\begin{equation}
\ddot{ \widetilde H}_{ij} = -\left[ - \partial^2 + g_3 \zeta^{-2} \partial^4 + g _8 \zeta^{-4} \partial^6 \right] \widetilde H_{ij}.
\end{equation}
Clearly, the spin-$2$ particle has the dispersion relation discussed previously and shown in eq.~(\ref{Eq:DispRelLIV}).
However, there exist an additional scalar mode: the trace $h$ of the perturbation to the metric $h_{ij}$, which satisfies the following equation
\begin{equation}
\label{xineq0scalar}
\left(1-{3\over2}\, \xi\right) \; \ddot h = - \xi \left\{ -{1\over2} \partial^2 + \left(-4 g_2 - {3\over2} g_3 \right) \zeta^{-2} \partial^4 + \left(4 g_7 -{3\over2} g_8\right) \zeta^{-4} \partial^6 \right\} h.
\end{equation}
This is a wave equation with a trans-Bogoliubov dispersion relation.
At low momentum the group and phase velocity approach the same limiting propagation speed,
\begin{equation}
c_\mathrm{spin-0}^2 = {\xi \; \over 2-3\xi},
\end{equation}
The low momentum linearized action for this scalar mode is
\begin{equation}
S_\mathrm{spin-0}~=-\int d^3x dt \left[\frac{1}{c_\mathrm{spin-0}^2} \dot{h}^2-(\partial h)^2\right].
\end{equation}
From the last two equations one can infer the following: for $\xi<0$ and $\xi>2/3$ the scalar modes is classically unstable. On the other hand, for $\xi\in(0, 2/3)$, the mode is classically stable but then the kinetic term has the wrong sign in the action. Therefore, for these values the mode is quantum mechanically unstable (ghost). This leaves us with only two special values allowed, $\xi=2/3$ and $\xi=0$.
The latter is the value $\xi$ has in general relativity, whereas the former is an order 1 deviation from this value and we will not consider it futher.
For $\xi=0$ the gauge used to derived the propagator presented above is no longer accessible and we have to do a separate special case study. Using a slightly different gauge led to a peculiar mode growing quadratically with time in Ref.~\cite{Sotiriou:2009kx}. As it was later demonstrated in Ref.~\cite{Wang:2009yz} this mode was actually a gauge artifact and the scalar appears to be frozen around flat background (see also Ref.~\cite{Gao:2009ht} for Friedmann--Lema\^itre--Robertson--Walker spacetimes).
It appears that there is no scalar mode for $\xi=0$ whereas instabilities plagued nearby values. However, there are two different reasons for which setting $\xi=0$ by fiat will not work. Firstly, $\xi$ is a running coupling and we are lacking a symmetry that would allow us to set it to any specific value. Secondly, appearances deceive and the $\xi=0$ choice is not as healthy as it seems. In fact, even though for this value the mode appears to freeze around maximally symmetric backgrounds, this is only due to the fact that it actually gets strongly coupled at all scales \cite{Charmousis:2009mz,Koyama:2009yq}. It seems that we have run out of available options for the value of $\xi$. On the other hand we know that we would have to drive $\xi$ to zero eventually if general relativity is to be recovered at the infrared.
\section{Discussion\label{Sec:Conclusions}}
After discussing the motivation and giving a brief overview of the various versions of Ho\v{r}ava--Lifshitz gravity, we focussed on one of them, the projectable version where the lapse is assumed to be only a function of time and not space. The main advantage of this version is that it drastically reduces the number of terms that one must consider in the action and it, therefore, makes it possible to perform explicit calculations using the most general action which respects the symmetries of the theory.
Unfortunately, projectable Ho\v{r}ava--Lifshitz gravity appears to be plagued with serious problems related to the existence of a scalar degree of freedom. This scalar mode turns out to be classically or quantum mechanically unstable in general, and when parameters of the theory are chosen such that the low energy limit corresponds to general relativity, the scalar exhibits strong coupling at all scales. One could still hope for a non-perturbative restoration of the limit to general relativity \`a la Vainshtein in massive gravity \cite{Vainshtein:1972sx}. This is an avenue worth exploring but so far such an effect has not been shown to be present.
In the light of this we could choose to view the projectable version of Ho\v{r}ava--Lifshitz gravity as a useful tool in order to gain insight in the more complete versions of the theory. Indeed,
some of the characteristics of this more limited version, such as the strong coupling, are still shared by other versions, but some are not ({\em e.g.}~the fact that the Hamiltonian constraint is not a local one). Therefore, its study can provide some insight in the other versions, but only once the subtle differences are carefully weighed as well.
To conclude, Ho\v{r}ava--Lifshitz gravity constitutes an interesting quantum gravity theory. It appears challenging to construct a viable model within this framework. In fact there are numerous issues that have not even been systematically considered yet, such as matter coupling, renormalization group calculations, \emph{etc}. Given the fact that sensible renormalizable gravity theories are not easy to construct, Ho\v{r}ava--Lifshitz gravity seems to deserve further study.
\section*{References\label{Sec:References}}
|
\subsection{#1}}}
\newcommand\VD[1]{}
\newcommand\VND[1]{#1}
\newcommand\VL[1]{#1}
\newcommand\VS[1]{}
\newcommand{\textrm{Tr}}{\textrm{Tr}}
\newcommand{\textrm{Herm}^+_{d^2}(\mathbb{C})}{\textrm{Herm}^+_{d^2}(\mathbb{C})}
\newcommand{\textrm{Herm}^+_{d}(\mathbb{C})}{\textrm{Herm}^+_{d}(\mathbb{C})}
\newcommand{\textrm{Herm}^+_{dm}(\mathbb{C})}{\textrm{Herm}^+_{dm}(\mathbb{C})}
\newcommand{\textrm{Herm}^+_{m}(\mathbb{C})}{\textrm{Herm}^+_{m}(\mathbb{C})}
\newcommand{\operatorname{ad}}{\operatorname{ad}}
\newcommand{\abs}[1]{\left|#1\right|}
\newcommand{\mathbb{C}}{\mathbb{C}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{B}}{\mathcal{B}}
\newcommand{\mathcal{C}}{\mathcal{C}}
\newcommand{\mathcal{H}}{\mathcal{H}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\pa}[1]{\left(#1\right)}
\newcommand{\acco}[1]{\left\{#1\right\}}
\newcommand{\norme}[1]{\left|\left|#1\right|\right|}
\newcommand{{\bf\mathcal{S}}}{{\bf\mathcal{S}}}
\newcommand{\emph{etc.\ }{}}{\emph{etc.\ }{}}
\newcommand{\emph{et al\ }{}}{\emph{et al\ }{}}
\newcommand{\emph{e.g.\ }{}}{\emph{e.g.\ }{}}
\newcommand{\emph{i.e.\ }{}}{\emph{i.e.\ }{}}
\newcommand{\emph{cf.\ }{}}{\emph{cf.\ }{}}
\newcommand{\emph{vide infra}{}}{\emph{vide infra}{}}
\newcommand{R($\frac{\pi}{4})${}}{R($\frac{\pi}{4})${}}
\newcommand{controlled-\Phase{}}{controlled-R($\frac{\pi}{4})${}{}}
\newcommand{\cells}[4]{
\centering
\Ket{
\,
\begin{tabular}{ | p{2.8mm} | p{2.8mm} | }
\hline
#1 & #2 \\ \hline
#3 & #4 \\ \hline
\end{tabular}\,
}
}
\newcommand{\cellcolor[gray]{0.6}}{\cellcolor{orange}}
\newtheorem{Iso}{Isomorphism}
\newtheorem{Th}{Theorem}
\newtheorem{Cl}{Claim}
\newtheorem{Lem}{Lemma}
\newtheorem{Proposition}{Proposition}
\newtheorem{Scenario}{Scenario}
\newtheorem{Def}{Definition}
\begin{document}
\title{A Simple $n$-Dimensional Intrinsically Universal Quantum Cellular Automaton}
\author{Pablo Arrighi \and Jonathan Grattage}
\institute{University of Grenoble, \VL{Laboratoire }LIG,\VL{\\}
\VL{B\^{a}timent IMAG C, }220 rue de la Chimie,\VL{\\}
38400 \VS{SMH}\VL{Saint-Martin-d'H\`eres}, France
\and
\VL{Ecole Normale Sup\'erieure de }\VS{ENS-}Lyon, \VL{Laboratoire }LIP,\VL{\\}
46 all\'ee d'Italie, 69364 Lyon cedex 07, France}
\titlerunning{A Simple $n$-Dimensional Intrinsically Universal QCA}
\authorrunning{P. Arrighi and J. Grattage}
\toctitle{A Simple $n$-Dimensional Intrinsically Universal Quantum Cellular Automaton}
\tocauthor{Pablo Arrighi and Jonathan Grattage}
\maketitle
\begin{abstract}
We describe a simple $n$-dimensional quantum cellular automaton (QCA) capable of simulating all others,
in that the initial configuration and the forward evolution of any $n$-dimensional QCA can be encoded within the initial configuration of the intrinsically universal QCA. Several steps of the intrinsically universal QCA then correspond to one step of the simulated QCA. The simulation preserves the topology in the sense that each cell of the simulated QCA is encoded as a group of adjacent cells in the universal QCA.
\end{abstract}
\section{Introduction}\label{sec:introduction}\label{subsec:CA}
Cellular automata (CA), first introduced by Von Neumann \cite{Neumann}, consist of an array of identical cells, each of which may take one of a finite number of possible states. The whole array evolves in discrete time steps by iterating a function $G$. This global evolution $G$ is shift-invariant (it acts everywhere the same) and local (information cannot be transmitted faster than some fixed number of cells per time step).
Because this is a physics-like model of computation \cite{MargolusPhysics}, Feynman \cite{FeynmanQCA}, and later Margolus \cite{MargolusQCA}, suggested
that quantising this model was important, for two reasons: firstly, because in CA computation occurs without extraneous (unnecessary) control, hence eliminating a source of decoherence; and secondly because they are a good framework in which to study the quantum simulation of a quantum system. From a computation perspective there are other reasons to study QCA, such as studying space-sensitive problems in computer science, \emph{e.g.\ }{} `machine self-reproduction' \cite{Neumann} or `Firing Squad Synchronisation', which QCA allow in the quantum setting. There is also a theoretical physics perspective, where CA are used as toy models of quantum space-time \cite{LloydQG}. The first approach to defining QCA \cite{ArrighiMFCS,DurrWell,Watrous} was later superseded by a more axiomatic approach \cite{ArrighiUCAUSAL,ArrighiLATA,SchumacherWerner} together with the more operational approaches \cite{BrennenWilliams,NagajWocjan,PerezCheung,Raussendorf,VanDam,Watrous}.
The most well known CA is Conway's `Game of Life', a two-dimensional CA which has been shown to be universal for computation, in the sense that any Turing Machine (TM) can be encoded within its initial state and then executed by evolution of the CA. Because TM have long been regarded as the best definition of `what an algorithm is' in classical computer science, this result could have been perceived as providing a conclusion to the topic of CA universality. This was not the case, because CA do more than just running any algorithm. They run distributed algorithms in a distributed manner, model phenomena together with their spatial structure, and allow the use of the spatial parallelism inherent to the model. These features, modelled by CA and not by TM, are all interesting, and so the concept of universality must be revisited in this context to account for space. This is achieved by returning to the original meaning of the word \emph{universality} \cite{AlbertCulik,Banks,DurandRoka}, namely the ability for one instance of a computational model to be able to simulate other instances of the same computational model. Intrinsic universality formalises the ability of a CA to simulate another in a space-preserving manner \cite{MazoyerRapaport,OllingerJAC,Theyssier}, and was extended to the quantum setting in \cite{ArrighiUQCA,ArrighiNUQCA,ArrighiPQCA}.
There are several related results in the CA literature. For example, refs. \cite{MargolusPhysics,MoritaCompUniv1D,MoritaCompUniv2D} provide computation universal Reversible Partitioned CA constructions, whereas ref. \cite{MoritaIntrinsicUniv1D} deals with their ability to simulate any CA in the one-dimensional case. The problem of minimal intrinsically universal CA was addressed, \emph{cf.\ }{} \cite{OllingerRichard}, and for Reversible CA (RCA) the issue was tackled by Durand-Lose \cite{Durand-LoseLATIN,Durand-LoseIntrinsic1D}. The difficulty is in having an $n$-dimensional RCA simulate all other $n$-dimensional RCA and not, say, the $(n-1)$-dimensional RCA, otherwise a history-keeping dimension could be used, as by Toffoli \cite{ToffoliConstruction}. There are also several other QCA related results. Watrous \cite{WatrousFOCS} has proved that QCA are universal in the sense of QTM. Shepherd, Franz and Werner \cite{ShepherdFranz} defined a class of QCA where the scattering unitary $U_i$ changes at each step $i$ (CCQCA). Universality in the circuit-sense has already been achieved by Van Dam \cite{VanDam}, Cirac and Vollbrecht \cite{VollbrechtCirac}, Nagaj and Wocjan \cite{NagajWocjan} and Raussendorf \cite{Raussendorf}. In the bounded-size configurations case, circuit universality coincides with intrinsic universality, as noted by Van Dam \cite{VanDam}. QCA intrinsic universality in the one-dimensional case is resolved in ref. \cite{ArrighiFI}. Given the crucial role of this in classical CA theory, the issue of intrinsic universality in the $n$-dimensional case began to be addressed in refs. \cite{ArrighiNUQCA,ArrighiPQCA}, where it was shown that a simple subclass of QCA, namely Partitioned QCA (PQCA), are intrinsically universal. Having shown that PQCA are intrinsically universal, it remains to be shown that there exists a $n$-dimensional PQCA capable of simulating all other $n$-dimensional PQCA for $n>1$, which is presented here.
PQCA are QCA of a particular form, where incoming information
is scattered by a fixed unitary $U$ before being redistributed. Hence the problem of finding an intrinsically universal PQCA reduces to finding some universal scattering unitary $U$ (this is made formal in section \ref{subsec:flat}, see Fig.\ref{fig:flattening34}). Clearly the universality requirement on $U$ is much more difficult than just quantum circuit universality. This is because the simulation of a QCA $H$ has to be done in a parallel, space-preserving manner. Moreover we must simulate not only an iteration of $H$ but several ($H^2$, \ldots), so after every simulation the universal PQCA must be ready for a further iteration.
From a computer architecture point of view, this problem can be recast in terms of finding some fundamental quantum processing unit which is capable of simulating any other network of quantum processing units, in a space-preserving manner. From a theoretical physics perspective, this amounts to specifying a scattering phenomenon that is capable of simulating any other, again in a space-preserving manner.
\VD{
\section{Definitions} \label{definitions}
\VL{\subsection{$n$-Dimensional (Partitioned) QCA}}\label{subsecdef}
First we recall the necessary definitions for $n$-dimensional QCA.
Config\-urations hold the basic states of an entire array of cells, and hence denote the possible basic states of the entire QCA:
\begin{Def}\textbf{(Finite configurations)}
A \emph{(finite) configuration} $c$ over $\Sigma$ is a function $c:
\mathbb{Z}^n \longrightarrow \Sigma$, with $(i_i,\ldots,i_n)\longmapsto
c(i_i,\ldots,i_n)=c_{i_i\ldots i_n}$, such that there exists a (possibly empty)
finite set $I$ satisfying $(i_i,\ldots,i_n)\notin I\Rightarrow c_{i_i\ldots i_n}=q$, where $q$ is a distinguished \emph{quiescent} state of $\Sigma$.
The set of all finite configurations over $\Sigma$ will be denoted $\mathcal{C}^{\Sigma}_f$.
\end{Def}
Since this work relates to QCA rather than CA, the global state of a QCA can be a superposition of these configurations.
To construct the separable Hilbert space of superpositions of configurations the set of configurations must be countable.
This is why finite, unbounded, configurations are considered.
\vspace{-2mm}\begin{Def}\textbf{(Superpositions of configurations)}\label{superp}
Let $\mathcal{H}_{\mathcal{C}^{\Sigma}_f}$ be the Hilbert space of configurations. Each finite configuration $c$ is associated with a unit vector $\ket{c}$, such that the family $\pa{\ket{c}}_{c\in\mathcal{C}^{\Sigma}_f}$ is an orthonormal basis of $\mathcal{H}_{\mathcal{C}^{\Sigma}_f}$. A \emph{superposition of
configurations} is then a unit vector in $\mathcal{H}_{\mathcal{C}^{\Sigma}_f}$.
\end{Def}
Detailed explanation of these definitions, as well as axiomatic definitions of QCA, are available \cite{ArrighiUCAUSAL,ArrighiLATA,SchumacherWerner}. Building upon these works, we have shown \cite{ArrighiNUQCA,ArrighiPQCA} that Partitioned QCA (PQCA) are intrinsically universal. Since they are the most canonical description of QCA, we will, without loss of generality, assume that all QCA are PQCA throughout this work.
\begin{Def}[Partitioned QCA]\label{def:pqca}
A partitioned $n$-dimensional quantum cellular automaton (PQCA) is defined by a scattering unitary operator $U$ such that $U:\mathcal{H}_{\Sigma}^{\otimes 2^n}\longrightarrow\mathcal{H}_{\Sigma}^{\otimes 2^n}$, and $U\ket{qq\ldots qq}=\ket{qq\ldots qq}$, \emph{i.e.\ }{} that takes
a hypercube of $2^n$ cells into a hypercube of $2^n$ cells and preserve quiescence. Consider $G=(\bigotimes_{2\mathbb{Z}^n} U)$, the operator over $\mathcal{H}$. The induced global evolution is $G$ at odd time steps, and $\sigma G$ at even time steps, where $\sigma$ is a translation by one in all directions, see Fig. \ref{fig:structure}\VS{ \emph{(left.)}}.
\end{Def}
\VL{
\begin{figure}[h]
\centering
\includegraphics[scale=.9, clip=true, trim=0cm 0cm 0cm 0cm]{img/structure.pdf}
\caption{\label{fig:structure} One-dimensional PQCA with scattering unitary $U$. Each line represents a quantum system, in this case a whole cell. Each square represents a scattering unitary $U$ which is applied to two cells. Time flows upwards.}
\end{figure}
}
\subsection{Intrinsic Simulation of $n$-Dimensional PQCA} \label{subsecsim}
\VL{Intrinsic simulation of one CA by another was discussed informally in section \ref{subsec:CA}. A
pedagogical discussion in the classical case is given in ref. \cite{OllingerJAC}.
Quantised intrinsic simulation was formalised in the one-dimensional case in ref. \cite{ArrighiUQCA}.}
The definition \VS{of intrinsic simulation} was extended to the quantum case in refs. \cite{ArrighiNUQCA,ArrighiPQCA}, with further discussion.
Here we simply recall these definitions.
\begin{Def}[Isometric coding]\label{isomcode}
Consider $\Sigma_G$ and $\Sigma_H$, two alphabets with distinguished quiescent states $q_G$ and $q_H$, and such that $|\Sigma_H|\leq|\Sigma_G|$. Consider $\mathcal{H}_{\Sigma_G}$ and $\mathcal{H}_{\Sigma_H}$ the Hilbert spaces having these alphabets as their basis, and $\mathcal{H}_{\mathcal{C}_f^{G}}$, $\mathcal{H}_{\mathcal{C}_f^{H}}$ the Hilbert spaces of finite configurations over these alphabets.\\
Let $E$ be an isometric linear map from $\mathcal{H}_{\Sigma_H}$ to $\mathcal{H}_{\Sigma_G}$ which preserves quiescence, \emph{i.e.\ }{} such that $E\ket{q_H}=\ket{q_G}$. It trivially extends into an isometric linear map $Enc=(\bigotimes_{\mathbb{Z}^n} E)$ from $\mathcal{H}_{\mathcal{C}_f^{H}}$ into $\mathcal{H}_{\mathcal{C}_f^{G}}$, which we call an isometric encoding.\\
Let $D$ be an isometric linear map from $\mathcal{H}_{\Sigma_G}$ to $\mathcal{H}_{\Sigma_H}\otimes\mathcal{H}_{\Sigma_G}$ which also preserves quiescence, in the sense that $D\ket{q_G}=\ket{q_H}\otimes\ket{q_G}$. It trivially extends into an isometric linear map $Dec=(\bigotimes_{\mathbb{Z}^n} D)$ from $\mathcal{H}_{\mathcal{C}_f^{G}}$ into $\mathcal{H}_{\mathcal{C}_f^{H}}\otimes\mathcal{H}_{\mathcal{C}_f^{G}}$, which we call an isometric decoding.\\
The isometries $E$ and $D$ define an isometric coding if the following condition is satisfied:\\
$\forall \ket{\psi}\in \mathcal{H}_{\mathcal{C}_f^{H}},\,\exists \ket{\phi}\in \mathcal{H}_{\mathcal{C}_f^{G}}\quad/\quad\ket{\psi}\otimes\ket{\phi}=Dec\pa{Enc \ket{\psi}}.$
\end{Def}
(Here $Dec$ is understood to morally be an inverse function of $Enc$, but some garbage $\ket{\phi}$ may be omitted.)
\begin{Def}[Direct simulation]\label{directsim}
Consider $\Sigma_G$ and $\Sigma_H$, two alphabets with distinguished quiescent states $q_G$ and $q_H$, and two QCA $G$ and $H$ over these alphabets. We say that $G$ directly simulates $H$, if and only if there exists an isometric coding such that\\
$\forall i\in\mathbb{N},\,\forall \ket{\psi}\in \mathcal{H}_{\mathcal{C}_f^{H}},\,\exists \ket{\phi}\in \mathcal{H}_{\mathcal{C}_f^{G}}\quad/\quad (G^i\ket{\psi})\otimes\ket{\phi}=Dec \pa{{H}^i\pa{Enc \ket{\psi}}}.$
\end{Def}
\begin{Def}[Grouping]\label{def:packmap}
Let $G$ be an $n$-dimensional QCA over alphabet $\Sigma$. Let $s$ and $t$ be two integers, $q'$ a word in $\Sigma'=\Sigma^{s^n}$. Consider the iterate global evolution $G^t$ up to a grouping of each hypercube of $s^n$ adjacent cells into one supercell. If this operator can be considered to be a QCA $G'$ over $\Sigma'$ with quiescent symbol $q'$, then we say that $G'$ is an $(s,t,q')$-grouping of $G$.
\end{Def}
\begin{Def}[Intrinsic simulation]\label{def:intsim}
Consider $\Sigma_G$ and $\Sigma_H$, two alphabets with distinguished quiescent states $q_G$ and $q_H$, and two QCA $G$ and $H$ over these alphabets. We say that $G$ intrinsically simulates $H$ if and only if there exists $G'$ some grouping of $G$ and $H'$ some grouping of $H$ such that $G'$ directly simulates $H'$.
\end{Def}
In other words, $G$ intrinsically simulates $H$ if and only if there exists some isometry $E$ which translates supercells of $H$ into supercells of $G$, such that if $G$ is iterated and then translated back, the whole process is equivalent an iteration of $H$, as in Fig. \VL{\ref{UsimV}}\VS{\ref{fig:structure} \emph{(right)}}.
\VL{
\begin{figure}[h]
\centering
\includegraphics[scale=0.75, clip=true, trim=0cm 0cm 0cm 0cm]{img/UsimV.pdf}
\caption{Intrinsic simulation of one PQCA by another.\label{UsimV}
The PQCA defined by $U$ simulates the PQCA defined by $V$. In this case, two cells of the $U$-defined PQCA are required to encode one cell of the $V$-defined PQCA, and the $U$-defined PQCA executes four time steps to simulate one time step of the $V$-defined PQCA. }
\end{figure}
}
\VS{
\begin{figure}[h]
\centering
\VS{\vspace{-10mm}}\includegraphics[scale=0.8, clip=true, trim=0cm 0cm 0cm 0cm]{img/structureandUsimV.pdf}
\caption{\label{fig:structure} \label{UsimV} \emph{Left:} Partitioned one-dimensional PQCA with scattering unitary $U$. Each line represents a quantum system, in this case a whole cell. Each square represents a scattering unitary $U$ which is applied to two cells. Time flows upwards. \emph{Right:} The PQCA defined by $U$ simulates the PQCA defined by $V$. In this case two cells of the $U$-defined PQCA encode one cell of the $V$-defined PQCA, and the $U$-defined PQCA is run for four time steps to simulate one time step of the $V$-defined PQCA. \VL{More generally the challenge is to define an initial configuration of the $U$-defined PQCA so that it behaves just as the $V$-defined PQCA with respect to the encoded initial configuration, after some fixed number of time steps. Clearly such an encoding must hold the configuration of the $V$-defined PQCA as well as a way of describing the scattering unitary $V$.}}
\end{figure}
}
}
\section{An Intrinsically Universal QCA}\label{sec:nuqca}
\VD{In section \ref{subsecdef} the formal definition of $n$-dimensional PQCA was discussed (Fig. \ref{fig:structure}), and the formal definition of intrinsic simulation was recalled (Fig. \ref{UsimV}). }
The aim \VD{now }is to find a particular $U$-defined PQCA which is capable of intrinsically simulating any $V$-defined PQCA, for any $V$.
In order to describe such a $U$-defined PQCA in detail, two things must be given:
the dimensionality of the cells (including the meaning attached to each of the states they may take), and the way the scattering unitary $U$ acts upon these cells.
\VL{First we discuss the general scheme used to solve this problem, and then describe the PQCA implementing it.} \VND{The necessary definitions for $n$-dimensional QCA are given in refs. \cite{ArrighiNUQCA,ArrighiPQCA}.
\VS{\medskip}
\subsection{Circuit Universality versus Intrinsic Universality in Higher Dimensions}
\VS{\indent} As already discussed, intrinsic universality refers to the ability for one CA to simulate any other CA\VS{,} \VL{in a way which preserves the spatial structure of the simulated CA.
Conversely, computation universality refers to the ability of a CA to simulate any TM, and hence run any algorithm.}\VS{whereas computation universality is about simulating a TM.} Additionally, circuit universality is the ability of one CA to simulate any circuit. \VL{These are \textsc{Nand} gate circuits for classical circuits and CA, and \textsc{Toffoli} gate circuits for reversible circuits and CA.} Informally, in a quantum setting, circuit universality is the ability of a PQCA to simulate \VL{any unitary evolution expressed as a}\VS{any finitary} combination of a universal set of quantum gates, such as the standard gate set: \textsc{Cnot, R($\frac{\pi}{4})${}} (also known as the $\frac{\pi}{8}$ gate), and the \textsc{Hadamard} gate.
\VL{
The relationships between these three concepts of CA universality have been noted previously \cite{DurandRoka}.
A computation universal CA is also a circuit universal CA, because circuits are finitary computations.
Moreover, an intrinsic universal CA is also a computation universal CA, because it can simulate any CA,
including computation universal CA. Hence intrinsic universality implies computation universality, which implies circuit universality.
In one-dimension this is not an equivalence. Intuitively, computation universality requires more than circuit universality, namely the ability to loop the computation, which is not trivial for CA. Similarly, intrinsic universality requires more than computation universality, such as the ability to simulate multiple communicating TM. In the classical setting there are formal results that distinguish these ideas \cite{OllingerJAC}.}
In $n$-dimensions, it is often assumed in the classical CA literature that circuit universality implies intrinsic universality, and \VL{hence these are all equivalent}\VS{that both are equivalent to computation universality} \cite{OllingerJAC}\VS{, without provision of an explicit construction}. Strictly speaking this is not true. Consider a two-dimensional CA which runs one-dimensional CA in parallel. If the one-dimensional CA is circuit/computation universal, but not computation/intrinsically universal, then this is also true for the two-dimensional CA. Similarly, in the PQCA setting, the two-dimensional constructions in \cite{PerezCheung} and \cite{Raussendorf} are circuit universal but not intrinsically universal.
However, this remains a useful intuition: Indeed, CA admit a block representation, where these blocks are permutations for reversible CA, while for PQCA the blocks are unitary matrices. Thus the evolution of any (reversible/quantum) CA can be expressed as an infinite (reversible/quantum) circuit of (reversible/ quantum) gates repeating across space. If a CA is circuit universal, and if it is possible to wire together different circuit components in different regions of space, then the CA can simulate the block representation of any CA, and hence can simulate any CA in a way which preserves its spatial structure. It is intrinsically universal.\VS{\medskip}
\VL{
This is the route followed next in constructing the intrinsically universal $n$-dimensional PQCA. First the construction of the `wires',
which can carry information across different regions of space, is considered. Here these are signals which can be redirected or delayed using barriers, with each signal holding a qubit of information. Secondly, the `circuit-pieces' are constructed, by implementing quantum gates which can be combined. One and two qubit gates are implemented as obstacles to, and interactions of, these signals.}
\subsection{Flattening a PQCA into Space}\label{subsec:flat}
\VL{In the classical CA literature it is considered enough to show that the CA implements some wires carrying signals, and some universal gates
acting upon them, to prove that an $n$-dimensional CA is in fact intrinsically universal. }
\VS{\indent}Any CA can be encoded into a `wire and gates' arrangement following the above argument, but this has never been made explicit in the literature.
This section makes more precise how to flatten any PQCA in space, so that it is simulated by a PQCA which implements quantum
wires and universal quantum gates. Flattening a PQCA means that the infinitely repeating, two-layered circuit is arranged in space so that at the beginning all the signals carrying qubits find themselves in circuit-pieces which implement a scattering unitary of the first layer, and then all synchronously exit and travel to circuit-pieces implementing the scattering unitary of the second layer, etc.
An algorithm for performing this flattening can be provided, however the process will not be described in detail,
for clarity and following the classical literature, which largely ignores this process.
The flattening process can be expressed in three steps:
Firstly, the $V$-defined PQCA is expanded in space by coding each cell into a hypercube of $2^n$ cells. This allows enough space for the scattering unitary $V$ to be applied on non-overlapping hypercubes of cells, illustrated in the two-dimensional case in Fig.~\ref{fig:flattening12}.
\begin{figure}
\centering
\VS{\includegraphics[scale=.85, clip=true, trim=0cm 0cm 0cm 0cm]{img/flattening1and2.pdf}}
\VL{\includegraphics[scale=.9, clip=true, trim=0cm 0cm 0cm 0cm]{img/flattening1and2.pdf}}
\caption{Flattening a PQCA into a simulating PQCA. \emph{Left}: Consider four cells (white, light grey, dark grey, black) of a PQCA having scattering unitary $V$.
The first layer PQCA applies $V$ to these four cells, then the second layer applies $V$ at the four corners. \emph{Right}: We need to flatten this so that the two-layers become non-overlapping. The first layer corresponds to the centre square, and the second layer to the four corner squares. At the beginning the signals (white, light grey, dark grey, black) coding for the simulated cells are in the centre square. \VL{They undergo $V$, and are directed towards the bottom left, top left, bottom right, and top right squares respectively, where they undergo $V$ but paired up with some other signals, etc.}
\label{fig:flattening12}}
\end{figure}
\noindent Secondly, the hypercubes where $V$ is applied must be connected with wires, as shown in Fig. \ref{fig:flattening12} $(right)$. Within these hypercubes wiring is required so that incoming signals are bunched together to undergo a circuit implementation of $V$, and are then dispatched appropriately, as shown in Fig. \ref{fig:flattening34} $(left)$. This requires both time and space expansions, with factors that depend non-trivially (but uninterestingly) upon the size of the circuit implementation of $V$ and the way the wiring and gates work in the simulating PQCA.
\begin{figure}
\centering
\VS{\includegraphics[scale=.8, clip=true, trim=0cm 0cm 0cm 0cm]{img/flattening3and4.pdf}}
\VL{\includegraphics[scale=.9, clip=true, trim=0cm 0cm 0cm 0cm]{img/flattening3and4.pdf}}
\caption{Flattening a PQCA into a simulating PQCA (cont'd). \emph{Left}: Within the central square the incoming signals are bunched together so as to undergo a circuit which implements $V$, and are then dispatched towards the four corners. This diagram does not make explicit a number of signal delays, which may be needed to ensure that they arrive synchronously at the beginning of the circuit implementing $V$. \emph{Right}: Within the central rectangle, the circuit which implements $V$ is itself a combination of smaller circuits for implementing a universal set of quantum gates such as \textsc{Cnot}, \textsc{Hadamard} and the \textsc{R($\frac{\pi}{4})${}}, together with delays. \VL{These are implemented as explained in sections \ref{subsec:onequbit} and \ref{subsec:gates}.}\label{fig:flattening34}}
\end{figure}
\noindent Next, an encoding of the circuit description of the scattering unitary $V$ is implemented in the simulating PQCA upon these incoming bunched wires, as shown in Fig. \ref{fig:flattening34} $(right)$. This completes the description of the overall scheme according to which a PQCA that is capable of implementing wires and gates is also capable of intrinsically simulating any PQCA, and hence any QCA. A particular PQCA that supports these wires and gates can now be constructed.\VS{\newpage}
\subsection{Barriers and Signals Carrying Qubits}\label{subsec:onequbit}
\VS{\indent} Classical CA studies often refer to `signals' without an explicit definition. In this context,
a signal refers to the state of a cell which may move to a neighbouring cell consistently, from one step to another, by the evolution of the CA.
Therefore a signal would appear as a line in the space-time diagram of the CA. These lines need to be implemented as signal redirections.
A $2$D solution is presented here, but this scheme can easily be extended to higher dimensions. Each cell has four possible basis states:
empty ($\epsilon$), holding a qubit signal ($0$ or $1$), or a barrier ($\blacksquare$). The scattering unitary $U$ of the universal PQCA acts on $2\times 2$ cell neighbourhoods.
Signals encode qubits which can travel diagonally across the 2D space (NE, SE, SW, or NW).
Barriers do not move, while signals move in the obvious way if unobstructed, as there is only one choice for any signal in any square of four cells.
Hence the basic movements of signals are given by the following four rules:
$$\cells{}{}{$s$}{} \mapsto \cells{}{$s$}{}{}, \qquad \cells{$s$}{}{}{} \mapsto \cells{}{}{}{$s$},$$
$$\cells{}{$s$}{}{} \mapsto \cells{}{}{$s$}{}, \qquad \cells{}{}{}{$s$} \mapsto \cells{$s$}{}{}{}.$$
where $s\in \{0,1\}$ denotes a signal, and blank cells are empty.
The way to interpret the four above rules in terms of the scattering unitary $U$ is just case-by-case definition, \emph{i.e.\ }{} they show that
$U\cells{}{}{$s$}{}=\cells{}{$s$}{}{}$.
Moreover, each rule can be obtained as a rotation of another, hence by stating that the $U$-defined PQCA is isotropic the first rule above suffices. This convention will be used throughout.
The ability to redirect signals is achieved by `bouncing' them off walls constructed from
two barriers arranged either horizontally or vertically:
$$
\cells{\cellcolor[gray]{0.6}}{$s$}{\cellcolor[gray]{0.6}}{} \mapsto \cells{\cellcolor[gray]{0.6}}{}{\cellcolor[gray]{0.6}}{$s$}.
$$
where $s$ again denotes the signal and the shaded cells denote the barriers which causes the signal
to change direction.
If there is only one barrier present in the four cell square being operated on then the signal simply propagates as normal and is not deflected:
$$\cells{\cellcolor[gray]{0.6}}{}{$s$}{} \mapsto \cells{\cellcolor[gray]{0.6}}{$s$}{}{}.$$
Using only these basic rules of signal propagation and signal reflection from barrier walls, signal delay (Fig. \ref{fig:delays}) and signal swapping (Fig. \ref{fig:swap})
tiles can be constructed. All of the rules presented so far are permutations of some of the base elements of the vector space generated by
$$\Set{\cells{$w$}{$x$}{$y$}{$z$}}_{w,x,y,z \in \{\epsilon,0,1,\blacksquare\}}$$
therefore $U$ is indeed unitary on the subspace upon which its action has so far been described.
\begin{figure}
\centering
\VS{\includegraphics[scale=.55, clip=true]{img/delayCirc.pdf}}
\VL{\includegraphics[scale=.60, clip=true]{img/delayCirc.pdf}}
\caption{The `identity circuit' tile, an $8\times 14$ tile taking 24 time-steps, made by repeatedly bouncing the signal
from walls to slow its movement through the tile. The dotted line gives the signal trajectory, with the arrow showing the
exit point and direction of signal propagation. The bold lines show the tile boundary.}
\label{fig:delays}
\end{figure}
\begin{figure}
\centering
\VS{\includegraphics[scale=.55, clip=true]{img/swapCirc.pdf}}
\VL{\includegraphics[scale=.60, clip=true]{img/swapCirc.pdf}}
\caption{The `swap circuit' tile, a $16\times 14$ tile, where both input signals are permuted and exit synchronously after 24 time-steps.
As the first signal (\emph{bottom left}) is initially delayed, there is no interaction.}
\label{fig:swap}
\end{figure}
\subsection{Gates}\label{subsec:gates}
\VS{\indent} To allow a universal set of gates to be implemented by the PQCA, certain combinations
of signals and barriers can be assigned special importance.
The Hadamard operation on a single qubit-carrying signal can be implemented by interpreting a
signal passing through a diagonally oriented wall, analogous to a semitransparent barrier in physics. This has the action
defined by the following rule:
$$\cells{\cellcolor[gray]{0.6}}{}{0}{\cellcolor[gray]{0.6}} \mapsto\frac{1}{\sqrt{2}}\cells{\cellcolor[gray]{0.6}}{0}{}{\cellcolor[gray]{0.6}} + \frac{1}{\sqrt{2}}\cells{\cellcolor[gray]{0.6}}{1}{}{\cellcolor[gray]{0.6}}$$
$$\cells{\cellcolor[gray]{0.6}}{}{1}{\cellcolor[gray]{0.6}} \mapsto \frac{1}{\sqrt{2}}\cells{\cellcolor[gray]{0.6}}{0}{}{\cellcolor[gray]{0.6}} - \frac{1}{\sqrt{2}}\cells{\cellcolor[gray]{0.6}}{1}{}{\cellcolor[gray]{0.6}} $$
This implements the Hadamard operation, creating a superposition of configurations with appropriate phases. Using this construction
a Hadamard tile can be constructed (Fig. \ref{fig:hadamard}) by simply adding a semitransparent barrier to the end of the previously
defined delay (identity) tile (Fig. \ref{fig:delays}).
\begin{figure}
\centering
\VS{\vspace{-0.5mm}\includegraphics[scale=.55, clip=true]{img/hadCirc.pdf}}
\VL{\includegraphics[scale=.60, clip=true]{img/hadCirc.pdf}}
\caption{The `Hadamard gate' tile applies the Hadamard operation to the input signal. It is a modification
of the identity circuit tile, with a diagonal (semitransparent) barrier added at the end which performs the Hadamard operation.}
\label{fig:hadamard}
\end{figure}
A way of encoding two qubit gates in this system is to consider that two signals which
cross paths interact with one another. The controlled-R($\frac{\pi}{4})${}{ }
operation can be implemented by considering signals that cross each other as interacting only if they are both $1$, in which case a global phase
of $e^{\frac{i\pi}{4}}$ is applied. Otherwise the signals continue as normal. This behaviour is defined by the following rule:
$$\cells{1}{}{1}{} \mapsto e^{\frac{i\pi}{4}}\cells{}{1}{}{1}, \qquad \cells{$x$}{}{$y$}{} \mapsto \cells{}{$y$}{}{$x$} otherwise$$
where $x,y \in \{0,1\}$. This signal interaction which induces a global phase change allows the definition of both a two signal
controlled-R($\frac{\pi}{4})${}{} tile (Fig. \ref{fig:cphase}) and a single signal R($\frac{\pi}{4})${}{} operation tile (Fig. \ref{fig:phase}).
\begin{figure}
\centering
\VS{\vspace{-0.5mm}\includegraphics[scale=.55, clip=true]{img/cPhaseCirc.pdf}}
\VL{\includegraphics[scale=.60, clip=true]{img/cPhaseCirc.pdf}}
\caption{The `controlled-\Phase{}{} gate' tile\VS{, with a signal interaction at the highlighted cell.}\VL{ applies the controlled-R($\frac{\pi}{4})${}{} operation to the two input qubits, by causing the signals to interact at the highlighted point (grey circle). The qubits are then synchronised so that they exit at the same time along their original paths. No swapping takes place.}}
\label{fig:cphase}
\end{figure}
\begin{figure}
\centering
\VS{\vspace{-0.5mm}\includegraphics[scale=.55, clip=true]{img/phaseCirc.pdf}}
\VL{\includegraphics[scale=.60, clip=true]{img/phaseCirc.pdf}}
\caption{The `R($\frac{\pi}{4})${}{} gate' tile. This tile makes use
of a signal, set to $\ket{1}$, which loops inside the grid every six time-steps, ensuring that it will interact with the signal that enters the
tile, and causing it to act as the control qubit to a controlled-\Phase{}{} operation. It therefore acts as a phase rotation on the input qubit, which passes directly through. \VL{After 24 time-steps the auxiliary control signal has returned to its origin, unchanged, hence the tile can be reused.}}
\label{fig:phase}
\end{figure}
These rules are simply a permutation and phase change of base elements of the form:
$$\Set{\cells{$x$}{}{$y$}{}}_{x,y \in \{0,1\}}$$
(and their rotations), therefore $U$ is a unitary operation on the subspace upon which its action has so far been described.
Wherever $U$ has not yet been defined, it is the identity. Hence $U$ is unitary.\VS{\medskip}
\subsection{Circuits: Combining Gates}\label{subsec:circuits}
\VS{\indent} A signal is given an $8 \times 14$ tile ($16 \times 14$ for two signal operations)
in which the action is encoded. The signals enter each tile at the fifth cell from the left, and propagate diagonally NE.
Each time step finds the tile shifted one cell to the right to match this diagonal movement, giving a diagonal tile.
The signal exits the tile $14$ cells North and East of where it entered. This allows
these tiles to be composed in parallel and sequentially with the only other requirement being that the signal exits at the appropriate point, \emph{i.e.\ }{}
the fifth cell along the tile, after $24$ time-steps. This ensures that all signals are synchronised as in Fig. \ref{fig:flattening34} (\emph{right}),
allowing larger circuits to be built from these elementary tiles by simply plugging them together. Non-contiguous gates can also be wired together
using appropriate wall constructions to redirect and delay signals so that they are correctly synchronised.
The implemented set of quantum gates, the identity, Hadamard, swap, R($\frac{\pi}{4})${}{} and controlled-\Phase{}{}, gives a universal set. Indeed the standard set
of \textsc{cNot}, \textsc{H}, \textsc{R($\frac{\pi}{4})${}} can be recovered as follows:
$$\textsc{cNot}\ket{\psi}=(\mathbb{I}\otimes H)(\textsc{cR(${\pi}\slash{4}$)})^4(\mathbb{I}\otimes H)\ket{\psi}$$
where $\textsc{cR($\frac{\pi}{4}$})^4$ denotes four applications of the controlled-\Phase{}{} gate, giving the controlled-\textsc{Phase} operation.
\VS{\vspace{-0.5mm}}\section{Conclusion}\label{sec:discussion}
This paper presents a simple PQCA which is capable of simulating all other PQCA, preserving the topology
of the simulated PQCA. This means that the initial configuration and the forward evolution of any PQCA can be encoded
within the initial configuration of this PQCA, with each simulated cell encoded as a group of adjacent cells in the PQCA,
\emph{i.e.\ }{} intrinsic simulation.
The construction in section \ref{sec:nuqca} is given in two-dimensions, which can be seen to generalise to $n>1$-dimensions.
The main, formal result of this work can therefore be stated as:
\begin{Cl}
There exists an $n$-dimensional $U$-defined PQCA, $G$, which is an intrinsically universal PQCA. Let $H$ be a $n$-dimensional $V$-defined PQCA such that $V$ can be expressed as a quantum circuit $C$ made of gates from the set $\textsc{Hadamard}$, $\textsc{Cnot}$, and $\textsc{R($\frac{\pi}{4})${}}$. Then $G$ is able to intrinsically simulate $H$.
\end{Cl}
Any finite-dimensional unitary $V$ can always be approximated by a circuit $C(V)$ with an arbitrary small error $\varepsilon=\max_{\ket{\psi}}||V\ket{\psi}-C\ket{\psi}||$. Assuming instead that $G$ simulates the $C(V)$-defined PQCA, for a region of $s$ cells over a period $t$, the error with respect to the $V$-defined PQCA will be bounded by $st\varepsilon$. This is due to the general statement that errors in quantum circuits increase, at most, proportionally with time and space \cite{NielsenChuang}.
Combined with the fact that PQCA are universal \cite{ArrighiNUQCA,ArrighiPQCA}, this means that $G$ is intrinsically universal, up to this unavoidable approximation.\VS{\medskip}
\subsection{Discussion and Future Work}
\indent QC research has so far focused on applications for more secure and efficient computing, with theoretical physics supporting this work in theoretical computer science. The results of this interdisciplinary exchange led to the assumptions underlying computer science being revisited,
with information theory and complexity theory, for example, being reconsidered and redeveloped.
However, information theory also plays a crucial role in the foundations of theoretical physics\VS{ (\emph{e.g.\ }{} deepening our understanding of entanglement \cite{DurVidalCirac} and decoherence \cite{PazZurek})}.
These developments are also of interest in theoretical physics studies where physical aspects such as particles and matter are considered; computer science studies tend to consider only abstract mathematical quantities.
Universality, among the many computer science concepts, is a simplifying methodology in this respect. For example, if the problem being studied crucially involves some idea of interaction, universality makes it possible cast it in terms of information exchanges \emph{together} with some universal information processing.
This paper presents an attempt to export universality as a tool for application in theoretical physics, a small step towards the goal of finding and understanding a \emph{universal physical phenomenon}, within some simplified mechanics.
Similar to the importance of the idea of the spatial arrangement of interactions in physics, intrinsic universality has broader applicability than computation universality and must be preferred. In short, if only one physical phenomenon is considered, it should be an intrinsically universal physical phenomenon, as it could be used to simulate all others.
The PQCA cell dimension of the simple intrinsically universal construction given here is four (empty, a qubit ($\ket{0}$ or $\ket{1}$), or a barrier). In comparison,
the simplest classical Partitioned CA has cell dimension two \cite {MargolusQCA}.
Hence, although the intrinsically universal PQCA presented here is the simplest found, it is not minimal. In fact, one can also manage \cite{Arrigh2UQCA} an intrinsically universal PQCA with a cell dimension of three, in two different ways. One way is to encode the spin degree of freedom (0 and 1) into a spatial degree of freedom, so that now the semitransparent barrier either splits or combines signals.
The second way is to code barriers as pairs of signals as in the Billiard Ball CA model \cite{MargolusQCA}.
These constructions may be minimal, but are not as elegant as the one presented here.
In future work we will show that there is a simple, greater than two-dimensional PQCA which is minimal, as it has a cell dimension of two.
\section*{Acknowledgements}
The authors would like to thank J\'er\^ome Durand-Lose, Jarkko Kari, Jacques Mazoyer, Kenichi Morita, Nicolas Ollinger, Guillaume Theyssier and Philippe Jorrand.
|
\section{Introduction}
\label{sc:introduction}
Particle production in thermal and non-thermal systems plays an
important role in relativistic heavy ion physics and in cosmology. Typically
there is some weakly interacting particle species with a low abundance in a
hot medium, from which these particles are then produced. Examples are photons
produced in a quark--gluon plasma, or dark matter candidates such as
gravitinos, axions, or axinos produced in an otherwise thermal universe and at
reheating after inflation.
The production of photons from a thermal quark-gluon plasma is remarkably
complicated. At {\em leading logarithmic order}, that is, at leading order not just
in the strong coupling $ \alpha _ { \rm S } $, but rather in $ \alpha _ { \rm S }
\log ( 1/\alpha _ { \rm S } ) $ only $ 2 \to 2 $ scattering processes
contribute and the rate has been computed by Kapusta {\em et al.}
\cite{kapusta} and by Baier {\em et al.} \cite{baier}.
However, already for the complete leading order contribution
the production cannot simply be understood in terms of scattering
processes involving only a handful of particles. Bremsstrahlung and pair
annihilation involve multiple interactions via soft gluon exchange.
Such processes are not suppressed, despite the large number
of interactions \cite{aurenche,arnoldPhoton}. The emission occurs almost collinearly so that
internal lines are nearly on-shell and compensate the suppression. Viewed in
position space the radiated photon and its source overlap over large distances, and the
interference of different interactions
cannot be neglected. This leads to the so-called Landau-Pomeranchuk-Migdal
(LPM) effect \cite{arnoldPhoton,landau,migdal,AurencheLPM}.
The complete leading order
photon production rate has been computed at leading order by Arnold, Moore
and Yaffe by summing all relevant diagrams.
In this paper we present a different approach. We consider hard
particles with momenta of order $ T $ which propagate through a
gauge field background. The gauge field momenta are either soft ($ k \sim g T $)
or are
almost on-shell with virtuality $ k ^ 2 \sim g ^ 2 T ^ 2 $ and almost collinear with
the loop momentum.
Formally we integrate out the hard modes
in this soft and collinear background, leaving us with an effective theory for
soft and collinear
modes. The effective theory is described by an equation which has a similar
structure as the non-abelian Vlasov equations \cite{blaizot93} describing Hard
Thermal Loops \cite{htl}.
Then we distinguish the soft and collinear fields and identify
them with the soft gluon field and the photon field, respectively.
In a second step we integrate out
the soft gluons to
obtain an effective theory for the
collinear gauge fields only. \footnote{This approach is similar to the one
used in
Ref.~\cite{effective}, where soft gauge
fields were integrated out to obtain an
effective theory for ultrasoft ($ k \sim g ^ 2 T $) fields.}
The effective theory takes the form of an integral equation
which has been obtained previously in
the calculation of the photon production rate \cite{arnoldPhoton}.
This paper is organized as follows. In Sec. \ref{sc:rate} we briefly recall the
relation between thermal photon production rate and the finite temperature
polarization tensor. The main part of the paper is contained in Sec.
\ref{sc:diagrams} with the calculation of the
1-loop diagrams in a soft or collinear gauge field background. The key steps
of our calculation are the approximation (\ref{propagator}) and the partial fractioning
(\ref{fractioning}). These are used both to compute the 2-point function
(Sec.~\ref{sc:2pointfunction})
and then to obtain the recursion relation for the $ n $-point functions in
terms of $ (n -1) $-point functions (Sec.~\ref{sc:npoint}). Then all $ n $-point functions are put
together in an effective action (Sec. \ref{sc:current}). In Sec.~\ref{sc:integrate}
we integrate out the gluon fields and in Sec.~\ref{sc:fermionic} we
discuss the generalization from
scalar quarks to spin-1/2 quarks. We summarize and conclude in
Sec.~\ref{sc:conclusions}. Finally, in
Appendix \ref{ap:connected} we show that the connected pieces which we
encounter in Sec.~\ref{sc:integrate} vanish.
\vspace{3mm
\noindent{\bf Notation and conventions:} We use the metric with
signature~$ + - - - $.
\section{The thermal photon production rate}
\label{sc:rate}
The rate per unit volume at which real photons with momentum $ \vec k $ and polarization vector $
\varepsilon _ \mu $ are produced in a hot QCD gas can be written as
\begin{equation} \label{prod_rate}
( 2 \pi ) ^ 3 2 k ^ 0 \frac{ d \Gamma }{d ^ 3 k} =
e ^ 2 \varepsilon _ \mu
\varepsilon _ \nu ^ * \Pi ^ { \mu \nu } _ < ( k )
\end{equation}
where $ k ^ 0 = |\vec k | $ is the photon energy and
\begin{align}
\Pi ^ { \mu \nu } _ < ( k ) =
\int d ^ 4 x e ^{ i k \cdot x } \langle J ^ \mu ( 0 ) J ^ \nu ( x )
\rangle
\label{2point}
\end{align}
is a 2-point function of electromagnetic current operators $ J ^ \mu $.
We have not included the electromagnetic coupling constant in $ J^\mu $, which
therefore appears in Eq.~(\ref{prod_rate}). The 2-point function (\ref{2point})
can be written in terms of the retarded polarization tensor
\begin{align}
\Pi ^ { \mu \nu } _ < ( k ) = 2 f _ { \rm B } ( k ^ 0 ) {\rm Im } \Pi
^ { \mu \nu } _ { \rm ret } ( k )
\label{retarded}
\end{align}
where $ f _ { \rm B } ( E ) = 1/( e ^{ E/T } -1 ) $ is the Bose-Einstein
distribution at temperature $ T $.
Among the leading order contributions to the production rate (\ref{prod_rate})
are QCD Compton and $ q \bar q $
annihilation. These are $ 2 \to 2 $ scattering
processes for which only a few diagrams contribute.
We will only be
concerned with bremsstrahlung and pair annihilation
contributions. There a quark undergoes multiple scattering with
soft momentum transfer $ q \sim g T $
and at some point emits a bremsstrahlung photon or annihilates with an
antiquark into a photon.
Therefore we want to evaluate $ \Pi ^ { \mu
\nu } $ with a quark interacting via soft gluons only. For simplicity we first consider
scalar quarks, the generalization of our method to spin-1/2 quarks is
straightforward and is described in Sec.\ref{sc:fermionic}. The
first step we will take is to integrate out the scalar
quark fields in a soft or collinear gauge field background.
That means that we have to calculate diagrams with two external photon lines
and an arbitrary number of
soft external gluon lines.
\section{1-loop diagrams with soft or collinear external gauge fields}
\label{sc:diagrams}
We consider thermal 1-loop diagrams
with an arbitrary number of external
gauge field legs. For the bremsstrahlung and pair annihilation contribution to
photon production, two of the gauge field momenta are hard ($ k \sim T $) and
correspond to the produced photon. The remaining ones are
soft ($ k \sim gT $) gluons. The particle in the loop
corresponds to a quark which suffers soft scattering via gluon exchange and
radiates the photon.
Even though the photons
and the gluons play very different roles, our approach allows for a
unified treatment, and in this section we do not have to distinguish these two.
First we review the relevant kinematics which has been
extensively discussed in Ref.~\cite{arnoldPhoton}. It allows us
to simplify propagators and
vertices. We obtain a
recursion relation between a diagram with $ n $ external gauge field lines
and the difference of two diagrams with $ n -1 $ external lines.
Then we consider the current
induced by the gauge fields which is the first derivative of
the generating functional of all $ n $-point
functions. The recursion relation turns into a generalized kinetic equation.
It can be viewed as a generalized Vlasov equation which contains
a convective term and a force term.
\subsection{Kinematics and power counting}
\label{sc:kinematics}
The hard momenta we are considering are all almost collinear. Up to
higher orders they all point into the same direction which we denote by $ \vec
v $ with $ \vec v ^ 2 = 1 $. The 3-momentum components in the $ \vec v
$-direction are denoted by
\begin{align}
p _ \| \equiv \vec p \cdot \vec v
\end{align}
We define the light-like vector $ v \equiv
( 1, {\vec v } ) $. One has to account for three distinct momentum scales.
\begin{enumerate}
\item The emitting charged particle, which corresponds
to the particle in the loop, and the emitted particle both have $ p _ \| $
of order $ T $, which is our hard scale.
\item All 3-momenta perpendicular to $ \vec v
$ are soft, $ \vec p _\perp \sim g T $. Furthermore, all momentum components
of the gluons are
soft,
$ k _ \mu \sim g T $
\item Finally, all 4-momenta $ k $ are `collinear',
$ v \cdot k = ( k _ 0 - k _ \| ) \sim g ^ 2 T $.
\end{enumerate}
This includes the case that the emitted photon is off-shell by an amount
$ k ^ 2 \sim g ^ 2 T ^ 2 $ which is relevant for dilepton production.
All momenta in the loop have $ k ^ 2 \sim g ^ 2 T^2 $. Therefore the propagators
are sensitive to the so-called
asymptotic mass $ m \sim g T $, \footnote {It is
oftentimes referred to as $ m_\infty $. }
which is given by the real part of the thermal self-energy, with hard loop momentum, of a
light-like hard particle \cite{Flechsig}. It is thus generated by integrating out the gluons
with hard momenta. This procedure does not yet yield a thermal width
which is only generated by integrating out the gluons with soft
momenta. \footnote{Therefore, in contrast to \cite{arnoldPhoton} we
never include the (IR divergent) thermal width in the quark propagators,
only the asymptotic mass.} For scalars and fermions in the representation $ r $ of the
gauge group
\begin{equation}
m^2 = \frac{1 }{4} C_2(r) g^2 T^2
\label{mass}
\end{equation}
with the quadratic Casimir of the representation $ r $. In a
SU($N$) gauge theory $
C_2(r) = ( N ^ 2 -1 ) /( 2 N ) $.
\subsection{2-point function}
\label{sc:2pointfunction}
In this section we explicitly compute the 2-point function. It turns out that
all $n$-point functions with $n > 2$ can be obtained from $ (n -1) $-point
functions through a simple recursion relation.
Furthermore, all kinematic approximations which are needed for
the general case already appear for the 2-point function.
We work in the imaginary time formalism, where the loop integral consists
of a sum over imaginary Matsubara frequencies and an integral over 3-momenta.
The external momenta must also be taken imaginary and can be continued to real
values only after the sum has been performed.
The Matsubara sum is performed as usual by writing it as a contour integral
\begin{align}
T \sum_{ p _ 0 = i n 2 \pi T} h( p _ 0 )
=
\int_C \frac{d p _ 0 }{2\pi i}
\left [ \frac12 + f _ { \rm B } (p _ 0 ) \right ] h(p _ 0 )
\label{matsubara}
\end{align}
where the integration contour goes up on the right of the imaginary axis
and goes down on the left of the imaginary axis. Then we close
the contour around the poles of the propagators.
After evaluating the integral \eqref{matsubara},
we can continue $ k _ 0 $ from a Matsubara frequency towards the real
axis. The retarded polarization tensor is obtained by taking
$ k_ 0 = \mbox{Re} ( k_ 0 ) + i \epsilon $.
Now we describe the two main steps of our calculation. In the imaginary time
formalism the scalar propagator appearing in Fig.~\ref{fg:npoint} is
\begin{align}
\Delta ( p ) \equiv \frac{ -1}{ p ^ 2 - m ^ 2 }
= \frac{ -1}{ \bar v \cdot p \,\, v \cdot p - \vec p _\perp ^ 2 - m ^ 2 }
\label{propagator}
\end{align}
with $ \bar v = (1, -\vec v ) $ .
The other propagator in
Fig.~\ref{fg:npoint}(a) will only be parametrically large, of order $ g ^ {-2 }
T ^{ -2 } $, when $ v \cdot p \sim v \cdot k $
because $ v \cdot k \sim g ^ 2 T $. In this case
$ p _ 0 $ approximately equals $ p _ \| $, and $ \bar v \cdot p \simeq 2 p
_ \| $. Therefore we can approximate
\begin{align}
\Delta ( p ) \simeq
\frac{ 1}{ 2 p _ \|} D ( p )
\label{propapp}
\end{align}
with
\begin{align}
D ( p )
=
\frac{ -1}{ v \cdot p
- ( \vec p _\perp ^ 2 + m ^ 2 ) / ( 2 p _ \| )
}
\label{D}
\end{align}
After this approximation the propagator, viewed as a function of the energy
variable, has only one pole. \footnote { Note that we have made the
approximation (\ref{propapp}) {\em before} the Matsubara summation, that is,
when $ p _ 0 $ is still purely imaginary and of order $ T $, even though
(\ref{propapp}) is valid only for real frequencies. This is justified,
however, because the second pole in $ p _ 0 $ of the propagator, which gets
lost through the approximation (\ref{propapp}), would give $ v \cdot p
\simeq -2 p _ \| \sim T $. With this value of $ v \cdot p $ the second
propagator $ \Delta ( k - p ) $ would be of order $ T ^ {-2 } $. This is
suppressed relative to the contribution we keep, for which we have $ \Delta
( k - p ) \sim g ^{ -2 } T ^ {-2 } $. The same type of argument applies to
the poles of $ \Delta ( k - p ) $. } For the following it turns out to be
very convenient to partial fraction the product of the two propagators,
\begin{align}
D ( p ) D ( p - k )
=
\frac1{ \epsilon ( k, \vec p ) }
\left [ D ( p )
- D ( p - k ) \right ]
\label{fractioning
\end{align}
where
\begin{align}
\epsilon ( k, \vec p )
\equiv
v \cdot k + \frac{ ( \vec p _\perp - \vec k _\perp ) ^ 2 + m ^ 2}
{ 2 ( p _ \| - k _ \| ) }
- \frac{ \vec p _\perp ^ 2 + m ^ 2}
{ 2 p _ \| }
\label{defoflambda}
\end{align}
denotes the difference of the poles of the two propagators which is of order
$ g ^ 2 T $.
Both vertices are proportional to $ ( 2 p - k ) _ \mu $. We
associate the factor $ ( 2 p _ \| ) ^{ -1 } $ of Eq.~(\ref{propapp}) with the left vertex in
Fig.~\ref{fg:npoint}(a) rather than with the propagator itself
and define the resulting vertex factor
\begin{align}
V ^ \mu ( p, p -k ) \equiv \frac{ 1 } { 2 p _ \| } ( 2 p - k ) ^ \mu
\label{vertexfactor}
\end{align}
The different components
of $ V $ have different orders of magnitude. The largest component is in the
direction of $ v $ and is $ O ( 1 ) $.
The transverse components are of order
$ g $, and the $ \bar v $-component is $ O ( g ^ 2 ) $. For
computing the production rate of real photons one needs the transverse
components. It is therefore not sufficient to take into account only the leading
order piece. We also include the transverse components, and the approximation
\begin{align}
V ^ \mu ( p, p -k ) \simeq
\frac{ 1 } { 2 p _ \| } \left [
( 2 p _ \| - k _ \| ) v ^ \mu + ( 2 p - k ) _\perp ^ \mu
\right ]
\label{vertex}
\end{align}
is implicitly understood in the following.
We write the polarization tensor as
\begin{align}
\Pi _ { \mu \nu } ^{ ab } ( k ) =
\int \frac{ d ^ 3 p } { ( 2 \pi ) ^ 3 }
V _ \mu ( p, p -k )
{\rm tr} \left ( t ^ a \widehat \Pi _ { \nu } ^ b ( k ,\vec p) \right )
\label{pimunu}
\end{align}
which turns out to be convenient for
discussing the general $ n $-point function, and define
\begin{equation}
\mathcal{F}(p_\parallel,k_\parallel) \equiv
f _ {\rm B}(p_\parallel)
- f _ {\rm B} (p_\parallel - k_\parallel)
\label{bose}
\end{equation}
Our result for the 2-point function is then
\begin{align}
\widehat \Pi _ { \nu } ^ b ( k , \vec p)
=
\frac{ 1 } { \epsilon ( k, \vec p ) }
\mathcal{F}(p_\parallel,k_\parallel)
V _ \nu ( p - k, p )
t ^ b
\label{54}
\end{align}
As we already stated below Eq.~(\ref{2point}), we do not include the gauge coupling
in $ \Pi _ { \mu \nu }$.
There is also the tadpole diagram containing the 4-point vertex which
contributes to the 1-loop polarization tensor. It does not
have the collinear enhancement since it does not depend on the external
momentum. But its transverse components are of the same order
as those of the diagram in Fig.~\ref{fg:npoint} (a). However,
due to the independence on
the external momentum the tadpole diagram
has no discontinuity and therefore does not
contribute to the production rate. Therefore we do not consider it here.
\subsection{Recursion relation for $n$-point diagrams}
\label{sc:npoint}
\begin{figure}[t]
\centering
\begin{picture}(50,90)(230,20)
\SetWidth{1}
\Vertex(180. ,60. ){2}
\Photon(180,60)(210,60){3}{4}
\DashArrowArc(150,60)(30,0,180){5}
\DashArrowArc(150,60)(30,180,0){5}
\Vertex(120. ,60. ){2}
\Text(80,60)[]{ $k$}
\Photon(120. , 60.)( 90., 60.){3}{4}
\Text(150,105)[]{ $p$}
\Text(150, -40)[]{(a)}
\end{picture}
\begin{picture}(50,90)(30,-80)
\SetWidth{1}
\Text(80,60)[]{ $k$}
\Photon(120. , 60.)( 90., 60.){3}{4}
\Vertex(120. ,60. ){2}
\DashArrowArc(150,60)(30,120,180){5}
\Vertex(138.343 ,87.6427 ){2}
\Text(119.479,126.382)[]{ $-k _ 1 $}
\Photon(129.479,116.382)(139.739,88.1908){3}{4}
\DashArrowArc(150,60)(30,60,120){5}
\Vertex(172.98 , 79.28 ){2}
\Text(203.96 , 107.57 )[]{ $-k _ 2 $}
\Photon(172.98 , 79.28 )(195.96 , 98.57 ){3}{4}
\DashArrowArc(150,60)(30,320,40){5}
\Vertex( 172.98, 40.7){2}
\Text(203.96 , 13.4 )[]{ $-k _ { n -2 } $}
\Photon( 172.98, 40.7)(195.96 , 21.4 ){3}{4}
\DashArrowArc(150,60)(30,240,320){5}
\Vertex( 139.739, 31.8092){2}
\Text(121.479 , -5.61844)[]{ $-k _ { n - 1 } $}
\Photon(139.739 ,31.8092 )(129.479 , 3.61844){3}{4}
\DashArrowArc(150,60)(30,180,240){5}
\Vertex(191.575 , 77.2208){1}
\Vertex( 195., 60.){1}
\Vertex( 191.575, 42.7792){1}
\Text(110,85)[]{ $p$}
\Text(150, -50)[]{(b)}
\end{picture}
\caption{1-loop diagrams with soft or collinear external gauge field lines.
Only the 2-point function (a) needs to be calculated explicitly.
The $ n $-point functions (b) are related to the $ ( n -1 ) $-point functions
by a recursion relation. All external momenta are
outgoing. 4-momentum
conservation implies
$
k = \sum _ { j = 1 } ^ {n - 1} k _ j
$.}
\label{fg:npoint}
\end{figure}
Now we consider the diagram with $n > 2$ external gauge field lines in
Fig.~\ref{fg:npoint}(b).
We will find that it can be recursively related to
diagrams with $ n - 1 $ external lines. To obtain
this relation we pick out one vertex, the leftmost one in
Fig.~\ref{fg:npoint}(b) which carries momentum $ k $.
The remaining ones have incoming momenta $ k _ j $
with $ j = 1, \ldots , n - 1$.
As for the 2-point function we can use the approximation
(\ref{propapp}) for the propagators.
After the partial fractioning (\ref{fractioning})
this diagram is proportional to the difference of the diagrams in which either
the vertex with $ k _ 1 $ or the vertex with $ k _ { n -1 } $ is omitted.
Each vertex carries a generator $ t ^ a $ of some gauge group. As before we do
not include the gauge couplings in the vertices. From the calculation of the
2-point function we know that in the vertices we have $ p ^ 0 \simeq p _
\| $. In analogy with Eq.~(\ref{pimunu}) we write
\begin{align}
\Pi ^ { ( n ) } { } ^{ a a _ 1 \cdots a _ {n - 1} } _ { \mu \mu _ 1 \cdots
\mu _ {n - 1} } ( k _ 1, \ldots , k _ { n-1 } )
= \int \frac{ d ^ 3 p } { ( 2 \pi ) ^ 3 } V _ \mu ( p, p - k )
{\rm tr} \left [ t ^ a
\widehat \Pi ^ { ( n ) }
{ } ^{ a _ 1 \cdots a _ {n - 1} } _ { \mu _ 1 \cdots
\mu _ {n - 1} } ( k _ 1, \ldots , k _ { n-1 }, \vec p ) \right ]
\label{52}
\end{align}
We only have to
consider the two propagators $ D ( p ) $ and $ D ( p - k ) $ which
are connected to the left vertex in Fig.~\ref{fg:npoint}(b).
We apply the same approximations as for the
2-point function including the
partial fractioning (\ref{fractioning}). The two terms in Eq.~(\ref{fractioning})
are then proportional to diagrams in which either the propagator with
momentum $ p $ or the one with momentum $ p - k $ has been omitted from
Fig.~\ref{fg:npoint}(b). Thus, if one also leaves out the vertex factors
connected to these propagators each of the two terms gives a
($ n -1 $)-point function.
For the second term in Eq.~(\ref{fractioning}) we obtain a contribution in
which the propagator with momentum $ p $ has been
omitted.
Here we perform a shift in the summation variable, $ p ^ 0 \to p ^ 0
+ k _ 1 ^ 0 $. Then the remaining propagators are the same which appear in the
$ ( n - 1 ) $-point function with incoming momenta $ k _ { 2 } , \ldots
, k _ { n - 1 } $, but with the loop 3-momentum $ \vec p $ replaced by
$ \vec p - \vec k _ 1 $.
Therefore we can write $ \widehat \Pi ^{ ( n ) }
( k _ 1, \ldots , k _ {n - 1} , \vec p ) $
in terms of the difference of
$ \widehat\Pi ^{ ( n - 1 ) } $ with either $ k _ 1 $ or $ k _ { n - 1 } $
omitted,
\begin{align}
\epsilon ( k, \vec p )
\widehat \Pi ^ { ( n ) } { } ^{ a _ 1 \cdots a _ {n - 1} }
_ { \mu _ 1 \cdots \mu _ {n - 1} }
& ( k _ 1, \ldots , k _ {n - 1} , \vec p )
\nonumber \\ =
& -
\widehat \Pi ^ { ( n - 1 ) } { } ^{ a _ 1 \cdots a _ { n - 2 }}
_ { \mu _ 1 \cdots \mu _ { n - 2 } }
( k _ 1, \ldots , k _ { n - 2 } , \vec p )
V _ { \mu _ {n - 1} } ( p - k, p - k + k _ { n -1 } ) t ^{ a _ {n -
1} }
\nonumber \\ &
{} + V _ {\mu _ 1 } ( p - k _ 1 , p )
t ^{ a _ 1 }
\widehat \Pi ^ { ( n - 1) } { } ^{ a _ 2 \cdots a _ {n - 1} }
_ { \mu _ 2 \cdots \mu _ {n - 1} }
( k _ 2, \ldots , k _ { n - 1 } , \vec p - \vec k _ 1 )
\label{53}
\end{align}
\subsection{The induced current}
\label{sc:current}
Now we attach a gauge field $ W_ \mu \equiv t^a W_\mu ^a $
to each vertex in Fig.~\ref{fg:npoint}(b) except to the one with momentum $ k $,
and then sum over all $ n $.
The result can be interpreted as the current $ J _ \mu $ which is induced by the
gauge field background. It can also be viewed as the first functional
derivative of the effective action or generating functional of our diagrams.
Working with the induced current is more convenient than with individual
diagrams because one does not have to worry about summing over all
permutations of external lines. Furthermore, it can be used for integrating
out the soft gauge fields. As in Eqs.~(\ref{pimunu}) and (\ref{52}) we write
\begin{equation}
J _ \mu ^ a ( k ) = \int \frac{ d ^ 3 p } { ( 2 \pi ) ^ 3 }
V _ \mu ( p, p - k ) {\rm tr} \left ( t ^ a \widehat J ( k, \vec p ) \right
)
\label{defofhatJ}
\end{equation}
The ``unintegrated'' current $ \widehat J $ is given by
\begin{align}
\widehat J ( k, \vec p ) = \sum _ { n = 2 } ^{ {\infty } } &
\prod _ { i = 1 } ^ {n - 1}
\left ( \int \frac{ d ^ 4 k _ i } { ( 2 \pi ) ^ 4 }
W ^{ \mu _ i } _ { a _ i } ( k _ i )
\right )
( 2 \pi ) ^ 4 \delta \left ( k - \sum _ { j = 1 } ^ {n - 1} k _ j \right )
\nonumber \\
\times & \widehat \Pi ^{ ( n ) }
{ } ^{ a _ 1 \cdots a _ {n - 1} } _ { \mu _ 1 \cdots
\mu _ {n - 1} } ( k _ 1, \ldots , k _ {n - 1} , \vec p )
\end{align}
Using Eqs.~\eqref{54} and \eqref{53} one obtains a relation
\footnote{
Fourier transformed with respect to $ k $
this relation has a similar structure as the non-abelian Vlasov \cite{blaizot93} equations from
which one obtains the generating functional of Hard Thermal Loops \cite{htl}.
The linear part of the convective term $v \cdot k$
in the Vlasov equation got replaced by
$\epsilon (k,\vec p)$.
The force term corresponds to the first term on the RHS. Finally,
the non-linear part of the covariant convective derivative
is replaced by the integral
on the RHS. In fact, if one could neglect $ k $ and $ q $ relative to
$ \vec p $, then the integral would be proportional to the Fourier transform
of the commutator $[ v \cdot W , \widehat J ]$. Unlike the Vlasov equation
our equation is non-local with two sources of non-locality. One is
the terms with $ k _ \| $ in the denominator,
and the other is due to the fact that the second $ \widehat J $ in the
integral depends on $ \vec p - \vec q $.
}
for $ \widehat J ( k, \vec p )$,
\begin{align}
\epsilon ( k, \vec p ) \widehat J ( k, \vec p )
= &
\mathcal{F}(p_\parallel,k_\parallel)
V (p - k, p ) \cdot W ( k )
\nonumber \\ &
{} - \int \frac{ d ^ 4 q } { ( 2 \pi ) ^ 4 }
\left [ \widehat J ( k - q , \vec p )
V( p - k, p - k + q ) \cdot W ( q )
\right .
\nonumber\\ &
\left.
\qquad \qquad \qquad
{} - V ( p - q, p) \cdot W ( q )
\widehat J ( k - q , \vec p - \vec q )
\right ]
\label{57}
\end{align}
\section{Integrating out soft gluons}
\label{sc:integrate}
The photon polarization tensor which enters the production rate
\eqref{prod_rate} can be obtained from the diagrams in Fig.~\ref{fg:npoint} by
identifying two external lines with photons and the remaining ones with gluons.
Connecting the gluon
vertices with propagators and integrating over the gluon momenta
will generate precisely the ladder
diagrams studied in \cite{arnoldPhoton}. In addition, it generates quark
self-energy insertions with soft gluon loops, which at LO are purely imaginary
and correspond to a thermal width. These two contributions by themselves would
be IR divergent, but their sum is IR convergent.
In terms of the current (\ref{defofhatJ}) it means that one of the background
gauge fields is the photon field and all others are gluons. The
gluon fields are integrated out, leaving only the photon.
Then the current is just $ \Pi ^{ \mu \nu }
A _ \nu $, from which one can read off the polarization tensor $ \Pi ^{ \mu
\nu } $.
We therefore distinguish between external photon and gluon fields $ A ^{ \mu
} $ and $ G ^ \mu $, and write
\begin{eqnarray}
W ^ \mu = A ^ \mu + G ^ \mu
\end{eqnarray}
The photon carries $ k _ \| $ of order $ T $, while the gluon field
has $ q _ \| \sim g T $. In order to compute the photon
polarization tensor we consider one external photon field. In $ \widehat J $ we
only need to consider terms zeroth and first order in $ A ^ \mu $, $
\widehat J = \widehat J _ 0 + \widehat J _ 1 $, and what we need to compute is
$\widehat J_1$.
In the equation for $ \widehat J _ 0 $ the function $ \mathcal {F}$
vanishes at leading order. Therefore $ \widehat J _ 0 $ is
suppressed compared to $ \widehat J _ 1 $ and it can be neglected in the
equation for $ \widehat J _ 1 $. Thus the equation for $ \widehat J _ 1$ takes
exactly the same form as Eq.~(\ref{57}) for $ \widehat J $, with $ W $
replaced by $ A $ in the inhomogeneous term, and with the two $ W $'s inside the
integral replaced by $ G $. We keep only the leading order piece of
the gluon vertex factors, so that $ V( p - k, p - k + q )
\simeq V (p - q, p ) \simeq v $ and thus
\begin{align}
\epsilon (k,\vec p) & \widehat J_1 ( k, \vec p )
=
\mathcal{F}(p_\parallel,k_\parallel)
V( p - k, p ) \cdot A ( k )
\nonumber \\ &
- \int \frac{ d ^ 4 q } { ( 2 \pi ) ^ 4 }
\left [ \widehat J_1 ( k - q , \vec p ) v \cdot G ( q )
- v \cdot G ( q )
\widehat J_1 ( k - q , p_\parallel,\vec p_\perp - \vec q _\perp ) \right ]
\label{action63}
\end{align}
Now we would like to integrate out the gluon field. We denote the
resulting current by $ \langle \widehat J _ 1 \rangle $.
In order
to see how it works we write Eq.~\eqref{action63} schematically as
$ \widehat J _ 1\sim A + G \widehat J _ 1$,
leaving out all terms and
all factors which are not relevant
for the present discussion. Now we iterate it once to obtain
$ \widehat J _ 1\sim A + G ( A + G\widehat J _ 1 ) $.
The term $ G A $
vanishes when we integrate out the gluons.
Thus we can drop this term and
write $ \widehat J _ 1 \sim A + G G \widehat J _ 1 $.
Integrating out the gluons then gives
$ \langle \widehat J _ 1 \rangle
\sim A + \langle GG \widehat J _ 1 \rangle $. The two gluon fields can either be
contracted with each other, or with the other gluon fields in $ \widehat J _ 1 $,
that is, $ \langle GG \widehat J _ 1 \rangle \sim \langle GG \rangle
\langle \widehat J _ 1 \rangle + \langle GG \widehat J _ 1 \rangle _ { \rm
connnected } $. In Appendix \ref{ap:connected} we show that the connected
part vanishes at leading order and can therefore be dropped.
Thus, by integrating out the gluons one obtains a closed
equation for $ \langle \widehat J _ 1 \rangle $ of the form
$ \langle \widehat J _ 1 \rangle \sim A +
\langle GG \rangle
\langle \widehat J _ 1 \rangle $.
Now we can become more explicit.
After iterating the integral equation once and
integrating out the soft gluons, they have disappeared as external particles
and appear only in terms of their propagator,
\begin{equation}
\left\langle G_\mu ^a(q) G_\nu ^b(q') \right\rangle = g ^ 2 \delta^{ab}
\Delta_{\mu \nu}(q) ( 2 \pi ) ^ 4 \delta(q + q')
\label{gluonprop}
\end{equation}
Since we are interested
in the electromagnetic current we put $ t ^ a = \mathds{1}$ in
Eq.~(\ref{defofhatJ}). Using ${\rm tr}
\mathds{1} = d(r)$ and $t^a t^a = C_2(r) \mathds{1}$, we obtain
\begin{eqnarray}
\lefteqn{
\epsilon ( k, \vec p ) \,\, {\rm tr} \left\langle \widehat J_1 (k,\vec p)
\right\rangle
=
d(r) \mathcal{F}(p_\parallel,k_\parallel)
V ( p - k, p ) \cdot A ( k ) } \nonumber \\
&& {} - 2 C_2(r) g ^ 2
\int\frac{ d ^ 4 q } { ( 2 \pi ) ^ 4 }
\frac{v^\mu
v^\nu \Delta_{\mu \nu}(q)}{v\cdot(k - q)} {\rm tr} \left[ \left\langle
\widehat J_1 (k,\vec p) \right\rangle - \left\langle \widehat J_1 (k,p_\parallel,\vec
p_\perp - \vec q _\perp ) \right\rangle \right] \hfill
\label{action72.1}
\end{eqnarray}
The square bracket neither depends on $ q _ 0 $ nor on $ q _ \| $, and the
integrals over $ q _ 0 $ and $ q _ \| $
can be performed.
Eq.~(\ref{action72.1}) is
thus an integral equation which determines the transverse momentum
dependence $ \mbox{tr} \langle
\widehat J _ 1 \rangle $.
Inside the integral we have approximated $ \epsilon(k - q,
\vec p ')$ \\ $ \simeq v \cdot ( k - q ) $, i.e., we have neglected the terms containing transverse
momenta and thermal masses, even though they are of the same order as the
term we kept. This is possible because the gluon
propagator separately depends on $ q _ 0 $ and $ q _ \| $, which are both of order $ g
T $, and not on their difference
$ v \cdot q $, which is of order $ g ^ 2 T
$. Therefore the terms we have omitted only contribute to a higher order
shift of the integration variable $ q _ \| $.
To
perform the integrals over $ q _ 0 $ and $ q _ \| $ in
Eq.~(\ref{action72.1}) we use again the imaginary time formalism. It means
that we replace
the integral $( 2 \pi ) ^{ -1 } \int d q _ 0 $ by a sum over Matsubara
frequencies like in
Eq.~(\ref{matsubara}). Furthermore, it means that $ k _ 0 $ in Eq.~(\ref{action72.1}) has to be
an (imaginary) Matsubara frequency. Only after performing the sum over $ q _ 0
$ one can
analytically continue $ k _ 0 $ towards the real axis. We
are interested in the production rate which is proportional to the imaginary
part of the retarded propagator. Therefore we have to give $ k _ 0 $
a small imaginary part, i.e., $ k _ 0 = \mbox{Re} ( k _ 0 ) + i \varepsilon $.
In Appendix
\ref{ap:connected} we show how we perform the Matsubara sum.
Using standard results for the HTL resummed propagators
\cite{LeBellac} and the sum rule of Ref.~\cite{Aurenche1} one then obtains
\begin{eqnarray}
T\sum _ { q _ 0 } \int \frac{ d q _ \| } { 2 \pi }
\,\,
\frac{v^\mu
v^\nu \Delta_{\mu \nu}(q)}{v\cdot(k - q)}
\simeq
\frac{ i } { 4 } T \left [ - \frac{ 1 } { \vec q _\perp ^ 2}
+ \frac{ 1 } { \vec q _\perp ^ 2 + m^2_{\rm D} } \right ]
\label{sumrule}
\end{eqnarray}
where we were able to neglect the dependence on $ v \cdot k \sim g ^ 2 T
$. That is, the only dependence on $ v \cdot k $ enters through the imaginary part
of $ k _ 0 $. The Debye mass squared enters through the gluon propagators and
is given by
\begin{equation} m^2_{\rm D} = \frac{g^2 T^2}{6} (2N + N_s
+ N_f)
\end{equation}
for a SU$(N)$ gauge theory
with $N_s$ complex scalars and $N_f$ Dirac fermions.
For the production rate \eqref{prod_rate} we need the polarization tensor
which we write as
\begin{equation} \label{pol_tensor}
\Pi_{\mu \nu}(k) = \int \frac{d^3 p}{(2\pi)^3} V_\mu(p , p - k) \widehat \Pi_{\nu}(k,\vec p)
\end{equation}
The reduced polarization tensor $\widehat \Pi_\nu$ is related to $ \widehat J
_ 1 $ through
\begin{eqnarray}
\mbox{tr} \langle \widehat J _ 1 ( k, \vec p ) \rangle = \widehat \Pi _ \nu ( k, \vec p ) A
^ \nu ( k )
\end{eqnarray}
Therefore it satisfies the integral equation
\begin{equation}
\begin{aligned}
\epsilon (k,\vec p) & \widehat \Pi_{\nu} (k,\vec p) =
d ( r ) \mathcal{F}(p_\parallel,k_\parallel) V_\nu ( p - k, p )
\hfill \\
& +
i C_2(r) g ^ 2 T
\int \frac{d^2 q _\perp }{(2\pi)^2}
\left [ \frac{ 1 } { \vec q _\perp ^ 2 }
- \frac{ 1 } { \vec q _\perp ^ 2 + m^2_{\rm D} } \right ]
\left[ \widehat \Pi_{\nu}(k,\vec p) - \widehat
\Pi_{\nu}(k,p_\parallel,\vec p_\perp -
\vec q _\perp ) \right ] \vphantom{\int} \hfill
\end{aligned}
\label{inteqphoton}
\end{equation}
We finally want to show that our integral equation \eqref{inteqphoton} can be reduced to the one
in Ref.~\cite{arnoldPhoton} for the production of real photons.
We choose $ \vec v $ in the direction
of $ \vec k $ so that $ \vec k _\perp = 0 $. Then we have $ v \cdot k = 0 $,
and thus $ \epsilon ( k , \vec p ) = ( \vec p _\perp ^ 2 + m ^ 2 )k _ \|
/[2 p _ \| ( p _ \| - k _ \| ) ] $. Only the transverse
components of $\widehat \Pi_\nu$ contribute since the polarization vectors in
\eqref{prod_rate} are purely transverse.
If we define a new function $\vec f$ via
\begin{equation} \label{Identification}
\widehat { \vec \Pi } _\perp(k,\vec p) = \frac{
id(r)\mathcal{F}(p_\parallel,k_\parallel) }
{2 ( p _ \| - k _ \| ) }
\vec f (k,\vec p)
\end{equation}
we obtain the equation of the same form as in \cite{arnoldPhoton}
\begin{align}
2 \vec p _\perp & = i\epsilon(k,\vec p) \vec f(k,\vec p)
\hfill \nonumber \\
& {} + g ^2 C_2(r) T
\int \frac{d^2 q_\perp}{(2\pi)^2} \left [ \frac{1}{\vec q_\perp ^2} -
\frac{1}{\vec q_\perp ^2 + m_ { \rm D } ^2} \right ]
[ \vec f(k,\vec p) - \vec f(k,p_\parallel,\vec p_\perp -
\vec q_\perp) ] \hfill
\end{align}
\section{Spin-1/2 quarks}
\label{sc:fermionic}
So far we have always been dealing with scalar quarks to avoid technical
complications. There is no need to
redo the entire calculation for spin-1/2 quarks. The results above are still valid up to a few
modifications which we now describe.
We now have to deal with
the resummed fermion propagator
\begin{equation}
S(p) = \frac{-1}{\cancel{p} - \Sigma(p)}
\end{equation}
in the high temperature limit when the zero temperature mass can be
neglected.
In the plasma rest frame the self-energy $ \Sigma(p)$ takes the general form
\cite{weldon}
\begin{equation} \label{Sigma_Ansatz}
\Sigma(p) = a(p) \cancel{p} + b(p) \gamma^0
\end{equation}
Therefore chiral symmetry is not broken by thermal effects, and the
left- and right-chiral fermions propagate independently.
Since $ \Sigma(p) \sim g^2 T$, we have $ a(p) \sim g^2$.
Thus $ a ( p ) \cancel{p} $ is small compared to the tree level
contribution and may be neglected.
For $ p \sim T $, $ p ^ 2 \sim g ^ 2 T ^ 2 $ the propagator can then be written as
\begin{equation} \label{FermionProp}
S(p) \simeq - \frac{ \cancel{p} + b(p) \gamma^0
}
{ p^2 - 2 b(p) p^0}
\simeq - \frac{ \cancel{p}
}
{ p^2 - m ^ 2 }
\end{equation}
Here we have neglected terms of order $ g ^ 2 T $ in the numerator and
we have identified
\begin{equation}
b(p) = \frac{m^2}{2p^0}
\end{equation}
at $ p ^ 2 = m ^ 2 $, where $ m $ is the asymptotic thermal mass in
Eq.~(\ref{mass}).
Since the left and right-handed fermions propagate independently, one
may consider the photon production from left-handed quarks only. The
complete rate is then twice as large. Thus one can deal with Weyl
instead of Dirac spinors, the vertices contain $ \bar \sigma ^ \mu
$ instead of $ \gamma ^ \mu $, and the propagator (\ref{FermionProp})
contains $ \sigma \cdot p $ instead of $ \cancel{ p } $. Up to terms
of order $ g ^ 2 T $, which we neglect in the numerator, $ p ^ \mu $ is
light-like. Therefore we can write
\begin{align}
\sigma \cdot p \simeq 2 p _ \| \eta ( p ) \eta ^\dagger ( p )
\end{align}
where $ \eta ( p ) $ is a normalized eigenvector of $ \vec p \cdot \vec \sigma
/p _ \| $ with negative eigenvalue.
Thus we find, similarly to Ref.~\cite{arnoldPhoton},
\begin{align}
S _ L ( p )
\simeq
\eta ( p ) \eta ^\dagger ( p ) D ( p )
\label{fermionpropagator}
\end{align}
with the same $ D ( p ) $ as in Eq.~(\ref{D}).
Note that, unlike in
Ref.~\cite{arnoldPhoton}, our result is valid for
both signs of $ p _ \| $.
We associate the spinors $ \eta ( p ) $ and $
\eta ^\dagger ( p ) $ with the vertices on either side of the propagator
rather than with the propagator itself. Therefore the vertex factor now reads
\begin{align}
V ^ \mu ( p, p - k ) = \eta ^\dagger ( p - k ) \bar \sigma ^ \mu \eta
( p )
\end{align}
instead of Eq.~(\ref{vertexfactor}). For real photon production one needs $ V
$ up to order $ g $,
\begin{align}
V ^ \mu ( p, p - k ) = v ^ \mu + V ^ \mu _\perp + O ( g ^ 2 )
\end{align}
In the helicity basis the transverse components $ V ^ \mu _\perp $ are
particularly simple. We choose $ \vec v $ as the 3-direction. Then
for $ V ^{ \pm } \equiv ( V ^ 1 \pm i V ^ 2 )/\sqrt{ 2 } $ one
finds
\begin{align}
V ^ + = \frac{ p ^ + } { p _ \| - k _ \| } + O ( g ^ 2 ) , \qquad
V ^ - = \frac{ p ^ - } { p _ \| } + O ( g ^ 2 )
\end{align}
The other difference compared to scalar quarks is due to the Fermi-Dirac
statistics, the Bose-Einstein functions get
replaced by Fermi-Dirac distributions, so that $ \mathcal{F} $ becomes
\begin{align}
\mathcal{F}(p _ \|, k _ \| ) = f _ {\rm F}(p_\parallel)
- f _ {\rm F} (p_\parallel - k_\parallel)
\label{fermi}
\end{align}
\section{Summary and Conclusions}
\label{sc:conclusions}
In this paper we have obtained an integral equation (Eq.~(\ref{57})) which sums all thermal
1-loop diagrams with an arbitrary number of soft or collinear external gauge
fields. We have applied it to compute the rate for real photon production by
bremsstrahlung and pair annihilation in a hot QCD plasma.
Compared to the original calculation of the photon production rate in
Ref.~\cite{arnoldPhoton} our approach is
significantly simplified by the fact the calculation is done in two
steps. In the first step we have integrated out the hard momentum modes at one
loop. The resulting effective theory is summarized by Eq.~(\ref{57}). It has a
similar structure as the non-abelian Vlasov equation which describes the Hard
Thermal Loops for soft external gauge fields. In a second step we have
integrated out the soft gauge fields corresponding to gluons. This results in the
integral equation obtained earlier in Ref.~\cite{arnoldPhoton} which sums all leading order
ladder and self energy contributions to the photon polarization tensor for
hard on-shell photons, and which thus describes the Landau-Pomeranchuk-Migdal
effect on thermal photon production.
Our approach should easily allow for generalizations. The
method can be adopted to the production of other particles than photons,
e.g. the production of spin-1/2-fermions.
We also hope for a possible generalization to non-equilibrium
situations, as they occur e.g. in heavy ion collision.
{\bf Acknowledgments}
This work was supported in part through the
DFG funded Graduate School GRK 881.
|
Subsets and Splits